Lecture Notes

in Computational Science
and Engineering 73
Editors
Timothy J. Barth
Michael Griebel
David E. Keyes
Risto M. Nieminen
Dirk Roose
Tamar Schlick

For further volumes:

http://www.springer.com/series/3527
•
Hans-Joachim Bungartz ● Miriam Mehl
Michael Schäfer
Editors

Fluid Structure Interaction II

Modelling, Simulation, Optimization

123
Editors
Hans-Joachim Bungartz Michael Schäfer
Miriam Mehl Technische Universität Darmstadt
Technische Universität München Numerische Berechnungsverfahren
Institut für Informatik Dolivostr. 15
Boltzmannstr. 3 64293 Darmstadt
85748 Garching Germany
Germany [email protected]
[email protected]
[email protected]

ISSN 1439-7358
ISBN 978-3-642-14205-5 e-ISBN 978-3-642-14206-2
DOI: 10.1007/978-3-642-14206-2
Springer Heidelberg Dordrecht London New York

Library of Congress Control Number: 2010936632

Mathematics Subject Classification Numbers (2010): 65-06, 65K10, 65Mxx, 65Nxx, 65Y05, 74F10,
74Sxx, 76D05, 76Fxx

© Springer-Verlag Berlin Heidelberg 2010

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current version, and permission for use must always be obtained from Springer. Violations are
liable to prosecution under the German Copyright Law.
The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply,
even in the absence of a specific statement, that such names are exempt from the relevant protective laws
and regulations and therefore free for general use.

Cover design: deblik, Berlin

Printed on acid-free paper

Springer is part of Springer Science + Business Media (www.springer.com)

Preface

Modern Computational Science and Engineering (CSE) is confronted with several

challenges of “multi-type”: Multi-physics problems involve more than one phys-
ical effect; multi-scale models involve different scales with respect to space or
time; multi-level methods are needed to efficiently tackle large linear systems;
multi-core architectures require a new access to parallelism; and much of the
research in CSE requires the collaboration of experts from several disciplines –
it is multi-disciplinary. Concerning the first issue, multi-physics problems such as
fluid-structure interactions (FSI), i. e. the interplay of some moveable or deformable
structure with an internal or surrounding flow field, are one of the most relevant
and most intensely studied coupled problems. Despite this high attention, FSI are
still not completely understood, and there is an obvious lack of reliable, robust, and
efficient computational methods.
Furthermore, there is a somewhat astonishing discrepancy between, on the one
hand, how complex specific scenarios have already been successfully simulated
(think of airbags or parachutes, e. g.) and, on the other hand, how big the problems
are that occur when those codes shall be used for different problems. Hence, there
hasn’t been any widely accepted numerical benchmark for FSI before this volume’s
predecessor LNCSE 53 in 2006. Also experimental validation has turned out to be
far from trivial: either the experimental setting is too complicated for the numeri-
cal tools, or the numerical setting is not feasible for experiments; if, finally, both
experiments and numerical simulations can deal with a certain scenario, the effects
intended to study often do not show up. All this shows that there are still challeng-
ing questions in FSI research, ranging from modelling via numerical treatment up
to implementation and software tools – and only their ensemble provides a key to
deeper insight in FSI.
The present volume contains selected contributions from the “First International
Workshop on Computational Engineering – special topic Fluid-Structure Interac-
tions” held in Herrsching, Germany, in October 2009. This three-day workshop
was jointly organized by three initiatives funded by the German Research Foun-
dation (Deutsche Forschungsgemeinschaft, DFG) – the “International Graduate
School of Computational Engineering” in Darmstadt, the “International Graduate
School of Science and Engineering” in Munich, and the Research Unit 493 “Fluid-
Structure Interaction: Modelling, Simulation, Optimization” (FOR 493). FOR 493

v
vi Preface

was established by the DFG in 2003. In this framework, researchers from seven
German universities working in the fields of mathematics, informatics, mechanical
engineering, chemical engineering, or civil engineering joined their forces to push
forward the state-of-the-art of fundamental computational research in FSI. Designed
as a forum for latest results in computational engineering in general and FSI in par-
ticular, the workshop in Herrsching brought together leading experts from all over
the world, allowing for three highly interesting days of tutorials, invited lectures, and
minisymposia, and, now, resulting in the fifteen papers collected in this volume –
which is the second book on FSI published by our Research Unit FOR 493.
We would like to thank the editors of Springer’s Lecture Notes in Computational
Science and Engineering (LNCSE) for admitting our collection to this series for the
second time, as well as Springer Verlag and, in particular, Dr. Martin Peters, for
their most valuable support from the first idea to the final layout. Furthermore, we
are deeply obliged to Michael Lieb, who did a great job in compiling the single
contributions to a finally harmonic ensemble. Last, but not least, we want to express
our thanks to DFG for more than six years of ongoing funding. Without this sup-
port, neither many of the results presented in the contributions of this book nor this
volume itself would have become reality.

Munich and Darmstadt Hans-Joachim Bungartz

May 2010 Miriam Mehl
Michael Schäfer
Contents

Multi-Level Accelerated Sub-Iterations for Fluid-Structure

Interaction . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . 1
A.H. van Zuijlen and H. Bijl

A Classification of Interface Treatments for FSI . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . 27

C.A. Felippa, K.C. Park, and M.R. Ross

Computer Modeling and Analysis of the Orion Spacecraft

Parachutes . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . 53
K. Takizawa, C. Moorman, S. Wright, and T.E. Tezduyar

Stability Issues in Partitioned FSI Calculations . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . 83

J. Vierendeels, J. Degroote, S. Annerel, and R. Haelterman

Hydroelastic Analysis and Response of Pontoon-Type

Very Large Floating Structures.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .103
C.M. Wang and Z.Y. Tay

Efficient Numerical Simulation and Optimization

of Fluid-Structure Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .131
M. Schäfer, D.C. Sternel, G. Becker, and P. Pironkov

An Adaptive Finite Element Method for Fluid-Structure

Interaction Problems Based on a Fully Eulerian Formulation.. . . . . . . .. . . . . . .159
R. Rannacher and T. Richter

Numerical Simulation and Benchmarking of a Monolithic

Multigrid Solver for Fluid-Structure Interaction Problems
with Application to Hemodynamics .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .193
S. Turek, J. Hron, M. Mádlı́k, M. Razzaq, H. Wobker,
and J.F. Acker

vii
viii Contents

Numerical Simulation of Fluid–Structure Interaction

Using Eddy–Resolving Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .221
M. Münsch and M. Breuer

Partitioned Simulation of Fluid-Structure Interaction

on Cartesian Grids . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .255
H.-J. Bungartz, J. Benk, B. Gatzhammer, M. Mehl, and T. Neckel

An Explicit Model for Three-Dimensional Fluid-Structure

Interaction using LBM and p-FEM. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .285
S. Geller, S. Kollmannsberger, M. El Bettah, M. Krafczyk,
D. Scholz, A. Düster, and E. Rank

An XFEM Based Fixed-Grid Approach for 3D Fluid-Structure

Interaction . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .327
W.A. Wall, A. Gerstenberger, U. Küttler, and U.M. Mayer

Fluid-Structure Interaction in the Context of Shape

Optimization and Computational Wind Engineering . . . . . . . . . . . . . . . . . . .. . . . . . .351
M. Hojjat, E. Stavropoulou, T. Gallinger, U. Israel, R. Wüchner,
and K.-U. Bletzinger

Experimental Benchmark: Self-Excited Fluid-Structure

Interaction Test Cases . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .383
J. Pereira Gomes and H. Lienhart

Numerical Benchmarking of Fluid-Structure Interaction:

A Comparison of Different Discretization and Solution
Approaches . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .413
S. Turek, J. Hron, M. Razzaq, H. Wobker, and M. Schäfer
•
Multi-Level Accelerated Sub-Iterations
for Fluid-Structure Interaction

A.H. van Zuijlen and H. Bijl

Abstract Computational fluid-structure interaction is most commonly performed

using a partitioned approach. For strongly coupled problems sub-iterations are
required, increasing computational time as flow and structure have to be resolved
multiple times every time step. Many sub-iteration techniques exist that improve
robustness and convergence, although still flow and structure problems have to be
solved a number of times every time step. In this paper we apply a multilevel accel-
eration technique, which is based on the presumed existing multigrid solver for the
flow domain, to a two-dimensional strongly coupled laminar and turbulent problem
and investigate the combination of multilevel acceleration with the Aitken underre-
laxtion technique. It is found that the value for the under-relaxation parameter is not
significantly different when performing sub-iterations purely on the coarse level or
purely on the fine level. Therefore coarse and fine level sub-iterations are used alter-
nately, where it is found that performing 3 or 4 coarse level sub-iterations followed
by 1 fine level sub-iteration resulted in the highest gain in efficiency. Although the
total number of sub-iterations increases slightly by 30%, the number of fine grid
iterations can be decreased by as much as 65–70%.

1 Introduction

In many engineering applications, fluid-structure interaction phenomena play a key

role in the dynamic stability of a structure (e.g. aircraft, wind-turbines, suspension
bridges, etc.). A fast and accurate computation of the dynamic interaction between

A.H. van Zuijlen

Delft University of Technology, Kluyverweg 1, 2629HS, Delft, The Netherlands
e-mail: [email protected]
H. Bijl
Delft University of Technology, Kluyverweg 1, 2629HS, Delft, The Netherlands
e-mail: [email protected]

H.-J. Bungartz et al. (eds.), Fluid Structure Interaction II, Lecture Notes 1
in Computational Science and Engineering 73, DOI 10.1007/978-3-642-14206-2 1,
c Springer-Verlag Berlin Heidelberg 2010
2 A.H. van Zuijlen and H. Bijl

flow and structure is therefore of the utmost importance. For fluid-structure inter-
action often a partitioned approach is chosen to resolve the coupled problem, as it
allows reusing existing flow and structure solvers and independent development and
optimization of the codes. The drawback of a partitioned approach over a mono-
lithic approach is that the coupling between the flow and structure domain needs
additional attention and for strongly coupled physical problems, sub-iterations are
required. Sub-iterating increases the computational expense as flow and structure
have to be resolved multiple times each time step. In order to reduce computational
costs, we propose a multilevel algorithm for reducing the costs of particularly the
flow solver.
Basic sub-iteration techniques include block-Gauss-Seidel iterations, which may
suffer instability or fixed under-relaxation methods, which are robust, but at the
price of slower convergence. In literature several methods can be found for per-
forming sub-iterations in an efficient and robust fashion. Especially when the flow
is incompressible, robustness of the sub-iteration technique is an issue [3]. One of
the most popular methods is the Aitken under-relaxation method [6], that tunes
the under-relaxation parameter to obtain faster convergence. Methods that require
more implementation effort are e.g. interface GMRES [5] or reduced order mod-
eling [9]. They have the advantage over the Aitken method that they have faster
convergence. Recently a quasi Newton method has been proposed [4] and applied to
fluid-structure interaction coupling [2] that has a similar implementation complexity
as the Aitken method but with superior convergence.
Whichever method is chosen, a (small) number of sub-iterations still has to be
performed. In these sub-iterations solving the fluid dynamic equations is generally
the most computationally expensive. In previous research [11] we investigated the
possibility of performing sub-iterations initially on a coarse level and showed that
a basic coarse level block-Gauss-Seidel sub-iteration would have much the same
convergence as sub-iterating on the fine level.
In this paper we apply this model to a more challenging two-dimensional strongly
coupled laminar [8] and turbulent problem and investigate the combination of mul-
tilevel acceleration with Aitken under-relaxation. We investigate the effectiveness
of performing initial sub-iterations on a coarse level and alternating coarse and fine
level sub-iterations.

2 Coupled problem

In this paper a fluid-structure interaction problem is addressed which is based on

the benchmark problem by Turek [8]. The original problem consists of an incom-
pressible fluid around a circular cylinder with a flexible trailing flap. The coupled
problem consists of a fluid domain ˝f , see Fig. 1, which is here modeled as a com-
pressible fluid, and a structure domain ˝s which is modeled as a linear elastic body.
The boundaries of the domain are given by s , f , which result in boundary con-
ditions for the structure and fluid dynamics respectively. In the partitioned approach
Multi-Level Accelerated Sub-Iterations for Fluid-Structure Interaction 3

Fig. 1 Computational domain.

the interface boundary between the fluid and the structure domains I is denoted
by two boundaries sf and f s which close the structure and fluid domains so that
each domain can be treated separately from the other. The coupling between flow
and structure is introduced in the boundary conditions that are imposed on sf and
f s and that should yield continuity of displacement (of the interface) and stresses.
Since the flap has the highest flexibility in y-direction and the shear stresses mainly
act in x-direction, we simplify the continuity of stresses to a continuity in pressure
so that the conditions at the interface are

df s D dsf ; (1)

psf D pf s ; (2)

where d denotes the displacement of the interface boundary and p the pressure. At
the moment it is still assumed that the spatial coupling is continuous and the tem-
poral coupling instantaneous. Since the domains have been split and the coupling is
performed using boundary conditions, readily available flow and structure solvers
can be used to discretize and resolve their own dynamics on their own domains.
Therefore, we do not address the specific spatial discretization of the solvers and
simply write

dws
C Ds .ws ; psf / D Ss ; (3)
dt
dwf
C Df .wf ; df s / D Sf ; (4)
dt
wherein ws and wf are the discrete state vectors for the structure state and fluid
state respectively. They contain e.g. the structural displacement or the fluid density.
The spatial discretization of the governing equation is simplified by the operator D,
which depends both on the state w and on the fluid-structure interface conditions
psf (the discrete pressure acting on the structure) and df s (the displacement of
the fluid domain boundary). The right hand side may contain terms S that may arise
from boundary conditions on s and f . Equations (3) and (4) are in semi-discrete
form. We assume that the time integration is performed by the same implicit scheme
in both domains. The structure and flow solver programs can then be described
as solution techniques that can find solutions wnC1
s and wnC1
f under the boundary
4 A.H. van Zuijlen and H. Bijl

conditions psf and df s such that they satisfy (or minimize)

rs .wnC1
s ; psf /  ss D 0; (5)
rf .wnC1
f ; df s /  sf D 0; (6)

wherein r the residual function (discretized representation of the governing equa-

tions), s a constant source term within the time step that can depend on e.g.
previous solutions or boundary conditions, p the discrete pressures in the boundary
nodes and d the discrete displacements of the boundary nodes. The subscript s; f
denotes that the discrete quantities belong to the structure and fluid domains respec-
tively. A Computational Structure Dynamics (CSD) package is capable of finding
a wnC1
s such that (5) is satisfied for a given pressure load ps f . A Computational
Fluid Dynamics (CFD) package is able to find a wnC1 f
such that (6) is satisfied for
a given boundary displacement df s . A fully implicit (or fully coupled) solution
would require

rs .wnC1
s ; pnC1
sf /  ss D 0; (7)
rf .wnC1
f ; dnC1
f s /  sf D 0; (8)

wherein the superscript nC1 denotes the discrete solution at the new time level tnC1 .
The coupling between (7) and (8) now poses a problem in a partitioned approach as
the pressure acting on the structure interface pnC1
sf depends on the fluid state wf
nC1

and the displacement of the fluid boundary dnC1

f s depends on the structure state
nC1
ws . Both the spatial coupling (transferring data from the flow to the structure
mesh and vice versa) and the temporal coupling (obtaining an implicitly coupled
solution) are addressed in the next sections.

2.1 Spatial coupling

The coupling between flow and structure takes place at the fluid-structure bound-
ary. In the continuous case, this boundary I would be identical for both fluid and
structure domains, however, at the discrete level, the boundary sf and f s do
not have to be matching and gaps or overlaps may occur. In [1] several interpola-
tion techniques have been studied for transfer of displacements and pressures at the
fluid-structure interface. It was concluded that a consistent mesh interpolation was
preferred over the conservative interpolation method. For the consistent interpola-
tion, two separate interpolations are defined for the transfer from fluid to structure
mesh and from structure mesh to fluid mesh. The interpolation methods do not have
to be the same kind, e.g. a radial basis function interpolation can be used for transfer
of displacements from the structure to the flow mesh and a simple nearest neighbor
algorithm can be used to transfer pressures from the flow mesh to the structure mesh.
Multi-Level Accelerated Sub-Iterations for Fluid-Structure Interaction 5

For the conservative approach on the other hand one can choose an interpolation
method for transferring displacements from the structure to the flow, but one has
to use the transposed of the interpolation to transfer forces from the flow to the
structure. First the coupling of displacements is performed by transferring displace-
ments from the discrete structure boundary to the discrete flow boundary through an
interpolation If s
df s D If s .dsf /: (9)
The displacements at the structure boundary follow directly from the structure state
vector ds D g.ws /. When the structure state vector ws already contains the dis-
placements of the boundary nodes as part of its degrees-of-freedom, g could simply
be represented by a Boolean matrix that extracts only the boundary displacement
from the structure state vector.
The second part of the coupling is the transfer of pressure loads from the flow to
the structure by an interpolation Isf

ps f D Isf .pf s /; (10)

and the pressure at the fluid boundary follows directly from the fluid state vector
pf D f .wf /. This time the function f may be more complicated as the fluid state
wf can be defined in cell centers (for a cell-centered finite volume solver), whereas
the pressures on the interface may be defined in face centers or vertex locations.
In that case an interpolation from cell centered values to boundary values is also
taking place in the function f . Additionally, the fluid state vector may only contain
the conservative variables and not the primitive variable p, in which case f also
computes the pressure from the conservative variables.

2.2 Temporal coupling

In partitioned fluid-structure interaction, obtaining the coupled solution described

by (7) and (8) would require sub-iterating, e.g. when a sequential algorithm is used

rs .wis ; pO isf /  ss D 0; (11)

rf .wif ; If s .g.wis ///  sf D 0; (12)

wherein the superscript i denotes the i -th sub-iteration and pO isf is the estimation of
the fluid pressure acting on the structure for the i -th iteration. The simplest choice
for the estimation is
pO isf D Isf .g.wfi 1 //; (13)
which results in a block-Gauss-Seidel type of iteration, but which is not guaranteed
to be stable. To increase robustness under-relaxation can be applied, but generally at
the expense of slower convergence rate. In this paper we focus on the widely applied
6 A.H. van Zuijlen and H. Bijl

Aitken method [6], which applies an adaptive under-relaxation to the estimation for
the next time step
pO iC1
sf
D pO isf C  i C1 .pisf  pO isf /; (14)

for which the under-relaxation parameter  i C1 is obtained from

 
i C1 i .ei /T .ei /
 D 1 ; (15)
.ei /T .ei /

with ei D pisf  pO isf the error between the estimated and the resulting pressure
after solving (11) and (12) for iteration i and ei D ei  ei 1 . For the first under-
relaxation step a  has to be chosen as ei 1 is not available yet. One can use last
known value and at the very start of the computation any (sufficiently small) value
can be taken.

3 Two-level acceleration

In order to reduce the computational time for the (initial) sub-iterations, the two-
level acceleration scheme solves for a correction of the solution on a coarsened
mesh. The way the method works is best demonstrated on a linear system first.

3.1 Linear fine grid operator

Therefore, the coupled system (7) and (8) is written as a linear system
   
wh;s sh;s
Lh  D 0; (16)
wh;f sh;f

wherein Lh is a linear operator (constant matrix) which consists of

 
Ah;s Ah;sf Ih;sf fh
Lh D ; (17)
Ah;f s Ih;f s gh Ah;f

which contains the discretization of the governing equations in the structure Ah;s
and fluid Ah;f domains and the discretization of the boundary conditions Ah;sf
and Ah;f s . The coupling also depends on the (linear) interpolation algorithms at the
fluid-structure interface (Ih;sf and Ih;f s ) and the interpolation functions fh and gh .
Suppose the approximation of the coupled solution is denoted by .w O h;f /T , we
O h;s w
obtain a residual rh    
wO h;s sh;s
Lh  D rO h : (18)
O h;f
w sh;f
Multi-Level Accelerated Sub-Iterations for Fluid-Structure Interaction 7

Subtraction of (18) from (16) yields

 
O h;s
wh;s  w
Lh D Orh ; (19)
O h;f
wh;f  w

which can written as

Lh "h D Orh ; (20)
wherein the correction "h is defined as
 
O h;s
wh;s  w
"h D : (21)
O h;f
wh;f  w

Instead of solving the original coupled equation (16), the coupled system (19) term
can be solved to obtain the correction term "h . This obviously involves solving a
coupled system of the same complexity as the original equation. Following the stan-
dard multigrid approach, the correction term is computed on a coarse mesh which
reduces computational costs.

3.2 Linear coarse grid operator

The coarse level system is obtained by restriction R and prolongation P operators

commonly used in the multigrid algorithm. Prolongation and restriction are only
applied to flow and structure domains individually, e.g.
 
Rs 0
RD ; (22)
0 Rf
 
Ps 0
P D ; (23)
0 Pf

so that there does not exist a coarse level variable that consists of a combination
of a fine grid flow variable and fine grid structure variable. Left multiplication of
(19) with R and denoting the correction on the fine mesh as a prolongation of the
correction on the coarse mesh "h D P "H yields

RLh P "H D ROrh ; (24)

wherein the coarse grid operator LH D RLh P is

 
Rs Ah;s Ps Rs Ah;sf Ih;sf fh Pf
LH D : (25)
Rf Ah;f s Ih;f s gh Ps Rf Ah;f Pf
8 A.H. van Zuijlen and H. Bijl

Several remarks can be made with respect to the coarse grid coupled system to be
solved in (24): first of all the right-hand-side term is directly related to the resid-
ual from the fine mesh, and therefore, when the fine grid solution is converged, the
coarse grid correction reduces to zero as well. Secondly, the coarse level operator is
often constructed by simply making the discretization of the governing equations on
a coarse grid, instead of constructing it from the fine grid operator by restriction and
prolongation, e.g. use AH;f instead of RAh;f P . The discretization of AH;f does
not necessarily have to be done by the same discretization, e.g. the fine grid dis-
cretization could be a second order central discretization, whereas the coarse grid
discretization could be based on a first order upwind scheme. Thirdly, the coarse
grid coupling terms (off-diagonal blocks in (25)), may be interpreted as follows: the
coupling term Rs Ah;sf Ih;sf fh Pf denotes that when we have a coarse grid correc-
tion term in the flow domain "H;f , that this term is prolonged to the fine mesh level
"h;f and its influence on the boundary value for the pressure ıph;f s is computed
through fh . This change in the boundary pressure field is interpolated to the fine
grid level of the structure with the Ih;sf operator to obtain ıph;sf and finally, its
influence Ah;sf ıph;sf on the coarse structure level is obtained by restriction to the
coarse level. This shows that the coupling at the fluid structure interface (both for
the mesh interpolation as the functions f and g takes place on the fine grid level.
Alternatively one could make a new coupling on the coarse level AH;sf IH;sf fH ,
which requires the definition of an interpolation between the coarse level meshes
IH;sf and fH . The two options, performing the coarse level coupling through the
fine grid level and performing the coarse level coupling on the coarse grid level
itself, are schematically depicted in Fig. 2(a) and Fig. 2(b) respectively. The defini-
tion of the coarse grid operator (25) reflects the coupling as depicted in Fig. 2(a),
whereas Fig. 2(b) is represented by the coarse grid operator
 
Rs Ah;s Ps AH;sf IH;sf fH
LQ H D : (26)
AH;f s IH;f s gH Rf Ah;f Pf

In this paper only the first option is explored in further detail as the second option
may have a drawback in the accuracy of the coupling when the interface data is
not constructed accurately using fH as mentioned in Sect. 4.1.1. In any case, the
resulting coarse level system is a coupled system again. Resolving it would require
again a partitioned or monolithic approach.

3.3 Coarse level sub-iterations

For the same reason as that the fine level was resolved in a partitioned fashion,
the coarse level system is also resolved in a partitioned fashion, e.g. for a standard
Gauss-Seidel iteration
Multi-Level Accelerated Sub-Iterations for Fluid-Structure Interaction 9

(a) Coupling through fine grid level.

(b) Coupling through coarse grid level.

Fig. 2 Schematic representation of coupled two-level algorithm.

  i
Rs Ah;s Ps 0 "H;s
Rf Ah;f s Ih;f s gh Ps Rf Ah;f Pf "H;f
  i 1
0 0 Rs Ah;sf Ih;sf fh Pf "H;s
D ROrh  ; (27)
0 0 "H;f

or the equivalent
  i
Rs Ah;s Ps 0 ı"H;s
D ROrhi 1 ; (28)
Rf Ah;f s Ih;f s gh Ps Rf Ah;f Pf ı"H;f
10 A.H. van Zuijlen and H. Bijl

wherein ı"iH D "iH  "H i 1

, the change in correction term for the next sub-iteration
i 1 O hi 1 .
and rO h the fine grid residual obtained for the estimated fine grid solution w
The two options for subiterating on the coarse level (27) and (28) are explained in
more detail.

Coarse grid sub-iteration options

The original equation to solve on the fine grid reads

Lh wh C sh D 0: (29)

Assume a first estimation at iteration 0 w0h , so that the first residual becomes

r0h D Lh w0h C sh : (30)

The coarse grid operator is obtained by restriction and prolongation LH D RLh P

and split into two matrices for the partitioned iterations on the coarse grid LH D
LH;L C LH;R , where LH;L remains on the left-hand-side of the equation and LH;R
is taken to the right-hand-side. Using the initialization "0H D 0, the first iteration on
the coarse grid following (27) provides the first coarse grid correction term

LH;L "1H D Rr0h  LH;R 0; (31)

or
"1H D L1 0
H;L Rrh : (32)

Iterating only on coarse grid

A second iteration according to (27) gives

LH;L "2H D Rr0h  LH;R "1H ; (33)

or
 0 
"2H D L1 1
H;L Rrh C LH;R "H ;

D "1H  L1 1
H;L LH;R "H ; (34)

so that after 2 coarse grid iterations the correction solution on the fine grid yields

w2h D w0h C P "2H D w0h C P "1H  PL1 1

H;L LH;R "H : (35)
Multi-Level Accelerated Sub-Iterations for Fluid-Structure Interaction 11

Iterating through fine grid

Starting from the first correction term (32), the solution on the fine grid is updated

w1h D w0h C P "1H : (36)

For the new estimation, the residual of the system is obtained

r1h D Lh w1h C sh
 
D Lh w0h C P "1H C sh
D r0h C Lh P "1H ; (37)

and the correction term of the previous iteration on the coarse grid is not required
when one solves for the second iteration on the coarse grid

LH;L ı"2H D Rr1h ; (38)

since this can be written as

ı"2H D L1 1
H;L Rrh
 
D L1 0 1
H;L R rh C Lh P "H

D L1 0 1 1
H;L Rrh  LH;L RLh P "H ; (39)

and since LH D RLh P and using (32) one obtains

ı"2H D "1H  L1 1

H;L LH "H

D "1H  L1 1
H;L .LH;L C LH;R /"H
D "1H  "1H  L1 1
H;L LH;R "H

D L1 1
H;L LH;R "H ; (40)

so that the solution on the fine grid after two coarse grid iterations becomes

w2h D w1h C P ı"2H D w0h C P "1H  PL1 1

H;L LH;R "H ; (41)

which shows that the same solution is obtained as with method one (35), although
no information needs to be retained at the coarse grid level to perform consecutive
iterations.
The advantage of the second option (28) is that there is no need to “remember”
the value of the coarse grid correction term to take into account in the right-hand-
side of (27). Sub-iterating on the coarse level may be unstable as well when the
fine grid sub-iterations are unstable. Therefore, a more advanced sub-iteration tech-
nique is required on the coarse level as well. Here we choose the same sub-iteration
12 A.H. van Zuijlen and H. Bijl

technique (Aitken) as applied for the fine level sub-iterations. Under-relaxation

could then be applied to the correction term "H , but we choose to apply the under-
relaxation to the equivalent fine level variable ph;sf (14). Essentially the coarse
grid correction term in the fluid domain is prolonged to the fine grid to update the
estimated solution on the fine grid

wih;f D wh;f
i 1
C P "ih;f : (42)

Next, the pressure at the fluid boundary is updated using the function fh

pih;f s D fh wih;f ; (43)

which allows the transfer of the boundary pressure to the structure mesh by

pih;sf D Isf pih;f s ; (44)

to which the Aitken under-relaxation is applied (14) to obtain the new estimation of
the pressure on the structure mesh

pO ih;
C1
sf
D pO ih;sf C  i C1 .pih;sf  pO ih;sf /; (45)

for which the under-relaxation  is computed by (15). Note that the computa-
tion of the Aitken under-relaxation does not differ in any sense from the Aitken
under-relaxation applied directly to fine level sub-iterations, i.e. the Aitken algo-
rithm cannot “distinguish” between coarse and fine level sub-iterations. Therefore,
coarse and fine level iterations can alternate each other without modification to the
Aitken algorithm. The question remains whether the computation of the underrelax-
tion parameter  by alternating coarse and fine level iterations affects the stability
and convergence properties of the Aitken method.

3.4 Nonlinear coarse level flow sub-iterations

In our applications, the structure requires only a small amount of computational time
to solve compared to the flow solver. Therefore, only the fluid domain is coarsened,
so that Ps D Rs D I . The coarse level sub-iteration (28) is now written as
  i !
i 1
Ah;s 0 ı"h;s rO h;s
D i 1 ; (46)
Rf Ah;f s Ih;f s gh Rf Ah;f Pf ı"H;f ROrh;f ;

which is graphically represented in Fig. 3. So far the system is still assumed to

be linear. To find an equivalent when non-linear operators (e.g. Af ) are used, the
change in correction is written in terms of the flow and structure states

ı"i D "i  "i 1 D wi  wi 1 ; (47)

Multi-Level Accelerated Sub-Iterations for Fluid-Structure Interaction 13

Fig. 3 Schematic representation of coarse grid flow – fine grid structure coupling.

and (46) is written in two consecutive steps

 
Ah;s wih;s  wh;s
i 1 i 1
D Orh;s ; (48)
  h  i
Rf Ah;f Pf wiH;f  wH;f
i 1
D Rf rO h;fi 1
C Ah;f s Ih;f s gh wih;s  wh;s
i 1
;
(49)

wherein (48) is equivalent to solving the structure with an Aitken prediction for the
pressure at the structure interface

Ah;s wih;s C Ah;sf pO ih;sf  sh;s D 0; (50)

and (49) is equivalent to a coarse grid solve for the fluid for a residual that consists
i 1
of the fluid at state wh;f and the coupling of the structure at state wih;s since using
(17) and (18) we obtain
 
rO ih;f
1
C Ah;f s Ih;f s gh wih;s  wh;s
i 1 i 1
D Ah;f wh;f C Ah;f s Ih;f s gh wih;s  sh;f ;
i 1
D Ah;f wh;f C Ah;f s dih;f s  sh;f ; (51)

so that (49) becomes

  h i
Rf Ah;f Pf wiH;f  wH;f
i 1 i 1
D Rf Ah;f wh;f C Ah;f s dih;f s  sh;f : (52)

Now, in order to perform the coarse grid iterations for nonlinear systems, we sub-
stitute the nonlinear operator rh;s .wih;s ; pO ih;sf / for the linear terms Ah;s wih;s C
Asf pO ih; to obtain
sf
rh;s .wih;s ; pO ih;s /  sh;s D 0; (53)
14 A.H. van Zuijlen and H. Bijl

for the structure solve. Next in (52) rh;f .wih;f ; dih; / is substituted for the lin-
fs
ear terms Ah;f wih;f C Ah;f s dih;f s and the coarse grid operator RAh;f P is
approximated by the nonlinear operator defined on the coarse grid rH;f to obtain
h i
rH;f .wiH;f ; diH;f s /rH;f .wH;f
i 1
; diH;f s / D Rf rh;f .wh;f
i 1 i
; dh;f s /  sh;f :
(54)
Note that on the left-hand-side of (54), the nonlinear coarse grid functions are
both evaluated with the boundary displacement diH;f s such that these terms would
cancel on the left-hand-side (in the linear case). Both the coarse level boundary dis-
placement as the fluid solution at iteration i  1 can be obtained through restriction
from the fine level
i 1 i 1
wH;f D Rf wh;f ; (55)
diH;f s D Rf dih;f s : (56)

The right-hand-side of (54) consists of the restriction of the residual on the fine fluid
i 1
level which is present for the solution wh;f under the boundary condition dih; .
fs
In order to obtain the coarse grid solution wiH;f , the flow solver has to solve

rH;f .wiH;f ; diH;f s /  sOH;f D 0; (57)

wherein sO is a source term on the coarse grid level equal to

h i
sOH;f D Rf rh;f .wih;f
1
; dih;f s /  sh;f C rH;f .wiH;f
1
; diH;f s /; (58)

which consists of the restriction of the fine grid residual and the coarse grid residual
for the solution wiH;f
1
. Notice that (57) is of the same form as the original fine grid
system (8), except with a different formulation for the source term. Therefore, an
existing solver (multigrid) can be used to resolve the coarse grid solution wiH;f .
Once the solution on the coarse grid is obtained at a computational expense much
smaller than resolving the flow on the fine level, the coarse grid correction term is

"iH;f D wiH;f  wH;f

i 1
; (59)

and the corrected fine grid solution becomes

wih;f D wh;f
i 1
C Pf "iH;f : (60)

i 1
The impact of rH;f .wH;f ; diH;f s / in the source term (58) can be indicated as fol-
lows: suppose the fully coupled solution is obtained, so that the fine grid residual
is identical to zero. In that case one would not want the coarse grid accelera-
tion to compute a correction other than zero. When the fine grid residual is zero,
Multi-Level Accelerated Sub-Iterations for Fluid-Structure Interaction 15

i 1
sOH;f D rH;f .wH;f ; diH;f s / and, therefore, (57) yields

rH;f .wiH;f ; diH;f s /  rH;f .wH;f

i 1
; diH;f s / D 0; (61)

which shows that the obvious solution is wiH;f D wH;f

i 1
, so that also the correction
term "H;f D 0 and the solution is not adversely affected.

3.5 Two-level accelerated sub-iteration algorithm

The algorithm that is used in this paper for acceleration of the Aitken sub-iterations
consists of the following steps:
1. Solve the structure (11) using Aitken for the prediction of the pressure force (14),

rh;s .wih;s ; pO ih;s /  sh;s D 0:

2. Obtain the fine grid residual for the fluid state still at the previous sub-iteration
(54)
rQ ih;f
1 i 1
D rh;f .wh;f ; Ih;f s .g.wis ///  sh;f :
3. Restrict the fine grid residual and the fine grid solution to the coarse level and
compute the coarse level modified source term (58).

rH;f .wiH;f ; RIh;f s .gh .wih;s ///  rH;f .wH;f

i 1
; RIh;f s .gh .wih;s /// D QrH;f :

4. Solve the coarse level flow problem (57)

rH;f .wiH;f ; diH;f s /  sOH;f D 0;

to obtain the coarse level correction term (59): "iH;f D wiH;f  wH;f i 1

5. Update the fluid state at the fine level by applying the correction by prolongation
(P ) of wiH;f to the fine grid (60)

wih;f D wh;f
i 1
C PwH;f :

6. Continue the procedure starting at step 1 until sufficiently converged or the

maximum number nc of coarse level iterations is performed.
7. Perform standard sub-iterations on the fine grid until sufficiently converged or
the maximum number nf of fine level iterations is performed.
8. When not sufficiently converged continue at step 1.
The algorithm allows the subsequent use of coarse and fine grid sub-iterations and
is denoted by CGPnc Cnf , e.g. CGP1 C1 means that 1 coarse level prediction step
is performed followed by 1 fine level correction (smoothing) sub-iteration, whereas
16 A.H. van Zuijlen and H. Bijl

CGP1 C1 (also denoted by CGP-FG) denotes that first all sub-iterations are per-
formed on a coarse level until converged, thereafter all sub-iterations are performed
on the fine level. At the moment only two level acceleration has been discussed, but
as with multigrid techniques, even coarser levels could be employed.

4 Results

The proposed multilevel acceleration technique is applied to two strongly coupled

two-dimensional test problems. The first is a laminar test case and the second a
turbulent test case. The flow solver is a unstructured, cell-centered, finite volume
solver with a second order central discretization and artificial Jameson dissipation.
Coupling of the CFD code to the linear structure code is performed by FLECS [7],
a open source flexible coupling shell that allows the coupling of different codes and
performs the interpolation between the interface meshes. The interpolation between
flow and structure meshes (Ih;sf and Ih;f s ) is a radial basis function interpolation
with a thin-plate-spline function [1]. The structure has its unknowns in the element
nodes which are already on the boundary of the elements. Therefore, the function
gh is simply selecting the boundary displacements from the structure state vector.
The fluid unknowns are located in the cell centers and, therefore, the function gh
performs an interpolation of the cell centered data to boundary vertex data.
At the moment, the flow solver re-initializes for every sub-iteration, meaning that
the coarse mesh levels are regenerated every sub-iteration. The multigrid levels are
obtained by agglomeration of fine level cells such that the coarsening ratio is close
to the theoretical optimal value of 4 (for two-dimensional applications), see Fig. 4.
However, as the mesh is deforming, the lay-out of the coarse grid may change and
even the amount of cells on the coarse level may differ from step to step. Therefore,
by performing the coupling still through the fine grid level, no information on the
coarse level correction term needs to be stored, which would be particularly diffi-
cult when the coarse level meshes change in lay-out. Of course, regenerating the
whole mesh structure does increase computational costs a lot. For the coarse level
solve, most of the computing time is spend on deforming the fluid mesh and regen-
erating the mesh data structure rather than the iterations for solving the coarse level
problem.

4.1 Two-dimensional laminar flow problem

The laminar case is based on [8] and consists of a circular cylinder of diameter
0.1m in a channel with height H D 0:41m, length L D 2:5m, with an elastic flap
behind it of length l D 0:35m and thickness h D 0:02m, see Fig. 5. The inflow is a
parabolic velocity profile (see [8] for details) with a mean velocity of 2 and a max-
imum velocity of 3m/s. Differently from the original test problem, compressibility
Multi-Level Accelerated Sub-Iterations for Fluid-Structure Interaction 17

(a) Undeformed fine grid level. (b) Undeformed coarse grid level.

(c) Deformed fine grid level. (d) Deformed coarse grid level.

Fig. 4 Fine and coarse level meshes for undeformed and deformed situation.

Fig. 5 Two-dimensional laminar testcase.

is allowed, since the solver we use is a compressible Reynolds Averaged Navier-

Stokes solver. The reference Mach number based on the mean flow velocity is set
to M0 D 0:14. The Reynolds number is based on the mean velocity and cylinder
diameter and is Re D 200. The structure is modeled as a linear elastic structure
with a density equal to the flow density  D 1000kg/m3 and a Young’s modulus of
E D 5:6  106 kg/(m.s2 ). Time integration is performed by an implicit, third-order
accurate, multistage Runge-Kutta scheme with a time step t D 0:01s (for details
on time integration see [10]). Each implicit Runge-Kutta stage is sub-iterated until
jjpis  pO is jj2  102. For this test problem the transient to a periodic state is sim-
ulated in 500 time steps. The flow solver uses 3 grid levels for the multigrid (MG)
solver. For the coarse level iterations, we use the intermediate MG mesh. The ratio
of fine to coarse level mesh cells is 20737 to 5442, which is  3:8. During the simu-
lations the averaged under-relaxation factor N is determined. This value is a measure
of how strongly coupled the simulation is. A value of N close to 1 means that hardly
any under-relaxation is required, whereas a small value of N indicates a strong cou-
pling with much under-relaxation. When only Aitken sub-iterations are performed
on the fine grid (FG), the averaged value of the under-relaxation N D 0:321, indicat-
ing that the problem is strongly coupled and that standard Gauss-Seidel iterations
18 A.H. van Zuijlen and H. Bijl

diverge. In a second simulation, first coarse level sub-iterations are performed until
the convergence criterion is met after which the sub-iterations are all performed on
the fine grid (CGP-FG). The averaged value of the under-relaxation on the coarse
level turned out to be NC G D 0:316, and on the fine level NF G D 0:319, from which
it is concluded that the fine and coarse grid sub-iterations perform “equally” with
respect to Aitken: therefore it would seem possible to alternately use coarse and fine
grid sub-iterations in combination with Aitken under-relaxation.

4.1.1 Sub-iteration behavior

Now we take a closer look at the sub-iterations performed for the first time step. The
initial solution consists of a developed vortex street and a structure that is at rest. In
the first computational step the structure is allowed to move. The Aitken algorithm
is initialized by a first value of  D 0:7, which is higher than the averaged value.
In Fig. 6 the resulting lift force (force in y direction) is plotted for the multigrid
iterations performed by the flow solver for the first time step. When the flow solver
is sufficiently converged or has done 20 MG iterations, a sub-iteration is carried
out. This coincides with a sharp change in lift as the structure has moved to a new
position. In Fig. 6(a) it can be noted that when only sub-iterating on the fine mesh
(Aitken-FG) the lift converges to around 2500 in the first flow solve, and to around
-7500 for the second flow solve, which shows an increase in error with respect to the
fully coupled solution which turns out to be around 150. The reason is that for the
second flow solve, Aitken is applied with an under-relaxation of  D 0:7, which is
in fact too high and for the next sub-iteration  is automatically lowered by Aitken
and the lift turns out at 250 (Fig. 6(b)) which is already much closer to the correct
value. Another striking difference is the value for the lift computed on the coarse
grid level and the fine grid level: this is best seen in Fig. 6(b), where the CGP-FG

(a) Full range. (b) Close up.

Fig. 6 History of lift for MG iterations performed by flow solver.

Multi-Level Accelerated Sub-Iterations for Fluid-Structure Interaction 19

Fig. 7 Computation of interface stresses.

algorithm, first only iterates the flow on the coarse grid level, after which only fine
grid iterations are used. The lift on the coarse level seems to converge to 70, whereas
the fine grid iterations converge to 150. Also for the CGP2 C1 algorithm, the coarse
and fine level sub-iterations are easily identified because of the large difference in
lift. The reason for this discrepancy is that the flow solver, when iterating on the
coarse level, computes the boundary forces based on the solution on the coarse level,
e.g. fH .wH;f /. In Fig. 7 the computation of the stress in y direction over the fluid
structure interface is shown. The results of the CGP-FG computation are used. The
’fine grid reference computation’ denotes the stress computed after the final fine grid
iteration (fully coupled solution). After 12 coarse grid sub-iterations, just before the
fine grid sub-iterations start, the computation of the force on the coarse grid and
fine grid is shown. It is clear that the computation of the interface stress on the
coarse level has a large discrepancy with the reference solution. However, when
for the same solution, the computation of the interface stress is computed on the
fine mesh, the error with respect to the reference is already much smaller. Should
the coupling between the coarse level flow and the structure be done directly, this
discrepancy will be introduced in the coupling and it may be conceivable that the
coarse level sub-iterations converge to a solution, very different from the coupled
fine grid solution and, therefore, may not contribute to a faster convergence.

4.1.2 Convergence in interface pressure and displacement

The next property that is investigated is the convergence in interface pressure and
displacements for the fine grid and two-level sub-iterations. It was concluded in [11]
20 A.H. van Zuijlen and H. Bijl

(a) Interface displacement difference for (b) Interface pressure difference for Aitken-FG.
Aitken-FG.

(c) Interface displacement difference for CGP- (d) Interface pressure difference for CGP-FG.
FG.

Fig. 8 Differences in interface displacement and pressure.

that coarse level sub-iterations perform just as well as fine grid sub-iterations when
the partitioning error is still large in the lower modes and that the error in the higher
modes that cannot be represented on the coarse level remain until a fine grid solver
is done. However, that test case was only a one-dimensional piston problem and the
fluid-structure interface only consisted of a single point. This time the coupling is
stronger and the fluid-structure interface is two-dimensional as well. To verify that
the same convergence behavior can be observed the difference in the interface pres-
sure and displacement with respect to the final (fully coupled) solution is shown in
Fig. 8. In 8(a) and 8(b) the results for the Aitken-FG sub-iterations are shown. It can
be seen that at first the error in pressure and displacement increases, as for the first
Aitken sub-iteration the under-relaxation parameter  D 0:7 which is too high for
this strongly coupled problem as explained in the previous section. Also the largest
errors in displacement and pressures are caused by low mode shapes, which should
enable the multilevel algorithm to efficiently tackle these modes first on the coarse
level before resolving the high mode shapes with a more expensive fine level solve.
Multi-Level Accelerated Sub-Iterations for Fluid-Structure Interaction 21

(a) Coarsening ratio 3.8. (b) Coarsening ratio 13.4.

Fig. 9 Convergence in interface pressure when switching from coarse to fine level sub-iterations.

This is shown in Fig. 8(c) and Fig. 8(d) which show the convergence of the interface
displacement and pressure for the first four coarse level sub-iterations. The conver-
gence on the coarse level has a close resemblance with the convergence on the fine
level, although there seem to be some small (high wave number) oscillations present
in the pressure distribution. Another conclusion from [11] was that a fine grid solve
effectively reduces the high wave number mode errors that cannot be represented on
the coarse level. To show this property for the more two-dimensional test problem,
in Fig. 9 the convergence for the pressure at the interface is shown for the last two
coarse level sub-iterations in the CGP-FG algorithm and the first two subsequent
fine level sub-iterations. In Fig. 9(b) the coarse level is represented by the coarsest
multigrid mesh generated by the flow solver, which, in this case, consists of only
1553 cells which yields a coarsening ratio of 13.4 with respect to the fine mesh. In
both cases the convergence on the coarse level stagnates as the high wave number
error cannot reduce any further. Performing a fine grid solve smooths out the high
wave number error and the remaining partitioning error is again smooth and has a
large low wave number component. Therefore, in order to avoid the stagnation of
convergence on the coarse level, it is beneficial to solve the fine level problem after a
few coarse level iterations. In order to increase computational efficiency, the coarse
level sub-iterations can be used after the fine level smoothing, as the remaining parti-
tioning error has a (relatively) large low wave number mode component again. This
is illustrated in Fig. 10 where for the CGP3 C1 all the sub-iterations to convergence
are shown in blocks of 4 sub-iterations: 3 coarse level iterations, followed by 1 fine
level smoothing. Fig. 10(d) only shows one coarse level iteration as the coarse level
iteration already reached the required tolerance after 1 iteration. The final fine level
iteration shows an error of zero as this result is taken as the reference value. The
effectiveness of alternating coarse and fine level sub-iterations is shown. Applica-
tion to other sub-iteration techniques seems straight forward, especially when these
techniques rely on the approximation of the low wave number modes first, e.g. the
quasi-Newton subiterations of [2], as these modes can be accurately represented on
the coarse level.
22 A.H. van Zuijlen and H. Bijl

(a) Sub-iterations 1–4. (b) Sub-iterations 5–8.

(c) Sub-iterations 9–12. (d) Sub-iterations 13–14.

Fig. 10 Convergence in interface pressure for CGP3 C1 .

4.1.3 Costs of sub-iterations

It must be noted that the first computational step is quite demanding in terms of
finding a coupled solution: as the solution starts from a flow and structure state
that are completely uncoupled, the initial transient to a periodic motion is quite
strong so that the first few time steps require more sub-iterations per time step then
during the remainder of the computation. Of course also the somewhat high value
of  D 0:7 for the first sub-iteration increases the partitioning error initially. To
compare the computational costs in terms of fine and coarse level sub-iterations,
we therefore determine the averaged number of sub-iterations per stage during the
whole simulation from initial condition to periodic motion. In Fig.11 the averaged
number of sub-iterations per stage in terms of coarse and fine level iterations are
shown. The first point in the plot gives the reference number of sub-iterations as
these correspond to Aitken sub-iterations on the fine grid (FG) only. The second
point (CGP-FG) first only performs sub-iterations on the coarse level, thereafter it
continues with sub-iterations only on the fine level. CGP-FG reduces the number
of fine grid sub-iterations by roughly 40–50%, but on the other hand increase the
 Multi-Level Accelerated Sub-Iterations for Fluid-Structure Interaction 23

12
Laminar test case
10
avrg. # subiterations / stage

6
Fine level
4 Coarse level
Total
2

0
FG CGP-FG CGP1C1 CGP2C1 CGP3C1 CGP4C1 CGP5C1

Fig. 11 Averaged number of coarse and fine grid sub-iterations per stage for the laminar case.

total number of sub-iterations more than 60%. Therefore, a lot of computational effi-
ciency gained by performing less fine grid flow solves is lost by the additional coarse
level sub-iterations to perform, which also include moving the fluid mesh, transfer
of data between flow and structure and solving the structure. Alternating coarse
level/fine level sub-iterations, reduces the number of fine level sub-iterations as well
as the total number of sub-iterations compared to CGP-FG. Increasing the number
of coarse grid sub-iterations before a fine grid sub-iteration reduces the number of
required fine grid solves, which seems to converge to 2. Increasing the number
of coarse grid sub-iterations even further, only increases the total number of sub-
iterations required, without an obvious gain in reduction of fine grid sub-iterations.
The most effective scheme is found to be CGP3 C1 , reducing the number of fine grid
iterations by 65% at the expense of an increase in total number of sub-iterations
by 30%.

4.2 Two-dimensional turbulent flow problem

The proposed method is applied to a two-dimensional laminar and a two-

dimensional turbulent test case, which, geometrically, differs slightly from the lam-
inar case as the circular cylinder is replaced by a square cylinder, see Fig. 12, with
edges of 0.1m to increase the shed vorticity. The Reynolds number is increased to
Re D 1  105 and the Spalart-Allmaras turbulence model is used. The Young’s mod-
ulus of the structure is increased to E D 22:4  106 kg/(m.s2 ). The initial condition
is taken after the transient to a periodic coupled solution, and only a single period
( 32 time steps) is simulated with a time step of t D 0:01. The ratio of fine to
coarse level mesh cells is 56746 to 15151, which is  3:7. Simulations are run for
different settings and Fig. 13 shows the averaged number of sub-iterations required
per stage in terms of coarse and fine level iterations. From the computation on only
24 A.H. van Zuijlen and H. Bijl

Fig. 12 Two-dimensional turbulent testcase.

14
Turbulent test case
12
avrg. # subiterations / stage

10

6 Fine level
Coarse level
4 Total

0
FG CGP-FG CGP1C1 CGP2C1 CGP3C1 CGP4C1 CGP5C1 CGP6C1

Fig. 13 Averaged number of coarse and fine grid sub-iterations per stage for the turbulent case.

the fine grid (FG) it was found that the average under-relaxation N D 0:396 and that
for CGP-FG: NC G D 0:373 and NF G D 0:343. Again these values are in close range
of each other, indicating a similar behavior of fine and coarse grid sub-iterations.
Much of the same behavior as for the laminar case can be found: performing more
coarse grid sub-iterations before a fine grid solve reduces the number of required
fine grid solves substantially, although at some point the increase in total number
of sub-iterations reduces computational efficiency. The most effective schemes is
found to be CGP4 C1 , reducing the number of fine grid iterations by 70% at the
expense of an increase in total number of sub-iterations by 30%.

5 Conclusions

In this paper we investigated the combination of multilevel acceleration with the

Aitken underrelaxtion technique applied to a two-dimensional strongly coupled
laminar and turbulent test problem. Since the value for the under-relaxation parame-
ter is not significantly different for coarse and fine level sub-iterations, they are used
alternately, where it is found that performing 3 or 4 coarse level sub-iterations fol-
lowed by 1 fine level sub-iteration results in the highest gain in efficiency. Although
the total number of sub-iterations increases slightly by 30%, the number of fine grid
iterations was decreased by as much as 65–70%.
Multi-Level Accelerated Sub-Iterations for Fluid-Structure Interaction 25

Performing the coupling in the coarse level sub-iterations over the fine mesh,
allows alternating fine and coarse level sub-iterations: 1) without modifications to
the Aitken algorithm, 2) without the necessity to store coarse level information and
3) with an automatic incorporation of the coupling terms. However, it comes at the
price of having to perform a fine grid mesh deformation, even for a coarse level
sub-iteration. It is worth while investigating whether a fully coarse level coupling
and sub-iteration can be used as well.
Up to this point only the Aitken under-relaxation was used in combination with
multilevel acceleration and only few strategies were tested for optimal convergence.
Application to other sub-iteration techniques seems straight forward, especially
when these techniques rely on the approximation of the low wave number modes
first, as these modes can be accurately represented on the coarse level.

References

1. A. de Boer, A.H. van Zuijlen, and H. Bijl. Review of coupling methods for non-matching
meshes. Comput. Methods Appl. Mech. Engrg., 196:1515–1525, 2007.
2. J. Degroote, K-J. Bathe, and J. Vierendeels. Performance of a new partitioned procedure versus
a monolithic procedure in fluidstructure interaction. Comput. Struct., 87:793–801, 2009.
3. C. Forster, W.A. Wall, and E. Ramm. Artificial added mass instabilities in sequential staggered
coupling of nonlinear structures and incompressible viscous flows. Comput. Methods Appl.
Mech. Engrg., 196(7):1278–1293, 2007.
4. R. Haelterman, J. Degroote, D. van Heule, and J. Vierendeels. The quasi-newton least squares
method: A new and fast secant method analyzed for linear systems. SIAM J. Numer. Anal.,
47:2347–2368, 2009.
5. C. Michler, E.H. Brummelen, and R. de Borst. An interface Newton-Krylov solver for fluid-
structure interaction. Int. J. for Num. Meth. in Fluids, 47(10-11):1189–1195, 2005.
6. D.P. Mok, W.A. Wall, and E. Ramm. Accelerated iterative substructure schemes for insta-
tionary fluid-structure interaction. In First MIT Conference on Computational Fluid and Solid
Mechanics, pages 1325–1328, 2001.
7. Margreet Nool, Erik Jan Lingen, Aukje de Boer, and Hester Bijl. Flecs, a flexible coupling
shell application to fluid-structure interaction. In PARA’06: State-of-the-Art in Scientific and
Parallel Computing, Umeøa, Sweden, 2006.
8. S. Turek, J.: Proposal for Numerical Benchmarking of Fluid-Structure Interaction between an
Elastic Object Hron, Schäfer-M. (eds.) Fluid-Structure Interaction Lecture Notes Comput.S̃ci.
Laminar Incompressible Flow. In: Bungartz, H.-J., and pp. 371-385-Springer Heidelberg Eng.,
LNCSE 53. 2006.
9. J. Vierendeels, L. Lanoye, J. Degroote, and P. Verdonck. Implicit coupling of partitioned flu-
idstructure interaction problems with reduced order models. Comput. Struct., 85:970–976,
2007.
10. A.H. van Zuijlen, A. de Boer, and H. Bijl. Higher order time integration through smooth mesh
deformation for 3d fluid-structure interaction simulations. J. Comput. Phys., 224:414–430,
2007.
11. A.H. van Zuijlen, S. Bosscher, and H. Bijl. Two level algorithms for partitioned fluid-structure
interaction computations. Comput. Methods Appl. Mech. Engrg., 196:1458–1470, 2007.
•
A Classification of Interface Treatments for FSI

C.A. Felippa, K.C. Park, and M.R. Ross

Abstract This paper proposes a taxonomy of methods for the treatment of the fluid-
structure interface in FSI coupled problems. The top-level classification is based on
the presence or absence of Additional Interface Variables (AIV) as well as their type.
Associated prototype methods: Direct Force Motion Transfer (DFMT), Mortar and
Localized Lagrange Multipliers (LLM) are defined. These are later studied in more
detail using a specific FSI benchmark problem used in Ross’ 2006 thesis. Desirable
attributes of the interfacing methods are stated and commented upon.

1 Introduction

Computational fluid-structure interaction (FSI) emerged in the late 1960. It was pri-
marily driven by linear (or linearized) problems in acoustics, flutter and vibration.
Given the speed and memory limitations of computers of that time, early analy-
sis techniques naturally favored analytic or semi-analytic modal formulations in the
frequency domain; see e.g. [23]. Structural discretizations were used primarily for
modal eigenextraction. Time-domain FSI began to make headway in the 1970s as
interest developed in exterior and nonlinear FSI problems, which lie outside the
scope of the frequency domain approach. A serious obstacle to sustained growth
was the great variety of fluid problems. While computer-based structural analysis
was dominated by the Finite Element Method (FEM) since the mid 1960s, no com-
parable universal treatment was available on the fluid side. For example, an acoustic
fluid model is useless for high speed gas dynamics, turbulent or multiphase flow.

C.A. Felippa and K.C. Park

Department of Aerospace Engineering Sciences and Center for Aerospace Structures,
University of Colorado at Boulder, Boulder, Colorado 80309-0429, USA
e-mail: [email protected]
M.R. Ross
Analytical Dynamics Department, Sandia National Laboratories, P. O. Box 5800, MS 0346,
Albuquerque, NM 87185-0346, USA
e-mail: [email protected]

H.-J. Bungartz et al. (eds.), Fluid Structure Interaction II, Lecture Notes 27
in Computational Science and Engineering 73, DOI 10.1007/978-3-642-14206-2 2,
c Springer-Verlag Berlin Heidelberg 2010
28 C.A. Felippa et al.

The lack of an universal discretization methodology for the fluid component

fostered specialization in modeling and computer implementation as technological
advances stoked the need for simulations. A key motivation behind the partitioned
methods developed in the 1970s and 1980s [13–16, 24–28] was to “divide and con-
quer” the coupled problem. This division facilitated reuse of FEM and CFD software
resources developed since the early 1950s, as well as the expertise and experience
accumulated during that time. But combinatorial complexity has remained in two
aspects: time marching methods, and treatment of the fluid-solid interface. To find a
path through that maze, a classification of those aspects is proposed here. Emphasis
is placed on the interface treatment because this aspect has experienced more rapid
development in recent years.
The present paper has evolved from M. R. Ross’ thesis work [36], which has
appeared in a recent article sequence [37, 38]. The comparison of interface treat-
ments outlined in [38] formed the skeleton for a tentative classification, which is
expanded upon in the next section.

2 Classification Ingredients

Four ingredients of computational FSI are sketched in Fig. 1: fluid, structure, inter-
face and time stepping. The first three are realized as discrete spatial models. The
last one identifies the time domain response method.
As far as the proposed classification is concerned, there is no need to rein-
vent labels for the fluid and structure models when viewed as isolated individual
entities, since pertinent terminology is well established by now. Echoing introduc-
tory remarks, the structural model will normally be FEM based, whereas the fluid
model will be problem dependent; for example a finite volume discretization in
aeroelasticity or a boundary-element discretization in acoustics.
Terminology for the other components: interface treatment and time stepping,
is not so well established. For the latter, the top-level division into monolithic and
partitioned, formally introduced in a 1983 review chapter [26], is gradually gain-
ing acceptance. But interface treatment nomenclature remains volatile since some
techniques, especially those introducing Lagrange multipliers, are of more recent
coinage. On this account we will focus on that component.

Interface xi

Structure
Fluid time
Time
Stepp
ing

Fig. 1 Two classification ingredients: interface treatment and time-stepping methods.

A Classification of Interface Treatments for FSI 29

2.1 Interface Treatment: Top Level Classification Criteria

There is a interface treatment feature that strongly drives computer implementation:

whether additional interface variables (AIV) are present or not. Accordingly we
shall categorize the top level interface treatment into three major classes.
1. Primal. There are no AIVs
2. Dual. AIVs of dual type (Lagrange multipliers) are introduced.
3. Primal-dual. AIVs of both dual and primal type are introduced.
Any of these may be associated with either monolithic or partitioned time-
stepping solution procedures, although admittedly some combinations might not
be practical. It follows that the foregoing classification may be viewed as “orthog-
onal” to one based on time stepping, a distinction that simplifies terminology. The
orthogonality is illustrated in Fig. 2. The gridlike arrangement places interface treat-
ments horizontally and time stepping solution procedures vertically. That figure also
displays important interface treatment realizations, as discussed next.

2.2 Interface Treatment Prototypes

Fig. 3 sketches practical realization schemes (referred to as “prototypes” for brevity)

for each of the three major interface treatment categories: primal, dual and primal-
dual. It pictures nonmatching meshes since that is often the case in FSI.
The prototype for primal methods is the Direct Force-Motion Transfer, or DFMT,
scheme:
F)S Interface fluid pressures are passed as forces to structural nodes,
F(S Structural motions are passed as interface fluid-particle velocities.

Interface Treatment
T Primal Dual Primal - Dual Major categories
i Direct Force - Motion Mortar LLM
m Important instances
Transfer (DFMT)
e or "prototypes" of
Local Global above categories
S
t
e Monolithic
p This paper focuses on
p Partitioned Interface Treatment
i (many variants:
n e.g., staggered,
g subcycled)

Fig. 2 Two “orthogonal,” top-level classification criteria.

30 C.A. Felippa et al.

(a) Primal (b) Dual (c) Primal - Dual

Direct Force-Motion Mortar Localized Lagrange
Transfer (DFMT) Multipliers (LLM)
Frame
Fluid Structure Fluid Structure Fluid Structure

Multipliers Multipliers

Fig. 3 Sketches for practically important realizations of three interface treatment prototypes.

The gapless sketch of Fig. 3(a) means that transfers are done directly; there are no
AIV intermediaries. Two DFMT subclasses are distinguished in Fig. 2 by qualifiers
local and global, which convey interface coupling extent. In a local DFMT scheme,
a fluid (structure) interface freedom interacts only with neighboring structure (fluid)
interface freedoms. In a global DFMT scheme, the coupling extends beyond that,
and possibly to the entire interface.
Over the past decade the Mortar scheme [5] has emerged as prototype for dual
methods. Interface freedoms are linked through Lagrange multipliers. Variants arise
primarily through the choice of multiplier discretization spaces. Figure 3(b) shows
delta-function multipliers collocated at interface nodes. This choice, which is that
studied in [36–38], has the advantage of implementation simplicity and immediate
physical interpretation. Smoother spaces of distributed multipliers may be used at
the cost of additional implementation complexity, as well as need of knowing inter-
face shape function details for spatial integration. For structure-structure interaction,
as in the contact problem, the idea behind this interfacing method can be traced way
back, to a variational principle proposed by Prager for solid mechanics [33].
The Localized Lagrange Multiplier (LLM) scheme, sketched in Fig. 3(c), is a
prototype for primal-dual methods. A kinematic “frame” is placed between the fluid
and structure. Multipliers are collocated at the interface between the fluid and frame,
as well as between structure and frame. The idea, originally introduced for contact
problems [30], synthesizes two trends: the FETI domain decomposition methods
developed in the early 1990s [8] and, as the background theory evolved, the varia-
tional treatment of hybrid finite elements proposed by Atluri [2] for solid mechanics.
For the three-decade evolution of this governing functional see [27, 31].

2.3 Hierarchical Specialization: From LLM To DFMT

The three major interface treatment classes: primal, dual and primal-dual, are not
impervious islands. Analytical and computational procedures may be used to move
over from the most general top class: primal-dual, to dual and finally to primal.
This “hierarchical cascade,” AIV-reduction process is flowcharted in Fig. 4. Links
A Classification of Interface Treatments for FSI 31

Localized Lagrange
Multipliers (LLM)
Wavelet
Frame Multiplier filtering?
elimination Primal Treatment Primal Treatment
elimination
in functional by Global DFMT by Local DFMT
by LS

Mortar

Fig. 4 Reduction “cascade” of additional interface variables (AIV) among interfacing prototypes.

displayed in that figure shed light on interrelations among seemingly disparate

schemes that may be found in the FSI literature.
More specifically: the LLM interface functional may be reduced to a Mortar
functional method by introducing constraints among the interface variables, as
described in Sect. 4.3. Upon discretization of either LLM or Mortar functionals,
algebraic elimination of interface variables by least-square (LS) techniques beget
global DFMT schemes. As of this writing it is not known whether one may subse-
quently precipitate a local DMFT scheme through further localization, although in
principle model reduction schemes such as wavelet filtering could be explored along
those lines. Accordingly the rightmost link in Fig. 4 bears a question mark.

2.4 Background

The hierarchy displayed in Fig. 4 did not historically develop in the left-to-right
“top-down” fashion. It germinated and grew in a more circuitous fashion, mostly
going from specific to general. At a fluid-solid “wet” interface one deals with
two interacting agents: fluid pressures that coerce a flexible structure into resisting
forces, and structural velocities that impel the fluid. Upon discretization, interface
pressures can be lumped to structure node forces whereas structure motions are pre-
scribed as velocities on the fluid interface, using appropriate interpolation should
meshes mismatch. This transparent physical interpretation is easily understood by
engineers. Understandably this simple scheme was the first one to be adopted in FSI
computations.
An early DFMT implementation was used in the development of partitioned solu-
tion procedures for underwater shock [17, 24], an application sketched in Fig. 5(a).
The sketch reflects mid-1980 capabilities, which included hull cavitation modeling
[14], later extended to free-surface bulk cavitation to simulate surface ship attacks.
The simulation coupled three computational models: a FEM-discretized submerged
structure, surrounded by a FEM-discretized, near-field, acoustic bilinear fluid able to
cavitate, closed with a BEM-discretized silent boundary implemented through the
Doubly Asymptotic Approximation (DAA) of Geers [18–20]. Meshes were usu-
ally nonmatching. Interface exchanges were done through the local DFMT scheme
32 C.A. Felippa et al.

PHYSICAL DOMAIN

Silent COMPUTATIONAL
DAA DOMAIN
boundary
shock
Structure wave
Cavitation
region Acoustic fluid
Acoustic fluid (near field)
(far field)

(a)
STRUCTURE PARTITION FLUID PARTITION
Pressures
Bilinear
fluid
volume
Structure Pressures
Silent
Fluid - particle DAA
velocities Boundary

(b)

Fig. 5 Early Local DFMT treatment: underwater shock analysis.

Fluid silent Cavitating LLM Frames

Cavitating region boundary region

Water

Concrete
Saturated soil
Unsaturated fill

Sandstone Soil silent

Base rock
boundary

Base ground motion

(a) (b)

Fig. 6 First LLM application to FSI: dam under seismic action; (a) section of gravity dam,
(b) LLM partitioning of computational domain.

flowcharted in Fig. 5(b). We know now that this treatment is not energy conserving.
However, for explosion-induced shockwave simulations in which only a few mil-
liseconds of early-time response are of interest to assess survivability, this defect is
of little importance.
The opposite in complexity is the primal-dual LLM scheme that evolved from
the sources described in Sect. 2.2. It was initially used only for structure-structure
interaction in contact problems. The first FSI application, described in [36–38], dealt
with a dam under seismic action, as sketched in Fig. 6. The multiple-partition model
pictured in Fig. 6 is, however, too complicated for systematic research studies.
A Classification of Interface Treatments for FSI 33

Fluid silent
STRUCTURE & SOIL PARTITION FLUID PARTITION
boundary

Structure
Fluid & Acoustic
(acoustic FEM) Soil Near Fluid
Structure Field Silent
(elastic FEM) Boundary
Soil
Silent Cavitating
(elastic FEM) Fluid
Boundary

Soil silent boundary

(a) (b)

Fig. 7 Two-partition decomposition of dam problem.

A simpler two-partition model of the dam problem is shown in Fig. 7(a). This sim-
plification gets rid of cross points, which may cause serious difficulties with some
interface treatments. The interaction diagram is flowcharted in Fig. 7(b).
The main motivation for trying this scheme was program modularity. Interposing
the frame conceals the fluid from the structure and vice-versa: each program com-
ponent only sees forces passed as Lagrange multipliers. It does not need to know
details of the opposite mesh, such as element shape functions. (In commercial codes
those may be unknown to users anyway.) Nonmatching meshes, which are impor-
tant in the contact problems for which LLM was first developed, are conveniently
handled. Model reduction of one of the components (for example, replacing a lin-
ear structure by a set of vibration or Ritz modes) is eased as such replacements are
fully hidden by the frame. Stand-alone testing and staged development are greatly
simplified.
What price LLM modularity? The introduction of a large number of AIVs, and
the fact that the frame must be discretized through special rules if certain conser-
vation attributes are enforced. The AIV impact can be lessened by hierarchical
elimination. Getting rid of the frame reduces LLM to Mortar, which carries less
AIVs. Eliminating all AIVs by least-square methods yields global DFMT schemes,
a reduction that has special appeal in linear dynamic problems (because it can be
preprocessed). Since full AIV elimination typically couples all interface variables,
interface processing costs remain high with respect to local DFMT schemes. As
discussed in following sections, tradeoffs of various nature inevitably emerge at all
stages.

3 Desirable Attributes

From a glance at the grid of Fig. 2, it is obvious that combining interface treatments
with a profusion of time stepping schemes begets a large number of possible FSI
computational methods for a given problem. Which one to pick? To help answer this
question one should consider two aspects: which method attributes are desirable,
34 C.A. Felippa et al.

and which priority is assigned to those attributes. Following is a commented, non-

exhaustive list of attributes.
The first one is essential:
 Handles nonmatching meshes. Methods that only work for matching meshes
are useless.
The next three pertain to conservation attributes; their relative priority is problem
dependent.
 Flux transmission consistency. This may be verified by the Interface Force
Patch Test (IFPT) described later.
 Rigid body motion transmission consistency. This may be verified by the
Interface Rigid Motion Patch Test (IRMPT) described later.
 Energy conservation. No artificial energy is gained or lost at interfaces. Testable
by duality.
The next set collects beneficial algorithmic attributes.
 Interface equation stability. Primarily germane to dual and primal-dual treat-
ments: interface matrices must have correct rank.
 Time stepping stability. Applicable to partitioned time stepping: temporal sta-
bility of monolithic solution procedure should not be degraded by partitioning.
The interface treatment can have an indirect effect on this property, as discussed
in Appendix A of [37].
 Time stepping accuracy. Similar to above, with “stability” replaced by “accu-
racy order”.
 Observer independence. Pertinent to dual and primal-dual methods that require
designation of master and slave faces at interfaces. Solution should not depend
on choice of master.
 Correct handling of cross points. A cross point (“cross line” in 3D) is one at
which three or more interfaces meet. Some treatments may fail there if meshes
do not match.
 Interface equations derivable from functional. Advantages are listed in
Sect. 4.1.
 Interface error monitoring for stepsize control. Equations obtained from inter-
face functionals befit this attribute. For a discussion of energy error measures of
this kind see [38].
Desirability of the following ones is affected by project goals and problem
features, as noted in the comments.
 Fast implementation. Crucial in time critical projects. Prioritization favors
primal methods.
 Overall computational efficiency. Qualifier “overall” means to stress here
that what matters is the efficiency with which the full coupled problem is
solved, rather than that attained for individual components. (Often one problem
component dominates processing time.)
A Classification of Interface Treatments for FSI 35

 Parallelization friendly. Obviously linked to previous one.

 Handles nonlinear problems. Relative weight of this attribute depends on
whether the interface treatment is affected by nonlinearities, as is the case in
contact problems.
 Handles multiscale phenomena. Of interest for a certain class of problems. For
example, turbulent flow or capillar effects.
 Handles gaps and interpenetration. Similar to above one in being limited to
problems that may exhibit volume mesh overlapping. For example, multiphase
flow or gas bubbles.
 Adapts well to Reduced Order Modeling (ROM) of a component. Here
the information-hiding and data-filtering properties of primal-dual methods are
beneficial.
 Facilitates software reuse. This applies to treatments in which open source soft-
ware is available for fluid, structure, or both. For example, ability to use separate
mesh generators.
 Facilitates black box (BB) reuse. Differs from above in that commercial or
proprietary software is used for fluid, structure, or both. As code modifications
are precluded, communication must be done via I/O. In addition, details such
as element shape functions might be unavailable, which can demerit interfacing
methods that rely on such knowledge.
As may be expected, no known FSI methodology meets all these desirable
attributes, or even a majority of them. Prioritization as per application and goals is
essential. For example, in shock-wave-excited problems correct flux (force) trans-
mission is important but energy conservation is not. Conservation attributes are
studied in more detail in following sections.

4 A Specific FSI Problem: Dam Under Seismic Action

As previously noted, the present classification emerged from a study of alternative

interface formulations in Ross’ thesis [36], summarized in a recent article sequence
[37, 38]. The compared formulations were used to simulate seismic actions on two
existing water-filled dams: the Koyna gravity dam located in India, and the Morrow
Point arch dam located in Colorado, USA. These were analyzed under standard digi-
tized earthquake records available from the COSMOS database [1] using 2D and 3D
models, respectively. The Koyna dam, pictured in Fig. 8(a), was more extensively
studied since the 2D plane strain model allowed systematic convergence studies,
quantified by various error mesaures, to be rapidly carried out using Matlab on a
PC. A typical 2D discrete model is pictured in Fig. 8(b). Note that fluid-structure
meshes are nonmatching.
36 C.A. Felippa et al.

Fluid silent Fluid 21 m

boundary
Dam 19 m

Soil silent
boundary
157 m 66 m
131 m
50 m
70 m
Interface frame
Rock foundation
(a) (b)

Fig. 8 LLM application to a FSI problem: Koyna gravity dam under seismic action; (a) dam
picture, (b) spatial discretization using a plane strain “slice”.

4.1 Variationally Based Interface Coupling

All interface coupling methods used in the dam problems are variationally based.
The following advantages can be cited for working within a variational framework:

1. Lagrange multipliers that appear in dual and primal-dual treatments are naturally
accommodated. Their presence in a variational statement guides decisions as to
the selection of multiplier interpolation spaces.
2. If space discretization of the physical problem components produce symmet-
ric equations, symmetry of the master coupled equations is guaranteed. This
facilitates the use of more efficient equation solvers as well as coupled sys-
tem eigensolvers. As regards the latter, maintaining symmetry excludes possible
appearance of nonphysical complex roots.
3. Testing for interface energy conservation is simplified.

Even if the interface treatment is not variationally based, as was the case with the
local DFMT scheme used in the underwater shock problem of Fig. 5, comparison
with DFMT equations derived through variational methods is illuminating.
In the present section the LLM variational framework of [37, 38] is recalled for
the reader’s convenience. Two other interface methods developed as LLM special-
izations: the Mortar and global DFMT methods, are discussed in Sects. 5 and 6,
respectively.

4.2 Equations of Motion

Below we summarize the coupled equations of motion (EOM) of the LLM dis-
cretization shown in Figs. 9 and 10 for the matching and nonmatching mesh case,
respectively. Structural damping, silent boundaries and nonlinear fluid-cavitation
effects are omitted from the governing EOM for brevity. Only elastic material
behavior and small displacement motions are considered. Fluid flow effects are
ignored as only the early time response under seismic actions is of interest. Under
A Classification of Interface Treatments for FSI 37

the foregoing assumptions the coupled EOM may be derived from the total func-
tional
˘ D ˘S C ˘F C ˘B : (1)
Here ˘S and ˘F are governing functionals used for the structure and fluid, respec-
tively, considered as separated individual entities. When existing FEM programs are
used, the choice of ˘S and ˘F is implicit from whatever models were used therein,
and need not be discussed further. The new ingredient is ˘B , called the interface
functional. This term produces the interface coupling equations studied below.
Separate discretization of the fluid and structure models, as well as LLM treat-
ment of the fluid-structure interface as illustrated in Figs 9 and 10 yields the
following semidiscrete matrix EOM in terms of displacements and interface forces:

Fluid

Structure

λF λ
S

Lagrange multipliers discretized Frame discretized with

as node collocated delta functions. piecewise linear elements
Only one multiplier on each with nodes collocated with
side of the frame needed, since boundary nodes
acoustic fluid is inviscid

Fig. 9 Zoom on interface elements coupled by a LLM treatment: matching mesh case. Frame
nodes coincide with fluid and structure boundary nodes on each side.

Fluid

Structure

λF
λS

Lagrange multipliers discretized Frame discretized with

as boundary node collocated delta piecewise linear elements
functions. Only one multiplier on each with nodes placed as per
side of the frame needed, since conservation requirements
acoustic fluid is inviscid

Fig. 10 Zoom on interface elements coupled by a LLM treatment: nonmatching mesh case. Frame
nodes do not generally coincide with fluid or structure boundary nodes on each side.
38 C.A. Felippa et al.
2 38 9 2 38 9
MS 0 0 0 0 ˆ uR S > KS 0 BSn 0 0 ˆ uS >
ˆ > ˆ >
6 0 MF 0 0 07ˆ< uR F >
= 6 0 KF 0 BF n 0 7ˆ< uF >
=
6 7 R 6 7
6 0 0 0 0 0 7 S C 6 BTSn 0 0 0 LSn 7 S
4 5ˆ R > 4 5ˆ >
0 0 0 0 0 ˆ:̂ F
>
>
;
T
0 BF n 0 0 LF n ˆ:̂ F
>
>
;
T T
0 0 0 0 0 uR B 0 0 LSn LF n 0 uB
8 9
ˆ f S >
ˆ
ˆ >
< fF >=
D 0 : (2)
ˆ
ˆ 0 > >
:̂ > ;
0

For the structure model, uS is the array of structural node displacements. whereas
MS , KS and fS denote the master mass matrix, stiffness matrix and applied force
vector, respectively, associated with uS . For the fluid model, uF is the array of
fluid node displacements. whereas MF , KF and fF denote the master mass matrix,
stiffness matrix and applied force vector, respectively, associated with uF . Over
the LLM-treated FSI interface, uB is the array of frame node displacements, S the
array of frame-to-structure interaction forces at wet structural nodes, F the array of
frame-to-fluid interaction forces at fluid nodes, BSn and BF n are Boolean matrices
that map S and F onto the full set of structural and fluid node forces, respectively,
LSn and LF n are matrices that map frame displacements uB to structural node
freedoms and fluid node freedoms, respectively. Structure and fluid nodes need not
coincide over the interface. The time dependence of the state vectors and forces
is omitted from (2) for brevity. A superposed dot denotes time differentiation with
respect to time t.
Fluid irrotationality is enforced by the transformation uF D DF , where
collects displacement potential degrees of freedom at fluid mesh nodes and DF is
a generally rectangular transformation matrix. (Since the displacement potential is
a scalar field, there is only one freedom per node.) Carrying out a congruential
transformation on fluid freedoms yields
2M 8 9 2 38 9
S 0 0 0 0 3ˆ uR S > KS 0 BSn 0 0 ˆ uS >
ˆ R > 0 BF n 0 7 ˆ
ˆ >
6 0 MF 0 0 0 7ˆ
< >
= 6 0 KF
6 T 7< >
=
6 7 R 0 LSn 7
6 0 0 0 0 0 7 S C66 BSn T0 0 7 S
4
0 0 0 0
5ˆ R
0 ˆ:̂ F
>
>
>
;
4 0 BF n 0 0 LF n 5ˆ
ˆ F
:̂
>
>
>
;
0 0 0 0 0 uR B 0 0 T T
LSn LF n 0 uB
8 9
ˆ f S >
ˆ
ˆ >
< fF > =
D 0 ; (3)
ˆ
ˆ 0 > >
:̂ >
;
0

in which MF D DTF MF DF , KF D DTF KF DF , BF n D DTF BF n and

T
fF D DF fF . For detailed tutorial examples on the configuration of these EOM,
see Appendix C in [37].
A Classification of Interface Treatments for FSI 39

4.3 Interface Functionals

The “wet” fluid-structure boundary is denoted by B . Since an acoustic fluid model

is inviscid, only the normal displacements uF n and uSn , as well as normal trac-
tions tF n and tSn , appear in the strong interaction conditions: uF n D uSn and
tF n C tSn D 0. (Note that tF n D pB , in which pB is the hydrodynamic fluid
pressure on B , positive if compressive.) Two weak forms, originally proposed for
elasticity problems by Prager [33], Pian and Tong [31] and Atluri [2], can be stated
in terms of the following interface functionals
Z
˘BM ŒBn  D .uF n uSn / Bn d;
B
Z n o
L
˘B ŒBn ; F n ; uBn  D .uF n uBn / F n C .uSn uBn / Sn / d: (4)
B

Here independently varied fields are identified in square brackets on the left-hand
side. In ˘BM , Bn is a global Lagrange multiplier function that connects directly
the fluid and structure faces. In ˘BL , F n and Sn are localized Lagrange multiplier
functions that link the independently varied normal displacement uBn of a frame
introduced between fluid and structure. The LLM treatment of the interface is based
on ˘BL whereas the variational-based Mortar method outlined later derives from
˘BM . Substituting

uBn ) 12 .uF n C uSn /; 1

2
.F n  Sn / ) Bn (5)

into ˘BL reduces it to ˘BM . Note that the replacements in (5) must be adjusted at
cross points where more than two partitions meet. At such locations special handling
is needed because no unique normal exists. Excluding that situation, ˘BL embodies
˘BM as special case.

4.4 Discrete Interface Equations

To produce matrix connection equations, functionals (4) are spatially discretized by

assuming shape functions for the independently varied fields: either Bn in ˘BM , or
F n , Sn and uBn in ˘BL . Boundary normal displacements uSn and uF n come from
elements used for the fluid and structure models, respectively, and are prescribed
data for both interface functionals. Restricting attention to the more general ˘BL ,
assume F n D NF F , Sn D NS S and uBn D NB uB . For the boundary
displacements take uF n D NFB uF and uSn D NFB uS , where NFB and NSB are
shape functions for fluid and displacement elements, respectively, evaluated on B
and projected over the normal n. Insert these interpolations in ˘BL , and integrate
over B to get the discretized functional

˘BL ŒF ; S ; uB  D TF .BF uF  LF uB / C TS .BS uS  LS uB / ; (6)

40 C.A. Felippa et al.

in which the connection matrices are defined as

Z Z
BF D NTF NFB d; BS D NTS NSB d;
ZB ZB
T
LF D NF NB d; LS D NTS NB d: (7)
B B

Some notational simplifications have been made in the foregoing expressions for
brevity: subscript n is dropped throughout whereas B denotes the discretized “wet”
interface surface, which may differ from the original one for curved geometries.
The integral evaluations in (7) are significantly simplified by taking NF and NS
to be delta functions collocated at the fluid and structure interface nodes, respec-
tively, a configuration illustrated in Fig. 10. (If meshes match, multipliers, frame and
boundary nodes coalesce as pictured in in Fig. 9.) If so BF and BS become Boolean
matrices that select and normal-project node boundary freedoms from the complete
state vectors: uBF D BF uF and uBS D BS uS . For interpolating frame displace-
ments uB we have so far only used piecewise linear, C 0 -continuous shape functions
NB with nodes placed according to distribution rules discussed in [29, 30]. Frame
interpolation spaces of lower (C 1 ) continuity remain to be investigated.
Setting the first variation ı˘BL to zero yields the three matrix equations

uBF D BF uF D LF uB ; uBS D BS uS D LS uB ; LTF F  LTS S D 0:

(8)

When adjoined to the fluid and structure uncoupled EOMs, these appear as the last
three matrix equations of the coupled EOM (2) in displacement coordinates. For
future use, introduce the matrices

QFF D LF LTF ; QSS D LS LTS ; RFF D LTF LF ; RSS D LTS LS : (9)

Inverses of the Q matrices appear in ensuing derivations. Since LF and LS are

generally rectangular for nonmatching meshes, one or more of (9) could become
singular, in which case ordinary inverses do not exist. In the equations below AG
denotes the Moore-Penrose generalized inverse of A, also popularly known as the
pseudoinverse [34].

4.5 LLM Interface Force-Motion Relations

Solving the last equation of (8) by least-squares methods yields

S D TSF F ; F D TF S S ; (10)

in which TSF D QG T G T

SS LS LF and TF S D QFF LF LS are called force transfer
matrices. Here pseudoinverses should be replaced by ordinary inverses if appropri-
ate. Products TF S TSF and TSF TF S are orthogonal projectors. Elimination of uB
A Classification of Interface Treatments for FSI 41

from the first two equations of (8): uBF D LF uB and uBS D LS uB , yields the
corresponding transformations between boundary displacement vectors:

uBS D USF F ; uBF D UF S S ; in which USF D TTF S ; UF S D TTSF :

(11)
Here USF D .LTS /G RSS RG T T
FF LF D TF S and UF S D .LF /
T G
RFF RG T
SS LS D
T T
TSF will be called motion transfer matrices. To prove that USF D TSF from lin-
ear algebra, start from the identity RG T T G
SS LS D LS QSS , premultiply both sides
by LF .LF / RFF and use projector properties. Likewise for UF S D TTSF . The
T T G

transformation duality (11) can be established more directly from work theorems,
noting that no energy is gained or lost at the interface, as follows. The comple-
mentary virtual work ıWB of interface displacements on their conjugate multiplier
variations is

ıWBS D uTBS ıS ; 
ıWBF D uTBF ıF ; ıWB D ıWBF
 
C ıWBS D 0:
(12)
Setting ıF D TF S ıS and uTBS D uTBF UTSF in ıWB D 0 gives uTBF .UTSF 
TF S / ıS D 0 for arbitrary uBF and ıS , whence TF S D UTSF . Setting ıS D
TSF ıF and uTBF D uTBS UTF S gives uTBS .UTF S  TSF / ıF D 0 for arbitrary
uBS and ıF , whence TSF D UTF S .

4.6 LLM Interface Patch Tests

The duality (11) holds for any LF and LS , even if those matrices were filled
with random numbers. Thus it is independent of the frame discretization choice,
or even of whether a frame is present. This generality can be used to advantage for
other treatments, e.g. the Mortar and Global DFMT methods derived through AIV
elimination. The placement of frame nodes may affect, however, the results of the
Interface Patch Test (IPT) as shown by the following example, which is a variant of
one studied in [30].
Consider the two coarse nonmatching 2D meshes pictured in Fig. 11(a). The
interface B of height H connects three bilinear rectangular fluid elements of height
1
3
H to two bilinear rectangular structural elements of height 12 H . All elements have
uniform thickness h. Of the four frame nodes, two (1 and 4) are placed at both
ends and two (2 and 3) symmetrically located at distance ˛ H from the middle
fluid boundary node positions, as indicated. Here  13  ˛  16 is a dimensionless
free parameter; conventionally ˛ > 0 if the middle frame nodes lie in the center
thirdspan, as drawn in the Fig. 11(a). If ˛ D 16 frame nodes 2 and 3 coalesce and
the frame has 3 nodes. If ˛ D  13 nodes 2 and 3 coalesce with the end nodes and
the frame has 2 nodes. The only displacement DOF considered are the interface
displacements shown in Fig. 11(b). These are collected in vectors

uTBS D ΠuS1 uS2 uS3  ; uTBF D ΠuF 1 uF 2 uF 3 uF 4  ;

uTB D ΠuB1 uB2 uB3 uB4  : (13)
42 C.A. Felippa et al.

Interface Frame (B)

uF1 uB1 uS1 λF1 λS1
4S
Note: 5F 1F 1B 1S
α>0 as
shown 6F 2F uF2 λF2
Structure (S) uB2
αH 2B uS2 λS2
H 2S
5S
αH uF3 λF3
3B uB3
7F 3F
Fluid (F)
8F 4F 4B 3S uF4 uB4 uS3 λF4 λS3
6S
x
(a) (b) (c)

Fig. 11 Example to illustrate interface patch test and Zero Moment Rule (ZMR).

The interface multiplier freedoms, displayed in Fig. 11(c), are collected in vectors

TS D ΠS1 S2 S3  ; TF D ΠF 1 F 2 F 3 F 4  : (14)

The connection and force-transfer matrices for arbitrary ˛, excluding the node-
coalescence cases ˛ D  13 and ˛ D 16 , are

12 0 0 0 3
2 3
6 3˛ 1 0 0 7 1 0 0 0
6 1 C 3˛ 1 C 3˛ 7
LF D 6
3˛ 7 ; LS D 4 0 1=2 1=2 0 5;
4 0 0 1 5
1 C 3˛ 1 C 3˛ 0 0 0 1
0 0 0 1
2 3˛ 3 2 3
1 1 C 3˛ 0 0 2 3˛ 0
6 7 6 0 1 C 3˛ 0 7
TSF D6 1 1
4 0 1 C 3˛ 1 C 3˛ 0 5 ;
7 TF S D 12 4 5: (15)
0 1 C 3˛ 0
0 0 3˛ 0 3˛ 2
1 C 3˛ 1
To check the Interface Force Patch Test (IFPT) one assumes that the 3 fluid
elements are under uniform pressure p, positive if compression. If displacement
shape functions vary linearly along the edge, as in the case of a bilinear fluid ele-
ment, the consistent interface fluid node forces are QfBF D  16 phH Π1 2 2 1 T .
Assuming a similar linear shape function variation over the structural elements and
uniform stress xx D p, others zero, the consistent structural node forces are
QfBF D 1 phH Π1 2 1 T . To apply the IFPT, set  O F D fBF , Q S D QfBS and use
4
the multiplier transformation equations (10) to compute

3 2 2 3
1  3˛
1 C 9˛
O F D pH h 4 4 5 ; O S D pH h 6 1 C 3˛ 7
S D TSF  F D TF S  4 5:
6.1 C 3˛/ 1 C 9˛ 4 1 C 3˛
1  3˛
(16)
A Classification of Interface Treatments for FSI 43

The IFPT is passed if S D  Q S and F D  Q F . Clearly this happens if and only if

1
˛ D 9 . This is the only 4-node-frame configuration that satisfies the Zero Moment
Rule (ZMR) [29, 30].
The interface rigid-motion patch test (IRMPT) checks whether a linearly varying
fluid boundary displacement field is correctly transmitted to the structure and vice-
versa. If d denotes uniform translation along x and  the rotation about z, the correct
node displacement values are uO BF D d Π1 1 1 1 T C 16 H Π3 1 1 3 T and
uO BS D d Π1 1 1 T C 12 H Π1 0 1 T . Application of the displacement transfor-
mations (11) gives a uBS D USF uO BF that reproduces uO BS exactly for any ˛. The
converse transformation, however, is exact only for ˛ D 16 .

5 Mortar Method

Since its inception in 1990 [5] the Mortar method has gained popularity as a inter-
facing scheme for multiphysics capable of handling non-matched meshes [3, 4, 6,
7, 12, 22, 40]. With growing acceptance the name has come to designate a set of
loosely related mesh coupling techniques. The common feature is the use of one
and only one “gluing” Lagrange multiplier field that directly links the two sides of
the interface (that is, without a kinematic frame). For coupling of an inviscid fluid
(in particular, an acoustic fluid) to an elastic structure the gluing field is a scalar,
which was called Bn in Sect. 4.3.
Mortar interface equations have been usually constructed from Galerkin or other
weighted residual methods, see e.g. [22]. This is inevitable for general fluid models,
which are not derivable from variational principles. In the present context, however,
we will restrict those equations to be based on the functional ˘BM of (4), a vari-
ational framework that preserves symmetry. The only field to be discretized is the
scalar multiplier function Bn on B . Physically this is the normal-to-the-interface
surface traction, also known as wet wall pressure.
In most of the published literature on Mortar methods, multipliers are distributed
functions interpretable as surface tractions. To make a fair comparison with LLM,
however, we take the multiplier space for Bn to be that of delta functions, as pic-
tured in Fig. 12; which can be viewed as interaction point forces. An important
question is, where should those forces be placed? For matching meshes the answer
is easy, see Fig. 12(a). For nonmatching meshes, as pictured in Fig. 12(b), the choice
is not obvious. One solution is to declare a face as master (typically that pertaining
to the finer mesh, as explicitly recommended in [38]), and collocate point forces
at the master nodes. This is illustrated in Fig. 12(c) with the fluid face picked as
master. A dual-master scheme, e.g. [21], collocates multipliers at all interface nodes
as shown in Fig. 12(d); this avoids master vs. slave decisions but may lead to sin-
gularity or ill-conditioning. Similar freedom placement decisions may be necessary
when using distributed Lagrange multiplier spaces. Regardless of choice, a master
mesh must know (in the nonmatching case) about the boundary shape functions of
the slave one and the modularity of the LLM treatment is lost.
44 C.A. Felippa et al.

? Fluid Fluid
Fluid Fluid
Structure Structure Structure Structure
?

?
λB
λB λB λB
Master
Slave

Matching meshes: Non-matching meshes: Non-matching meshes: Non-matching meshes:

only one multiplier normal several possibilities for multipliers at master multipliers at each
to interface needed since the ?, depending on face nodes (fluid chosen face node (dual
acoustic fluid is inviscid master-slave face choices as master in figure) master faces)

(a) (b) (c) (d)

Fig. 12 Mortar treatment with node-collocated delta function multipliers.

5.1 Mortar Equations Of Motion

From now on subscript n will be again omitted for brevity. With the discrete multi-
plier assumption written as B D NB B , insertion into ˘BM gives the discretized
functional  
O F  uTS BO S B :
˘BM ŒB  D uTF B (17)
R R
where BO F D B NTFB NB and BO S D B NTSB NB . Since NB consists of delta
functions, the integrals reduce to collocation on the master face and interpolation on
the other. For nonmatching meshes BO F and B O S are not Boolean and so will gener-
ally differ from the BF and BS of the LLM treatment. Setting ı ˘BM D 0 yields the
matrix connection equation BO TS uS D BO TF uF . For interpretation, let fBS D BO TS B
and fBF D BO TF B be the boundary force arrays conjugate to uBS and uBF ,
respectively. Hence fBS C fBF D 0, which expresses discrete interface force equi-
librium. Adjoining this to the dynamic equations of the uncoupled fluid and structure
gives the matrix equations of motion in terms of structural and fluid displacements.
If damping is ignored the EOMs are
2 38 9 2 38 9 8 9
MS 0 0 < uR S = KS 0 OS
B < uS = < f S =
4 0 MF 0 5 uR F C 4 0 KF BO F 5 uF D fF : (18)
:R ; : ; : ;
0 0 0 B BO TS BO TF 0 B 0

As before fluid irrotationality is enforced by the transformation uF D DF , where

collects displacement potential degrees of freedom at fluid mesh nodes. Carrying
out a congruential transformation on fluid freedoms yields
2 38 9 2 38 9 8 9
MS 0 0 < uR S = KS 0 BO S < uS = < fS =
4 0 MF 05 R C 4 0 KF BO F 5 D fF ;
:R ; : ; : ;
0 0 0 B O
BTS BO TF 0  B 0
(19)
A Classification of Interface Treatments for FSI 45

in which MF D DTF MF DF , KF D DTF KF DF , BO F n D DTF BF n and fF D

DTF fF . A response analysis can be carried out by the partitioned analysis procedure
described in [37]. The numerical stability of that procedure in conjunction with the
Newmark time integrator is studied in Appendix A of that reference.

5.2 Mortar Interface Force-Motion Relations

As done in Sect. 4.5 a relationship between the fluid boundary displacements uBF
and the structure boundary displacements uBS can be found on solving B O T uS D
S
BO F uF by least-square methods. Defining Q
T O FF D BO F BO and Q
T O SS D BO S BO T we
F S
have

O G BO F BO T uBS D TO F S uBS ;
uBF D Q O G BO S BO T uBF D T
uBS D Q O SF uBF :
FF S SS F
(20)
The dual force transformations

fBS D TSF fBF ; fBF D TF S fBS ; (21)

follow from interface energy conservation as follows. The complementary

 virtual

work ıWB over B is ıWB D uTBF ıfBF CuTBS ıfBS D uTBS TO TF S ıfBF CıfBS D
   
uTBS TO TF S  U
O SF ıfBF D uT TO T  U
BS FS
O SF BO T ıB D 0 for arbitrary uBS and
F
ıB , whence U O SF D TO T . Similarly U
O F S D TO T .
FS SF

5.3 Mortar Method Assessment

The Mortar method based on delta-function multipliers (point interaction forces)

produces simpler interaction equations than LLM, as a comparison of the coupled
EOMs (2)–(3) versus (18)–(19) makes obvious. For nonmatching meshes, forming
BO F and BO S involves more work than the Boolean BF and BS since shape function
interpolations are generally needed in the former, but there is no need for the LLM
connection matrices LS and LF . This simplicity, however, is counterbalanced by
two features:
(1) The interface force patch test may be violated if collocated face displacements
are used.
(2) Interaction equations may become singular at cross points (2D) or cross lines
(3D), if meshes are nonmatching.
Weakness (1) is illustrated by a 2D example given in [38], in which two iso-P
bilinear elements are linked to a 9-node biquadratic iso-P element. See Fig. 5 of that
reference. It can be also verified that the IRMTPT is also violated whence motion
transfer is incorrect. To pass the patch test it is necessary to pair the multipliers with
weighted node displacements.
46 C.A. Felippa et al.

Weakness (2) has no simple cure within the Mortar context. In fact it does not
depend on whether multipliers are distributed or lumped as point forces, because the
problem is caused by the lack of unique normal.

6 Direct Force-Motion Transfer Methods

The term Direct Force-Motion Transfer or DFMT was introduced in [38] as col-
lective label for a wide class of interfacing methods with historical and practical
importance. Their common feature is that no AIV are introduced. Boundary forces
and displacements (or velocities) are moved directly from fluid to structure and
vice-versa. It is convenient to distinguish two subclasses, qualified here as Local
and Global.

6.1 Local DFMT for Blast Simulations

As described in Sect. 2.4, local DFMT methods for FSI are the oldest as well as
the easiest to understand because they are based on simple physics. This matrix-
free process was used in simulations of submerged structures subject to underwater
shock (UWS) using staggered time integration. The initial implementation, studied
in [24], coupled a coarse BEM fluid model to a finer structural FEM mesh. BEM to
FEM force lumping was based on contributing wet-surface areas. Interface energy
conservation was not enforced or even monitored: this is unimportant in shock wave
transient simulations that span milliseconds. Despite that shortcoming this DFMT
performed well for the chief goal of assessing structural vulnerability.
For the hull-cavitation simulations of [14] a fluid-volume FE mesh was placed
between the structure and the BEM model operating as silent boundary; see Fig. 5.
Many local DFMT flavors exist. For expediency in [38] they were identified by
the code in which implemented. Examples: DFMT-CFA and DFMT-CASE for
those used in the UWS code CFA (Cavitation Fluid Analyzer) of [14] and CASE
(Cavitation Spectral Elements) of [39], respectively.

6.2 Consistent Interpolation

A similar transfer method was initially used by Farhat and his team [9, 10] for
the aeroelasticity problem. It was driven by the need to determine aircraft sur-
face boundary forces from the aerodynamic pressure provided by a gas-dynamics
fluid-volume (FV) code. Here the fluid pressure may experience rapid changes in
space, especially with Navier-Stokes fluid models. Unlike UWS simulations, the
near field fluid mesh is typically finer than the structure mesh to facilitate capture of
those spatial gradients. The aeroelasticity DFMT process incorporated refinements:
fluid pressures were transferred to structural mesh Gauss points, from which struc-
ture node forces are obtained by quadrature. The motion transfer is carefully done
A Classification of Interface Treatments for FSI 47

by consistent interpolation, as described in more detail below. Application of this

method to the example of Fig. 11 gives the following fluid-to-structure force transfer
fBS D TSF fBF and the structure-to-fluid motion transfer uBF D UF S uBS :
8 9 2 1 1 0 0 3 8 fF 1 9
< fS1 = 3 ˆ
< >
=
4 2 2 5 fF 2
fBS D fS2 D  0 3 3 0 ;
: ; 1 :̂ fF 3 >
;
fS3 0 0 3 1 fF 4
8 9 2 3
u 1 0 0 8 9
ˆ
< F1 > = 6 1 3 0 7 < uS1 =
uF 2
uBF D D6 4 4 7
4 0 3 1 5 : uS2 ; : (22)
:̂ uF 3 >
; 4 4 uS3
uF 4 0 0 1

It is easily checked that (22) pass both the stress and rigid-motion patch tests. How-
ever since TSF is not the transpose of UF S , interface energy conservation is not
verified. In aeroelastic calculations that may span minutes of real time, correct inter-
face energy balance is important if the principal aim is detecting flutter or divergence
onset, so as to establish flight envelopes. Accordingly the original transfer scheme
was modified [11, 32] into the Consistent Interpolation (CI) method as follows.
First, motion transfer was carefully implemented to take into account peculiari-
ties of FV fluid codes. Each FV grid point j on the fluid boundary is paired with
the closest structural element e S (or elements if j is equally distant from more than
one). The structural element natural coordinates j of j (or of its projection on the
structure element if j is offset from the interface) are determined. (In cell-centered
FV schemes grid point j may be offset from the fluid-structure interface B ; in that
case the projection on B is used.) The fluid displacement uFj is determined using
the structural element shape functions evaluated at j . A structure-to-fluid displace-
ment transformation matrix UF S is built by repeating this procedure over all j so
that uFB D UF S uSB . Second, the transformation from fluid to structure forces:
fSB D TSF fSB , follows from duality: TF S D UTSF . This enforces interface
energy conservation. In typical aeroelastic problems the fluid mesh is more refined
then the structure. Thus, the CI method interpolates refined mesh values from coarse
mesh values, which helps to produce well conditioned coupling matrices.
Application of the CI method to the example of Fig. 11 yields
8 9 2 1 1 0 0 3 8 fF 1 9
< fS1 = 3 ˆ
< >
=
fF 2
fBS D fS2 D  4 0 23 32 0 5 ;
: ; 1 :̂ fF 3 >
;
fS3 0 0 3 1 fF 4
8 9 2 3
u 1 0 0 8 9
ˆ
< F1 > = 6 1 2 0 7 < uS1 =
uF 2
uBF D D6 3 3 7
4 0 2 1 5 : uS2 ; : (23)
:̂ uF 3 >
; 3 3 uS3
uF 4 0 0 1
48 C.A. Felippa et al.

To check the interface force patch test (IFPT), apply uniform pressure p over
the 3 fluid elements. Then fBF D  16 phH Π1 2 2 1 T and fBS D .1=18/phH
Œ 5 8 5 T . Since fBS ¤ QfBS D 14 phH Œ 1 2 1 T , the test is not passed. Notice
that the transformation matrices in (23) are the same as those provided by LLM
treatment in the example of Sect. 4.6 if ˛ D 16 . This configuration does not sat-
isfy the Zero Moment Rule (ZMR) stated in [30]. Thus it is not surprising that the
IFPT fails.

6.3 Global DFMT Methods

Aother subclass of DFMT is that of global DFMT methods. These are not based on
local interface physics. They are instead built in two stages. First a multiplier based
discretization such as LLM or point-force Mortar is constructed. Interface unknowns
are then eliminated by least-square methods as described in Sects. 4.5 and 5.2 to
yield the force transfer matrices TSF and TF S . The motion transfer matrices follow
from duality.
The qualifier global indicates that transfer matrices are generally fully populated,
meaning that each interface DOF is coupled to every other one. By construc-
tion interface energy conservation is satisfied a priori, but interface patch tests
are not necessarily passed, as previous examples make clear. The following Table
summarizes method labeling used for the application examples discussed in [38].
Label Fluid-structure interface treatment
LLM LLM with step-by-step solution of interface equations
Mortar Point-force Mortar with step-by-step solution of interface equations
DFMT-CASE Local DFMT procedure used in CASE spectral code [39]
DFMT-LLM Global DFMT with LLM-derived transfer matrices (10) and (11)
DFMT-Mortar Global DFMT with Mortar-derived transfer matrices (20) and (21)
Do LLM and DFMT-LLM produce identical results? Only under very special
conditions. For instance, if connection matrices LF and LS are square and of full
rank. Otherwise the least-squares elimination of interface unknowns can be expected
to work as a low pass filter that projects interface patterns on the column span of
the Q matrices. This will typically mollify the computed response. A similar remark
applies to Mortar versus DFMT-Mortar. Whether this kind of filtering is acceptable
or desirable can be expected to be problem and goal dependent.

7 Conclusions

The Introduction of the survey article [17] notes that one obstacle to rapid progress
in computational multiphysics is combinatorial complexity in method design and
implementation. Restricting ourselves to the FSI problem, model-based simulation
involves making decisions on several computational ingredients:
S Structure spatial model
F Fluid spatial model
A Classification of Interface Treatments for FSI 49

FSI Fluid-structure interface treatment

TS Time stepping
Aux Auxiliary components for some problems, especially those involving
exterior domains; for example silent boundaries, ALE mesh mover, : : :
One difficulty in coming to terms with ingredients FSI and TS is agreement on
terminology, especially as regards the former. (As noted in the Introduction, some
convergence on the classification of time stepping methods appears to be emerg-
ing.) Although the literature on coupled problem simulation is growing steadily,
lack of coherent nomenclature may hinder the impact of individual contributions. It
is hoped that the proposed classification of interface treatments in Sect. 2 may help
to alleviate textual obstacles to result dissemination. More specifically, it is hoped
that the proposed terminology and in particular understanding the “orthogonality”
between FSI and TS, may foster comparisons and correlation of results obtained by
separate research teams.
Assuming some consensus on terminology is eventually achieved, another area
that deserves attention is the association of desirable attributes with specific cou-
pled problems, project objectives and model choices. From the long list compiled
in Sect. 3 it should be obvious that no “universal” method that possesses even a
majority of those attributes can be hoped to be found. Consequently, prioritization
by linking attributes to methods and objectives is necessary.
To conclude, the following FSI research areas seem worth exploring:
 Develop interface treatments that satisfy the largest number of conservation
conditions for arbitrary nonmatching meshes, and examine related tradeoffs.
 Study which attributes are inherited in the hierarchical “cascading” flowcharted
of Fig. 4.
 Investigate links between interface treatments and stabilityCaccuracy of parti-
tioned time-stepping procedures.
 Find out which interface treatment(s) best support reliable and effective energy
error measures for adaptive timestep control.
 Find effective LLM frame configuration rules for interacting three-dimensional
nonmatching meshes, thus generalizing the ZMR of [30] to 3D.
 Study the possibility of “localizing” global DFMT treatments by matrix filtering
to generate more computationally efficient local DFMT schemes.
 Investigate the suitability of the various interface treatments to support Reduced
Order Models for one or more problem components.
 Generalize dual and primal-dual interface treatments to handle multiscale effects.

Acknowledgements The first author (CAF) expresses his thanks to the organizers of the Inter-
national Workshop on Computational Engineering 2009 for their invitation to present a keynote
lecture and write the present contribution. The contribution of the second author (KCP) was partly
supported by the WCU Program of the Korea Science and Engineering Foundation funded by the
Ministry of Education, Science and Technology, Republic of Korea, through Grant R31-2008-000-
10045-0. The work of the third author (MRR) was part of his doctoral dissertation while at the
Aerospace Engineering Sciences Department, University of Colorado at Boulder. That research
was supported by the US National Science Foundation under Grant High-Fidelity Simulations for
Heterogeneous Civil and Mechanical Systems, CMS-0219422.
50 C.A. Felippa et al.

References

1. Anonymous. COSMOS Virtual Data Center for Strong Motion, Consortium of Organizations
for Strong Motion Observation Systems (COSMOS), Pacific Earthquake Engineering Research
Center. Available from http://db.cosmos-eq.org, University of California, Berkeley.
2. S. N. Atluri. On “hybrid” finite-element models in solid mechanics, in: Advances in Computer
Methods for Partial Differential Equations. Ed. by R. Vichnevetsky, AICA, Rutgers University,
346–356, 1975.
3. F. Baaijens. A fictitious domain/mortar element method for fluid-structure interaction. Int. J.
Numer. Meths. Fluids. 35, 743–761, 2001.
4. F. Belgacem. The mortar finite element method with Lagrange multipliers. Numer. Math.. 84,
173–197, 1999.
5. C. Bernardi, Y. Maday, and A. T. Patera. A new nonconforming approach to domain decom-
position: the mortar element method. Technical report, Université Pierre at Marie Curie, Paris,
France, 1990.
6. D. Braess, W. Dahmen and C. Wieners. A multigrid algorithm for the mortar finite element
method. SIAM Journal on Numerical Analysis. 37, 48–69, 2000.
7. F. Casadei, E. Gabellini, G. Fotia, F. Maggio and A. Quarteroni. A mortar spectral/finite ele-
ment method for complex 2D and 3D elastodynamic problems. Comp. Meth. Appl. Mech.
Engrg. 191, 5119–5148, 2002.
8. C. Farhat and F.-X. Roux. Implicit parallel processing in structural mechanics, Comput. Mech.
Advances, 2, 1–124, 1994.
9. C. Farhat, M. Lesoinne and N. Maman. Mixed explicit/implicit time integration of coupled
aeroelastic problems: three-field formulation, geometric conservation and distributed solution.
Int. J. Numer. Meth. Engrg., 21, 807–835, 1995.
10. C. Farhat, S. Piperno and B. Larrouturu. Partitioned procedures for the transient solution of
coupled aeroelastic problems; Part I: model problem, theory and two-dimensional application.
Comp. Meths. Appl. Mech. Engrg., 124, 79–112, 1995.
11. C. Farhat, M. Lesoinne and P. LeTallec. Load and motion transfer algorithms for fluid/structure
interaction problems with nonmatching discrete interfaces: Momentum and energy conserva-
tion, optimal discretization and application to aeroelasticity. Comp. Meths. Appl. Mech. Engrg.,
157, 95–114, 1998.
12. V. Faucher and A. Combescure. A time and space mortar method for coupling linear modal
subdomains and non-linear subdomains in explicit structural dynamics. Comp. Meths. Appl.
Mech. Engrg., textbf192, 509–533, 2003.
13. C. A. Felippa and K. C. Park. Staggered transient analysis procedures for coupled dynamic
systems: formulation, Comp. Meths. Appl. Mech. Engrg., 24, 61–112, 1980.
14. C. A. Felippa and J. A. DeRuntz. Finite element analysis of shock-induced hull cavitation,
Comp. Meths. Appl. Mech. Engrg., 44, 297–337, 1984.
15. C. A. Felippa and T. L. Geers. Partitioned analysis of coupled mechanical systems, Engrg.
Comput., 5, 123–133, 1988.
16. C. A. Felippa, K. C. Park and C. Farhat. Partitioned analysis of coupled mechanical systems,
Comp. Meths. Appl. Mech. Engrg., 190, 3247–3270, 2001.
17. C. A. Felippa and K. C. Park. Model-based partitioned analysis of coupled problems, chapter
4 in Computational Aspects of Structural Dynamics and Vibrations, ed. by G Sandberg and R.
Ohayon, CISM Courses and Lectures, vol. 505, Springer-Verlag, Berlin, 2008, 171–216.
18. T. L. Geers. Residual potential and approximate methods for three-dimensional fluid-structure
interaction, J. Acoust. Soc. Am., 49, 1505–1510, 1971.
19. T. L. Geers. Doubly asymptotic approximations for transient motions of general structures,
J. Acoust. Soc. Am., 64, 1500–1508, 1978.
20. T. L. Geers. Boundary element methods for transient response analysis, in: Chapter 4 of
Computational Methods for Transient Analysis, ed. by T. Belytschko and T. J. R. Hughes,
North-Holland, Amsterdam, 221–244, 1983.
A Classification of Interface Treatments for FSI 51

21. B. Herry, L. Di Valentin and A. Combescure. An approach to the connection between subdo-
mains with nonmatching meshes for transient mechanical analysis. Int. J. Numer. Meth. Engrg.,
55, 973–1003, 2002.
22. L. A. Jakobsen. A finite element approach to analysis and sensitivity analysis of time dependent
fluid-structure interaction systems. Ph.D. Dissertation, Aalborg University, Denmark, 2002.
23. M. C. Junger and D. Feit, Sound, Structures and Their Interaction, MIT Press, Cambridge,
Massachussets, 1972.
24. K. C. Park, C. A. Felippa and J. A. DeRuntz. Stabilization of staggered solution procedures
for fluid-structure interaction analysis, in: Computational Methods for Fluid-Structure Inter-
action Problems, ed. by T. Belytschko and T. L. Geers, AMD Vol. 26, American Society of
Mechanical Engineers, New York, 95–124, 1977.
25. K. C. Park. Partitioned transient analysis procedures for coupled-field problems: stability
analysis, J. Appl. Mech., 47, 370–376, 1980.
26. K. C. Park and C. A. Felippa. Partitioned analysis of coupled systems, Chapter 3 in Computa-
tional Methods for Transient Analysis, T. Belytschko and T. J. R. Hughes, eds., North-Holland,
Amsterdam–New York, 157–219, 1983.
27. K. C. Park and C. A. Felippa. A variational principle for the formulation of partitioned
structural systems, Int. J. Numer. Meth. Engrg., 47, 395–418, 2000.
28. K. C. Park, C. A. Felippa and R. Ohayon. Partitioned formulation of internal fluid-structure
interaction problems via localized Lagrange multipliers, Comp. Meths. Appl. Mech. Engrg.,
190, 2989–3007, 2001.
29. K. C. Park, C. A. Felippa and G. Rebel, A simple algorithm for localized construction of
non-matching structural interfaces, Int. J. Numer. Meth. Engrg., 53, 1261–1285, 2002.
30. K. C. Park, C. A. Felippa and G. Rebel. Interfacing nonmatching finite element discretizations:
the Zero Moment Rule. In Trends in Computational Mechanics, ed. by W. A. Wall et. al.,
CIMNE, Barcelona, Spain, 355–367, 2001.
31. T. H. H. Pian and P. Tong. Basis of finite element methods for solid continua, Int. J. Numer.
Meth. Engrg., 1, 3–29, 1969
32. S. Piperno and C. Farhat. Partitioned procedures for the transient solution of coupled aeroe-
lastic problems - Part II: energy transfer analysis and three-dimensional applications. Comp.
Meth. Appl. Mech. Engrg., 190, 3147–3170, 2001.
33. W. Prager. Variational principles for linear elastostatics for discontinous displacements, strains
and stresses, in Recent Progress in Applied Mechanics, The Folke-Odgvist Volume, ed. by B.
Broger, J. Hult and F. Niordson, Almqusit and Wiksell, Stockholm, 463–474, 1967.
34. C. R. Rao and S. K. Mitra. Generalized Inverse of Matrices and its Applications. Wiley, New
York, 1971.
35. G. Rebel, K. C. Park, C. A. Felippa. A contact formulation based on localized Lagrange multi-
pliers: formulation and applications to two-dimensional problems, Int. J. Numer. Meth. Engrg.,
54, 263–297, 2002.
36. M. R. Ross. Coupling and simulation of acoustic fluid-structure interaction systems using
localized Lagrange multipliers. PhD thesis, University of Colorado at Boulder, 2006. Available
from http://caswww.colorado.edu/courses/FSI.d
37. M. R. Ross, C. A. Felippa, K. C. Park and M. A. Sprague. Treatment of acoustic fluid-structure
interaction by localized Lagrange multipliers: Formulation, Comp. Meths. Appl. Mech. Engrg.,
197, 2008, 3057–3079.
38. M. R. Ross, M. A. Sprague, C. A. Felippa and K. C. Park. Treatment of acoustic fluid-structure
interaction by localized Lagrange multipliers and comparison to alternative interface coupling
methods, Comp. Meths. Appl. Mech. Engrg., 198, 2009, 986–1005.
39. M. A. Sprague and T. L. Geers. A spectral-element method for modeling cavitation in transient
fluid-structure interaction. Int. J. Numer. Meth. Engrg.. 60, 2467–2499, 2004.
40. B. Wohlmuth. Hierarchical a posteriori error estimators for finite element methods. SIAM
Journal of Numerical Analysis. 36, 1636–1658, 1999.
•
Computer Modeling and Analysis of the Orion
Spacecraft Parachutes

K. Takizawa, C. Moorman, S. Wright, and T.E. Tezduyar

Abstract We focus on fluid-structure interaction (FSI) modeling of the ringsail

parachutes to be used with the Orion spacecraft. The geometric porosity of the
ringsail parachutes with ring gaps and sail slits is one of the major computational
challenges involved in FSI modeling. We address the computational challenges
with the latest techniques developed by the Team for Advanced Flow Simula-
tion and Modeling (T?AFSM) in conjunction with the Stabilized Space–Time
Fluid–Structure Interaction (SSTFSI) technique. We investigate the performance
of the three possible design configurations of the parachute canopy, carry out
parametric studies on using an over-inflation control line (OICL) intended for
enhancing the parachute performance, discuss rotational periodicity techniques for
improving the geometric-porosity modeling and for computing good starting condi-
tions for parachute clusters, and report results from preliminary FSI computations
for parachute clusters. We also present a stability and accuracy analysis for the
Deforming-Spatial-Domain/Stabilized Space–Time (DSD/SST) formulation, which
is the core numerical technology of the SSTFSI technique.

1 Introduction

Various types of fluid–structure interaction (FSI) problems have been addressed and
numerous FSI solution techniques have been developed in recent decades (see, for
example, [2–11, 15, 17–33, 35, 38, 40, 42–45, 50, 54–57, 59–64, 66–78]). The Team
for Advanced Flow Simulation and Modeling (T?AFSM) has addressed many of the
challenges involved in FSI modeling of parachutes (see [20,35–37,39,41,45,57–59,
61,66]), with parallel, 3D computations going as far back as 2000. The core technol-
ogy used in all this FSI modeling is the the Deforming-Spatial-Domain/Stabilized
Space–Time (DSD/SST) formulation [46, 47, 51, 52], which was introduced as a

K. Takizawa, C. Moorman, S. Wright, and T.E. Tezduyar

Mechanical Engineering, Rice University, 6100 Main Street, Houston, TX 77005, USA
e-mail: [email protected]

H.-J. Bungartz et al. (eds.), Fluid Structure Interaction II, Lecture Notes 53
in Computational Science and Engineering 73, DOI 10.1007/978-3-642-14206-2 3,
c Springer-Verlag Berlin Heidelberg 2010
54 K. Takizawa et al.

general-purpose interface-tracking (moving-mesh) technique for flows with mov-

ing boundaries and interfaces, including FSI. In early years of parachute modeling
by the T?AFSM, a block-iterative FSI coupling technique [48] (see [54, 57] for
the terminology) was used (see, for example, [20, 35, 36, 45]). The quasi-direct FSI
coupling technique was introduced in [56, 57] and became part of the core tech-
nology used in the subsequent parachute FSI simulations of the T?AFSM (see,
for example, [49, 57, 61, 62]). The stabilized space–time FSI (SSTFSI) technique
was introduced in [54]. It is based on the new-generation DSD/SST formulations,
which were also introduced in [54], increasing the scope and performance of the
space–time FSI techniques developed earlier. The SSTFSI technique is now the core
technology used in the parachute FSI computations of the T?AFSM (see, for exam-
ple, [54, 58, 59]). A number of special FSI techniques were introduced in [54, 59]
in conjunction with the SSTFSI technique and the DSD/SST formulation, including
the the Homogenized Modeling of Geometric Porosity (HMGP). With the HMGP,
we bypass the intractable complexities of the geometric porosity by approximating
it with an “equivalent”, locally-varying fabric porosity.
The HMGP and some of the other special FSI techniques were motivated by the
task the T?AFSM has undertaken: computer modeling and analysis of the ring-
sail parachutes to be used with the Orion spacecraft. In this paper we focus on the
performance analysis of those parachutes, comparison of different canopy design
configurations, parametric studies on using an over-inflation control line (OICL)
considered for enhancing the parachute performance, and modeling of parachute
clusters. We also describe the FSI technique we have introduced for more effec-
tive parachute modeling and analysis, such as techniques for building a consistent
starting condition for the FSI computations and rotational-periodicity techniques
for improving the HMGP and for computing good starting conditions for parachute
clusters. We include a stability and accuracy analysis for the DSD/SST formulation.

2 Starting Condition

A consistent starting condition is essential for making accurate comparisons in many

applications using FSI modeling. Starting conditions are especially important when
investigating the transient response. A number of techniques for building FSI start-
ing conditions are reported in [58, 66]. These techniques mostly focus on starting
the FSI computations gently. The purpose of further influencing the starting con-
dition with the methods introduced here is primarily related to making the starting
conditions consistent and matching the physical conditions observed during NASA
drop tests. To build an appropriate starting point for comparing performance of var-
ious parachute designs, we first analyzed the parachute drop test data and compared
the test results to our earlier computations [66]. Based on this, we concluded that
a fully inflated parachute with all sails behaves as follows: the parachute exhibits
a periodic breathing motion caused by vortex shedding, and this dynamic nature
results in a fluctuating descent speed. Furthermore, during drop tests for a given
Computer Modeling and Analysis of the Orion Spacecraft Parachutes 55

Table 1 Computational result reported in [66]. V and T denote velocities and periods. Subscripts
“D” and “RH” denote descent and relative horizontal. Subscripts “B” and “S” denote breathing
and swinging. NA: Not applicable.
VD (ft/s) VRH (ft/s) TB (s) TS (s)
Computation 22.0 4 to 12 7.0 NA

range of altitudes, there is a corresponding wind direction, meaning the parachute

sees a variable wind. In steady wind, however, the parachute settles to a nearly con-
stant gliding direction after a few seconds. Moreover, when this is the case, the
speed of the parachute relative to the surrounding air is the dominant consideration
for the flow condition, and any lateral force the parachute sees from the wind is only
a function of the relative speed. The computation reported in [66] does not include
any side wind, but the parachute exhibits a comparable relative velocity due to the
parachute gliding. Table 1 provides the earlier computations reported in [66], which
are in close agreement with the drop test data.
Based on the argument above, we infer that the main influence of a side wind
is inducing a payload swing. To induce a payload swing in our computations how-
ever, instead of adding a variable wind to our model, we introduce an instantaneous
“payload swing” effect by giving the payload a one-time horizontal-velocity hike.
We now provide more details on the sequence for arriving at a starting point
for FSI computations. After 100 s of “symmetric FSI”, which is described in [66],
the parachute reaches a settled periodic (“breathing”) stage, which we deem to be
consistent between the various parachute designs computed. The payload and the
parachute have no horizontal speed at the end of the “symmetric FSI” step, which
does not match what is observed in the tests. To imitate the naturally-occurring
swinging motion, we instantaneously hike the horizontal speed of the payload to
20 ft/s. Simultaneously, we begin the de-symmetrization (see [66]) using a Cosine
form which lasts for one breathing period (7 s). Although the vortex shedding
pattern behind the parachute is not exactly the same in each computation, the
momentum added to the payload with this velocity hike is consistent. We believe
imitating the payload momentum in this fashion represents this aspect of the starting
condition with a reasonable closeness to the actual conditions.

3 Various Canopy Configurations

The steady descent of a ringsail parachute is dynamic. Through this investigation we

seek to reduce oscillations in the descent speed of the payload and reduce the hori-
zontal gliding of a single parachute. The oscillations in descent speed have a direct
impact on the maximum descent speed constraint, and the horizontal gliding of the
parachute has cluster performance implications. Adjusting the geometric porosity
by reconfiguring the canopy can impact both stability and drag performance of the
parachute. The baseline and two alternate canopy configurations are investigated.
The two alternate configurations are “missing” the 5th and 11th sail, respectively,
56 K. Takizawa et al.

Table 2 Number of nodes and elements for each of the three parachute configurations. Here nn
and ne are number of nodes and elements, respectively. The structural mechanics mesh consists
of four-node quadrilateral membrane elements, two-node cable elements and one-node payload
element. The structure interface mesh consists of four-node quadrilateral elements. The fluid vol-
ume mesh consists of four-node tetrahedral elements, while the fluid interface mesh consists of
three-node triangular elements.
PA PM5 PM11
Structure
nn 30;722 28;642 28;082
Membrane ne 26;000 24;080 23;600
Cable ne 12;521 11;401 12;121
Payload ne 1 1 1
Interface nn 29;200 27;120 26;560
ne 26;000 24;080 23;600
Fluid
Volume nn 178;270 192;412 180;917
ne 1;101;643 1;192;488 1;119;142
Interface nn 2;140 2;060 2;060
ne 4;180 3;860 3;860

when numbered starting from the top including the rings. We will refer to the three
parachutes as “PA” (all sails are in place), “PM5” (missing the 5th sail) and “PM11”
(missing the 11th sail).

3.1 Computational Conditions

The computational conditions are the same as those described in [66]. The number
of nodes and elements are given in Table 2. The payload weight is approximately
5,570 lbs. The total weights, including the parachute for each of the three configu-
rations: PA, PM5 and PM11, are approximately 5,725 lbs, 5,720 lbs and 5,715 lbs,
respectively. Fig. 1 shows the parachute shape and structure mesh for each configu-
ration. The homogenized porosity distribution used here is based on the computation
reported earlier by the T?AFSM, which is given in Table 1 in [66]. The porosity at
the edges facing a missing sail is calculated in the same way as the porosity at the
edges of the vent and skirt. The porosity coefficient for the edge nodes is set to
the fabric porosity, linearly progressing to the homogenized value for the adjacent
patch. Figure 2 shows the porosity distribution for each of the three cases.

3.2 Results

An important dynamical feature in parachute flight is the breathing in steady

descent. The flow separates near the skirt of the parachute and a ring vortex forms.
Computer Modeling and Analysis of the Orion Spacecraft Parachutes 57

Fig. 1 Parachute shape and structure mesh for PA, PM5 and PM11, shown from left to right and
top to bottom.

The vortex creates a large pressure differential near the skirt of the parachute at
maximum inflation. The vortex then moves upward causing the maximum pressure
differential to move from the skirt to the crown. Figure 3 shows the parachute and
flow field for each configuration when the tilt angle, as measured between the ver-
tical axis and a line connecting the payload and vent, is at a maximum. Table 3
provides the results for the three parachute configurations.
Figure 4 shows the payload descent speed and skirt diameter of the three
parachutes. The descent speed is directly dependent on the diameter because drag
production is dictated by the projected area of the canopy. It can be seen from the
plots that the PA and PM5 have similar maximum diameters. Though the diameters
are close, PM5 has a slightly higher average descent speed due to the decreased
projected area with a missing sail. The similar maximum diameter is due to the fact
that the pressurization of the lower sails for PA and PM5 are unaffected by the mod-
ification. The PM11 parachute drag performance is hindered largely by a lack of
projected area due to the loss of pressurization of the bottom three sails. Figure 5
58 K. Takizawa et al.

314

40

Fig. 2 Porosity distribution (in CFM) applied to the fluid interface for the three different parachute
geometries, “PA”, “PM5” and “PM11”, from left to right and top to bottom.

shows the horizontal-velocity magnitudes for the three configurations. In this case
we define improved static stability as having lower gliding speed. PA and PM5 have
larger gliding speeds while the glide speed of PM11 is very low. The static stability
of PM11 is much improved over that of PM5 and PA. This may have implications
in parachute cluster applications. The horizontal-velocity magnitude for the PM11
payload exhibits sharp kinks near 0 ft/s, indicating that the horizontal velocity is
reversing direction. This reversal happens when, in a swinging cycle, the tilt angle
is maximum. On the other hand, PA and PM5 payloads are moving on an ellipti-
cal trajectory with respect to the vent, as clearly seen in Fig. 6, and therefore do
not exhibit sharp kinks. Overall, the study shows that the more stable configura-
tions exhibit a loss of drag. However, it shows that geometric porosity can influence
the descent characteristics. Therefore a parachute’s performance might be tuned by
readjusting the geometric porosity.
Computer Modeling and Analysis of the Orion Spacecraft Parachutes 59

Fig. 3 Parachute and flow field for PA. PM5 and PM11 when the tilt angle is at a maximum.

Table 3 Computational results for the three parachute configurations.

VD (ft/s) VRH (ft/s) TB (s) TS (s)
PA 21.4 4 to 13 6.7 16.4
PM5 24.0 4 to 13 5.8 16.6
PM11 29.0 0 to 4 NA 17.0

4 Over-Inflation Control Line (OICL)

From drop tests and computations, we observe that the change in projected area due
to canopy breathing causes large fluctuations in descent speed. We test several over
inflation control lines (OICL) attached at the canopy skirt, which limit inflation.
Figure 7 shows the configuration of the OICL.
60 K. Takizawa et al.

32
30

Speed (ft/s)
28
26
24
22
20
18
16
0 20 40 60 80 100 120
Time (s)
PA PM5 PM11
95

90
Diameter (ft)

85

80

75

70

65
0 20 40 60 80 100 120
Time (s)
PA PM5 PM11

Fig. 4 Payload descent speed and parachute skirt diameter for the three configurations. The thin
vertical line at 7 s marks the end of the de-symmetrization.

4.1 Computational Conditions

The computational conditions are the same as those described in [66]. The number
of nodes and elements are given in Table 2. The PA canopy configuration is used
for this study. We present results for 76, 80 and 84 ft OICL cases. In the case of an
OICL smaller than the starting skirt diameter, for example 76 ft, we first compute
symmetric FSI with a zero-stiffness OICL. Then we turn on the stiffness when the
diameter is at a minimum. In this way we avoid the sudden changes of the internal
force balance for the structure.

4.2 Results

Figure 8 shows the parachute and flow field for PA and OICL1 , each at maximum
diameter. The descent speed and diameter are shown in Fig. 9. The average inflated

1
The acronym OICL is used also to indicate “PA with OICL”
Computer Modeling and Analysis of the Orion Spacecraft Parachutes 61

30

25

Speed (ft/s)
20

15

10

0
0 20 40 60 80 100 120
Time (s)
PA PM5 PM11
16
14
Speed (ft/s)

12
10
8
6
4
2
0
0 20 40 60 80 100 120
Time (s)
PA PM5 PM11

Fig. 5 Horizontal-velocity magnitude for the payload (top) and vent (bottom) for the three
parachute configurations. The thin vertical line at 7 s marks the end of the de-symmetrization.

diameter of the unrestricted canopy is 79 ft. We demonstrate that an OICL smaller

than the average inflated diameter causes breathing to cease completely. Meanwhile,
breathing continues in a damped fashion if the OICL is larger than the average
diameter. The descent speed illustrated in Fig. 9 is the superposition of several
parachute-dynamics factors. The first one is force oscillations produced by the vary-
ing area of the canopy, which leads to acceleration of the entire system. The second
one is the geometric coupling between the payload and parachute. As the parachute
expands in diameter, the payload moves toward the canopy and vice versa. The third,
and the dominant, one is the payload swing.
Figure 10 shows the horizontal-velocity magnitudes for the three OICL cases.
The magnitude of the canopy horizontal velocity suggests that the OICL has lit-
tle impact on the static stability of the parachute. Figure 11 shows the axial force
and projected area. It can be seen that the OICL smooths the oscillations in pro-
jected area and hence the axial force. The OICL accomplishes its goal of damping
the breathing oscillations that cause descent speed spikes, which are undesirable in
terms of the maximum descent speed constraint.
We propose that the descent speed of the payload be examined using a different
approach. The goal of this effort was to reduce the oscillations in descent speed due
62 K. Takizawa et al.

15

(ft)
0

-15
-60 -30 0 30 60
(ft)

15
(ft)

-15
-30 0 30
(ft)
15
(ft)

-15
-60 -30 0 30 60
(ft)

Fig. 6 Payload trajectories (relative to the vent) for PA, PM5 and PM11. Lines are drawn from
60 to 120 s. Dots are placed every 2.3 s. We note that the y scale is stretched to twice that of the x
scale.

to breathing. This has been accomplished, but the dominant swinging factor over-
whelms the descent speed data. We suggest attempting to separate out the dynamical
factors that contribute to the overall payload descent speed to provide a more clear
metric for comparison of the parachutes. One has to be careful in trying to choose an
optimum value for the OIC, because a shorter OICL reduces average drag as well as
descent speed peaks. Reducing the average drag too much would result in a descent
speed with little or no oscillation but a maximum descent speed that exceeds the
oscillatory peaks of the less constrained cases. Future analysis of the descent data
will provide more insight into the optimum length of the OICL.
Computer Modeling and Analysis of the Orion Spacecraft Parachutes 63

Fig. 7 OICL (red) is attached to the suspension lines. Views from outside (left) and inside (right).

Fig. 8 Parachute and flow field for PA and OICL (76 ft).

5 Rotational Periodicity

In general, periodic domains are useful for reducing computational cost. We bene-
fit from that by applying rotational periodicity to a highly refined parachute model.
In the case of parachute clusters, we also benefit from another aspect of periodic
domains, namely, their ability to stabilize the computation prior to full modeling
by enforcing solution periodicity. We already use similar constraints to stabilize
pre-FSI parachute computations. For example, we are using symmetric FSI to gen-
erate a good starting condition. A symmetric FSI computation is more likely to
remain stable because of the symmetric deformation of the parachute, even if we
apply a local artificial perturbation, such as the instantaneous horizontal velocity
hike described in Sect. 2. By imposing a similar stabilizing constraint, periodic
domains are useful for pre-FSI computations.
64 K. Takizawa et al.

32
30

Speed (ft/s)
28
26
24
22
20
18
16
0 20 40 60 80 100 120
Time (s)

95
Diameter (ft)

90
85
80
75
70
65
0 20 40 60 80 100 120
Time (s)

PA OICL80
OICL76 OICL84

Fig. 9 Payload descent speed and parachute skirt diameter for PA and the three OICL cases. The
thin vertical line at 7 s marks the end of the de-symmetrization.

5.1 Periodic n-gore Model

Here our goal is to carry out a fluid mechanics computation with the actual geo-
metric porosity. We use a periodic domain to reduce the computational cost of such
high refinement. We realize that the parachute wake and deformation are completely
asymmetric. Moreover, the parachutes glides horizontally. Therefore, a periodic
assumption is not suitable in terms of global quantities, such as descent speed and
drag coefficient. However, it is meaningful when we focus on local phenomena such
as flow through the parachute gaps and slits. It is also helpful to improve or verify
the HMGP (see [58, 59, 65]). We can afford much higher resolution near the surface
because of the limited domain. Figure 12 shows a comparison of the mesh resolution
between a HMGP mesh and a periodic four-gore model. In previous computations,
slip conditions were applied on the boundaries intersecting the canopy. Here we
employ rotational periodicity on those boundaries. One boundary is a master bound-
ary and the other is a slave boundary. The nodes on the slave boundary do not contain
any unknowns. Because of this, we changed the mesh generation strategy.
First we extract one gore of the canopy. Figure 13 shows an extracted one-gore
surface mesh. Using this surface mesh, we generate a volume mesh with a pie-slice-
shaped domain, which is only a 4:5ı angle. Then, we repeat the one-gore model n
Computer Modeling and Analysis of the Orion Spacecraft Parachutes 65

30
25

Speed (ft/s)
20
15
10
5
0
0 20 40 60 80 100 120
Time (s)

16
14
Speed (ft/s)

12
10
8
6
4
2
0
0 20 40 60 80 100 120
Time (s)
PA OICL80
OICL76 OICL84

Fig. 10 Horizontal-velocity magnitude for the payload (top) and canopy (bottom) for PA and the
three OICL cases. The thin vertical line at 7 s marks the end of the de-symmetrization.

times and merge them as an n-gore model as shown in Fig. 14. A one-gore model
cannot be used for computations because the volume mesh may have an element
consisting of both master and slave nodes. By merging n one-gore models, the mesh
will have at least n elements between the master and slave boundaries.
Figure 15 shows the flow around the rings, which is unsteady. Figure 16 shows
the flow around the sails. We observe that some of the jets exit the slits in a down-
ward direction. This means that the momentum exchange between the surrounding
air and the canopy is not in the same direction as that represented by the smoothed-
surface normal-vector in HMGP. Thus, HMGP should be improved by incorporating
the directional nature of geometric porosity, as shown in the periodic n-gore results.

5.2 Periodic FSI

A cluster of three main parachutes will be used to recover the Orion spacecraft. It is
necessary to compute this cluster in FSI since there is no direct correlation between
single main and cluster performance. Again, starting conditions for FSI computa-
tions are very important, especially for highly dynamic problems such as a parachute
cluster. We generate a single main parachute mesh at an appropriate coning angle
66 K. Takizawa et al.

7.0

Area (×103 ft2)

6.5
6.0
5.5
5.0
4.5
4.0
3.5
0 20 40 60 80 100 120
Time (s)
7.0
Force (×103 lbf)

6.5

6.0

5.5

5.0

4.5
0 20 40 60 80 100 120
Time (s)
PA OICL80
OICL76 OICL84

Fig. 11 Projected area (top) and axial force (bottom) for PA and the three OICL cases. The thin
vertical line at 7 s marks the end of the de-symmetrization.

Fig. 12 Mesh resolution comparison. Meshes for the HMGP (top) and four-gore model (bottom).
Computer Modeling and Analysis of the Orion Spacecraft Parachutes 67

Fig. 13 One-gore surface mesh around the rings (top) and sails (bottom). The surface mesh is
shaded and the mesh across the ring gaps and sail slits is unshaded.

Fig. 14 Four-gore domain (left) and the enlarged view of the periodic boundary around the canopy
(right).

in a one-third domain as shown in Fig. 17. We employ rotational periodicity on

the fluid dynamics and limit the payload motion to the vertical direction. Figure 18
shows the flow patterns and parachute deformation for this periodic cluster FSI. We
merge three one-third periodic domains to form a full-domain solution, which is a
good starting condition for full FSI. Figure 19 shows the result from a preliminary
full FSI computation.
68 K. Takizawa et al.

Fig. 15 Periodic four-gore model. Flow around the rings.

Fig. 16 Periodic four-gore model. Flow around the sails.

6 Stability and Accuracy Analysis of the DSD/SST Formulation

6.1 Space–Time Shape Functions

A space–time shape function can be written as a product of its spatial and temporal
parts:

Na˛ D T ˛ ./ Na . / ; a D 1; 2; : : : ; ne n ; ˛ D 1; 2; : : : ; ne nt ; (1)

where  and  are the spatial and temporal element coordinates, and ne n and ne nt
are the numbers of spatial and temporal nodes. More details can be found in [57].
Computer Modeling and Analysis of the Orion Spacecraft Parachutes 69

Fig. 17 One-third domain (top-right) and the rest of the whole domain (top-left), which is not
used in the computation. The single main parachute at specified coning angle in the one-third
domain (bottom).

Both the spatial and temporal shape functions can be either classical finite element
or NURBS shape functions. NURBS shape functions can be found in [3, 14].

6.2 Temporal Shape Functions

For ne nt D 2, both the finite element and B-spline shape functions are the linear
kind. Figure 20 shows the temporal shape functions for ne nt D 3. The quadratic
shape functions can be written as

1 1
T 1 ./ D   .1  / ; T 2 ./ D 1   2 ; T 3 ./ D  .1 C / ; (2)
2 2
70 K. Takizawa et al.

Fig. 18 Periodic FSI computation of cluster using a one-third domain. Pressure on the cutting
plane and velocity vectors colored by magnitude.

Fig. 19 A cluster of three main parachutes. Preliminary FSI computation.

Computer Modeling and Analysis of the Orion Spacecraft Parachutes 71

1.0 1.0
1
T
2
T1
0.8 T T2
3 0.8
T T3
0.6
0.6
0.4
0.4
0.2

0.0 0.2

-0.2 0.0
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1

Fig. 20 Quadratic temporal shape functions. Finite element functions (left) and B-splines (right).

and the quadratic B-spline shape functions can be written as

.1  /2 1  2 .1 C /2
T 1 ./ D ; T 2 ./ D ; T 3 ./ D : (3)
4 2 4
Since the quadratic shape functions can be represented by the quadratic B-splines,
we focus on the B-splines.

6.3 Advection Equation

Our stability and accuracy analysis is based on the advection equation:

@'
C u  r ' D 0; (4)
@t
where ' is a scalar and u is a constant advection velocity. The DSD/SST formulation
of Eq. (4) is written as follows:
Z ! Z
@' h  
w h h
r
C u r ' dQ C
h
.wh /C h C h 
n .' /n  .' /n d˝
Qn @t ˝n
.nel /n Z
! !
X @wh @' h
C SUPG r wh
C uh r C uh  r ' h dQ D 0; (5)
eD1 Qn
e @t @t

where wh is the test function, Qn is the slice of the space–time domain between
the time levels tn and tnC1 , the notation .  / C
n and .  /n denotes the values as tn is
72 K. Takizawa et al.

approached from below and above respectively, .nel /n is the number of elements at
time level tn , and the stabilization parameter
ˇ
ne n ˇ ˇ !ˇ!1
X
ne nt X ˇ
ˇ @T ˛ ˇˇ ˇ
SUPG D SUGN12 D ˇ N C T ˛
.u  v/  r N a ˇ : (6)
ˇ @t ˇ a ˇ
˛D1 aD1

Here v  dx
dt
is the mesh velocity. The SUPG test function can also be written as
ˇ
hˇ
@wh @w ˇ
C u  r wh D ˇ C .u  v/  r wh : (7)
@t @t ˇ

ˇ
@wh ˇ
@t ˇ
The option where the term is excluded was called “WTSE” in Remark 2
ˇ
hˇ
of [54], and the option where the @w
@t ˇ term is active “WTSA”.

Remark 1 The  definition given by Eq. (6) is coming from [47], which is the
space–time version of the original definition given in [53]. These definitions sense,
in addition to the element geometry, the order of the interpolation functions. Some
 definitions do that and some do not. The definitions given in Sect. 3.3.1 and 3.3.2
of [34], for example, are among those that do not.
Remark 2 Remark 1 is applicable also when the interpolation functions are NURBS
functions. This includes classical p-refinement and also k-refinement, except when
used in conjunction with periodic B-splines.

6.3.1 Fourier Analysis for One-dimensional Uniform Mesh

A stability and accuracy analysis for the case of exact time-integration and no sta-
bilization was given in [1] for B-splines. In the analysis we present here, we use
different temporal and spatial shape function combinations at different CFL values,
where CFL D ct h
. Here, h is the element length, t is the time step size, and
c D u  v is the relative velocity. The element length is equivalent to the distance
between the control points for periodic B-splines. Therefore the element length is
directly related to the number of unknowns. We test three temporal shape functions:
“SV”, reported in [54], linear functions and quadratic B-splines, with three spatial
shape functions: linear functions and C 1 and C 2 B-splines. Figures 21–28 show, at
different CFL values, the phase error, j1  !=!j,
Q and the algorithmic damping ratio,
Q !,
= Q as function of the dimensionless wave number kh. More details on this type
of analysis can be found in [12, 13, 16]. We note that this type of analysis can only
measure phase error within ˙.
Remark 3 Although not shown here, except for the SV option, the space–time
formulation involves non-zero algorithmic damping even for the Galerkin method.
Computer Modeling and Analysis of the Orion Spacecraft Parachutes 73

100 100
10-2 10-2
10-4 10-4
|1 - ω/ω|

x/ω
~ ~
10-6 10-6
∼

CFL=0.1 CFL=0.1
10 -8
CFL=0.5 10-8 CFL=0.5
CFL=1.0 CFL=1.0
10-10 CFL=2.0 10-10 CFL=2.0
CFL=4.0 CFL=4.0
-12
10 10-12
0.01 0.1 1 0.01 0.1 1
k h/ π k h/ π

Fig. 21 Phase error (left) and algorithmic damping ratio (right). Linear functions in space and SV.

100 100
10-2 10-2
10-4 10-4
|1 - ω/ω|

x/ ω
~~

10-6 10-6
∼

CFL=0.1 CFL=0.1
10-8 CFL=0.5 10-8 CFL=0.5
CFL=1.0 CFL=1.0
10-10 CFL=2.0 10-10 CFL=2.0
CFL=4.0 CFL=4.0
10-12 10-12
0.01 0.1 1 0.01 0.1 1
k h/ π k h/ π

Fig. 22 Phase error (left) and algorithmic damping ratio (right). Linear functions in space
and time.

100 100
10-2 10-2
10-4 10-4
|1 - ω/ω|

x/ ω
~ ~
∼

10-6 10-6
CFL=0.1 CFL=0.1
10-8 CFL=0.5
10-8 CFL=0.5
CFL=1.0 CFL=1.0
10-10 CFL=2.0
10-10 CFL=2.0
CFL=4.0 CFL=4.0
10-12 10-12
0.01 0.1 1 0.01 0.1 1
k h/ π k h/ π

Fig. 23 Phase error (left) and algorithmic damping ratio (right). C 1 B-splines in space and SV.
74 K. Takizawa et al.

100 100
10-2 10-2
10-4 10-4
|1 - ω/ω|

x/ω
~ ~
10-6 10-6
∼

CFL=0.1 CFL=0.1
10-8 CFL=0.5 10-8 CFL=0.5
CFL=1.0 CFL=1.0
10-10 CFL=2.0 10-10 CFL=2.0
CFL=4.0 CFL=4.0
10-12 10-12
0.01 0.1 1 0.01 0.1 1
k h/ π k h/ π

Fig. 24 Phase error (left) and algorithmic damping ratio (right). C 1 B-splines in space and linear
functions in time.

100
100
-2
10
10-2
10- 4 10-4
|1 - ω/ω|

~ ~
ξ/ ω

-6
10 10-6
CFL=0.1
10 -8 CFL=0.1
CFL=0.5 10-8 CFL=0.5
CFL=1.0 CFL=1.0
10- 10 CFL=2.0 10-10 CFL=2.0
CFL=4.0 CFL=4.0
- 12
10 10-12
0.01 0.1 1 0.01 0.1 1
k h/ π k h/ π

Fig. 25 Phase error (left) and algorithmic damping ratio (right). C 1 B-splines in space and
quadratic B-splines in time.

100 100

10-2 10-2

10-4 10-4
|1 - ω/ω|

~ ~
ξ/ ω

10-6 10-6
CFL=0.1 CFL=0.1
10-8 CFL=0.5 10-8 CFL=0.5
CFL=1.0 CFL=1.0
10-10 CFL=2.0 10-10 CFL=2.0
CFL=4.0 CFL=4.0
10-12 10-12
0.01 0.1 1 0.01 0.1 1
k h/ π k h/ π

Fig. 26 Phase error (left) and algorithmic damping ratio (right). C 2 B-splines in space and linear
functions in time.
Computer Modeling and Analysis of the Orion Spacecraft Parachutes 75

100
100
10- 2
10- 2
-4
10 10- 4
|1 - ω/ω|

~ ~
ξ/ ω
10- 6 10- 6
CFL=0.1
-8
10 CFL=0.5 10- 8 CFL=0.1
CFL=1.0 CFL=0.5
10 - 10 - 10 CFL=1.0
CFL=2.0 10 CFL=2.0
CFL=4.0 CFL=4.0
10- 12 10- 12
0.01 0.1 1 0.01 0.1 1
k h/ π k h/ π

Fig. 27 Phase error (left) and algorithmic damping ratio (right). C 2 B-splines in space and
quadratic B-splines in time.

100 100
-2
10 10- 2

10-4 10- 4
|1 - ω/ω|

~ ~
ξ/ ω

-6
10 10- 6
CFL=0.1 CFL=0.1
10-8 CFL=0.5 10- 8 CFL=0.5
CFL=1.0 CFL=1.0
10-10 CFL=2.0 10- 10 CFL=2.0
CFL=4.0 CFL=4.0
10-12 10- 12
0.01 0.1 1 0.01 0.1 1
k h/ π k h/ π

Fig. 28 Phase error (left) and algorithmic damping ratio (right). Linear functions in space and
time with the WTSE option.

Remark 4 Linear functions and C 1 B-splines in space are potentially fourth-order

and sixth-order accurate, respectively. But we quickly lose the higher-order accu-
racy at higher CFL numbers, unless we also use higher-order interpolation in time.
For example, SV with C 1 B-spline in space is basically second-order accurate
when CFL > 0:1, which is what we already have when we use SV with just linear
functions in space.

Remark 5 Linear functions in space and time with the WTSE option is second-
order accurate, as can be seen in Fig. 28. This is consistent with what was reported
in [12] using a stabilization equivalent to the WTSE option. The WTSE option is
consistently lower-order accurate compared to the WTSA option, except for with
linear functions in space, in which case the accuracies are comparable but the phase
error directions are different in some regions, as can be seen in Fig. 29.
76 K. Takizawa et al.

10.00 10.00
6.40 6.40
3.20 3.20
1.60 0.9 1.60
0.9
CFL

CFL
0.99 0.99
0.80 0.999 0.80 0.999
0.9999 0.9999
0.40 1.0001 0.40 1.0001
1.001 1.001
0.20 1.01 0.20 1.01
1.1 1.1
0.10 0.10
0.01 0.025 0.05 0.1 0.2 0.4 0.81 0.01 0.025 0.05 0.1 0.2 0.4 0.81
k h/ π k h/ π

Fig. 29 Frequency ratio !=!Q as function of the dimensionless wave number and CFL. Values
larger than one indicate that the numerical solution propagates faster than the exact one. Linear
functions in space and SV with the WTSE (left) and WTSA (right) options.

7 Concluding Remarks

We described how we have addressed the computational challenges involved in

fluid-structure interaction (FSI) modeling of the ringsail parachutes to be used
with the Orion spacecraft. Our computations are based on the Stabilized Space–
Time Fluid–Structure Interaction (SSTFSI) technique and some recent techniques
developed by the T?AFSM in conjunction with the SSTFSI. We described some
of those techniques. We analyzed the performance of the three possible design con-
figurations of the parachute canopy and carried out a parametric study on using
an over-inflation control line (OICL) intended for enhancing the parachute perfor-
mance. We showed that we have a good computational capability for evaluating
parachute design and performance. We compared several aspects of our parachute
FSI computations and NASA drop test data, and they are in reasonably good agree-
ment. We described how we can use rotational periodicity techniques for improving
the geometric-porosity modeling and for computing a good starting condition for
FSI computation of parachute clusters. We presented results from preliminary FSI
computations for a cluster of three parachutes. The core numerical technology
used in the SSTFSI technique is the Deforming-Spatial-Domain/Stabilized Space–
Time (DSD/SST) formulation, which was developed by the T?AFSM earlier for
flow problems with moving boundaries and interfaces, including FSI problems. We
presented a stability and accuracy analysis for various options of the DSD/SST for-
mulation, explaining the interplay between using higher-order interpolation in space
and higher-order integration in time.

Acknowledgements This work was supported in part by NASA Johnson Space Center under
grants NNJ06HG84G and NNX09AM89G. It was also supported in part by the Rice Computational
Research Cluster funded by NSF under Grant CNS-0821727. We thank Jason Christopher for
his contributions at various stages of the parachute computations. We also thank Joel Martin for
carrying out some of the PM5 computations.
Computer Modeling and Analysis of the Orion Spacecraft Parachutes 77

References

1. Y. Bazilevs, V. M. Calo, J. A. Cottrel, T. J. R. Hughes, A. Reali, and G. Scovazzi. Variational

multiscale residual-based turbulence modeling for large eddy simulation of incompressible
flows. Computer Methods in Applied Mechanics and Engineering, 197:173–201, 2007.
2. Y. Bazilevs, V. M. Calo, T. J. R. Hughes, and Y. Zhang. Isogeometric fluid–structure
interaction: theory, algorithms, and computations. Computational Mechanics, 43:3–37, 2008.
3. Y. Bazilevs, V. M. Calo, Y. Zhang, and T. J. R. Hughes. Isogeometric fluid–structure interaction
analysis with applications to arterial blood flow. Computational Mechanics, 38:310–322, 2006.
4. Y. Bazilevs, J. R. Gohean, T. J. R. Hughes, R. D. Moser, and Y. Zhang. Patient-specific iso-
geometric fluid–structure interaction analysis of thoracic aortic blood flow due to implantation
of the Jarvik 2000 left ventricular assist device. Comp. Meth. in Appl. Mech. and Engng.,
198:3534–3550, 2009.
5. Y. Bazilevs, Ming-Chen Hsu, D. Benson, S. Sankaran, and A. Marsden. Computational
fluid–structure interaction: Methods and application to a total cavopulmonary connection.
Computational Mechanics, 45:77–89, 2009.
6. Y. Bazilevs, Ming-Chen Hsu, Y. Zhang, W. Wang, X. Liang, T. Kvamsdal, R. Brekken,
and J. Isaksen. A fully-coupled fluid–structure interaction simulation of cerebral aneurysms.
Computational Mechanics, 2009. Published online. DOI:10.1007/s00466-009-0421-4.
7. Kai-Uwe Bletzinger, R. Wüchner, and A. Kupzok. Algorithmic treatment of shells and free
form-membranes in FSI. In H.-J. Bungartz and M. Schäfer, editors, Fluid–Structure Interac-
tion, volume 53 of Lecture Notes in Computational Science and Engineering, pages 336–355.
Springer, 2006.
8. M. Brenk, H.-J. Bungartz, M. Mehl, and T. Neckel. Fluid–structure interaction on Cartesian
grids: Flow simulation and coupling environment. In H.-J. Bungartz and M. Schäfer, edi-
tors, Fluid–Structure Interaction, volume 53 of Lecture Notes in Computational Science and
Engineering, pages 233–269. Springer, 2006.
9. W. Dettmer and D. Peric. A computational framework for fluid-structure interaction: Finite
element formulation and applications. Comp. Meth. in Appl. Mech. and Engng., 195:5754–
5779, 2006.
10. W. G. Dettmer and D. Peric. On the coupling between fluid flow and mesh motion in the
modelling of fluid–structure interaction. Computational Mechanics, 43:81–90, 2008.
11. Jean-Frederic Gerbeau, M. Vidrascu, and P. Frey. Fluid–structure interaction in blood flow on
geometries based on medical images. Computers and Structures, 83:155–165, 2005.
12. G. Hauke and M. H. Doweidar. Fourier analysis of semi-discrete and space–time stabilized
methods for the advective-diffusive-reactive equation: I. SUPG. Computer Methods in Applied
Mechanics and Engineering, 194:45–81, 2005.
13. T. J. R Hughes. The Finite Element Method. Linear Static and Dynamic Finite Element
Analysis. Prentice-Hall, Englewood Cliffs, New Jersey, 1987.
14. T. J. R. Hughes, J. A. Cottrell, and Y. Bazilevs. Isogeometric analysis: CAD, finite elements,
NURBS, exact geometry, and mesh refinement. Comp. Meth. in Appl. Mech. and Engng.,
194:4135–4195, 2005.
15. T. J. R Hughes, W. K. Liu, and T. K. Zimmermann. Lagrangian–Eulerian finite element formu-
lation for incompressible viscous flows. Comp. Meth. in Appl. Mech. and Engng., 29:329–349,
1981.
16. T. J. R Hughes and T. E. Tezduyar. Finite element methods for first-order hyperbolic systems
with particular emphasis on the compressible Euler equations. Comp. Meth. in Appl. Mech. and
Engng., 45:217–284, 1984.
17. J. G. Isaksen, Y. Bazilevs, T. Kvamsdal, Y. Zhang, J. H. Kaspersen, K. Waterloo, B. Romner,
and T. Ingebrigtsen. Determination of wall tension in cerebral artery aneurysms by numerical
simulation. Stroke, 39:3172–3178, 2008.
18. A. A. Johnson and T. E. Tezduyar. Parallel computation of incompressible flows with complex
geometries. International Journal for Numerical Methods in Fluids, 24:1321–1340, 1997.
78 K. Takizawa et al.

19. A. A. Johnson and T. E. Tezduyar. Advanced mesh generation and update methods for 3D flow
simulations. Computational Mechanics, 23:130–143, 1999.
20. V. Kalro and T. E. Tezduyar. A parallel 3D computational method for fluid–structure inter-
actions in parachute systems. Comp. Meth. in Appl. Mech. and Engng., 190:321–332,
2000.
21. R. A. Khurram and A. Masud. A multiscale/stabilized formulation of the incompress-
ible Navier–Stokes equations for moving boundary flows and fluid–structure interaction.
Computational Mechanics, 38:403–416, 2006.
22. U. Küttler, C. Förster, and W. A. Wall. A solution for the incompressibility dilemma in parti-
tioned fluid–structure interaction with pure Dirichlet fluid domains. Computational Mechanics,
38:417–429, 2006.
23. U. Küttler and W. A. Wall. Fixed-point fluid–structure interaction solvers with dynamic
relaxation. Computational Mechanics, 43:61–72, 2008.
24. R. Löhner, J. R. Cebral, C. Yang, J. D. Baum, E. L. Mestreau, and O. Soto. Extending the
range of applicability of the loose coupling approach for FSI simulations. In H.-J. Bungartz and
M. Schäfer, editors, Fluid–Structure Interaction, volume 53 of Lecture Notes in Computational
Science and Engineering, pages 82–100. Springer, 2006.
25. M. Manguoglu, A. H. Sameh, F. Saied, T. E. Tezduyar, and S. Sathe. Preconditioning tech-
niques for nonsymmetric linear systems in computation of incompressible flows. Journal of
Applied Mechanics, 76:021204, 2009.
26. M. Manguoglu, A. H. Sameh, T. E. Tezduyar, and S. Sathe. A nested iterative scheme for
computation of incompressible flows in long domains. Computational Mechanics, 43:73–80,
2008.
27. M. Manguoglu, K. Takizawa, A. H. Sameh, and T. E. Tezduyar. Solution of linear systems in
arterial fluid mechanics computations with boundary layer mesh refinement. Computational
Mechanics, published online, DOI: 10.1007/s00466-009-0426-z, October 2009.
28. A. Masud, M. Bhanabhagvanwala, and R. A. Khurram. An adaptive mesh rezoning scheme for
moving boundary flows and fluid–structure interaction. Computers & Fluids, 36:77–91, 2007.
29. C. Michler, E. H. van Brummelen, and R. de Borst. An interface Newton–Krylov solver
for fluid–structure interaction. International Journal for Numerical Methods in Fluids,
47:1189–1195, 2005.
30. S. Mittal and T. E. Tezduyar. Massively parallel finite element computation of incompressible
flows involving fluid-body interactions. Comp. Meth. in Appl. Mech. and Engng., 112:253–282,
1994.
31. S. Mittal and T. E. Tezduyar. Parallel finite element simulation of 3D incompressible
flows – Fluid-structure interactions. International Journal for Numerical Methods in Fluids,
21:933–953, 1995.
32. R. Ohayon. Reduced symmetric models for modal analysis of internal structural-acoustic and
hydroelastic-sloshing systems. Comp. Meth. in Appl. Mech. and Engng., 190:3009–3019, 2001.
33. T. Sawada and T. Hisada. Fuid–structure interaction analysis of the two dimensional flag-
in-wind problem by an interface tracking ALE finite element method. Computers & Fluids,
36:136–146, 2007.
34. F. Shakib, T. J. R. Hughes, and Z. Johan. A new finite element formulation for computational
fluid dynamics: X. The compressible euler and navier-stokes equations. Comput. Methods
Appl. Mech. and Engrg., 89:141–219, 1991.
35. K. Stein, R. Benney, V. Kalro, T. E. Tezduyar, J. Leonard, and M. Accorsi. Parachute
fluid–structure interactions: 3-D Computation. Comp. Meth. in Appl. Mech. and Engng.,
190:373–386, 2000.
36. K. Stein, R. Benney, T. Tezduyar, and J. Potvin. Fluid–structure interactions of a cross
parachute: Numerical simulation. Comp. Meth. in Appl. Mech. and Engng., 191:673–687,
2001.
37. K. Stein, T. Tezduyar, and R. Benney. Computational methods for modeling parachute systems.
Computing in Science and Engineering, 5:39–46, 2003.
38. K. Stein, T. Tezduyar, and R. Benney. Mesh moving techniques for fluid–structure interactions
with large displacements. Journal of Applied Mechanics, 70:58–63, 2003.
Computer Modeling and Analysis of the Orion Spacecraft Parachutes 79

39. K. Stein, T. Tezduyar, V. Kumar, S. Sathe, R. Benney, E. Thornburg, C. Kyle, and

T. Nonoshita. Aerodynamic interactions between parachute canopies. Journal of Applied
Mechanics, 70:50–57, 2003.
40. K. Stein, T. E. Tezduyar, and R. Benney. Automatic mesh update with the solid-extension mesh
moving technique. Comp. Meth. in Appl. Mech. and Engng., 193:2019–2032, 2004.
41. K. R. Stein, R. J. Benney, T. E. Tezduyar, J. W. Leonard, and M. L. Accorsi. Fluid–structure
interactions of a round parachute: Modeling and simulation techniques. Journal of Aircraft,
38:800–808, 2001.
42. K. Takizawa, J. Christopher, T. E. Tezduyar, and S. Sathe. Space–time finite element com-
putation of arterial fluid–structure interactions with patient-specific data. Communications in
Numerical Methods in Engineering, published online, DOI: 10.1002/cnm.1241, March 2009.
43. K. Takizawa, C. Moorman, S. Wright, J. Christopher, and T. E. Tezduyar. Wall shear stress
calculations in space–time finite element computation of arterial fluid–structure interactions.
Computational Mechanics, published online, DOI: 10.1007/s00466-009-0425-0, October
2009.
44. T. Tezduyar, S. Aliabadi, M. Behr, A. Johnson, and S. Mittal. Parallel finite-element computa-
tion of 3D flows. Computer, 26(10):27–36, 1993.
45. T. Tezduyar and Y. Osawa. Fluid–structure interactions of a parachute crossing the far wake of
an aircraft. Comp. Meth. in Appl. Mech. and Engng., 191:717–726, 2001.
46. T. E. Tezduyar. Stabilized finite element formulations for incompressible flow computations.
Advances in Applied Mechanics, 28:1–44, 1992.
47. T. E. Tezduyar. Computation of moving boundaries and interfaces and stabilization parameters.
International Journal for Numerical Methods in Fluids, 43:555–575, 2003.
48. T. E. Tezduyar. Finite element methods for fluid dynamics with moving boundaries and inter-
faces. In E. Stein, R. De Borst, and T. J. R. Hughes, editors, Encyclopedia of Computational
Mechanics, Volume 3: Fluids, chapter 17. John Wiley & Sons, 2004.
49. T. E. Tezduyar. Interface-tracking and interface-capturing techniques for finite element com-
putation of moving boundaries and interfaces. Comp. Meth. in Appl. Mech. and Engng.,
195:2983–3000, 2006.
50. T. E. Tezduyar, S. K. Aliabadi, M. Behr, and S. Mittal. Massively parallel finite element sim-
ulation of compressible and incompressible flows. Comp. Meth. in Appl. Mech. and Engng.,
119:157–177, 1994.
51. T. E. Tezduyar, M. Behr, and J. Liou. A new strategy for finite element computations involv-
ing moving boundaries and interfaces – the deforming-spatial-domain/space–time procedure:
I. The concept and the preliminary numerical tests. Comp. Meth. in Appl. Mech. and Engng.,
94(3):339–351, 1992.
52. T. E. Tezduyar, M. Behr, S. Mittal, and J. Liou. A new strategy for finite element computations
involving moving boundaries and interfaces – the deforming-spatial-domain/space–time proce-
dure: II. Computation of free-surface flows, two-liquid flows, and flows with drifting cylinders.
Comp. Meth. in Appl. Mech. and Engng., 94(3):353–371, 1992.
53. T. E. Tezduyar and Y. J. Park. Discontinuity capturing finite element formulations for nonlinear
convection-diffusion-reaction equations. Comp. Meth. in Appl. Mech. and Engng., 59:307–325,
1986.
54. T. E. Tezduyar and S. Sathe. Modeling of fluid–structure interactions with the space–time
finite elements: Solution techniques. International Journal for Numerical Methods in Fluids,
54:855–900, 2007.
55. T. E. Tezduyar, S. Sathe, T. Cragin, B. Nanna, B. S. Conklin, J. Pausewang, and M. Schwaab.
Modeling of fluid–structure interactions with the space–time finite elements: Arterial fluid
mechanics. International Journal for Numerical Methods in Fluids, 54:901–922, 2007.
56. T. E. Tezduyar, S. Sathe, R. Keedy, and K. Stein. Space–time techniques for finite element com-
putation of flows with moving boundaries and interfaces. In S. Gallegos, I. Herrera, S. Botello,
F. Zarate, and G. Ayala, editors, Proceedings of the III International Congress on Numerical
Methods in Engineering and Applied Science. CD-ROM, Monterrey, Mexico, 2004.
80 K. Takizawa et al.

57. T. E. Tezduyar, S. Sathe, R. Keedy, and K. Stein. Space–time finite element techniques
for computation of fluid–structure interactions. Comp. Meth. in Appl. Mech. and Engng.,
195:2002–2027, 2006.
58. T. E. Tezduyar, S. Sathe, J. Pausewang, M. Schwaab, J. Christopher, and J. Crabtree. Fluid–
structure interaction modeling of ringsail parachutes. Computational Mechanics, 43:133–142,
2008.
59. T. E. Tezduyar, S. Sathe, J. Pausewang, M. Schwaab, J. Christopher, and J. Crabtree. Interface
projection techniques for fluid–structure interaction modeling with moving-mesh methods.
Computational Mechanics, 43:39–49, 2008.
60. T. E. Tezduyar, S. Sathe, M. Schwaab, and B. S. Conklin. Arterial fluid mechanics modeling
with the stabilized space–time fluid–structure interaction technique. International Journal for
Numerical Methods in Fluids, 57:601–629, 2008.
61. T. E. Tezduyar, S. Sathe, and K. Stein. Solution techniques for the fully-discretized equations
in computation of fluid–structure interactions with the space–time formulations. Comp. Meth.
in Appl. Mech. and Engng., 195:5743–5753, 2006.
62. T. E. Tezduyar, S. Sathe, K. Stein, and L. Aureli. Modeling of fluid–structure interactions
with the space–time techniques. In H.-J. Bungartz and M. Schäfer, editors, Fluid–Structure
Interaction, volume 53 of Lecture Notes in Computational Science and Engineering, pages
50–81. Springer, 2006.
63. T. E. Tezduyar, M. Schwaab, and S. Sathe. Sequentially-Coupled Arterial Fluid–Structure
Interaction (SCAFSI) technique. Comp. Meth. in Appl. Mech. and Engng., 198:3524–3533,
2009.
64. T. E. Tezduyar, K. Takizawa, and J. Christopher. Multiscale Sequentially-Coupled Arterial
Fluid–Structure Interaction (SCAFSI) technique. In S. Hartmann, A. Meister, M. Schaefer, and
S. Turek, editors, International Workshop on Fluid–Structure Interaction — Theory, Numerics
and Applications. Kassel University Press, 2009.
65. T. E. Tezduyar, K. Takizawa, J. Christopher, C. Moorman, and S. Wright. Interface projection
techniques for complex FSI problems. In T. Kvamsdal, B. Pettersen, P. Bergan, E. Onate, and
J. Garcia, editors, Marine 2009, Barcelona, Spain, 2009. CIMNE.
66. T. E. Tezduyar, K. Takizawa, C. Moorman, S. Wright, and J. Christopher. Space–time finite ele-
ment computation of complex fluid–structure interactions. International Journal for Numerical
Methods in Fluids, published online, DOI: 10.1002/d.2221, 2009.
67. T. E. Tezduyar, K. Takizawa, C. Moorman, S. Wright, and J. Christopher. Multiscale
sequentially-coupled arterial FSI technique. Computational Mechanics, published online, DOI:
10.1007/s00466-009-0423-2, October 2009.
68. R. Torii, M. Oshima, T. Kobayashi, K. Takagi, and T. E. Tezduyar. Influence of wall elasticity
on image-based blood flow simulation. Japan Society of Mechanical Engineers Journal Series
A, 70:1224–1231, 2004. in Japanese.
69. R. Torii, M. Oshima, T. Kobayashi, K. Takagi, and T. E. Tezduyar. Computer modeling
of cardiovascular fluid–structure interactions with the Deforming-Spatial-Domain/Stabilized
Space–Time formulation. Comp. Meth. in Appl. Mech. and Engng., 195:1885–1895, 2006.
70. R. Torii, M. Oshima, T. Kobayashi, K. Takagi, and T. E. Tezduyar. Fluid–structure interac-
tion modeling of aneurysmal conditions with high and normal blood pressures. Computational
Mechanics, 38:482–490, 2006.
71. R. Torii, M. Oshima, T. Kobayashi, K. Takagi, and T. E. Tezduyar. Influence of wall elasticity
in patient-specific hemodynamic simulations. Computers & Fluids, 36:160–168, 2007.
72. R. Torii, M. Oshima, T. Kobayashi, K. Takagi, and T. E. Tezduyar. Numerical investigation of
the effect of hypertensive blood pressure on cerebral aneurysm — Dependence of the effect
on the aneurysm shape. International Journal for Numerical Methods in Fluids, 54:995–1009,
2007.
73. R. Torii, M. Oshima, T. Kobayashi, K. Takagi, and T. E. Tezduyar. Fluid–structure inter-
action modeling of a patient-specific cerebral aneurysm: Influence of structural modeling.
Computational Mechanics, 43:151–159, 2008.
74. R. Torii, M. Oshima, T. Kobayashi, K. Takagi, and T. E. Tezduyar. Fluid–structure interaction
modeling of blood flow and cerebral aneurysm: Significance of artery and aneurysm shapes.
Comp. Meth. in Appl. Mech. and Engng., 198:3613–3621, 2009.
Computer Modeling and Analysis of the Orion Spacecraft Parachutes 81

75. R. Torii, M. Oshima, T. Kobayashi, K. Takagi, and T. E. Tezduyar. Influence of wall thick-
ness on fluid–structure interaction computations of cerebral aneurysms. Communications in
Numerical Methods in Engineering, published online, DOI: 10.1002/cnm.1289, July 2009.
76. R. Torii, M. Oshima, T. Kobayashi, K. Takagi, and T. E. Tezduyar. Role of 0D peripheral vascu-
lature model in fluid–structure interaction modeling of aneurysms. Computational Mechanics,
published online, DOI: 10.1007/s00466-009-0439-7, November 2009.
77. E. H. van Brummelen and R. de Borst. On the nonnormality of subiteration for a fluid-structure
interaction problem. SIAM Journal on Scientific Computing, 27:599–621, 2005.
78. W. A. Wall, S. Genkinger, and E. Ramm. A strong coupling partitioned approach for fluid–
structure interaction with free surfaces. Computers & Fluids, 36:169–183, 2007.
•
Stability Issues in Partitioned FSI Calculations

J. Vierendeels, J. Degroote, S. Annerel, and R. Haelterman

Abstract In this chapter a short review will be given on stability issues for fluid-
structure interaction (FSI) problems we encountered and studied in the last decade.
Based on this, the ideas behind two implicit coupling algorithms, developed in the
department, will be explained. The first algorithm is the Interface Quasi-Newton
coupling method and the second is the Interface Artificial Compressibility cou-
pling method. Most of the applications that are shown are in the biomechanical
field. These are representative for more general strongly coupled problems with
incompressible fluids and flexible structures.

1 Stability issues when computing bileaflet heart valve motion

When the motion of the heart valve is simulated with two separate solvers, one has
to calculate a new position of the heart valve leaflets for each time step and the
flow field has to be updated each time step taking into account the motion of the

J. Vierendeels
Ghent University, Department of Flow, Heat and Combustion Mechanics,
Sint-Pietersnieuwstraat 41, 9000 Gent, Belgium
e-mail: [email protected]
J. Degroote
Ghent University, Department of Flow, Heat and Combustion Mechanics,
Sint-Pietersnieuwstraat 41, 9000 Gent, Belgium
e-mail: [email protected]
S. Annerel
Ghent University, Department of Flow, Heat and Combustion Mechanics,
Sint-Pietersnieuwstraat 41, 9000 Gent, Belgium
e-mail: [email protected]
R. Haelterman
Royal Military Academy, Department of Mathematics, Renaissancelaan 30,
1000 Brussels, Belgium
e-mail: [email protected]

H.-J. Bungartz et al. (eds.), Fluid Structure Interaction II, Lecture Notes 83
in Computational Science and Engineering 73, DOI 10.1007/978-3-642-14206-2 4,
c Springer-Verlag Berlin Heidelberg 2010
84 J. Vierendeels et al.

v(t)

u(t) x(t)
xt
()
Ab A

v(t)

Fig. 1 Motion of a rigid body in a tube, generic one-dimensional test case.

leaflets. Let us assume that a flow solver can calculate the flow field at the new time
level when a prescribed motion of the leaflet is provided and that the moment on
the leaflet, computed by the flow solver, can be used to integrate the motion of the
leaflet in time. The motion of a leaflet with a frictionless hinge can be described by

R
M D I ; (1)

where M is the moment computed from the flow solver, I is the moment of inertia
of the leaflet and R is the angular acceleration of the leaflet. The motion of a heart
valve with the motion of the surrounding blood is a strongly coupled problem and
implicit coupling is necessary to obtain a stable solution [10, 11, 22]. However, with
a partitioned method, where subiterations are necessary between the flow solver and
the structural solver, it was observed that the stability of these subiterations strongly
depends on the density of the fluid, the density of the valve material, the moments of
inertia of the leaflets and the gap size between the leaflets and the housing, whereas
the time step size had no influence. This can be explained with a stability analysis
on a generic one-dimensional test case: the motion of a rigid cylindrical body in a
cylindrical tube, shown in Fig. 1.
Between the cylinder and the tube there is a small gap which connects the fluid at
the front and the back of the cylinder. There exists an analogy between the moving
valve leaflet problem and the 1D test case because during certain phases of the leaflet
motion also a small gap is present that connects the fluid on both sides of the leaflets.
Also, both problems are described with one degree of freedom. In Fig. 1, u.t/ is the
velocity of the oncoming fluid, v.t/ is the velocity in the gaps, xP is the velocity of
the rigid body, A and Ab are the cross sectional area of the tube and the front area
of the rigid body, respectively.
For this generic test case, the conservation of mass is given by

P C .A  Ab /v.t/
Au.t/ D Ab x.t/ (2)

or
1
v.t/ D .u.t/  .1  a/x.t//
P ; (3)
a

with a D Ag =A and Ag D A  Ab , the gap size.

Stability Issues in Partitioned FSI Calculations 85

Conservation of momentum in the gap is expressed by

@v 1 @p
C D0 (4)
@t f @x

for inviscid flow. f denotes the density of the fluid, p is the pressure. In the analysis
the influence of the viscous terms is neglected. The driving force for the rigid body
motion is given by the pressure difference between the left and right wall of the
cylinder. This pressure difference is

@v
pl  pr D Lf ; (5)
@t
with L the length of the gap. The force on the solid body is then given by

F D Ab .pl  pr / D .1  a/A.pl  pr /: (6)

The equation of motion for the solid body becomes

mxR D F (7)

or

@u f 1 1  a f
xR D  xR ; (8)
@t s a a s

with s the density of the solid. Furthermore, K is denoted by

1  a f
KD : (9)
a s

The stability of different coupling strategies for this problem was studied in [22]
and the main conclusions are given below. The study was done using a backward
Euler time integration scheme for the flow solver. For the structural solver, a class
of time integration schemes given by

xP nC1 D xP n C .1  ˇ/ xR n t C ˇ xR nC1 t
x nC1 D x n C xP n t C  xR n t 2 C ˛ xR nC1 t 2 (10)

was considered. For  D 1=2  ˛, this corresponds to the class of Newmark

schemes. For  D 0 and ˛ D ˇ D 1, this corresponds to the backward Euler
scheme. Indeed, with the latter choice, the scheme can be rewritten as

xP nC1 D xP n C xR nC1 t;

x nC1 D x n C xP nC1 t: (11)
86 J. Vierendeels et al.

1.1 Explicit coupling

When explicit coupling is used (˛ D ˇ D 0), it was shown that the stability con-
dition for  can only be fulfilled if K is smaller than 3, which means that for large
K an explicit scheme is unstable and cannot be used. K becomes large for this case
when the gap size is small or when the density of the fluid is large relative to the
density of the solid. For this model, the size of the time step has no influence on the
stability of this partitioned scheme.

1.2 Implicit coupling

When implicit coupling is used, it was shown that even for large values of K stable
schemes could be obtained. However, only when the backward Euler scheme for the
structure was used, corresponding to the time integration scheme used in the flow
solver, the spurious modes are damped critically, which means that spurious modes
will disappear in the next time step. However, when other time integration schemes
are used, the spurious modes are not damped critically and it can take a lot of time
steps before a spurious mode disappears, depending of the amount of numerical
damping in the time integration scheme. This was verified for several time integra-
tion schemes of the class mentioned above, including the second-order Newmark
scheme, the fully implicit Newmark scheme and a damped Newmark scheme. The
spurious oscillations were due to the choice of a different time integration scheme
for the flow and the structural solver and not because of the partitioned approach
since the coupling iterations within the time steps were completely converged. This
behaviour was also numerically verified with the heart valve simulation [22]. Below,
the influence of the choice of the coupling algorithm on the convergence of the
coupling iterations is described.

1.2.1 Gauss-Seidel coupling iterations

When Gauss-Seidel iterations are used to converge the partitioned problem within
one time step (explicit coupling iterations), it was shown that the underrelaxation
factor that has to be used for optimal convergence of these subiterations scales pro-
portionally with 1=K for large values of K. An optimal value can be derived for this
generic test case, but when the gap size is not constant as a function of time, e.g.
with leaflets moving around a hinge, the optimal value is changing with time.

1.2.2 Newton and quasi-Newton coupling iterations

When the underrelaxation factor is computed, which can e.g. be done by comput-
ing a numerical derivative of the change in moment with respect to a change in
Stability Issues in Partitioned FSI Calculations 87

angular acceleration, the coupling iterations become implicit. When this numeri-
cal derivative is computed in each coupling iteration, the Newton-Raphson method
is obtained. Otherwise, also a secant method can be used. If the derivative is esti-
mated from the previous time steps or if it is only computed once, a quasi-Newton
method is obtained. This idea will be extended further when dealing with more than
one degree of freedom at the interface between the flow problem and the structural
problem.

1.2.3 Example of divergence with Gauss-Seidel coupling iterations

It is shown in the following example why Gauss-Seidel coupling iterations can

become unstable. Consider the computation of the motion of the leaflets of a bileaflet
mechanical heart valve starting from a closed position and the fluid at rest. Consider
the computation of a time step when the valve is opening but when the gaps are still
small. A velocity profile changing with time is applied at the inlet. In Fig. 2, the
geometry and the plane in which pressure and velocity fields will be analyzed are
shown. In Fig. 3, pressure and velocity contours are shown in this plane on the old
time level and the new time level, obtained with a convergent coupling scheme.
Figure 4 shows the velocity and the pressure field during Gauss-Seidel coupling
iterations without underrelaxation. When the flow solver is called the first time for

3.00e+03
2.84e+03
2.68e+03
2.52e+03
2.36e+03
2.20e+03
2.04e+03
1.88e+03
1.72e+03
1.56e+03
1.40e+03
1.24e+03
1.08e+03
9.20e+02
7.60e+02
6.00e+02
4.40e+02
2.80e+02
1.20e+02
4.00e+01
2.00e+02

3.00e+03
2.84e+03
2.68e+03
2.52e+03
2.36e+03
2.20e+03
2.04e+03
1.88e+03
1.72e+03
1.56e+03
1.40e+03
Normal flow direction
1.24e+03
1.08e+03
9.20e+02
7.60e+02
6.00e+02
4.40e+02
2.80e+02
1.20e+02
4.00e+01
2.00e+02

Fig. 2 Bileaflet mechanical heart valve, 3D geometry and cut plane in which the results of Fig. 3
and Fig. 4 are shown.
88 J. Vierendeels et al.

1.60e+00 3.00e+03
1.52e+00 2.84e+03
1.44e+00 2.68e+03
1.36e+00 2.52e+03
1.28e+00 2.36e+03
1.20e+00 2.20e+03
1.12e+00 2.04e+03
1.04e+00 1.88e+03
9.60e+01 1.72e+03
8.80e+01 1.56e+03
8.00e+01
7.20e+01
Old time level 1.40e+03
1.24e+03
6.40e+01 1.08e+03
5.60e+01 9.20e+03
4.80e+01 7.60e+03
4.00e+01 6.00e+03
3.20e+01 4.40e+03
2.40e+01 2.80e+03
1.60e+01 1.20e+03
8.00e+02 -4.00e+03
8.00e+00 -2.00e+03

1.60e+00 3.00e+03
1.52e+00 2.84e+03
1.44e+00 2.68e+03
1.36e+00 2.52e+03
1.28e+00 2.36e+03
1.20e+00 2.20e+03
1.12e+00 2.04e+03
1.88e+03
1.04e+00
9.60e+01 1.72e+03 Normal flow direction
8.80e+01 1.56e+03
8.00e+01 1.40e+03
7.20e+01 1.24e+03
6.40e+01 New time level 1.08e+03
5.60e+01 9.20e+03
4.80e+01 7.60e+03
4.00e+01 6.00e+03
3.20e+01 4.40e+03
2.40e+01 2.80e+03
1.60e+01 1.20e+03
8.00e+02 -4.00e+03
8.00e+00 -2.00e+03

Fig. 3 Velocity and pressure field in the cut plane on the old time level and on the new time level
(converged result).

the computation of this time step, the leaflets are put in a position extrapolated from
the positions of the previous time steps. However, the inlet condition has changed
and more fluid is entering the fluid domain than can be stored in the volume swept
by the moving leaflets. Therefore, the incompressible fluid is squeezed through the
gaps between the leaflets, such that high flow velocities are present in the gaps,
giving rise to high accelerations. These accelerations can be related to high pressure
gradients, since in an accelerated or decelerated flow the pressure gradient is mainly
caused by the acceleration or deceleration and not by the viscous terms, i.e.

Dv
f  rp; (12)
Dt

where D=Dt denotes the material or substantial derivative. Thus, the computed
pressure in the region near the inlet will be much higher than the pressure com-
puted behind the valve which can be seen in the pictures. This will result in a large
moment around the hinge, resulting in a large angular acceleration. Thus, in the next
coupling iteration, when this predicted value for the angular acceleration has been
used to compute a new position of the leaflets, the position of the leaflets can be too
far. In this case, the computed flow field shows a backflow through the gaps because
then the volume that is swept by the leaflets is larger than the amount of fluid enter-
ing the domain, thus part of the fluid in the volume that is swept by the leaflets
has to flow back towards the inlet side. The corresponding pressure field shows a
large gradient in the opposite direction because in the gap between the leaflets, the
flow is accelerated now in the opposite direction. In the next coupling iteration, the
leaflets will now be moved in the opposite direction with a certain error with respect
Stability Issues in Partitioned FSI Calculations 89

valve did not move

far enough

valve moved
too far

valve did not move

far enough

valve moved
too far

valve did not move

far enough

Fig. 4 Velocity and pressure field in the cut plane during Gauss-Seidel coupling iterations without
underrelaxation. Divergence is detected.
90 J. Vierendeels et al.

to the correct position. When the amplitude of this error is increasing during the
coupling iterations, the coupling method is unstable and a solution for the next time
step cannot be found, although it exists. Underrelaxation can be used to make these
coupling iterations convergent, but the underrelaxation factor cannot be estimated
in advance, thus one has to use a dynamic underrelaxation factor (e.g. Aitken relax-
ation [15,16]) or an implicit coupling method within the coupling iterations in order
to obtain a good convergence.

1.3 Spurious oscillations due to initial conditions

Spurious oscillations can also be introduced in a calculation due to non-physical ini-

tial conditions. Consider as an example the impact of a cylinder on a water surface
(slamming). The cylinder is hollow and rather thin and can deform during impact,
giving rise to smaller peak pressures than when the cylinder is rigid [8]. When one
wants to estimate this peak pressure, one has to take into account the fluid-structure
interaction. The interaction takes place during impact, thus it would be interesting if
the calculation could be started when the cylinder is just above the surface in order
to reduce the computational time. However, it has then already an initial velocity
and also the air around the cylinder is not at rest, but when starting a calculation,
typically the air around the cylinder is initialized to be at rest and the cylinder is
given an initial velocity. What happens then is that during the first time step the air
in the neighbourhood of the cylinder is accelerated to a velocity corresponding to
the velocity of the falling cylinder. If the time step is taken small, the correspond-
ing acceleration is rather high and also the pressures acting on the cylinder are high,
although the cylinder is only in contact with air. Of course this situation is not physi-
cal, but the result is that the cylinder starts to oscillate because of this pressure pulse.
In the next time steps these high pressures are absent, because the air is moving after
the first time step and the accelerations in the air are rather small until impact occurs.
However, the oscillations of the cylinder remain, because there is little damping in
the structure or in the flow (due to viscosity) and when looking at the force act-
ing on the cylinder as a function of time, these spurious oscillations remain present
(Fig. 5). When keeping the cylinder as a rigid body during the first time steps and by
accelerating it until it has the correct speed, the spurious oscillations are eliminated.
In these simulations no damping in the material is modelled.

2 Stability issues when computing flexible structures

Consider the following 2D problem: a plate is present in a channel as shown in

Fig. 6. The flow and the plate are at rest and should remain at rest when performing
fluid-structure interaction calculations, since the velocity is kept zero at the inlet and
the pressure at the outlet is kept constant. We calculate from this initial condition the
next time step, but we perturb the position of the plate before starting the coupling
Stability Issues in Partitioned FSI Calculations 91

3.0E+3
Spurious oscillations due to initialization
2.5E+3
No spurious oscillations
2.0E+3

1.5E+3
Force [N]

1.0E+3

5.0E+2

0.0E+0

–5.0E+2

–1.0E+3
0 0.001 0.002 0.003 0.004 0.005 0.006
Time [s]

Fig. 5 Force acting on a hollow cylinder during slamming simulations. The spurious oscillations
present during the whole simulations are initiated due to the initial conditions.

u= 0 p = cst

Fig. 6 Plate in a channel. The flow is initially at rest and the velocity is kept zero at the inlet. The
pressure is initialized with the outlet pressure and the outlet pressure is kept constant. Everything
should stay at rest during FSI calculations with these initial and boundary conditions.

iterations. What should happen is that the perturbation should damp out during the
coupling iterations.
Let’s consider a perturbation with a small wave number as seen in the left panel
of Fig. 7. When the flow solver is called with this perturbed position, it computes a
flow through the gaps, of course, the fluid initially present in the volume swept by
the plate must be transfered to the other side of the plate within this time step, again
giving rise to a large acceleration in the gap and a large pressure difference across the
plate. With this pressure load, the structural solver is called, and depending on the
properties of fluid and solid and geometry the plate can move in the other direction
but with an increased amplitude of the perturbation, leading to divergence in subse-
quent coupling iterations. However, when a perturbation with a high wave number
is applied as seen in the middle and the right panel of Fig. 7, it can be observed
that the computed flow field results in more local flow behaviour and local pressure
differences with much smaller amplitudes. Thus, with the same fluid and structural
parameters, the coupling iterations can be convergent. What can be concluded from
this numerical experiment is that errors on the position of the interface with small
92 J. Vierendeels et al.

perturbation with low wave number

perturbation with higher wave number

with adapted scale

Pressure contours

Velocity vectors Velocity vectors Velocity vectors

Pressure contours Pressure contours Pressure contours

Fig. 7 Plate in a channel. The velocity and pressure field is shown for the first coupling iteration.
Before we start the coupling iteration, the position of the plate is perturbed with a small wave
number perturbation (left panel) and a high wave number perturbation (middle and right panels).
The right panel shows the same velocity and pressure fields as the middle panel but on a different
scale.

L
x

a i -1 i-1/2 i i+1/2 i+1

r x
Dx

Fig. 8 Flexible tube used for the 1D Fourier analysis.

wave number are likely to be unstable in Gauss-Seidel coupling iterations, while

errors with high wave number are more likely to be stable.
The same conclusion can be drawn from a stability analysis on the flow through
an elastic tube (Fig. 8). Such an analysis was performed in [2, 6, 18, 19]. The anal-
ysis was based on a Fourier decomposition of the error modes with respect to the
Stability Issues in Partitioned FSI Calculations 93

Fig. 9 Amplification factor

102
as a function of the wave τN = 0.1
number computed with a 1D τN =1
Fourier analysis for the 100 τN =10
flexible tube problem. A high
dimensionless stiffness is
used. Mass in the structure is 10−2
neglected.  denotes the

μ
dimensionless time step and
10−4
N is the number of grid cells.

10−6

10−8
0 π /4 π /2 3π /4 π
ϑ

Fig. 10 Amplification factor

102
as a function of the wave
number computed with a 1D
Fourier analysis for the 100
flexible tube problem. A low
dimensionless stiffness is
used. Mass in the structure is 10−2
neglected.  denotes the
μ

dimensionless time step and

10−4
N is the number of grid cells.

10−6 τN = 0.1
τN=1
τN =10
10−8
0 π /4 π /2 3π /4 π
ϑ

coupled solution. The behaviour of the amplification factor  of these error modes
is studied for a coupling algorithm with Gauss-Seidel iterations without underrelax-
ation. Linearisation was done around a state with constant pressure p0 , radius r0
and velocity v0 . The results are obtained as a function of a dimensionless stiffness
 and a dimensionless time step  D u0 t=L with L the length of the tube, u0 the
velocity inside the tube and t the time step size.  represents the dimensionless
wave number and ranges between 0 and , with  D the highest wave number
that can be represented on the computational grid. From the results in Fig. 9 and
Fig. 10, it can be indeed observed that error modes with small wave number are
less stable than error modes with higher wave number. The influence of the stiffness
and the time step is also shown. When the stiffness is increased, the amplification
factor becomes smaller and when the time step decreases, the amplification factor
increases. The same happens when the length L of the tube increases. A similar
behaviour was also observed in [1, 12]. In Fig. 9 and Fig. 10, the influence of the
mass in the structure is neglected. When taking into account the mass of the struc-
ture, it can be seen that for a decreasing time step, the amplification factor stops
increasing (Fig. 11).
94 J. Vierendeels et al.

Fig. 11 Amplification factor

as a function of the
dimensionless time step  for 100
different wave numbers
computed with a 1D Fourier
analysis for the flexible tube
10−2
problem. A small
ϑ = 2π/100

μ
dimensionless stiffness is
ϑ = 2π /50
used. Mass in the structure is
ϑ = 2π/20
taken into account. N is the 10−4
number of grid cells. ϑ = 2π/10
ϑ = 2π/5
ϑ = 2π /2
10−6 − 3
10 10−2 10−1 100 101 102
τN

This behaviour can be explained as follows. For very small time steps the dis-
placement is mainly controlled by the acceleration term in the structural equations
and this term starts to balance the load term, resulting in smaller displacements for
smaller time steps. Thus, for very small time steps, an error in the load will result in
an error on the displacement, scaled with the square of the time step. On the other
hand, in the flow solver, an error in the displacement will introduce an error on the
load, amplified inversely with the square of the time step size. Thus, for small time
steps there is no amplification related to the time step size on the displacement error
after one coupling iteration. For larger time steps the displacement computed by the
structural solver is rather controlled by a balance between the stiffness and the load
term. However, in the flow solver the displacement stays in balance with the load
with a factor inversely with the square of the time step. This is because the viscos-
ity term is typically very small in comparison with the acceleration term. Thus, for
larger time steps an error in the displacement will be amplified after a coupling iter-
ation with an amplification factor depending on the time step size. This analysis is
verified with numerical simulations. In Fig. 12 the required number of coupling iter-
ations is shown as a function of the time step size for this flexible tube problem. This
demonstrates that Gauss-Seidel iterations become unstable for smaller time steps.
Therefore, a coupling method which treats at least the error modes with small
wave number implicitly has a chance to be a convergent method. Since implicitness
has to be obtained by computing derivatives numerically (when black box solvers
are used), such semi-implicit methods (such as quasi-Newton methods which make
use of increasing-rank Jacobians) can be interesting to use, compared to full implicit
methods such as Newton-Raphson methods (e.g. Newton-GMRES). The Interface
Quasi Newton technique with approximation of the Inverse of the Jacobian form
a Least-Squares model (IQN-ILS) [7] is such a method. For the approximation of
the Jacobian, data of previous time steps can be reused. Figure 12 shows that for
the IQN-ILS method the number of coupling iterations rises if the time step size
is decreased because the number of unstable modes rises which have to be treated
implicitly. For very small time steps the number of coupling iterations does not
Stability Issues in Partitioned FSI Calculations 95

10
GS
9 IQN − ILS
IQN − ILS(4)
8

7
Iterations [−]

0
10−5 10−4 10−3 10−2 10−1 100
Δ t [s]

Fig. 12 Number of coupling iterations as a function of the time step size for the flexible tube
problem. A small dimensionless stiffness is used. Mass in the structure is taken into account. GS:
Gauss-Seidel, IQN-ILS(x): Interface Quasi-Newton with approximation of the Inverse of the Jaco-
bian from a Least-Squares model, x: number of time steps of which data is reused when calculating
the inverse of the Jacobian.

vary anymore, which confirms the analysis above. With reuse of data of previous
time steps (IQN-ILS(x), x denotes the number of previous time steps of which data
is reused), the number of coupling iterations becomes rather independent of the
number of unstable modes in this test case. In this chapter the number of coupling
iterations corresponds to the number of flow and structural solver calls, so it is
directly related to the runtime of the computation, since the amount of work outside
the solvers is negligible.

3 Coupling methods

A short overview of some coupling methods is given below together with a discus-
sion on the performance of these methods, related to the stability issues mentioned
above. The solution of the following fixed point problem is sought:

X D S.F .X //; (13)

with X the position of the nodes at the interface between the fluid and the structure,
F the flow solver which computes the load on the interface and S the structural
solver which computes the new position of the interface. In Fig. 13 the Gauss-Seidel
96 J. Vierendeels et al.

Fig. 13 Fixed point

iterations with the Xk
Gauss-Seidel coupling
method without k ++
underrelaxation. Xk = S(F(X k ))
X k +1 = X k +DX k

Rk = Xk – Xk
DX k = R k

No
Rk < e
(X k+1 = X k )
Yes

Fig. 14 Fixed point

iterations with the Xk
Gauss-Seidel coupling
k ++
method with underrelaxation.
The underrelaxation factor X k = S (F(X k ))
can be dynamically computed X k +1 = X k +DX k
with Aitken’s method.
Rk = Xk – Xk
DX k = wR k
No
Rk < e
(X k+1 = X k +w (X k – X k))
Yes

method is shown.1 Here, X k =X nC1;k where n C 1 denotes the new time level and
k denotes the coupling iteration level when looking for the solution on time level
nC1. The notation nC1 will be omitted. R denotes the residual of the coupling iter-
ation. In the Gauss-Seidel method an underrelaxation parameter ! can be used.2 The
Gauss-Seidel coupling method with Aitken underrelaxation [15, 16] is an example
where dynamic underrelaxation is used (Fig. 14).
Two classes of methods developed in our research group are the Interface Artifi-
cial Compressibility (IAC) methods [9, 18, 19, 21] and the Interface Quasi-Newton
(IQN) methods [3, 7] which were originally developed as a coupling method based
on Reduced Order Models for the behaviour at the interface of the flow and the
structural solver [23, 24]. In the IAC method (Fig. 15) the flow solver is altered dur-
ing the coupling iterations in such a way that the altered flow solver converges to
the original flow solver together with the convergence of the coupling iterations.
Of course it is important to alter the flow solver in such a way that convergence of
the coupled problem is enhanced. This can be done by introducing information from

1
This corresponds to the Richardson’s method for a fixed point problem, but in partitioned FSI
calculations the iterations are also called Gauss-Seidel iterations.
2
This corresponds to the modified Richardson’s method for fixed point problems.
Stability Issues in Partitioned FSI Calculations 97

Fig. 15 Fixed point

iterations with the Interface Xk
Artificial Compressibility
(IAC) method. The flow k ++
solver is altered in such a way X k = S (F k (X k ))
that convergence of the X k+1 = X k +DX k
coupling iterations is
enhanced. The altered flow Rk = Xk – Xk
solver also converges towards DX k = R k
the real flow solver during the
No
convergence of the coupling Rk < e
iterations [9]. (X k+1 = X k )
Yes

the structural solver in the flow solver. The idea on which the IAC method is based
can be best explained when looking at the problem of the flexible tube. Consider an
equilibrium flow in a straight flexible tube. From a certain moment on the inflow
condition is changed, such that a wave starts travelling from the inlet to the outlet.
Consider the first time step where e.g. a rise of the velocity at the inlet occurs. In
this time step the wave will travel over a certain distance and locally near the inlet
the velocity and the pressure will rise. If the time step is small enough, the wave
will not be noticed yet at the outlet. However, when computing this solution with a
partitioned approach with Gauss-Seidel iterations, the flow solver does not know yet
that the tube wall has to move and due to the incompressibility constraint the extra
amount of fluid entering the tube will immediately flow out. Therefore, the fluid in
the whole tube has to accelerate which results in a pressure gradient in the whole
tube. This results in a solution of the first coupling iteration where a big pressure
difference will be present between the inlet and the outlet, the pressure difference
needed to accelerate all the fluid between inlet and outlet. When the structural solver
is called, the large pressure can cause the tube wall to move much too far, resulting
in unstable coupling iterations. When the flow solver is modified in such a way that
the incoming fluid does not have to leave the tube immediately, but can be stored
locally, the coupling method can be stabilized. When a source term is added to the
continuity equation of the cells adjacent to the fluid-structure interface

V nC1;k  V n @V p nC1;kC1  p nC1;k

 C ˙f .vf  vg rid;f /nf Sf D  ;
t @p t
(14)
the computed pressure field p nC1;kC1 will be altered in such a way that it takes into
account the effect of storing fluid near the boundary. In this equation, V denotes
the control volume, v and vg rid;f denote the flow velocity and the grid velocity
at the face f with area Sf of the control volume. n and n C 1 denote the old and
the new time level, k denotes the coupling iterations level (not the iteration level in
the flow solver). In the flow solver, V nC1;k is known (computed from the structure
solver) and a solution is sought for p nC1;kC1 and vnC1;kC1 . The structural solver
98 J. Vierendeels et al.

then computes the new position of the interface, resulting in an updated position
of the control volume, now at coupling iteration level k C 1: V nC1;kC1 . When due
to the source term, this altered pressure field is much closer to the coupled solution
of the time step, and when this pressure field is subsequently applied to the struc-
tural solver, the computed tube wall displacement will also be close to the coupled
solution. The question is: how to alter the flow solver in order to obtain such a pres-
sure field? The answer can be found by analyzing the source term in eq. (14), which
can be approximated as

@V p nC1;kC1  p nC1;k V nC1;kC1  V nC1;k

   ; (15)
@p t t

such that the continuity equation that is solved for is approximated by

V nC1;kC1  V n
 C ˙f .vf  vg rid;f /nf Sf  0: (16)
t
This equation takes into account that the interface will move and, consequently,
that the control volume will change due to a change in the pressure field. When
the coupling iterations converge, the source term becomes zero, the approximation
becomes an equality and the incompressible solution is found.
Thus, the source term is a vanishing term which approximates the amount of
mass that will be stored near the boundary due to the motion of the boundary in one
coupling iteration, although the boundary motion has not been computed yet when
the new pressure and velocity fields are computed. Therefore, one has to estimate
what the displacement of the boundary will be. When the applied load changes and
when the pressure is the main reason for the boundary displacement (which is typ-
ically the case), one has to know how a local pressure change will cause a local
volume change. This information is derived from the structural solver by perturbing
the current pressure distribution with a small value and by calculating the corre-
sponding local volume change. From this information @V =@p is computed for each
boundary cell (Fig. 16).

Fig. 16 Computation of
@V =@p by perturbing the
current pressure distribution,
resulting in a motion of the
boundary from Xa to Xb ,
computed by the structure
solver. The computed volume
change (vol) in each cell,
together with @V =@p that are
used in the continuity
equation of the flow solver.
Stability Issues in Partitioned FSI Calculations 99

Fig. 17 Fixed point

iterations with the Interface Xk
Quasi-Newton method with
k ++
approximation of the Inverse
of the Jacobian from a X k = S (F (X k ))
Least-Squares model X k+1 = X k +DX k
(IQN-ILS). [3, 7]
Rk = Xk – Xk
( )
–1
dR
DX k = DR k
dX

Rk < e No
Yes

For a one-dimensional problem it was analyzed in [9, 18, 19] that this method
stabilized the Gauss-Seidel iterations and in such a way that convergence is rapidly
obtained. Also in 3D calculations [9], this was observed.
The Newton-Raphson technique can also be applied to this fixed point problem as
shown in Fig. 17. The problem here is the calculation of the inverse of the Jacobian
of the residual with respect to the position of the interface nodes, especially when
black box solvers are used. In the Newton GMRES method, X k is computed with
GMRES from a linearized system, but in each iteration within the GMRES solver,
a call to the linearized flow and structural solver is needed. In the interface GMRES
method [20], X k is computed with GMRES applied on the original black box
flow and structural solver. Thus, GMRES is applied on a non-linear problem and
also within each iteration of the GMRES solver a call to the flow and structural
solver is needed.
When the update X k is computed with an approximation for the inverse of
the Jacobian, the method is called a quasi-Newton method. We developed a method
[3, 7] in which the inverse of the Jacobian is directly computed with a least squares
approach in such a way that, if the method is applied to a linear problem, the full
Jacobian is obtained after N C 1 iterations with N the number of degrees of free-
dom at the interface (size of X ) [13, 14]. It was also shown that the method is a
generalized secant method. The method can be summarized as follows [7]: calculate
 1
dR
X D
k
Rk ; R k D R kC1  Rk D 0  Rk D Rk (17)
dX

by approximating the inverse of the Jacobian as

 1
dR dX d.XQ  R/ d XQ
D D D  I: (18)
dX dR dR dR

With
h i h i
V k D R0 : : : Rk1 ; W k D XQ 0 : : : XQ k1 ; ˛ k D Œ˛0 : : : ˛k1
; (19)
100 J. Vierendeels et al.

solve Rk D V k ˛ k for ˛ k with the least squares method:

 T 1 T
˛k D V k V k V k Rk : (20)

Then XQ k  W k ˛ k , which results in an expression for X k as a function of Rk :

 T 1 T
X k D XQ k  Rk D W k V k V k V k Rk  Rk ; (21)

such that an approximation of the inverse of the Jacobian .dR=dX /1 is given by
 1  T 1
dR T
 Wk Vk Vk V k  I: (22)
dX

However, this matrix with size corresponding to the number of degrees of freedom
on the interface is never computed nor stored, since only a multiplication of this
matrix with a vector is needed. In practice a QR decomposition of V is used when
calculating X [3, 7]. Remark that Eq. (21) can also be written as
!
d XQ
X D R C
k k
Rk ; (23)
dR

which shows that the Gauss-Seidel iteration is used when the second term is zero.
One can also decompose the residual in small wave number errors and high wave
number errors. From previous analysis it is expected that the errors with small
wave number will be amplified in the Gauss-Seidel iteration, but once they are
incorporated in the Jacobian, they will be treated implicitly in subsequent coupling
iterations. The error modes that are not incorporated yet, will not give a contribution
to the second term and thus these modes are treated with the Gauss-Seidel approach.
Since typically only a few error modes are unstable, one does not need a full Jaco-
bian in order to converge the coupling iterations. When the Jacobian is being built,
it is also easy to include information from the previous time steps under the assump-
tion that the Jacobian is not changing much with time. When small time steps are
taken, the coupling problem can become tougher as shown above, but with small
time steps, it takes more time steps for the Jacobian to change, thus with small time
steps also more time steps can be used to compute the Jacobian. The effect of the
reuse of data of previous time steps was already shown in Fig. 12 for the flexible
tube problem. Other examples of the performance of the method can be found in
[3–5, 8].
Stability Issues in Partitioned FSI Calculations 101

4 Conclusion

Stability analysis performed on coupling methods for partitioned coupled problems

revealed that one must be cautious when using different time integration schemes
for the flow and the structural solver. It revealed also that typically only a few error
modes of the interface problem are unstable when using Gauss-Seidel iterations.
These can be stabilized by using (dynamic) underrelaxation. Based on the insights
of stability analysis, one can understand why the Interface Artificial Compressibility
method can perform well and why the Interface Quasi-Newton method can converge
well within a limited amount of coupling iterations, especially with reuse of data of
previous time steps.

Acknowledgements J. Degroote gratefully acknowledges a PhD fellowship of the Research

Foundation-Flanders (FWO).

References

1. Causin, P., Gerbeau, J.F., Nobile, F.: Added-mass effect in the design of partitioned algorithms
for fluid-structure problems. Comput. Meth. Appl. Mech. Eng. 194 4506–4527 (2005)
2. Degroote, J., Bruggeman, P., Haelterman, R., Vierendeels, J.: Stability of a coupling technique
for partitioned solvers in FSI applications. Comput. Struct. 86 2224–2234 (2008)
3. Degroote, J., Bathe, K.J., Vierendeels, J.: Performance of a new partitioned procedure versus a
monolythic procedure in fluid-structure interaction. Comput. Struct. 87 793–801 (2009)
4. Degroote, J., Bruggeman, P., Vierendeels, J.. A coupling algorithm for partitioned solvers
applied to bubble and droplet dynamics. Comput. Fluids 38 613–624 (2009)
5. Degroote, J., Bruggeman, P., Haelterman, R., Vierendeels, J.: Bubble simulations with an inter-
face tracking technique based on a partitioned fluid-structure interaction algorithm. J. Comput.
Appl. Math. (2010) doi: 10.1016/j.cam.2009.08.096
6. Degroote, J., Annerel, S., Vierendeels, J.: Stability analysis of Gauss-Seidel iterations in a
partitioned simulation of fluid-structure interaction. Comput. Struct. (88) 263–271 (2010)
7. Degroote, J., Haelterman, R., Bruggeman, P., Vierendeels, J.: Performance of partitioned
procedures in fluid-structure interaction. Comput. Struct. (88) 446–457 (2010)
8. Degroote, J., Souto-Iglesias, A., Van Paepegem, W., Annerel, S., Bruggeman, P., Vierendeels
J.: Partitioned simulation of the interaction between an elastic structure and free surface flow.
Comput. Meth. Appl. Mech. Eng. (2010) In press.
9. Degroote, J., Swillens, A., Bruggeman, P., Haelterman, R., Segers, P., Vierendeels, J.: Simula-
tion of fluid-structure interaction with the interface artificial compressibility method. Commun.
Numer. Methods Eng. 26 276–289 (2010)
10. Dumont, K., Vierendeels, J.A.M., Segers, P., Van Nooten, G.J., Verdonck, P.R.: Predicting ATS
Open PivotTM heart valve performance with computational fluid dynamics. J. Heart Valve Dis.
14, 393–399 (2005)
11. Dumont, K., Vierendeels, J., Kaminsky, R., Van Nooten, G., Verdonck, P., Bluestein, D.: Com-
parison of the hemodynamic and thrombogenic performance of two bileaflet mechanical heart
valves using a CFD/FSI model. J. Biomech. Eng.-Trans. ASME 129 558–565 (2007)
12. Förster, C., Wall, W.A., Ramm, E.: Artificial added mass instabilities in sequential staggered
coupling of nonlinear structures and incompressible viscous flows. Comput. Meth. Appl. Mech.
Eng. 196 1278–1293 (2007)
102 J. Vierendeels et al.

13. Haelterman, R., Degroote, J., Van Heule, D., Vierendeels, J.: The Quasi-Newton Least Squares
Method: A New and Fast Secant Method Analyzed for Linear Systems. SIAM J. Numer. Anal.
47 2347–2368 (2009)
14. Haelterman, R., Degroote, J., Van Heule, D., Vierendeels, J.: On the similarities between the
quasi-Newton inverse least squares method and GMRes. SIAM J. Numer. Anal. 47 4660–4679
(2010)
15. Küttler U., Wall W.A.: Fixed-point fluid-structure interaction solvers with dynamic relaxation.
Comput. Mech. 43 61–72 (2008)
16. Mok, D.P., Wall, W.A., Ramm, E.: Accelerated iterative substructuring schemes for instation-
ary fluid-structure interaction. In: Bathe, K.-J. (ed.) Computational Fluid and Solid Mechanics,
pp. 1325–1328. Elsevier (2001)
17. Riemslagh, J., Vierendeels, J., Dick, E.: Coupling of a Navier-Stokes solver and an elastic
boundary solver for unsteady problems. In: Proc. of the Fourth ECCOMAS Computational
Fluid Dynamics Conference, Athens, September 1998, Computational Fluid Dynamics ’98,
pp. 1040-1045, Chichester, John Wiley & Sons. (2002) ISBN: 0-471-98579-1
18. Riemslagh, K., Vierendeels, J., Dick, E.: A simple but efficient coupling procedure for flex-
ible wall fluid-structure interaction. In: Proc. of the Fluids 2000 AIAA Computational Fluid
Dynamics Conference, Denver, CO, 2000. Paper AIAA 2000-2336, 8 pp. (2000) CD-ROM
ISBN: 1-56347-433-6
19. Riemslagh, K., Vierendeels, J., Dick, E.: An efficient coupling procedure for flexible wall
fluid-structure interaction. In: Proc. of the Fifth ECCOMAS Computational Fluid Dynamics
Conference, Barcelona, September 2000, Computational Fluid Dynamics ’00, 2000. 13 pp.
(2000) CD-ROM ISBN: 84-89925-70-4
20. van Brummelen, E.H., Michler, C., de Borst, R.: Interface-GMRES(R) acceleration of subiter-
ation for fluid-structure-interaction problems. Report DACS-05-001, January 2005. Available
from: http://www.em.lr.tudelft.nl/downloads/DACS-05-001.pdf.
21. Vierendeels, J.A., Riemslagh, K., Dick, E., Verdonck, P.R.: Computer simulation of intra-
ventricular flow and pressure gradients during diastole. J. Biomech. Eng.-Trans. ASME 122
667–674 (2000)
22. Vierendeels, J., Dumont, K., Dick, E., Verdonck, P.: Analysis and stabilization of fluid-structure
interaction algorithm for rigid-body motion. AIAA J. 43 2549–2557 (2005)
23. Vierendeels, J., Lanoye, L., Degroote, J., Verdonck, P.: Implicit coupling of partitioned fluid-
structure interaction problems with reduced order models. Comput. Struct. 85 970–976 (2007)
24. Vierendeels, J., Dumont, K., Verdonck, P.: A partitioned strongly coupled fluid-structure
interaction method to model heart valve dynamics. J. Comput. Appl. Math. 215 602–609
(2008)
Hydroelastic Analysis and Response
of Pontoon-Type Very Large Floating Structures

C.M. Wang and Z.Y. Tay

Abstract Pontoon-type very large floating structures (VLFS) are giant plates rest-
ing on the sea surface. As these structures have a large surface area and a relatively
small depth, they behave elastically under wave action. This type of fluid-structure
interaction has being termed hydroelasticity. Hydroelastic analysis is thus neces-
sary to be carried out for VLFS designs in order to assess the dynamic motion and
stresses due to wave action. This paper presents the mathematical formulation for
the hydroelastic analysis of VLFS. Hydroelastic responses and mitigation methods
in reducing the structural response are discussed using some example problems.

1 Introduction

The rising interest in very large floating structures (VLFS) over the past two decades
was triggered by the corresponding increase in urbanization and human population
in land-scarce countries such as Japan, Singapore and The Netherlands. The con-
ventional solution of creating additional land through land reclamation is not cost
effective when the water depth is large. Moreover, land reclamation works generally
have a negative environmental impact on coastlines and marine eco-system. VLFS
on the other hand has advantages over the traditional land reclamation solution with
respect to the following aspects: they are cost effective when the water depth is large
and the seabed is soft; environmentally friendly as they do not damage the marine
eco-system, or silt-up deep harbours or disrupt the ocean currents; they are easy and
fast to construct and therefore the investment may be monetized more rapidly; they

C.M. Wang
Professor, Centre for Offshore Research and Engineering, Department of Civil Engineering,
National University of Singapore, Kent Ridge, Singapore, 119260
e-mail: [email protected]
Z.Y. Tay
Research Fellow, Centre for Offshore Research and Engineering, Department of Civil Engineering,
National University of Singapore, Kent Ridge, Singapore, 119260
e-mail: [email protected]

H.-J. Bungartz et al. (eds.), Fluid Structure Interaction II, Lecture Notes 103
in Computational Science and Engineering 73, DOI 10.1007/978-3-642-14206-2 5,
c Springer-Verlag Berlin Heidelberg 2010
104 C.M. Wang and Z.Y. Tay

can be easily removed or expanded; and the structure on VLFS is protected from
seismic shocks since VLFSs are inherently base isolated.
VLFS may be categorised into the semisubmersible-type and the pontoon-type.
The semisubmersible-type VLFS such as the mobile offshore base (MOB) had been
proposed by the US Navy as early as in the pre-cold war era to support military oper-
ations where conventional land bases were not available. Such semisubmersible-
type VLFS with a raised platform above sea level using column tubes are suitable
for deployment in high seas with large waves. In contrast, the pontoon-type VLFS
are intended for deployment in calm water such as in a cove, lagoon or harbour.
Examples of these pontoon-type VLFS are the Mega Float (a runway test model in
the Tokyo Bay, see Fig. 1a), emergency rescue bases (moored in Tokyo Bay, Ise Bay
and Osaka Bay), floating storage facilities (in Shirashima and Kamigoto Islands, see
Fig. 1b), floating bridges (in Seattle, USA, see Fig. 1c) and floating ferry piers (in
Ujina Port, Hiroshima, see Fig. 1d).
Singapore has also recently constructed the world’s largest floating performance
stage at the Marina Bay (Fig. 2a) and feasibility studies are underway to build
a mega floating fuel storage facility (FFSF) (Fig. 2b) to cater for the increasing
demand for oil storage capacity. Such FFSF may double up as bunker cum mooring

(a) Mega-Float, Tokyo Bay. (b) Floating oil storage base at Kamigoto Island
(Source:http://www.mhi.co.jp/en/products/
detail/oil storage terminal kamigoto.html).

(c) Lacey V. Murrow Bridge and the Third (d) Floating ferry pier at Ujina Port, Hiroshima.
Washington Bridge at Seattle (Source:http://
en.wikipedia.org/Lacey V. Murrow Memorial
Bridge).

Fig. 1 Applications of pontoon-type VLFSs.

Hydroelastic Analysis and Response of Pontoon-Type Very Large Floating Structures 105

(a) Floating performance stage @ Marina Bay, (b) Proposed floating fuel storage facility
Singapore. (Photo courtesy of JCPL).

(c) Proposed floating cruise terminal. (d) Proposed mega floating crab restau-
rant.

Fig. 2 VLFSs in Singapore.

system for ships, thereby relieving traffic congestions in the Singapore harbour and
decreasing the turnaround time for ships. The first author has proposed the use of
VLFS as a floating-type cruise terminal (Fig. 2c) and a mega floating crab restau-
rant (Fig. 2d) as opposed to the conventional onshore design in order to create iconic
structures for Singapore to attract tourists.
VLFS technology has also made possible future large human habitation on the
ocean surface. The Lilypad Floating Ecopolis (Fig. 3), proposed by the Belgium
architect Vincent Callebaut, is an example of a visionary proposition to house the
city population on a huge floating lily-shaped island. More concepts of floating cities
are given in a recent paper by Pernice [1].
As VLFSs have a large surface area and a relatively small depth, they behave
elastically under wave action. The fluid-structure interaction has been termed as
hydroelasticity. Hydroelastic analysis is thus necessary to be carried out for VLFS
in order to assess the dynamic motion and stresses due to wave action design [2].
This paper focuses on the hydroelastic analysis of pontoon-type VLFSs based on
the frequency domain approach. Frequency domain analysis has been applied exten-
sively to problems of floating structure dynamics and is particularly useful for the
determination of long term responses [3].
106 C.M. Wang and Z.Y. Tay

Fig. 3 Lilypad Floating Ecopolis. Source: www.vincent.callebaut.org.

2 Literature Review on Hydroelastic Analysis of Floating

Structures

The conventional method of solving the motion of floating bodies is to convert

the Laplace equation for the velocity potential into a boundary value problem. The
boundary conditions are the Neumann condition at the seabed and the wetted surface
of the floating body, the linearised free surface condition and the radiation condi-
tions at infinity. The earliest solution to this boundary value problem was given
by John [4, 5] in which he used the Green’s function within a boundary integral
formulation to solve for the wave scattering from floating bodies. A detailed descrip-
tion of the linear wave theory was published by Wehausen and Laiton [6] in their
remarkable review article ’Surface Waves’. This review article contains benchmark
solutions for wave-structure interactions problems [3,7–11]. However, earlier works
on the wave-structure interactions problems only consider the floating structure as
a rigid body. With the increasing interest in VLFS as one of the future solutions
for creating land for land-scarce countries, hydroelastic analysis on floating struc-
tures emerged as a new research area in the early 1990s. To name a few, among
the pioneers working on the hydroelastic theory of VLFS are Ertekin et al. [12],
Yago and Endo [13], Utsunomiya et al. [14], Kashiwagi [15] and Ohmatsu [16].
The development of hydroelastic theory should also be attributed to Meylan and
Squires [17] and Meylan [18, 19] who studied ice-floe problems which are similar
to VLFS problems [20].
The Mega-Float project in Japan from 1995 to 2001 triggered extensive studies
on the hydroelastic response of the VLFS. As the pontoon-type VLFS has a small
draft compare to its length, the common approach is to model the entire floating
structure by a single plate based on the Kirchhoff plate theory (see for example,
refs. [14, 15, 21]) while the water wave is modelled by using the linear wave theory.
Hydroelastic Analysis and Response of Pontoon-Type Very Large Floating Structures 107

Usually the frequency domain approach is used instead of the time domain approach
when determining the hydroelastic response amplitude operator of the plate model
because of its simplicity and ability to capture the pertinent response parameters in
a steady state condition.
More sophisticated hydroelastic analyses have also been performed for a com-
plete 3D floating structure, in order to obtain the deflections and stresses for the
secondary structural components in the VLFS [22–25] as well as the local strength
of the structural members [26, 27]. The hydroelastic analyses of the VLFS in the
the presence of breakwaters [14, 16] and even attached anti-motion devices in the
form of submerged plates [28–31] have been carried out. Researchers found that
breakwaters and anti-motion devices have a profound effect in reducing the struc-
tural motion due to attenuated wave forces. In order to solve for aircraft landing
or wave impact problems on VLFS, Kim and Webster [32], Watanabe et al. [33]
and Kashiwagi [34, 35] investigated the transient response using the time-domain
approach.
Mega floating fuel storage modules, as shown in Figs. 1b and 2b, have larger
draft to length ratios as opposed to the mat-like VLFS. This necessitates the mod-
elling of the floating modules as a thick plate according to the Mindlin plate theory
[36–39]. The use of the Mindlin plate theory not only leads to more accurate
prediction of the deflections, but it also provides a better prediction for the stress-
resultants. Studies on hydroelastic interactions of two large box-like floating storage
modules such as the one shown in Fig. 2b have also been carried out recently by Tay
et al. [39]. The adjacent module affects each other due to diffracted waves, radi-
ated waves and waves been squeezed into a channel formed by the floating modules
being placed side-by-side. Khabakhpasheva and Korobkin [40] propounded an aux-
iliary floating structure that is connected to the main floating structure in order to
reduce the deflection of the main floating structure. Wang et al. [41] then proposed
a novel way in reducing the hydroelastic response of VLFS through the design of
semi-rigid mechanical joints that connect the auxiliary structure to the main float-
ing structure. The design involves suitable rotational stiffness and location of such
mechanical joints.

3 Problem Definition

Fig. 4 shows the components of a typical pontoon-type very large floating struc-
ture, which consist of the floating structures, access-bridge, breakwater and mooring
facility. Breakwaters are constructed around the floating structure to attenuate the
wave forces that impact the floating structure. The VLFS is usually moored by the
dolphin-frame guide mooring system that restrains the horizontal movement of the
floating structure but allows the structure to freely move in the vertical direction
according to the tidal variation and varying payload.
The problem at hand is to determine the hydroelastic response of a VLFS under
wave action. Also we aim to show the reduction of the hydroelastic responses of
floating structures with the attachment of submerged plates and the appropriate
design of semi-rigid connections.
108 C.M. Wang and Z.Y. Tay

Fig. 4 Components of a
pontoon-type VLFS.

Fig. 5 Figure depicting the pontoon-type very large floating structure in (a) plan view and (b) side
view.

4 Mathematical Formulations

4.1 Plate-Water Model

Consider a pontoon-type very large floating structure with a rectangular plan area of
L1 L2 as shown in Fig. 5a. The box-like floating structure has a constant height h,
top, bottom and side wall thickness tslab , and draft d as shown in Fig. 5b. The
seabed is assumed to be level and the water depth is H. The structure is subjected to
an incoming wave of wave period T and a wave height 2A that impacts the structure
at a wave angle  with respect to the x-axis.
For hydroelastic analysis, the box-like floating structure is commonly modelled
as an equivalent solid plate (see Fig. 6a), by keeping the lengths and the height to
be the same as the actual structure (Fig. 5) but its Young’s modulus E and Poisson’s
ratio  of the equivalent solid plate are tweaked to match the natural frequencies
and vibration modes of the actual structure. The equivalent plate is assumed to
be perfectly flat with free edges. The plate material is commonly assumed to be
isotropic or orthotropic and obeys Hooke’s law. For the hydroelastic analysis, most
researchers modelled the plate using the classical thin plate theory, but more recently
the use of Mindlin plate theory is getting popular due to its ability to provide a better
Hydroelastic Analysis and Response of Pontoon-Type Very Large Floating Structures 109

Fig. 6 Schematic diagram depicting (a) cross section of equivalent solid plate (b) deflection of
plate with respect to water surface.

prediction to the stress resultants and its allowance for the effects of transverse shear
deformation and rotary inertia. Unlike the classical thin plate theory where the equa-
tions of motion are described solely by the deflection w variable, the motion of the
Mindlin plate is represented by the vertical deflection W(x, y; t), the rotation about
the x-axis y .x; yI t/ and the rotation about the y-axis x .x; yI t/. In the anal-
ysis, the plate is assumed to be restrained in the x-y plane by the station keeping
system and the plate can only deformed in the vertical direction (i.e. z-direction).
The equations for plate motion will be given in Sect. (4.3.1).
The water is assumed to be a perfect fluid with no viscosity and incompressible
and the fluid motion to be irrotational. Based on these assumptions, the fluid motion
may be represented by a velocity potential ˚. It is also assumed that the fluid motion
is small so that the linear potential theory can be applied for formulating the fluid-
motion problem. We also consider the steady-state harmonic motions of the fluid
and the structure, with the circular frequency !. The equations for fluid motion will
be given in Sect. (4.3.2).
The water-plate interaction is realized by assuming the deformed surface of the
plate to be coincident with the water surface in contact with the plate, i.e. there are
no gaps between the plate surface and the water surface.

4.2 Symbols and Non-Dimensional Variables

Fig. 6a shows the cross section of the equivalent solid plate. The symbol  denotes
the floating plate domain while the water domain is denoted by ˝. The symbols SF ,
SB , SHB and SHS represent the free surface, the seabed, the wetted bottom surface
of the floating body and the wetted side surface of the floating body, respectively.
The free and undisturbed water surface is at z = 0. The deflection w is the verti-
cal elevation of the plate measured from the free and undisturbed water surface as
shown in Fig. 6b.
In the subsequent mathematical formulation, we shall use the following non-
dimensionalised spatial variables: xN D x=L, yN Dpy=L, zN D z=L, wN pD W=L,
hN D h=L, N N
p d D d=L and the time variables: t D t g=L, 'N D ˚=.L gL/ and
!N D ! L=g, where L D L2 of the floating structure as shown in Fig. 5a, g is the
gravitational acceleration and ! the wave frequency.
110 C.M. Wang and Z.Y. Tay

We consider the steady state motions of a harmonically excited system at a circu-

N Thus, we can separate the time variable from the spatial variables
lar frequency !.
as follows:

W .x; y; t/
D ReŒw.
N x; N i !N tN;
N y/e (1)
L
x .x; y; t/ D ReΠN x .x; N i !N tN;
N y/e (2)
y .x; y; t/ D ReΠN y .x; N i !N tN;
N y/e (3)
˚.x; y; z; t/
p D ReŒ'.
N x; N zN/e i !N tN;
N y; (4)
gL

where w.N x; N Nx .x;

N y/, N Ny .x;
N y/, N y/
N denote the dimensionless Mindlin plate deflec-
tion, rotation about the y-axis and rotation about the x-axis, respectively, and
N x;
'. N y;
N zN/ is the dimensionless water velocity potential. Re denotes the real part
of the complex number.

4.3 Mathematical Formulation for Plate and Water Motions

4.3.1 Eq. of Motion and Boundary Conditions for Plate

The equations of motion of the Mindlin plate are given as [42]

" ! !#
@2 w N @2 wN @ Nx @ Ny
 G hN
2
C 2 C C C  hN!N 2 wN D wN  i !N ';
N (5)
@xN 2 @yN @xN @yN
! !  
D .1  / @2 N x @2 N x D .1 C / @2 N x @2 N y 2 N N @wN
C C C   G h x C
2 @xN 2 @yN 2 2 @xN 2 N yN
@x@ @xN
 hN3 2 N
D !N x ;
12
(6)
! !  
D .1  / @2 Ny @2 Ny D .1 C / @2 Ny @2 Nx N
C C C   2 G hN N y C @w
2 @xN 2 @yN 2 2 @yN 2 N xN
@y@ @yN
 hN3 2 N
D !N y ;
12
(7)

where D D E hN 3 =Œ12.1   2 / is the flexural rigidity, E the Young’s modulus,

G D E=Œ2.1 C / the shear modulus,  the Poisson ratio,  the specific gravity of
the plate material and  2 the shear correction factor (normally taken as 5/6).
Hydroelastic Analysis and Response of Pontoon-Type Very Large Floating Structures 111

Since all the edges of the plate are free, the bending moment, twisting moment
and transverse shear force must vanish at the edges, i.e.
 N 
@ n @ Ns
Bending moment Mnn D D C D 0; (8)
@n @s
  N 
1 @ n @ Ns
Twisting moment Mns D D C D 0; (9)
2 @s @n
 
@wN
Shear force Qn D  2 G hN C N n D 0; (10)
@n

where s and n denote the tangential and normal directions to the cross-section of the
plate.
For convenience, the overbar embellishment for the variables in the foregoing
equations will be dropped from hereon.

4.3.2 Eq. of Motion for Water

The linear wave potential theory is used to model the fluid-motion problem by using
the velocity potential '. The single frequency velocity potential '.x; y; z/ of the
water must satisfy the Laplace’s equation [43],

r 2 ' D 0 in ˝; (11)

where r 2 ./ D @2 ./ =@x C @2 ./ =@y C @2 ./ =@z and the boundary conditions
are
@'
D i !w on SHB ; (12)
@z
@'
D 0 on SHS ; (13)
@n
@'
D ! 2 ' on SF ; (14)
@z
@'
D 0 on SB ; (15)
@z

where n is the unit normal vector to the surface S. The wave velocity potential must
also satisfy the Sommerfeld radiation condition as jxj ! 1 [44],

p  @ 
lim jxj  i k .'  'I n / D 0 on S1 ; (16)
jxj!1 @jxj
where x D .x; y/ and S1 the artificial fluid boundary at infinity. The wave number k
satisfies the dispersion relationship

ktanh.kH / D ! 2 ; (17)
112 C.M. Wang and Z.Y. Tay

and 'I n is the incident velocity potential given by

A cosh .k .z C H // i k.x cos  Cy sin  /

'I n D e ; (18)
! cosh kH
where A is the wave amplitude.

4.4 Modal Expansion Method for Solving Coupled

Water-Plate Problem

The plate equations (5) to (7) indicates that the response of the plate w is cou-
pled with the fluid motions (or velocity potential '). On the other hand, the fluid
motion can only be obtained when the plate deflection w is specified in the boundary
condition (12). In order to decouple this interaction problem into a hydrodynamic
problem in terms of the velocity potential and a fluid-plate vibration problem in
terms of the generalized displacement, we adopt the modal expansion method as
proposed by Newman [7]. In this method, the plate deflection is expanded by a series
of the products of the modal functions clw .x; y/ and their corresponding complex
amplitudes lw ,
XN
w.x; y/ D lw clw .x; y/; (19)
lD1

where N denotes the total number of modes.

As the problem is linear, the total velocity potential can be represented by a linear
superposition of the diffracted part 'D (D 'I n C 'S , where 'S is the scattered
potential) and the radiated part 'R . By using the modal expansion method, the total
velocity potential ' may be expressed as

X
N
'
'.x; y; z/ D 'D .x; y; z/ C 'R .x; y; z/ D 'D .x; y; z/ C l l 'l .x; y; z/; (20)
lD1

where 'lD1;2;:::;N is the radiated potential corresponding to the unit-amplitude

motion of the l-th modal function of the freely vibrating plate. Note that the complex
'
amplitudes l l in Eq. (20) have been assumed to be equal to lw in Eq. (19).
By substituting Eqs. (19) and (20) into the Laplace equation (11) and the fluid
boundary condition [Eqs. (12) to (16)], we arrive at the following governing equa-
tion and boundary conditions for each of the unit-amplitude radiated potential (i.e.
for l D 1; 2; :::; N ) and the diffracted potential (i.e. for l D D).

r 2 ' D 0 in ˝; (21)
Hydroelastic Analysis and Response of Pontoon-Type Very Large Floating Structures 113

@'l i !clw .x; y/ for l D 1; 2; :::; N
D on SHB ; (22)
@z 0 for l D D
@'l
D 0 on SHS ; (23)
@n
@'l
D ! 2 'l on SF ; (24)
@z
@'l
D 0 on SB ; (25)
@z

and the Sommerfeld radiation condition

p  @ 
lim jxj  i k .'  'I n / D 0 on S1 : (26)
jxj!1 @jxj

Now, the boundary value problems for each of the unit-amplitude radiated potential
and diffracted potential are given explicitly in an uncoupled form.

5 Method of Solution

The method of solution for solving the hydroelastic problem is shown in the
appendices and involves the following steps:
Step 1 Transform the Laplace equation (21) with the boundary conditions given
by Eqs. (22) to (26) into a boundary integral equation. This step can be best
done by using the boundary element method (BEM) because this method
can easily handle the Sommerfeld radiation condition using the free sur-
face Green function. For details on the transformation of the Laplace
equation into a boundary integral equation using the free surface Green
function, please refer to Appendix 1.
Step 2 The governing equation for the plate [(5) to (7)] is solved by using the
finite element method. The plate is discretised into a finite number of
8-node Mindlin plate elements. Substitute the velocity potential 'l
obtained from Step 1 into the discretised equation of motion (assembled
in the global form). This step is shown in Appendix 2.
Step 3 Expand the plate displacement field w O as a series of products of vibration
modes cOl and complex amplitudes l . The vibration modes (eigenvectors)
could be obtained by performing a free vibration analysis of the plate.
Solve the assembled equation of motion for the plate in Step 2 for the
complex amplitudes l . This step is shown in Appendix 2.
Step 4 Multiply the complex amplitudes l computed in Step 3 with the vibration
modes cOl to obtain the plate deflection (hydroelastic response). The stress
resultants may be readily obtained from taking appropriate derivatives of
the deflection and rotations.
114 C.M. Wang and Z.Y. Tay

6 Results and Discussions

Some example problems are treated to demonstrate the modelling and the hydroe-
lastic analysis of pontoon-type VLFS with and without the presence of breakwaters.
Moreover, the effects of submerged-type anti-motion device and semi-rigid connec-
tion joints in reducing the hydroelastic response of the VLFS will be discussed.

6.1 Numerical Modelling of VLFS

The detailed model of an actual VLFS (with bulkheads and stiffeners) consists of
a huge number of degrees of freedoms, thereby making the hydroelastic analysis
on such a model very time-consuming. In order to reduce computational time, the
detailed structural model is usually replaced by an equivalent solid plate with the
same dimensions (i.e. length, width and height being kept the same as the actual
structure), but its Young’s modulus and Poisson’s ratio are tweaked in order to match
the vibration modes and natural frequencies of the actual structure.
We shall use the example problem of the Marina Bay floating performance
stage [38] (see Fig. 2a), to describe the procedure in obtaining the equivalent solid
plate that modeled the actual floating structure. The floating performance stage con-
sists of 15 pontoons of 40m  16.6m  1.2m laid out in a 5 by 3 configuration to
produce a large platform measuring 120m  83m  1.2m. The thickness of the top
surface slab is 12mm and all the side and bottom slabs are 8mm. The pontoon has a
system of stiffeners to stiffen the structure as shown in Fig. 7.
Wang et al. [38] created a 3-D finite element model of the floating stage in the
software ABAQUS as shown in Fig. 8a. The stiffeners, top and bottom slabs, side
slabs as well as interior webs are discretized with a combination of plate and beam
elements. The side and corner (diamond shape) connectors are used to hold the
pontoons together and these connectors are also modeled in detail.
Figure 8b shows the simplified model of the floating performance stage, con-
sisting of 15 equivalent solid isotropic plates that model the pontoons, 20 corner
connectors and plate strips along their boundaries that simulate the connector
regimes. The properties of the components in the simplified plate model are given
in Table 1. These properties are obtained by keeping the length, width and height
dimensions of the simplified model to be the same but the Young’s modulus and
Poisson’s ratio adjusted to match the first three vibration frequencies fi and modes
of the 3-D finite element model as shown in Fig. 9.
By establishing the validity of the simplified plate model, the hydroelastic anal-
ysis of the floating performance stage may be carried out with a considerably less
computational effort. Studies have shown that hydroelastic responses obtained from
such simplified plate models are in good agreement with experimental test results
(see for example, refs. [13–15]).
Hydroelastic Analysis and Response of Pontoon-Type Very Large Floating Structures 115

Fig. 7 System of stiffeners in the floating performance platform.

(a) 3-D FEM model. (b) Simplified model.

Fig. 8 Fig. depicting plan view of (a) 3-D FEM model (b) simplified model.

Table 1 Young’s modulus, Poisson’s ratio and equivalent mass density for the equivalent plate
model components.
Material Pontoon Strip Connector
Young’s modulus (GPa) 10 0:05 0:3
Poisson’s ratio 0:1 0:1 0:1
Mass density (kg m3 ) 233:22 233:22 233:22

6.2 Hydroelastic Response of VLFS

The hydroelastic response of the floating structure depends on the operating sea
states and water depths. By using the same example problem of the Marina Bay
floating performance stage, we shall now discuss the hydroelastic response of the
116 C.M. Wang and Z.Y. Tay

(a) 1st mode, f1 D 0:995Hz.

(b) 2nd mode, f2 D 1:023Hz.

(c) 3rd mode, f3 D 1:072Hz.

Fig. 9 First three vibration modes and natural frequencies of (a) 3-D FEM model (b) simplified
plate model.

simplified model (Fig. 8b) operating at a sea state with a constant water depth H of
3m and a wave period T of 2.4s. The deflection for a unit wave amplitude wmax =A of
the simplified model under head sea and beam sea conditions are shown in Fig. 10.
The hydroelastic responses due to the wave loads (shown in Fig. 10) are rather small
compared to the deflections resulting from the heavy static loads as reported in [38].
This is because of the benign sea state condition on site. Note that the wave length

=L is only 0.07.
Next, we shall show the effect of wavelength on the hydroelastic response of
VLFS treated by Yago and Endo [13]. The VLFS measures 300m  60m  2m
and floats in very deep waters (i.e. H D 1). Figure 11 shows the deflection
Hydroelastic Analysis and Response of Pontoon-Type Very Large Floating Structures 117

0.15
0.3
0.1
0.2
w

w
0.05 0.1
0 0
0 20 0 20
0 40 0 40
20 60 20 60
40 x 40 x
80 80
y y
60 100 60 100
80 120 80 120

(a) Head sea, wmax =A D 0:089. (b) Beam sea, wmax =A D 0:344.

Fig. 10 Deflection at the bottom plate of floating stage under (a) head sea (b) beam sea. Wave
period T = 2.4s. Water depth H = 3m.

(a)
=L D 0:4. (b)
=L D 0:5.

(c)
=L D 0:6. (d)
=L D 1:0.

Fig. 11 Vertical deflection along the centreline of the VLFS at (a)

=L D 0:2 (b)
=L D 0:4
(c)
=L D 0:6 (d)
=L D 1:0.
118 C.M. Wang and Z.Y. Tay

along the centreline of the VLFS under a head sea condition. The response of the
floating structures increases with respect to
=L. This shows that the hydroelastic
response of the VLFS is significantly affected by the wave lengths
=L found at the
operational sea state.
Utsunomiya and Watanabe [45] also investigated the effect of water depth on the
hydroelastic response of VLFS. They considered a 1500m  150m  1m pontoon-
type VLFS floating on a relatively shallow constant water depth H = 8m as well
as a variable water depth. The contour plot of the variable water depth is shown in
Fig. 12.
A comparison between the deflection (for unit wave amplitude) of the VLFS
under the wave angle  D =4 in constant and variable water depth is presented in
Fig. 13. A considerable difference in the response characteristic is observed when
the VLFS is located in either a constant or a variable water depth. Thus, in the
relatively shallow water case, the effect of a water depth under the VLFS becomes
significant and must be accounted for in the hydroelastic analysis.

Fig. 12 Contour plot of the variable water depth.

(a) Constant water depth (H = 8m). (b) Variable water depth.

Fig. 13 Deflection of VLFS under wave angle  D =4 and wavelength

D 156:8m at
(a) constant water depth (b) variable water depth.
Hydroelastic Analysis and Response of Pontoon-Type Very Large Floating Structures 119

6.3 Mitigation of Hydroelastic Response

The previous section shows that the hydroelastic response of VLFS could be very
large especially when subjected to sea states with large wavelengths or when the
VLFS is floating on shallow water with a variable seabed topology. This has prompt
researchers to figure out ways to mitigate the response of the structure under wave
action. The conventional way of reducing the response of the structure under wave
action is by constructing breakwaters close to the VLFS in order to attenuate the
wave forces. Utsunomiya et al. [14] carried out hydroelastic analysis on VLFS
with and without the presence of bottom-founded type breakwater. Tay et al. [39]
showed that the floating-type breakwater is effective in attenuating the wave forces
impacting on the two adjacently placed floating fuel storage modules. The effects of
breakwaters on the responses of their respective VLFS are shown in Fig. 14.
The use of breakwaters to attenuate the wave forces impacting on VLFS is rela-
tively expensive and requires more time for construction. This prompted engineers
to invent anti-motion devices that are attached to the VLFS. For example, Takagi
et al. [29] proposed a submerged box-shape device attached to the fore-end of the
floating structure (see Fig. 15a). He carried out hydroelastic analysis and showed
that the anti-motion performance of the box-shape device is effective in reducing
the deformation, shearing force and bending moment of the floating structure (see
Figs. 15b-d).
Ohta et al. [28] also carried out hydroelastic analysis as well as experimental
tests to investigate the effect of a submerged vertical plate (Fig. 16a) on the hydroe-
lastic deflection of VLFS. The reduction in deflection was found to increase with
increasing submerged plate depth and decreasing wave length. Ohta et al. [28] and

(a) Very large floating structure (VLFS). Wave (b) Floating fuel storage facility (FFSF).
length
=L D 0:125. Wave length
=L D 0:4.

Fig. 14 Comparison of deflection amplitude with and without breakwater for (a) very large
floating structure (b) floating fuel storage facility.
120 C.M. Wang and Z.Y. Tay

(a) Box-shape anti-motion device. (b) Beflection.

(c) Shearing force. (d) Bending moment.

Fig. 15 (a) box-shape anti-motion device. Effect of box-shape anti-motion device towards reduc-
tion in (b) deflection (c) shearing force (d) bending moment. Wave period T = 6s. Wave direction:
Head sea.

Fig. 16 Very large floating structure with attached submerged anti-motion device in the form of
(a) vertical plate (b) horizontal plate.

Watanabe et al. [30,31] carried out hydroelastic analysis on VLFS with a submerged
horizontal plate attached at the fore-end of the structure (Fig. 16b). They found that
the horizontal plate is effective in reducing the deflection of the floating structure if
the plate is placed at an appropriate depth from the floating structure. However, the
depth requirement of these box-shape and submerged plate anti-motion devices will
result in an increase in steady drift forces. Hence the size and cost of the mooring
dolphin system will also increase correspondingly.
Khabakhpasheva and Korobkin [40] proposed a novel design in reducing the
hydroelastic response of the VLFS by connecting the VLFS with an auxiliary
module using hinge or semi rigid joints. They found that the auxiliary structure
is able to reduce the deflection of the main floating structure significantly. Wang
et al. [41] then proposed the use of semi-rigid mechanical joints to connect the
modules as shown in Fig. 17a. They found that by using an appropriate rotational
stiffness " and location of such a mechanical joint, the hydroelastic response for the
interconnected floating beams can be significantly reduced as shown in Fig. 17b.
Hydroelastic Analysis and Response of Pontoon-Type Very Large Floating Structures 121

Fig. 17 (a) Schematic diagram of interconnected beams (b) Normalised maximum deflection of
interconnected beams with varying rotational stiffness .

Wang et al. [41] claimed that the semi-rigid mechanical joint is more effective in
reducing the hydroelastic response when compared to the rigid joint or a simple
hinge mechanical joint.

7 Summary and Future Studies on Hydroelastic

Analysis/Response of VLFS

We have presented the mathematical formulation for the hydroelastic analysis of

pontoon-type VLFS in the frequency domain framework. In order to reduce compu-
tational time, a simplified plate model having the same dynamic properties (in terms
of vibration modes and natural periods) as the actual VLFS is generally used for the
hydroelastic analysis.
In the hydroelastic analysis, we have employed the coupled finite-element boun-
dary-element (FE-BE) method to solve the coupled water-plate problem. We assume
the water to be a perfect fluid and the fluid motion is irrotational, so that a velocity
potential ' exists. The fluid motion is governed by the Laplace equation r 2 ' D 0
and the potential ' must satisfy the Neumann boundary conditions on the wetted
surface of the floating bodies, the free surface and the seabed as well as the Som-
merfeld radiation condition at the far field. On the other hand, the plate is assumed
to be perfectly flat and may be modelled using the Mindlin plate theory. To obtain
122 C.M. Wang and Z.Y. Tay

the deflection w of the floating plate, one needs to solve the plate equation of motion
that unfortunately involves the fluid velocity potential '. In order to determine the
velocity potential ', one needs to solve the Laplace equation which in turn has
a surface boundary condition in contact with the floating plate that contains the
unknown deflection w. In order to decouple this water-plate interaction problem,
we have adopted the modal expansion method. In this method, we take a series of
modal functions obtained from a free vibration analysis of the plate and substitute
them into the boundary condition associated with the water surface and plate sur-
face interface, thereby removing the unknown deflection w. We solve the Laplace
equation which was transformed into a boundary integral equation via the free sur-
face Green’s function for the potential '. The computed potential ' is substituted
into the equation of motion of the floating plate as the hydrodynamic force and the
equation is solved using the finite element method for the complex amplitudes l of
the plate. The deflection w of the plate can then be obtained by taking the product of
the modal functions cl of the freely vibrating plate and the complex amplitudes l .
The hydroelastic responses computed from this coupled finite-element boundary
element method were found to be in good agreement with experimental test results.
Parametric studies by researchers have shown that the hydroelastic response of
VLFS is significantly affected by the flexural rigidity of the floating structure, wave
length and variation in water depth due to the undulating sea bed.
Various methods for reducing the hydroelastic response of the VLFS were pre-
sented. Breakwaters were found to be effective in attenuating the wave forces but are
expensive and time consuming to construct, especially the bottom-founded type. On
the other hand, the submerged anti-motion devices that are attached to the floating
structure are relatively easy to install and effective in reducing motion. Herein, we
have discussed three different types of submerged anti-motion devices (i.e. the box-
shape device, submerged vertical and horizontal plate). These anti-motion devices
were found to be effective in reducing the hydroelastic responses of floating struc-
tures if properly designed. However, these anti-motion devices may induce large
steady drift forces on the VLFS due to their relatively large submerged surfaces. A
novel method in mitigating the response of the structure is to attach auxiliary hor-
izontal plates to the VLFS using mechanical joints of varying rotational stiffness.
Preliminary studies have shown that semi-rigid mechanical joints are more effective
in reducing the hydroelastic deflection of the VLFS as compared to purely hinge
connection or rigid connection.
Despite considerable research effort being made in developing hydroelastic
analysis tools for VLFS, studies are still required. These future studies include
 Non-linear analysis The frequency domain analysis is not valid in extreme sit-
uations such as in a large storm. Such extreme events have to be considered in
the VLFS design for safety and survivability reasons. More work should be done
to investigate the non-linear responses such as the transient response of VLFS
under large wave impact [46].
 Arbitrary shaped VLFS As opposed to the conventional rectangular shaped
VLFS, architects prefer aesthetically pleasing shapes of VLFS such as the com-
plicated shape for the floating crab restaurant as shown in Fig. 2d. Research work
Hydroelastic Analysis and Response of Pontoon-Type Very Large Floating Structures 123

is needed to develop efficient hydroelastic analysis method to handle complicated

shapes of VLFS.
 Wave-structure-liquid interaction For VLFS used as floating fuel storage mod-
ules, it is necessary to consider the free surface effect due to the sloshing behavior
of fuel in the storage modules. The free surface effect might be crucial towards
the response of the floating storage module especially when resonance occurs.
Some research studies have been made on the sloshing of liquid in tankers but
without considering the coupled interaction of liquid-structure with the wave
action. Moreover, the structure is usually assumed to be a rigid body in these stud-
ies. Wave-structure-fluid interaction that takes into account the elastic behavior
of the structure is a potential line of investigation.
 Articulated VLFS While studies on the interconnected floating beams (Fig. 17)
only focused on 2-D problems and a study has been done on articulated floating
plates [47], future studies could be made on optimiziting the rotational stiff-
ness of the connections and their positions in order to develop more effective
connector system for reducing the hydroelastic responses.
 Hydroelastic analysis using Navier-Stokes equation Vorticity and viscous effects
are usually neglected in the hydroelastic analysis using the potential theory. How-
ever, high vorticity of fluid is expected from tsunami waves impacting VLFS
or in the case of VLFS fitted with anti-motion devices such as bilge keels and
submerged plates. In such situations, the fluid has to be represented by the
Navier-Stokes (NS) equation that allows for rotational flow.
 Slowly varying drift forces The mooring dolphin systems that hold the floating
modules in place are usually designed based on the drift/mean forces. Further
research studies are needed to include the slowly varying drift forces in the design
of the mooring dolphin system for the VLFS located in multi-directional random
seas.

References

1. Pernice R (2009) Japanese urban artificial islands: an overview of projects and schemes for
marine cities during 1960-1990s. J Archit Plann, AIJ, 74(641):1847–1855
2. Suzuki H. and Yoshida K. Design flow and strategy for safety of very large floating structure.
Proc of Int Wksp on Very large Floating Struct, VLFS’96, Hayana, Japan, 1996, 21–27
3. Chakrabarti SK (1988) Hydrodynamics of Offshore Structures. WIT Press, UK
4. John F (1949) On the motion of floating bodies. Com on Pure and Appl Math, Part I, 2:13–57
5. John F (1950) On the motion of floating bodies. Com on Pure and Appl Math, Part II, 3:45–100
6. Wehausen J and Laitone E (1960) Surface wave. Fluid Dynamics III, Handbuch der Physik,
Springer, Berlin
7. Newman JN (1977) Marine Hydrodynamics. The MIT Press, Cambridge
8. Newman JN (1985) Algorithm for the free-surface Green function. J Engrg Math, 19:57–67
9. Newman JN (1994) Wave effects on deformable bodies. Appl Ocean Res, 16:47–59
10. Noblesse F (1979) Potential theory of steady motion of ships. Department of Ocean Engineer-
ing Report, MIT
11. Mei CC (1989) The Applied Dynamics of Ocean Surface Waves - Advance Series on Ocean
Engineering, Volume 1, World Scientific, Singapore
124 C.M. Wang and Z.Y. Tay

12. Ertekin RC, Riggs HR, Che XL and Du SX (1993) Efficient methods for hydroelastic analysis
of very large floating structures. J Ship Res, SOEST No. 3155, 37(1):58–76
13. Yago K and Endo H (1996) On the hydroelastic response of box-shaped floating structure with
shallow draft. J Soc Nav Arch Japan, 180:341–352
14. Utsunomiya T, Watanabe E and Eatock Taylor R. Wave response analysis of a box-like VLFS
close to a breakwater. Proc of 17th Int Conf on Offshore Mech and Arctic Engrg, Lisbon,
Portugal, July 5-9, 1998, OMAE98-4331
15. Kashiwagi M (1998) A b-spline Galerkin scheme for calculating the hydroelastic response of
a very large floating structure in waves. J Mar Sci Technol, 3:37–49
16. Ohmatsu S. Numerical calculation method of hydroelastic response of a pontoon-type VLFS
close to a breakwater. In: Ertekin RC, Kim JW (eds). Proc 3rd Int Wksp Very Large Float-
ing Struct, University of Hawaii at Manao, Honolulu, Hawaii, USA, September 22-24, 1999,
2:805–811
17. Meylan MH and Squire VA (1996) Response of a circular ice floe to ocean waves. J Geophys
Res, 101(C4):8869–8884
18. Meylan MH (1997) The forced vibration of a thin plate floating on an infinite liquid. J Sound
and Vib, 205(5):581–591
19. Meylan MH (2001) A variation equation for the wave forcing of floating thin plates. J Appl
Ocean Res, 23(4):195–206
20. Squire VA (2008) Synergies between VLFS hydroelasticity and sea ice research. Int J Offshore
and Polar Engrg, 18(3):1–13
21. Hermans AJ (2000) A boundary element method for the interaction of free-surface waves with
a very large floating flexible platform. J Fluids Struct, 14:943–956
22. Inoue K (2001) Stress analysis of Mega-Float in waves by two-step method. Proc of 20th
Int Conf on Offshore Mech and Arctic Engrg, Rio de Janeiro, Brazil, June 3-8, 2001,
OMAE2001/OSU-5209
23. Kada K, Fujita T and Kitabayashi K (2002) Stress analysis for structurally discontinuous parts
in a Mega-Float structure. Int J Offshore and Polar Engrg, 12:48–55
24. Seto H, Ochi M, Ohta M and Kawakado M. Hydrodynamic response analysis of real very large
floating structures in regular waves in open/sheltered sea. Proc of Int Symp on Ocean Space
Utilization Tech, Tokyo, 2003, 85–93
25. Seto H, Ohta M, Ochi M and Kawakado M (2005) Integrated hydrodynamic-structural analysis
of very large floating structures VLFS. Mar Struct, 18(2):181–200
26. Sasajima H. Local structural analysis of large floating structures. In: Ertekin RC, Kim JW
(eds). Proc 3rd Int Wksp Very Large Floating Struct, University of Hawaii at Manao, Honolulu,
Hawaii, USA, September 22-24, 1999, 2:602–606
27. Inoue K, Nagata S and Niizato H (2003) Stress analysis of detailed structures of Mega-Float in
irregular waves using entire and local structural models. Proc 4th Int Wksp Very Large Floating
Struct, Tokyo, Japan, 219–228
28. Ohta H, Torii T, Hayashi N, Watanabe E, Utsunomiya T, Sekita K and Sunahara S. Effect
of attachment of a horizontal/vertical plate on the wave response of a VLFS. In: Ertekin RC,
Kim JW (eds). Proc 3rd Int Wksp Very Large Floating Struct, University of Hawaii at Manao,
Honolulu, Hawaii, USA, September 22-24, 1999, 265–274
29. Takagi K, Shimada K and Ikebuchi T (2000) An anti-motion device for a very large floating
structure. Mar Struct, 13:421–436
30. Watanabe E, Utsunomiya T, Kuramoto M, Ohta H, Torii T and Hayashi N. Wave response
analysis of VLFS with an attached submerged plate. Proc 12th Int Offshore Polar Engrg,
Kitakyushu, Japan, May 26-31, 2002:319–326
31. Watanabe E, Utsunomiya T, Ohta H and Hayashi N. Wave response analysis of VLFS with
an attached submerged plate: verification with 2-D model and some 3-D numerical examples.
Proc Int Symp Ocean Space Utilisation technol, National Maritime Research Institute, Tokyo,
Japan, January, 2003, 147–154
32. Kim JW and Webster WC (1998) The drag on an airplane taking off from a floating runway. J
Mar Sci Tech, 3(2):76–81.
Hydroelastic Analysis and Response of Pontoon-Type Very Large Floating Structures 125

33. Watanabe E, Utsunomiya T and Tanigaki S (1998) A transient response analysis of a

very large floating structure by finite element method. Structural Engrg/Earthquake Engrg,
JSCE;15(2):155–163
34. Kashiwagi M (2000) A time-domain mode-expansion method for calculating transient elastic
responses of a pontoon-type VLFS. J Mar Sci Technol, 5:89–100
35. Kashiwagi M (2004) Transient responses of a VLFS during landing and take-off of an airplane.
J Mar Sci Tech, 9(1):14–23.
36. Watanabe E, Utsunomiya T, Wang CM and Le Thi Thu Hang (2000) Benchmark hydroelastic
responses of circular VLFS under wave action. Engrg Struct, 28(6):423–30
37. Wang CD and Wang CM. A comparative study on the linear wave response of a very large
floating body modeled by a plate based on Kirchhoff and Mindlin plate theories. Proc of
ICETECH: Structures in ice/iceberg populated waters, Banff, Canada, July 16-19, 2006,
ICETECH06-157-RF
38. Wang CM, Song JH, Utsunomiya T, Koh HS and Lim YB. Hydroelastic analysis of floating
performance stage @ Marina Bay, Singapore. Proc of 27th Int Conf on Offshore Mech and
Arctic Engrg, Estoril, Portugal, June 15-20, 2008, OMAE2008-57324
39. Tay ZY, Wang CM and Utsunomiya T (2009) Hydroelastic responses and interactions of
floating fuel storage modules placed side-by-side with floating breakwater. Mar Struct,
22(3):633–658
40. Khabakhpasheva TI and Korobkin AA (2002) Hydroelastic behavior of compound floating
plate in waves. J Engrg Maths, 44:21–24
41. Wang CM, Muhammad R and Choo YS. Reducing hydroelastic response of interconnected
floating beams using semi-rigid connections. Proc of 28th Int Conf on Offshore Mech and
Arctic Engrg, Honolulu, Hawaii, May 31-June 5, 2009, OMAE2009-79692
42. Mindlin RD (1951) Influence of rotary inertia and shear on flexural motions of isotropic, elastic
plates. J Appl Mech, 18:31–38
43. Wang CM, Watanabe E and Utsunomiya T (2008) Very Large Floating Structures. Routledge,
UK: Taylor and Francis
44. Sarpkaya T and Isaacson M (1981) Mechanics of Wave Forces on Offshore Structrures. Van
Nostrand Reinhold, New York, USA
45. Utsunomiya T and Watanabe E (2006) Fast multipole method for wave diffraction/radiation
problems and its application to VLFS. Int J Offshore and Polar Engrg, 16(4):253–260
46. Watanabe E, Utsunomiya T and Wang CM (2003) Hydroelastic analysis of pontoon-type
VLFS: a literature survey. Eng Struc, 26:245–256
47. Xia D, Kim JW and Cengiz Ertekin R (2000) On the hydroelastic behavior of two-dimensional
articulated plates. Mar Struct, 13:261–278
48. Linton CM (1999) Rapidly convergent representation for Green’s functions for Laplace’s
equation. Proc Roy Soc, London, A(455):1767–1797
49. Petyt M.(1990) Introduction to Finite Element Vibration Analysis. Cambridge University
Press, Cambridge, UK
50. Cengiz Ertekin R, Wang SQ, Che SL and Riggs HR (1995) On the application of the Haskind-
Hanaoka relations to hydroelastic problems. Mar Struct, 8:617–629

Appendix 1 Solution for Velocity Potential Components:

Radiated and Diffracted Potentials

The boundary value problem defined by Eqs. (21) – (26) can be solved by using
the boundary element method. To transform the Laplace equation into a boundary
integral equation, we use the free-surface Green function that satisfies the bound-
ary conditions given by Eqs. (22) – (26). The boundary integral equation for each
velocity potential component can be defined as [43]
126 C.M. Wang and Z.Y. Tay
Z Z
@G.x; x0 / @'l .x0 /
 2'l .x/ C 'l .x0 / dS D G.x; x0 / dS; (27)
S @n S @n

where x D .x; y; z/ is the source point and x0 D .x 0 ; y 0 ; z0 / the field point. G.x; x0 /
is a free-surface Green’s function for water of finite depth and is given by [48]
1
X 2K0 .km R/
G.x; x0 / D   cos km .z C H / cos km .z0 C H /; (28)
sin 2km H
mD0
h
2
1 C 2km H

where km is a positive root number satisfying km tan km H D ! 2 , with .m  1/

and k0 D i k. K0 is the modified Bessel function of the second kind and R represents
the horizontal distance between x and x0 .
We refer to Eq. (27) as the integral equation for the water. By further introducing
the boundary conditions given by Eq. (22) for @'=@n into Eq. (27), Wang et al. [43]
obtained
Z
@G.x; x0 /
 2'l .x/ C S 'l .x0 / dS
S S @n
( HS R HB
i ! SHB G.x; x0 /clw .x0 / dS for l D 1; 2; :::; N
D ; (29)
4'I n for l D D

where SHS is the wetted side surface of the floating body and SHB the wetted
bottom surface of the floating body. The integral equation (29) is then solved for 'l
by using the constant panel method [44]. The integration in Eq. (29) is performed
using the Gauss quadrature method.
With the computed 'l , the radiated potential 'R and diffracted potential 'D
(introduced in Eq. 20) can be written in the matrix form as

f'R g D f'l gf l g D i !f'Ql gfcOlw gf l g

   !
1 1 @G 1
D i !  ŒI C ŒG fcOlw gf l g;
2 2 @n
(30)
  
1 @G 1
f'D g D 2  ŒI C f'I ng; (31)
2 @n

where ŒG is the global matrix for the free surface Green function, ŒI the identity
matrix, fcOlw g the mode shapes (eigenvectors) obtained by performing a free vibration
analysis on the plate and
  
1 1 @G 1
f'Ql g D  ŒI C ŒG : (32)
2 2 @n
Hydroelastic Analysis and Response of Pontoon-Type Very Large Floating Structures 127

Appendix 2 Solving Plate Eq. using Finite Element Method

The governing equations [(5) to (7)] of motion of the Mindlin plate are solved by
using the finite element method (FEM). For a rectangular plate, consider using the
8-node rectangular element of area 2a  2b as shown in Fig. (18a) (see ref. [49] for
details). By mapping into the  -space [Fig. (18b)] where  D x=a and D y=b,
the displacement field vector we in the element domain e may be discretised as
2 3
X 8 h i X 8 Nje 0 0
6 7 e Oe Oe T
we D NO je fw
O ej g D O j xj yj g ;
4 0 Nje 0 5 fw (33)
j D1 j D1 0 0 Nje

where wO ej is the transverse deflection, O xj

e
the rotation about the -axis and O yj
e
the
rotation about the -axis at the j -th node of the plate element e [see Fig. (18b)]. The
basis functions Nje are given by
1

Nje .; / D 1 C j  1 C j j  C j  1 for nodesj D 1; 2; 3; 4;

4 (34)
1

Nj .; / D
e
1 2
1 C j for nodesj D 5; 6; (35)
2
1

Nje .; / D 1 C j  1  2 for nodesj D 7; 8: (36)

2
In view
 of the assumed displacement field, the elemental matrices for flexural
stiffness kf , shear stiffness Œks  and mass Œm are given by
Z h i
  h3 h e iT
kf 2424 D Bf ŒD Bfe dAe ; (37)
e 12
Z
 T  
Œks 2424 D  2 h Bse ŒDs  Bse dAe ; (38)
e
2 3
Z h iT h 03 0 h i
6 7
Œm2424 D  NO e 4 0 h12 0 5 NO e dAe ; (39)
e 3
0 0 h12

Fig. 18 8-node rectangular Mindlin plate element (a) in physical space and (b) mapped into  -
space.
128 C.M. Wang and Z.Y. Tay

where dAe D d d .h Thei flexural rigidity ŒD, shear rigidity ŒDs , flexural strain-
 
displacement matrix Bfe , the shear strain-displacement matrix Bse and the basis
h i
function matrix NO e are given by

2 3
1 0
E 4 1 0 5;
ŒD33 D (40)
.1   2 /
0 0 1
2
 
E 10
ŒDs 22 D ; (41)
2 .1 C / 0 1
2 dN e dN e
3
h i 0 0  a1 d 1          0 0  a1 d 8
6 e
dN e 7
Bfe D6 1 dN1
4 0 b de 0          0 b1 d8 0 7;
5 (42)
324 e e e
1 dN1 1 dN1 1 dN8 1 dN8
0 a d   b d          0 a d   b d
" dN e dN e
#
  1 1
0 N1e          a1 d 8 0 N8e
Bse 224
D 1 dN e
a d
dN e ; (43)
b d
1
N1e 0          b1 d8 N8e 0
2 e 3
h i N1 0 0          N8e 0 0
NO e D 4 0 N1e 0          0 N8e 0 5 : (44)
324
0 0 N1e          0 0 N8e

The equation of motion for the Mindlin plate element is given by

   
e
O g D ff g;
kf C Œks  C krf  ! 2 Œm fw (45)

  R h iT h e i R
where krf D e NO e NO dAe is the restoring force, ff g D i ! e f' e gdAe
the hydrodynamic force and ! the angular velocity of the wave. By assembling
Eq. (45) into the global form, one obtains
   

O D fFg;
Kf C ŒKs  C Krf  ! 2 ŒM C ! 2 ŒMa  fwg (46)
   
where Kf , ŒKs , Krf , ŒM and ŒMa  are the global flexural stiffness matrix,
the global shear stiffness matrix, the global restoring force matrix, the global mass
matrix and the global added mass matrix, respectively. The displacement vector fwg O
may be expanded in an appropriate set of modes fcOl g as

O D fcOl gf l g:
fwg (47)

Note that fcOl g could be obtained by performing a free vibration test on the Mindlin
plate. By substituting Eq. (47) into Eq. (46), one obtains
   

Q
fcOlT g ŒKs  C Kf C Krf  ! 2 ŒM C ! 2 ŒMa  fcOl gf l g D fcOlT gfFg D fFg;
(48)
Hydroelastic Analysis and Response of Pontoon-Type Very Large Floating Structures 129

where l D 1; 2; :::; N and f FQ g D fcOlT gfFg is the exciting force. By using the
computed velocity potentials from Eqs. (30) and (31), the global added mass ŒMa 
and the exciting force fFQ g can be calculated from

Z X
M
ŒMa  D f'Ql gdS D ui f'Ql .xi /g; (49)
SHB i D1
Z X
M
Q D i !fcOT g
fFg f'D gdS D i !fcOlT g ui f'D .xi /g; (50)
l
SHB i D1

where x 2 e ; xi and ui are sets of M integration points and their corresponding

weights over e . The generalized added mass is given as a complex matrix, and
thus it includes the effect of radiation damping [43].
It should be noted that fcOl g is a P  N matrix, fFg is a P  1 matrix and the
bracketed ./ term on the left hand side of Eq. (48) is a P  P matrix, where P
is the number of nodes in the plate domain and N the total number of modes. By
multiplying the inverse of the bracketed ./ term on the left hand side of Eq. (48)
Q the complex amplitudes f l g which is a P  1 matrix could be obtained.
with fFg,
These complex amplitudes f l g are then back-substituted into Eq. (47) for the plate
O
deflection fwg.
Note that one can also express the exciting force in Eq. (50) in terms of the
unit-amplitude radiated potential 'r D 'lD1;2;:::;N and the incident potential 'I n ,
thereby eliminating the need to solve for the scattered potential 'S . The alternative
expression of the exciting force is derived as follows. The exiting force in Eq. (50)
can be written in the integral form as
Z
F D i !clT 'D dSHB : (51)
SHB

By substituting the bottom hull surface boundary condition for the unit-amplitude
radiated potential (i.e. @'r =@n D i !cl D i !clT ) into Eq. (51), the exciting force
F can be re-written as
Z Z
@'r @'r
F D 'D dSHB D  .'I n C 'S / dSHB
@n SHB @n SHB
Z  
@'r @'r
D 'I n C 'S dSHB : (52)
SHB @n @n

The application of the Green’s second identity over the unit-amplitude radiated
potential 'r and the scattered potential 'S gives
Z   Z
@'r @'S @'r
'S  'r dSHB D 0 H) 'S dSHB
SHB @n @n SHB @n
Z
@'S
D 'r dSHB : (53)
SHB @n
130 C.M. Wang and Z.Y. Tay

By substituting Eq. (53) into Eq. (52) and using the bottom hull surface boundary
condition for the scattered potential @'S =@n D @'I n =@n, one obtains the exciting
force in the form of
Z  
@'I n @'r
F D 'r  'I n dSHB : (54)
SHB @n @n

This exciting force expression (54), in terms of the unit-amplitude radiated poten-
tial and incident potential, is referred to as the Haskind-Hanaoka relation [50]. The
Haskind-Hanaoka relation may be used to check the exciting force obtained from
the pressure distribution method.
Efficient Numerical Simulation and
Optimization of Fluid-Structure Interaction

M. Schäfer, D.C. Sternel, G. Becker, and P. Pironkov

Abstract The paper concerns the efficient numerical simulation and optimiza-
tion of fluid-structure interaction (FSI) problems. The basis is an implicit parti-
tioned solution approach involving the finite-volume flow solver FASTEST, the
finite-element structural solver FEAP, and the coupling interface MpCCI. Special
emphasis is given to the grid moving techniques for which algebraic and elliptic
approaches are considered. The possibilities for accelerating the computations by
the usage of multigrid methods, adaptive underrelaxation, and displacement predic-
tion are discussed. A concept for integrating the FSI solver into an optimization
procedure for FSI problems is presented. Numerical results are given to illustrate
the capabilities of the approaches considered.

1 Introduction

Fluid-structure interaction (FSI) problems occur in many applications in industry

and science. The numerical efficiency of the underlying coupled solution algorithm
is essential for a realistic numerical simulation of such kind of problems. In particu-
lar, this is important for a simulation based optimization of FSI problems requiring
multiple runs of the coupled solver. In the present paper we consider an implicit par-
titioned solution approach, which combines the advantages of weakly and strongly
coupled schemes. The method is realized based on the finite-volume flow solver
FASTEST [1] involving an ALE formulation, the finite-element structural solver
FEAP [28], and the coupling interface MpCCI [2]. For each time step the implicit
solution procedure consists in the application of different nested iteration processes
for linearization, pressure-velocity coupling, and linear system solving, which are
linked by an iterative fluid-structure coupling procedure. Several techniques are
considered to improve the efficiency and stability of the coupled solution procedure:

M. Schäfer, D.C. Sternel, G. Becker, and P. Pironkov

Institute of Numerical Methods in Mechanical Engineering, Technische Universität Darmstadt,
Dolivostr. 15, 64293 Darmstadt, Germany
e-mail: [email protected], www.fnb.tu-darmstadt.de

H.-J. Bungartz et al. (eds.), Fluid Structure Interaction II, Lecture Notes 131
in Computational Science and Engineering 73, DOI 10.1007/978-3-642-14206-2 6,
c Springer-Verlag Berlin Heidelberg 2010
132 M. Schäfer et al.

 advanced grid moving techniques,

 adaptive under-relaxation of the structural displacements,
 prediction of structural displacements,
 multigrid methods.
Special emphasis is given to the grid moving which is a crucial issue when large
deformations of the structure are involved. Different strategies based on linear inter-
polation, transfinite mapping, and Laplace equations are considered and combined
in order to ensure a dynamic grid movement procedure, which preserves the qual-
ity of the block-structured fluid grid even for large three-dimensional structure
deformations.
Based on the FSI solver an integrated approach for the shape optimization of FSI
problems is presented. Corresponding approaches have not been deeply investigated
yet in the literature [15]. We employ NURBS surfaces for the shape representa-
tion and a derivative-free mathematical optimization technique. A key feature of the
integrated method is the unified treatment of structural deformation due to FSI and
shape variation for fulfilling the optimization objectives.
For all aspects mentioned, results for representative test cases are given which
illustrate the functionality and performance of the approaches considered.

2 Governing Equations

For the fluid domain part ˝f we assume a flow of an incompressible Newtonian

fluid. In this case the basic conservation equations governing transport of mass and
momentum for a fluid control volume Vf with surface Sf are given by
Z
v  n dSf D 0 ; (1)
Sf
Z Z Z Z
@
f v dVf C f .v  vg /.v  n/ dSf D f ff dVf C Tf  n dSf ; (2)
@t
Vf Sf Vf Sf

where v denotes the velocity vector with respect to Cartesian coordinates x, t

denotes the time, f the fluid density, n the outward normal vector, and ff are external
volume forces (e.g., buoyancy forces). vg is the velocity with which Sf may move
(grid velocity) due to displacements of solid parts. The Cauchy stress tensor Tf for
incompressible Newtonian fluids is defined by
 
Tf D f rv C rvT  pI (3)

with the pressure p, the dynamic viscosity f , the vector gradient r, and the identity
tensor I.
Simulation and Optimization of Fluid-Structure Interaction 133

For the structure we denote a material point in the reference configuration as X

whose position in the current configuration is given by

x D .X; t/ : (4)

The displacements are evaluated by

u D x  X: (5)

For more details see [18, 30]. The basic balance equation for momentum for the
solid domain ˝s can be written as

r  .Fs STs / C s fs D s 
R; (6)

where  R D @2 .X; t/=@t 2 denotes the acceleration, Ss the second Piola-Kirchhoff

stress tensor, s the density of the solid, and fs external volume forces acting on the
solid (e.g., gravitational forces). Fs D @=@X represents the deformation gradient.
In the present investigation we consider the Saint Venant-Kirchhoff material law

Ss D s tr.E/ I C 2s E (7)

with the Green-Lagrangian strain tensor

1 T
ED .F Fs  I/ ; (8)
2 s

as kinematic property. s and s are the two Lamé constants.

The problem formulation has to be closed by prescribing suitable boundary and
interface conditions. On solid and fluid boundaries standard conditions as for indi-
vidual solid and fluid problems can be prescribed. For the velocities and the stresses
on a fluid-solid interface we have the conditions

v D P D vb and Tf n D Ts n ; (9)

where vb is the velocity of the interface and Ts D Fs Ss FTs = det Fs is the Cauchy
stress tensor of the solid.

3 Basic Numerical Techniques

Solid parts of the problem domain are treated by the finite-element solver FEAP
(see [28]). For the fluid parts the block-structured parallel multigrid finite-volume
flow solver FASTEST is employed (see [1, 23]). Both solvers involve second-order
spatial discretizations and fully implicit second-order time discretizations.
The fluid-structure coupling is realized via an implicit partitioned approach (see
also [22]). In Fig. 1 a schematic view of the iteration process, which is performed for
134 M. Schäfer et al.

Start End

yes
no
t = tmax? yes
next time step
no FSI
convergency
next FSI iteration

Fluid field Grid Movement

− FVM FASTEST − grid generation
− block−structured grid − flux correction

u
v,p

Coupling Interface Underrelaxation

F u

INTERPOLATION (MPCCI)

FEAP

− FEM
− unstructured grid

Fig. 1 Flow chart of coupled solution procedure.

each time step, is given. After the initializations the flow field is determined in the
actual flow geometry. Based on this the friction and pressure forces on the interact-
ing walls are computed and passed to the structural solver as boundary conditions.
The structural solver computes the deformations, which are used to modify the fluid
mesh. Then the flow solver is started over again.
For the mesh deformation algebraic and elliptic approaches are employed which
are described in detail in Sect. 4. Within the fluid solver a discrete form of the space
conservation law Z Z
@
dVf  vg  n dSf D 0 ; (10)
@t
Vf Sf

is used to compute the additional convective fluxes in Eqs. (1) and (2). This is done
using the swept volumes ıVc of the control volume faces for which one has the
relation (see [9]):

X ıV n Vfn  Vfn1 X
c
D D .vg  nSf /nc ; (11)
c
tn tn c
Simulation and Optimization of Fluid-Structure Interaction 135

where the summation index c loops over the faces of the control volume, the
index n denotes the time level tn and tn is the time step size. By this way inter-
face displacements enter the fluid problem part in a manner strictly ensuring mass
conservation.
The fluid-structure interaction (FSI) iteration routine is repeated until a conver-
gence criterion is reached, which is defined by the change of the mean displace-
ments:
1 X kuk;m1  uk;m k1
N
< DFSI
< "; (12)
N kuk;m k1
kD1

where m counts the FSI iterations, N denotes the number of interface nodes, and
k  k1 the infinity norm.
The data transfer between the flow and the solid solvers within the partitioned
solution procedure is performed via an interface realized by the coupling library
MpCCI (see [2]). MpCCI is used for controlling the data communication as well as
for carrying out the interpolations of the data within (possibly) non-matching fluid
and solid grids. The initialization routine provides MpCCI the geometry informa-
tion at the fluid-solid interface for both grids. Using these geometry informations
the forces at the control volume centers of the fluid grid interface are interpolated
onto the strucural grid and passed to the structural solver. The displacements from
the structural solver at the nodes are transfered via MpCCI to the control volume
vertices of the fluid grid interface. Afterwards the complete fluid grid is adapted
(see Sect. 4) and the corresponding coordinates of the control volume centers are
computed. An update of the geometry information of the solid grid is not necessary,
since the structural finite-element computations always relate to the original solid
grid. Note that nearly independent discretizations for the fluid and solid subprob-
lems can be used. For the necessary interpolations MpCCI employs a conservative
linear approach (see [2]).

4 Grid movements strategies

Several different strategies for the grid movement based on linear interpolation,
transfinite mapping, and Laplace equations are used. Aim is to provide a dynamic
grid distortion procedure, which preserves the grid quality even for large three
dimensional structural deformations. As illustrated in Fig. 1 every FSI iteration
causes a shape altering of the FSI interface which is followed by an update of the
spatial discretization of the flow domain. This implies a blockwise movement of the
grid which also takes the parallel implementation into account.
The information delivered by MpCCI from FEAP to FASTEST are the displace-
ments of the coupling faces which are the basis of the grid movement. At first the
grid of all blocks which contain a moved coupling face (primary blocks) should be
updated. In addition, it is essential that the grid of the blocks sharing an edge with
the coupling face are updated, too. Furthermore, a conditional face movement can
136 M. Schäfer et al.

Fig. 2 (a) Primary and secondary blocks (initial block positions). (b) Order of grid adaption
(distorted block positions).

be useful. For instance, the opposite face of a primary face can be distorted parallel
to the primary face or angles of edges can be preserved. This causes also the grid
updates within blocks containing no coupling face or an edge of a coupling face
(secondary blocks). In Fig. 2 an example for primary and secondary blocks of a
moveable grid is shown. The movement is induced by the structure inside the gap,
the coupling faces are the three faces around the gap. Also the order of the treatment
of the blocks must be defined (see Fig. 2).
In Fig. 3 the structure of the complete grid distortion procedure is shown. It
works blockwise in the defined order and also takes parallelization into account. The
procedure starts with the distortion of the edges of the moving block. The second
step is the distortion of the block faces. Finally, the new coordinates of the faces
are used as boundary conditions for the distortion of the internal grid. We briefly
describe the components of the procedure. More details can be found in Yigit [33].
Let us consider a block with nodal coordinates xi;j;k D .xi;j;k ; yi;j;k ; zi;j;k / with
i D 1; : : : ; Ni , j D 1; : : : ; Nj , and k D 1; : : : ; Nk . A simple approach for the dis-
tortion of an edge is linear interpolation. For instance, for the edge along i -direction
with j D k D 1, when the end points x1;1;1 and xNi ;1;1 are already distorted the
linear interpolation for the internal points reads

xi;1;1 D ki .xNi ;1;1  x1;1;1 / C x1;1;1; (13)

where

1 X X
i Ni
ki D kxm;1;1  xm1;1;1 k; LD kxi;1;1  xi 1;1;1 k: (14)
L mD2
i D2

A cubic spline interpolation is employed, which preserves the angle between the
edge and the connecting face. We define

a1 D 3x1;1;1 C 3xNi ;1;1  2T1 C T2 ; (15)

a2 D 2x1;1;1  2xNi ;1;1 C T1  T2 (16)
Simulation and Optimization of Fluid-Structure Interaction 137

Fig. 3 Flow chart of the grid distortion routine.

and distort the internal points according to

xi;1;1 D x1;1;1 C ki3 a2 C ki2 a1 C ki T1 : (17)

T1 and T2 are the direction vectors at the end points of the edge which are chosen
such that the angles between the edge and connecting face does not change.
For the distortion of both, faces and blocks three different approaches are
employed. The methods for the faces are two-dimensional variations of the (three-
dimensional) methods used for blocks. The simplest technique is the linear inter-
polation, where Eq. (13) is used along each of the parametrical directions. Another
138 M. Schäfer et al.

algebraic approach is the transfinite interpolation (TFI). We consider a mapping

between the physical domain with coordinates x and the unit cube as computational
domain with coordinates .; ; /. The mapping function according to transfinite
interpolation can be written as

x ./ D U1 C U2 C U3  U1  U2  U2  U3  U3  U1 C U1  U2  U3 ; (18)

with

U1 D .1  /x.0; ; / C .1 C /x.1; ; /;

U2 D .1  /x.; 0; / C .1 C /x.; 1; /; (19)
U3 D .1  /x.; ; 0/ C .1 C /x.; ; 1/;

where the dot denotes the binary product, e.g.,

U1  U2 D .1  /.1  /x.0; 0; / C .1 C /.1  /x.1; 0; / (20)

C .1  /.1 C /x.0; 1; / C .1 C /.1 C /x.1; 1; /:

The third technique implemented is an elliptic grid generation method on the

minimal surface defined by the surface bounded by six bounding faces (four edges
in the two-dimensional variant for the faces) with zero mean curvature. Follow-
ing Spekreijse [26] we define the unit cube as a parameter space with Cartesian
coordinates .s; t; u/, where s, t, and u are the normalized arclengths along the four
corresponding edges with the same direction. The mapping between computational
and parameter spaces .; ; / ! .s; t; u/ is defined as

s D s1 ./.1  t/.1  u/ C s2 ./t.1  u/ C s3 ./.1  t/u C s4 ./tu;

t D t1 ./.1  s/.1  u/ C t2 ./s.1  u/ C t3 ./.1  s/u C t4 ./su; (21)
u D u1 ./.1  s/.1  t/ C u2 ./s.1  t/ C u3 ./.1  s/t C u4 ./st;

where si , ti , and ui (i D 1; : : : ; 4) are the grid point distributions along the edges.
According to this bilinear transformation the new coordinates of the interior nodes
of the 6 faces are computed and used as a boundary conditions. The edges are
computed via a linear interpolation; thus closing the system of equations.
The grid is obtained by solving the following equation:

a11 x  C 2a12 x C a13 x  C a22 x C 2a23 x C a33 x 

C.a11 P11
1
C 2a12 P12
1
C 2a13 P13
1
C a22 P22
1
C 2a23 P23
1
C 2a33 P33
1
/x
C.a11 P11
2
C 2a12 P12
2
C 2a13 P13
2
C a22 P22
2
C 2a23 P23
2
C 2a33 P33
2
/x (22)
C.a 11 3
P11 C 2a 12 3
P12 C 2a 13 3
P13 Ca 22 3
P22 C 2a 23 3
P23 C 2a 33 3
P33 /x D 0;
Simulation and Optimization of Fluid-Structure Interaction 139

with the control functions

0 1 01 0 1
s  s s 
P11 D T 1 @ t  A ; P12 D T 1 @ t A ; P13 D T 1 @ t  A ;
u  u u 
0 1 0 1 0 1
s s s 
P22 D T 1 @ t A ; P23 D T 1 @ t A ; P33 D T 1 @ t  A ; (23)
u u u 

where 0 1
s s s
T D @ t t t ; A (24)
u u u 
and aij are the contravariant metric base vectors of the mapping .; ; / ! x.
The Poisson problem (22) is discretized by a central difference scheme and the
resulting discrete system is linearized by a Picard iteration scheme. The resulting
linear system is solved by the Gauß-Seidel method in each iteration.
All grid movement techniques described for edges, faces, and blocks can be
combined flexibly via an user input file. As an example the corresponding combina-
tion of techniques for the configuration of the FSI reference experiment (see [10])
is shown in Fig. 4. We achieve fixed wall distances by definining parallel edge
movements, preserved angles near the rear of the structure, transfinite interpolation,
elliptically smoothed grids, and fixed grids in not affected blocks.
Due to the employment of these techniques reliable simulations involving turbu-
lent flows and large deformations are possible (that would fail if just simple grid
movement were employed). As an example, in Fig. 5 results within the flow domain
of a large eddy simulation of the turbulent FSI reference experiment are shown.

Fig. 4 (a) Configuration for turbulent FSI test case. (b) Grid of the turbulent FSI test case.
140 M. Schäfer et al.

Fig. 5 Snapshot of isolines of the mean velocity for turbulent FSI reference experiment obtained
with LES.

Further examples and systematic investigations for the grid movement techniques
can be found in [34].

5 Acceleration Techniques

In this section some investigations concerning the possibilities of accelerating

partitioned FSI computations are summarized. Further results can be found in [27].

5.1 Multigrid Method

In order to reduce the computational effort multigrid methods can be invoked into
the coupled solution procedure. This can be done at the subproblem level by solving
the fluid and/or solid subproblem with a multigrid technique. Here, we employ such
a kind of partitioned multigrid method for the fluid subproblem (see [32]).
Concerning the multigrid components a standard geometric approach is
employed, i.e., a full approximation storage (FAS) scheme for the treatment of non-
linearities, a V-cycle strategy for the movement through the grid levels, and bilinear
interpolation for prolongation and restriction. Details can be found, for instance,
in [11].
Simulation and Optimization of Fluid-Structure Interaction 141

Fig. 6 Configuration of driven lid

cavity test case.
uin(t) outlet

0.5 L 2L

y
x L
z
L
flexible
membrane L

As test case for the partitioned multigrid method a lid driven cavity with a flexible
membrane at the bottom is considered (see Fig. 6). The side length of the cube is
L D 1 m. The membrane thickness is tmem D 0:1 m. The membrane is fixed at the
edges and the lid moves with a time dependent lid velocity vlid defined by

vlid .t/ D 0:5 .1:5  cos .2 t=T0 // (25)

with period T0 D 5 s. The inlet and outlet have 10 % of the total height hin D hout D
0:1 m and are situated in the upper part on the left and right sides, respectively. The
inflow velocity corresponds to the lid velocity. At the outlet a zero gradient condition
for the velocity is used.
The material parameters for the structure and the fluid are defined as:
 Youngs modulus Es D 50000 N=m2
 Poisson ratio
s D 0:3
 Structure density s D 100 kg=m3
 Fluid density f D 100 kg=m3
 Fluid dynamic viscosity f D 0:01 kg=.ms/
 static pressure pstatic D 0:1 Pa
Es and
s are related to the Lamé constants in Eq. (7) by

Es
s Es
s D and s D : (26)
.1 C
s /.1  2
s / 2.1 C
/

For the following investigation, three successive refined fluid grids with 2496,
19968, and 159744 control volumes, respectively, are considered. The multigrid
computations are performed with three grid levels for each grid set. The structure
discretization is fixed for all computations, i.e., linear solid brick elements are uti-
lized with a discretization of 20  2  20 for the x-, y-, and z-directions. In each test
case a simulation time of 1.5 s from a predefined flow field is computed. The com-
puter employed for all computations was a Intel Pentium4 PC with a clock rate of
2539 MHz. The computing times for the single grid and partitioned multigrid com-
putations are indicated in Table 1. One can observe a typical multigrid performance
with increasing acceleration on an increasing problem size.
142 M. Schäfer et al.

Table 1 Computing times [s] Control volumes 2496 19968 159744

for single grid and partitioned
multigrid methods for Single grid 2278 15294 260483
different fluid grid sizes. Partitioned multigrid 1202 1943 21445
Acceleration 1.9 7.9 12.1

5.2 Structural Underrelaxation

The partitioned coupling scheme might be rather sensitive with respect to the defor-
mations, in particular, within the first FSI iterations. This may lead to instabilities
or even the divergence of the iteration process. In order to counteract this effect
an underrelaxation is employed. The actually computed displacements uact are (lin-
early) weighted with the values uold
i from the preceding iteration to achieve the new
displacements unew
i :
unew D ˛FSI uact C .1  ˛FSI /uold ;
where 0 < ˛FSI  1 is the underrelaxation factor.
m
For the adaptive determination of ˛FSI different methods are known. We uti-
lize an approach based on the Aitken method, which is an extrapolation approach
frequently applied in the context of Newton-Raphson iterations. The basis of this
approach was proposed by Aitken [3] and later improved by Irons and Tuck [12].
It was identified as very efficient for computations in the field of fluid-structure
interaction by Mok [16].
Employing the values from two preceding iterations the Aitken factor  m is
extrapolated by:
 m1 T
m m1 m1 u  um  um
 D C .  1/ (27)
.um1  um /2

with um1 D um2  uQ m1 and um D um1  uQ m . The actual underrelaxation
m
factor ˛FSI is defined by
˛FSI D 1   m : (28)
As the first Aitken factor in each time step  0 the last one from the preceding time
step can be taken. For the first time step  0 D 0 can be chosen.
As test case for the performance of the adaptive underrelaxation the lid driven
cavity problem (see previous section) is considered. The results for the single grid
and partitioned multigrid methods are indicated in Table 2. The average numbers
of FSI iterations per time step hardly vary between the single grid and multigrid
methods. In both cases the number is nearly halved by using the adaptive Aitken
method. The computation using multigrid and adaptive underrelaxation is more than
14 times faster than the single grid computation with fixed underrelaxation.
Simulation and Optimization of Fluid-Structure Interaction 143

Table 2 Average number of Single grid Multigrid

FSI iterations for single grid
and multigrid and ˛FSI D 0:8 fixed 15294 1943
acceleration with fixed and ˛FSI adaptive 8550 1057
adaptive underrelaxation. Acceleration 1.8 1.8

Fig. 7 (a) Configuration for turbulent FSI test case. (b) Turbulent kinetic energy and deformation
for turbulent FSI test case.

5.3 Displacement Prediction

A straightforward and simple approach to reduce the number of FSI iterations is the
extrapolation of the starting value for the structural displacement at the beginning
of a new time step based on values from previous time steps. Employing two time
steps, the extrapolation (for equidistant time step size t) reads:

t n
unC1 D un C .u  un1 / : (29)
2
In particular, the extrapolation is useful for large deformations and/or large time step
sizes.
To illustrate possible acceleration effects we consider the FSI of a turbulent flow
around a structure plate with one rotational degree of freedom and an elastic mem-
brane (see Fig. 7). The Reynolds number based on the total length of the structure
is ReD170 000. 640 hexaeder elements with enhanced strain formulation discretize
the structural domain and 25 000 control volumes discretize the fluid domain. For
turbulence modelling the k-" model with wall functions is employed. Details of
the configuration can be found in [33]. Figure 7 shows a snapshot of the turbulent
kinetic energy indicating the large structural movement.
Figure 8 shows the convergence of the FSI iterations within one time step with
and without displacement prediction. An acceleration factor of more than 5 can be
achieved in this case.
144 M. Schäfer et al.

Fig. 8 Convergence of FSI iterations with and without displacement prediction.

6 Optimization

This chapter presents a shape optimization approach for fluid-structure interaction

problems. Mathematically a general optimization problem can be written as follows:

min J .' .a/ ; a/ subject to F .' .a/ ; a/ D 0 (30)

with the objective functional J and the constraints F . The state variables ' 2 RN
are internal values that can not be manipulated directly. They depend on the design
variables a 2 RM , which can be controlled from the outside. Often, additional
side constraints for the control variables have to be fulfilled:

il  ai  iu for i D 1; : : : ; M (31)

with constants il and iu .

In case of FSI problems the state variables can be flow and structure proper-
ties like velocity, pressure, temperature, turbulent kinetic energy, deformation, et
cetera. For shape optimization the control variables are geometrical parameters, for
instance, the coordinates of the control points of NURBS surfaces (see Sect. 6.2.1).
An objective functional can be, for example, the minimization of pressure drop or
structural deformation. The constraints are the partial differential equations describ-
ing the flow and structural problem. Typical side constraints are, e.g., restrictions to
the inlet velocity to prevent transition to turbulence or limitations to the maximal
deformation due to geometrical reasons.
In Fig. 9 a schematic view of a shape optimization for FSI problems is indicated.
In the following we first give a description of the chosen NURBS surface-based
approach for shape representation. Then, the theoretical background of the utilized
Simulation and Optimization of Fluid-Structure Interaction 145

Improving a system’s
functionality and efficiency

Shape optimization
• Minimize objective function J
using design parameters a to
alter the body’s shape
• min J(a) = F (Js (a),Jf (a))
a∈Rn
Update a Update a
Jf (a) Js (a)

Computational structural
Computational fluid dynamics F mechanics

• Computation of fluid forces F • Computation of structural

displacements δ(F )
• Computation of the state
variables J f (a) δ • Computation of the state
variables Js(a)

Fig. 9 Schematic shape optimization of a fluid-structure interaction problem.

optimization method is outlined. Finally, results for a test problem involving shape
optimization for a steady fluid-stucture interaction are presented.

6.1 Surface Representation

For the representation of surfaces NURBS (e.g., [20]) are employed. NURBS sur-
faces are a powerful technique to change the grid point position and therefore the
shape of the objects by adjusting the control points. Compared to the amount of grid
points the amount of control points is very small. NURBS are used in two ways
within this work. On the one hand to calculate an approximation of given surface
grid points, based on a predefined boundary condition enabling the easy definition of
an optimization surface. On the other hand during the optimization the coordinates
146 M. Schäfer et al.

of the NURBS control points serve as design parameters. We briefly outline the basic
concept. A detailed description of the whole implemented shape altering routine can
be found in [4, 5, 25].
A NURBS surface is a B-spline surface with a nonuniform knot vector [29],
which is defined as follows:

P
nO P
O
m
wi;j Pi;j Nip .u/Njq .v/
i D1 j D1
S.u; v/ D ; u; v 2 Œ0; 1 : (32)
P
nO P
O
m
wr;s Nrp .u/Nsq .v/
rD1 sD1

w D .wi;j /i;j denotes the weights. ;  denote nonuniform knotvectors (their

derivation is presented in Sect. 6.2.1). N p ; N q W Œ0; 1 ! R; with Nip WD NŒpi ;i C1 
(Njq using  analogously) denote the B-Spline basis functions of orders p; q com-
puted recursivley with the Cox-deBor recurrence (see [7]). Note, that we follow the
convention of Farin et al. [8]: 0=0 WD 0. Pi;j 2 R3 denote the control points.
One of the main advantages using NURBS surfaces is the possibilty to compute
them based on a given set of surface grid points X 2 Rn Rm R3 . This allows to fit
the NURBS surface onto the optimization boundary automatically (i.e. computing
the control points P) based on given X and additionally the orders p; q and number
O m.
of control points n; O
To simplify matters and according to Piegl and Tiller [20] very little has been
published on setting the weights in a fitting process, therefore we set w D 1 in this
paper. This leads us, based on Eq. (32), to a simpler system of equations

X O
nO X
m
S.uk ; vl / D Nip .uk /Njq .vl /Pi;j ; (33)
i D1 j D1

where uk ; vl denotes the parameterized points of X.

To finish the data fitting process Eq. (33) needs to be solved. Since there are
notably less control points compared to surface grid points, we need to solve an
overdetermined linear system of the form

Ax D b; with A 2 Rm  Rn ; x 2 Rn ; b 2 Rm ; m >> n: (34)

According to [19] the corresponding linear regression problem can be defined as

follows:
kAx  bk2 D minn kAx  bk2 : (35)
x2R

A suitable method to solve this problem is the QR-decomposition with Householder

reflection, which offers a stable algorithm. A detailed overview about this algorithm
is given in Becker et al. [4]. According to Dahmen and Reusken [6] an important
condition to achieve a unique solution is that A has a full rank, i.e., rank.A/ D n.
Schoenberg and Whitney [24] stated that this is always fulfilled if at least one of n
Simulation and Optimization of Fluid-Structure Interaction 147

parameterized grid points uk exists between every pair of subsequent knots, i.e.,

i < uk < i C1 ; i D 1; : : : ; nO C p; k D 1; : : : ; n; (36)

and analogously for vl . A more detailed description of the computation of the

nonuniform knot vectors considering the condition (36) is given later in Sect. 6.2.1.

6.2 Shape Optimization in an FSI framework

To implement a shape optimization within our FSI framework, three basic compo-
nents are needed: an initialization routine, an optimizer, and an evaluation routine
(see Fig. 10).

6.2.1 Initialization

The first step in the shape optimization procedure is the identification of the opti-
mization surface and the automatic approximation with a NURBS surface. This is
done by marking the surface during the meshing. This set of points is the input for
the NURBS surface fitting procedure as shown in Fig. 11.
To set up the system of equations (33) the surface grid points on the designated
optimization surface must be identified. Their coordinates in R3 are stored in the
array X. Although when the grid is blockstructured, as it is in FASTEST, the array

Fig. 10 Implementation of a shape optimization within our FSI framework.

Fig. 11 Initialization routine of the optimization framework.

148 M. Schäfer et al.

X can be build across block boundaries and regardless of the orientation of the local
coordinate system of each block. The next step is the construction of a well chosen
parameterization of X to derive u D .u1 ; : : : ; un / and v D .v1 ; : : : ; vm /. For this we
use an approach based on an algorithm of Ma and Kruth [14], where, for instance,
u (similar for v) is defined as

X
m
u1 D 0 ; uk D uO k;j ; k 2 f2; : : : ; ng; (37)
j D1

with

X
k1
.jjXi C1;l  Xi;l jj/e
i D1
uO 1;l D 0 ; uO k;l D ; k 2 f2; : : : ; ng; l 2 f1; : : : ; mg:
X
n1
.jjXi C1;l  Xi;l jj/ e

i D1
(38)
The parameter e 2 Œ0; 1 , introduced by Jung and Kim [13], offers the possibilty
to influence the type of parameterization. For e D 0, a uniform parameterization,
for e D 1, a chordlength parameterization and for e D 0:5, a centripedal parame-
terization is performed. However, e can be chosen within Œ0; 1 , dependent on the
setting.
In order to calculate the knot vectors, several boundary conditions have to be con-
sidered. To obtain a unique set of control points, it is necessary that A has full rank.
Hence, at least one grid point needs to be located between every pair of subsequent
knots (Eq. (36)). Furthermore, the grid density should be considered and the cor-
ner points of the NURBS surface should coincide with the corresponding control
points. For this the following algorithm is employed. Let us consider the discrete
parameters uk ; k 2 f1; : : : ; ng and vl ; l 2 f1; : : : ; mg as a data basis to compute
the knot vectors ;  using

 D .0; : : : ; 0; pC1 ; : : : ; nO ; 1; : : : ; 1/;

„ ƒ‚ … „ ƒ‚ …
p p

 D .0; : : : ; 0;
qC1 ; : : : ;
mO ; 1; : : : ; 1/; (39)
„ ƒ‚ … „ ƒ‚ …
q q

with lengths p C nO and q C m,O respectively. Recall that n; O m

O denote the numbers of
control points and p; q 2 N are the orders of the B-spline basis functions in each
boundary direction. Additional conditions are n > nO  p C 1 and m > m O  q C 1, to
assure having a grid point between two subsequent knots. As we can see in Eq. (39),
we need to calulate .nO  p/ entries for the knot vector  ( analogously). Therefore,
we split all entries of u D .u1 ; : : : ; uk ; : : : ; un / into .nO  p/ intervals. The .nO  p/
knot vector entries equal the mean average build within each of the .np/ O intervals.
Simulation and Optimization of Fluid-Structure Interaction 149

With known knot vectors  and , as well as all discrete grid points uk ; k D
1; : : : ; n and vl ; l D 1; : : : ; m we are able to calculate the B-spline basis functions
by the Cox-deBor recurrence (see [7]) and store them into the following matrices:

Np 2 RnO  Rn ; in u-direction and (40)

O
N 2 R  R ; in v-direction.
q m m
(41)

Now we can solve the system of equations (33) as described in Sect. 6.1 and find
the initial positions of the control points P.
To finish the preparation phase a relation between the set of control points P.g/ ,
where P.g/i;j 2 R denotes the .i; j /-th entry at the g-th evaluation call, and the
3

following set of design variables has to be established:

a.g/ 2 RM ; subject to l  a.g/  u ;  2 f1; : : : ; M g; and a D .a / ; (42)

where M 2 N denotes the number of design variables. The initial set of design
variables is always defined as a.0/ D 0. The control points of the g-th evaluation
call are derived as follows:
0 .g/ 1
˛i;j
.g/ B .g/ C .g1/
Pi;j D @ˇi;j A C Pi;j : (43)
.g/
i;j

For each pair .i; j / we distinguish four cases for the usage of the control points:
 Case I: Fixed control points:

.g/ .g/ .g/

˛i;j D ˇi;j D i;j D 0: (44)

 Case II: One design variable per control point utilizing a direction vector: s D
.si /i 2f1;2;3g

.g/ .g/ .g/

˛i;j D s1 a.g/ ; ˇi;j D s2 a.g/ ; i;j D s3 a.g/ ; ks k2 D 1: (45)

 Case III: Three design variables per control point in each spatial direction:

.g/ .g/ .g/ .g/ .g/

˛i;j D a.g/ ; ˇi;j D aC1 ; i;j D aC2 : (46)

 Case IV: The control point depends on another control point’s design variable
with a scaling factor t D .ti /i 2f1;2;3g :

.g/ .g/ .g/ .g/ .g/ .g/

˛i;j D t1 ˛k;l ; ˇi;j D t2 ˇk;l ; i;j D t3 k;l ; .i; j / ¤ .k; l/: (47)

This is an important feature to reduce the number of design variables, e.g., in

symmetric cases.
150 M. Schäfer et al.

The above definitions allow to connect the design variables intuitively with the con-
trol points. In this approach, the design variable’s value a.g/ 2 Œl ; u represents
the displacement along either a direction vector (case II) or a single direction in
space (case III).

6.2.2 Optimizer

As optimizer the SIMPLEX algorithm according to Nelder and Mead [17] is cho-
sen, which is a derivative-free optimization method for unconstrained minimization
problems. The method is widespread due to the fact that it makes no assump-
tions about the objective functional except that it is continuous. Furthermore, it is
quite numerically robust. Figure 12 shows the basic procedure of the algorithm.
The general idea of the algorithm is to compare the objective functional values
E.a.1/ /; : : : ; E.a.M C1/ / at the .M C 1/ vertices of an initial general simplex in the
M -dimensional space of design parameters and to move it towards a local minimal
point.
The following notations are used:
 a.l/ is the set of design parameters resulting in the lowest evaluation value E.a.l/ /
 a.h/ is the set of design parameters resulting in the highest evaluation value
E.a.h/ /
 a.s/ is the set of design parameters resulting in the second highest evaluation
value E.a.s/ /
 a.t / is the set of design parameters at the centroid of all sets without the highest
1 PM C1
and is derived with a.t / D M i D1;i ¤h a
.i /

 a is the reflected set of design parameters at the centroid a.t / and is derived
.t /

with a.r/ D a.t /  .a.h/  a.t / /

 a.e/ is the expanded set of design parameters at the reflected set a.r/ and is
derived with a.e/ D a.r/ C .a.r/  a.t / /
 a.c/ is the contracted set of design parameters and is derived with a.c/ D a.t / C
0; 5.a.h/  a.t / /
Three basic operations (reflection, contraction, and expansion) move the simplex
by substituting the vertex with the highest objective functional value by a point with
a lower value. Therefore, the centroid of the simplex is calculated and its functional
value is compared with the other values at the vertices of the simplex. Depending
on the centroid value the vertex with the highest functional value is exchanged, or
alternatively the complete simplex becomes shrinked around the vertex with the
lowest value with
a.i / D a.l/ C 0; 5.a.i /  a.l/ /; 8i: (48)
The method represents a robust algorithm for unconstrained optimization prob-
lems. However, it is not able to handle side constraints, since it is possible that the
simplex collapses and degenerates due to the boundaries. Afterwards the simplex is
caught in a subspace and will fail to find the optimum within the complete parameter
Simulation and Optimization of Fluid-Structure Interaction 151

a(0)
Initialization
Compute E(a(1)),
...,E(a(M+1))

Centroid E (r(r )) ≥ No Replace a(h) with

A
a(t) E(a(h))? a(r )

Reflection Contraction
No Yes
Compute a(r) and Compute a(c) and
E(a(r)) E (a(c))

E(a(r)) < No E(a(r )) < E (a(c)) ≥

E(a(l ))? E(a(s))? E(a(h))?
No
Yes Yes
Expansion Shrinking
Replace a(h) with
Compute a(e) and Yes Compute a(i ), ∀i
E(a(e)) a(c)
and E (a(i )),∀i,i ≠ l

E (a(e)) < No Replace a(h) with A

E(a(l ))? a(r )

No
Yes

Replace a(h) with Conv.?

a(e)

Yes

Fig. 12 SIMPLEX algorithm (Dashed boxes includes the call of the evaluation routine).

space. Thus, for constrained optimization it is necessary to ensure that the simplex
does not collapse. This is done by transforming the constrained optimization prob-
lem into an unconstrained problem by introducing penalty functions. The penalty
function can be defined as

X
M
P D .max.ai  iu ; 0/ C max.il  ai ; 0//: (49)
i D1
152 M. Schäfer et al.

where iu and il are the upper and lower bounds of the i -th design parameter ai and
 is a scaling factor. We set this factor usually to a value of  D 1000.

6.2.3 Evaluation

During the processing of the evaluation routine (illustrated in Fig. 13) a new set of
control points P.g/ 2 RnO  RmO  R3 is calculated with Eq. (43) using a new set
of design variables a.g/ as input. The B-spline basis functions already calculated
during the initialization (see Sect. 6.2.1) can be reused in this routine. Now, a new
set of surface grid points X.g/ 2 Rn  Rm  R3 can be calculated using the NURBS
.g/
equation (33), where Xi;j;k denotes the .i; j; k/th entry of the tensor X.g/ .
Note that this method generates the new surface grid X.g/ only. To move the
remaining computation grid within the flow domain according to the displaced
optimization surface, grid generation techniques already used for FSI grid displace-
ments are applied. Within the FSI approach introduced in Sect. 3, the FSI grid
generation routine receives computed displacements of the coupling surface from
the structural solver FEAP [21]. To utilize this routine for the purpose of shape

a(g) E (a(g))

Compute P(g) using

Compute E (a(g) )
a(g)

Yes
P(g)

Compute a new FSI

NURBS surface Conv.?
No
X(g) Force

Compute
Di,j,k ,∀i,j,k Solve flow MPCCI

D(g) Displ. Force Displ.

Grid distortion Grid distortion Solve structure

Fig. 13 Evaluation routine containing a steady FSI.

Simulation and Optimization of Fluid-Structure Interaction 153

optimization we need to compute the displacements between X.g/ and the original
optimization surface X0 , i.e.

.g/ 0
i;j;k D jXi;j;k  Xi;j;k j; 8 i; j; k: (50)

The grid generation routine uses  D .i;j;k /i;j;k to generate new grids within
the optimization surface’s neighboring blocks. This routine is explained in detail in
Sect. 3.
Now the flow variables like velocity, pressure or temperature are solved and
the forces at the FSI coupling interface are computed. These forces are send via
the interpolation tool MpCCI to the structural solver, which computes based on the
forces a structural displacment. This displacment serves as input to the grid distor-
tion routine as described in Sect. 4. This computes us a new flow domain which can
be solved again until a equlibrium between force and displacment is reached. Finally
the evaluation value E.a.g/ / is calculated an serves as input for the optimizer.

6.3 Optimization example

As test case we consider the shape optimization of a 2D steady FSI problem based on
the benchmark configuration of Turek and Hron [31] with altered outer boundaries
as shown in Fig. 14. The velocity at the inlet is defined as

y.H  y/
Ui n D 1:5U ; (51)
. H2 /2

where U D 0:1 m/s denotes the characteristic average velocity in x-direction and
H denotes the channel height. The density of the fluid is f D 1000 kg=m3 and
its dynamic viscosity is f D 1 kg/(ms). Thus, the Reynolds number is Re D 10
based on the cylinder diameter. The fluid domain is discretized by 50000 hexahedral
control volumes in the xy-plane.

(a) (b)

Fig. 14 (a) Setup of the steady FSI optimization example. (b) Relation between CP and DV.
154 M. Schäfer et al.

The structural density is set to s D 1000 kg=m3, the Youngs-Modulus Es D

5000 N=mm2 and the Possions Ratio to
s D 0:4. The structural domain has the
dimensions as set in the paper of Turek and Hron [31] and is disretized with 120
finite brick elements in the xy-plane.
The rigid cylinder is the shape optimization body and the objective is the mini-
mization of the displacement of point A (cf. Fig. 14). Figure 14 shows the NURBS
surface representing the optimization surface. The NURBS surface of order four
consists of seven control points (roman numbers in Fig. 14) in the xy-plane. The
control points I and V II are fixed at the end points, III to V are directly attached
to three design variables ai Df1;2;3g ; il  a  iu with
0 1
0
D @1A a3
.gC1/ .gC1/ .g/
PIII C PIII ;  0:05  a3  0:05 (52)
0
0 1
1
D @ 0 A a2
.gC1/ .gC1/ .g/
PI V C PI V ;  0:04  a2  0:05 (53)
0
0 1
0
D @C1A a1
.gC1/ .gC1/ .g/
PV C PV ;  0:05  a1  0:05: (54)
0

Control points II and V I are dependent with following relation

0 1
0
D @0:3A a3
.gC1/ .gC1/ .g/
PII C PII (55)
0
0 1
0
D @C0:3A a1
.gC1/ .gC1/ .g/
PV I C PV I : (56)
0

At first the results of the steady FSI simulation of the non-optimized cylinder are
presented. Figure 15a and 15b show the development of the displacement of point A
in x- and y-direction at the end of the elastic beam as well as the drag and lift force
during the FSI iterations. After 30 fluid-structure coupling iterations a steady state
is reached. At the converged solution the beam has experienced a final displacement
at point A of 0.04216 m (equal to 5.2% of channel height H ).
Next the optimization is considered. For each evaluation call of the SIMPLEX
algorithm a complete steady fluid-structure coupling routine needs to be executed.
The SIMPLEX required 115 evaluations to achieve the optimal result within 63
SIMPLEX loops. Thus, the algorithm needed an average of nearly two evaluations
per SIMPLEX loop (recall Fig. 12). Figure 16 shows the progress of the objective
function value for each SIMPLEX loop. The displacement was reduced by 29 % to
a final value of 0.02981 m (equal to 3.7% of channel height H )
Simulation and Optimization of Fluid-Structure Interaction 155

0.11
0.045 0.1
Displacement A [mm]

0.04 0.09
0.035

Force [N]
0.03 0.08
0.025 0.07
0.02
0.015 0.06
0.01 0.05
0.005
0 0.04
– 0.005 0.03
– 0.01 5 10 15 20 25
5 10 15 20 25
FSI coupling iterations FSI coupling iterations
x-direction y-direction Drag Lift
(a) (b)

Fig. 15 (a) Displacement of point A in x- and y-direction. (b) Drag and lift force onto the elastic
beam.
Objective functional value J(a)

0.044
0.042
0.04
0.038
0.036
0.034
0.032
0.03
0.028
0 5 10 15 20 25 30 35 40 45 50 55 60
Optimization runs

Fig. 16 Development of the objective function value during the optimization.

Fig. 17 Initial shape of the

cylinder and its final
displacement of the elastic
beam (gray) as well as the
final shape and displacement
(black). Please note that the
initial and deformed
blockstructure is shown as
well.

The final shape of the cylinder is illustrated in Fig. 17. The upper half of the
cylinder got flattend; this increases the velocity above the elastic beam. The lower
half got larger to decrease the velocity. The result is a smaller displacement of the
elastic beam.
To support the above mentioned statements concerning the optimal result,
Fig. 18a and Fig. 18b show the relative pressure distribution of the inital and final
156 M. Schäfer et al.

Fig. 18 (a) Relative pressure distribution of the initial configuration. (b) Relative pressure
distribution of the final configuration.

configuration. One can easily see the reduction of the pressure below the elastic
beam and the increase above.

7 Conclusions

A partitioned solution approach for the simulation and optimization of FSI problems
has been presented. Special emphasis has been given to the grid moving techniques
for which algebraic and elliptic approaches as well as special treatments of edges
and faces have been considered. A proper combination of these methods ensures
a high grid quality also for large deformations of the structure which, in partic-
ular, is of high importance for reliably simulating FSI involving turbulent flows.
The possibilities for accelerating the computations by the usage of multigrid meth-
ods, adaptive underrelaxation, and displacement prediction have been discussed.
The results indicate the high potential of such approaches. Finally, a concept for
integrating the FSI solver into an optimization procedure for FSI problems has been
presented which is based on NURBS surface representation and an unified treatment
of grid movement and shape variation. Results for an exemplary test case illustrate
the capabilities of the integrated optimization procedure.

Acknowledgements The financial support of the work by the Deutsche Forschungsgemeinschaft

within the Research Unit 493 Fluid-Structure Interaction: Modelling, Simulation, Optimization is
gratefully acknowledged. Thanks are addressed to Stephen Sachs and Michael Werner (grid move-
ment techniques), Johannes Siegmann (NURBS implementation), and Julian Michaelis (SIMPLEX
implementation) for their valuable contributions.

References

1. FASTEST – User Manual. Institute of Numerical Methods in Mechanical Engineering,

Technische Universität Darmstadt (2004)
2. MpCCI – Mesh-Based Parallel Code Coupling Interface. User Guide V2.0. Fraunhofer, SCAI
(2004)
3. Aitken, C.: Studies in practical mathematics. II. the evaluation of the latent roots and latent
vectors of a matrix. Royal Societiy of Edinburgh 57, 269–304 (1937)
Simulation and Optimization of Fluid-Structure Interaction 157

4. Becker, G., Falk, U., Schäfer, M.: Shape optimization with higher-order surfaces in considera-
tion of fluid-structure interaction. In: S. Hartmann, A. Meister, M. Schäfer, S. Turek (eds.)
International Workshop on Fluid-Structure Interaction: Theory, Numerics and Application.
Kassel Univ. Press, Kassel (2009)
5. Becker, G., Siegmann, J., Michaelis, J., Schäfer, M.: Comparison of a derivative-free and a
derivative-based shape optimization method. In: H.C. Rodrigues, J.M. Guedes, P.R. Fernandes,
M.M. Neves (eds.) Proceedings of the 8th World Congress on Structural and Multidisciplinary
Optimization, Lisbon, Portugal (2009)
6. Dahmen, W., Reusken, A.: Numerik für Ingenieure und Naturwissenschaftler. Springer-Verlag
Berlin Heidelberg, Berlin, Heidelberg (2008)
7. De Boor, C.: A practical guide to splines. Springer (2001)
8. Farin, G.E., Hoschek, J., Kim, M.S.: Handbook of computer aided geometric design. Elsevier,
Amsterdam (2002)
9. Ferziger, J., Perić, M.: Computational Methods for Fluid Dynamics. Springer, Berlin (1996)
10. Gomes, J., Lienhart, H.: Experimental study on a fluid-structure interaction reference test
case. In: H.J. Bungartz, M. Schäfer (eds.) Fluid-Structure Interaction - Modelling, Simulation,
Optimisation, LNCSE, vol. 53, pp. 356–370. Springer (2006)
11. Hackbusch, W.: Multi-Grid Methods and Applications. Springer-Verlag, Berlin (1985)
12. Irons, B., Tuck, R.: A version of the aitken accelerator for computer iteration. International
Journal of Numerical Methods in Engineering 1, 275–27 (1969)
13. Jung, H., Kim, K.: A new parameterisation method for NURBS surface interpolation. The
International Journal of Advanced Manufacturing Technology 16(11), 784–790 (2000)
14. Ma, W., Kruth, J.P.: NURBS curve and surface fitting for reverse engineering. The International
Journal of Advanced Manufacturing Technology 14(12), 918–927 (1998)
15. Mohammadi, B., Pironneau, O.: Shape optimization in fluid mechanics. Annual Review of
Fluid Mechanics 36(1), 255–279 (2004)
16. Mok, D.: Partitionierte Lösungsansätze in der Strukturdynamik und der Fluid-Struktur-
Interaktion. Ph.D. thesis, Institut für Baustatik, Universität Stuttgart (2001)
17. Nelder, J.A., Mead, R.: A simplex method for function minimization. The Computer Journal
7(4), 308–313 (1965)
18. Ogden, R.W.: Non-Linear Elastic Deformations. Dover Publications (1997)
19. Peters, G., Wilkinson, J.H.: The least squares problem and pseudo-inverses. The Computer
Journal 13(3), 309–316 (1970)
20. Piegl, L., Tiller, W.: The NURBS book. Springer Verlag, New York (1996)
21. Schäfer, M., Heck, M., Yigit, S.: An implicit partitoned method for the numerical simulation
of fluid-structure interaction. In: H.J. Bungartz, M. Schäfer (eds.) Fluid-Structure Interaction:
Modelling, Simulation, Optimization, 53, pp. 171–194. Springer, Berlin, Heidelberg (2006)
22. Schäfer, M., Lange, H., Heck, M.: Implicit partitioned fluid-structure interaction coupling. In:
Proceedings of 1er Colloque du GDR Interactions Fluide-Structure, pp. 31–88 (2005)
23. Schäfer, M., Meynen, S., Sieber, R., Teschauer, I.: Efficiency of multigrid methods for the
numerical simulation of coupled fluid-solid problems. Scientific Computing and Applications,
Advances in Computation: Theory and Practice pp. 257–266 (2001)
24. Schoenberg, I.J., Whitney, A.: On Polya frequency function. III. The positivity of translation
determinants with an application to the interpolation problem by spline curves. Transactions
of the American Mathematical Society 74(2), 246–259 (1953)
25. Siegmann, J., Becker, G., Michaelis, J., Schäfer, M.: A general sensitivity based shape opti-
mization approach for arbitrary surfaces. In: H.C. Rodrigues, J.M. Guedes, P.R. Fernandes,
M.M. Neves (eds.) Proceedings of the 8th World Congress on Structural and Multidisciplinary
Optimization, Lisbon, Portugal (2009)
26. Spekreijse, S.: Elliptic grid generation based on laplace equations and algebraic transforma-
tions. Journal of Computational Physics 118, 38–61 (1995)
27. Sternel, D.C., Schäfer, M., Heck, M., Yigit, S.: Efficiency and accuracy of fluid-structure inter-
action simulations using an implicit partitioned approach. Computational Mechanics, 43(1),
103–113 (2008)
158 M. Schäfer et al.

28. Taylor., R.: FEAP – A Finite Element Analysis Programm. Version 7.4 User Manual.
University of California at Berkeley (2002)
29. Tiller, W.: Rational B-splines for curve and surface representation. Computer Graphics and
Applications, IEEE 3(6), 61–69 (1983)
30. Truesdell, C., Noll, W.: The Non-Linear Field Theories of Mechanics, Third Edition. Springer
(2004)
31. Turek, S., Hron, J.: Proposal for numerical benchmarking of fluid-structure interaction between
an elastic object and laminar incompressible flow. In: H.J. Bungartz, M. Schäfer (eds.) Fluid-
Structure Interaction: Modelling, Simulation, Optimization, 53, pp. 371–385. Springer, Berlin,
Heidelberg (2006)
32. Yigit, S., Heck, M., Sternel, D.C., Schäfer, M.: Efficiency of Fluid-Structure Interaction Sim-
ulations with Adaptive Underrelaxation and Multigrid Acceleration. International Journal of
Multiphysics, 1, 85–99 (2007)
33. Yigit, S.: Phänomene der Fluid-Struktur-Wechselwirkung und deren numerische Berechnung.
Ph.D. thesis, Institute of Numerical Methods in Mechanical Engineering, TU Darmstadt (2008)
34. Yigit, S., Schäfer, M., Heck, M.: Grid movement techniques and their influence on laminar fluid
structure interaction computations. Journal of Fluids and Structures, 24(6), 819–832 (2008)
An Adaptive Finite Element Method
for Fluid-Structure Interaction Problems
Based on a Fully Eulerian Formulation

R. Rannacher and T. Richter

Abstract In this article, we continue the investigation of the general variational

framework for the adaptive finite element approximation of fluid-structure interac-
tion problems proposed in Dunne [12] and Dunne & Rannacher [13]. The modeling
is based on an Eulerian description of the (incompressible) fluid as well as the
(elastic) structure dynamics. This approach uses a technique which is similar to
the Level Set method in the simulation of two-phase flows in so far that it also
tracks initial data and from this determines to which “phase” a point belongs. The
advantage is that, in contrast to the common ALE approach, the computation takes
place on fixed, though dynamically adapted, reference meshes what avoids the crit-
ical degeneration in case of large deformation or boundary contact of the structure.
Based on this monolithic model of the fluid-structure interaction, we apply the Dual
Weighted Residual (DWR) Method for goal-oriented a posteriori error estimation
and mesh adaptation to fluid-structure interaction problems. Several test examples
are presented in order to illustrate the potential of this approach.

1 Introduction

Computational fluid mechanics and structure mechanics are two major areas of
numerical simulation of physical systems. With the introduction of high perfor-
mance computing it has become possible to tackle coupled systems with fluid and
structure dynamics. General examples of such fluid-structure interaction (FSI) prob-
lems are flow transporting elastic particles (particulate flow), flow around elastic
structures (e.g., airfoils) and flow in elastic structures (e.g., heart valves). In all

R. Rannacher
Institute of Applied Mathematics, University of Heidelberg
e-mail: [email protected]
T. Richter
Institute of Applied Mathematics, University of Heidelberg
e-mail: [email protected]

H.-J. Bungartz et al. (eds.), Fluid Structure Interaction II, Lecture Notes 159
in Computational Science and Engineering 73, DOI 10.1007/978-3-642-14206-2 7,
c Springer-Verlag Berlin Heidelberg 2010
160 R. Rannacher and T. Richter

these settings the dilemma in modeling the coupled dynamics is that the fluid model
is normally based on an Eulerian perspective in contrast to the usual Lagrangian
description of the solid model. This makes the setup of a common variational for-
mulation difficult. However, such a variational formulation of FSI is needed as the
basis of a consistent approach to residual-based a posteriori error estimation and
mesh adaptation as well as to the solution of optimal control problems by the Euler-
Lagrange method. The subject of the present paper is a novel monolithic finite
element (FE) formulation for fluid-structure interaction problems in a fully Eule-
rian setting. This work is largely based on the doctoral dissertation of Dunne [12]
and the survey articles Dunne & Rannacher [13], Bönisch, Dunne & Rannacher [7]
and Dunne, Rannacher & Richter [14].
To achieve a closed variational formulation for the fluid-structure interaction
problem, usually an auxiliary unknown coordinate transformation function Tf is
introduced for the fluid domain. With its help the fluid problem is rewritten on
the transformed domain. Then, all computations are done on the reference domain
which is fixed in time. As part of the computation, the auxiliary transforma-
tion function Tf has to be determined at each time step. Such, so-called “arbi-
trary Lagrangian-Eulerian” (ALE) methods are described for instance in Dunne,
Rannacher & Richter [14].
We however follow the alternative way of posing the fluid as well as the structure
problem in a fully Eulerian-Eulerian (Eulerian) framework. A similar approach has
been used by Lui & Walkington [22] in the context of the transport of visco-elastic
bodies in a fluid. In the Eulerian setting a phase variable is employed on the fixed
mesh to distinguish between the different phases liquid and solid. This approach
to identifying the fluid-structure interface is generally referred to as “interface
capturing”, a method commonly used in the simulation of multiphase flows.
Our method has similarities to the Level Set (LS) method of Osher & Sethian
[25]. In the classical LS approach the distance function has to continually be reini-
tialized, due to the smearing effect by the convection velocity in the fluid domain.
This makes the use of the LS method delicate for modeling FSI problems particu-
larly in the presence of cornered structures. To cope with this difficulty, in Dunne
[11, 12] a variant of the LS method, the Initial Position (IP) method, has been pro-
posed that makes reinitialization unnecessary and which easily copes with cornered
structures. This approach does not depend on the specific structure model. In the
present paper, we formulate a modification of the original fully Eulerian method
avoiding the additional IP-set, which has been proposed by Richter & Wick [30].
The equations we use are based on the momentum and mass conservation equa-
tions for the flow of an incompressible Newtonian fluid and the deformation of a
compressible St. Venant-Kirchhoff or likewise incompressible neo-Hookean solid.
The spatial discretization is accomplished by a second-order finite element method
with conforming equal-order (bilinear) trial functions using “local projection stabi-
lization” as proposed by Becker & Braack [3, 4]. The time discretization is based
on a Galerkin approach similar to the Crank-Nicolson scheme (see Schmich &
Vexler [29]). This method allows for rigorous error estimation and time step control
in a coupled space-time setting.
Adaptive FE method for FSI 161

Based on the fully Eulerian variational formulation of the FSI system, we use
the “dual weighted residual” (DWR) method, as described in Becker & Rannacher
[5, 6] and Bangerth & Rannacher [1], to derive “goal-oriented” a posteriori error
estimates. The evaluation of these error estimates requires the approximate solution
of a linear dual variational problem. The resulting a posteriori error indicators are
then used for automatic local mesh adaptation. The rigorous application of the DWR
method to FSI problems requires a Galerkin discretization in space as well as in
time, which was one of the motivations for introducing the monolithic fully Eulerian
formulation.
The method for computing FSI described in this paper has been validated at
several stationary model problems, for example a lid-driven cavity involving the
interaction of an incompressible Stokes fluid with a linearized incompressible neo-
Hookean solid (see Dunne & Rannacher [13]). Further, as a more challenging test
the self-induced oscillation of a thin elastic bar immersed in an incompressible fluid
has been treated (FLUSTRUK-A benchmark, Hron & Turek [21]). For this test prob-
lem, our method has also been compared against a standard “arbitrary Lagrange
Eulerian” (ALE) approach. The potential of the fully Eulerian formulation of the
FSI problems is indicated by its good behavior for large structure deformations.
All computations and visualizations have been done using the flow-solver package
GASCOIGNE [16] and the graphics package VISUSIMPLE [32]. The details on the
software implementation can be found in Dunne [11, 12].
The outline of this paper is as follows. Sect. 2 (“Formulation”) introduces the
basic notation for the flow and structure model and also the fully Eulerian for-
mulation of the FSI Problem, which will be used throughout this paper. Sect. 3
(“Discretization”) describes the discretization in space and time as well as the tech-
niques for solving the resulting algebraic systems. Sect. 4 (“Numerical Examples”)
contains the results obtained for different test-cases comprising large deformation
and free movement of the structure in the fluid. For completeness also some of the
results from Dunne & Rannacher [13] are recalled concerning the comparison of
the fully Eulerian and the ALE approach at the FSI benchmark FLUSTRUK-A. The
paper is closed by Sect. 5 (“Summary”).

2 Formulation

In this section, we present the fully Eulerian formulation of the FSI problem. The
Eulerian approach was first proposed in Dunne [11], then simplified by Richter &
Wick [30].

2.1 Notation

We begin with introducing some notation which will be used throughout this paper.
By ˝  Rd (d D 2 or d D 3), we denote the domain of definition of the
162 R. Rannacher and T. Richter

FSI problem. The domain ˝ is supposed to be time independent but to consist of

two possibly time-dependent subdomains, the fluid domain ˝f .t/ and the struc-
ture domain ˝s .t/. Unless needed, the explicit time dependency will be skipped
in this notation. The boundaries of ˝, ˝f , and ˝s are denote by @˝, @˝f ,
and @˝s , respectively. The common interface between ˝f and ˝s is i .t/, or
simply i .
The initial structure domain is denoted by b̋ s . Spaces, domains, coordinates,
values (such as pressure, displacement, velocity) and operators associated to b̋ s (or
b̋ f ) will likewise be indicated by a “hat”.
Partial derivatives of a function f with respect to the i -th coordinate are denoted
by @i f , and the total time-derivative
P by dt f . The
P divergence of a vector and tensor
is written as divf D i @i fi and .divF /i D j @j Fij . The gradient of a vector
valued function v is the tensor with components .rv/ij D @j vi .
By Œf , we denote the jump of a (possibly discontinuous) function f across an
interior boundary, where n is always the unit vector n at points on that boundary.
For a Lebesgue measurable set X , we denote by L2 .X / the Lebesque space
of square-integrable functions on X equipped with the usual inner product and
norm
Z
.f; g/X WD fg dx; kf k2X D .f; f /X ;
X

respectively, and correspondingly for vector- and matrix-valued functions. Mostly

the domain X will be ˝, in which case we will skip the domain index in products
and norms. For ˝f and ˝s , we similarly indicate the associated spaces, products,
and norms by a corresponding index “f” or “s”.
We will generally use roman letters, V , for denoting spaces of functions depend-
ing only on spatial variables and calligraphic letters, V , for spaces of functions
depending additionally on time. Let LX WD L2 .X / and L0X WD L2 .X /=R.
The functions in LX (with X D ˝, X D ˝f .t/, or X D ˝s .t/) with first-
order distributional derivatives in LX make up the Sobolev space H 1 .X /. Further,
H01 .X / D fv 2 H 1 .X / W vj@XD D 0g, where @XD is that part of the boundary @X at
which Dirichlet boundary conditions are imposed. We will use the function spaces
VX WD H 1 .X /d , VX0 WD H01 .X /d , and for time-dependent functions

LX WD L 2 Œ0; T I LX ; VX WD L 2 Œ0; T I VX  \ H 1 Œ0; T I VX ;

LX0 WD L 2 Œ0; T I L0X ; VX0 WD L 2 Œ0; T I VX0  \ H 1 Œ0; T I VX ;

where VX is the dual of VX0 , and L 2 and H 1 indicate the corresponding properties
in time. Again, the X -index will be skipped in the case of X D ˝, and for X D ˝f
and X D ˝s a corresponding index “f” or “s” will be used.
Adaptive FE method for FSI 163

2.2 Fluid

For the liquid part, we assume Newtonian incompressible flow governed by the
usual Navier-Stokes equations, i.e., the equations describing conservation of mass
and momentum. The (constant) density and kinematic viscosity of the fluid are f
and f , respectively.
The equations are written in an Eulerian framework in the time-dependent
domain ˝f .t/. The physical unknowns are the scalar pressure field pf 2 Lf
and the vector velocity field vf 2 vD f
C Vf0 . Here, vD
f
is a suitable extension of
the prescribed Dirichlet data on the boundaries (both moving or stationary) of ˝f ,
and g1 is a suitable extension to all of @˝f of the Neumann data for f  n on the
boundaries. We have “hidden” the fluid-structure interface conditions of continuity
of velocity and normal stress in parts of the boundary data vDf and g1 .
The variational form of the Navier-Stokes equations in an Eulerian framework is
obtained by multiplying them with suitable test functions from the test space Vf0 for
the momentum equations and Lf for the mass conservation equation.

Problem 1 (Fluid model in Eulerian formulation).

Find fvf ; pf g 2 fvD
f
C Vf0 g  Lf , such that vf .0/ D v0f , and

.f .@t C vf  r/vf ; v

/f C .f ; ". v
//f D .g1 ; v
/@˝f C .f; v
/f ;
(1)
.div vf ; p
/f D 0;

for all f v
; p
g 2 Vf0  Lf , where

f WD pf I C 2f f ".vf /; ".v/ WD 12 .rv C rvT /:

2.3 Structure

In the examples in Sect. 4, we consider two different types of materials, an incom-

pressible neo-Hookean (INH) material and a compressible elastic material described
by the St. Venant-Kirchhoff (STVK) model. These two models will be described by
a set of two parameters, the Poisson ratio s and the Young modulus Es , or alter-
natively, the Lamé coefficients s and s . These parameters satisfy the following
relations:

s 3s C 2s
s D ; Es D s ;
2.s C s / s C s
Es s Es
s D ; s D ;
2.1 C s / .1 C s /.1  2s /
164 R. Rannacher and T. Richter

where s D 12 for incompressible and s < 12 for compressible material. uO D and vO D

are suitable extensions of the prescribed Dirichlet data on the boundaries (usually
uO D D vO D D 0) and gO 2 is the extension of the Neumann data to the complete
boundary @˝O s .

2.3.1 Incompressible neo-Hookean (INH) material

Problem 2 (INH structure model in Lagrangian formulation).

Find fOus ; vO s ; pOs g 2 fOuD C VOs0 gfOvD O0 O
s C Vs g Ls , such that u
O s .0/ D uO 0s , vO s .0/ D vO 0s ,
and

.s dt vO s ; O u /sO C .O s FO T ; "O. O u //sO D .gO 2 ; O u /@˝O s C .fO2 ; O u /sO ;

.dt uO s  vO s ; O v /sO D 0; (2)
.det FO ; O p /sO D .1; O p /sO ;

O s , where FO WD I C rO uO s and
for all f O u ; O v ; O p g 2 VOs0  VOs0  L

O s WD pOs I C s .FO FO T  I /; "O. O u / WD 12 .rO O u C rO O uT /:

2.3.2 Compressible St. Venant-Kirchhoff (STVK) material

Problem 3 (STVK structure model in Lagrangian formulation).

Find fOus ; vO s g 2 fOuD C VOs0 g  fOvD O0
s C Vs g, such that u
O s .0/ D uO 0s , vO s .0/ D vO 0s , and

O O u //sO D .gO 2 ; O u /@˝O s C .fO2 ; O u /sO ;

.s dt vO s ; O u /sO C .JO O s FO T ; ".
(3)
.dt uO s  vO s ; O v /sO D 0;

for all f O u ; O v g 2 VOs0  VOs0 , where "O. O u / WD 12 .rO O u C rO O uT /.

2.3.3 Conversion of the structure models to the Eulerian framework

To rewrite the above conservation equations in an Eulerian frame, we need the pres-
sure pOs , the displacement uO s and its gradient rO uO s in the Eulerian sense, which
are denoted by ps , us and rus , respectively. There holds ps .x/ D pOs .x/ O and
us .x/ D uO s .x/,
O or more precisely

ps .x/ D pOs .Ts .x// D pOs .x/;

O us .x/ D uO s .Ts .x// D uO s .x/;
O

where Ts is the (inverse) displacement function of points in the deformed domain ˝s

back to points in the initial domain ˝O s . The corresponding displacement function is
TOs .x/:
O
Adaptive FE method for FSI 165

TOs W ˝O s ! ˝s ; TOs .x/

O D xO C uO s D x;
(4)
Ts W ˝s ! ˝s ; Ts .x/ D x  us D x:
O

Since detrO TOs D detFO ¤ 0, the displacements Ts and TOs are well defined.
To access the Eulerian deformation, we use the identity Ts .TOs .x//
O D xO for the
displacement functions obtained from (4). Differentiating this yields:

.I  ru/.I C rO uO / D I , rO uO D .I  ru/1  I:

Thus, the Cauchy stress tensor s can be written for INH and STVK materials in an
Eulerian framework as follows:

ps I C s .FF T  I / (INH material);
s D
J 1 F .s .trE/I C 2s E/F T (STVK material);
1
F D I C rO uO D .I  ru/1; J D detF; E D .F T F  I /:
2
Finally, we write the structure equations in Eulerian framework for both types of
material, INH and STVK.
Problem 4 (INH structure model in Eulerian formulation).
Find fus ; vs g 2 fuD
s C Vs g  fvs C Vs g and ps 2 Ls , such that us .0/ D us ,
0 D 0 0

vs .0/ D vs , and
0

.Os Js @t vs ; v
/s C .Os Js vs  rvs ; v
/s C .s ; r v
/s
D .g2 ; v
/sN C .s ns ; v
/i C .Os Js fs ; v
/s ;
(5)
.@t us C vs  rus  vs ; /s D 0;
u

.1  det Fs ; p
/ D 0;

for all f u ; v g 2 Vs0  Vs0 and p

2 Ls where Fs WD I  rus , Js WD det Fs ,
E WD 12 .FsT Fs1  I /, and

s WD ps I C s .Fs1 FsT  I /:

Problem 5 (STVK structure model in Eulerian formulation).

Find fus ; vs g 2 fuD
s C Vs g  fvs C Vs g, such that us .0/ D us , vs .0/ D vs , and
0 D 0 0 0

.Os Js @t vs ; v
/s C .Os Js vs  rvs ; v /s C .s ; r v /s ;
D .g2 ; v /sN C .s ns ; v /i C .Os Js fs ; v
/s ;
(6)
.@t us C vs  rus  vs ; u
/s D 0;
.1  det Fs ; p
/ D 0;

for all f u ; v g 2 Vs0  Vs0 , where Fs WD I  rus , Js WD det Fs , E WD

1
2
.FsT Fs1  I /, and
s WD Js Fs1 .s trE I C 2s E/FsT :
166 R. Rannacher and T. Richter

2.4 The FSI problem in Eulerian formulation

Now, we can combine the Eulerian formulations of the flow and the structure part of
the problem into a complete variational formulation of the nonstationary as well as
the stationary FSI problem in Eulerian framework. Here, the continuity of velocity
across the fluid-structure interface i is strongly enforced by requiring one common
continuous field for the velocity v 2 V in ˝. In the case of STVK material the
(non-physical) pressure ps in the structure subdomain is determined as harmonic
extension of the flow pressure pf , together denoted by p 2 Ł in ˝. The Dirichlet
boundary data vD f
and vDs on parts of @˝ are merged into a suitable velocity field
v 2 V . The deformation us is defined on all of ˝ by a continuous extension
D

to the flow domain. We denote the deformation by u 2 V and will describe this
(non-physical) extension later.
The balance of forces on the interface f  n D s  n on i now implicitly
appears as part of the combined variational formulation in form of a boundary
integral: Z
.Π n; v
/i D .f  s /  nf v
do:
i

The remaining parts of the Neumann boundary data g1 and g2 on @˝ are combined
to g3 . For ease of notation, we introduce the characteristic functions f of ˝f
and s of ˝s respectively. We write the Cauchy stress tensor and the density in the
whole domain as follows:

 WD f f C s s ;  WD f Of C s Os : (7)

The characteristic functions depend on the time t and - in case of the solid part -
on the deformation u. Let O f and O s be the characteristic functions of the reference
domains ˝O f and ˝O s respectively. Then, on the moving Eulerian domains ˝f .t/
and ˝s .t/ the characteristic functions can be expressed by

f .t; x/ D O f .x/;
O s .t; x/ D O s .x/:
O

Since the layout of the Eulerian subdomains is not known a priori, the characteristic
functions must be given implicitly by the solution. To close this gap, Dunne [11]
introduces the set of initial points (IP-Set) '.˝/ for all points of ˝ at time t. This
function indicates the origin of a point with coordinate x 2 ˝ at time t in the
reference domain xO 2 ˝O by xO D '.t; x/. The IP-Set is transported in the full domain
with a certain velocity w. While in the structure domain the structure’s velocity
vs itself is used to transport the IP-Set, using the fluid’s velocity would lead to
very entangled displacements and breakdown of the scheme. Instead, a harmonic
continuation of the structure’s velocity into the flow domain is used.
Here, a different approach, more simular to the ALE coordinates is used. No
further velocity field w is necessary. In the structure domain ˝s .t/ the deformation
has exactly the meaning of the IP-Set, by
Adaptive FE method for FSI 167

xO D ˚.t; x/j˝s .t / WD x  us .t; x/;

the mapping between reference and deformed coordinate is given. Using the har-
monic extension of us to the whole domain, the characteristic functions on the
moving domains are given by:

s .t; x/ D O s .x/
O D O s .x  u.t; x//; f .t; x/ D O f .x/
O D O f .x  u.t; x//:

With u being available on all of ˝, we can decide for every t and x the correspon-
dence to the subdomain by
(
x  u.t; x/ 2 ˝O s ) x 2 ˝s .t/
x2˝ ) (8)
x  u.t; x/ 62 ˝O s ) x 2 ˝f .t/:

The extension of u to the flow domain does not necessarily define a mapping Tf .t/ W
˝f .t/ ! ˝O f from the moving flow domain to the reference domain, only the
information Tf .t; x/ 62 ˝O s is used. Hence, we have more freedom to specify the
boundary values of the extension of u on @˝=@˝O s , e.g. by choosing:


u D 0 in ˝f .t/; u D us on i .t/; n  u D 0 on @˝f .t/=i .t/:

This allows for arbitrary large deformation and movement of the structure in the
flow domain.
Now, we can combine formulas (1), (5) and (6) to obtain a complete variational
formulation of the FSI problem in Eulerian coordinates.
In case of the incompressible material (INH), we use one continuous pressure
variable for both subproblems. Since the “physical pressure” does not need to be
continuous across the interface one could instead decouple the pressure variables
and harmonically extend each one to the other domain:

Problem 6 (FSI Problem in Eulerian formulation, INH material).

Find fu; v; pg 2 fuD C V 0 g  fvD C V 0  L g, such that u.0/ D u0 , v.0/ D v0 , and

..@t v C v  rv/; / C .; / D .g3 ; /@˝ C .f3 ; /;

.div v; / D 0;
(9)
. s .@t u C v  ru  v/; '/
C˛u f. f ˛u ru; r'/  .@n u; '/i g D 0;

for all f ; '; g 2 V 0  V 0  L, and with

f D pf I C f f .rv C rvT /;

p
2 Ls where Fs WD I  rus , Js WD det Fs , E WD 12 .FsT Fs1  I /, and

s WD ps I C s .Fs1 FsT  I /:

168 R. Rannacher and T. Richter

The parameter ˛u  ˛0 h2 needs to be chosen large enough to impose the impor-
tant condition dt u D v. The boundary integral is necessary to prevent feedback from
the continuation. When dealing with the compressible STVK material, the pressure
is harmonically extended to the structure domain.
Problem 7 (FSI Problem in Eulerian formulation, STVK material).
Find fu; v; pg 2 fuD C V 0 g  fvD C V 0  L g, such that u.0/ D u0 , v.0/ D v0 , and

..@t v C v  rv/; / C .; / D .g3 ; /@˝ C .f3 ; /;

. f div v; / C ˛p f. s rp; r/  .@n p; /i g D 0;
(10)
. s .@t u C v  ru  v/; '/
C˛u f. f ˛u ru; r'/  .@n u; '/i g D 0;

for all f u ; v g 2 V 0  V 0 and p

2 L where Fs WD I  rus , Js WD det Fs ,
E WD 12 .FsT Fs1  I /, and

s WD Js Fs1 .s trE I C 2s E/FsT :

In these variational formulations the location of the interface i is given implic-

itly by the deformation:

i .t/ WD fx 2 ˝; x  u.t; x/ 2 Oi g:

The resulting system is nonlinear even if linear models are used for the two
subproblems, e.g. a Stokes fluid and a linear elastic structure.
In some situations the solution of an FSI problem may tend to a “steady state”
as t ! 1. The equations for stationary fluid structure interaction in Eulerian for-
mulation are given in Dunne [11]. Corresponding stationary equations without the
IP-Set are developed in Richter & Wick [30].

2.5 Comparison to ALE formulations

In the classical ALE formulation for FSI problems the structure problem is for-
mulated on the static reference domain ˝O s . The flow problem is transformed back
to the reference flow domain ˝O f by an artificial transformation TOf W ˝O f ! ˝
usually derived by the continuation of the deformation to the flow domain. The situ-
ation in Eulerian coordinates is similar: here, the flow problem resides in the natural
coordinate system while the structure problem needs to be transformed. This trans-
formation Ts W ˝s ! ˝O s however is a “natural transformation” since it is given by
the deformation itself. If the structure solver is capable of handling a certain defor-
mation, the transformation is well behaved and cannot deteriorate in the context of
FSI. The deformation field is extended to the flow domain in both formulations, but
while the extended (and artificial) field is used to transform the flow variables in
Adaptive FE method for FSI 169

the ALE formulation, it is only necessary to lookup the domain of influence in the
Eulerian formulation. Here, no gradient evaluation is required and the regularity is
of lower importance. In the fully Eulerian formulation, the extension of the defor-
mation to the flow domain does not have to define a mapping between the reference
and the deformed fluid domain. Hence, more freedom is possible in modeling this
extension.
Both formulations contain equations for velocity, deformation and pressure, a
total of five solution variables in two spatial dimensions. The Eulerian formulation
is strongly nonlinear due to the implicit dependence of the domain on the deforma-
tion. In the ALE formulation the transformation of the flow domain imposes strong
nonlinearities which can prohibit large deformation. The Eulerian formulation tends
to be slightly more costly due to the implicit definition of the moving fluid-structure
interface i .t/.

3 Discretization

In this section, we detail the discretization in space and time of the FSI problem
based on its different variational formulations. Our method of choice is the Galerkin
finite element (FE) method with “conforming” finite elements, both in space and
time. For a general introduction to the FE method, we refer to Carey and Oden [10],
Girault & Raviart [17], Brenner & Scott [9], or Braess [8]. Space-time Galerkin
methods are described in Eriksson, Estep, Hansbo & Johnson [15], Schmich &
Vexler [29] and Meidner & Vexler [23] and Schmich [28] for the Navier-Stokes
equations. Having a Galerkin method for the completely discretized scheme at hand
rigorous error estimation is accessible.
First, we introduce a semi-discrete formulation incorporating a continuous or
discontinuous Galerkin method in time. The resulting methods will be variants of
the implicit Euler method and of the Crank-Nicolson scheme.
Then, we provide the framework for the spatial finite element method and we
describe the complete variational formulations which are the basis of the Galerkin
discretization. Since we are using a so-called “equal-order” approximation of all
physical quantities, additional pressure stabilization has to be incorporated which is
done here by the “local projection” technique of Becker and Braack [3].
At each time step a nonlinear algebraic problem is solved using a Newton-like
method. This relies on solving the linear defect-correction problem, which requires
the evaluation of the corresponding Jacobi matrix.

3.1 Galerkin formulation

For the discretization of Problem 6 (INH) and Problem 7 (STVK) we choose a

Galerkin approach in space and time using finite dimensional FE subspaces
170 R. Rannacher and T. Richter

Łkh  Ł WD L2 .I; L/; Vkh  V WD C.I; V /:

Within the present abstract setting the discretization in time is likewise by a Galerkin
method such as the dG(r) (“discontinuous” Galerkin) or the cG(r) (“continuous”
Galerkin) method of degree r  0. The dG(0) method is closely related to the
first-order backward Euler scheme and the cG(1) method to the second-order Crank-
Nicolson scheme (indeed even algebraically identical for autonomous problems).
The full space-time Galerkin framework is mainly introduced as the basis for a
systematic approach to residual-based a posteriori error estimation.
At first, we introduce a compact form of the variational formulation of the FSI
problem. For arguments U D fv; u; pg and  D f v ; u ; p g 2 W WD V  V  Ł,
we introduce the space-time semilinear form
Z T ˚
A.U /. / WD ..@t v C v  rv/; v
/ C ..U /; ". v
//
0
(
p
.div v; / (INH material)
C (11)
. f div v; p
/ C . s ˛p rp; r p
/ (STVK material)
C . s ˛v .@t u  v C v  ru/; u / C . f ru; r u
/

 .g3 ; v /@˝  .f3 ; v / dt:

With this notation, we can write the variational problems 6 and 7 in compact form
as follows.
Problem 8 (Compact Eulerian formulation of the FSI problems).
Find U 2 U D C W , such that

A.U /. / D 0 8 2 W ; (12)

where U D is an appropriate extension of the Dirichlet boundary and initial data and
the space W is defined by
˚ 
W WD  D f v
; u
; p
g 2 V  V  Ł; u
.0/ D v
.0/ D 0 :

3.2 Semidiscretization in time

For time discretization, we consider a mixed continuous Galerkin (cG(1)) method

for velocity and deformation and a discontinuous Galerkin (dG(0)) method for the
pressure. This Galerkin method can be interpreted as a variant of the Crank-Nicolson
scheme. A detailed introduction to Galerkin methods for time discretization is given
in the textbook of Eriksson, Estep, Hansbo & Johnson [15] or in the thesis of
Schmich [28]. We partition the time interval IN D Œ0; T  into M 2 N subintervals,

IN D f0g [ I1 [ I2 [    [ IM ;
Adaptive FE method for FSI 171

where Im D .tm1 ; tm , tm > tm1 , and km WD tm  tm1 . For the velocity and
deformation, we consider the following two time-discrete spaces:
ˇ
ˇ
VkcG WD fvk 2 C.IN; V /ˇvk jIm 2 P1 .Im ; V /; m D 1; 2; : : : ; M g;
ˇ
ˇ
VkdG WD fvk 2 L2 .I; V /ˇvk jIm 2 P0 .Im I V /; m D 1; 2; : : : ; M; vk .0/ 2 Vk g;

where P1 .I; V / and P0 .I; V / are the spaces of linear and constant functions,
respectively, on I with values in V . The first space VkcG contains functions which are
continuous and piecewise linear in time, the second one functions which are discon-
tinuous and constant on every subinterval Im . The choice of VkcG as trial and VkdG
as test space leads to a variant of the trapezoidal rule, using VkdG for both test and
trial spaces a variant of the implicit Euler scheme can be derived. For the pressure,
we choose piecewise constant functions in time from the space
ˇ
2 ˇ
k WD fpk 2 L .I; L/ˇpk jIm 2 P0 .Im I L/; m D 1; 2; : : : ; M g:
ŁdG

Then, by WkcG WD VkcG  VkcG  ŁdG k and likewise by Wk

dG
WD VkdG  VkdG 
dG
Łk , we denote the time discrete trial and test spaces. In compact notation, the time
discretized Galerkin formulation of our problem reads: find Uk D .vk ; uK ; pk / 2
UKD C WkcG such that
A.Uk /.k / D 0 8k 2 WkdG ;

with the space-time semilinear form as given in (11).

Since the test function from WkdG are discontinuous in time, this Galerkin scheme
splits into a time-stepping method. Setting

vm
k WD vk .tm /; um
k WD uk .tm /; pkm WD pk jIm ; km WD k jIm ;

and Ukm WD .vm k

; um
k
; pkm /, we can approximate the integrals in the semilinear
form (11) with the trapezoidal rule and solve for the approximation U m with U m1
given from the preceding time step. Here, one has to be careful with the implicit
association of points x 2 ˝ to the specific subdomain. A fluid point x 2 ˝f .tm1 /
can change to a structure point x 2 ˝s .tm / within one time interval Im . We thus
define corresponding to (7)

m
f WD f .tm /; m
s WD s .tm /;

 m WD m
f f .U / C s s .U /;
m m m
m WD m
f Of C s Os ;
m

and, for example, the momentum part in (11) approximated on Im takes the form
Z  m1 C m 
..@t v C v  rv/; v /dt  .vm  vm1 /; ' v
Im 2

km m1 m1 
C  v  rvm1 C m vm  rvm ; ' v :
2
172 R. Rannacher and T. Richter

3.3 Spatial discretization

The spatial discretization is by a conforming finite element Galerkin method on

meshes ˝h consisting of cells denoted by K,
[
˝N D KN i ;
i D1;:::;N

which are (convex) quadrilaterals in 2d or hexahedrals in 3d. Such a decomposition

˝h is referred to as “regular” if any cell edge is either a subset of the domain bound-
ary components D ; N , or a complete face or edge of another cell. However, to
facilitate mesh refinement and coarsening, we allow the cells to have a certain num-
ber of nodes that are at the midpoint of sides or faces of neighboring cells. These
“hanging nodes” do not carry degrees of freedom and the corresponding function
values are determined by linear or bilinear interpolation of neighboring “regular”
nodal points. For more details on this construction, we refer to Carey & Oden [10]
or Bangerth & Rannacher [1].
The mesh parameter h is a scalar cellwise constant function defined by hjK WD
hK D diam.K/. We set hmax WD maxK2˝h hK . For a cell K, we denote by K the
diameter of the maximal inscribed ball in K. To ensure proper approximation prop-
erties of the finite element spaces which are constructed based on the meshes ˝h ,
we require the “uniform-shape” and the “uniform-size” conditions to be fulfilled:
Mesh regularity condition: Each cell K 2 Th is the image of the reference unit
cube KO D Œ0; 1d under some d -linear mapping K W KO ! K. This mapping is
uniquely described by the 2d coordinate values of the corners of K, if the ordering
0
of the corners is preserved, see Fig. 1. The Jacobian tensors K of these mappings
are invertible and satisfy the uniform bounds
0 0 1
sup max kK k c; sup max k.K / k c: (13)
h>0 K2Th h>0 K2Th

This condition is satisfied if the cells K 2 Th possess the usual structural properties
of uniform “non-degeneracy”, “uniform shape”, and “uniform size property”.
To increase the number of cells in a decomposition ˝h , we employ “mesh refine-
ment”, which consists of subdividing a cell into 2d subcells. Cell subdivision is done
by connecting the midpoints of opposing edges or faces of a cell. A refinement is

Fig. 1 Reference mapping K W KO ! K.

Adaptive FE method for FSI 173

Fig. 2 A regular mesh after two cycles of global refinement.

Fig. 3 A regular mesh after two cycles of local refinement.

Fig. 4 A regular mesh after two cycles of patchwise local refinement.

“global” if this is done for every cell. An example of a regular mesh and two global
refinements is shown in Fig. 2. Each of the resulting meshes after refinement is
also regular. “Coarsening” of 2d cells is possible if they were generated by prior
refinement of some “parent cell”. A group of 2d such cells is referred to as a “(cell)
patch”.
In addition to “global” refinement, we will also use “local” refinement. This con-
sists of only subdividing some cells in a given decomposition. Such refinement leads
to nodes that are placed on the middle of the neighboring cells’ edges or faces, i.e., to
“hanging nodes”, where only one hanging node is allowed per edge or face. In Fig. 3
local refinement is applied twice leading to hanging nodes indicated by “bullets”.
Sometimes, we will require that a decomposition ˝h is organized in a patch-
wise manner. This means that ˝h is the result of global refinement of the coarser
decomposition ˝2h , as shown in Fig. 4.
Given a function space V the decomposition ˝h and the cellwise space of poly-
nomial functions Q.K/, we construct the corresponding finite element subspace
Vh  V by
Vh WD f' 2 V; 'jK 2 Q.K/; K 2 ˝h g:
Each polynomial function space Q.K/ is actually defined on a “reference cell”
O K/.
KO WD .0; 1/d as the reference function space Q. O We use the function space
of polynomial degree p  0 on KO defined by
˚ 
O WD span xO ˛ ; xO D .xO 1 ; : : : ; xO d / 2 K;
QO p .K/ O ˛ D .˛1 ; : : : ; ˛d /; ˛i 2 f0; : : : ; pg ;
174 R. Rannacher and T. Richter

with the usual multi-index notation. In the numerical tests presented below, only
finite elements with p D 1 (“d-linear elements”) are used. Therefore, we omit the
O K/.
superscript p and simply refer to Q. O The reference function space Q. O K/
O is
mapped to the corresponding cell K with the help of the d-linear mapping K W
KO ! K, ˚ 
Q.K/ D '.x/ D '. O K 1 O K/;
.x//; 'O 2 Q. O x2K :

The d-linear mapping K W KO ! K is uniquely described by the 2d coordinate

values of the corners of K, if the ordering of the corners is preserved, see Fig. 1.
Hence, in this case, the reference function space and the mapping function space are
the same. Accordingly, the resulting finite elements are called “isoparametric”.
To allow for dynamically changing meshes, we associate with ˝mh the decom-
position of the domain ˝ used in the time interval Im . The corresponding finite
element subspace is denoted by Vmh  V . The fully discrete trial and test spaces
are then defined as
ˇ
ˇ
Vkh
cG
WD fvkh 2 C.IN; V /ˇvkh jIm 2 P1 .Im ; Vmh /; m D 1; 2; : : : ; M g;
ˇ
ˇ
Vkh
dG
WD fvkh 2 L2 .I; V /ˇvkh jIm 2 P0 .Im ; Vmh /; m D 1; 2; : : : ; M; vkh .0/ 2 V0h g;
ˇ
ˇ
kh WD fpkh 2 L .I; L/ˇpkh jIm 2 P0 .Im ; Vmh /; m D 1; 2; : : : ; M g:
2
ŁdG

The spatial discretization by “equal-order” finite elements for velocity and pressure
needs stabilization in order to compensate for the missing “inf-sup stability”. We
use the so-called “local projection stabilization” (LPS) introduced by Becker and
Braack [3,4]. An analogous approach is also employed for stabilizing the convection
in the flow model as well as in the transport equation for the displacement u. To this
end, we define the mesh-dependent bilinear form
X
.'; /ı WD ıK .'; /K ;
K2˝h

where the parameter ıK is adaptively determined by

˛h2K
ıK WD :
f f f C s s C ˇjvh j1IK hK

Further, we introduce the “fluctuation operator” h W Vh ! V2h on the finest mesh

level ˝h by h WD I  P2h , where P2h W Vh ! V2h denotes the L2 -projection. The
operator h measures the fluctuation of a function in Vh with respect to its projection
into the next coarser space V2h . With this notation, we define the stabilization form
Z T ˚ p
Sı .Uh /.h / WD .r h ph ; r h C .vh  r h vh ; vh  r h
/
h ı
v
h /ı
0

C . s vh  r h uh ; vh  r h hu /ı dt;
Adaptive FE method for FSI 175

where the first term stabilizes the fluid pressure, the second one the INH structure
pressure, the third one the transport in the flow model, and the fourth one the trans-
port of the displacement uh . The LP stabilization has the important property that it
acts only on the diagonal terms of the coupled system and that it does not contain
any second-order derivatives. However, it is only “weakly” consistent, as it does
not vanish for the continuous solution, but it tends to zero with the right order as
h ! 0. The choice of the numbers ˛; ˇ in the stabilization parameter ıK is, based
on practical experience, in our computations ˛ D 1=2, and ˇ D 1=6.
With this notation the stabilized Galerkin finite element approximation of Prob-
lem 8 reads as follows.
Problem 9 (Space-Time discrete Galerkin approximation of the FSI problem
in Eulerian framework).
Find Ukh D fvkh ; ukh ; pkh g 2 Ukh
D
C Wkh
cG
, such that

Aı .Ukh /.kh / WD A.Ukh ; kh / C Sı .Ukh /.kh / D 0 8h 2 WhdG ; (14)

where the “discrete” finite element spaces Wkh

cG dG
and Wkh are defined analogously
as their “continuous” counterparts.
As on the spatially continuous level the existence of solutions to this semi-
discrete problem is not guaranteed and has to be justified separately for each
particular situation.

3.4 Error estimation and mesh adaptation

The main issues of this article is the automatic mesh adaptation within the Eulerian
finite element approximation of the FSI problem. The computations presented below
have been done on three different types of meshes:
 globally refined meshes obtained using several steps of uniform refinement of a
coarse initial mesh,
 locally refined meshes obtained using a purely geometry-based criterion by
marking all cells for refinement which have certain prescribed distances from
the fluid-structure interface,
 locally refined meshes obtained using a systematic residual-based criteria by
marking all cells for refinement which have error indicators above a certain
threshold.
The goal is to employ the so-called “Dual Weighted Residual Method” (DWR
method) for the adaptive solution of FSI problems. This method has been developed
in Becker & Rannacher [5, 6] (see also Bangerth & Rannacher [1]) as an exten-
sion of the duality technique for a posteriori error estimation described in Eriksson,
Estep, Hansbo & Johnson [15]. The DWR method provides a general framework for
the derivation of “goal-oriented” a posteriori error estimates together with criteria
176 R. Rannacher and T. Richter

of mesh adaptation for the Galerkin discretization of general linear and nonlinear
variational problems, including optimization problems. It is based on a complete
variational formulation of the problem, such as (12) for the FSI problem. In fact, this
was one of the driving factors for deriving the Eulerian formulation underlying (12).
We begin with a brief outline of the DWR method for the special case of an
FSI problem governed by an abstract variational equation posed on a space-time
domain such as (12). We restrict us to a simplified version of the DWR method,
which suffices for the present purposes. For a more elaborated version, which is
particularly useful in the context of optimization problems, we refer to the literature
stated above. For notational simplicity, we think the nonhomogeneous boundary
and initial data U D to be incorporated into a linear forcing term F .  /, or to be
exactly representable in the approximating space Wh . Then, we seek U 2 U D CW 0
such that

A.U /. / D F . / 8 2 W 0 : (15)

The corresponding (stabilized) Galerkin approximation seeks Uh 2 UhD C Wh0 such

that

A.Uh /.h / C Sı .Uh /.h / D F . / 8h 2 Wh0 : (16)

Suppose that the goal of the computation is the evaluation of the value J.U / for
some functional J.  / (for simplicity assumed to be linear) which is defined on W .
We want to control the quality of the discretization in terms of the error

J.U  Uh / D J.U /  J.Uh /:

To this end, we introduce the directional derivative A0 .U /.˚;  / the existence of

which is assumed. With the above notation, we introduce the bilinear form
Z 1
L.U; Uh /.˚;  / WD A0 .Uh C s.U  Uh //.˚;  / ds;
0

and formulate the “dual problem”

L.U; Uh /.˚; Z/ D J.˚/ 8˚ 2 W 0 : (17)

In the present abstract setting the existence of a solution Z 2 W 0 of the dual prob-
lem (17) has to be assumed. Now, taking ˚ D U  Uh 2 W 0 in (17) and using the
Galerkin orthogonality property

A.U /.h /  A.Uh /.h / D Sı .Uh /.h /;  2 Wh0 ;

Adaptive FE method for FSI 177

yields the error representation

J.U  Uh / D L.U; Uh /.U  Uh ; Z/

Z 1
D A0 .Uh C s.U  Uh //.U  Uh ; Z/ ds
0
D A.U /.Z/  A.Uh /.Z/
D F .Z  h /  A.Uh /.Z  h /  Sı .Uh /.h /
DW .Uh /.Z  h /  Sı .Uh /.h /;

where h 2 W 0 is an arbitrary element, usually taken as the generic nodal inter-

polant Ih Z 2 Wh0 of Z. For the evaluation of the terms on the right-hand side, we
split the integrals in the residual term .Uh /.Z  h / into their contributions from
the single time intervals Im and spatial mesh cells K 2 Thm and integrate by parts.
This results in an estimate of the error jJ.U  Uh /j in terms of computable local
residual terms K .Uh / multiplied by certain weight factors !K .Z/ which depend
on the dual solution Z,

X
M X
m m
jJ.U  Uh /j K .Uh / !K .Z/ C jSı .Uh /.h /j: (18)
mD1 K2T m
h

The explicit form of the terms in the sum on the right-hand side is rather compli-
cated and will not be repeated here. The second term due to the regularization is
assumed to be small and is therefore neglected. For details, we refer to Dunne [12]
and Dunne & Rannacher [13].
Since the dual solution Z is unknown, the evaluation of the weights !K .Z/
requires further approximation. First, we linearize by assuming

L.U; Uh /.˚;  /  L.Uh ; Uh /.˚;  / D A0 .Uh /.˚;  /;

and use the approximate “discrete” dual solution Zh 2 Wh0 defined by

A0 .Uh /.˚; Zh / D J.˚h / 8˚h 2 Wh0 : (19)

From Zh , we generate improved approximations to Z in a post-processing step

by patchwise higher-order interpolation. For example in 2d on 2  2-patches of
cells in Th the 9 nodal values of the piecewise bilinear Zh are used to construct a
patchwise biquadratic function ZQ as indicated in Fig. 5. This is then used to obtain
the approximate error estimate

X
M X
m Q
jJ.U  Uh /j   WD m
K .Uh / !K .Z/; (20)
mD1 K2T m
h

which is the basis of automatic mesh adaptation.

178 R. Rannacher and T. Richter

(2)
Zh I2hZh

.2/
Fig. 5 Local postprocessing by patchwise “biquadratic” interpolation, I2h Zh , of the “bilinear”
discrete solution Zh in 2D.

Remark 1. The dual solution Z has the features of a “generalized” Green function
G.K; K 0 /, as it describes the dependence of the target error quantity J.U  Uh /,
which may be concentrated at some cell K, on local properties of the data, i.e. in
this case the residuals K 0 on cells K 0 .

Remark 2. The solvability of the primal and dual problems (15), (16) and (17), (19),
respectively, is not for granted. This is a difficult task regarding the rather few exis-
tence results in the literature for general FSI problems. Further, the assumption of
differentiability cause concerns in treating the FSI problems in the Eulerian frame-
work since the dependence of the characteristic function f .x; u/ D O f .x  u/
on the deflection u is generally not differentiable (only Lipschitzian). However, this
non-differentiability can be resolved by the “Hadamard structure theorem”, on the
assumption that the interface between fluid and structure forms a lower dimensional
manifold and the differentiation is done in a weak variational sense. In essence this
has the same effect as discretizing along the interface and replacing the directional
derivative by a mesh-size dependent difference quotient, a pragmatic approach that
has proven itself in similar situations, e.g. for Hencky elasto-plasticity in Rannacher
and Suttmeier [27].

For the primal problem the directional (Gâteaux) derivative of the complete FSI
problem does not need to be exact, it only needs to be “good enough” for Newton
iteration to ensure convergence, leading to a reduction of the residuals of the non-
linear system. Thus, for the primal problem the nonlinear system is used to measure
the “quality” of the approximation. For the dual problem though things may initially
seem less clear, since the dual problem is simply a linear problem directly based on
the Gâteaux derivative. Of course, an immediate “measure of quality” of the discrete
dual solution is the residual of the linear system. But there is no immediate measure
for the quality of the discrete dual solution in relation to the continuous dual solu-
tion. This uncertainty stems from the highly nonlinear influence of the displacement
u in the Gâteaux derivative, which is seen in the transformed fluid equations in the
ALE framework. For the Eulerian framework, additional boundary Dirac integrals,
which stem from the shape derivatives are causing these nonlinearities. This seem-
ingly lack of clarity though is not typical to FSI problems. It is only more obvious
in such problems since everything visible depends on the position of the interface.
Generally though this uncertainty concerning the discrete dual solution is present
Adaptive FE method for FSI 179

in all nonlinear problems, since in such problems the Gâteaux derivatives depend
on the primal solution and can only be approximated by using the discrete primal
solutions.
In the case of FSI problems, we assume that the interface obtained on the current
mesh is already in good agreement with the correct one, hi  i , and set up the
dual problem formally with hi as a fixed interface. In all test computations, we did
not encounter difficulties in obtaining the discrete solutions. In fact the performance
of the error estimator for a given goal functional was always good for both the ALE
and the Eulerian framework.

3.5 Mesh adaptation algorithm

The approach which is used for the adaptive refinement of the spatial mesh is
straightforward. Let an error tolerance TOL be given. Then, on the basis of the
(approximate) a posteriori error estimate (20), the mesh adaptation proceeds as
follows:
1. Compute the primal solution Uh from (16) on the current mesh, starting from
some initial state, e.g., that with zero deformation.
2. Compute the solution ZQ h of the approximate discrete dual problem (19).
m Q
3. Evaluate the cell-error indicators m K WD K .Uh / !K .Zh /.
m

4. If  < TOL then accept Uh and evaluate J.Uh /, otherwise proceed to the next
step.
5. Determine the 30% cells with largest and the 10% cells with smallest values of
K . The cells of the first group are refined and those of the second group coars-
ened. Then, continue with Step 1. (Coarsening usually means canceling of an
earlier refinement. Further refinement may be necessary to prevent the occur-
rence of too many hanging nodes. In two dimensions this strategy leads to about
a doubling of the number of cells in each refinement cycle. By a similar strat-
egy it can be achieved that the number of cells stays about constant during the
adaptation process within the time stepping procedure.

Remark 3. The error representation (18) has been derived assuming the error func-
tional J.  / as linear. In many applications nonlinear, most often quadratic, error
functionals occur. An example is the spatial L2 -norm error

J.Uh / WD k.U  Uh /.T /k;

at the end time T . For nonlinear (differentiable) error functionals the DWR approach
can be extended to yield an error representation of the form (18); see Becker &
Rannacher [6], Bangerth & Rannacher [1].
180 R. Rannacher and T. Richter

3.6 Evaluation of directional derivatives and solution

of the algebraic problems

The time-discretized scheme as derived in Sect. 3.2 can be formulated as a time-

stepping scheme, similar to the Crank-Nicolson scheme. The nonlinear problem in
every time-step is solved by a Newton-like method. The resulting linear subprob-
lems are then solved by the “Generalized Minimal Residual (GMRES)” method
with preconditioning by a geometric multigrid method with block-ILU smoothing.
Since such an approach is rather standard nowadays, we omit its details and refer to
some relevant literature, e.g., Turek [31], Rannacher [26], or Hron and Turek [20].
For the implementational details of using the multigrid method on locally refined
meshes, we refer to Becker and Braack [2].
The linear operators are essentially (if time-stepping parts and factors stemming
from the approximation of the temporal derivatives are neglected) the directional
derivatives of the governing semilinear form of the variational formulation, i.e. the
form A.U /. / in the Eulerian framework,

d
A0 .U /.˚;  / WD A.U C "˚/. /j"D0 : (21)
d"
For only “weakly nonlinear” systems such as the original Navier-Stokes equations
in Eulerian framework obtaining the directional derivative is a straight forward task
and can be done analytically “by hand”. For structure mechanical systems (for
example based on the St. Vernant-Kirchhoff material law) though writing down
the explicit directional derivative can become cumbersome. For example in the
b'O v / is strongly nonlinear in uO ,
Lagrangian case the scalar product .JOFO T ; r

O
JO O FO T D FO .s trEO I C 2s E/; buO ;
FO D I C r EO D 12 .FO T FO  I /:

However, in the Eulerian framework the corresponding scalar product takes the form
.; r'/, which does not become any easier since the Cauchy stress tensor  is based
on the inverse of the “reverse deformation gradient” I  ru,

 D JF 1 .s trE I C 2s E/F T ; F D I  ru; E D 12 .F T F 1  I /:

To alleviate this problem one may use a method that is the basis of “Automatic
Differentiation” such as described in Rall [24] and Griewank [18]. The method is
used to determine the derivative of a function at a given position. It is based on the
technique of mechanically applying the basic rules of differentiation to the “seri-
alized evaluation” of a function. This is achieved by breaking down the evaluation
of the function for a given value into a sequence of basic elementary evaluations.
Consequently, since evaluation is done in a sequence the resulting values from one
evaluation are used in a later evaluation. To these elementary parts the rules of differ-
entiation (i.e. the chain rule, the sum rule and the product rule) are applied. For the
details of the application of this techniques within the simulation of FSI problems,
we refer to Dunne [12].
Adaptive FE method for FSI 181

4 Numerical examples

For the verification of the Eulerian formulation of fluid-structure interaction prob-

lems different numerical test-cases have been studied by Dunne [12], Dunne &
Rannacher [13] and Richter & Wick [30]. Being slightly more expensive, the Eule-
rian formulation yields the same results (and the same convergence order) as the
well-established ALE approach. In Dunne & Rannacher [13] the FLUSTRUK-A
benchmark by Hron & Turek [21] is studied. These results are compared with differ-
ent approaches, both monolithic and partitioned. For completeness, we recall some
of the most important observations obtained by these test calculation. In Richter &
Wick [30] at several numerical test-cases the behavior of the Eulerian and the ALE
formulation for large deformations is investigated, finding that the Eulerian formu-
lation is capable to model large deformations easier without the need of remeshing
or reinitialization. In this section, we present numerical examples which partic-
ularly demonstrate the capability of the Eulerian formulation to allow for large
deformation and free movement of the structure.

4.1 The FSI benchmark FLUSTRUK-A revisited

From Dunne & Rannacher [13], we recall the essential results for the FSI bench-
mark FLUSTRUK-A described in Hron & Turek [21]. A thin elastic bar immersed
in an incompressible fluid develops self-induced time-periodic oscillations of dif-
ferent amplitude depending on the material properties assumed. The configuration
of this benchmark is shown in Fig. 6. The test cases FSI-2* and FSI-3* are slight
variations from the original settings FSI-2 and FSI-3 using different material param-
eters for a better demonstration of special properties using the Eulerian approach.
This benchmark has been defined to validate and compare the different compu-
tational approaches and software implementations for solving FSI problems. In
order to have a fair comparison of our Eulerian-based method with the traditional
Eulerian-Lagrangian approach, also an ALE method has been implemented for this
benchmark problem.

A=(0.6,0.2)
H

Fig. 6 Configuration of the FSI benchmark FLUSTRUK-A.

182 R. Rannacher and T. Richter

Configuration: The computational domain has length L D 2:5, height H D 0:41,

and left bottom corner at .0; 0/. The center of the circle is positioned at C D
.0:2; 0:2/ with radius r D 0:05. The elastic bar has length l D 0:35 and height
h D 0:02. Its right lower end is positioned at .0:6; 0:19/ and its left end is clamped
to the circle. Control points are A.t/ fixed at the trailing edge of the structure with
A.0/ D .0:6; 0:20/, and B D .0:15; 0:2/ fixed at the cylinder (stagnation point).
Boundary and initial conditions: The boundary conditions are as follows: Along the
upper and lower boundary the usual “no-slip” condition is used for the velocity. At
the (left) inlet a constant parabolic inflow profile,

4y.H  y/
v.0; y/ D 1:5 UN ;
H2
is prescribed which drives the flow, and at the (right) outlet zero-stress   n D 0
is realized by using the “do-nothing” approach in the variational formulation (see
Heywood, Rannacher & Turek [19] and Rannacher [26]). This implicitly forces the
pressure to have zero mean value at the outlet. The initial condition is zero flow
velocity and structure displacement.
Material properties: The fluid is assumed as incompressible and Newtonian, the
cylinder as fixed and rigid, and the structure as (compressible) St. Venant-Kirchhoff
(STVK) material.
Discretization: The first set of computations is done on globally refined meshes
for validating the proposed method and its software implementation. Then, for the
same configuration adaptive meshes are used where the refinement criteria are either
purely heuristic, i.e., based on the cell distance from the interface, or are based on a
simplified stationary version of the DWR approach. In all cases a uniform time-step
size of 0:005 s is used. The curved cylinder boundary is approximated to second
order by polygonal mesh boundaries.
Below, we recall the results obtained for three of the most important test cases
within the benchmark FLUSTRUK-A. The corresponding characteristic problem
parameters are listed in Table 1. These tests are intended to demonstrate that the
fully Eulerian-based simulation method developed in this project is comparable and

Table 1 Parameter settings for the FSI test cases.

parameter FSI-2* FSI-3 FSI-3*
structure model STVK STVK INH
f Œ103 kg m3  1 1 1
f Œ103 m2 s 1  1 1 1
s 0.4 0.4 0.5
s Œ103 kg m3  20 1 1
s Œ106 kg m1 s 2  0.5 2 2
UN Œm s 1  0 2 2
Adaptive FE method for FSI 183

competitive with the common ALE approach with respect to accuracy including
mass conservation, while being sufficiently flexible in dealing with large deforma-
tion.

4.1.1 Moderate deformation: ALE versus Eulerian Method

We begin with the FSI-3 test case. Some snapshots of the results of these simulations
are shown in Fig. 7. The corresponding time-dependent behavior of the displace-
ments are shown in Fig. 8. With the Eulerian approach we estimated an oscillation
with a frequency of 5:48 s 1 and a lower frequency of 5:04 s 1 using ALE coordi-
nates. Hron and Turek [20] identified a frequency of 5:3 s 1 as reference value. As
maximum amplitude of the vertical displacement, the Eulerian approach gave the

Fig. 7 FSI-3 Test: Some snapshots of results obtained by the ALE (top two) and the Eulerian
(bottom two) approaches.

0.02 0.02
uy of point A

uy of point A

0.01 0.01

0 0

– 0.01 – 0.01

– 0.02 – 0.02

– 0.03 – 0.03

– 0.04 – 0.04
0 0.5 1 1.5 2 2.5 3 3.5 0 0.5 1 1.5 2 2.5 3 3.5
time [s] time [s]

Fig. 8 FSI-3 Test: Vertical Displacement of the control point A, obtained by the Eulerian approach
(left, N D 3876 cells) with maximum amplitude 3:01  102 and frequency 5:48 s 1 , and by
the ALE approach (right, N D 2082 cells) with maximum amplitude 3:19  102 and frequency
5:04 s 1 .
184 R. Rannacher and T. Richter

4
relative mass error [%]

–2

–4

–6

–8
0 1 2 3 4 5 6 7
time [s]

Fig. 9 FSI-3*: Relative mass error of the bar.

value 3:01  102 versus 3:19  102 for ALE coordinates. Here, the corresponding
reference value are given by Hron and Turek [20] as 3:438  102 . While our results
differ from these reference values by less than 5% for the frequency, the maximum
amplitude shows a larger deviation of up to 10%.
The FSI-3* test case is used to illustrate the mass conservation properties of the
Eulerian solution approach. Here, the incompressible neo-Hookean (INH) material
is used instead of the compressible St. Venant-Kirchhoff material. Since in the Eule-
rian approach the structure deformations are not in a Lagrangian framework, it is
not immediately clear, due to the coupling with the fluid, how well the mass of the
structure is conserved in an Eulerian approach, especially in the course of an insta-
tionary simulation comprising hundreds of time steps. In Fig. 9, we display the bar’s
relative mass error as a function of time. Except for certain initial jitters, the relative
error is less than 1%.

4.1.2 Large deformation: local versus zonal mesh refinement

In the test case FSI-2* the fluid is initially at rest and the bar is subjected to a verti-
cal (gravitational) force. This causes the bar to bend downward until it touches the
bottom wall. The density of the bar has been chosen larger than in the original FSI-
2 benchmark to yield a bigger deformation of the bar. We compare the efficiency
of zonal refinement versus local refinement by the DWR method for this test case
within the Eulerian methods. Figure 10 shows corresponding sequences of snap-
shots, while the position of the trailing tip A is displayed in Fig. 11. We see that
by sensitivity-driven local refinement within the DWR method on only 1 900 cells
Adaptive FE method for FSI 185

Fig. 10 A sequence of snap-shots of the bar’s large deformation under gravitational loading
obtained by the Eulerian approach using zonal refinement with N  3 000 and N  12 000 cells
(left and middle), and local mesh refinement by the DWR method (right) with only N  1 900
cells.

0.2 0.03
Heuristic Refinement, N approx. 12300 Heuristic Refinement, N approx. 12300
Heuristic Refinement, N approx. 3000 Heuristic Refinement, N approx. 3000
Adaptive Refinement, N approx. 1900 Adaptive Refinement, N approx. 1900
0.025
y position of point A

0.15
y position of point A

0.02

0.1
0.015

0.05
0.01

0 0.005
0 0.2 0.4 0.6 0.8 1 0.6 0.7 0.8 0.9 1 1.1
time [s] time [s]

Fig. 11 Time varying position of the point xA .t / over the full time interval Œ0; 1:1 (left) and over
a zoomed interval Œ0:6; 1:1 (right).

almost the same accuracy can be achieved as by zonal refinement on 12 300 cells.
The gain in CPU time needed is almost 85 % (about 30 h for the zonal versus about
4 h for the local refinement).

4.2 Free movement

Limiting factor for large deformation using the ALE framework is the implicit
transformation of the fluid mesh to artificial coordinates, usually defined via
186 R. Rannacher and T. Richter

^
Ωf
F

^
Ωs

Fig. 12 Rectangular elastic structure in unit square.

TOf D id C uO f , where uO f is an extension of the deformation to the flow domain

˝O f . In the Eulerian formulation, the flow problem is not transformed but instead
given in natural coordinates. Here, the artificial extension of the deformation field u
to ˝f is not used as transformation, but just needed to trace the domain of influence
in the definition of the characteristic functions (8).
In the middle of the unit square ˝ D .0; 1/2 , we prescribe an elastic object
˝O s D f.x; y/ W x 2 .0:3; 0:7/; y 2 .0:45; 0:55/g (see Fig. 12). On the boundary
@˝, the do-nothing condition is given for the velocity and a homogenous Neumann
condition for the deformation:

f @n v C p  n D 0 on @˝; @n u D 0 on @˝:

The dynamics of the system is driven by a force acting on the elastic structure:

.s .@t v C v  rv/; '/˝s .t / C .s ; r'/˝s .t / D .s Js F; '/;

with the volume force F given by

 
y C 12
F D :
x  12

This force results in a rotation of the rectangle and a rotational flow evolving in the
domain ˝. By the standard ALE approach (without remeshing and reinitialization)
this problem is not computable, since the extension of the deformation u to the flow
domain would lead to an increasingly large distortion of the meshes. In Fig. 13 we
show snapshots of the solution for different time points.
For comparison we show corresponding results obtained using an ALE formu-
lation. For the construction of the mapping TOf W ˝O f ! ˝f .t/ a bi-harmonic
extension of the deformation u to the flow domain is used. This usually yields the
Adaptive FE method for FSI 187

Fig. 13 Rotation of the elastic rectangle by a volume force.

Fig. 14 Implicit mesh deformation using ALE coordinates for the rotating rectangular structure.

best results and allows for the largest possible deformation, see Dunne [12] for
details. In Fig. 14 we show the deformation and the resulting distorted meshes for
different time-steps. The last snapshot indicates the largest possible deformation of
the flow domain before breakdown of the solution scheme in the next time-step. For
the ALE comparison a very coarse mesh is used for a better visualization of the
effects.
188 R. Rannacher and T. Richter

Γwall
^
Ωf
Γin
^
Ωs Γout
Γwall

Fig. 15 Narrowing channel with elastic obstacle. Flow driven by pressure drop: p D 0 on in and
p D 1 on out .

4.3 Boundary contact of the structure

The ALE formulation has problems when large deformation appears close to the
boundary of the domain. In particular the contact of the elastic structure with the
boundary is not possible within a monolithic formulation using simple ALE coor-
dinates without remeshing techniques. The transformation TOf .t/ W ˝O f ! ˝f .t/
of the flow domain would degenerate. Here, we model a pressure induced flow in a
narrowing channel, blocked by an elastic structure, see Fig. 15.
At the inflow and outflow boundary the do-nothing condition is prescribed with
a pressure drop causing a flow to the right:

f @n v C p  n D 0 on in ; f @n v C p  n D 1 on out :

Prescribing a pressure drop instead of a Dirichlet inflow condition, the problem

stays well-posed even if the channel is closed. Along the remaining boundary wall
homogenous Dirichlet condition v D 0 is prescribed. For the extension of the dis-
placement u to the flow domain a slip-like boundary condition is enforced on the
outer boundary,
u  n D 0 on wall :
In Fig. 16, we show snapshots of the deformed structure ˝s .t/ for different time
points. First, by the evolving flow the structure is bended to the right. “Contact” of
the structure with the wall is realized up to one layer of mesh elements. After mesh
refinement, the structure gets closer to the boundary of the domain.

5 Summary

In this paper, we presented a monolithic, fully Eulerian variational formulation

for “fluid-structure interaction (FSI)” problems. This approach allows us to treat
FSI problems with free bodies and large deformations. This is the main advan-
tage of this method compared to interface tracking methods such as the arbitrary
Lagrangian-Eulerian (ALE) method. For the FSI benchmark FLUSTRUK-A the
Eulerian approach turns out to yield results which are in good agreement with those
obtained by the ALE approach. In order to have a “fair” comparison both methods
have been implemented using the same numerical components and software library
Adaptive FE method for FSI 189

Fig. 16 Near-contact of the elastic structure with the wall. ”Contact” is possible up to one layer
of mesh elements.

GASCOIGNE [16]. The method based on the Eulerian approach is inherently more
expensive than the ALE method, by about a factor of two, but it allows to treat also
large deformations and boundary contact of the structure. This potential has been
investigated for several model configurations.
The monolithic variational formulation of the FSI problem forms the basis for the
application of the “Dual Weighted Residual (DWR)” method for “goal-oriented” a
posteriori error estimation and mesh adaptation. In this method inherent sensitivities
of the FSI problem are utilized by solving linear “dual” problems, similar as in the
Euler-Lagrange approach to solving optimal control problems. The feasibility of
this approach, in a simplified quasi-stationary version, has been demonstrated for
the FSI benchmark FLUSTRUK-A.

Acknowledgements We gratefully acknowledge the financial support of this project by the

Deutsche Forschungsgemeinschaft within the Research Unit 493 Fluid-Structure Interaction:
Modelling, Simulation, Optimization.

References

1. W. Bangerth and R. Rannacher, Adaptive Finite Element Methods for Differential Equations,
Birkhäuser, 2003.
2. R. Becker and M. Braack, Multigrid techniques for finite elements on locally refined meshes.
Numer. Linear Algebra Appl. 7, 363–379, 2000.
3. R. Becker and M. Braack, A finite element pressure gradient stabilization for the Stokes
equations based on local projections, Calcolo 38, 173-199, 2001.
190 R. Rannacher and T. Richter

4. R. Becker and M. Braack, A two-level stabilization scheme for the Navier-Stokes equations,
Proc. ENUMATH-03, pp. 123-130, Springer, 2003.
5. R. Becker and R. Rannacher, Weighted a-posteriori error estimates in FE methods, Lecture
ENUMATH-95, Paris, Sept. 18-22, 1995, in: Proc. ENUMATH-97, (H.G. Bock, et al., eds),
pp. 621-637, World Scientific Publ., Singapore, 1998.
6. R. Becker and R. Rannacher, An optimal control approach to error estimation and mesh adap-
tion in finite element methods, Acta Numerica 2000 (A. Iserles, ed.), pp. 1-101, Cambridge
University Press, 2001.
7. S. Bönisch, Th. Dunne, and R. Rannacher, Lecture on numerical simulation of liquid-structure
interaction, in Lecture Notes of Oberwolfach Seminar Hemodynamical Flows: Aspects of Mod-
eling, Analysis and Simulation, Nov. 20-26, 2005, Mathematical Resarch Institute Oberwolfach
(G.P. Galdi, R. Rannacher, A.M. Robertson, S. Turek, eds), Birkhäuser, Basel, 2007.
8. D. Braess, Finite Elements, Cambridge University Press, Cambridge, United Kingdom, 1997.
9. S. Brenner and R. L. Scott, The Mathematical Theory of Finite Element Methods, Springer,
Berlin Heidelberg New York, 1994.
10. G. Carey and J. Oden, Finite Elements, Computational Aspects, volume III. Prentice-Hall,
1984.
11. Th. Dunne, An Eulerian approach to fluid-structure interaction and goal-oriented mesh refine-
ment, Proc. “Finite Elements for Flow Problems (FEF05)”, IACM Special Interest Conference
supported by ECCOMAS, April 4-6, 2005, Swansea, Wales, UK, Int. J. Numer. Methods Fluids
51, 1017–1039, 2006.
12. Th. Dunne, Adaptive Finite Element Simulation of Fluid Structure Interaction Based on an
Eulerian Formulation, Doctoral dissertation, Institute of Applied Mathematics, University of
Heidelberg, 2007.
13. Th. Dunne and R. Rannacher, Adaptive finite element simulation of fluid-structure interac-
tion based on an Eulerian variational formulation, in “Fluid-Structure Interaction: Modeling,
Simulation, Optimization” (H.-J. Bungartz and M. Schäfer, eds), pp. 371–386, Springer,
Berlin-Heidelberg, 2006.
14. Th. Dunne, R. Rannacher and T. Richter, Numerical simulation of fluid-structure interaction
based on monolithic variational formulations, in Series “Comtemporary Challenges in Math-
ematical Fluid Mechanics” (G.P. Galdi, R. Rannacher, eds.), World Scientific, Singapore, to
appear 2010.
15. K. Eriksson, D. Estep, P. Hansbo, and C. Johnson, Computational differential equations,
Cambridge University Press, Cambridge, 1996.
16. GASCOIGNE, A C++ numerics library for scientific computing, Institute of Applied Mathe-
matics, University of Heidelberg, URL http://www.gascoigne.uni-P2.de/.
17. V. Girault and P.-A. Raviart, Finite Element Methods for the Navier-Stokes Equations,
Springer: Berlin-Heidelberg-New York, 1986.
18. A. Griewank, On automatic differentiation in mathematical programming: recent developments
and applications, (M. Iri and K. Tanabe, eds), pp. 83–108, Kluwer Academic Publishers, 1989.
19. J. Heywood, R. Rannacher, and S. Turek, Artificial boundaries and flux and pressure conditions
for the incompressible Navier-Stokes equations, Int. J. Numer. Math. Fluids 22, 325–352, 1992.
20. J. Hron and S. Turek, Proposal for numerical benchmarking of fluid-structure interaction
between an elastic object and laminar incompressible flow, in “Fluid-Structure Interac-
tion: Modelling, Simulation, Optimisation” (H.-J. Bungartz and M. Schäfer, eds), Springer,
Berlin-Heidelberg, 2006.
21. J. Hron and S. Turek, A monolithic FEM/multigrid solver for an ALE formulation of fluid-
structure interaction with applications in biomechanics, in “Fluid-Structure Interaction: Mod-
elling, Simulation, Optimisation” (H.-J. Bungartz and M. Schäfer, eds), pp. 146–170, Springer,
Berlin-Heidelberg, 2006.
22. C. Liu and N. J. Walkington, An Eulerian description of fluids containing visco-elastic
particles, Arch. Rat. Mech. Anal. 159, 229–252, 2001.
23. D. Meidner und B. Vexler, Adaptive Space-Time Finite Element Methods for Parabolic
Optimization Problems, SIAM Journal on Control and Optimization 46(1), S. 116 142, 2007.
Adaptive FE method for FSI 191

24. L. B. Rall, Automatic Differentiation - Techniques and Application, Springer, New York, 1981.
25. S. Osher and J. A. Sethian, Propagation of fronts with curvature based speed: algorithms based
on Hamilton-Jacobi formulations, Journal of Computational Physics 79, 12, 1988.
26. R. Rannacher, Finite element methods for the incompressible Navier-Stokes equations, in
Fundamental Directions in Mathematical Fluid Mechanics (G. P. Galdi, J. Heywood, R. Ran-
nacher, eds), pp. 191–293, Birkhäuser, Basel-Boston-Berlin, 2000.
27. R. Rannacher and F.-T. Suttmeier Error estimation and adaptive mesh design for FE models
in elasto-plasticity, in Error-Controlled Adaptive FEMs in Solid Mechanics (E. Stein, ed.),
pp. 5–52, John Wiley, Chichster, 2002.
28. M. Schmich, Adaptive Finite Element Methods for computing nonstationary incompressible
flows, Doctoral dissertation, Institute of Applied Mathematics, University of Heidelberg, 2009.
29. M. Schmich and B. Vexler, Adaptivity with Dynamic Meshes for Space-Time Finite Element
Discretizations of Parabolic Equations, SIAM J. Sci. Comput., Vol. 30(1), pp. 369-393, 2008.
30. T. Richter and T. Wick, Finite Elements for Fluid-Structure Interaction in ALE and Fully Eule-
rian Coordinates, submitted to Computer Methods in Applied Mechanics and Engineering,
2009.
31. S. Turek, Efficient solvers for incompressible flow problems: an algorithmic and computational
approach, Springer, Heidelberg-Berlin-New York, 1999.
32. VISUSIMPLE, An open source interactive visualization utility for scientific computing, Insti-
tute of Applied Mathematics, University of Heidelberg, URL http://visusimple.
uni-P2.de/.
•
Numerical Simulation and Benchmarking
of a Monolithic Multigrid Solver
for Fluid-Structure Interaction Problems
with Application to Hemodynamics

S. Turek, J. Hron, M. Mádlı́k, M. Razzaq, H. Wobker, and J.F. Acker

Abstract An Arbitrary Lagrangian-Eulerian (ALE) formulation is applied in a

fully coupled monolithic way, considering the fluid-structure interaction (FSI) prob-
lem as one continuum. The mathematical description and the numerical schemes
are designed in such a way that general constitutive relations (which are realistic for
biomechanics applications) for the fluid as well as for the structural part can be eas-
ily incorporated. We utilize the LBB-stable finite element pairs Q2 P1 and P2C P1
for discretization in space to gain high accuracy and perform as time-stepping the
2nd order Crank-Nicholson, respectively, a new modified Fractional-Step--scheme
for both solid and fluid parts. The resulting discretized nonlinear algebraic system is
solved by a Newton method which approximates the Jacobian matrices by a divided
differences approach, and the resulting linear systems are solved by direct or itera-
tive solvers, preferably of Krylov-multigrid type.
For validation and evaluation of the accuracy and performance of the proposed
methodology, we present corresponding results for a new set of FSI benchmark con-
figurations which describe the self-induced elastic deformation of a beam attached
to a cylinder in laminar channel flow, allowing stationary as well as periodically
oscillating deformations. Then, as an example of FSI in biomedical problems, the
influence of endovascular stent implantation on cerebral aneurysm hemodynamics
is numerically investigated. The aim is to study the interaction of the elastic walls
of the aneurysm with the geometrical shape of the implanted stent structure for
prototypical 2D configurations. This study can be seen as a basic step towards the
understanding of the resulting complex flow phenomena so that in future aneurysm
rupture shall be suppressed by an optimal setting of the implanted stent geometry.

S. Turek, M. Razzaq, H. Wobker, and J. F. Acker

Institute for Applied Mathematics, TU Dortmund, Vogelpothsweg 87, 44227 Dortmund, Germany
e-mail: [email protected]
J. Hron and M. Mádlı́k
Mathematical Institute, Charles University Prague, Sokolovska 83, 18675 Prague, Czech republic
e-mail: [email protected]

H.-J. Bungartz et al. (eds.), Fluid Structure Interaction II, Lecture Notes 193
in Computational Science and Engineering 73, DOI 10.1007/978-3-642-14206-2 8,
c Springer-Verlag Berlin Heidelberg 2010
194 S. Turek et al.

1 Introduction

In this paper, we consider the general problem of viscous flow interacting with an
elastic body which is being deformed by the fluid action. Such a problem is of great
importance in many real life applications, typical examples are the areas of biomed-
ical fluids which include the influence of hemodynamic factors in blood vessels,
cerebral aneurysm hemodynamics, joint lubrication and deformable cartilage, and
blood flow interaction with elastic veins [2, 9, 25, 26, 34]. The theoretical investiga-
tion of fluid-structure interaction problems is complicated by the need of a mixed
description for both parts: While for the solid part the natural view is the material
(Lagrangian) description, for the fluid it is usually the spatial (Eulerian) description.
In the case of their combination some kind of mixed description (usually referred
to as the Arbitrary Lagrangian-Eulerian description or ALE) has to be used which
brings additional nonlinearity into the resulting equations (see [17]).
The numerical solution of the resulting equations of the fluid-structure interaction
problem poses great challenges since it includes the features of structural mechan-
ics, fluid dynamics, and their coupling. The most straightforward solution strategy,
mostly used in the available software packages (see for instance [16]), is to decouple
the problem into the fluid part and solid part, for each of those parts using some well
established solution method; then the interaction process is introduced as external
boundary conditions in each of the subproblems. This has the advantage that there
are many well tested numerical methods for both separate problems of fluid flow
and elastic deformation, while on the other hand the treatment of the interface and
the interaction is problematic due to high stiffness and sensitivity. In contrast, the
monolithic approach discussed here treats the problem as a single continuum with
the coupling automatically taken care of as internal interface. This on the other hand
requires more robust nonlinear and linear solvers for the global problem.
Besides a short description of the underlying numerical aspects regarding dis-
cretization and solution procedure for this monolithic approach (see [17, 24]), we
present corresponding results for a set of FSI benchmarking test cases (‘channel
flow around cylinder with attached elastic beam’, see [30]), and we concentrate
on prototypical numerical studies for 2D aneurysm configurations and first steps
towards full 3D models. The corresponding parameterization is based on abstrac-
tions of biomedical data (i.e., cutplanes of 3D specimens from New Zealand white
rabbits as well as computer tomographic and magnetic resonance imaging data of
human neurocrania). In our studies, we allow the walls of the aneurysm to be elastic
and hence deforming with the flow field in the vessel. Moreover, we examine several
configurations for stent geometries which clearly influence the flow behavior inside
the aneurysm such that a very different elastic displacement of the walls is observed,
too. We demonstrate that both the elastic modeling of the aneurysm walls as well
as the proper description of the geometrical details of the shape of the aneurysm
and particularly of the stents are of great importance for a quantitative analysis of
the complex interaction between structure and fluid. This is especially true in view
of more realistic blood flow models and anisotropic constitutive laws for the elastic
walls.
Numerical simulation and benchmarking of fluid-structure interaction 195

2 Fluid-structure interaction problem formulation

The general fluid-structure interaction problem consists of the description of the

fluid and solid parts, appropriate interface conditions at the interface and conditions
for the remaining boundaries, respectively. Here, we consider the flow of an incom-
pressible Newtonian fluid interacting with an elastic solid. We denote the domain
f
occupied by the fluid by ˝t and the solid part by ˝ts at the time t 2 Œ0; T . Let
t0 D ˝N t \ ˝N ts be the part of the boundary where the elastic solid interacts with
f

the fluid. In the following, the description for both fields and the interface conditions
are introduced. Furthermore, discretization aspects and computational methods are
described.

2.1 Fluid mechanics

The fluid flow is assumed to be laminar. It can be described by the Navier-Stokes

equations for incompressible flows

@vf f
f . C v  rv/  r   f D 0; r  v D 0 in ˝t ; (1)
@t

where f is the constant density. The state of the flow is described by the veloc-
ity and pressure fields vf , p f , respectively. The external forces, for example due
to gravity or human motion, are assumed to be not significant and are neglected.
Although the blood is known to be non-Newtonian in general, we assume it to be
Newtonian in this study. This is because we consider large arteries with radii of more
than 2 mm, where the velocity and shear rate are high and the kinematic viscosity
 f is nearly constant [20], such that the non-Newtonian effects can be neglected.
The constitutive relation for the stress tensor reads

 f D p f I C 2D.vf /; (2)

where  is the dynamic viscosity of the fluid, pf is the Lagrange multiplier cor-
responding to the incompressibility constraint in (1), and D.vf / is the strain-rate
tensor:
1
D.vf / D .rvf C .rvf /T /: (3)
2
For the fluid-structure interaction we use the ALE form of the balance equations.
The corresponding discretization techniques are discussed in Sect. 3. Let us remark
that also non-Newtonian flow models can be used for modeling blood flow, for
instance of Power Law type or even including viscoelastic effects (see [7]) which is
planned for future extensions.
196 S. Turek et al.

2.2 Structural mechanics

The governing equations for the structural mechanics are the balance equations

@vs
s . C .rvs /vs  g/  r   s D 0; in ˝ts ; (4)
@t

where the superscript s denotes the structure, s is the density of the material, gs
represents the external body forces acting on the structure, and  s is the Cauchy
stress tensor. The deformation of the structure is described by the displacement us ,
s
with velocity field vs D @u@t
. Written in the more common Lagrangian description,
i.e. with respect to some fixed reference (for example initial) state ˝ s , we have

@2 us
0s .  g/  r  ˙ s D 0; in ˝s; (5)
@t 2

where ˙ s D J  s FT is the first Piola-Kirchhoff stress tensor. J denotes the deter-
minant of the deformation gradient tensor F, defined as F D I C rus . Unlike the
Cauchy stress tensor  s , ˙ s is non-symmetric. Since constitutive relations are often
expressed in terms of symmetric stress tensor, it is natural to introduce the second
Piola-Kirchhoff tensor Ss

Ss D FT ˙ s D J F1  s FT ; (6)

which is symmetric. For elastic material the stress is a function of the deformation
(and possibly of thermodynamic variables such as the temperature) but it is inde-
pendent of deformation history and thus of time. The material characteristics may
still vary in space. In a homogeneous material mechanical properties do not vary, the
strain energy function depends only on the deformation. A material is mechanically
isotropic if its response to deformation is the same in all directions. The constitutive
equation is then a function of F. More precisely, it is usually written in terms of the
Green-Lagrange strain tensor, as

1
E D .C  I/; (7)
2

where I is the identity tensor and C D FT F is the left Cauchy-Green strain tensor.
For the subsequent FSI benchmark calculations we employ the St. Venant-
Kirchhoff material model as an example for homogeneous isotropic material whose
reference configuration is the natural state (i.e. where the Cauchy stress tensor
is zero everywhere). The St. Venant-Kirchhoff material model is specified by the
following constitutive law

1
s D F. s .tr E/I C 2s E/FT ; Ss D s .tr E/I C 2s E; (8)
J
Numerical simulation and benchmarking of fluid-structure interaction 197

where s denotes the first Lamé coefficient, and s the shear modulus. More com-
plex constitutive relations for hyperelastic materials may be found in [14], and
particular models for biological tissues and blood vessels are reported in [11, 15].
The material elasticity is characterized by a set of two parameters, the Poisson ratio
 s and the Young modulus E. These parameters satisfy the relations

s s .3 s C 22 /
s D ; ED ; (9)
2. s C s / . s C s /

E sE
s D ; s D ; (10)
2.1 C  s / .1 C  s /.1  2 s /

where  s D 1=2 for incompressible and  s < 1=2 for compressible material. In
the large deformation case it is common to describe the constitutive equation using
a stress-strain relation based on the Green Lagrange strain tensor E and the second
Piola-Kirchhoff stress tensor S.E/ as a function of E. However, also incompressible
materials can be handled in the same way (see [17]).
For the hemodynamic applications, a Neo-Hooke material model is taken which
can be used for compressible or incompressible (for  s ! 1=2 ) s ! 1 )
material and which is described by the constitutive laws:

s
 s D p s I C .FFT  I/; (11)
J
s 1
0 D p s C .J  /: (12)
2 J
Both models, the St. Venant-Kirchhoff and the Neo-Hooke material model, share
the isotropic and homogenous properties, and both can be used for the computation
of large deformations. However, the St. Venant-Kirchhoff model does not allow for
large strain computation, while the Neo-Hooke model is also valid for large strains.
In the case of small strains and small deformations, both material laws yield the
same linearized material model. We implemented the St. Venant-Kirchhoff mate-
rial model as the standard model for the compressible case, since the setup of the
benchmark does not involve large strains in the oscillating beam structure. Its imple-
mentation is simpler and, therefore, the FSI benchmark will hopefully be adopted
by a wider group of researchers. If someone wants or has to use the Neo-Hooke
material, the results for a given set of E and  or and  are comparable, if the
standard Neo-Hooke material model as in (11), (12) is used. Similarly as in the case
of more complex blood flow models, also more realistic constitutive relations for
the anisotropic behavior of the walls of aneurysms can be included which however
is beyond the scope of this paper.
198 S. Turek et al.

2.3 Interaction conditions

The boundary conditions on the fluid-solid interface are assumed to be

 f n D  s n; vf D vs ; on t0 ; (13)

where n is a unit normal vector to the interface t0 . This implies the no-slip
condition for the flow and that the forces on the interface are in balance.

3 Discretization and solution techniques

For the moment, we restrict our considerations to two dimensions which allows sys-
tematic tests of the proposed methods for biomedical applications in a very efficient
way such that the qualitatitive behavior can be carefully analyzed. The correspond-
ing fully implicit, monolithic treatment of the fluid-structure interaction problem
suggests that an A-stable second order time stepping scheme and that the same finite
elements for both the solid part and the fluid region should be utilized. Moreover, to
handle the fluid incompressibility constraints, we have to choose a stable finite ele-
ment pair. For that reason, the conforming biquadratic, discontinuous linear Q2 P1
pair is used. Let us define the usual finite dimensional spaces U for displacement,
V for velocity, P for pressure approximation as follows

U D fu 2 L1 .I; ŒW 1;2 .˝/2 /; u D 0 on @˝g;

V D fv 2 L2 .I; ŒW 1;2 .˝t /2 / \ L1 .I; ŒL2 .˝t /2 /; v D 0 on @˝g;
P D fp 2 L2 .I; L2 .˝//g:

Then the variational formulation of the fluid-structure interaction problem is to find

.u; v; p/ 2 U  V  P that satisfy the corresponding weak form of the balance
equations including appropriate initial conditions. The spaces U; V; P on an interval
Œt n ; t nC1  would be approximated in the case of the Q2 ; P1 pair as

Uh D fuh 2 ŒC.˝h /2 ; uh jT 2 ŒQ2 .T /2 8T 2 Th ; uh D 0 on @˝h g;

2 2
Vh D fvh 2 ŒC.˝h / ; vh jT 2 ŒQ2 .T / 8T 2 Th ; vh D 0 on @˝h g;
Ph D fph 2 L .˝h /; ph jT 2 P1 .T /
2
8T 2 Th g:

Let us denote by unh the approximation of u.t n /, vnh the approximation of v.t n /
and phn the approximation of p.t n /. Consider for each T 2 Th the bilinear
transformation T W TO ! T , where TO is the unit square. Then, Q2 .T / is defined as
˚ 1

Q2 .T / D q ı T W q 2 span < 1; x; y; xy; x 2 ; y 2 ; x 2 y; y 2 x; x 2 y 2 > ; (14)
Numerical simulation and benchmarking of fluid-structure interaction 199

with nine local degrees of freedom located at the vertices, midpoints of the edges
and in the center of the quadrilateral. The space P1 .T / consists of linear functions
defined by ˚ 
P1 .T / D q ı T1 W q 2 span < 1; x; y > ; (15)
with the function value and both partial derivatives located in the center of the
quadrilateral, as its three local degrees of freedom, which leads to a discontinuous
pressure. The inf-sup condition is satisfied (see [5]); however, the combination of
the bilinear transformation with a linear function on the reference square P1 .T b/
would imply that the basis on the reference square did not contain the full basis. So,
the method can at most be first order accurate on general meshes (see [3, 5])

kp  ph k0 D O.h/: (16)

The standard remedy is to consider a local coordinate system .

; / obtained by
joining the midpoints of the opposing faces of T (see [3, 22, 29]). Then, we set on
each element T
P1 .T / WD span < 1;
;  > : (17)
For this case, the inf-sup condition is also satisfied and the second order approx-
imation is recovered for the pressure as well as for the velocity gradient (see
[5, 13])
kp  ph k0 D O.h2 / and kr.u  uh /k0 D O.h2 /: (18)
For a smooth solution, the approximation error for the velocity in the L2 -norm is
of order O.h3 / which can easily be demonstrated for prescribed polynomials or for
smooth data on appropriate domains.
In the last section we present first results for 3-dimensional computation where
we use the P2C P1 pair which also satisfies the Babuška–Brezzi stability condition
and yields a stable discretization of the incompressible problems (see [10]).

3.1 Time discretization

In view of a more compact presentation, the applied time discretization approach

is described only for the fluid part (see [23] for more details). In the following, we
restrict to the (standard) incompressible Navier-Stokes equations

vt  v C v  rv C rp D f; r  v D 0; in ˝  .0; T  ; (19)

for given force f and viscosity , with prescribed boundary values on the boundary
@˝ and an initial condition at t D 0.
200 S. Turek et al.

3.1.1 Basic--scheme

The basic -scheme for time discretization reads:

Given vn and t D tnC1  tn , then solve for v D vnC1 and p D p nC1

v  vn
C Œv C v  rv C rp D gnC1 ; divv D 0; in ˝ (20)
t

with right hand side gnC1 WD fnC1 C .1  /fn  .1  /Œvn C vn  rvn . The
parameter  has to be chosen depending on the time-stepping scheme, e.g.,  D 1
for the Backward Euler (BE), or  D 1=2 for the Crank-Nicholson-scheme (CN)
which we prefer. The pressure term rp D rp nC1 may be replaced by rp nC1 C
.1  /rp n , but with appropriate postprocessing, both strategies lead to solutions
of the same accuracy. In all cases, we end up with the task of solving, at each time
step, a nonlinear saddle point problem of given type which has then to be discretized
in space as described above.
These two methods, CN and BE, belong to the group of One-Step--schemes.
The CN scheme can occasionally suffer from numerical instabilities because of its
only weak damping property (not strongly A-stable), while the BE-scheme is of
first order accuracy only (however: it is a good candidate for steady-state simu-
lations). Another method which has proven to have the potential to excel in this
competition is the Fractional-Step--scheme (FS). It uses three different values for
 and for the time step t at each time level. In [24, 31] we additionally introduced
a modified Fractional-Step--scheme which particularly for fluid-structure interac-
tion problems seems to be advantageous. A brief description is given below and a
detailed description will appear in the thesis [23].
In the following, we use the more compact form for the diffusive and advec-
tive part:
N.v/v D v C v  rv : (21)

3.1.2 Fractional-Step--scheme
p
For the Fractional-Step--scheme we proceed as follows. Choosing  D 1  22
 0 D 1  2, and ˛ D 121
; ˇ D 1  ˛, the macro time step tn ! tnC1 D tn C t
is split into the three following consecutive substeps (with Q WD ˛t D ˇ 0 t):

ŒI C Q N.vnC /vnC C rp nC D ŒI  ˇtN.vn /vn C tfn ;

divvnC D 0;

Q
ŒI C N.vnC1
/vnC1 C rp nC1 D ŒI  ˛ 0 tN.vnC /vnC
C 0 tfnC1 ;
divvnC1 D 0;
Numerical simulation and benchmarking of fluid-structure interaction 201

ŒI C Q N.vnC1 /vnC1 C rp nC1 D ŒI  ˇtN.vnC1 /vnC1

CtfnC1 ;
divv nC1
D 0:

3.1.3 A modified Fractional-Step--scheme

Consider an initial value problem of the following form, with X.t/ 2 Rd ; d  1:

8
< dX D f .X; t/ 8t > 0;
dt (22)
:
X.0/ D X0 :

Then, a modified -scheme (see [31]) with macro p time step t can be written again
as three consecutive substeps, where  D 1  1= 2, X 0 D X0 , n  0 and X n is
known:
X nC  X n  
D f X nC ; t nC ;
t
1   nC 2  1 n
X nC1 D X C X ;
 
X nC1  X nC1  
D f X nC1 ; t nC1 :
t
As shown in [31], the ‘classical’ and the modified Fractional-Step--schemes are
 fully implicit,
 strongly A-stable, and
 second order accurate (in fact, they are ‘nearly’ third order accurate [31]).
These important properties promise some advantageous behavior, particularly in
implicit CFD simulations for nonstationary incompressible flow problems. Apply-
ing one step of this scheme to the Navier-Stokes equations, we obtain the following
variant of the scheme:
8 nC
ˆ
ˆ v  vn
< C N.vnC /vnC C rp nC D fnC ;
1: t
ˆ
:̂
divvnC D 0:

2 1 n
2: vnC1 D 1 nC

v C 
v :
8 nC1
ˆ
ˆ v  vnC1
< C N.vnC1 /vnC1 C re
p nC1 D fnC1 ;
3: t
ˆ
:̂
divvnC1 D 0:

4: p nC1 D .1  /p nC C e

p nC1 :
202 S. Turek et al.

These 3 substeps build one macro time step and have to be compared with
the previous description of the Backward Euler, Crank-Nicholson and the classi-
cal Fractional-Step--scheme which all can be formulated in terms of a macro time
step with 3 substeps, too. Then, the resulting accuracy and numerical cost are better
comparable and the rating is fair. The main difference to the previous ‘classical’
Fractional-Step--scheme is that substeps 1. and 3. look like a Backward Euler step
while substep 2. is an extrapolation step only for previously computed data such that
no operator evaluations at previous time steps are required.
Substep 3b. can be viewed as postprocessing step for updating the new pressure
which however is not mandatory. In fact, in our numerical tests [31] we omit-
ted this substep 3b. and accepted the pressure from substep 3. as final pressure
approximation, that means p nC1 D e p nC1 .
Summarizing, the numerical effort of the modified scheme for each substep is
cheaper at least for ‘small’ time steps (treatment of the nonlinearity) and complex
right hand side evaluations, while the resulting accuracy is similar. The modified
-scheme is a Runge-Kutta scheme; it has been derived in [12] as a particular case
of the Fractional-Step--scheme.

3.2 Solution algorithms

After applying the standard finite element method, the system of nonlinear algebraic
equations arising from the governing equations described in Sects. 2.1 and 2.2 reads
(for incompressible solid material)
0 10 1 0 1
Suu Suv 0 u fu
B C
@ Svu Svv kB A @ v A D @ fv A ; (23)
cu BsT cv BfT 0 p fp

which is a typical saddle point problem, where S describes the diffusive and convec-
tive terms from the governing equations. The above system of nonlinear algebraic
equations (23) is solved using the Newton method as basic iteration which can
exhibit quadratic convergence provided that the initial guess is sufficiently close
to the solution. The basic idea of the Newton iteration is to find a root of a function,
R.X/ D 0, using the available known function value and its first derivative. One
step of the Newton iteration can be written as
 1
@R.Xn /
XnC1 D Xn C ! n R.Xn /; (24)
@X

/n
where X D .uh ; vh ; ph / and @R.X
@X
is the Jacobian matrix. To ensure the conver-
gence globally, some improvements of this basic iteration are used. The damped
Newton method with line search improves the chance of convergence by adaptively
Numerical simulation and benchmarking of fluid-structure interaction 203

changing the length of the correction vector (see [17, 29] for more details). The
damping parameter ! n 2 .1; 0/ is chosen such that

R.XnC1 /  XnC1  R.Xn /  Xn : (25)

The damping greatly improves the robustness of the Newton iteration in the case
when the current approximation Xn is not close enough to the final solution. The
n/
Jacobian matrix @R.X
@X
can be computed by finite differences from the residual
vector R.X/
 
@R.Xn / ŒRi .Xn C ˛j ej /  ŒRi .Xn  ˛j ej /
 ; (26)
@X ij 2˛j

where ej are the unit basis vectors in Rn and the coefficients ˛j are adaptively cho-
sen according to the change in the solution in the previous time step. Since we know
the sparsity pattern of the Jacobian matrix in advance, which is given by the used
finite element method, this computation can be done in an efficient way such that
the linear solver remains the dominant part in terms of the CPU time (see [29, 33]).
A good candidate, at least in 2D, seems to be a direct solver for sparse systems
like UMFPACK [8] or MUMPS [1]; while this choice provides very robust linear
solvers, its memory and CPU time requirements are too high for larger systems (i.e.
more than 20 000 unknowns). Large linear problems can be solved by Krylov-space
methods (BiCGStab, GMRes [4]) with suitable preconditioners. One possibility is
the ILU preconditioner with special treatment of the saddle point character of our
system, where we allow certain fill-in for the zero diagonal blocks, see [6]. As an
alternative, we also utilize a standard geometric multigrid approach based on a hier-
archy of grids obtained by successive regular refinement of a given coarse mesh. The
complete multigrid iteration is performed in the standard defect-correction setup
with the V or F-type cycle. While a direct sparse solver [8] is used for the coarse
grid solution, on finer levels a fixed number (2 or 4) of iterations by local MPSC
schemes (Vanka-like smoother) [17, 29, 35] is performed. Such iterations can be
written as
0 1 0 l1 0 11 0 1
ulC1 u X B Suuj˝i Suvj˝i 0 deflu
@ vlC1 A D @ vl A  ! C B l C
@ Svuj˝i Svvj˝i kBj˝i A @ defv A :
p lC1 pl element˝i cu B T c BT 0 def lp
sj˝i v f j˝i

The inverse of the local (39  39) systems can be computed by hardware optimized
direct solvers. The full nodal interpolation is used as the prolongation operator P
with its transposed operator used as the restriction R D PT , see [16, 29] for more
details.
204 S. Turek et al.

4 FSI benchmarking

In order to validate and to analyze different techniques to solve such FSI problems,
also in a quantitative way, a set of benchmark configurations has been proposed
in [30] (also see the contribution in this volume). The configurations consist of
laminar incompressible channel flow around an elastic object which results in self-
induced oscillations of the structure. Moreover, characteristic flow quantities and
corresponding plots are provided for a quantitative comparison.
The domain is based on the 2D version of the well-known CFD benchmark in
[32] and by omitting the elastic bar behind the cylinder one can easily recover the
setup of the ‘classical’ flow around cylinder configuration which allows for vali-
dation of the flow part by comparing the results with the older flow benchmark.
The setting is intentionally nonsymmetric to prevent the dependence of the onset of
any possible oscillation on the precision of the computation. The mesh used for the
computations is shown in Fig. 1. A parabolic velocity profile is prescribed at the left
channel inflow

y.H  y/ 4:0
vf .0; y/ D 1:5UN  H 2 D 1:5UN 0:1681 y.0:41  y/; (27)
2

such that the mean inflow velocity is UN and the maximum of the inflow velocity pro-
file is 1:5UN . The no-slip condition is prescribed for the fluid on the other boundary
parts. i.e. top and bottom wall, circle and fluid-structure interface t0 . The outflow
condition can be chosen by the user, for example stress free or do nothing con-
ditions. The outflow condition effectively prescribes some reference value for the

level #refine #el #dof

0 0 62 1338
1 1 248 5032
2 2 992 19488
3 3 3968 76672
4 4 15872 304128
5 5 63488 1211392
6 6 253952 4835328
7 7 1015808 19320832

Fig. 1 Coarse mesh with number of degrees of freedom for refined levels.
Numerical simulation and benchmarking of fluid-structure interaction 205

pressure variable p. While this value could be arbitrarily set in the incompressible
case, in the case of compressible structure this will have influence on the stress and
consequently the deformation of the solid. In this description, we set the reference
pressure at the outflow to have zero mean value. Suggested starting procedure for
the non-steady tests is to use a smooth increase of the velocity profile in time as
( 1cos. 
2 t/
f vf .0; y/ if t < 2:0;
v .t; 0; y/ D 2 (28)
f
v .0; y/ otherwise,

where vf .0; y/ is the velocity profile given in (27). The following FSI tests are per-
formed for three different inflow speeds. FSI1 is resulting in a steady state solution,
while FSI2 and FSI3 result in periodic solutions. The parameter values for the FSI1,
FSI2 and FSI3 are given in the Table 1. Here, the computed values are summarized
in Table 2 for the steady state test FSI1. In Figs. 2 and 3, plots of the resulting x- and
y-displacements of the trailing edge point A (see [30]) of the elastic bar and plots
of the forces (lift, drag) acting on the cylinder and the bar are drawn. Furthermore,
computed values for three different mesh refinement levels and two different time
steps for the nonsteady tests FSI2 and FSI3 are presented respectively, which show
the (almost) grid independent solution behavior (for more details see [30]).

Table 1 Parameter settings for the FSI benchmarks.

parameter FSI1 FSI2 FSI3
kg parameter FSI1 FSI2 FSI3
s [103 m3 ] 1 10 1 s
s ˇ D f 1 10 1
 0.4 0.4 0.4
kg
s [106 ms2 ] 0:5 0:5 2:0 s 0.4 0.4 0.4
s
kg Ae D fEUN 2 3:5  104 1:4  103 1:4  103
f [103 m3 ] 1 1 1 N
2 Re D U fd 20 100 200
 f [103 ms ] 1 1 1
UN 0.2 1 2
UN [ ms ] 0.2 1 2

Table 2 Results for FSI1.

level nel ndof ux of A [103 m] uy of A [103 m] drag [N] lift [N]
2 992 19488 0:02287080 0:8193038 14:27359 0:7617550
3 3968 76672 0:02277423 0:8204231 14:29177 0:7630484
4 15872 304128 0:02273175 0:8207084 14:29484 0:7635608
5 63488 1211392 0:02271553 0:8208126 14:29486 0:7636992
6 253952 4835328 0:02270838 0:8208548 14:29451 0:7637359
7 1015808 19320832 0:02270493 0:8208773 14:29426 0:7637460
ref. 0:0227 0:8209 14:294 0:7637
206 S. Turek et al.

FSI2: x & y displacement of the point A [m].

0
fsi2 0.1
fsi2
-0.005 0.08
displacement x

0.06

displacement y
-0.01 0.04
-0.015 0.02
0
-0.02 -0.02
-0.04
-0.025
-0.06
-0.03 -0.08
34 34.2 34.4 34.6 34.8 35 34 34.2 34.4 34.6 34.8 35
time time

FSI2: lift and drag force [N] on the cylinder+elastic bar.

250 300
200 fsi2 fsi2
280
150 260
100 240
50 220
drag
lift

0
200
-50
-100 180
-150 160
-200 140
-250 120
34 34.2 34.4 34.6 34.8 35 34 34.2 34.4 34.6 34.8 35
time time

lev. ux of A [103 m] uy of A [103 m] drag [N] lift [N]

2 14:02 ˙ 12:03Œ3:85 1:25 ˙ 79:3Œ1:93 210:10 ˙ 72:62Œ3:85 0:25 ˙ 227:9Œ1:93
3 14:54 ˙ 12:50Œ3:86 1:25 ˙ 80:7Œ1:93 212:83 ˙ 75:89Œ3:86 0:92 ˙ 234:3Œ1:93
4 14:88 ˙ 12:75Œ3:86 1:24 ˙ 81:7Œ1:93 215:06 ˙ 77:76Œ3:86 0:82 ˙ 237:1Œ1:93
lev. ux of A [103 m] uy of A [103 m] drag [N] lift [N]
2 14:01 ˙ 12:04Œ3:86 1:25 ˙ 79:3Œ1:93 210:09 ˙ 72:82Œ3:86 0:52 ˙ 228:6Œ1:93
3 14:54 ˙ 12:48Œ3:86 1:25 ˙ 80:7Œ1:93 213:06 ˙ 75:76Œ3:86 0:85 ˙ 234:4Œ1:93
4 14:87 ˙ 12:73Œ3:86 1:24 ˙ 81:7Œ1:93 215:18 ˙ 77:78Œ3:86 0:87 ˙ 238:0Œ1:93
lev. ux of A [103 m] uy of A [103 m] drag [N] lift [N]
2 14:01 ˙ 12:04Œ3:86 1:28 ˙ 79:2Œ1:93 210:14 ˙ 72:86Œ3:86 0:49 ˙ 228:7Œ1:93
3 14:48 ˙ 12:45Œ3:86 1:24 ˙ 80:7Œ1:93 213:05 ˙ 75:74Œ3:86 0:84 ˙ 234:8Œ1:93
4 14:85 ˙ 12:70Œ3:86 1:30 ˙ 81:6Œ1:93 215:06 ˙ 77:65Œ3:86 0:61 ˙ 237:8Œ1:93
ref. 14:85 ˙ 12:70Œ3:86 1:30 ˙ 81:7Œ1:93 215:06 ˙ 77:65Œ3:86 0:61 ˙ 237:8Œ1:93

Fig. 2 Results for FSI2 with time step t D 0:002; t D 0:001; t D 0:0005 [s].
Numerical simulation and benchmarking of fluid-structure interaction 207

FSI3: x & y displacement of the point A [m].

0 0.04
fsi3 fsi3
-0.001 0.03
displacement x

displacement y
0.02
-0.002
0.01
-0.003 0
-0.004 -0.01
-0.02
-0.005
-0.03
-0.006 -0.04
19.5 19.6 19.7 19.8 19.9 20 19.5 19.6 19.7 19.8 19.9 20
time time

FSI3: lift and drag force [N] on the cylinder+elastic bar.

200 485
fsi3 480 fsi3
150 475
100 470
465
50
drag

460
lift

0 455
450
-50 445
-100 440
435
-150 430
19.5 19.6 19.7 19.8 19.9 20 19.5 19.6 19.7 19.8 19.9 20
time time

lev. ux of A [103 m] uy of A [103 m] drag [N] lift [N]

2 3:02 ˙ 2:83Œ10:75 1:41 ˙ 35:47Œ5:37 458:2 ˙ 28:32Œ10:75 2:41 ˙ 145:58Œ5:37
3 2:78 ˙ 2:62Œ10:93 1:44 ˙ 34:36Œ5:46 459:1 ˙ 26:63Œ10:93 2:41 ˙ 151:26Œ5:46
4 2:86 ˙ 2:70Œ10:95 1:45 ˙ 34:93Œ5:47 460:2 ˙ 27:65Œ10:95 2:47 ˙ 154:87Œ5:47
lev. ux of A [103 m] uy of A [103 m] drag [N] lift [N]
2 3:02 ˙ 2:85Œ10:75 1:42 ˙ 35:63Œ5:37 458:7 ˙ 28:78Œ10:75 2:23 ˙ 146:02Œ5:37
3 2:78 ˙ 2:62Œ10:92 1:44 ˙ 34:35Œ5:46 459:1 ˙ 26:62Œ10:92 2:39 ˙ 150:68Œ5:46
4 2:86 ˙ 2:70Œ10:92 1:45 ˙ 34:90Œ5:46 460:2 ˙ 27:47Œ10:92 2:37 ˙ 153:75Œ5:46
lev. ux of A [103 m] uy of A [103 m] drag [N] lift [N]
2 3:02 ˙ 2:85Œ10:74 1:32 ˙ 35:73Œ5:36 458:7 ˙ 28:80Œ10:74 2:23 ˙ 146:00Œ5:36
3 2:77 ˙ 2:61Œ10:93 1:43 ˙ 34:43Œ5:46 459:1 ˙ 26:50Œ10:93 2:36 ˙ 149:91Œ5:46
4 2:88 ˙ 2:72Œ10:93 1:47 ˙ 34:99Œ5:46 460:5 ˙ 27:74Œ10:93 2:50 ˙ 153:91Œ5:46
ref. 2:88 ˙ 2:72Œ10:93 1:47 ˙ 34:99Œ5:46 460:5 ˙ 27:74Œ10:93 2:50 ˙ 153:91Œ5:46

Fig. 3 Results for FSI3 with time step t D 0:001; t D 0:0005; t D 0:00025 [s].
208 S. Turek et al.

5 FSI Optimization benchmarking

The objective of the following benchmarking scenario is to extend the validated

FSI benchmark configurations towards optimization problems such that minimal
drag/lift values of the elastic object, minimal pressure loss or minimal nonstationary
oscillations through boundary control of the inflow, change of geometry or optimal
control of volume forces can be achieved. The main design aim for the subsequent
fluid structure interaction optimization problem is to minimize the lift on the beam
with the help of boundary control of the inflow data. Here, the simulation is based on
the described FSI1 configuration. Further extensions of this optimization problem
will be to control minimal pressure loss or minimal nonstationary oscillations of the
elastic beam through boundary control of the inflow section, change of the geome-
try (elastic channel walls or length/thickness of elastic beam) or optimal control of
volume forces.
As described before, the position of the beam is not symmetric such that the lift
is not zero. To minimize the lift value, we allow additional parabolic inflow at the
top and additional parabolic outflow at the bottom of the domain. The location of
the additional inlet and outlet is shown in the schematic diagram of the geometry in
Fig. 4. With V1 and V2 we denote the magnitude of the additional inflow and outflow
velocity, respectively. Then, the aim is to
 
minimize lift2 C ˛.V12 C V22 /
w.r.t. V1 ; V2 :

The arbitrary but fixed parameter ˛ > 0 reflects the ‘costs’ of the additional in-
and outflow and has to be prescribed. The domain is based on the 2D version of the
described FSI benchmark, shown in Fig. 4; however, the thickness of the beam is
increased from 0:02 to 0:04. Fluid and structural parameter values are based on the
FSI1 benchmark. Table 3 provides an overview of the geometry parameters.

Fig. 4 Geometry and configuration of the problem.

Numerical simulation and benchmarking of fluid-structure interaction 209

Table 3 Overview of the geometry parameters.

geometry parameters value [m]
channel length L 2:5
channel width H 0:41
cylinder center position C .0:2; 0:2/
cylinder radius r 0:05
elastic structure length l 0:35
elastic structure thickness h 0:04

Table 4 Results for FSI-Optimization.

level 1 level 2
˛ iter optimal values .V1 ; V2 / lift iter optimal values .V1 ; V2 / lift
1e+0 57 (3.74e-1, 3.88e-1) 8.1904e-1 59 (3.66e-1, 3.79e-1) 7.8497e-1
1e-2 60 (1.04e+0, 1.06e+0) 2.2684e-2 59 (1.02e+0, 1.04e+0) 2.1755e-2
1e-4 73 (1.06e+0, 1.08e+1) 2.3092e-4 71 (1.04e+0, 1.05e+01) 2.2147e-4
1e-6 81 (1.06e+0, 1.08e+1) 2.3096e-6 86 (1.04e+0, 1.05e+01) 2.2151e-6

Fig. 5 No displacement is visible of the beam due to optimal boundary flow control: Level 1 (left),
Level 2 (right).

Optimal points are those .V1 ; V2 / values which result in minimal lift values on
the beam depending on the parameter ˛. As ˛ decreases we get the reduction of the
lift on the beam, and the optimal values read .1:06; 10:8/ for level 1 and .1:04; 10:5/
for level 2 using the simplex algorithm proposed by Nelder and Mead [21]. Results
are given in Table 4 which show the optimal values for the velocities V1 and V2
providing also the resulting lift on the beam as compared with the FSI1 benchmark
values in Table 2. In Fig. 5, it is visible that the displacement of the beam decreases
with decreasing ˛, as well as the lift value decreases due to the boundary control.
Keep in mind that the original lift on the beam is approximately 7.6e-1 for the FSI1
benchmark while in the case of FSI1-Optimization, it reduces to approximately
2.3e-6 for ˛=1e-6. Interested readers are referred to [18] for a comprehensive sur-
vey of the original Nelder-Mead simplex algorithm and for its advantages and dis-
advantages.
210 S. Turek et al.

6 Applications to hemodynamics

In the following, we consider the numerical simulation of special problems encoun-

tered in the area of cardiovascular hemodynamics, namely flow interaction with
thick-walled deformable material (here: the arterial walls) and rigid parts (here:
stents), which can become a useful tool for deeper understanding of the onset
of diseases of the human circulatory system, as for example blood cell and inti-
mal damages in stenosis, aneurysm rupture, evaluation of new surgery techniques
of the heart, arteries and veins (see [2, 19, 34] and the literature cited therein).
In this contribution, prototypical studies are performed for brain aneurysms. The
word ‘aneurysm’ comes from the latin word aneurysma which means dilatation. An
aneurysm is a local dilatation in the wall of a blood vessel, usually an artery, due
to a defect, disease or injury. Typically, as the aneurysm enlarges, the arterial wall
becomes thinner and eventually leaks or ruptures, causing subarachnoid hemorrhage
(SAH) (bleeding into brain fluid) or formation of a blood clot within the brain. In
the case of a vessel rupture, there is a hemorrhage, which is particularly rapid and
intense in case of an artery. In arteries the wall thickness can be up to 30% of the
diameter and its local thickening can lead to the creation of an aneurysm. The aim
of numerical simulations is to relate the aneurysm state (unrupture or rupture) with
wall pressure, wall deformation and effective wall stress. Such a relationship would
provide information for the diagnosis and treatment of unruptured and ruptured
aneurysms by elucidating the risk of bleeding or rebleeding, respectively.
As a typical example for the related CFD simulations, a real view is provided
in Fig. 6 which also contains the automatically extracted computational domain and
(coarse) mesh in 2D, however without stents. In order to use the proposed numerical
methods for aneurysm hemodynamics, in a first step only simplified two-dimension-
al examples, which however include the interaction of the flow with the deformable
material, are considered. Flow through a deformable vein with elastic walls of a
brain aneurysm is simulated to analyze qualitatively the described methods; here,
the flow is driven by prescribing the flow velocity at the inflow section (Poiseuille
flow) while the solid part of the boundary is either fixed or stress-free. Both ends of
the walls are fixed, and the flow is driven by a periodical change of the inflow at the
right end.

Fig. 6 Left: Real view of aneurysm. Right: Schematic drawing of the mesh.
Numerical simulation and benchmarking of fluid-structure interaction 211

Fig. 7 Left: Schematic drawing of the measurement section. Middle: Mesh without stents (776
elements). Right: Mesh with stents (1431 elements) which are part of the simulations.

6.1 Geometry of the problem

For convenience, the geometry of the fluid domain under consideration is currently
based on simplified 2D models, see Fig. 7, which allows to concentrate on the
detailed qualitative evaluation of our approach based on the described monolithic
ALE formulation. The underlying construction of the (2D) shape of the aneurysm
can be explained as follows:
 The bent blood vessel is approximated by quarter circles around the origin.
 The innermost circle has the radius 6mm, the next has 8mm, and the last one has
8:25mm.
 This results in one rigid inner wall and an elastic wall between 8mm and 8:25mm
of thickness 0:25mm.
The aneurysm shape is approximated by two arcs and lines intersecting the arcs
tangentially. The midpoints of the arcs are the same .6:75I 6/, they have the radius
1:125mm and 1:25mm. They are intersected tangentially by lines at angular value
1:3 radians. This results in a wall thickness of 0:125mm for the elastic aneurysm
walls (see Fig. 7). The examined stents are of circular shape, placed on the neck of
the aneurysm, and we use three, respectively five stents (simplified ‘circles’ in 2D
as cutplanes from 3D configurations) of different size and position. Such stents (in
real life) are typically used to keep arteries open and are located on the vessel wall
while ‘our’ (2D) stent is located only near the aneurysm (Fig. 7). The purpose of
this device is to reduce the flux into and within the aneurysm in order to occlude it
by a clot or rupture.

6.2 Boundary and initial conditions

The (steady) velocity profile, to flow from the right to the left part of the channel, is
defined as parabolic inflow, namely

vf .0; y/ D UN .y  6/.y  8/: (29)

212 S. Turek et al.

Correspondingly, the pulsatile inflow profile for the nonsteady tests for which peak
systole and diastole occur for t D 0:25s and t D 0:75s respectively, is
prescribed as
vf .t; 0; y/ D vf .0; y/.1 C 0:75si n.2 t//: (30)
The natural outflow condition at the lower left part effectively prescribes some ref-
erence value for the pressure variable p, here p D 0. While this value could be
arbitrarily set in the incompressible case, in the case of a compressible structure this
might have influence on the stress and consequently the deformation of the solid.
The no-slip condition is prescribed for the fluid on the other boundary parts, i.e. top
and bottom wall, stents and fluid-structure interface.

6.3 Numerical results

The Newtonian fluid used in the tests has a density f D 1:035  106kg=mm3
and a kinematic viscosity  f D 3:38mm2=s which is similar to the properties of
blood. If we prescribe the inflow speed UN D 50mm=s, this results in a Reynolds
number Re  120 based on the prescribed peak systole inflow velocity and the
width of the veins which is 2mm such that the resulting flow is within the laminar
region. Parameter values for the elastic vein in the described model are as follows:
The density of the upper elastic wall is s = 1:12  106 kg=mm3 , solid shear
modulus is s D 42:85kg=mms 2 , Poisson ratio is  s D 0:4, Young modulus is
E D 120kN=mm2. As described before, the constitutive relations used for the
materials are the incompressible Newtonian model (2) for the fluid and a hypere-
lastic Neo-Hooke material for the solid. This choice includes most of the typical
difficulties the numerical method has to deal with, namely the incompressibility and
significant deformations.
From a medical point of view, the use of stents provides an efficient treatment for
managing the difficult entity of intracranial aneurysms. Here, the thickness of the
aneurysm wall is attenuated and the aneurysm hemodynamics changes significantly.
Since the purpose of this device is to control the flux within the aneurysm in order
to occlude it by a clot or rupture, the resulting flow behavior into and within the
aneurysm is the main objective, particularly in view of the different stent geome-
tries. Therefore, we decided for the 2D studies to locate the stents only in direct
connection to the aneurysm.
Comparing our studies with the CFD literature (see [2, 9, 27, 28, 34]), several
research groups focus on CFD simulations with realistic 3D geometries, but typi-
cally assuming rigid walls. In contrast, we concentrate on the complex interaction
between elastic deformations and flow perturbations induced by the stents. At the
moment, we are only able to perform these simulations in 2D. However, with these
studies we should be able to analyze qualitatively the influence of geometrical
details onto the elastic material behavior, particularly in view of more complex
blood models and constitutive equations for the structure. Therefore, the aims of
our current studies can be described as follows:
Numerical simulation and benchmarking of fluid-structure interaction 213

1. What is the influence of the elasticity of the walls onto the flow behavior inside
the aneurysm, particularly w.r.t. the resulting shape of the aneurysm?
2. What is the influence of the geometrical details of the (2D) stents, that means
shape, size, position, on the flow into and inside the aneurysm?
3. Do both aspects, small-scale geometrical details as well as elastic fluid-structure
interaction, have to be considered simultaneously or is one of them negligible in
first order approximation?
4. Are modern numerical methods and corresponding CFD simulation tools able to
simulate qualitatively the multiphysics behavior of such biomedical configura-
tions?
In the following, we show some corresponding results for the described prototypical
aneurysm geometry, first for the steady state inflow profile, followed by nonsteady
tests for the pulsatile inflow, both with rigid and elastic walls, respectively.

6.3.1 Steady configurations

Due to the given inflow profile, which is not time-dependent, and due to the low
Reynolds numbers, the flow behavior leads to a steady state which only depends on
the elasticity and the shape of the stents. Moreover, for the following simulations,
we only treat the aneurysm wall as elastic structure. Then, the aneurysm undergoes
some slight deformations which can hardly be seen in the following figures. How-
ever, they result in a different volume of the flow domain (see Fig. 9) and lead to
a significantly different local flow behavior since the spacing between stents and
elastic walls may change (see Fig. 8).
In Fig. 11, we visualize the different flow behavior by coloring corresponding
to the velocity magnitude and by showing corresponding vector plots inside the
aneurysm. Particularly the influence of the number of stents on the complete fluid
flow through the channel including the aneurysm can be clearly seen.
Summarizing these results for steady inflow, the simulations show that the stent
implantation across the neck of the aneurysm prevents blood penetration into the
aneurysm fundus. Moreover, the elastic geometrical deformation of the wall is
slightly reduced by implanting the stents while the local flow behavior inside the
aneurysm is more significantly influenced by the elastic properties of the outer wall,
particularly due to different width between stents and walls of the aneurysm. In the
next section, we will consider the behaviour of more realistic flow configurations
with time-dependent pulsatile inflow which will be analyzed for the case of elastic
behaviour of the aneurysm walls.

6.3.2 Pulsatile configurations

For the following pulsatile test case, we have taken again the aneurysm part as elastic
while the other parts of the walls belonging to the channel are rigid. First of all, we
show again (see Fig. 10) the resulting volume of the flow domain for five, three and
214 S. Turek et al.

Fig. 8 Magnitude of the blood flow velocity for four configurations. Top to bottom: Rigid walls
without stents, elastic walls without stents, rigid walls with stents, elastic walls with stents. Left:
Overall view. Right: Scaled view of the aneurysm.
Numerical simulation and benchmarking of fluid-structure interaction 215

26.65
no stents, elastic fundus
no stents, rigid walls
3 stents, elastic fundus
5 stents, elastic fundus
26.6

26.55
volume

26.5

26.45

26.4

26.35
1 2 3 4 5 6 7 8
time step

Fig. 9 Resulting volume of the fluid domain for different configurations.

no stents. In all cases, the oscillating behavior due to the pulsatile inflow is visible
which also leads to different volume sizes. Looking carefully at the resulting flow
behavior, we see differences w.r.t. the channel flow near the aneurysm, namely due
to the different flow rate into the aneurysm, and significant local differences inside
the aneurysm.

6.3.3 Extension to 3D

Finally, we show first results of extending the monolithic formulation to 3D. A sim-
ilar problem of pulsatile flow in an elastic tube with an aneurysm-like cavity is
solved. The material parameters are the same as in the previous section and the
resulting deformation and flow field at different times are shown in Fig. 12.

7 Summary and future developments

We presented a monolithic ALE formulation of fluid-structure interaction problems

suitable for applications with large structural deformations and laminar viscous
flows, particularly arising in biomechanics. The corresponding discrete nonlinear
216 S. Turek et al.

26.63
no stents, elastic fundus
3 stents, elastic fundus
5 stents, elastic fundus
26.62 rigid walls

26.61

26.6
volume

26.59

26.58

26.57

26.56
1 1.5 2 2.5 3 3.5 4
time step
26.62
no stents, elastic fundus
3 stents, elastic fundus
5 stents, elastic fundus
26.615

26.61

26.605
volume

26.6

26.595

26.59

26.585

26.58
1.9 2 2.1 2.2 2.3 2.4
time step

Fig. 10 Domain volume with rigid and elastic behavior of the aneurysm wall.
Numerical simulation and benchmarking of fluid-structure interaction 217

Fig. 11 Left column: no stent. Middle column: three stents. Right column: five stents. Figures
demonstrate the local behavior of the fluid flow inside the aneurysm during one cycle.
218 S. Turek et al.

Fig. 12 Pulsatile fluid flow through an elastic tube with cavity. Flow field represented by velocity
vectors and velocity magnitude at different times.

systems result from the finite element discretization by using the high order Q2 P1
FEM pair. The systems are solved monolithically via a discrete Newton iteration
and special Krylov-multigrid approaches.
While we restricted our studies to the simplified case of Newtonian fluids and
small deformations, the used numerical components allow the system to be coupled
with additional models of chemical and electric activation of the active response of
the biological material as well as power law models used to describe the shear thin-
ning property of blood. Further extension to viscoelastic models and coupling with
mixture based models for soft tissues together with chemical and electric processes
allow to perform more realistic simulations for real applications.
We applied the presented numerical techniques to FSI benchmarking settings
(‘channel flow around cylinder with attached elastic beam’, see [30]) which allow
the validation and also evaluation of different numerical solution approaches for
fluid-structure interaction problems. Moreover, we examined prototypically the
influence of endovascular stent implantation on aneurysm hemodynamics. The aim
was, first of all, to study the influence of the elasticity of the walls on the flow
behavior inside the aneurysm. Moreover, different geometrical configurations of
implanted stent structures have been analyzed in 2D. These 2D results are far from
providing quantitative results for such a complex multiphysics configuration, but
they allow a qualitative analysis w.r.t. both considered components, namely the
elastic behavior of the structural parts and the multiscale flow behavior due to the
geometrical details of the stents. We believe that such basic studies are helpful for
the development of future ‘Virtual Flow Laboratories’ which individually assist to
design personal medical tools.
Numerical simulation and benchmarking of fluid-structure interaction 219

Acknowledgements The authors want to express their gratitude to the German Research Asso-
ciation (DFG), funding the project as part of FOR493 and TRR30, the Jindřich Nečas Center
for Mathematical Modeling, project LC06052 financed by MSMT, and the Higher Education
Commission (HEC) of Pakistan for their financial support of the study. The present material
is also based upon work kindly supported by the Homburger Forschungsförderungsprogramm
(HOMFOR) 2008.

References

1. P. R. Amestoy, I. S. Duff, and J. Y. L’Excellent. Multifrontal parallel distributed symmetric

and unsymmetric solvers. Computer Methods in Applied Mechanics and Engineering, 184
(2-4):501 – 520, 2000.
2. S. Appanaboyina, F. Mut, R. Löhner, E. Scrivano, C. Miranda, P. Lylyk, C. Putman, and
J. Cebral. Computational modelling of blood flow in side arterial branches after stenting
of cerebral aneurysm. International Journal of Computational Fluid Dynamics, 22:669–676,
2008.
3. D. N. Arnold, D. Boffi, and R. S. Falk. Approximation by quadrilateral finite element. Math.
Comput., 71:909–922, 2002.
4. R. Barrett, M. Berry, T. F. Chan, J. Demmel, J. Donato, J. Dongarra, V. Eijkhout, R. Pozo,
C. Romine, and H. Van der Vorst. Templates for the solution of linear systems: Building blocks
for iterative methods. SIAM, Philadelphia, PA, second edition, 1994.
5. D. Boffi and L. Gastaldi. On the quadrilateral Q2 P1 element for the Stokes problem. Int. J.
Numer. Meth. Fluids., 39:1001–1011, 2002.
6. R. Bramley and X. Wang. SPLIB: A library of iterative methods for sparse linear systems.
Department of Computer Science, Indiana University, Bloomington, IN, 1997. http://
www.cs.indiana.edu/ftp/bramley/splib.tar.gz.
7. H. Damanik, J. Hron, A. Ouazzi, and S. Turek. A monolithic FEM approach for non-isothermal
incompressible viscous flows. Journal of Computational Physics, 228:3869–3881, 2009.
8. T. A. Davis and I. S. Duff. A combined unifrontal/multifrontal method for unsymmetric sparse
matrices. ACM Trans. Math. Software, 25(1):1–19, 1999.
9. M. A. Fernandez, J-F. Gerbeau, and V. Martin. Numerical simulation of blood flows through
a porous interface. ESAIM: Mathematical Modelling and Numerical Analysis, 42:961–990,
2008.
10. M. Fortin. Old and new finite elements for incompressible flows. International Journal for
Numerical Methods in Fluids, 1(4), 1981.
11. Y. C. Fung. Biomechanics: Mechanical Properties of Living Tissues. Springer-Verlag, New
York, 1993.
12. R. Glowinski. Finite element method for incompressible viscous flow. In P. G Ciarlet and
J. L. Lions, editors, Handbook of Numerical Analysis, Volume IX,, pages 3–1176. North-
Holland, Amsterdam, 2003.
13. P. M. Gresho. On the theory of semi-implicit projection methods for viscous incompressible
flow and its implementation via a finite element method that also introduces a nearly consistent
mass matrix, part 1: Theory. Int. J. Numer. Meth. Fluids., 11:587–620, 1990.
14. G. A. Holzapfel. A continuum approach for engineering. John Wiley and Sons, Chichester,
UK, 2000.
15. G. A. Holzapfel. Determination of meterial models for arterial walls from uni-axial exten-
sion tests and histological structure. International Journal for Numerical Methods in Fluids,
238(2):290–302, 2006.
16. J. Hron, A. Ouazzi, and S. Turek. A computational comparison of two FEM solvers for nonlin-
ear incompressible flow. In E. Bänsch, editor, Challenges in Scientific Computing, LNCSE:53,
pages 87–109. Springer, 2002.
220 S. Turek et al.

17. J. Hron and S. Turek. A monolithic FEM/multigrid solver for ALE formulation of fluid struc-
ture interaction with application in biomechanics. In H.-J. Bungartz and M. Schäfer, editors,
Fluid-Structure Interaction: Modelling, Simulation, Optimisation, LNCSE:53. Springer, 2006.
18. J. C. Lagarias, J. A. Reeds, M. H. Wright, and P. E. Wright. Convergence properties of the
Nelder-Mead simplex method in low dimensions. SIAM J. Optim., 9:112–147, 1998.
19. R. Löhner, J. Cebral, and S. Appanaboyina. Parabolic recovery of boundary gradients.
Communications in Numerical Methods in Engineering, 24:1611–1615, 2008.
20. D. A. McDonald. Blood flow in arteries. In second ed. Edward Arnold, 1974.
21. J. A. Nelder and R. Mead. A simplex method for function minimization. Computer Journal,
7(4):308–313, 1965.
22. R. Rannacher and S. Turek. A simple nonconforming quadrilateral Stokes element. Numer.
Methods Partial Differential Equations., 8:97–111, 1992.
23. M. Razzaq. Numerical techniques for solving fluid-structure interaction problems with appli-
cations to bio-engineering. PhD Thesis, TU Dortmund, to appear, 2010.
24. M. Razzaq, J. Hron, and S. Turek. Numerical simulation of laminar incompressible fluid-
structure interaction for elastic material with point constraints. In R. Rannacher and
A. Sequeira, editors, Advances in Mathematical Fluid Mechancis-Dedicated to Giovanni paolo
Galdi on the Occasion of his 60th Birthday. Springer, in print, 2009.
25. T. E. Tezduyar, S. Sathe, T. Cragin, B. Nanna, B.S. Conklin, J. Pausewang, and M. Schwaab.
Modeling of fluid structure interactions with the space time finite elements: Arterial fluid
mechanics. International Journal for Numerical Methods in Fluids, 54:901–922, 2007.
26. T. E. Tezduyar, S. Sathe, M. Schwaab, and B.S. Conklin. Arterial fluid mechanics modeling
with the stabilized space time fluid structure interaction technique. International Journal for
Numerical Methods in Fluids, 57:601–629, 2008.
27. R. Torri, M. Oshima, T. Kobayashi, K. Takagi, and T.E. Tezduyar. Influence of wall elasticity
in patient-specific hemodynamic simulations. Computers and Fluids, 36:160–168, 2007.
28. R. Torri, M. Oshima, T. Kobayashi, K. Takagi, and T.E. Tezduyar. Numerical investigation of
the effect of hypertensive blood pressure on cerebral aneurysm dependence of the effect on the
aneurysm shape. International Journal for Numerical Methods in Fluids, 54:995–1009, 2007.
29. S. Turek. Efficient Solvers for Incompressible Flow Problems: An Algorithmic and Computa-
tional Approach. Springer-Verlag, 1999.
30. S. Turek and J. Hron. Proposal for numerical benchmarking of fluid-structure inter-
action between an elastic object and laminar incompressible flow. In H.-J. Bungartz
and M. Schäfer, editors, Fluid-Structure Interaction: Modelling, Simulation, Optimisation,
LNCSE:53. Springer, 2006.
31. S. Turek, L. Rivkind, J. Hron, and R. Glowinski. Numerical study of a modified time-steeping
theta-scheme for incompressible flow simulations. Journal of Scientific Computing, 28:533–
547, 2006.
32. S. Turek and M. Schäfer. Benchmark computations of laminar flow around cylinder. In E.H.
Hirschel, editor, Flow Simulation with High-Performance Computers II, volume 52 of Notes
on Numerical Fluid Mechanics. Vieweg, 1996. co. F. Durst, E. Krause, R. Rannacher.
33. S. Turek and R. Schmachtel. Fully coupled and operator-splitting approaches for natural
convection flows in enclosures. International Journal for Numerical Methods in Fluids,
40:1109–1119, 2002.
34. A. Valencia, D. Ladermann, R. Rivera, E. Bravo, and M. Galvez. Blood flow dynamics and
fluid–structure interaction in patient-specific bifurcating cerebral aneurysms. International
Journal for Numerical Methods in Fluids, 58:1081–1100, 2008.
35. S. P. Vanka. Implicit multigrid solutions of Navier-Stokes equations in primitive variables. J.
of Comp. Phys., 65:138–158, 1985.
Numerical Simulation of Fluid–Structure
Interaction Using Eddy–Resolving Schemes

M. Münsch and M. Breuer

Abstract Eddy–resolving schemes such as large–eddy simulation (LES) or

detached-eddy simulation (DES) have become popular due to their favorable capa-
bilities of predicting complex turbulent flows. That is especially true for instan-
taneous flow processes involving large–scale flow structures such as separation,
reattachment and vortex shedding. Flow phenomena of such kind are very often
encountered when the flow around or through a device enforces the structures to be
deformed or displaced, i.e. for fluid–structure interaction (FSI). The present study
deals with several aspects which have to be taken into account when LES is married
to FSI. That comprises the coupling scheme, the handling of moving or deformable
grids and the question how their quality requirements can be achieved. A coupling
scheme leading to strong coupling among flow and structure, but also maintaining
the advantageous properties of explicit time–marching schemes used for LES, was
set up and analyzed. Thus a new and favorable coupling procedure for FSI within
the LES context was developed. This and other issues of the numerical methods
applied such as the measures to maintain the grid quality are discussed in detail.
Results of validation test cases of FSI in the context of rigid body motions (e.g. an
elastically mounted cylinder or a swiveling flat plate) as well as benchmark results
with flexible structures computed with a finite–element code are presented.

M. Münsch
Institute of Fluid Mechanics, Friedrich-Alexander University Erlangen-Nuremberg, Cauerstraße 4,
D–91058 Erlangen, Germany
e-mail: [email protected]
M. Breuer
Department of Fluid Mechanics, Institute of Mechanics, Helmut-Schmidt-University Hamburg,
Holstenhofweg 85, D–22043 Hamburg, Germany
e-mail: [email protected]

H.-J. Bungartz et al. (eds.), Fluid Structure Interaction II, Lecture Notes 221
in Computational Science and Engineering 73, DOI 10.1007/978-3-642-14206-2 9,
c Springer-Verlag Berlin Heidelberg 2010
222 M. Münsch and M. Breuer

1 Introduction

The interaction between fluids and structures is a topic of interest in many fields
such as mechanical engineering (e.g. airfoils), civil engineering (e.g. towers) or
medicine technique (e.g. artificial heart valves). Beside experimental investigations
also numerical simulations have become an important and valuable tool for solving
this kind of problems. For numerical investigations two basic approaches can be
used. The monolithic approach is based on a unique numerical formulation of the
whole FSI problem (e.g. a finite–element formulation) which can be solved by one
sole code. By using a partitioned approach the entire FSI problem is divided into a
fluid and a structure domain. For both fields separate numerical formulations are set
up. The solution of the subproblems is done separately by the usage of highly spe-
cialized codes for each subtask. However, the coupling between both domains has
to be done via an additional coupling interface. Here aspects such as code-to-code
communication or grid-to-grid interpolation, e.g., of loads and displacements, have
to be considered.
In general the disparity of the scales is also characteristic for FSI problems.
The largest scales can be in the range of several hundred meters, e.g. suspension
bridges, whereas the smallest scales of turbulent flows, i.e. the Kolmogorov scale
lk , are in a range of 0.1 millimeters. In terms of flow simulation among others three
categories of methodologies can be chosen for turbulent flows: DNS, RANS and
LES. The direct numerical simulation (DNS) is the most accurate one for turbu-
lence computation, see e.g., [1]. The Navier–Stokes equations are solved without
averaging or approximations for all motions in the flow field. For that purpose a
high number of grid points and time steps are required which leads to a costly
and time–consuming simulation restricted to low or moderate Reynolds numbers.
A second group of methods are the so-called Reynolds–Averaged Navier–Stokes
(RANS) equations which arise from averaging the Navier–Stokes equations in time.
In contrast to the DNS approach the RANS equations have to be closed by a statis-
tical turbulence model. The resulting flow field is usually time–averaged. Previous
studies investigated laminar and also turbulent FSI applications using this RANS
approach. In [13, 14] a partitioned fully implicit FSI scheme was applied to couple
a three–dimensional finite–volume based multi–block flow solver for incompress-
ible fluids [9] with a finite–element code for the structural subproblem. Efficient
coupling was performed for large time step sizes used for implicit time–stepping
schemes within RANS simulations. Depending on the flow problem RANS sim-
ulations are not always suitable for the prediction of the flow field. Large–scale
flow structures such as vortex shedding or instantaneous separation and reattach-
ment require more advanced methodologies such as large–eddy simulations (LES)
[1–3] leading to the third category of methodologies for the simulation of turbulent
flows. In contrast to RANS the LES technique is based on a spatial filtering of the
Navier–Stokes equations and not on temporal averaging. Thus by computing the
large scales directly and modeling solely the small scales an unsteady field of physi-
cal values is obtained and the prediction of complex turbulent flow fields such as the
flow past bluff bodies is enabled. Therefore, LES is also a preferential method in the
Numerical Simulation of Fluid–Structure Interaction Using Eddy–Resolving Schemes 223

context of FSI applications when large–scale flow structures are expected. The task
of the project was to marry FSI and LES by considering all relevant aspects playing
a significant role for the (new) combination of these advanced techniques.

2 Governing Equations for the Fluid Flow

For the present study the inhouse code FASTEST-3D [9, 10] was used and extended
for the FSI–LES application intended. The code is based on a finite–volume scheme
which is used to discretize the filtered Navier–Stokes equations for an incompress-
ible fluid. The discretization is done on a curvilinear, blockstructured body–fitted
grid with colocated variable arrangement by applying standard schemes. Linear
interpolation of the flow variables to the cell faces and a midpoint rule approxi-
mation for the integrals is used to obtain a second-order accurate central scheme.
Within a FSI application the fluid forces acting on the structure lead to the dis-
placement or deformation of the structure. Thus the computational domain is no
longer fixed but changes in time, which has to be taken into account. Besides other
numerical techniques, the most popular one is the so-called Arbitrary Lagrangian–
Eulerian (ALE) formulation. Here the conservation equations for mass, momentum
(and energy) are re-formulated for a temporally varying domain, i.e., control vol-
umes (CV) with time-dependent volumes V .t/ and surfaces S.t/. The governing
equations in ALE formulation read:
Mass Conservation:
Z Z
d
 dV C .uj  ug;j /  nj dS D 0: (1)
dt V .t / S.t /

Momentum Conservation:
Z Z Z Z
d
ui dV C ui .uj  ug;j /  nj dS D ij  nj dS  p  ni dS:
dt V .t / S.t / S.t / S.t /
(2)
The density is denoted by , the pressure by p and the three Cartesian com-
ponents of the velocity vector by ui . The molecular momentum transport tensor is
indicated by ij whereas nj describes the unit normal vector pointing outwards.
In case of LES, an additional tensor denoted ijSGS has to be taken into account
in the surface integral on the right-hand side of Eq. (2) describing the influence of
the non–resolved subgrid–scales (SGS) on the resolved flow field. Since the grid is
deformable, the grid velocity with which the surface of a CV is moving is taken into
account via ug;j . Here, the volume integrals now describe local changes in a CV of
variable shape and thus the additional mass and momentum fluxes due to ug;j . Since
the system of equations given by (1) and (2) is not closed anymore, the unknown
grid velocity ug;j has to be determined. To consider the conservation principle and
224 M. Münsch and M. Breuer

to avoid the loss of mass and momentum the so-called space conservation law (SCL)
[7, 8] is applied to compute the unknown grid velocities ug;j .
Space Conservation Law:
Z Z
d
dV  ug;j  nj dS D 0: (3)
dt V .t / S.t /

This extra conservation law assures that within a change of the position or the
shape of a CV no space is lost. In discretized form the SCL is expressed by the swept
volumes of the corresponding cell faces. Inserting Eq. (3) into (1) the mass conserva-
tion equation for a fixed grid is obtained. Therefore, the original pressure–correction
scheme applied for the solution of the incompressible Navier–Stokes equations on
fixed grids has not to be changed concerning the mass conservation equation in
the context of moving grids. The additional grid fluxes in the momentum equation
are consistently determined by applying the SCL. Here, an incompressible fluid
with temperature–independent fluid properties is considered. Thus the conservation
equation for the energy is not taken into account.
In principle, the LES concept leads to a closure problem similar to that obtained
by RANS. However, the non–resolvable small scales in a LES are much less
problem–dependent than the large–scale motion so that the subgrid–scale turbulence
can be represented by relatively simple models, e.g., zero–equation eddy–viscosity
models. Like other eddy–viscosity models the well-known and most often used
Smagorinsky model [21] is based on Boussinesq’s approximation which describes
the stress tensor ijSGS as the product of the strain rate tensor S ij and an eddy
viscosity T ,
 
1 @ui @uj
ijSGS;a D ijSGS  SGS
ıij kk =3 D 2 T S ij with S ij D C ; (4)
2 @xj @xi

where ijSGS;a is the anisotropic (traceless) part of the stress tensor ijSGS and ıij is
the Kronecker delta. The trace of the stress tensor is added to the pressure resulting
SGS
in the new pressure P D p C kk =3. The eddy viscosity T itself is a function of
the strain rate tensor S ij and the subgrid length l, as
" 3 #0:5

2 y C
T D l j S ij j with l D Cs  1  exp ,
AC
r
y u w
 D .x  y  z/1=3 , C
y D , u D and AC D 25 . (5)
 

Cs is the well-known Smagorinsky constant which has to be prescribed as a

fixed value or can be determined as a function of time and space by the dynamic
procedure originally proposed by Germano et al. [12] and later improved by sev-
eral authors, e.g. Lilly [20]. In the first case, a Van Driest damping function is
required [see Eq. (5)] in order to take the reduction of the subgrid length l near solid
Numerical Simulation of Fluid–Structure Interaction Using Eddy–Resolving Schemes 225

walls into account. In the present investigations, both the fixed parameter version
of the Smagorinsky model was applied using the well established standard constant
Cs D 0:1 but also the dynamic version. The latter applies a test filter of filter width
2 taking the 27 neighboring control volumes into account.
The different structural models applied will be described in Sect. 6 in conjunction
with the corresponding test cases.

3 Coupling Schemes for Partitioned FSI

For the coupling between the fluid solver and the structure solver in general two
different methodologies can be applied in the context of a partitioned approach.
Within the so-called loose coupling approach the fluid and the structure subproblem
are only solved once per time step leading to severe stability problems in most appli-
cations. For the strong coupling approach, fluid and structure solutions are repeated
in a staggered manner. Here, a coupling scheme was finally developed meeting the
requirements for an efficient large–eddy simulation on the one hand and a stable
coupling algorithm on the other hand. This approach is characterized by a strong
coupling between fluid and structure and explicit time marching for the fluid sub-
problem. For comparison also a coupling scheme for implicit time marching used
in previous studies [13, 14] is presented.

3.1 Strong Coupling with Explicit Time Marching

For LES small time steps are required to resolve the turbulent flow field in time
which leads to a preferred usage of explicit time-marching schemes. Here, for exam-
ple, a predictor–corrector scheme of second–order accuracy forms the kernel of the
fluid solver. In the predictor step an explicit three substep low–storage Runge–Kutta
scheme advances the momentum equation in time leading to a prediction of the
velocities u , v and w . These predicted velocities do not satisfy mass conserva-
tion. Thus, in the following corrector step the mass conservation equation has to be
fulfilled by solving a Poisson equation for the pressure–correction based on a solver
relying on an incomplete LU decomposition. The pressure-correction algorithm is
repeated until a predefined convergence criterion " for the mass conservation is
reached leading to the final velocities u, v, w and the corresponding pressure field p
of the new time step. For that purpose, typically 5 to 10 pressure–correction itera-
tions are required until the mass conservation equation is numerically satisfied, e.g.
m P < " D O.108 /. For FSI computations this algorithm was extended towards a
code coupling scheme [4–6] which is shown in Fig. 1. Receiving the corresponding
velocity and pressure field from the pressure–correction step, the resulting forces,
i.e. pressure and shear stresses, or the associated moments are computed. These
forces or moments are transferred to an equation of motion if rigid body motion is
226 M. Münsch and M. Breuer

Estimation X t–1, X t–2, X t–3

~ ~. . t = t + Dt
X, X Xt –1

Grid
adaption ~
Xt = Xn

Predictor-Corrector
scheme
Momentum Eq.
Runge-Kutta yes
u ∗, v ∗, w ∗
Grid no FSI
adaption convergence

Poisson Eq. FSI

u, v, w, p subiteration loop n

. ~ ~ ~
X t–1, X t–1 X n = Xn–1 + a ( Xnew − X n –1 )
Forces and Moments CSD / Eq. of Motion

Fig. 1 Strong coupling algorithm with explicit time marching as used in the present study.

considered. For more complex structures such as shells or membranes the forces
are transferred to an appropriate computational structure dynamics code (CSD),
here a finite–element solver (Carat [24]). The response of the structure, i.e. the dis-
placements Xnew , are transferred to the fluid solver. In the case of a loose coupling
approach the displacements are considered to be the physically correct displace-
ments of the actual time step. Thus the computation passes on to the successive time
step (t D t C t). This loose coupling between fluid and structure is in general only
stable for low density ratios of the fluid to the structure, i.e., a low so-called added
mass effect. For cases in which the added mass effect plays a dominant role, the
loose coupling scheme, especially in connection with incompressible flows, tends
to fail.
To overcome this difficulty, a so-called FSI subiteration loop is introduced in the
scheme preferred [4–6]. If the FSI convergence criterion is not fulfilled, i.e., the
dynamic equilibrium between fluid and structure is not achieved, the computational
grid of the fluid domain is re-adapted. The grid adaptation is based on an underre-
laxation of the structural response Xnew by taking an underrelaxation factor ˛ and
the displacement of the previous subiteration loop (n-1) into account. Subsequent
to the grid adaptation solely the corrector step of the predictor–corrector scheme is
performed again and a new velocity and pressure field is obtained. Afterwards new
loads for the structure solver are generated leading to an update of the corresponding
displacements. The dynamic equilibrium between fluid and structure is numerically
obtained if the change of the resulting displacements within the subiteration cycle
Numerical Simulation of Fluid–Structure Interaction Using Eddy–Resolving Schemes 227

reaches a convergence criterion. Then the subiteration process is stopped and the
computation for the next time step is started. The new time step begins with an
estimation of the displacement XQ and velocity XPQ of the structure which serves for
convergence acceleration within the whole coupling process. Here, a second–order
extrapolation for the displacements is applied by taking the displacement values of
three former time steps indicated by the superscripts t 1 , t 2 , and t 3 into account:

XQ t D 3X t 1  3X t 2 C X t 3 : (6)

For consistency, a first–order extrapolation is applied on the structural velocity XPQ :

XPQ t D 2XP t 1  XP t 2 : (7)

According to these estimated boundary values, the entire computational grid has to
be adapted as it is done in each FSI subiteration loop. Presently, this grid adjustment
is performed based on a transfinite interpolation (see also Sect. 4). Once the grid
is adapted, the predictor–corrector scheme is applied and the first cycle of the FSI
subiteration loop is entered.
In contrast to this strong coupling scheme relying on explicit time marching,
a strong coupling scheme based on implicit time stepping as used in [13, 14]
is depicted in Fig. 2. Since the flow is solved with a fully implicit scheme, an

Estimation Xt–1, Xt–2, Xt–3

~ ~. . t = t + Dt
X, X Xt–1
~
Xt = Xn yes

Grid no FSI
adaption convergence

Momentum Eq. SIMPLE

u*, v*, w* scheme

Correction inner FSI

u, v, w, p CFD loop subiteration loop n

CFD no
convergence

yes . ~ ~ ~
X t–1, Xt–1 Xn = Xn–1+ α (Xnew–Xn–1)
Forces and Moments CSD / Eq. of Motion

Fig. 2 Strong coupling algorithm with implicit time marching shown for comparison.
228 M. Münsch and M. Breuer

inner CFD loop becomes necessary within the applied SIMPLE scheme [15]. Both
schemes significantly differ with respect to the momentum equations within the
FSI–subiteration loop. For the implicit scheme the momentum equations are solved
repeatedly in each subiteration sweep. In contrast they are only solved once per time
step for the explicit case which strongly reduces the computational effort. Further-
more, the number of FSI–subiterations required to reach the convergence criterion
is typically at least one order of magnitude smaller for the explicit scheme than for
the implicit variant. In conclusion, instabilities due to the added mass effect known
from loose coupling schemes are avoided by the newly developed coupling scheme
and the explicit character of the time–stepping scheme beneficial for LES is still
maintained.

3.2 Data Interpolation and Transfer

In the elementary case of a rigid body motion, here described via an equation of
motion embedded within the fluid solver, the forces and/or moments are computed
based on the pressure and the shear stresses on the surface. These pressures and
shear stresses are located in the face center nodes of the control volumes which
stick to the fluid–structure interface. Depending on the direction of the forces or
moments they are applied as source terms in the corresponding equations of state
for driving the displacement. The predicted displacements are directly transferred to
the corresponding grid nodes.
When dealing with more complex structures, a more sophisticated structure
solver, i.e. a finite–element solver, has to be considered. Due to different discretiza-
tion techniques applied for the subtasks (finite volumes vs. finite elements) also
different types of grids and different grid resolutions are used leading to non–
matching surfaces meshes. Consequently, a grid to grid data interpolation and
transfer becomes necessary. The fluid solver FASTEST-3D is based on a cell–
centered variable arrangement for the pressure and the velocities, whereas the grid
coordinates and displacements are cell–vertex bound. Therefore, in an initial step,
the fluid loads FCi predicted based on the pressure and shear forces at the cell face
nodes (CFN, center of the cell) are conservatively interpolated to the face vertices,
i.e., the grid node (GN), see Fig. 3. That step is necessary since the displacements
are required at GN. The whole interpolation procedure is as follows:
1. Conservative interpolation of the fluid loads located in the cell centers to the cell
vertices of the fluid domain (CFD!CFD):

nj ni
X j
X
FGN;CFD D FiC;CFD (8)
j D1 i D1
Numerical Simulation of Fluid–Structure Interaction Using Eddy–Resolving Schemes 229

F iGN F jGN

FiC
1 1
CFN

CFD

1 1
FkGN FlGN

2 3

r=3
r=4 r=2

r=1
η CSD
ξ

Fig. 3 Interpolation of loads and displacements between the fluid domain (CFD) and the structure
domain (CSD).

2. Conservative interpolation of the fluid loads from the cell vertices of the fluid
domain to the grid nodes of the structure domain (CFD!CSD):

nk nj
X X j
FkCSD D FGN;CFD (9)
kD1 j D1

3. Bilinear interpolation of the displacements from the structure domain ur;CSD to

the cell vertices of the fluid domain (CSD!CFD):

4
X
.Nre .j ; j /ur;CSD /
j
xGN;CFD D (10)
rD1

j
Here, the resulting displacements in the fluid domain are denoted by xGN;CFD
and are obtained by the displacements of the structure domain ur;CSD weighted
by the shape functions Nre of the CSD element e. The interpolation steps 2 and
3 are done via the coupling interface CoMA [11] developed by the Chair of Struc-
tural Analysis of the Technical University Munich. This interface is based on the
Message–Passing–Interface (MPI) and thus runs in parallel to the fluid and struc-
ture solver. The communication in-between the codes is performed via standard
MPI commands. Since MPI is also used for code parallelization in FASTEST-3D,
effective coupling with respect to high–performance computing is enabled.
230 M. Münsch and M. Breuer

4 Grid Smoothing

Within the FSI algorithm the computational grid of the fluid domain is adjusted to
the actual position of the structure. This is done during each subiteration cycle or at
least once in a time step. To maintain grid quality, i.e., orthogonality and smooth-
ness, and according to these properties the accuracy of the numerical prediction
for time–dependent and moving grids, is a further challenge for FSI algorithms,
especially for turbulent flows using the LES approach (see Sect. 5). Beside using
algebraic approaches based on transfinite interpolations also developments on ellip-
tic grid smoothing [22, 25] have been carried out within the project. The objective
is to maintain some designated grid quality, in order to resolve boundary layers
for example with a minimum computational effort. Here, the approach chosen is
based on a composite mapping which is performed separately on every block of
the flow field for the reason of parallelization. The composite mapping consists of
a nonlinear transfinite algebraic transformation and an elliptic transformation. The
algebraic transformation maps the computational space C one-to-one onto a param-
eter space P , with P  R3 . The computational space C , with C  R3 , as well as
the parameter space P are introduced as a Cartesian unit cube with  D .; ; /
and s D .s; t; u/, respectively. By applying the algebraic transformation the grid
points in the inner region of P are set based on the prescribed boundary grid point
distribution on @C . Due to the one-to-one mapping relation between the boundary
@C and the boundary @D of the physical space D, the grid point distribution in P
is directly related to the prescribed boundary point distribution in D. The physical
space D  R3 , a block of a block-structured grid for example, is spanned via Carte-
sian coordinates x D .x; y; z/. Once discretized, the nonlinear system of equations
for calculating the position of the points Pi D .sP ; tP ; uP /i in space P is solved by
the Newton approach. The resulting grid distribution defines 18 control functions
(for the 3D case) of a Poisson system which has to be solved to obtain the final grid
point distribution in D. This differentiable one-to-one mapping of the parameter
space P onto the physical space D is defined by the elliptical transformation. The
resulting nonlinear equation system is linearized by the use of Picard iterations and
then solved via the strongly implicit procedure, i.e., the so-called SIP-solver already
available in FASTEST-3D. The grid obtained in D is basically non–orthogonal at
the boundaries.
In [22] an extended approach is suggested. In a first step the basic algorithm
described above is applied to generate an initial grid. In a following step the Laplace
equations s D 0, t D 0 and u D 0 are solved. Here, Neumann boundary con-
ditions, i.e., @s=@n D 0 for example, are applied for the coaxial edges of P and
boundary conditions of Dirichlet type, i.e., s D 1 for example, for the remaining
edges. In a final step, again the Poisson system with control functions based on
cubic Hermite interpolations are solved leading to the desired orthogonality at the
boundaries. Here, the computational effort is very high and even higher in the con-
text of FSI applications, where these steps have to be repeated several times during
one time step. Therefore, investigations on two alternative approaches have been
started to obtain a satisfying grid on the one hand with less computational effort
Numerical Simulation of Fluid–Structure Interaction Using Eddy–Resolving Schemes 231

on the other hand. The first approach follows the idea of readjusting the source
terms, in the vicinity of walls for example, to modify the grid with respect to grid
point distances or angles between the grid lines. Here, the method of Hilgenstock
[18] is applied and added to the basic algorithm described above. In theory the
computational effort is reduced in contrast to the extended algorithm described in
[22]. But special care has to be taken since grid folding may occur and conver-
gence may restricted due to severe manipulations of the source terms. The second
approach starts from an initially more or less perfect grid in respect to orthogonal-
ity and refinement, from which the corresponding source terms are computed. Then
these source terms are adjusted, in the course of an FSI computations, to adapt the
grid to the moving boundaries and to maintain grid quality. Here, only principle
investigations have been performed up to now.
Beside the grid quality close to the walls also the grid quality close to the block-
to-block connectivities is of interest. The basic approach described in [22] will lead
to a smooth grid in the inner region of the corresponding block. Since the grid point
on the boundary @D are not moved, the grid quality of the block-to-block intersec-
tion may be reduced. The extended approach in [22] leads to an orthogonal grid
at the boundaries and thus to a smooth block-to-block intersection of grid lines.
However, the computational effort is very high and total orthogonality may not be
necessary.
To overcome this drawback, in the present study the source terms in parameter
space P , which are influencing the grid quality also at the boundaries, are computed
with the help of overlapping grids at the block-to-block connectivities. Due to this
procedure a smooth but non–orthogonal grid is generated at the connectivities. Since
the computation of these source terms can be incorporated into the basic approach
of [22], the total computational effort is of the same order than the basic algorithm
which is strongly advantageous for FSI–LES applications.

5 Large–Eddy Simulation on Moving Grids

For a finite–volume scheme as applied in the present investigation, the grid and thus
the size of the control volumes varies in time. In the context of LES relying on an
implicit filtering approach as used in many practically relevant LES predictions, the
grid movement means that the filter width defined by Eq. (5) now additionally varies
in time, i.e.  D f .x; t/. That leads to additional commutation errors depending on
d=dt and d=dx [19]. In order to investigate the influence of the temporally
varying filter width on the quality of the results, predictions of a plane channel
flow at Re D 590 on oscillating grids (domain: 2 ı  2ı  ı where ı denotes
the channel half–width; grid: 128  128  128 CVs) were carried out applying the
classical Smagorinsky model [21] with Van Driest damping near solid walls as well
as the dynamic model by Germano et al. [12] and Lilly [20].
232 M. Münsch and M. Breuer

D X (t–t0) / T = 1 / 4
1.6
D X (t–t0) / T = 3 / 4

A = 0.4
1.4

1.2
DX

0.8

0.6

0.4
–1 0 1 2 3 4 5 6 7
X

Fig. 4 Distribution of the grid spacing x for .t  t0 / D 14 T and .t  t0 / D 34 T over the channel
length x.

5.1 Grid Movement

The sinusoidal grid movement was defined by

xmove =ı D A  sin.2 .t  t0 /=T /  sin.x0 =ı/: (11)

Here, the new coordinate of the grid point is denoted with xmove whereas x0 is
the reference coordinate of the fixed grid computation. The grid movement and with
it the effective grid spacing x is controlled via the amplitude A and the period T
of the sinusoidal deformations. In Fig. 4 the resulting grid spacing x along the
channel length x is plotted for two different instants in time within a period of
motion. For an amplitude of A D 0:4 the grid is compressed to 60% of the initial
grid spacing in the middle of the channel at .t  t0 / D 14 T , see also Fig. 5(a). At
.t  t0 / D 34 T the grid is stretched to 140% of the initial grid spacing in the middle
of the channel. The time step t was equal to 0.01 s for all computations.
A variety of cases was considered (see Table 1). In one subset of cases, the ampli-
tude A of the internal grid deformation was varied from 0.1 to 0.85 and the period T
of the deformations was set to 1 flow–through time (TF T ). In a second subset,
computations for different cycle periods T at a constant amplitude were performed.

5.2 Results

The results were compared to LES predictions on a fixed grid (reference) by comput-
ing the root-mean-square deviations (rmsd) of the time- and space–averaged stresses
Numerical Simulation of Fluid–Structure Interaction Using Eddy–Resolving Schemes 233

0 1 2 3 4 5 6
(a)

0 1 2 3 4 5 6
(b)

Fig. 5 Reference grid (a) and deformed grid (b) at .t  t0 / D 34 T .

Table 1 Overview of moving grid computations performed showing amplitudes A and cycle peri-
ods T with the resulting root-mean-square deviations (rmsd) of u0 u0 , v0 v0 , w0 w0 , u0 v0 and U in
[%].
Smagorinsky model: variation of amplitude A
A T=TFT rmsd(u0 u0 ) rmsd(v0 v0 ) rmsd(w0 w0 ) rmsd(u0 v0 ) rmsd(U )
0.10 1 2.63 0.81 0.85 8.93 0.16
0.25 1 2.98 1.11 1.16 5.22 0.21
0.50 1 2.55 2.68 2.36 9.36 0.48
0.65 1 2.63 3.71 2.53 5.74 0.67
0.75 1 3.34 5.13 3.63 13.73 0.81
0.85 1 4.19 6.35 3.92 12.06 1.06
Dynamic model of Germano et al.: variation of amplitude A
A T=TFT rmsd(u0 u0 ) rmsd(v0 v0 ) rmsd(w0 w0 ) rmsd(u0 v0 ) rmsd(U /
0.25 1 1.70 4.37 2.46 33.69 0.41
0.50 1 2.71 6.79 4.28 20.02 0.68
Smagorinsky model: variation of cycle period T
A T=TTF rmsd(u0 u0 ) rmsd(v0 v0 ) rmsd(w0 w0 ) rmsd(u0 v0 ) rmsd(U /
0.50 1 2.55 2.68 2.36 9.36 0.48
0.50 0:2 4.98 10.58 5.70 12.80 1.40
0.50 0:1 10.15 26.57 17.17 18.54 3.04

u0 u0 , v0 v0 , w0 w0 , u0 v0 and of the mean flow velocity U . As can be seen in Table 1

for a fixed period T an increasing amplitude A leads to increasing deviations of
the investigated flow properties for the classical Smagorinsky model. Beside the
shear stresses all other quantities reach their maximum deviations at an amplitude of
A D 0:85. The rmsd for u0 v0 reaches a maximum for an amplitude of A D 0:75. The
234 M. Münsch and M. Breuer

mean velocity component U shows in general the smallest deviation from the refer-
ence test case. These results are also shown in Fig. 6. The results of the amplitude
variation for the dynamic model of Germano et al. yield the same trend. Compar-
ing the results of both models, the Smagorinsky model shows less deviations from
the reference case than the dynamic model. The second test case, i.e. the variation
of the period T , shows a bigger impact on the rmsd–values than the first test case.
For a decreasing period, i.e. an increasing frequency of the grid motion, the devia-
tions from the reference case increase. A maximum rmsd–value of 26.57% for v0 v0
is obtained for T =TF T D 0:1. Here, again the mean flow velocity U is the least
sensitive variable with a deviation of 3.04%. Exemplarily, Fig. 7 depicts the results
for the variation of the period T . In the first case, one cycle of the grid deformation
is equivalent to one flow–through time. In that case (T =TF T D 1) the deviations
from the reference case are negligible. If the grid oscillates faster (i.e. 5 oscillations
per flow–through, T =TF T D 0:2) first deviations from the reference case appear.

1
2.5 A = 0.1
A = 0.5
0.8 A = 0.85
2
reference
RMS (u’u’)

0.6 1.5
U

A = 0.1
0.4 A = 0.5
A = 0.85 1
reference
0.2 0.5

0
–1 – 0.5 0 0.5 1 –1 –0.5 0 0.5 1
Y Y
(a) RMS .U / (b) RMS .u0 u0 /

1
1.2

0.8 1

0.8
RMS (w’w’)
RMS (v’v’)

0.6

0.6
0.4 A = 0.1
A = 0.5 0.4 A = 0.1
A = 0.85 A = 0.5
0.2 refrence A = 0.85
0.2
reference

–1 –0.5 0 0.5 1 –1 – 0.5 0 0.5 1

Y Y
(c) RMS .v0 v0 / (d) RMS .w0 w0 /

Fig. 6 LES predictions (Smagorinsky model) of plane channel flow on temporally varying grids,
three different amplitudes A, grid cycle periods T D TF T (RMS values normalized by u ).
Numerical Simulation of Fluid–Structure Interaction Using Eddy–Resolving Schemes 235

(a) RMS .U / (b) RMS .u0 u0 /

(c) RMS .v0 v0 / (d) RMS .w0 w0 /

Fig. 7 LES prediction of plane channel flow on temporally varying grids, A D 0:5, three different
grid cycle periods T (RMS values normalized by u ).

A further increase to 10 grid oscillations per flow–through leads to strong deviations.

In that case the errors introduced by the grid movement are obvious. Although dif-
ferent effects such as time–varying local resolution and the commutation error itself
cannot clearly be distinguished, in summary, these additional errors due to moving
grids significantly affect the LES results. The impact on the mean flow properties
is a minor one as long as the amplitudes and corresponding oscillations are not too
high. Therefore, LES was already applied to fully coupled FSI computations within
this study (see Sect. 6). The effect on the secondary moments is much stronger lead-
ing to severe restrictions on tolerable amplitudes and cycle periods. These results
provide hints on what kind of grid motions within an FSI application using LES can
be accepted.
236 M. Münsch and M. Breuer

6 FSI Test Cases

In this section, a subset of test cases is presented which have been computed to verify
the FSI coupling algorithm developed. For the sake of simplicity only rigid body
motion is considered. The test cases are dealing with laminar as well as turbulent
flows. Laminar flows are considered here because of less computational effort to
derive data sets for comparison, although this work focuses on turbulent flows.

6.1 Elastically Mounted Cylinder at Re D 200

The flow around an elastically supported circular cylinder was computed for dif-
ferent values of the reduced damping parameter Sg (defined below) according to
the paper of Zhou et al. [26]. The Reynolds number based on the reference velocity
U1 D 0:514 m/s and the cylinder diameter of D D 6  103 m was equal to 200
and thus in the laminar regime. Hence the subgrid-scale model was switched off (no
LES). The configuration is depicted in Fig. 8(a).
For the prediction of the flow field a grid with 262,144 control volumes was used.
The spanwise direction was resolved with 8 CVs since the flow is assumed to be
two–dimensional. In spanwise direction symmetry boundary conditions are applied.
At the inflow a constant velocity U1 was specified. A no–slip boundary condition
was set for the cylinder surface. At the outflow a convective outflow condition was
prescribed. The time-step size applied was t D 1:5  105 seconds. Assuming the
cylinder to be fixed, i.e. the classical test case of the flow around a cylinder, the well–
known vortex shedding phenomenon in the wake of the cylinder was obtained. The
corresponding shedding frequency was computed to be fs D 16:7 Hz leading to a
Strouhal number of S t D fs  D=U1  0:2. The result obtained by the fixed cylin-
der computation was taken as initial conditions for the fluid–structure interaction
case, i.e., the released and thus oscillating cylinder. The response on the resulting
fluid forces is described by equations of motions [Eq. (12)] which are solved numer-
ical by a classical Runge–Kutta scheme. Thus the resulting displacements in x and
y directions are controlled by the mass of the cylinder m, a damping coefficient d , a
spring constant c and the related forces Fx .t/ and Fy .t/. The forces are determined
by numerical integration of the resulting pressure field p and shear stresses  acting
on the cylinder surface:

R
m x.t/ C d x.t/
P C c x.t/ D Fx .t/ ;
(12)
R
m y.t/ C d y.t/
P C c y.t/ D Fy .t/ :
In the work of Zhou et al. [26] a so-called reduced damping parameter Sg D
8 2 S t 2 d  M  with the mass ratio M  and the normalized damping parameter
d  was introduced. The reduced mass ratio is defined as M  D m =D 2 with
the normalized cylinder mass m D m= lc . The normalized damping parameter is
given by d  D d= lc . Here, the fluid density is denoted by  and lc describes the
cylinder length. Additionally, the relation between the natural cylinder frequency fn
Numerical Simulation of Fluid–Structure Interaction Using Eddy–Resolving Schemes 237

c d 0.02
0.015 U / U∞
c 1.5
0.01 1.4
1.3
1.2
0.005 1.1
1

Y [m]
Uoo m 0
0.9
0.8
0.7
D –0.005 0.6
0.5
0.4
–0.01 0.3
0.2
0.1
y d
–0.015
–0.02
0 0.02 0.04
x X [m]
(a) (b)

Fig. 8 (a) Spring–mass–damper model of the cylinder test case; (b) Flow field of the normalized
velocity magnitude for Sg D 1:0 and Re D 200.

(a) Sg D 0.01 (b) Sg D 1.0

Fig. 9 Normalized displacements of the cylinder center for (a) Sg D 0:01 and (b) Sg D 1:0.

p
and the shedding frequency fs given by fn D 1=.2 / c=m D 1:3fs provides the
corresponding values for the spring constant c and the cylinder mass m.
In Fig. 8(b) a snapshot of the flow field depicted by the normalized velocity
magnitude is exemplarily shown for Sg D 1:0. Downstream of the cylinder the vor-
tex shedding phenomenon occurring is visible. Furthermore, the recirculation area
behind the cylinder is deflected due to the cylinder movement. In Fig. 9 the nor-
malized traces of the cylinder are shown. As observed by Zhou et al. the X/D–Y/D
plots describe an asymmetric 8. The point of intersection is located off-centered for
both cases. In Table 2 the predicted results of the absolute normalized locations are
presented in comparison with the results by Zhou et al. [26]. A decreasing reduced
damping parameter Sg leads to an increase of the displacements. The agreement
between the predicted and the reference Y/D values is better than for the X/D val-
ues. Overall the results show satisfactory correspondence with the reference values.
The deviations maybe generally attributed to the different solution techniques,
238 M. Münsch and M. Breuer

Table 2 Absolute normalized displacements for Sg D 0:01 and Sg D 1:0 at Re D 200.

Sg D 0:01:
.X=D/mi n .X=D/max .Y =D/mi n .X=D/max
Present prediction 0.33 0.55 0.64 0.61
Zhou et al. [26] 0.31 0.47 0.62 0.56
Sg D 1:0:
.X=D/mi n .X=D/max .Y =D/mi n .X=D/max
Present prediction 0.32 0.34 0.29 0.28
Zhou et al. [26] 0.28 0.39 0.27 0.27

i.e. finite–volume method versus discrete vortex method, and/or to the unsteadiness
of the flow leading to different solutions due to bifurcation effects. For this test case
the maximum number of FSI subiterations nmax was very low, i.e., equal to 1.

6.2 Elastically Mounted Square Cylinder at Re D 13,000

Further testing of the FSI coupling scheme was performed by the investigation of
the sub-critical flow around an elastically mounted square cylinder as shown in
Fig. 10(a). With a free–stream velocity of U1 D 10 m/s and a cylinder edge length
of b = 0:02 m the Reynolds number is equal to Re D 13; 000. The extension of the
cylinder in spanwise direction is equal to 4b, whereas the dimension of the whole
domain is specified by 40b  11b  4b. The cylinder center is placed at 10b down-
stream of the inflow section. The grid applied consists of 2,228,224 CVs using 130
CVs to resolve the spanwise direction.
At the inflow a velocity block profile was specified with U1 . For the upper and
lower face of the domain slip boundary conditions were set. In spanwise direction a
periodic boundary condition was used. A convective outflow condition was applied
with Uc D U1 at the outflow section. On the cylinder surface a no–slip boundary
condition is applied. For the computation of the turbulent flow field a LES predic-
tion using the Smagorinsky model was performed. The Smagorinsky constant was
set to CS D 0:1. In order to compare the coupling schemes presented in Figs. 1
and 2 the time-step size has been chosen such that the step size fits to the numer-
ical characteristics of the considered coupling scheme. Thus, a time-step size of
t D 104 seconds was applied for the implicit time–marching algorithm lead-
ing to a CFL number of 20. A CFL number of 0:59 was obtained for the explicit
time marching algorithm with a time step size of t D 3  106 which is more
appropriate for the simulation of turbulent flows with LES than the implicit coun-
terpart. The displacement of the structure are considered by equations of motions
fitting to a spring–mass–damper system [Eq. (12)]. The mass of the cylinder is
equal to m D 0:005 kg, the damping coefficient is set to d D 0:1 Ns/m and a
spring constant of c D 5 N/m is applied. As before, the equations are solved by
Numerical Simulation of Fluid–Structure Interaction Using Eddy–Resolving Schemes 239

c d
0.18
c 0.16 U / U∞
m 0.14 1.9
1.7
Uoo 0.12
1.5

Y [m]
1.3
b 0.1
1.1
0.9
0.7
0.08 0.5
0.3
0.1
0.06
y d 0.04
0.18 0.2 0.22 0.24 0.26 0.28 0.3 0.32 0.34 0.36 0.38 0.4 0.42
x X [m]

(a) (b)

Fig. 10 (a) Spring–mass–damper model of the square cylinder test case; (b) Normalized velocity
magnitude and streamlines of the oscillating cylinder test case at Re D 13; 000.

0.15

2
0.1

1.5 0.05
Δx / b

Δy / b

0
1
Δx / b Expl.TM
–0.05
Δx / b Impl.TM
0.5
–0.1 Δy / b Expl.TM
Δy / b Impl.TM
0 –0.15
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
time [s] time [s]
(a) (b)

Fig. 11 Normalized displacements obtained by the strong coupling algorithm with either explicit
or implicit time–marching schemes: (a) x–component and (b) y–component.

a classical Runge–Kutta scheme. As initial conditions the displacements, velocities

and accelerations of the structure were set to zero. q
A snapshot of the resulting normalized velocity magnitude Ux2 C Uy2 C Uz2 =
U1 and the corresponding streamlines is plotted in Fig. 10(b). For this case the
flow at the inlet is laminar. Flow separation occurs at the leading and partially at
the trailing edge of the cylinder forming free shear layers in which the transition to
turbulence takes place. Oscillations of the cylinder in x and y–direction are excited
due to the phenomenon of alternating vortex shedding from the cylinder which is
visible in the wake of the cylinder. In addition, a drag–induced movement of the
cylinder in x–direction occurs. The resulting normalized displacements x=b and
x=b are shown in Fig. 11.
Starting from the initial conditions x=b D 0 and y=b D 0 a transitional
behavior is visible for the x–component for both FSI coupling schemes replying
240 M. Münsch and M. Breuer

either on explicit or implicit time–marching. After a real time of about 0:1 seconds
a maximum displacement is reached. The displacement reduces until the transi-
tional time interval ends at about 0:5 seconds. In the following, the x–components
vary around x=b D 1:6. The y–component oscillates around y=b D 0 caused
by the vortex shedding cycle. The maximum number of subiterations was equal
to nmax D 3 for the explicit time marching. In Fig. 11 the displacement values
are not perfectly matching for both time–stepping schemes. The displacement in
x–direction predicted by the explicit time–marching scheme is slightly higher than
for the implicit scheme. The y-components do not correspond in their chronological
sequence. Here, only a correspondence in the displacement magnitudes can be seen.
The differences can be attributed to different starting time steps when the cylinder
is released, used for both LES computations. Furthermore, the non–deterministic
nature of turbulence has to be taken into account. Thus only statistical values such
as mean values or standard deviations can be directly compared but not the time
histories of instantaneous processes.
For further verification of the results a Fast–Fourier–Transformation was applied
on the displacement data sets by neglecting the phase of transition, i.e., data sets
for real time values greater than 0:8 seconds. Whereas no characteristic peak is vis-
ible in the frequency spectra of the x–displacements, such peaks show up for the
y–displacements. For the data predicted by the implicit time–marching a peak was
found for a frequency of f D 69 Hz which equals the shedding frequency of the
vortices. The explicit time–marching scheme delivers a slightly shifted character-
istic frequency of f D 71 Hz. Summing up, the results obtained by the different
coupling strategies relying on different time–marching schemes show good agree-
ment. Thus, the functionality of the developed strong coupling scheme using an
explicit time–marching method within a LES prediction is confirmed.

6.3 Swiveling Plate in the Turbulent Regime

The fluid–structure interaction of a pivot–mounted flat plate within a turbulent chan-

nel flow [6] was considered for further testing of the developed FSI scheme. For that
purpose also experimental investigations have been performed at LSTM Erlangen
[17]. In Fig. 12 the setup of the computational domain is plotted. A flat plate of
dimensions 177mm  64mm  2mm with a chord length c D 64 mm is mounted
in a channel with rectangular cross-section.
The turbulent flow is simulated by LES using the Smagorinsky model with a
Smagorinsky constant of CS D 0:1 and Van Driest damping near solid walls. At the
inflow a block velocity profile is set. Here, two inflow velocities have been consid-
ered with U1 D 1:07 m/s and U1 D 0:6 m/s leading to Reynolds numbers based
on the chord length c of Re D 68; 000 and 38; 000. At the outlet a convective out-
flow condition is set with an outflowing velocity Uc D U1 . At the upper and lower
channel wall as well as for the flat plate no–slip conditions have been applied. In
spanwise direction two cases have been considered. On the one hand only a section
Numerical Simulation of Fluid–Structure Interaction Using Eddy–Resolving Schemes 241

1.7c

I, C y
c
g

x
ϕ

3c

3.75c

Fig. 12 Setup of the domain for the swiveling flat plate inside a channel.

Table 3 Grid resolutions and spanwise extensions used to resolve the turbulent flow field.
Grid Number of CVs Spanwise Resolution Spanwise Extension [mm]
1 434,176 4 64
2 9,409,536 64 64
3 31,757,184 177 177

of one chord length of the full spanwise extension of the plate used in the experi-
mental setup is taken into account. On the other hand, i.e. for one computation, the
spanwise direction is set as in the experimental setup. For both cases periodic bound-
ary conditions have been applied. Thus, the influence of the side walls is neglected.
This is at least partially justified due to the fact that in the experimental setup a
small gap between the plate ends and the spanwise channel walls is installed. The
simulations have been performed on three different grids (see Table 3). The prelim-
inary grid 1 has a total number of 434,176 CVs resolving the spanwise direction
(64 mm) with only 4 CVs. Grid 2 has a total number of 9,409,536 CVs. Here, the
spanwise direction (64 mm) is resolved with 64 CVs. Grid 3 has a total number
of 31,757,184 CVs with 177 CVs in spanwise direction taking the experimental
extension in spanwise extension (177 mm) into account.
For the non–deflected plate a wall–normal resolution given by y C D 3 for Re D
68; 000 and by y C D 0:7 for Re D 38; 000 is achieved for the wall–nearest control
volume center to resolve the flow at the trailing edge of the plate. The time–step
size was t D 105 s for all cases considered. The response of the plate on the
outer moment Mz .t/ imposed by pressure and shear forces is described by a spring–
mass–model leading to an ordinary differential equation for the plate angle '. This
equation is solved by a classical Runge–Kutta scheme (RK) on the one hand and,
for verification reasons, with an implicit Euler scheme (IE) on the other hand.
242 M. Münsch and M. Breuer

40 40

front body angle [deg]

20 20
angle [deg]

0 0

– 20 –20

– 40 – 40
1 2 3 4 5 6 7 0 60 120 180 240 300 360
time [sec] time phase angle [deg]
(a) (b)

Fig. 13 Computed angular displacements of the plate: (a) as a function of time, (b) for one period
as a function of the time–phase angle for Re D 68; 000 computed on grid 1 and solving the
equation of motion with the Runge–Kutta scheme.

R C C sin.'.t// D Mz .t/
I '.t/ (13)

Here, I describes the moment of inertia of the plate and the axle and C is equal
to the resulting moment due to gravity and buoyancy forces. The latter results from
density difference between the fluid and the structure. The outer moment Mz .t/
was linearly scaled for the cases with reduced spanwise extension for fitting the
experimental setup.
In Fig. 13(a) the temporal development of the plate angle is shown. For initial-
ization purposes of the turbulent flow field the plate is kept in fixed position, i.e.
' D 0ı , for about 0:36 s real time. Then the plate is released and the coupling
between fluid and structure is performed. After a period of transition, which takes
about 2 s, amplitudes of 'mi n D 45ı to 'max D 45ı are reached. In Fig. 13(b)
the instantaneous time–phase resolved plate angle for an exemplarily chosen period
is shown. In Table 4 the computed results for the different Reynolds numbers, com-
putational grids and solvers used for the equation of motion are shown. Obviously,
the swiveling frequency obtained increases with the Reynolds number. Furthermore,
the absolute magnitude of the angular displacements reduces by refining the grid in
spanwise direction. For grid 3 only preliminary results, i.e., not averaged results, are
shown but the tendency is confirmed. The choice of an implicit Euler or a classical
Runge–Kutta scheme for solving the equation of motion barely does not effect the
angular amplitudes nor the frequencies. Performing the FSI subiterations in com-
bination with this classical Runge–Kutta scheme, an averaged number of n D 5
subiterations was required to obtain dynamic equilibrium between fluid and struc-
ture. When taking the experimental results into consideration, a mismatch between
the computed and the measured results becomes obvious. The angular displace-
ments are measured to be in a range of ' D ˙22ı and ' D ˙27ı , respectively.
The effect of increasing angular amplitudes with increasing Reynolds number was
Numerical Simulation of Fluid–Structure Interaction Using Eddy–Resolving Schemes 243

Table 4 Overview of the simulations performed concerning to different grid sizes, Reynolds num-
bers Re and solvers for the equation of motion with the corresponding findings in comparison with
the experimental results.
Grid 1: 434,176 CVs with 4 CVs in spanwise direction
U1 [m/s] Re Structure 'max [ı ] 'mi n [ı ] Frequency[Hz]
0.6 38,000 RK C49 49 1.94
1.07 68,000 RK C49 45 3.14
1.07 68,000 IE C46 47 3.45

Grid 2: 9,409,536 CVs with 64 CVs in spanwise direction

U1 [m/s] Re Structure 'max [ı ] 'mi n [ı ] Frequency[Hz]
0.6 38,000 RK C44 44 1.95
1.07 68,000 RK C44 45 3.23

Grid 3: 31,757,184 CVs with 177 CVs in spanwise direction

U1 [m/s] Re Structure 'max [ı ] 'mi n [ı ] Frequency[Hz]
1.07 68,000 RK >C40 <40 >3.0

Experiment:
U1 [m/s] Re 'max [ı ] 'mi n [ı ] Frequency[Hz]
0.6 38,000 C22 22 1.5
1.07 68,000 C27 27 2.49

not detected in the computations. The behavior of the frequency as a function of

the Reynolds number is predicted in the right way, but the absolute frequencies
computed are higher than the experimental ones. So far, only the frequency ratio
ŒFreq.Re D 38; 000/=Freq.Re D 68; 000/
EXP D 0:602 of the measured data show
good agreement with the ratio ŒFreq.Re D 38; 000/=Freq.Re D 68; 000/
CFD D
0:604 obtained from the numerical investigations (grid 2).
In Fig. 14 the computed instantaneous velocity field and the phase–averaged
velocity field obtained by measurements are shown for different time–phase angles.
The time–phase angle is defined as tpa D t=T  360ı with T denoting the swiveling
period and t describing the real time passed within the period. Being aware of com-
paring averaged with instantaneous data sets, nevertheless basic differences in the
results can be detected. The vortex generated on the suction side of the inclined plate
is of larger scale in the computations than for the experimental case for tpa  300ı.
This vortex travels downstream and with it a zone of low pressure. When passing
the center of rotation the low pressure region together with the high pressure on the
pressure side of the plate reduces the rotational speed of the plate. Finally, the direc-
tion of rotation is changed and the plate tends towards the next angular amplitude
passing the point of origin at tpa D 360ı . Here, the vortex is going to leave the
plate region for the experimental case (Fig. 14(d)). However, in the simulation the
vortex still maintains on the former suction side, close to the trailing edge, affecting
the movement of the plate. Obviously for this kind of test case, there is a strong cou-
pling between the vortices generated by the angular displacements on the one hand
and the effect of the vortices on the angular displacement within the next instant of
244 M. Münsch and M. Breuer

75 75

50 50
U / U*[–] U / U*[–]
25 25
1.6

y [mm]
1.6
y [mm]

1.2
0 1.2 0
0.8
0.8
0.4
–25 0.4 – 25
0
0
–50 – 50

–75 – 75
–25 0 25 50 75 100 125 150 175 200 225 – 25 0 25 50 75 100 125 150 175 200 225
x [mm] x [mm]
(a) Computation: tpa D 303ı (b) Computation: tpa D 360ı

75 75

50 50
U / U*[–] U / U*[–]
25 25
1.6 1.6

y[mm]
y [mm]

1.2
0 0 1.2
0.8 0.8
–25 0.4 –25 0.4
0 0
–50 –50

–75 –75
–25 0 25 50 75 100 125 150 175 200 225 –25 0 25 50 75 100 125 150 175 200 225
x[mm] x[mm]
ı ı
(c) Experiment: tpa D 300 (d) Experiment: tpa D 360

Fig. 14 Computed (instantaneous) and experimentally measured (phase–averaged) velocity fields

and plate deflections at two different time instants of the swiveling motion at Re D 68; 000.

time on the other hand. In spite of extensive numerical investigations taking different
resolutions, different numerical algorithms for the structure and a variety of other
issues not presented here into account, up to now, the reason for the over-prediction
of angular displacements and frequencies has not been found. Beside further numer-
ical improvements such as grid quality or pressure–correction on curved grids, also
the starting procedure for recomputing the experimental cases has to be considered.
Whereas the fluid and the movement of the plate is accelerated over a certain time
interval in the experimental case, the fluid is already initialized with the desired flow
velocities and the plate is released not until the flow computation is settled for the
numerical predictions. Furthermore, the influence of damping effects in the bear-
ings of the swiveling plate was investigated but could not fully excluded up to now.
Finally, the small gap between the plate and the channel walls mentioned above for
the experimental setup could not be replicated in the simulations and might be a
reason for the deviations observed.

7 Benchmark Test Cases

Within the activities of the DFG research group FOR 493 several benchmark test
cases have been developed (see [23]). Here, pure CFD as well as pure CSD test
cases, and numerical FSI test cases within the laminar flow regime are available [23]
Numerical Simulation of Fluid–Structure Interaction Using Eddy–Resolving Schemes 245

to be solved with the fluid–structure interaction algorithms developed. Furthermore,

also experimental test cases for laminar as well as turbulent flows were investi-
gated and are available for benchmark purposes, see [16, 17] and the corresponding
contribution in this issue. First verifications of the codes developed have been per-
formed on the pure fluid dynamic applications (named CFD1–CFD3 in [23]) or
structure mechanics benchmarks (named CSM1–CSM3, see also the paper in this
issue). Since good agreement with the proposed reference data sets was obtained,
fully coupled FSI computations on numerical as well as experimental benchmark
test cases have been carried out in the present project.

7.1 Numerical Benchmark FSI3

For testing the coupling algorithm developed (see Sect. 3.1) as well as the data
transfer and interpolation strategies (see Sect. 3.2), the benchmark test case named
FSI3 in [23] and depicted in Fig. 15 was considered. In a channel of length L D 2:5
m and a height of H D 0:41 m a cylinder is mounted. The cylinder position is
slightly off–centered, with the cylinder center located at 0:2 m downstream of the
inflow section and with a distance of 0:2 m from the lower lateral channel wall. The
cylinder has a diameter of D D 0:1 m. With a mean inflow velocity of U1 D 2 m/s
and a kinematic viscosity of  f D 0:001 m2 /s a Reynolds number of Re D 200
is obtained. The density of the incompressible and Newtonian fluid is set to f D
1000 kg/m3 . In the wake of the cylinder, a flexible beam is attached to the cylinder.
The beam is of length l D 0:35 m and has a thickness of h D 0:02 m. The cylinder
is assumed to be fixed and rigid, whereas the material of the beam is of St. Venant–
Kirchhoff type characterized by a Poisson ratio of  s D 0:4, a Young modulus of
E D 5:6  106 kg/(m s2 / and a density of s D 1000 kg/m3 .
For the flow predictions two different grids with either only 27; 040 or 108; 160
CVs in a 2D plane were used, respectively. Since the code is three–dimensional,
in total 108; 160 CVs or 865; 280 CVs were used in 3D. Since for low Reynolds
numbers the stability of explicit time–marching schemes is dominated by viscous
effects, either the computational grid has to be coarse or the time–step size has to be

U
H
l

Fig. 15 Setup of the computational domain for the FSI 1–3 test case.
246 M. Münsch and M. Breuer

rigorously reduced or a combination of both has to be applied to stay in the stable

limits1 . Therefore, relatively coarse grids in combination with time–step sizes of
tf D 9  105 s (coarse grid) and tf D 2  105 s (medium grid) were chosen to
predict the flow field. At the inflow a parabolic velocity profile was set with:

y.H  y/
ui n .y/ D 1:5U1 : (14)
. H2 /2

At the lateral boundaries as well as for the structure, i.e. the fixed cylinder and the
beam, no–slip boundary conditions were applied. In spanwise direction slip con-
ditions were chosen. At the channel outlet a convective outflow condition with
Uc D U1 was specified. The response of the structure was computed with the
finite–element code Carat [24] provided by the Chair of Structural Analysis of TU
Munich. Here, the beam was modeled using in total 30 four–noded shell elements.
The time–step size of the solver for the structure was adapted to the time–step size
of the fluid solver, i.e., ts D tf . For the code coupling the coupling tool CoMA
[11] was used as described in Sect. 3.2. Within the coupled FSI prediction a constant
underrelaxation factor of ˛ D 0:5 was considered for transferring the computed dis-
placements from the structure domain to the fluid domain. For this case, a first–order
extrapolation for the estimation of the structural displacements at the beginning of
each time step was performed (see Sect. 3.1).
In the present configuration, vortex shedding occurs at the cylinder. These vor-
tices travel downstream and start to interact with the flexible structure which leads
to an oscillation of the beam. The resulting displacements are depicted in Fig. 16(a)
and 16(c) for point A located at the trailing edge on the center line of the beam
(medium grid). After some transition phase the amplitude of the oscillation reaches
constant values for the x– and y–components (see Fig. 16(b) and 16(d)). The mean
values and the amplitude of the displacements in both direction are given in Table 5
for both grids. It is obvious that the displacements in y–direction are about one order
of magnitude larger than those in x–direction. Furthermore, an improvement of the
results of the refined grid is visible with respect to the reference data. The frequen-
cies of the displacements were determined to fx D 10:12 Hz and fy D 5:05 Hz
on the medium grid, respectively. The actual reference value [23] is specified with
fx;ref D 10:93 Hz for the x–component and fy;ref D 5:46 Hz for the y–component.
Due to the off–centered location, the structure is affected by lift forces right from
the beginning of the computation. The resulting lift and drag forces are depicted
in Fig. 17. For the lift minimum and maximum values of Lmi n D 172:5 N and
Lmax D 204:9 N are predicted on the medium grid. The DFT delivers a frequency of
fL D 5:05 Hz, which is equal to the oscillation frequency of y. For the drag force
corresponding results of Dmi n D 428 N and Dmax D 507 N have been obtained.
Similar to the predicted frequency value of y the DFT leads to fD D 10:12 Hz.

1
Since the FSI coupling scheme is mainly intended for high–Re flows within the LES context, this
restriction solely applies to the present test case.
Numerical Simulation of Fluid–Structure Interaction Using Eddy–Resolving Schemes 247

0 0

–0.002
– 0.002
–0.004

–0.006 – 0.004

Δx(m)
Δx (m)

–0.008
– 0.006
–0.01

–0.012
– 0.008
–0.014

16.5 17 17.5 18 18.5 19 19 19.1 19.2 19.3 19.4

t (s) t (s)
(a) x–Displacement: (b) zoomed x–Displacement:

0.06

0.04
0.04

0.02 0.02
Δy (m)

Δy(m)

0 0

–0.02
– 0.02

–0.04
– 0.04

–0.06
16.5 17 17.5 18 18.5 19 18.95 19 19.05 19.1 19.15 19.2 19.25 19.3 19.35 19.4
t(s) t(s)
(c) y-Displacement: (d) zoomed y-Displacement:

Fig. 16 Time histories of the displacements x and y obtained for the FSI3 benchmark test
case on the medium grid.

The actual reference values are Lmi n;ref D 151:4 N and Lmax;ref D 156:4 N for
the lift force with a frequency of fL;ref D 5:46 Hz and Dmi n;ref D 432:7 N and
Dmax;ref D 488:2 N for the drag with a reference frequency of fD;ref D 10:93 Hz.
The deviations between the predicted results and the reference values are obvious.
They in general decrease with increasing grid resolution. Taking the restrictions on
the grid resolution into account, the predicted results are within the expected range.
These restrictions on grid resolution and time–step size are, by far, less severe when
applying the algorithm to turbulent flows connected with higher flow velocities (see
Sect. 7.2). Nevertheless, by performing this test case the proper behavior of the
whole partitioned FSI setup was proven.
248 M. Münsch and M. Breuer

Table 5 Results of numerical benchmark FSI3 on two different grid levels; forces are given for a
reference width of 1 m (values in parenthesis are the corresponding frequencies in Hz).
x [103 m] y [103 m] Drag [N] Lift [N]
Coarse 5.18 ˙ 5.04 [10.14] 1.12 ˙ 45.1 [4.99] 477.0 ˙ 48.0 [10.14] 7.0 ˙ 223.0 [4.99]
Medium 4.54 ˙ 4.34 [10.12] 1.50 ˙ 42.5 [5.05] 467.5 ˙ 39.5 [10.12] 16.2 ˙ 188.7 [5.05]
Ref. [23] 2.88 ˙ 2.72 [10.93] 1.47 ˙ 34.9 [5.46] 460.5 ˙ 27.74 [10.93] 2.50 ˙ 153.91 [5.46]

400
200

150
200
100

50
Lift (N)

Lift (N)
0
0

–50
–200 –100

–150

–400 –200
16.5 17 17.5 18 18.5 19 19 19.05 19.1 19.15 19.219.2519.319.3519.4
t (s) t (s)
(a) Lift (b) zoomed Lift

550
500

480
Drag (N)
Drag (N)

500

460

450
440

420
16.5 17 17.5 18 18.5 19 19 19.05 19.1 19.15 19.2 19.25 19.3 19.35 19.4
t (s) t (s)
(c) Drag (d) zoomed Drag

Fig. 17 Time histories of lift and drag obtained for the FSI3 benchmark test case on the medium
grid.

7.2 Turbulent Experimental Benchmark

First steps have been taken to compute a FSI case which was defined to serve as
a benchmark test case supported by experimental investigations on FSI in laminar
and turbulent flow regimes [16]. On the centerline of a channel a rigid cylindrical
front body with diameter D D 22 mm is mounted which is joined to a flexible steel
membrane of 0:04 mm thickness. A rigid rectangular rear mass is attached to the
membrane and forms the trailing edge of the whole structure. The rear mass has a
Numerical Simulation of Fluid–Structure Interaction Using Eddy–Resolving Schemes 249

0.1 U / U∞
1.6
1.5
0.05 1.4
1.3
1.2
1.1

Y [m]
1
0 0.9
b 0.8
U D
0.7
H 0.6
0.5
l – 0.05 0.4
0.3
0.2
0.1
– 0.1

– 0.05 0 0.05 0.1 0.15 0.2 0.25

L X [m]

(a) (b)

Fig. 18 (a) Setup of the computational domain; (b) Normalized velocity magnitude and stream-
lines of the fixed structure at Re D 14; 885.

thickness of 4 mm and a length of b D 10 mm. In total, the length of the membrane

and the rear mass is specified with l D 60 mm. The monovariant structure is free to
rotate around an axle located in the center point of the cylindrical front body. The
densities are specified with 2828 kg/m3 for the aluminum front body, 7855 kg/m3
for the stainless steel membrane and 7800 kg/m3 for the stainless steel rear mass. In
[16] the Young modulus of the membrane is given with 200 kN/mm2. The channel,
i.e. the fluid domain in the simulation, has a length of L D 15:4D and a height of
H D 10:9D. The center of the cylindrical front body is located at 2:5D downstream
of the inflow section. The computational domain is depicted in Fig. 18(a).
Water with a density of  D 997kg/m3 and a dynamic viscosity of  D
1:002  103 kg/ms was chosen. With an inflow velocity of U1 D 0:68 m/s a
Reynolds number of Re D 14; 885 is obtained by taking the cylinder diameter
D as reference length. The fluid domain was resolved by 6; 503; 760 control vol-
umes considering a resolution of 121 control volumes in spanwise direction. The
wall–nearest control volume center is located at y C D 3 at the trailing edge of the
structure, i.e. at the rear mass. The prediction of the turbulent flow field was done
via LES applying the Smagorinsky model (Cs D 0:1) with Van-Driest damping.
At the inflow section a velocity block profile with U1 was applied. An convective
outflow boundary condition was used with a convective velocity of Uc D U1 . The
channel walls as well as the structure have been set to no–slip boundary conditions.
In a first step, the structure is assigned to be fixed. The resulting normalized velocity
magnitude is depicted in Fig. 18(b). In a next step, the membrane and the rear mass
are released and allowed to move freely. The computation of the structure response
is done as in the FSI3 test case via the finite–element code Carat [24]. The cou-
pling was again done via the coupling tool CoMA [11]. For a first check of the code
coupling, a reduced grid resolution of 812; 970 control volumes was applied for the
fluid domain to compute the whole FSI problem on an eight–core local machine.
As consequence spurious pressure oscillations appeared in the region of the flexible
membrane and thus no valid FSI data set was obtained yet. In a prospective step the
grid size will be refined again and the whole structure will be released, i.e. also the
250 M. Münsch and M. Breuer

cylindrical front body. As a result of this preliminary test, the whole setup was tested
to be ready for high–performance computing now, which will become necessary for
solving this problem properly.

8 Conclusions

The computation of fluid–structure interaction in turbulent flows requires advanced

techniques such as LES. To resolve the turbulent flow properly in time and to
account for the spectrum of spatial scales in the flow field, which also might have
an influence on the response of the structure, for LES small time–step sizes are
inevitable. Thus, explicit time–marching algorithms are favored. Within this basic
framework a coupling scheme was developed which conserves the character of this
explicit time–marching schemes on the one hand and serves for a strong coupling
between fluid and structure on the other hand. This is achieved by directly connect-
ing the computation of the structure subproblem with the corrector step of the fluid
subproblem leading to a dynamic equilibrium between fluid and structure.
Furthermore, to step towards FSI applications replying on LES, the influence
of deforming grids on the quality of LES predictions in the context of the ALE
approach have been investigated. For this purpose, grid movements with specific
amplitudes and periods have been forced to a plane channel flow at Re D 590.
This results provide hints on what kind of grid motions are tolerable within an FSI
application using LES.
Since the grid quality, which should be maintained during the movement or defor-
mation of the structure, is found to be an important issue, a lot of effort was spent on
grid smoothing techniques. Here the dilemma is twofold. On the one hand orthog-
onality and smoothness of the grid are known to be of special interest in the LES
context and therefore it is worth to spend CPU–time for this issue. On the other hand
adjustments of the grids might be necessary within each subiteration. Thus the pro-
cedure for each smoothing step is not tolerated to be highly CPU–time consuming.
Besides methods relying on transfinite interpolations also elliptic grid smoothing
techniques based on composite mapping were taken into consideration. The com-
posite mapping consists of a nonlinear transfinite algebraic transformation and an
elliptic transformation. Special emphasis had to be put on maintaining the grid qual-
ity near boundaries and block connectivities, everything under the constraint of low
computational costs per iteration or time step.
Results of the new coupling scheme developed have been presented for various
cases, e.g., the flow around cylindrical structures for laminar (Re D 200) and sub-
critical (Re D 13; 000) flows. Here, the response of the structure, i.e. the cylinders,
was determined by solving equations of motion. For the laminar test case the results
are found to be in good agreement with [26]. The sub-critical case was designed to
compare the results obtained by the explicit scheme developed with the result deliv-
ered by a classical scheme based on implicit time-marching [13,14]. Here, satisfying
results have been obtained for the displacement magnitudes and the frequencies of
Numerical Simulation of Fluid–Structure Interaction Using Eddy–Resolving Schemes 251

the cylinder oscillations. Aiming at the verification with experimental results, the
response of a loosely mounted flat plate confined to turbulent flows at Reynolds
numbers of Re D 38; 000 and 68; 000 was computed. Here, different grid reso-
lutions, extensions of the plate in spanwise direction and solution algorithms for
solving the structure subproblem were investigated. The presented results deviate
from the experimental findings. Consequently, further investigations considering the
influence of the channel length, the influence of the gaps between plate and chan-
nel walls and the proper numerical realization of the experimental startup procedure
have to be performed.
The coupling algorithm developed was extended towards code-to-code commu-
nication with external codes, since proper computation of the structure domain
requires more convenient approaches, such as finite-element solvers. Here, coupling
with the finite-element solver Carat [24] by the use of the coupling tool CoMA [11]
was enabled leading to the computation of benchmark test cases [23]. For the numer-
ical benchmark case of the laminar flow (Re D 200) around a cylinder with an
attached flexible beam, satisfactory results have been achieved. Here, restrictions on
the grid resolution and the time step have to be taken into account to apply explicit
time–marching on low Reynolds number flows. Nevertheless, the proper function-
ality of the developed partitioned fluid-structure approach was shown. Thus, these
developments were finally applied to compute an experimental benchmark test case
in turbulent flow. Due to the complexity of the structure and a huge number of CVs
in the fluid domain this extensive simulation is still in progress. In summary, sig-
nificant steps have been done to marry LES with FSI. Reasonable results have been
obtained for numerical test cases. The computational reproduction of experimentally
obtained results is still challenging in connection with FSI. Here, further investiga-
tions are necessary to investigate the influence of parameters, which are usually
simplified or not considered within a CFD approach up to now, i.e. the experimental
startup processes or small gaps or structural tolerances.

Acknowledgements The project is financially supported by the Deutsche Forschungsgemein-

schaft within the research group ’Fluid–Struktur–Wechselwirkung: Modellierung, Simulation,
Optimierung’ (FOR 493) under contract number BR 1847/6. We gratefully acknowledge the coop-
eration with the Chair of Structural Analysis of the Technical University Munich providing the
codes Carat [24] and CoMA [11] including intensive support, especially Dipl.–Ing. Th. Gallinger
and Dr.–Ing. R. Wüchner.

References

1. Breuer, M.: Direkte Numerische Simulation und Large–Eddy Simulation turbulenter

Strömungen auf Hochleistungsrechnern. Habilitationsschrift, Universität Erlangen–Nürnberg,
Berichte aus der Strömungstechnik, ISBN 3–8265–9958–6, Shaker Verlag, Aachen, (2002).
2. Breuer, M.: A Challenging Test Case for Large–Eddy Simulation: High Reynolds Number
Circular Cylinder Flow. Int. J. Heat Fluid Flow 21, 648–654 (2000)
3. Breuer, M.: Large–Eddy Simulation of the Sub-Critical Flow Past a Circular Cylinder:
Numerical and Modeling Aspects. Int. J. Numer. Meth. Fluids 28, 1281–1302 (1998)
252 M. Münsch and M. Breuer

4. Breuer, M., Münsch, M.: LES Meets FSI – Important Numerical and Modeling Aspects, Proc.
of the Seventh Int. ERCOFTAC Workshop on DNS and LES: DLES-7, Trieste, Italy, Sept. 8–
10, 2008, ERCOFTAC Series, vol. 13, Direct and Large–Eddy Simulation VII, eds. Armenio,
V., Geurts, B., Fröhlich, J. Springer, Netherland (2010)
5. Breuer, M., Münsch, M.: Fluid–Structure Interaction Using LES – A Partitioned Cou-
pled Predictor–Corrector Scheme. Proc. in Applied Mathematics and Mechanics, PAMM, 8,
10515–10516 (2008)
6. Breuer, M., Münsch, M.: FSI of the Turbulent Flow around a Swiveling Flat Plate Using Large-
Eddy Simulation, In: Proc. of the Int. Workshop on Fluid-Structure Interaction (2008), eds.
Hartmann, S., Meister, A., Schäfer, M., Turek, S., 31–42, Kassel University Press, ISBN 978-
3-89958-666-4, Kassel (2009)
7. Demirdžić, I. and Perić, M.: Finite–Volume Method for Prediction of Fluid Flows in Arbitrarily
Shaped Domains with Moving Boundaries. Int. J. Numer. Meth. Fluids 10, 771–790 (1990)
8. Demirdžić, I. and Perić, M.: Space Conservation Law in Finite–Volume Calculations of Fluid
Flow. Int. J. Numer. Meth. Fluids 8, 1037-1050 (1988)
9. Durst, F., Schäfer, M.: A Parallel Block–Structured Multigrid Method for the Prediction of
Incompressible Flows. Int. J. Numer. Meth. Fluids 22, 549–565 (1996)
10. Durst, F., Schäfer, M., Wechsler, K.: Efficient Simulation of Incompressible Viscous Flows
on Parallel Computers. In: Hirschel, E.H. (ed.) Flow Simulation with High–Performance
Computers II, pp. 87–101. Vieweg, Braunschweig (1996)
11. Gallinger, T., Kupzok, A., Israel, U., Bletzinger, K.-U., Wüchner, R.: A Computational
Environment for Membrane–Wind Interaction. Int. Workshop on Fluid–Structure Interaction:
Theory, Numerics and Applications, Herrsching am Ammersee, Germany, Sept. 29 – Oct. 1,
(2008)
12. Germano, M., Piomelli, U., Moin, P., Cabot, W.H.: A Dynamic Subgrid Scale Eddy Viscosity
Model. Phys. Fluids A 3, 1760–1765 (1991)
13. Glück, M., Breuer, M., Durst, F., Halfmann, A., Rank, E.: Computation of Wind–Induced
Vibrations of Flexible Shells and Membranous Structures. J. Fluids Structures 17, 739–765
(2003)
14. Glück, M., Breuer, M., Durst, F., Halfmann, A., Rank, E.: Computation of Fluid–Structure
Interaction on Lightweight Structures. Int. J. of Wind Eng. Indus. Aerodyn. 89, 1351–1368
(2001)
15. Patankar, S.V., Spalding, D.B.: A Calculation Procedure for Heat, Mass and Momentum
Transfer in Three–Dimensional Parabolic Flows, Int. J. Heat Mass Transfer, 15, 1787–1806
(1972)
16. Pereira Gomes, J., Lienhart, H.: Experimental Study on a Fluid–Structure Interaction Reference
Test Case. In: Bungartz, H.-J., Schäfer, M. (eds.) Fluid–Structure Interaction, Lecture Notes
Comput. Sci. & Eng., LNCSE 53, pp. 356–370, Springer, Heidelberg (2006)
17. Pereira Gomes, J., Münsch, M., Breuer, M., Lienhart, H.: Flow–induced Oscillation of a Flat
Plate — A Fluid–Structure Interaction Study Using Experiment and LES, 16. DGLR–Fach–
Symposium der STAB, RWTH Aachen, Germany, Nov. 3–5, 2008, In: Notes on Numerical
Fluid Mechanics and Multidisciplinary Design, in press, Springer, Berlin
18. Hilgenstock, A.: A Method for the Elliptic Generation of Three-Dimensional Grids with Full
Boundary Control. Deutsche Forschungsanstalt für Luft– und Raumfahrt (DFLVR), DFVLR-
IB A 09, 221–87 (1989)
19. Leonard, S., Terracol, M., Sagaut, P.: Commutation Error in LES with Time–Dependent Filter
Width. Computers & Fluids 36, 513–519 (2007)
20. Lilly, D.K.: A Proposed Modification of the Germano Subgrid Scale Closure Method. Phys.
Fluids A 4, 633–635 (1992)
21. Smagorinsky, J.: General Circulation Experiments with the Primitive Equations. I: The Basic
Experiment. Month. Weath. Rev. 91, 99–165 (1963)
22. Spekreijse, S.P.: Elliptic Grid Generation Based on Laplace Equations and Algebraic Transfor-
mations. J. of Comput. Physics 118, 38–61 (1995)
Numerical Simulation of Fluid–Structure Interaction Using Eddy–Resolving Schemes 253

23. Turek, S., Hron, J.: Proposal for Numerical Benchmarking of Fluid–Structure Interaction
between an Elastic Object and Laminar Incompressible Flow. In: Bungartz, H.-J., Schäfer,
M. (eds.) Fluid–Structure Interaction, Lecture Notes Comput. Sci. & Eng., LNCSE 53, pp.
371–385, Springer, Heidelberg (2006)
24. Wüchner, R., Kupzok, A., Bletzinger, K.-U.: A Framework for Stabilized Partitioned Analysis
of Thin Membrane-Wind Interaction. Int. J. Numer. Meth. Fluids 54, 945–963 (2007)
25. Yigit, S., Schäfer, M., Heck, M.: Numerical Investigation of Structural Behavior During Fluid
Excited Vibrations. REMN 16, 491–519 (2007)
26. Zhou, C.Y., So, R.M.C., Lam, K.: Vortex-Induced Vibrations of an Elastic Circular Cylinder.
J. Fluids & Structures 13, 165–189 (1999)
•
Partitioned Simulation of Fluid-Structure
Interaction on Cartesian Grids

H.-J. Bungartz, J. Benk, B. Gatzhammer, M. Mehl, and T. Neckel

Abstract This contribution describes recent developments and enhancements of

the coupling tool preCICE and the flow solver Peano used for our partitioned
simualtions of fluid-structure interaction scenarios. Peano brings together hard-
ware efficiency and numerical efficiency exploiting advantages of tree-structured
adaptive Cartesian computational grids that, in particular, allow for a very memory-
efficient implementation of parallel adaptive multilevel solvers – an efficiency
which is crucial facing the large computational requirements of multi-physics
applications and the recent trend in computer architectures towards multi- and
many-core systems. preCICE is the successor of our coupling tool FSI❄ce and
offers a solver-independent implementation of coupling strategies and data mapping
functionalities for general multi-physics problems. The underlying client-server-
like concept maintains the full flexibilty of the partitioned approach with respect
to exchangeability of solvers. The data mapping relies on fast spacepartitioning
tree algorithms for the detection of geometric neighbourhood relations between
components of non-matching grids.

1 Introduction

The importance of multi-physics simulations becomes evident looking at applica-

tions such as aerodynamic design optimisation, noise control, biomedical flow and
structure simulations, and many others. Within the last years, computing power has
reached a state that allows to compute more and more realistic and detailed models
for such problems – provided that efficient simulation codes are available as stabil-
ity and accuracy issues involved still are very challenging for standard approaches.
Efficiency here refers to numerical methods as well as hardware-efficient

H.-J. Bungartz, J. Benk, B. Gatzhammer, M. Mehl, and T. Neckel

Department of Computer Science, Technische Universität München, Boltzmannstr. 3, 85748
Garching, Germany
e-mail: fbungartz,gatzhamm,mehl,[email protected]

H.-J. Bungartz et al. (eds.), Fluid Structure Interaction II, Lecture Notes 255
in Computational Science and Engineering 73, DOI 10.1007/978-3-642-14206-2 10,
c Springer-Verlag Berlin Heidelberg 2010
256 H.-J. Bungartz et al.

implementations and efficient software engineering. The latter is crucial to be able to

flexibly and easily establish, extend, or maintain simulation environments according
to new requirements such as improved models.
In terms of the software engineering aspect, partitioned approaches for the
simulation of multi-physics scenarios offer a great potential: They reuse existing,
established, tested, and trusted codes for single-physics problems and couple them
to a multi-physics simulation environment via suitable software tools and numerical
coupling strategies. In comparison to monolithic approaches tackling the whole sys-
tem in one code as one large set of equations, the partitioned approach thus allows
for a fast prototyping and for exchanging solvers according to new developments or
modified scenarios. This advantage, of course, does not come for free but raises the
new challenge of developing numerical coupling strategies that yield stable tran-
sient coupled simulations. The coupling of incompressible fluid flow to (elastic)
structures is one example where this holds in particular. Another numerical aspect
that has to be kept in mind – in particular with respect to physical conservation
laws – is the mapping of data between the in general non-matching solver grids. But
also the technical realisation of the code coupling, which may seem to be a minor
issue, is non-trivial if the whole flexibility of the partitioned approach is to be main-
tained. This means, for example, that the independence and black-box property of
the involved codes must not be destroyed by data mapping routines that directly con-
nect the grid information of two solvers. Also the implementation of the coupling
strategy and control in one (or even both) of the solvers would contradict the idea
of a partitioned approach. Thus, developing a new, separate coupling component is
favourable.
This paper focusses on the simulation of fluid-structure interactions as a spe-
cial case of surface-coupled multi-physics applications. We developed the coupling
tool preCICE (precise Code Interaction Coupling Environment) for such kinds of
coupled simulations which aims at maintaining the full flexibility of the partitioned
approach and provides a variety of coupling and data mapping functionalities as well
as clearly defined interfaces for the implementation of additional variants. From the
point of view of computing time and memory requirements, the simulation of the
fluid side is the most costly part of fluid-structure interaction simulations. Our flow
solver, implemented in the partial differential equation framework Peano, achieves
very low memory requirements and a high data access efficiency – which is essential
with memory access becoming a more and more severe bottleneck of any data-
intensive simulation – on the basis of octree-like adaptively refined Cartesian grids.
In addition to its memory efficiency, the Peano flow solver offers a state-of-the-art
numerical efficiency: dynamical grid adaptivity, multigrid solvers, load balanced
domain decomposition, and accurate force computations at the coupling surface.
These features are not trivial to combine with hardware efficiency.
In Sect. 2, we describe the coupling environment preCICE. Section 3 presents
the essential basics of the PDE framework Peano and details of the implementation
of the Navier-Stokes solver. Numerical results achieved with these two codes are
shown in Sect. 4.
Partitioned Simulation of Fluid-Structure Interaction on Cartesian Grids 257

2 The Coupling Environment preCICE

The development of the coupling environment preCICE (Precise Code Interaction

Coupling Environment) has been guided by three main principles: preserving the
inherent flexibility of the partitioned approach for multi-physics simulations, imple-
menting hardware- and numerically efficient methods, and providing a reasonable
user comfort. Of course, preCICE is not the first tool capable of coupling different
simulation codes. MpCCI [2] is probably the best known tool offering an MPI-like
programming interface and a rich data mapping and communication functional-
ity, but no implementation of coupling numerics. In addition, particular scientific
solutions for specific fields of applications exist such as The Model Coupling
Toolkit MCT [36], which is widespread in climate simulation, or Uintha [47], a
component-oriented realisation of a coupling unit. Although Uintha is restricted
to solvers based on Cartesian grids, the successful integration of a fluid-structure
interaction model [47] shows the general potential of component-based approaches.
These approaches establish the coupling tool as a separate software component with
functionalities going far beyond pure data mapping and communication, including
the coupling control and, for some coupling strategies, also the solution of interface
equations. The Component Template Library [45] is a computer science
oriented approach to realise distributed component-based simulations and provides
data communication for a wide variety of computer architectures and programming
languages.
In contrast to these existing tools, preCICE aims at bringing together the com-
ponent-based approach with an efficient data mapping technology, a general appli-
cability for a wide range of solvers and solver grids, and easy plug-in mechanisms
and exchangeability of solvers.

2.1 FSI❄ce – the First-Generation Coupling Tool

preCICE is an improved and generalised version of FSI❄ce (Fluid-Structure Inter-

action Coupling Environment). Developed and validated in [12] and [14], the main
goal of FSI❄ce is to separate coupling functionalities from the solvers to be coupled
and to pertain the flexibility of exchanging coupled solvers readily and indepen-
dently from each other. The implementation of the data mapping is left to the solver
codes, in order to allow for full customization. However, support is provided by
spacetree data structures accelerating neighbour search and projection tasks occur-
ing in data mapping. Coupling with FSI❄ce is realised in a client-server approach as
depicted in Fig. 1. FSI❄ce provides a coupling supervisor, an extra program instance
acting as client in the coupled simulation. The supervisor steers the coupled simu-
lation by sending simulation requests to the solvers which are seen as servers and
have to deliver a corresponding result back to the supervisor. The coupling supervi-
sor holds an explicit representation of the coupling interface in form of a triangulated
mesh. Solvers map their data to and from this coupling mesh only. Together with the
client-server approach, this allows to separate the internals of coupled solvers from
each other, such that an independent exchange of a solver can be performed.
258 H.-J. Bungartz et al.

Fig. 1 Client-server approach realised in the coupling tool FSI❄ce. The solvers are solely inter-
acting with the coupling supervisor. The coupling partner is, hence, logically decoupled from the
solver and can thus be exchanged independently.

2.2 Coupling Approach of preCICE

preCICE provides its functionality in form of a modular programming library with

defined interfaces for all main coupling tasks (cf. Sects. 2.3–2.6). Multi-physics sce-
narios to be simulated are specified by xml files. This makes preCICE applicable to a
wide range of surface-coupled multi-physics – not only fluid-structure interaction –
simulation scenarios. As FSI❄ce, preCICE uses an explicit representation of the
coupling surface by a triangulated mesh. In addition, preCICE provides a polygon
variant for 2D scenarios. In contrast to FSI❄ce, the data mapping is implemented
as a black-box functionality in preCICE. Except for data mapping, the geometric
information of the coupling surface can also be used as a geometry interface for
grid generation and modification of flow and structural solvers as it is done for the
Peano CFD solver (cf. Sect. 3).
An important difference between FSI❄ce and preCICE is related to the physical
deployment of the coupling components. In contrast to the client-server approach
of FSI❄ce, preCICE realises a peer-to-peer communication layout. The coupling
supervisor dedicated to the whole coupled simulation is replaced by local realisa-
tion of a coupling supervisor for every solver involved. (see Fig. 2 (a)). This local
control instance has all necessary information to steer the associated solver and to
exchange data with other solvers. The peer-to-peer communication layout avoids
duplicated communication of coupling data as in FSI❄ce and allows for a unified
implementation of the coupling control logic. In contrast to MpCCI, which also
works with a peer-to-peer communication concept, preCICE still includes both the
convergence control and the implementation of the coupling strategy and, thus, acts
as an independent numerical compenent. Despite of the peer-to-peer communication
Partitioned Simulation of Fluid-Structure Interaction on Cartesian Grids 259

Fig. 2 (a) Peer-to-peer concept of preCICE with local realizations of coupling supervisors and
direct communication between coupled solvers. (b) Parallel coupling concept of preCICE. P0–P3
are solver processes, while P4 and P5 are processes running preCICE. P4 acts as master, similar to
the coupling supervisor of FSI❄ce, while P5 is a worker taking up jobs from the master.

layout, the actual deployment of resources to computing processes is flexible. When

running a coupled simulation with parallelised solvers, e.g., the coupling supervisor
can be realised in separate processes as depicted in Fig. 2 (b). Via the proxy software
design pattern, this behaviour is hidden from a solver using preCICE.

2.3 Data Mapping

preCICE supports the exchange of data on the coupling surface of the involved
solvers. This can be a simple copying of data in case of matching grids or projec-
tions and interpolations if grids are non-matching. An overview and comparison
of data mapping methods can be found in [22]. There, the most important proper-
ties of a data mapping are defined to be (i) global conservation of energy over the
interface, (ii) global conservation of loads over the interface, (iii) accuracy, (iv) con-
servation of the order of the coupled solvers, and (v) computational efficiency. For
fluid-structure interaction, displacements u and forces f have to be mapped between
the discrete coupling surface representation of the fluid F and of the structure S .
The mappings can be written as

u F D H u uS ; fS D Hf fF ; (1)

where the matrix Hu denotes the mapping from the vector uS of discrete displace-
ment values at S to the vector uF of discrete displacement values at F , and Hf
determines the mapping of discrete forces fF on the fluid surface to forces fS on
the structure side. Global conservation of energy WF D fFT uF D fST uS D WS , of
total forces, and of rigid body motion is achieved if

Hf D H u T ; (2)
X
.Hf /ij D 1 8j; and (3)
i
X
.Hu /ij D 1 8i: (4)
j
260 H.-J. Bungartz et al.

Fig. 3 Mapping of data from a Cartesian grid to a preCICE coupling mesh in 2D, with step 1,
finding the closest projection point, and step 2, interpolating the data of the mapping point.

As preCICE uses its own, central explicit representation of the coupling surface,
it only has to provide data mapping from this coupling mesh to arbitrary solver grids
and vice versa. These mappings consist of two steps: first, given a grid point x from
the solver grid, neighbouring elements of the coupling mesh have to be determined.
Second, the data at vertices of the determined elements on the coupling mesh are
mapped to data d.x/ at point x of the solver grid (or vice versa, respectively).
Currently, a mapping with linear accuracy is implemented (see Fig. 3): for each
grid point x from the solver grid, a closest projection point xp on the preCICE
coupling mesh has to be found. The barycentric coordinates ˛; ˇ; and  of xp are
used to linearly interpolate the value d.x/ at x from the vertex values d1 ; d2 ; and
d3 of the coupling mesh triangle including the projection point xp :

d.x/ D ˛d1 C ˇd2 C d3 ; ˛ C ˇ C  D 1: (5)

˛; ˇ, and  denote the barycentric coordinates of xp in the respective triangle.1 For

the mapping from the solver grid to the coupling mesh, we use the transpose of this
mapping such that condition (2) is fulfilled. In addition, also the fulfillment of (3)
and (4) can be shown.

2.4 Spacetree Acceleration for Geometry Queries

The data mapping described in the previous section involves determining neigh-
bouring elements of the preCICE coupling mesh for solver grid points2 . preCICE
uses spacetrees to speed up such geometrical queries. The basic idea of a spacetree

1
In case xp is a vertex  D ˇ D 0 and in case xp lies on an edge  D 0, e.g.
2
A similar situation appears for other geometry queries used to set up the Cartesian solver grid of
our flow solver in Peano.
Partitioned Simulation of Fluid-Structure Interaction on Cartesian Grids 261

(a) (b) (c)

Fig. 4 (a) Part of a quadtree refined during a query for the closest projection point on the preCICE
coupling mesh. The query point is drawn as a circle, and the projection and interpolation steps of
the mapping are indicated by arrows. (b) Invalid situation where the spacetree cell with the mapping
point is not intersected by the coupling mesh due to a too deep spacetree refinement. (c) Special
case where the closest projection point found in the spacetree cell is actually not the closest on the
coupling mesh. The circle indicates the distance to the projection point.

is to recursively divide the space under consideration into smaller elements and to
use information inheritance to restrict expensive operations to local areas of interest.
The spacetree variants employed in preCICE are quad- and octrees as well as binary
kd-trees with different subdivision rules. We limit the discussion here to the case of
quadtrees in 2D.
In the beginning of a simulation, the quadtree consists of the root cell only, cov-
ering the whole computational domain. Depending on the location of the incoming
geometrical (projection) queries (cf. Fig. 4(a)), arbitrary, dynamical, and recursive
refinements are performed. A leaf cell of the quadtree is then split into four equally
sized children and the elements of the coupling mesh contained in the father cell
are distributed to the newly created children cells according to their location. An
element being partly contained in several quadtree cells is owned by all of these
cells. For the computation of a projection point, the quadtree is traversed in a top-
down manner. The tree structure allows to exclude cells from the search already at a
very high level such that small cells have to be examined only in a small area. This
leads to a reduction of the computational complexity from O.NM / in the case of
N solver grid points and M coupling mesh elements to O.N log M /.
The underlying locality principle works only if the solver grid points and the
coupling mesh are close to each other when compared to the cell size of the quadtree.
Fig. 4(b) shows an example, where the coupling mesh elements and the solver grid
point are not in the same quadtree cell, which is an invalid situation and has to be
avoided.
A special case appearing when searching for closest projection points is illus-
trated in Fig. 4(c): If the distance to the projection point is larger than the one to the
sides of the spacetree cell, there might be a closer projection point in a neighbouring
cell. This situation is circumvented by comparing the closest distance found to the
distance to the sides of the cell, and to redo the projection point search within all
cells of the parent quadtree-node, if required. This can lead to an uprising cascade
262 H.-J. Bungartz et al.

within the hierarchical structure of the quadtree. However, according to our experi-
ences (cf. Sect. 4.3), such cases occur only in O.1/ cases and are negligible for the
overall complexity.

2.5 Coupling Schemes

The partitioned coupling approach separates the system of equations describing the
whole multi-physics scenario into subsystems describing only single fields. In fluid-
structure interaction, we denote the overall equation system by M.f; u/ D 0 and
divide it into the fluid subsystem F .u/ D f and structure subsytem S.f / D u,
where u are the structural displacements and f are the fluid forces. The goal of cou-
pling schemes is to find a solution of the original system M.u; f / D 0 by solving
the subsystems F .u/ D f and S.f / D u only.
Explicit coupling schemes solve the systems F .u/ D f and S.f / D u once per
discrete simulation time step, either synchronised or staggered3. The updated cou-
pling surface values u and f are exchanged and used for the next time step. While
these schemes show very good computational efficiency compared to solving the
full system M.u; f / D 0, they become instable for many FSI scenarios, in particular
for incompressible fluids. This is due to the added-mass-effect described in [46] and
[24], which is predominant in incompressible flow FSI. Implicit coupling schemes
set up a fixed-point iteration for each timestep. The iteration can be written as

uinC1
C1
D Sn .Fn .uinC1 //; (6)

where n denotes the timestep and i the iteration count. The solution of M.u; f / D 0
is retrieved when convergence is achieved. However, these schemes are insta-
ble for the same problems as the explicit schemes. The remedy is to apply an
underrelaxation 0 < !  1 in modifying the structural displacements:

uQ i C1 D .1  !/ui C !ui C1 : (7)

Since convergence can be achieved only with very small relaxation factors ! result-
ing in a slow convergence, strategies for accelerating the convergence speed have to
be sought after. One such technique is the Aitken acceleration for vector sequences,
which is described in [35] and applied to FSI in [21]. The Aitken technique is an
adaptive underrelaxation with variable !. [57] applies Newton’s method to the cou-
pling surface values and uses reduced order models for fluid and structure solver to
compute the required Jacobians in a black box manner.
These coupling schemes can be enhanced by subcycling in either of the involved
solvers. This allows non-coupled intermediate timesteps and, thus, decouples the

3
Note that this may involve also implicit time integration within a field (intrafield).
Partitioned Simulation of Fluid-Structure Interaction on Cartesian Grids 263

time scales of individual fields. In fluid-structure interaction, the fluid solver usually
has much more restrictive stability requirements on the timestep length than the
structure solver. Subcycling is, hence, typically applied to the fluid solver.
Since preCICE provides its functionality in form of a black box library, only cou-
pling schemes fulfilling this requirement can be implemented. This has the benefit
that all implemented coupling schemes can be used for any solver combination. pre-
CICE currently implements staggered explicit and implicit coupling schemes with
constant or Aitken based relaxation. Implicit coupling iterations can be performed
with a constant number of iterations per timestep, with an absolute convergence
limit, or with a relative convergence limit. A solver is allowed to do arbitrary sub-
cycling, but has to obey the synchronised global timestep length given by preCICE.
All the implemented coupling schemes are decoupled from the solvers and can be
configured by providing xml-files to specific multi-physics scenarios such as FSI.
The real potential of preCICE is in the implementation of more sophisticated cou-
pling schemes involving more complicated numerics such as solvers for interface
equations. The realisation of such schemes will be the focus of future work in
preCICE.

2.6 Communication

A rather technical task in the coupling of partitioned multi-physics applications is

the exchange of coupling data (forces and displacements for FSI) and control infor-
mation between the coupled solvers. Typical communication means are file-based,
use the Message Passing Interface (MPI) [42], or socket communication libraries.
preCICE currently offers communication based on MPI. To meet the requirements
of different types of solver codes as good as possible, solvers can be coupled in three
different ways, which can be configured by providing xml-files.
The first way of coupling two solvers with preCICE is rather untypical for par-
titioned multi-physics simulations. It requires to compile the solver codes into one
executable and to start the executable with several processes, at least as many as
coupled solvers. The startup code has to separate the processes into the respective
solver codes, the rest of the communication between the solver processes is setup
by preCICE automatically.
The second approach works with separate solver codes which are run with one
MPI call. MPI then puts the processes of all solvers into the global communication
space and preCICE automatically separates the processes into the right groups and
setting up the connections between them. This approach does not work for parallel
solvers using the global communication space. Such solvers need to be modified to
set up a separate communication space for its processes in that case, which might
be tedious in some cases and can be impossible for commercial codes.
A third approach uses MPI 2.0 features and works similar to socket-based com-
munciation. The solvers can be started individually, and a connection is achieved
264 H.-J. Bungartz et al.

automatically by preCICE through MPI ports. This approach works fine with
parallelized solvers, but requires MPI 2.0 features which are not commonly installed
on supercomputers.

2.7 Application Programming Interface

The functionality described in Sects. 2.3 - 2.6 is provided to a solver code by the
application programming interface (API) of preCICE. The main interface is written
in C++. A C wrapper is also available to make preCICE work with solvers written
in C code. A simple application example for the preCICE API can be as follows:
precice::CouplingInterface cplInterface ( solverName );
cplInterface.configure ( xmlConfigurationFileName );
double limitDt = cplInterface.initialize ();
while ( cplInterface.isCoupledSimulationOngoing() ) {
limitDt = cplInterface.advance ( computedDt );
}
cplInterface.finalize ();

An object of class CouplingInterface has to be created and used to access

any coupling functionality. The coupled simulation is configured from an xml file,
where the coupling data, coupling mesh, spacetrees, involved solvers, and cou-
pling schemes are specified. The call of initialize() creates all required data
structures and initiates the contact to the other coupled solver(s). If a solver is not
starting the coupling simulation, coupling data is received at that point already. A
limit for the length of the first timestep is returned back. This limit is a maximum
limit, i.e. a solver is always allowed to do a smaller timestep, but will be eventu-
ally synchronized to a global timestep length by the limits set by preCICE. The
while loop represents the main timestepping loop of the solver. Here, the control
of the overall length of the coupled simulation is given to preCICE by using the
isCoupledSimulationOngoing() method. The computed data is exchanged
when calling advance(). The solver makes the computed timestep length avail-
able to preCICE, which is necessary to steer the simulation in time. Finally, the
communication and data structures are terminated in the call of finalize().
In order to apply implicit coupling schemes, additional routines need to be used,
since a caching of the solver state becomes necessary within the coupling iterations.
The API functions to be used have the following syntax:
cplInterface.requireAction ( actionName );
cplInterface.fulfilledAction ( actionName );
The first function makes queries of preCICE known to a solver. The second function
tells preCICE that the required action has been fulfilled by a solver and prevents
the invalid use of coupling functionalities. For implicit coupling, two actions are
required by a solver. The first action is to store the state of a solver, the second to
load the state of the solver.
Partitioned Simulation of Fluid-Structure Interaction on Cartesian Grids 265

The examples above do not show how a solver can acccess the coupling data.
This can be done by using the following methods
int forcesID = cplInterface.getDataID (nameForces);
cplInterface.readData (forcesID, pointCoords, force);
cplInterface.writeData (forcesID, pointCoords, force);
In order to access data, an identifier has to be retrieved from preCICE. Then, a
simple call to readData() or writeData() transfers coupling data from or to
the coupling mesh of preCICE. The actually applied mapping is hidden from the
solver code.
As can be seen from these examples, the preCICE API can be integrated in solver
codes with minimal effort and, in particular, requires neither the (re-)implementation
of data mapping nor of the coupling strategy4 .

3 The Cartesian Flow Solver in Peano

Our flow solver has been implemented in the PDE framework Peano. The idea of
Peano is to exploit the potential of octree-like adaptively refined Cartesian grids for
hardware and numerically efficient implementations of PDE solvers. Cartesian grids
have undergone a revival within the last years due to their simplicity and structured-
ness [26, 40, 51, 54, 58]. In some sense, Peano is the successor of our flow solver
F3F (see for example [14] for a description). In contrast to F3F, Peano allows for
adaptively refined grids as well as dynamical grid refinement, for the generic imple-
mentation of solvers for further PDEs and varying discretisations, and for multilevel
algorithms. We describe the underlying algorithmical concept of Peano in a very
short manner. For more detailed descriptions, refer to [13, 18, 19, 38, 44, 59].

3.1 Octree-Like Cartesian Grids

Octree-like grids are constructed in a (locally) recursive way: in each (local) refine-
ment step, each grid cell can be subdivided into a fixed and constant number of
child cells. This way, they correspond to a tree of grid cells, the so-called spacetree.
Octrees represent three-dimensional grids that divide each grid cell into eight equal
child cells. Peano uses slightly different grids that emerge from a partitioning into
three parts per coordinate direction and refinement step. We call the corresponding
trees (k D 3)-spacetrees. Fig. 5 shows examples of a two- and a three-dimensional
Peano grid together with the associated trees.
In Peano, data are assigned to grid vertices and grid cells. It is up to the
application to decide which data such as degrees of freedom and grid refinement

4
This holds as long as standard methods are sufficient. Otherwise, the preCICE interfaces can be
used to implement user-defined methods.
266 H.-J. Bungartz et al.

Fig. 5 Examples for (k D 3)-spactree grids together with the corresponding spacetrees. Left: two-
dimensional grid with maximal refinement depth three; right: three-dimensional grid with maximal
refinement depth two.

information are stored at these points. Hanging nodes, that is vertices at bound-
aries between different refinement depths that do not have the regular number of
neighbouring cells, are not allowed to hold degrees of freedom.
The discretisation and evaluation of all operators involved in a simulation is
realised in a strictly cell-oriented way. That is, the algorithm processes the grid
cell by cell and computes the part of the respective operator that can be calculated
solely with vertex and midpoint data of the current cell. For adaptively refined grids,
this offers the great advantage that these operator cell parts are evaluated in an equal
manner for all cells no matter whether they are located at a boundary between dif-
ferent refinement depths or not. For finite element discretisations, the cell part of the
operator is given by the integral of the discrete weak form over the current cell [10].
During a grid traversal, the cell-parts of the operators are accumulated to the com-
plete operator values. Hanging nodes get values of the respective unknown function
by interpolation before the cell-wise operator evaluation starts and deliver restricted
values of their cell-part operator values to father nodes after the operator evaluation
is completed.

3.2 Data and Parallelisation Algorithms

As a basis for the data access algorithms in Peano, we use the space-filling Peano
curve [50]. Space-filling curves have been established as a simple and efficient
tool for traversal and domain decomposition of various types of adaptively refined
grids (see for example [11, 18, 26, 27, 43, 48, 49, 53, 59, 60] and citations therein).
In fact, not the space-filling curve itself but discrete iterates of the curve define a
(processing) order of the grid cells. Figure 6 displays this for a two-dimensional
(k D 3)-spacetree grid traversed along the Peano curve.
Partitioned Simulation of Fluid-Structure Interaction on Cartesian Grids 267

Fig. 6 Two-dimensional adaptive (k D 3)-spacetree grid with the corresponding iterate of the
Peano curve.

The locality properties of the Peano curve ensure a quasi-optimal time locality
of data access. This leads to improved cache-efficiency in various codes described
in literature (see for example [4, 61]). The PDE framework Peano uses the Peano
curve not only as a processing order for the grid cells but also as a tool supporting
the construction of highly efficient data structures and data access algorithms. For
that purpose, further properties of the curve – that could not be shown to hold for
any Hilbert curve in 3D, e.g. – are decisive: The projection property (projections
to lower-dimensional manifolds are lower-dimensional Peano curves) and the palin-
drome property (projections to lower-dimensional manifolds from different sides in
space result in the same Peano curve but with changing direction). These two prop-
erties allow to base the whole data management on a few streams and stacks. Such
data-structures inherently enforce a spatially highly local data access, which further
improves the cache-hitrates of the code.
For parallelisation, Peano uses a parallel and balanced tree-based domain par-
titioning [18, 59] – again exploiting the locality properties of the Peano curve to
achieve both load-balancing and quasi-minimal communication costs (compare also
[11, 53, 61]).
As the second focus in the implementation of the Peano framework was on
numerical efficiency, it offers state-of-the art numerical methods: The hierarchi-
cal tree structure of the grid inherently includes multilevel data that are traversed
by a recursive version of the Peano curve in a top-down-bottom-up order. On this
basis, Peano provides geometric multigrid solvers in an efficient and natural way.
In addition, this traversal and the stack concept allow for an easy implementa-
tion of dynamical grid adaptivity without affecting the data access efficiency of the
code [39].
268 H.-J. Bungartz et al.

To fully exploit all benefits of the spacetree grid and the sophisticated data admin-
istration based on the Peano curve, the framework idea is not based on an iterator
concept as used for example in [5, 6] and [37], but relies on an event concept. That
is, the user can trigger calls of his functions at certain points during a grid traversal,
for example when entering or leaving a grid cell [59].

3.3 Discretisation of the Navier-Stokes Equations

Various approaches exist and are successfully applied which allows to compute
numerical solutions for incompressible flows represented by the incompressible
Navier-Stokes equations

@u
 u C .u  r/u C rp D 0 ; (8)
@t
r u D 0 : (9)

In Peano, we are just starting to investigate space-time methods (see [41], e.g.) dis-
cretising both space and time simultaneously in a single mesh via four-dimensional
adaptive Cartesian grids available in the framework. So far, Peano’s flow solver uses
the more common separated discretisation in space and time with a d -dimensional
computational grid (regular or adaptive, with d D 2 or d D 3) for the domain and
an additional 1D “mesh” for time integration where only the necessary active time
levels are actually stored.

3.3.1 Space Discretisation

To discretise the Navier-Stokes equations in space, two different approaches exist in

Peano’s flow solver: a low-order finite element method (FEM) and the interpolated
differential operators (IDO). The Q1 Q0 elements of the FEM approach with con-
stant pressure and d -linear velocities (see Fig. 7(a)) allow for a straigtforward and
efficient cell-wise evaluation of the operators and fit well with local and dynamical
adaptivity. Various boundary conditions are supported. For 2D, a divergence-free
variant of ansatz functions is available (for details see [7, 14, 44]). The concept of
IDO represents a higher-order discretisation scheme for parabolic, hyperbolic, and
elliptic PDEs originally proposed in [3] and further developed in [31–34], e.g. In
spite of a moderate support of the discretisation stencil of IDO, a wide range of
wave-numbers up to high values can numerically be resolved on a given grid result-
ing in advantageous stability properties. The basic idea of this scheme is to use the
governing equations in point-wise and in integrated form. Hence, a coupled system
of point-values, line-integrated data, and cell-integrated values is solved. From the
technical point of view, the successful realisation of IDO in Peano represents the
Partitioned Simulation of Fluid-Structure Interaction on Cartesian Grids 269

xQ
Qi,j+1 i,j+1 Qi+1,j+1

yQ xyQ yQ
i,j e i+1,j

xQ
Qi,j i,j Qi+1,j

(a) (b)

Fig. 7 Storage locations for the low order FEM approach in three dimensions (a): Trilinear veloc-
ities (dark-blue spheres) and constant pressure (light-blue box) stored on vertices and the center of
the cell, respectively. 2D IDO discretisation scheme (b) with vertex data, face- and cell-centered
values to be stored.

proof of concept that larger stencils with additional degrees of freedom on faces
(see Fig. 7(b)) may be evaluated efficiently via interlayed iterations over the grid.
Both discretisation types for the Navier-Stokes equations (8) and (9) in space
yield the following semi-discrete equations:
d uh
A C Duh C C.uh /uh  M T ph D 0 ; (10)
dt
M uh D 0 : (11)

The operators D; C.uh /, M T , and M correspond to diffusion, convection, gra-

dient, and divergence. A represents the mass matrix. The concrete form of these
operators depends, of course, on the type of discretisation (FEM or IDO, respec-
tively).

3.3.2 Time Discretisation

The continuity and momentum equations in their semi-discrete form (10), (11) rep-
resent a system of ordinary differential equations (ODE). Various ODE integration
methods exist which differ in their order (accuracy), their computational complexity,
and their storage demands (see [23, 28, 29, 52]). For efficiency reasons, spatial and
temporal error contributions have to be balanced. Therefore, the order of the ODE
integration method in use depends on the order of the spatial discretisation scheme
and vice versa.
The Peano flow solver implements an explicit time discretisation following
Chorin’s projection method (see [20]) decoupling the solution of the velocity and
270 H.-J. Bungartz et al.

pressure fields in two separate steps, similar to a predictor-corrector scheme. First,

an intermediate velocity is computed via (10) that does not yet respect the mass-
conservation constraint. Second, the pressure is computed from a Poisson-type
equation generated from the continuity equation (11) applied to the right hand side
of the momentum equation (10). This pressure is used to project the intermediate
velocity onto a mass-conservative velocity field according to (10). For the computa-
tion of the preliminary velocity, Peano uses forward Euler for FEM and the classical
explicit Runge-Kutta method of fourth order (in a low-storage variant, cf. [8, 9]) for
IDO space discretisation.
Besides the Chorin projection method, an implicit  method with automatic time
step size control (as proposed in [25]) is implemented depending on the scenario
under consideration, the solution of the resulting non-linear system in each time
step may pay off as it allows larger time steps avoiding stability restrictions.

3.4 Memory Consumption and Performance

Peano in general and its flow solver in particular possess low memory requirements
due to the efficient storage format of the grid data in the spacetree. The vast majority
of data is stemming from the degrees of freedom (DoF) such as velocity or pressure
data of the simulation. The automatic generation of the corresponding data classes
via the in-house tool DaStGen (see [16, 17]) allows for an additional packing of
data: While C++ stores specific data types in technically larger blocks of memory
than necessary (one byte for boolean values, e.g., depending on the computer archi-
tecture), DaStGen can compress several of such variables in one block and save
memory.
Of course, different space or time integration schemes for the flow solver such as
the ones described above require a different number of local data in vertices or cells.
Peano’s flow solver minimises the amount of data by choosing the concrete data in
use at runtime. Table 1 shows a survey on the memory consumption for a single
cell and vertex in two and three dimensions in case of the FEM space discretisation
and different time discretisation methods. We emphasise the low amount of mem-
ory limited by 200 bytes compared to typical requirements of several kilobytes in
comparable simulation packages using unstructured grids.

Table 1 Memory requirement for FEM discretisation with three different time integration schemes
on adaptive grids in Peano: forward Euler (FE), classical Runge-Kutta (RK), and adaptive 
method. The values indicated in bytes include, besides necessary attributes for grid management
and flow simulation, all variables used in the program.
dimension cell in bytes vertex in bytes
default  adap. FE RK  adap.
2D 40 48 92 124 144
3D 44 52 124 172 200
Partitioned Simulation of Fluid-Structure Interaction on Cartesian Grids 271

Concerning the performance of Peano and its flow solver, different benchmark
comparisons have been evaluated (see [44]). To summarise, a comparison of the
adaptive grid implementation and a standard regular grid implementation with lexi-
cographic grid traversal resulted in runtime overheads of 30–80% only (decreasing
with increasing problem size) for two-dimensional simulations, dropping to less
than 3% in 3D. Comparing such an adaptive setup with a tuned, specific-purpose
2D IDO code on merely regular grids showed an overhead of a factor of about 1.7–
7.9 (again decreasing with increasing problem size) which is in the order of what
can be expected. Thus, Peano’s flow solver is able to efficiently simulate adaptive
scenarios in two and three dimensions. Note that the grid adaptivity pattern — even
steep adaptivity near geometric boundaries — does not influence the runtime per
degree of freedom or grid cell.

3.5 Moving Geometry

Peano’s spacetree representation of the Cartesian grid is combined with a marker-

and-cell approach to define obstacles in the fluid. If such obstacles move in the
domain, a clever handling of subtrees allows to create new and delete old cells as
well as to switch markers (inner to outer or vice versa) by local computations only.
In addition to these grid updates, three additional numerical features have been
implemented in the flow solver: First, a mass-conservative interpolation is done
for velocities on newly created cells. Second, velocity data are restricted to parent
nodes when deleting cells/vertices. Third, an additional projection step similar to the
Chorin projection method may be carried out to generate a fully mass-conservative
flow field after velocity changes at boundaries (see [13] for details).
Figure 8 shows two snapshots of the computational grid in a scenario with an
obstacle moving in a channel5 .

3.6 Consistent Forces

Accurate forces at the surface between fluid and structure are crucial for the compu-
tation of FSI simulations. Therefore, we adapted the elegant and efficient method of
consistent forces proposed by Gresho and Sani in [25] to our requirements. We will
restrict all formulations to the case of a finite element space discretisation in the fol-
lowing. Note that there are analogue formulations for finite volume discretisations
(see [14, 15]). The basic idea is to use the residual of the momentum equations on
the structure surface S which has to be in equilibrium with the sustaining forces
Fs exerted by the structure on the fluid. This results in one minor drawback and two
major advantages: It is not possible to separate the pressure and viscous parts of the

5
This scenario has been used as a test case to compare the case of flow around a fixed obstacle
with its identical counterpart, a moving obstacle in a non-moving liquid (see [56]).
272 H.-J. Bungartz et al.

Fig. 8 2D adaptive Cartesian grid with moving, spherical obstacle at start (a) and end (b) position.

consistent forces. Typically, however, such a separation is not necessary for FSI sim-
ulations. Concerning the advantages, we get the desired force or stress values with
nearly no additional computation. All components in (15) have already been calcu-
lated by the flow solver for the boundary cells. Furthermore, all properties of the
discretisation concerning for example discretisation order or physical conservation
laws carry over to the force computation (see [25]).
To apply the method of consistent forces, the semi-discrete momentum equa-
tions (10) are enlarged to hold not only for the nodes i D 1; : : : ; N inside the fluid
domain, but also for nodes i D N C 1; : : : ; NS on the boundary S . Virtual degrees
of freedom are introduced which possess ansatz functions fQj and corresponding
coefficients FQj on these boundary points6 . The ansatz functions fQj are of the same
type as used for the velocities (bilinear/ trilinear, e. g.) and represent the sustaining
forces Fs as follows:

NS
X
Fs .x/ D FQj fQj .x/ 8x 2 S : (12)
j DN C1

Using the equilibrium between sustaining forces and residual of the momentum
equations for a boundary node i 2 fN C 1; : : : ; NS g in the weak form, we obtain

Z NS
X Z
fQi .x/  Fs .x/ ds D FQj fQi .x/  fQj .x/ ds;
S j DN C1 S
 
d uh
D A C Duh C C.uh /uh  M T ph ; (13)
dt i

6
For nodal basis functions, the coefficients FQj denote the local forces on the boundary nodes j .
Partitioned Simulation of Fluid-Structure Interaction on Cartesian Grids 273

which shows the coupling of the force coefficients FQj by a boundary mass matrix.
To avoid the solution of the linear system of equations (13) in each time step, this
mass matrix is lumped, resulting in
Z Z
fQi .x/  Fs .x/ ds D FQi fQi .x/ ds : (14)
S S

Hence, we obtain an explicit formula for the consistent local force coefficients
 
1 d uh
FQi D R A C Duh C C.uh /uh  M T ph ; i D N C1; : : : ; NS :
Q
fi .x/ ds dt i
S
(15)

To determine the global force on a part of the boundary S , one has to inte-
grate (12) using the coefficients FQi calculated by (15). This integration of the ansatz
functions can be realised analytically and is, thus, reduced to summing up weighted
force coefficients FQi .
From the point of view of implementation, the consistent forces hardly require
additional computations: The program only adds up contributions to the semi-
discrete momentum Eq. (10) which have already been calculated before during the
time-step.

4 Numerical Results/Coupled Simulations

4.1 Flow Simulations

Within the different benchmarks for fluid-structure interaction, a suite of three 2D

scenarios are included which are intended to validate the flow solver separately
(i.e. without any interaction). The setup consists of a channel with a fixed solid
obstacle (a cylinder joint with a cantilever) near the inlet (see [30]). The three scenar-
ios vary in the corresponding Reynolds number Re. The setups CFD1 and CFD2 at
Re D 20 and Re D 100 exhibit a stationary behaviour whereas CFD3 at Re D 200
is fully time-dependent. As reference values, the total forces on the obstacle (and
the frequency of the oscillations in the time-dependent case) are examined.
Peano uses the coupling mesh provided by the coupling library preCICE
described in Sect. 2 as a basis of the generation of its adaptive Cartesian grid and
also transmits the resulting force data to this mesh. To judge the influence of differ-
ent grid adaptivity patterns and also the general potential of aptively refined grids to
save degrees of freedom and, thereby, also computational costs, we used six different
setups concerning the underlying, a priori refined grids with a maximum spacetree
level of eight (see [44]). Fig. 9 and 10 show simulated pressure distributions for all
274 H.-J. Bungartz et al.

pressure
0.14
0.13
0.12
0.11
0.1
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01

(a) (b)
pressure pressure
1.8
1.6 5
1.4
1.2 3
1 1.5
0.8
0.6 0.5
0.4
0.2 – 0.12
0 – 0.5
– 0.2
– 0.4 –2
– 0.6
– 0.8 –4

(c) (d)
Fig. 9 Pressure disctributions in a part of the computational domain around the obstacle for the
CFD1–CFD3 benchmarks. (a) Adaptive mesh in use with maximum and minimum refinement level
eight and seven, respectively. (b) Steady-state solution of the CFD1 benchmark with Re D 20.
(c) Steady-state solution of the CFD2 benchmark with Re D 100. (d) Solution of the CFD3
benchmark with Re D 200 at t D 8:12.

Fig. 10 Pressure distribution of the FSI benchmarks CFD3 (Re D 200, cf. [30]) at t D 8:12. The
minimum and maximum tree level is 7 and 8.

three benchmark cases on adaptive grids with minimum and maximum refinement
depth seven and eight and a purely geometry-driven refinement.
Table 2 contains the resulting drag and lift forces of the two steady-state scenar-
ios CFD1 and CFD2. Up to the last line in this table, the grids used are refined only
at the boundary of the geometry as shown in Fig. 9. The force coefficients converge
towards the reference values. These results can be improved by using manually
designed refinement boxes (see Fig. 11 and the last line of Table 2). A compari-
son of the third (regular grid with depth seven) and the last line of Table 2 shows
the gain in accuracy of an adaptive grid compared to a regular grid with comparable
number of cells. Dynamical grid adaptivity, which is currently being implemented in
the Peano CFD solver will further reduce the number of cells required for a certain
accuracy.
Partitioned Simulation of Fluid-Structure Interaction on Cartesian Grids 275

Table 2 Results for the steady-state benchmark scenarios CFD1 and CFD2 at Re D 20 and
Re D 100 (cf. [30]). The force values in horizontal and vertical direction are denoted by fx and
fy . The “box” grid setup corresponds to the setup in Fig. 11.
max. min. #cells #vertices #total #hang. Re D 20 Re D 100
level level DoF nodes
fx fy fx fy
6 6 4294 4096 12486 0 15.35 1.854 141.7 15.60
7 6 5096 4737 14570 340 15.58 1.243 149.3 11.44
7 7 38680 38083 114846 0 15.57 1.248 149.5 11.57
8 6 7636 6833 21302 1256 14.48 1.137 137.1 10.46
8 7 41220 40179 121578 916 14.48 1.146 137.3 10.81
8 6 (box) 39268 38344 115956 1744 14.46 1.143 137.2 10.74
ref. data – – – – 14.29 1.119 136.7 10.53

Fig. 11 Visualisation of the adaptive grid with levels 6 to 8 using manual refinement boxes for the
FSI benchmark flow (cf. [30]). The colouring corresponds to the horizontal velocity values.

Table 3 Results of the time-dependent benchmark scenario CFD3 at Re D 200 (cf. [30]) for dif-
ferent adaptive grids. The force values in horizontal and vertical direction, fx and fy , are indicated
with mean value and oscillation amplitude, while the frequency is shown in the last column. The
“box” grid setup corresponds to the grid shown in Fig. 11.
max. min. #cells #total #hang. fx fy frequ.
level level DoF nodes
6 6 4294 12486 0 – – –
7 6 5096 14570 340 – – –
7 7 38680 114846 0 494.8 ˙ 7.57 12.97 ˙ 479.8 4.407
8 6 7636 21302 1256 433.0 ˙ 4.79 23.33 ˙ 328.7 3.997
8 7 41220 121578 916 434.6 ˙ 5.14 8.43 ˙ 381.7 4.332
8 6 (box) 39268 115956 1744 439.3 ˙ 6.25 7.25 ˙ 447.6 4.421
ref. data – – – 439.5 ˙ 5.62 11.89 ˙ 437.8 4.396

For the time-dependent setup at Re D 200 (CFD3), resulting forces are given in
Table 3. The computations have been carried out using identical grid layouts as for
the two steady-state setups CFD1 and CFD2. The first two runs with level combi-
nations (6,6) and (7,6) are not resulting in a stable oscillation due to the very coarse
mesh sizes. For finer resolutions, we observe good convergence towards the refer-
ence force data both in terms of mean values and amplitudes of the oscillations. The
276 H.-J. Bungartz et al.

time time
8.932 9.338 9.744 10.15 8.932 9.338 9.744 10.15
442
400
440 300

438 200

100
436
drag

lift
0
434
-100

432 -200

-300
430
-400
428
10000 15000 20000 25000 10000 15000 20000 25000
timestep timestep
(a) (b)

Fig. 12 Drag (a) and lift (b) force of the laminar flow around the FSI obstacle at Re D 200
(i.e. CFD3) in the time interval [8.932,10.15]. The adaptive Cartesian grid in use possesses a
maximum and minimum level of refinement of eight and seven, respectively.

correct frequency of the simulation results is indirectly given by the resulting values
of the Strouhal number. The time-dependent behaviour of the forces for refinement
levels (8,7) is visualised in Fig. 12. Note that there are tiny oscillations in the drag
force of the second minimum in each period. They show an amplitude of 0.0125
corresponding to 0.003% of the absolute drag values. Two different reasons might
cause this behaviour (unknown from Peano’s behaviour for the corresponding clas-
sical cylinder benchmark [55]): not accurate enough pressure solutions at those time
intervals (in fact, the limiting number of 200 CG iterations is reached), or a too large
time step size for the temporal discretisation error. These oscillations do not repre-
sent a major problem for the stability of the overall flow as they disappear quickly
after the second minimum.
The fluid-structure interaction benchmark scenarios are merely two-dimensional.
Peano and its flow solver support, of course, also 3D setups. A snapshot of stream-
lines of the well-known, laminar benchmark flow around a cylinder (see [55]) is
shown in Fig. 13.

4.2 Structure Simulations

As for the fluid solver, [30] also describes a set of pure computational structural
mechanics (CSM) benchmark scenarios to validate structural solvers. The scenarios
consist of a 2D setup with the same geometry as placed into the fluid domain of
the CFD tests, which is a fixed circle (cylinder) with a flexible attached cantilever.
Only a gravitational force is applied to the structure, which results in a bending
Partitioned Simulation of Fluid-Structure Interaction on Cartesian Grids 277

Fig. 13 Time-depending 3D flow field around a cylinder at Re D 100 (cf. [55]). Streamlines are
plotted in black over the coloured horizontal velocity distribution. The adaptive grid possesses a
refinement of level 6 in the total domain and of level 7 at the surface of the cylinder.

Table 4 Results and reference values of benchmark CSM1.

no elements no dofs Ax Œ103 m Ay Œ103 m
183 840 7.147 65.89
723 3142 7.177 66.04
2928 12138 7.184 66.08
11712 47698 7.186 66.10
187392 752962 7.187 66.10
ref. values 7.187 66.10

Table 5 Results and reference values of benchmark CSM2.

no elements no dofs Ax Œ103 m Ay Œ103 m
69 348 0.4659 16.91
276 1246 0.4682 16.96
1104 4698 0.4686 16.97
4416 18226 0.4688 16.97
17664 71778 0.4690 16.97
ref. values 0.4690 16.97

of the cantilever. The displacements Ax and Ay of point A, which is located in

the center of the back edge of the cantilever, are provided as benchmark reference
values. Scenarios CSM1 and CSM2 are evaluated stationary, while scenario CSM3
is without damping and results in a periodic oscillation.
We used the commercial simulation tool COMSOL Multiphysics [1] to com-
pute the CSM scenarios. The results obtained for CSM1 and CSM2 are listed in
Tables 4 and 5. They are computed with quadratic order finite elements and uni-
formly decreasing mesh size. CSM3 is computed with a backward-Euler method.
The results are presented in form of a mean value, an oscillation amplitude, and the
oscillation frequency, which are computed from the last oscillation period. Table 6
and Fig. 14 show the results of CSM3.
278 H.-J. Bungartz et al.

Table 6 Results and reference values of benchmark CSM3.

dt no elements no dofs Ax;mean ˙ Ax;ampl fx ŒH z Ay;mean ˙ Ay;ampl fy ŒH z
Œ103 m Œ103 m
1e-3 183 840 14.141 ˙ 14.139 1.0965 64.359 ˙ 63.441 1.0965
1e-3 723 3142 14.161 ˙ 14.159 1.0941 64.374 ˙ 63.626 1.0929
1e-3 2928 12138 14.180 ˙ 14.180 1.0953 64.351 ˙ 63.649 1.0953
1e-3 11712 47698 14.224 ˙ 14.224 1.0941 64.388 ˙ 63.812 1.0941
1e-3 46848 189090 14.155 ˙ 14.155 1.0952 64.274 ˙ 63.626 1.0941
5e-4 46848 189090 14.375 ˙ 14.375 1.0941 64.503 ˙ 64.297 1.0941
ref. values 14.305 ˙ 14.305 1.0995 63.607 ˙ 65.160 1.0995

0.005 0
-0.02

displacement y of A [m]
0
displacement x of A [m]

-0.005 -0.04

-0.01 -0.06

-0.015 -0.08

-0.02 -0.1

-0.025 -0.12

-0.03 -0.14
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
time [sec] time [sec]
(a) (b)

Fig. 14 Oscillating x- (a) and y-displacement (b) of point A, located at the back end of the
cantilever, for benchmark scenario CSM3.

4.3 Data Mapping

In Sect. 2.4, we have described the spacetree algorithms used to answer geometri-
cal queries occuring during data mapping between non-matching solver grids and
the coupling mesh in preCICE. To analyse the performance of these algorithms, a
simple 2D setup with a quadtree has been considered. We measured computational
runtimes for finding nearest projection points onto a coupling mesh polygon. The
coupling mesh approximates a circle with radius r D 1:0, the query points that are to
be projected on the coupling mesh are located equispaced on a concentric circle with
radius rquery D r C d , with d > 0. A quadtree with sidelengths l D 2:0 is covering
both the circle geometry and the query points. To study the influence of different
parameter settings, we varied the maximal cell size hquad at the coupling surface,
the distance of the query points to the coupling mesh d , and the discretisation width
of the circle hgeo . We first studied the influence of different refinement levels hquad
on the runtime for different geometry discretisation widths hgeo and corresponding
projection distances d . The choice of hgeo D d mimics a convergence study where
both, the solver grid and the coupling mesh, are refined repeatedly. The results in
Fig. 15(a) illustrate the different effects occuring in spacetree accelerated projection
Partitioned Simulation of Fluid-Structure Interaction on Cartesian Grids 279

70 1000
d = 7.813e-4 no tree
d = 3.906e-4 quadtree
60 d = 1.953e-4
d = 9.766e-5
d = 4.883e-5 100
50
runtime t [sec]

runtime t [sec]
40
10
30

20
1

10

0 0.1
0.0001 0.001 0.01 0.1 1 1e-05 0.0001 0.001 0.01 0.1
spacetree h sphere h

(a) (b)

Fig. 15 (a) Runtimes t over different quadtree refinements hquad . The different curves correspond
to different values of d , which is set to equal hgeo . Optimal runtimes are achieved for hquad 
100 hgeo . (b) Comparison of the runtime of a quadtree with optimally chosen hquad to a projection
point computation without using a spacetree.

point search. The optimal choice of hquad is limited from below by the special case
described in Sect. 2.4, leading to repeated projection point searches. However, also
if hquad is chosen to large, the coupling mesh elements are not subpartitioned fine
enough. For our test scenario hquad  100hgeo yields an optimal acceleration effi-
ciency. Figure 15(b) compares the runtimes of a quadtree with optimal hquad to that
of using no spacetree.

4.4 Coupled Simulations

Corresponding to the three fluid benchmarks CFD1–CFD3, three benchmark cases

for the full fluid-structure interaction scenario are proposed in [30]. The first sce-
nario, FSI1 results in a stationary deflection of the cantilever, FSI2 and FSI3 in
periodic oscillations. Accurate simulations for these scenarios are currently work in
progress using the flow solver in Peano (cf. Sect. 3), the coupling tool preCICE (cf.
Sect. 2), and the commercial structural solver in COMSOL [1]. Figure 16 shows
snapshots of a simulation of the FSI2 benchmark at Re D 100.
In addition, various other scenarios such as particle transport in microscopic
devices and deflection of a ’tower’ have been simulated (see Fig. 17 and [13, 38]
for some examples).
280 H.-J. Bungartz et al.

Fig. 16 (a) Snapshot of FSI2 benchmark Peano fluidsolver grid with velocity magnitudes.
(b) Snapshot of FSI2 benchmark preCICE coupling mesh and channel geometry used to set up
the Peano fluidsolver grid. Velocity vectors show the movement of the cantilever.

Fig. 17 (a) 3D coupled simulation with F3F and AdhoC, coupled by FSI❄ce. (b) Simulation of a
drift ratchet with Peano and Comsol, coupled by preCICE.

5 Summary and Outlook

The results shown in this paper demonstrate the suitability of Cartesian grids for par-
titioned fluid-structure interaction simulations both in terms of achievable accuracy
and in particular in terms of hardware performance of the respective codes. We pro-
posed a Cartesian flow solver implemented in our PDE framework Peano with very
low memory requirements and high cache hit-rates as a result of the structuredness
of the chosen type of adaptive Cartesian grids. In spite of this hardware efficiency,
the Peano CFD solver involves state-of-the-art numerical features such as highly
local grid adaptivity, multigrid solvers, and accurate and efficient force computa-
tions via the method of consistent forces. Even large geometry or topology changes
that frequently occur in fluid-structure interaction scenrios can be handled very nat-
urally and cheaply based on a fixed grid (Eulerian) approach. The parallelisation of
the solver is work in progress with some first promising results for simple exam-
ples. The same holds for dynamical, criterion-based grid adaptivity. The second
tool described is the coupling tool preCICE that currently offers standard coupling
functionality concerning data mapping, code-to-code communication, and coupling
strategies. In contrast to other available coupling tools, preCICE combines map-
ping, communication, and coupling numerics in a central component. With this and
Partitioned Simulation of Fluid-Structure Interaction on Cartesian Grids 281

the central coupling surface description, preCICE fully exploits the flexibility of the
partitioned simulation approach that is allows for an easy and independent exchange
of flow and structure solvers or coupling strategies.
In the near future, the basic functionality of both codes will be enhanced to
further improve the efficiency in various aspects. The Cartesian geometry repre-
sentation in Peano shall be improved by at least second order accurate methods
such as cut-cell or immersed boundary methods. At the same time, completing the
work on dynamical grid adaptivity and parallelisation with dynamical load bal-
ancing will allow for higher (and more focussed) grid resolutions. The solution
of the involved systems of equations will be accelerated by improved matrix-free
geometric multigrid solvers.
In the coupling tool preCICE, the most important step will be the implementation
of further, non-standard coupling strategies such as reduced order methods, non-
staggered coupling iterations, or multigrid methods. These methods strongly profit
from or even are possible only by the component-based approach with the coupling
tool as a separate numerical and not a mere data transfer component. Of course, also
other, more specific mapping methods will be included as needed.
In the end, we will be able to provide both a flow solver fully exploiting the
numerical and hardware-related advantages of Cartesian grids as well as a cou-
pling tool offering maximum flexibility in terms of the choice of flow and structural
solvers as well as in terms of (also non-standard highly sophisticated) coupling
strategies.

Acknowledgements This work has partially been funded by the German Research Founda-
tion (DFG) within the framework of the Research Group FOR493 – Fluid-Structure Interaction:
Modelling, Simulation, and Optimisation. This funding is thankfully acknowledged.

References

1. COMSOL Multiphysics GmbH. COMSOL: Multiphysics Modeling and Simulation. http://

www.comsol.de/.
2. R. Ahrem, P. Post, B. Steckel, and K. Wolf. MpCCI: A tool for coupling CFD with other
disciplines. In Proceedings of the 5th World Conference in Applied Fluid Dynamics, CFD -
Efficiency and the Economic Benefit in Manufacturing, 2001.
3. T. Aoki. Interpolated Differential Operator (IDO) scheme for solving partial differential
equations. Comput. Phys. Comm., 102:132–146, 1997.
4. M. Bader and Ch. Zenger. Cache oblivious matrix multiplication using an element ordering
based on a Peano curve. Linear Algebra and its Applications, 417(2–3):301–313, 2006.
5. P. Bastian, M. Blatt, A. Dedner, C. Engwer, R. Klöfkorn, M. Ohlberger, and O. Sander.
A Generic Grid Interface for Parallel and Adaptive Scientific Computing. Part I: Abstract
Framework. Computing, 82(2–3):103–119, 2008.
6. P. Bastian, M. Blatt, A. Dedner, C. Engwer, R. Klöfkorn, M. Ohlberger, and O. Sander. A
Generic Grid Interface for Parallel and Adaptive Scientific Computing. Part II: Implementation
and Tests in DUNE. Computing, 82(2–3):121–138, 2008.
7. C. Blanke. Kontinuitätserhaltende Finite-Element-Diskretisierung der Navier-Stokes-
Gleichungen. Diploma Thesis, Technische Universität München, 2004.
282 H.-J. Bungartz et al.

8. E. K. Blum. A modification of the Runge-Kutta fourth-order method. Mathematics of

Computation, 1962.
9. M. E. Brachet, P. D. Mininni, D. L. Rosenberg, and A. Pouquet. High-order low-storage
explicit Runge-Kutta schemes for equations with quadratic nonlinearities. Technical Report
arXiv:0808.1883, CERN, Aug 2008.
10. D. Braess. Finite Elements—Theory, Fast Solvers, and Applications in Solid Mechanics.
Cambridge University Press, 3rd edition, 2007.
11. V. Brázdová and D.R. Bowler. Automatic data distribution and load balancing with space-
filling curves: implementation in conquest. Journmal of Physics: Condensed Matter, 20, 2008.
12. M. Brenk, H.-J. Bungartz, M. Mehl, R.-P. Mundani, A. Düster, and D. Scholz. Efficient inter-
face treatment for fluid-structure interaction on cartesian grids. In ECCOMAS COUPLED
PROBLEMS 2005, Proc. of the Thematic Conf. on Computational Methods for Coupled Prob-
lems in Science and Engineering. International Center for Numerical Methods in Engineering
(CIMNE), 2005.
13. M. Brenk, H.-J. Bungartz, M. Mehl, I. L. Muntean, T. Neckel, and T. Weinzierl. Numeri-
cal simulation of particle transport in a drift ratchet. SIAM Journal of Scientific Computing,
30(6):2777–2798, 2008.
14. M. Brenk, H.-J. Bungartz, M. Mehl, and T. Neckel. Fluid-structure interaction on cartesian
grids: Flow simulation and coupling environment. In H.-J. Bungartz and M. Schäfer, edi-
tors, Fluid-Structure Interaction, number 53 in Lecture Notes in Computational Science and
Engineering, pages 233–269. Springer, 2006.
15. M. Brenk, H.-J. Bungartz, and T. Neckel. Cartesian discretisations for fluid-structure interac-
tion – consistent forces. In P. Wesseling, E. Oñate, and J. Périaux, editors, ECCOMAS CFD
2006, European Conference on Computational Fluid Dynamics. TU Delft, 2006.
16. H.-J. Bungartz, W. Eckhardt, M. Mehl, and T. Weinzierl. Dastgen - A Data Structure Generator
for Parallel C++ HPC Software. In M. Buback, G. D. van Albada, P. M. A. Sloot, and J. J.
Dongarra, editors, ICCS 2008 Proceedings, Lecture Notes in Computer Science, Heidelberg,
Berlin, 2008. Springer-Verlag.
17. H.-J. Bungartz, W. Eckhardt, T. Weinzierl, and C. Zenger. A precompiler to reduce the mem-
ory footprint of multiscale PDE solvers in C++. Future Generation Computer Systems, 2009.
(accepted).
18. H.-J. Bungartz, M. Mehl, T. Neckel, and T. Weinzierl. The pde framework peano applied to
computational fluid dynamics. Computational Mechanics, 2009. accepted.
19. H.-J. Bungartz, M. Mehl, and T. Weinzierl. A parallel adaptive Cartesian PDE solver using
space–filling curves. In E. W. Nagel, V. W. Walter, and W. Lehner, editors, Euro-Par 2006,
Parallel Processing, 12th International Euro-Par Conference, volume 4128 of Lecture Notes
in Computer Science, pages 1064–1074, Berlin Heidelberg, 2006. Springer-Verlag.
20. A. J. Chorin. Numerical solution of the Navier-Stokes equations. Math. Comp., 22:745–762,
1968.
21. W. A. Wall D. P. Mok and E. Ramm. Accelerated iterative substructuring schemes for instation-
ary fluid-structure interaction. Computational Fluid and Solid Mechanics, pages 1325–1328,
2001.
22. A. de Boer, A. H. van Zuijlen, and H. Bijl. Review of coupling methods for non-matching
meshes. Comput. Methods Appl. Mech. Engrg., 196(8):1515–1525, 2007.
23. P. Deuflhard and F. Bornemann. Numerische Mathematik II. Integration gewöhnlicher Differ-
entialgleichungen. de Gruyter, 1994.
24. Ch. Förster, W. A. Wall, and E. Ramm. Artifcial added mass instabilities in sequential staggered
coupling of nonlinear structures and incompressible viscous flows. Comput. Methods Appl.
Mech. Engrg., 196(7):1278–1293, 2007.
25. P. M. Gresho and R. L. Sani. Incompressible Flow and the Finite Element Method. John Wiley
& Sons, 1998.
26. M. Griebel and G. Zumbusch. Parallel multigrid in an adaptive PDE solver based on hashing
and space-filling curves. Parallel Computing, 25(7):827–843, 1999.
Partitioned Simulation of Fluid-Structure Interaction on Cartesian Grids 283

27. F. Günther, A. Krahnke, M. Langlotz, M. Mehl, M. Pögl, and Ch. Zenger. On the parallelization
of a cache-optimal iterative solver for pdes based on hierarchical data structures and space-
filling curves. In Recent Advances in Parallel Virtual Machine and Message Passing Interface:
11th European PVM/MPI Users Group Meeting Budapest, Hungary, September 19 - 22, 2004.
Proceedings, volume 3241 of Lecture Notes in Computer Science, Heidelberg, 2004. Springer-
Verlag.
28. E. Hairer, S. P. Nørset, and G. Wanner. Solving Ordinary Differential Equations I.Nonstiff Prob-
lems., volume 8 of Springer Series in Comput. Mathematics. Springer-Verlag, second edition,
1993.
29. E. Hairer, S. P. Nørset, and G. Wanner. Solving Ordinary Differential Equations II.Stiff and
Differential-Algebraic Problems., volume 14 of Springer Series in Comput. Mathematics.
Springer-Verlag, second edition, 1996.
30. J. Hron and S. Turek. Proposal for numerical benchmarking of fluid-structure interaction
between elastic object and laminar incompressible flow. In H.-J. Bungartz and M. Schäfer,
editors, Fluid-Structure Interaction, number 53 in Lecture Notes in Computational Science
and Engineering, pages 371–385. Springer-Verlag, 2006.
31. Y. Imai and T. Aoki. Accuracy study of the IDO scheme by Fourier analysis. Journal of
Computational Physics, 217:453–472, 2006.
32. Y. Imai and T. Aoki. A higher-order implicit IDO scheme and its CFD application to local
mesh refinement method. Comput. Mech., 38:211–221, 2006.
33. Y. Imai and T. Aoki. Stable coupling between vector and scalar variables for the IDO scheme
on collocated grids. Journal of Computational Physics, 215:81–97, 2006.
34. Y. Imai, T. Aoki, and K. Takizawa. Conservative form of interpolated differential opera-
tor scheme for compressible and incompressible fluid dynamics. Journal of Computational
Physics, 227:2263–2285, 2008.
35. B. M. Irons and R. C. Tuck. A version of the aitken accelerator for computer iteration.
International Journal of Numerical Methods in Engineering, 1:275–277, 1969.
36. J. Larson, R. Jacob, and E. Ong. The model coupling toolkit: A new fortran90 toolkit for
building multiphysics parallel coupled models. Int. J. High Perf. Comp. App., 19(3):277–292,
2005.
37. K. Long. Sundance: A rapid prototyping toolkit for parallel pde simulation and optimiza-
tion. In M. Heinkenschloss L. T. Biegler, O. Ghattas and B. van Bloemen Waanders, editors,
Large-Scale PDE-Constrained Optimization, Lecture Notes in Computational Science and
Engineering, volume 30 of Lecture Notes in Computational Science and Engineering, pages
331–339. Springer-Verlag, 2009.
38. M. Mehl, M. Brenk, I.L. Muntean, T. Neckel, and T. Weinzierl. Benefits of structured cartesian
gris for the simulation of fluid-structure interactions. In Proceedings of the Third Asian-Pacific
Congress on Computational Mechanics, Kyoto, Japan, 2007.
39. M. Mehl, T. Weinzierl, and C. Zenger. A cache-oblivious self-adaptive full multigrid method.
Numerical Linear Algebra with Applications, 13(2-3):275–291, 2006.
40. D. Meidner, R. Rannacher, and J. Vihharev. Goal-oriented error control of the iterative solution
of finite element equations. Journal of Numerical Mathematics, 17:143–172, 2009.
41. D. Meidner and B. Vexler. Adaptive space-time finite element methods for parabolic optimiza-
tion problems. SIAM Journal on Control and Optimization, 46(1):116–142, 2007.
42. Message Passing Interface Forum. MPI: A Message-Passing Interface Standard, Version 1.1.
http://www.mpi-forum.org/docs/docs.html.
43. W. F. Mitchell. A refinement-tree based partitioning method for dynamic load balancing with
adaptively refined grids. Journal of Parallel and Distributed Computing, 67(4):417–429, 2007.
44. T. Neckel. The PDE Framework Peano: An Environment for Efficient Flow Simulations. Verlag
Dr. Hut, 2009.
45. R. Niekamp. Component template library (ctl). http://www.wire.tu-bs.de/forschung/projekte/
ctl/e ctl.html.
46. J. F. Gerbeau P. Caussin and f. Nobile. Added-mass effect in the design of partitioned algo-
rithms for fluid-structure problems. Computer Methods in Applied Mechanics and Engineering,
194:4506–4527, 2005.
284 H.-J. Bungartz et al.

47. St. G. Parker and J. Guilkey und T. Harman. A component-based parallel infrastructure for
the simulation of fluid-structure interaction. Engineering with Computers, 22(3–4):277–292,
2006.
48. A. Patra and J.T. Oden. Problem decomposition for adaptive hp finite element methods.
Computing Systems in Engineering, 6(2):976–109, 1995.
49. S. Roberts, S. Klyanasundaram, M. Cardew-Hall, and W. Clarke. A key based parallel adap-
tive refinement technique for finite element methods. In Proc. Computational Techniques and
Applications: CTAC ’97, pages 577–584, Singapore, 1998.
50. H. Sagan. Space-filling curves. Springer-Verlag, New York, 1994.
51. R. S. Sampath, S. S. Adavani, H. Sundar, I. Lashuk, and G. Biros. Dendro: parallel algorithms
for multigrid and amr methods on 2:1 balanced octrees. In SC ’08: Proceedings of the 2008
ACM/IEEE conference on Supercomputing, pages 1–12, Piscataway, NJ, USA, 2008. IEEE
Press.
52. J. Stoer and R. Bulirsch. Numerische Mathematik 2. Springer-Verlag, 4th edition, 2000.
53. H. Sundar, R.S. Sampath, and G. Biros. Bottom-up construction and 2:1 balance refinement of
linear octrees in parallel. SIAM J. Sci. Comput., 30(5):2675-2708, 2008.
54. T. Tu, D. R. O’Hallaron, and O. Ghattas. Scalable parallel octree meshing for terascale applica-
tions. In SC ’05: Proceedings of the 2005 ACM/IEEE conference on Supercomputing, page 4,
Washington, DC, USA, 2005. IEEE Computer Society.
55. S. Turek and M. Schäfer. Benchmark computations of laminar flow around a cylinder. In E. H.
Hirschel, editor, Flow Simulation with High-Performance Computers II, number 52 in NNFM.
Vieweg, 1996.
56. K. Unterweger. CFD simulations of moving geometries using Cartesian grids. Diploma Thesis,
Institut für Informatik, Technische Universität München, 2009.
57. J. Vierendeels. Implicit coupling of partitioned fluid-structure interaction solvers using
reduced-order models. In H.-J. Bungartz and M. Schäfer, editors, Fluid-Structure Interac-
tion, number 53 in Lecture Notes in Computational Science and Engineering, pages 1–18.
Springer-Verlag, 2006.
58. P. Gamnitzer W.A. Wall and A. Gerstenberger. Fluid-structure interaction approaches on
fixed grids based on two different domain decomposition ideas. International Journal of
Computational Fluid Dynamics, 22(6):411–427, 2008.
59. T. Weinzierl. A Framework for Parallel PDE Solvers on Multiscale Adaptive Cartesian Grids.
Verlag Dr. Hut, 2009.
60. G. W. Zumbusch. On the quality of space-filling curve induced partitions. Z. Angew. Math.
Mech., 81:25–28, 2001.
61. G.W. Zumbusch. Parallel Multilevel Methods. Advances in Numerical Mathematics. Teubner
Verlag, 2003.
An Explicit Model for Three-Dimensional
Fluid-Structure Interaction
using LBM and p-FEM

S. Geller, S. Kollmannsberger, M. El Bettah, M. Krafczyk, D. Scholz,

A. Düster, and E. Rank

Abstract An explicit coupling model for the simulation of surface coupled fluid-
structure interactions with large structural deflections is introduced. Specifically,
the fluid modeled via the Lattice Boltzmann Method (LBM) is coupled to a high-
order Finite Element discretization of the structure. The forces and velocities are
discretely computed, exchanged and applied at the interface. The low compressibil-
ity of the Lattice Boltzmann Method allows for an explicit coupling algorithm. The
proposed explicit coupling model turnes out to be accurate, very efficient and stable
even for nearly incompressible flows. It was implemented in three software compo-
nents: V IRTUAL F LUIDS (fluid), A DHO C (structure) and FSI❄ce (a communication
library). The validity of the approach is demonstrated in two dimensions by means
of comparing numerical results to measurements of an experiment. This experiment
involves a flag-like structure submerged in the laminar flow field of an incompress-
ible fluid where the structure exhibits large, geometrically non-linear, self excited,
periodic motions. The methodology is then extended to three dimensions. Its per-
formance is first demonstrated via the computation of a falling sphere in a pipe.
The close correspondence of the results obtained by application of the numerical
scheme compared to a semi-analytic solution is demonstrated. The proposed explicit
coupling model is then extended to a plate in a cross flow. We verify the results

S. Geller and M. Krafczyk

Institute for computational modeling in civil engineering, TU Braunschweig, Muehlenpfordtstraße
4-5, 38106 Braunschweig, Germany
e-mail: [email protected]
S. Kollmannsberger, D. Scholz, and E. Rank
Computation in Engineering, TU München, Arcisstraße 21, 80333 München, Germany
e-mail: [email protected]
A. Düster
Numerische Strukturanalyse mit Anwendungen in der Schiffstechnik (M-10),
Technische Universität Hamburg-Harburg, Schwarzenbergstraße 95 C, D-21073 Hamburg
e-mail: [email protected]
M.E. Bettah
Department of Ocean Engineering, University of Rhode Island, Narragansett, RI 02882, USA
e-mail: [email protected]

H.-J. Bungartz et al. (eds.), Fluid Structure Interaction II, Lecture Notes 285
in Computational Science and Engineering 73, DOI 10.1007/978-3-642-14206-2 11,
c Springer-Verlag Berlin Heidelberg 2010
286 S. Geller et al.

by comparing them to results obtained by application of the commercial ALE-

Finite Volume—h-FEM Fluid-Structure interaction solver A NSYS M ULTIPHYSICS.
Additional examples demonstrate the applicability of the proposed methodology to
problems of (arbitrarily) large deformations and of large scale.

1 Introduction

Examples for Fluid-Structure interaction (FSI) problems are found in various dis-
ciplines. In civil engineering bridges, sky scrapers, membrane structures, off shore
platforms and many other constructions are exposed to loads that require FSI inves-
tigations. In mechanical engineering, valves, pumps, airfoils, ship propellers and
many other devices have to be dimensioned considering FSI. The design of stir-
rers, extruders and injection systems in process engineering or artificial heart valves
and blood vessels in medicine require the consideration of the bidirectional interac-
tion between fluid and structure as well. The realistic transient simulation of such
processes is rather challenging with respect to both, the numerical techniques and
the required computer capacities. In addition to the effort resulting from the com-
bination of structural and fluid mechanical computations, the interaction brings in
a variety of complex numerical and mechanical additional phenomena which need
special attention as well. This comprises questions of the adequate consideration of
the dynamic and nonlinear interaction itself as well as the capabilities of partitioned
solution approaches. Additionally, the suitability of methods which have established
themselves in the individual disciplines may now suffer drawbacks when applied
in the interaction. The robustness and efficiency of the methods, and the adequate
description of complex changing geometries must be addressed as well.
In general, two approaches to solve fluid-structure interaction (FSI) problems
exist: The monolithic approach [38,39,82] discretizes the two separate domains with
a similar discretization scheme and solves the resulting, coupled system of equations
within one solver. The compatibility conditions at the interface are treated inherently
within this system of equations. By contrast, the partitioned approach [53] uses sep-
arate solvers for the fluid and the structural system. Strong coupling methods [48,83]
as well as loose coupling methods [19, 20, 54, 61, 62, 62, 63, 63] exist. In the parti-
tioned solution, the solvers need to communicate physical properties of their mutual
boundary to fulfill the interface conditions. Each domain may utilize any type of
discretization considered efficient for its field. For the fluid, mainly two approaches
are used in the context of partitioned FSI for large deformations. The fluid is usually
described either on an arbitrarily moving grid (arbitrary Lagrangian Eulerian (ALE)
formulation) or on a fixed Cartesian grid. At first sight the primary advantage of the
ALE method is its capability to handle boundary layers more naturally, though this
is not true in all cases. By contrast, large displacements impose virtually no extra
effort when coupling to fluid solvers based on fixed Cartesian grids but do cause
difficulties for methods based on an ALE approach.
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 287

The first LB algorithm for an interaction problem between a fluid and rigid
obstacles has been developed by Ladd [49, 50] for the simulation of particulate sus-
pensions. A special treatment of boundary conditions and the activation/deactivation
of fluid nodes have been developed in this work. Krafczyk et al. [47] extended
these ideas to perform two-dimensional FSI of an artificial heart-valve with mov-
ing leaflets. The flow in the channel was driven by transient boundary conditions
representing a physiologically relevant regime. Previous contributions to the sub-
ject matter include [27, 71], where the LB fluid-solver V IRTUAL F LUIDS based
on unstructured quadtree type grids was coupled to the high order finite element
(p-FEM) structural dynamics solver A DHO C by an eigenvalue approach to pre-
dict bidirectional fluid-structure interaction for geometrically linear problems in two
dimensions.
The methodology applied can be classified as a discrete forcing approach of an
Embedded Domain Method. Here, the Fluid-Structure interface is represented as a
sharp interface and the boundary conditions are directly imposed on the structure
and on the fluid at this discrete interface. As such, this contribution is a continua-
tion of the work presented in [27, 45, 71]. In [45], this methodology was layed out
in detail, benchmarked and verified against the two-dimensional, numerical exam-
ples proposed in [80, 81]. By contrast, the present paper shows that the method can
reproduce the experimental results described in [32] where a detailed comparison of
the results obtained in comparison to the application of other methods is presented
as well.
Additionally, the methods applied in two dimensions are now extended to three
dimensions and verified against a semi-analytic solution for a falling sphere in a
cylinder and the three-dimensional benchmark suggested in [1, 2].
An immersed boundary method lattice Boltzmann approach was presented in [21,
34, 35] for solving fluid-particles interaction problems. The method was extended
by a multi-block strategy allowing for mesh refinement [76]. Other approaches for
computing Fluid-Structure interactions with the Lattice Boltzmann Method have
been developed recently as well. In [73] a distributed-Lagrange-multiplier/fictitious-
domain method (DLM/FD) is introduced to deal with the fluid/elastic–solid interac-
tions. In [72, 74] this approach is extended to three dimensions. The fluid motion is
solved by LBM and the deformation of the solid body is solved by a low order finite
element method.
The contribution at hand is structured as follows: Sect. 2 recapitulates the Lat-
tice Boltzmann Method followed by a description of the necessary transformation
of the physical quantities into the lattice Boltzmann system in Sect. 3. A brief
description of the p-Version of the Finite Element Method applied to structural
dynamics is given in Sect. 4. The coupling approach itself is described in Sect. 5
and the boundary conditions are laid out in Sect. 6. Section 7 presents the validation
of the methodology via comparison with the experimental setup given in [32]. It
consists of a rigid cylinder in the cross flow of a Newtonian, incompressible fluid
with a flexible tail attached to its downstream side. The method is then extended
in Sect. 8 to three dimensions and applied to an example of a rigid body motion,
namely a falling sphere in a cylinder. Section 9 then applies the methodology to
288 S. Geller et al.

a three-dimensional plate in a cross flow which exhibits geometrically non-linear

deformations induced by the fluid. Section 10 then demonstrates the applicability of
the developed methods to problems of very large deflections in FSI. The full scale
computation of a bridge in Fluid-Structure interaction is also laid out in Sect. 10 to
discuss an application to problems of large scale before conclusions are drawn in
Sect. 11.

2 The Lattice Boltzmann Method for single-phase

flow on non-uniform grids

A smoothed tree-type grid (quadtrees and octrees) is used. Typical quadtree meshes
are shown in Fig. 1. In the interest of algorithmic simplicity, we employ smoothed
trees i.e. the neighboring cells can only differ by at most one grid level. In Fig. 1,
a non-smoothed quadtree, a smoothed quadtree and a smoothed quadtree with a
minimum width of 3 cells per grid level are shown. Regions around arbitrarily
shaped objects can be refined locally. In essence, the LBM is a discretization of
the Boltzmann equation in space and time on these grids. The basic, physical notion
of the method is to let fluid-mass fractions1 collide on the nodes formed by this
lattice and propagate them only among the links formed between the nodes of this
lattice. To introduce the method formally, the following abbreviations are used: x
represents a 3D vector in space and f is a b-dimensional vector, where b is the
number of microscopic velocities along the links formed by the grid. The d 3q19
model [64] with the following microscopic velocities is considered:
8 9
< 0 c c 0 0 0 0 c c c c c c c c 0 0 0 0 >
ˆ =
fei ; i D 0; : : : ; 18g D 0 0 0 c c 0 0 c c c c 0 0 0 0 c c c c ;
:̂ 0 0 0 0 0 c c 0 0 0 0 c c c c c c c c >
;

Fig. 1 Quadtree, smoothed quadtree and quadtree with minimum width of 3 cells.

1
i.e. particles
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 289

where c D x l
tl
is a constant velocity determining the speed of sound cs and cs2 D
2
c =3. The lattice Boltzmann equation using the multi-relaxation time model [10,52]
for the grid level l is then given by

fi;l .t C tl ; x C ei tl / D fi;l .t; x/ C ˝i;l ; i D 0; : : : ; b  1; (1)

where tl is the time step and ˝i;l the collision operator. The collision operator on
grid level l is given by:

˝ l D M1 Sl Œ.Mf /  meq  : (2)

M represents the transformation matrix, which is composed of 19 orthogonal basis

vectors f˚ i ; i D 0; : : : ; b  1g. Its definition can be found in the appendix. They
are orthogonal with respect to the inner product h˚ i ; ˚j wi (in contrast to [11],
where h˚ i ; ˚j i D 0; if i ¤ j ). The vector w constitutes the weights fwi ; i D
0 : : : ; b  1g:

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
w D . ; ; ; ; ; ; ; ; ; ; ; ; ; ; ; ; ; ; /:
3 18 18 18 18 18 18 36 36 36 36 36 36 36 36 36 36 36 36

The moments m D Mf are labeled as

m D .; e; ; jx ; qx ; jy ; qy ; jz ; qz ; 3pxx ; 3xx ; pww ; ww ; pxy ; pyz ; pxz ; mx ; my ; mz /:

meq is the vector symbolizing the equilibrium moments given in Eqs. (4) and Sl D
fsl;i;i ; i D 0; : : : ; b  1g is the diagonal collision matrix. The nonzero collision
parameters sl;i;i (the eigenvalues of the collision matrix M1 Sl M) are:

sl;1;1 D sl;a ;
sl;2;2 D sl;b ;
sl;4;4 D sl;6;6 D sl;8;8 D sl;c ;
sl;10;10 D sl;12;12 D sl;d ;
tl
sl;9;9 D sl;11;11 D sl;13;13 D sl;14;14 D sl;15;15 D  D sl;! ;
l
sl;16;16 D sl;17;17 D sl;18;18 D sl;e :

The relaxation time l is defined as:

 1
l D 3 2
C tl ; (3)
c 2
where  is the kinematic viscosity. The parameters sa ; sb ; sc ; sd and se can be cho-
sen in the range Œ2; 0 and tuned to improve stability [52]. The optimal values
depend on the specific characteristics of the system (considering geometry, initial
and boundary conditions) and can not be computed in advance. Some reasonable
290 S. Geller et al.

values for these parameters are given in [11]. We choose sa D sb D sc D

sd D se D maxfsl;! ; 1:0g. The nonzero equilibrium distribution functions
eq
fmi ; i D 0; : : : ; 18g are given by:

meq
0 D ; meq
3 D 0 ux ; meq
5 D 0 uy ; meq
7 D 0 uz ; (4a)
meq
1 De eq
D 0 .u2x C u2y C u2z /; (4b)
eq eq
m9 D 3pxx D 0 .2 u2x  u2y  u2z /; (4c)
meq
11 D pzz
eq
D 0 .u2y  u2z /; (4d)
meq
13 D eq
pxy D 0 ux uy ; (4e)
meq
14 D pyz D 0 uy uz ;
eq
(4f)
meq
15 D eq
pxz D 0 ux uz ; (4g)

where 0 is a constant density and  a density variation.

The macroscopic quantities (density and momentum) and the stress tensor are
given by (omitting the index l):
X
D fi ; (5a)
i
X
0 u D ei fi ; (5b)
i
t 1
Sxx D .1  /. e C pxx  0 u2x /; (5c)
2 3
t 1 1 1
Syy D .1  /. e  pxx C pww  0 u2y /; (5d)
2 3 2 2
t 1 1 1
Szz D .1  /. e  pxx  pww  0 u2z /; (5e)
2 3 2 2
t
Sxy D .1  /.pxy  0 ux uy /; (5f)
2
t
Syz D .1  /.pyz  0 uy uz /; (5g)
2
t
Sxz D .1  /.pxz  0 ux uz /: (5h)
2
Using the Chapman-Enskog expansion [6, 42], which in essence is a gradient
expansion around the local equilibrium state, it can be shown that the LB Method:
 is a second order scheme both in space and time for the compressible Navier-
Stokes equations in the low Mach number limit using the advective scaling
 is a scheme of first order in time and second order in space for the incompressible
Navier-Stokes equations using the diffusive scaling
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 291

f = fine grid

gridinterface

Xf
c = coarse grid

Xc

Fig. 2 Grid-interface and scale directions.

In order to model spatial refinement where one level of refinement is represented

by one grid level, nested time-stepping is used in this paper. If the speed of sound,
the kinematic viscosity, the Mach and the Reynolds number are assumed to be equal
on all grids, then two time steps on the finer grid have to be performed during one
time step on the coarser level [22]. An approach where the Mach number is lowered
on the finer grid levels to ensure a faster convergence to the incompressible Navier-
Stokes equations can be found in [66]. This approach requires to perform four time
steps on the finer grid during one time step on the coarser mesh.
A typical grid interface between two grid levels is shown in Fig. 2. The overlap
of the interface is due to the fact that missing distributions on one grid level have
to be computed from the adjacent grid level. Ensuring the continuity of pressure,
neq
the velocity, and also of their derivatives, the non-equilibrium parts (f i [9, 22,
neq
85] or alternatively mi ) have to be rescaled. Performing the rescaling after the
propagation step one obtains:

sl;i;i tl1 neq

mneq
i;l1
D mi;l : (6)
sl1;i;i tl

The relaxation parameters sl;a D sl;b D sl;c D sl;d D sl;e can in principle be
arbitrarily chosen in the range Œ2; 0, but we used the same scaling as for sl;! .

sl;a sl;b sl;c sl;d sl;e sl;!

D D D D D : (7)
slC1;a slC1;b slC1;c slC1;d slC1;e slC1;!

Different space and time interpolations have to be used because of the different
mesh spacings xl and time steps tl . A cubic interpolation in space is used for
the ’hanging’ nodes. Fig. 3 shows the interpolation for the ’hanging’ node P using
the nodes P1, P2, P3 and P4. Linear interpolation is used in time. Details of the
algorithm can be found in [8, 85]. The necessary boundary conditions are described
in Sect. 6.
292 S. Geller et al.

P1 P2 P P3 P4

Fig. 3 Hanging node P and nodes P1-P4 required for interpolation.

3 Transformation of physical quantities

A reasonable set of dimensionless quantities which uniquely define an FSI problem

are given e.g. by
Es
 the Aerolastic number Ae D 2 ,
f U0f
U0f H
 the Reynolds number Re D f
,

 the density ratio ˇ D  s ,
f
 and the Poisson ratio of the elastic structure s .

Here Es is the Young’s modulus, s and f the density of structure and fluid
respectively, U0f a reference velocity, H a reference length and f the kinematic
viscosity.
The fluid solver rescales the system in such a way, that numerical errors are
reduced (avoiding very small and/or large numerical values). Since this rescaling
is inherent to most LB implementations, the procedure for rescaling two different
systems labeled as LB and real is presented below. The scaling of the forces and
the time from the fluid solver system (LB) to the real system is done by setting the
dimensionless drag coefficient equal in both systems. The result is:
2
Hreal  real  u2real
Freal D FLB  ; (8)
2
HLB  LB  u2LB

with reference height H , reference density  and reference velocity u. By mul-

tiplying the drag formula with the square of the Reynolds number we obtain the
following force scaling formula:

real  real
2
Freal D FLB  : (9)
LB  LB
2
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 293

This force transformation is necessary, if no reference velocity is given. By equat-

ing the Reynolds number in both systems and using u D T H
we obtain the following
time scaling:
LB  Hreal
2
Treal D TLB  : (10)
real  HLB
2

Computing the displacements as well as the velocity of the structure the wall
velocity (boundary condition) for the LB system is

LB  Hreal
uLB D ureal  : (11)
real  HLB

Further dimensionless quantities are the weight coefficients:

 D g H
3
2 ,
 ı D g u2H ,
The formula for rescaling the weights due to both systems is:

u2LB  Hreal
gLB D greal  : (12)
u2real  HLB

If there is no reference velocity the scaling formula is:

2
LB  Hreal
3
gLB D greal  : (13)
2
real  HLB
3

4 Structural Solver

The 3D LB solver V IRTUAL F LUIDS was coupled with two structural models imple-
mented in PE [41] and A DHO C [12]. PE is a rigid multi-body physics engine. It is a
framework designed for both virtual reality scenarios with real time requirements as
well as large-scale computer simulations with millions of interacting rigid bodies.
In contrast to the physics engines for computer games, PE primarily aims at real-
istic simulation environments comprising rigid multi-body dynamics, as they arise
for example in material sciences or fluid mechanics. For instance, for some years,
rigid multi-body dynamics has been used in combination with lattice Boltzmann
flow simulations to simulate the behavior of rigid bodies in a flow [41]. This cou-
pled simulation system is used to simulate fluidization processes with a high number
of buoyant rigid bodies or sedimentation processes with large numbers of floating
bodies.
For modeling the interaction with fully three-dimensional thin- and thick walled
structures exhibiting transient deformations and deflections, a high-order Finite Ele-
ment Method (p-FEM) is employed. This has been implemented in A DHO C [12].
Our implementation of high order finite elements in three dimensions utilizes a
294 S. Geller et al.

hexahedral element formulation, using the hierarchic shape functions introduced

in [77].
These three-dimensional shape functions can be classified into four groups: the
nodal or vertex modes, the edge modes, the face modes, and the internal modes.
The nodal or vertex modes are defined by the standard trilinear shape functions,
well known from the isoparametric eight-noded brick element. The edge and face
modes are non-zero on the edges and faces which they are associated to and vanish
on all other edges and faces, whereas the internal modes are purely local being zero
on all faces and edges of the hexahedral element.
p ;p ;p
Three different types of trial spaces can be defined: the trunk space Sts   
p ;p  ;p 
.˝sth /, the tensor product space Sps .˝sth / and an anisotropic tensor prod-
p;p;q h
uct space S .˝st /. A detailed description of these trial spaces can be found
p ;p ;p
in [13, 14, 77, 78]. The polynomial degree for the trial spaces Sts    .˝sth / and
p ;p ;p
Sps   .˝sth / can be varied separately in each local direction. It is possible to
construct discretizations where the polynomial degree for the in-plane and thick-
ness direction of thin-walled structures can be treated differently. High-order solid
elements can therefore provide a fully three-dimensional solution also including
arbitrary three-dimensional stress states, and, nevertheless can cope with high aspect
ratios of thin-walled structures. Moreover, they are less prone to locking effects,
in contrast to classical low order elements. Thus, only one element type for thin-
as well as thick-walled structures is sufficient, without making transition elements
between thin-walled and massive parts of the structure necessary. A detailed inves-
tigation of the advantages of high order solid elements for thin-walled (nonlinear)
continua can be found in [13, 15–17, 36, 37, 65, 71, 77, 78].
Time integration is performed using classical second order accurate finite differ-
ence methods such as the generalized-˛ method [7] or the method of Newmark [59].
Arising nonlinearities are treated by a Newton-Raphson procedure.
For Fluid-Structure Interaction problems, one advantage of three-dimensional
models is especially obvious, i.e. the inherently correct representation of the wet
surface being in contact with the fluid. Unlike for dimensionally reduced beam,
plate and shell models, there is no necessity to reconstruct this skin surface from the
middle surface and certain kinematic assumptions. Curved geometries are described
by means of the blending function method. This method also works well for three-
dimensionally curved surfaces. A detailed description is given in [78]. Moreover, by
using an adaptive selection of the polynomial degree in the three directions, a very
cost-efficient discretization can be found while maintaining control of the approxi-
mation and modeling errors. See [17, 56] for elastostatic problems and [68–70] for
elastodynamic problems as further references.

5 Description of the Coupling approach

The solution of the two-field problem of Fluid-Structure interaction is achieved by a

partitioned approach solving each field separately. The coupling is performed by
exchange of information at the boundary. The coupling framework is described
below.
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 295

5.1 Previous work

In the course of this work we investigated several coupling algorithms which were
especially devised to cope with difficulties we faced at the beginning of the research
work. These are mentioned here for completeness. However, only the most efficient
one will be described in more detail in Sect. 5.2.
A first version used an eigenvalue approach for the structural integration in time
[27, 43, 71] which by design could only account for geometrically linear structural
deformations. The coupling algorithm was of explicit nature since the boundary
information was only exchanged once at each time step.
A second, completely different and more stable implicit version which could
also take geometrically nonlinear structural displacements into account is described
in [28]. This concept proved to be quite involved from a computational point of view
and was therefore not further investigated.
The most efficient algorithm turned out to be the version described next.

5.2 Coupling algorithm

Setup process

At the boundary, two completely different discretizations are present. The structural
boundary is described by coarse and curved p-elements while the fluid boundary is
described by relatively small cubes and a locally refined octree type grid. We there-
fore introduce an interface mesh to couple both discretizations. The interface mesh
is a moving surface mesh consisting of flat triangles. On each node the values for the
velocity, the load vector and other physical quantities required for the exchange are
stored. The mesh is constructed as follows. The nodes are defined by the Gaussian
integration points on the surface of the p-Elements or as a set of equidistant points
situated on the surface of the structure. A triangulation is then carried out to obtain
the interface mesh. The interface mesh can be adapted by the p-FEM solver as well
as by the LB solver. A Gaussian interface mesh and the coupling is shown in Fig. 4.
The framework (for details see [4]) is based on a client-server concept and real-
ized through MPI [23]. A master process manages the exchange of data between the
structural and fluid code.
The values to be transferred (i.e. the tractions and displacements) are only
allowed to vary linearly between the nodes of the interface mesh.
As part of the setup, the structural solver provides the supervisor process with the
interface mesh in its initial configuration. This mesh is handed to the fluid solver.

Core algorithm

The coupling coupling algorithm is depicted in Fig. 5. It incorporates nested time

stepping. While n fluid steps are performed corresponding to one structural step,
296 S. Geller et al.

Structure FEM−Mesh Interface Mesh Fluid LBM−Grid

displacements displacements

forces forces

Fig. 4 Coupling using interface mesh.

Lattice Boltzmann p-FEM

fluid solver structure solver
tractions
fluid step = 0 structure step = 0

internal step = 1 interpolated

surface geometry

level
internal step = 2
Dt structure
nDt fluid

fluid step = 1

internal step = 1
interpolated
surface geometry

level
internal step = 2

new
surface geometry
fluid step = n structure step = 1
displacements

Fig. 5 Explicit coupling algorithm with multilevel nested time-stepping.

further sub-steps (called fluid-internal steps) need to be performed in each of these

fluid steps, in order to propagate the fluid unknowns fI .t; x/ through the hierar-
chical, non-uniform grid defined by the LBM solver (see Sect. 2). After the setup
process (Sect. 5.2), the algorithm follows these steps:
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 297

1. The fluid solver computes the traction vector on the interface mesh points.
2. The tractions are handed through the interface mesh to the structural solver.
3. The structure integrates these tractions in its load vector either the standard way
or with a load conservative scheme [45]. The structural solver calculates the
displacements which are then exchanged through the interface mesh.
4. The fluid solver performs an interpolation of the positions of the interface mesh
in time and calculates the solution.
5. Step 4 is repeated for all fluid-internal steps.
6. Step 5 is repeated for the number n of fluid-subcycling steps. The fluid stresses
of the subcycling steps plus the fluid internal steps are averaged.
Thus, the algorithm is a staggered coupling algorithm with a subiteration for the
fluid with the specialty that the fluid-solver needs to perform fluid-internal steps
on its subgrids. As such, the presented algorithm is only conditionally stable and
first order accurate in time. It was first reported by [19] and has been used by a
countless number of authors since then in different contexts. Review articles such
as e.g. [20, 61–63] have summarized and analyzed a large number of possible algo-
rithms for the partitioned solution of coupled mechanical systems and classified the
scheme utilized in this article as a conventional serial staggered (CSS) procedure.
Their analysis extends far beyond the version of the CSS procedure presented here.
Some of these algorithms were implemented and evaluated in terms of their energy
conservation properties in the context of coupling to a lattice Boltzmann method
in [45].
However, and although subject to improvement, the simple version utilized in
this paper suffices to compute all FSI examples presented below.

6 Boundary Conditions

6.1 Kinematic Boundary conditions

The kinematic boundary conditions for the LBM-solver are imposed by the dis-
placements and velocities at the surface of the structure. They can be obtained
in a straight forward manner by the usual postprocessing-step for Finite Element
Methods in structural dynamics.
In the LBM-flow solver, however, macroscopic flow quantities are gained via
forming the corresponding moments of the microscopic flow quantities, i.e. the par-
ticle distribution functions. In turn, macroscopic boundary conditions can only be
set implicitly via the particle distribution functions. A well known and simple way to
introduce no-slip walls is the so-called bounce back scheme. It allows spatial second
order accuracy if the boundary is aligned with one of the lattice vectors eO i D ei =c
and is first order accurate otherwise. To maintain second order accuracy for arbitrar-
ily shaped, moving boundaries we use the modified bounce back scheme developed
in [3, 51] for the velocity boundary conditions.
298 S. Geller et al.

q<0.5 q>=0.5
fIAt+1
fIFt+1 fIAt+1 fIDt+1

fiFt fiAt fIAt fIAt

2q 2q-1
1 q 1 q

Fig. 6 Interpolations for second order bounce back scheme.

In Fig. 6 two cases along a link i can be identified:

(a) wall–node distance qi < 0:5 jei tj and
(b) wall–node distance qi  0:5 jei tj.
The modified bounce back scheme is

t C1 e i uw
fIA D .1  2q/  fiF
t
C 2q  fiA
t
 2wi ; 0:0 < q < 0:5; (14)
cs2
t C1 2q  1 t 1 t ei uw
fIA D  fIA C  fiA  wi ; 0:5  q  1:0: (15)
2q 2q qcs2

The recovery of second order accuracy even for curved boundaries is clearly demon-
strated in [26]. A detailed discussion of LBE boundary conditions can be found in
[30]. In contrast to the simple bounce-back scheme, the use of these interpolation
based no-slip boundary conditions results in a notable mass loss across the no-slip
lines. Yet, the results obtained with bounce-back were inferior which highlights the
importance of a proper geometric resolution of the flow domain. Pressure boundary
conditions are implemented by setting the incoming distributions to [79]

fI .t C t; x/ D fi .t; x/ C fIeq .p0 ; u.tB ; xB // C fieq .p0 ; u.tB ; xB //; (16)

where p0 is the given pressure, tB D t C 12 t, xB D x C 12 ei , .fI ; fi / are an

anti-parallel pair of distributions with velocities ei D eI , and fI is the incoming
and fi the outgoing distribution function value. u is obtained by extrapolation.

Activation/Deactivation of new fluid nodes

If a new fluid node is created due to the moving structure, linear inter- or extrapo-
lation (depending on the geometrical configuration) are used to compute the local
velocity at the concerned node. A local Poisson type iteration described in [57] is
then applied locally to compute the corresponding pressure and the higher order
moments.
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 299

6.2 Force Boundary Conditions

Force boundary conditions are imposed onto the structural boundary by the fluid.
There are two possibilities to evaluate forces on boundaries using the LB Method.
The first one is the momentum-exchange based method and the second one is the
pressure/stress integration based method. A detailed description and comparison of
both methods can be found in [58] and the details for the implementation in 2D can
be found in [27, 45]. The boundary of the structure is given by the interface mesh
consisting of flat triangles in 3D as discussed in Sect. 5.
The momentum exchange method works well if one is only interested in the
value of the integrated forces acting on structural elements, but it leads to noisy
results for the computation of the forces if the boundary of the elements of the
structure or the interface mesh have a size comparable to the grid distance of the
fluid mesh. Incorrect local forces are created if too few links contribute to the force
computation. In this case the stress integration method is favorable.
An advantage of the LBM compared to conventional CFD solvers is the local
availability of the stress tensor. The stress tensor S˛ˇ with scalar pressure is:

S˛ˇ D cs2 ı˛ˇ C

˛ˇ : (17)

˛ˇ is computed from the non equilibrium part of the distributions and ı is the
Kronecker-Delta.
  b1  
t X neq 1

˛ˇ D 1  fi ei ˛ eiˇ  ei ei ı˛ˇ : brium with the resisting force

2 D
i D1
(18)
In three dimensions, 256 different cases emerge of how a point may lie inside the
grid. These cases conform with the Marching-Cubes (3-D) algorithm [55] which
was developed for computer graphics to compute isosurfaces from Cartesian grids
(see Fig. 7).
Two methods for the force evaluation were implemented. The first method we
term stress integration method. It computes and selects the corresponding marching
cubes cases. Then the stresses are inter- and/ or extrapolated onto the boundary via
recovering the stresses on the nodes inside the structure as described in [28,45]. The
second method sets the stresses on the boundary point (i.e. on the interface mesh)
equal to the ones at the nearest neighboring node in normal direction. This is also
termed next-neighbor interpolation [33].
Once the tractions are evaluated on the structural boundary they need to be
integrated in the structural load vector. Again, there are two possibilities, both
of which were already described in [45] and repeated here briefly for complete-
ness. The first possibility is to directly integrate the transferred tractions in the
deformed configuration. This possibility assumes a p-continuous load function
across element boundaries. The second possibility is to apply a composed integra-
tion scheme. Kinks in the load function can then be taken into account. Details for
the convergence rate of such a scheme were also given in [71].
300 S. Geller et al.

Fig. 7 Representation of the 15 base configurations of the Marching Cube algorithm.

7 Experimental benchmark

In [45] the numerical benchmark (see Figs. 8 and 9) of the DFG research unit 493
[81] was studied with the two-dimensional version of V IRTUAL F LUIDS. Addition-
ally, we verified the methodology against the experimental benchmark described in
[32], where the results of our approach are not only compared to the experimental
values, but also put into context to the results obtained by other research groups.
In contrast to the numerical benchmark, the flag is very thin and an additional
mass is attached at its end to add inertia to the system and thereby maintain stable,
periodic motions. The cylinder may rotate freely around its center. Two different
cases (Re D 140 and Re D 190) are considered each of which excites a different
structural mode.
kg
The cylinder is made of aluminium with a density of  D 2828Πm 3  while the
kg
membrane and the attached mass are made of steel with a density of  D 7855Πm 3 .

Young’s modulus of the steel was measured to be E D 200Œ mm2 . kN

The fluid used in the experiment is Polyethyleneglycolsyrup with a density of

kg m2
1050Πm 3  and a kinematic viscosity of 0:000164Πs .
The center of the cylinder is situated at a distance of 55Œmm from the inlet (9).
The origin of the coordinate system is set at the center of the cylinder. The channel
is 338Œmm long and 240Œmm high. The gravitional force acts along the x-axis.
The gravity is equal to 9:81Πsm2  and is scaled to the LB system using Eq. (13). In
comparison to the numerical benchmark, the stiffness (EI) of the flag is reduced by
a factor of 1=1000 and the structural mass is reduced by a factor of 1=700, while the
fluid mass is approximately the same. The system is, therefore, badly conditioned
and a partitioned solution approach is very prone to instabilities.
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 301

Fig. 8 Geometrical definition of the structure in mm [31].

(– 55;120) (284;– 120)

y

(– 36;85) (236;– 85)

w’

w
x

(– 36;85) (236;– 85)

L’

(– 55;120) (284;– 120)

Fig. 9 Physical domain and measurement domain (dashed) in mm [31].

On coarse LB grids, a stable simulation can only be obtained if the mass of the
membrane is increased. For the grid with 340240 nodes the mass of the membrane
needed to be increased by a factor of four. For a higher resolution of the fluid grid
with 1360  960 nodes, it was not necessary to increase the mass to obtain a stable
simulation. For 100 hours of computational time, 1.6 seconds realtime could be
simulated for the grid with 240 nodes in the height. This might seem little, but is
considered to be quite fast compared to other methods. For the grid with 480 nodes
per height, the time was 245 hours for 1.6 seconds realtime on a 2.0 GHz AMD
Opteron 64-Bit Processor. The Mach number in the simulation was equal to 0:05.
The close agreement of the computed trajectories at the end point of the flag with
the ones measured in the experiments is depicted in Fig. 10 and Fig. 11. For a direct
302 S. Geller et al.

Experiment 1: Displacement of Point A

25
Experiment
20 Virtual Fluids - AdhoC

15
displacement y [mm]

10
5
0
–5
–10
–15
–20
–25
63 64 65 66 67 68 69 70 71
displacement x [mm]

Fig. 10 Trajectories of the endpoint compared with experimental data for Re D 140.

Experiment 2: Displacement of Point A

25
Experiment
20 Virtual Fluids - AdhoC

15
displacement y [mm]

10

–5

–10

–15

–20
57 58 59 60 61 62 63 64 65 66 67
displacement x [mm]

Fig. 11 Trajectories of the endpoint compared with experimental data for Re D 190.

comparison to the results obtained by other groups see [32] which is presented in a
different chapter of this book.
The corresponding frequencies of the trajectories are given in Table 1 where the
experimental results were taken as the reference for the computation of the relative
deviation.
The low Reynolds number benchmark can be reproduced very well. Only the
higher Reynolds number benchmark exhibits a notable deviation in the frequency of
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 303

Table 1 Experiments: computed versus measured eigenfrequencies.

case experiment 1: Re D 140 experiment 2: Re D 190
measured eigenfrequency ŒH z 6.38 13.58
computed eigenfrequency ŒH z 6.71 16.70
rel. deviation [%] 5.17 22.98

Fig. 12 Screenshot of a sphere in channel.

about 22Œ%. Additionally, a small unsymmetry is present in the trajectories as the

zero crossing is slightly offset. It is pointed out that other methods, such as an ALE
approach did have severe difficulties with the mesh deformation in this benchmark
due to the fact that the end mass exhibits a rotation of more than 45 degrees at each
cycle. This is no issue in our approach.

8 Validation of 3D-FSI: The falling sphere

We now set out to extend the methodology to three dimensions and evaluate its
performance against a three-dimensional numerical testcase consisting of a falling
sphere in a pipe (see Fig. 12). As only rigid body motions are involved, the fluid
solver V IRTUAL F LUIDS was coupled with the dynamics engine P HYSICS E NGINE
( PE ) [40] to simulate the fluid-structure interactions.
The benchmark is defined as follows: a sphere of radius r and density s falls in
a fluid of density f and viscosity  under the action of gravity. When the sphere
reaches its terminal velocity, its weight FG and the statical buoyancy FA are in
equilibrium with the resisting force FD .
304 S. Geller et al.

4 3
FG D  r s g; (19)
3
4
FA D  r 3 f g; (20)
3
1
FD D CDW f  r 2 U 2 : (21)
2
In an infinite domain, the terminal velocity is then defined as:
s
4 .s  f / 2r
U D g: (22)
3 f CDW
The terminal velocity is evaluated iteratively, as the drag coefficient CDW
depends on the Reynolds number Re D Ud= defined for the sphere.
The drag coefficient CD for laminar flow in the Stokes regime can be computed
as follows:
CD D 24=Re: (23)
Schiller and Naumann [67] extended this equation for higher Reynolds numbers.

24
CD D .1 C 0:15Re 0:687 /: (24)
Re
In case the sphere is falling in a finite domain represented by a circular pipe, addi-
tional corrections have to be considered in order to take into account the wall
effects. These were given by Fayon und Happel [18] as a function of a perturbation
expansion of  D rsphere =rcyli nder :
24
CDW D CD C .K  1/; (25)
Re
with
1  0:758575
KD : (26)
1  2:105 C 2:08653  1:70685 C 0:726036
Eq. (25) is valid for  < 0:6 and Re < 50 and can predict the drag coefficient with
˙5Œ% accuracy.

8.1 Fixed sphere

A sphere of diameter D is submerged into a channel of cylindrical shape and diame-

ter 2D. The sphere is fixed at the centerline of the cylinder and situated at a distance
of 2:4D measured from the inlet of the channel. To simulate a moving sphere,
no-slip boundary conditions are specified on the sphere while velocity boundary
conditions are defined at the inlet of the channel as well as at its walls. The outlet
of the channel is supplied with a fixed pressure boundary condition with p D 0.
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 305

2.4D
D 2D
U=(0,0,u)
U=(0,0,0)
s
P 0
x

8D
y z

Fig. 13 Numerical simulation setup.

Table 2 Deviation from the theoretical value of the drag coefficient acting on the sphere for
different mesh resolutions considering Re D 1.
domain size u0 [m s 1 ]  [m2 s 1 ] D=x CDW [-] rel. deviation [%]
302  120 0.0111 0.166 15 141.19 2.3
602  240 0.0055 0.166 30 142.00 1.7
902  360 0.0037 0.166 45 142.30 1.5

Table 3 Deviation from the theoretical value of the drag coefficient acting on a regular moving
sphere for different mesh resolutions considering Re D 1.
domain size u0 [m s 1 ]  [m2 s 1 ] D=x CDW [-] rel. deviation [%]
302  120 0.0111 0.166 15 139.53 3.4
602  240 0.0055 0.166 30 141.64 1.9
902  360 0.0037 0.166 45 142.38 1.4

Figure 13 depicts the domain schematically. The drag forces are computed via the
momentum exchange method. The simulation was computed with three different
resolutions for Reynolds number Re D 1. The reference value for the drag coeffi-
cient is cd;W D 144:48. Results are given in Table 2. The computed drag coefficient
converges to the reference values as the mesh resolution increases.

8.2 Moving sphere

Having shown that a fixed sphere can be computed with a deviation from the the-
oretical value of the drag coefficient of approximately two percent, the same setup
is now utilized for the simulation of a moving sphere. Therefore, no-slip boundary
conditions are applied on the walls of the channel as well as at its inlet and outlet.
As before, the resulting drag coefficient is measured.
The sphere is submerged into the fluid domain and positioned in the center of the
cylinder at a distance of one sphere diameter from the top of the cylinder. After 500
iterations in the coarse setup and 4000 iterations in the fine setup the drag coefficient
converges. The results are depicted in Table 3.
306 S. Geller et al.

Fig. 14 Partitioning of the fluid domain for the parallel computation mit 14 subdomains.

8.3 Falling Sphere

The falling sphere in a fluid was computed with different mesh resolutions and
different sphere densities. The simulation parameter for this test case are defined
2
as follows: rCyli nder D 0:1Œm, rSphere D 0:05Œm,  D 0:01Œ ms  and f D
kg
1000Œ m 3 . The length of the pipe is L D 0:8Œm. It is thus sufficiently long such that
the terminal velocity is reached before the sphere hits the bottom. No-slip bound-
ary conditions are specified on the cylinder’s wall. Gravity is taken into account
by applying a volume force on the fluid (see Eq. (13)). The boundary condition
on the sphere’s wall is defined via the modified bounce-back rule of second order
accuracy described in Eq. (15) to take into account the sphere’s velocity. As a the
computation is very time consuming and huge amount of memory is needed for the
high resolution configuration, this testcase was simulated with the parallel version
(see Fig. 14) of V IRTUAL F LUIDS. This extension is described in the PhD thesis of
Freudiger [24]. In the highest resolution the computation had 37 million nodes2 .
For the grid and Mach number parameter study the density of the sphere was
kg kg
set to s D 2000Πm 3  and the density of the fluid was set to s D 1000Πm3 . The
viscosity in the LB grid was set to 0:1666. The terminal velocity of the sphere was
determined iteratively from the analytical solution to be: uref D 0:0906. In this case
the Reynolds number is Re D 0:91. The resulting forces at the sphere fluctuated
around 0:1Œ% from their theoretical value. The sphere’s velocity was measured for
different mesh resolutions. The values are listed in Table 4.
The deviation of the sphere’s velocity from the theoretical value for different
Reynolds numbers (i.e. different densities of the sphere) is given in Table 5. The

2
This is equivalent to about seven hundred million degrees of freedom in the D3Q19 model which
was applied here.
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 307

Table 4 Deviation of the sphere’s velocity from its theoretical value for different mesh resolutions.
domain size D=x u [m s 1 ] rel. deviation [%]
302  120 15 0.09080 0.2
602  240 30 0.09215 1.7
902  360 45 0.09186 1.4
1202  480 60 0.09190 1.4
1502  600 75 0.09180 1.3
1802  720 90 0.09177 1.3
2102  840 105 0.09174 1.2

Table 5 Deviation of the sphere’s terminal velocity from the theoretical value for different
Reynolds numbers.
s Re uanalytical Œm s 1  ucomputed Œm s 1  rel. deviation [%]
1200 0.18 0.0184 0.0187 1.6
1500 0.46 0.0457 0.0468 2.4
2000 0.91 0.0906 0.0896 1.1
4000 2.65 0.2653 0.2740 3.3
6000 4.34 0.4340 0.4500 3.7
8000 5.98 0.5977 0.6250 4.5
10000 7.58 0.7578 0.7800 2.9

grid resolution was 302  120 nodes. The deviation w.r.t. the theoretical value is in
the uncertainty range of Eq. (25).

9 Plate in a cross flow

Bathe and Ledezma suggest a plate in a cross flow for an FSI benchmark with geo-
metrically non-linear structural deformations [1, 2]. This benchmark measures the
displacements of the top of the plate at different Reynolds numbers ranging from
Re D 500 to Re D 5000. In [1, 2], the fluid was simulated with an Arbitrary-
Lagrangian-Eulerian (ALE) formulation coupled to a Lagrangian description of the
structure. The computation results in a quasi-stationary deflection of the plate.
For the coupling of the lattice Boltzmann solver V IRTUAL F LUIDS with the
p-FEM solver A DHO C the case Re D 500 and Re D 2500 is considered.
Additionally, a laminar test case at Re D 10 is investigated.
The configuration parameters are as follows: The plate has a Young’s modulus
of E D 7  1010 Œ mN2  with a Poisson’s ratio of s D 0:3 while the thickness of the
plate is h D 0:00125Œm. The fluid domain is block shaped. The bottom as well as
the rectangular plate have no-slip boundary conditions. The velocity at the inlet is
defined via Eq. (27) and the outlet has a constant pressure of p D 0ŒP a.

3V
v.z/ D .2H z  z2 /: (27)
2H 2
308 S. Geller et al.

solid fluid
E = 70 GPa ρ = 1000 kg / m3
ν = 0.3 μ = 0.1 kg / m s
shell thickness = 1.25 mm

inlet
3V (2H z - z)2 slip
v(z) = slip
2H2
0.01 < V < 0.1 m/s

outlet
τn= 0
2m
slip
H=5m
3m

Z 3m
Y
X 2m
8m 12 m

no-slip

Fig. 15 Domain with integrated rectangular plate [1].

Slip boundary conditions are imposed as depicted in Fig. 15. The fluid has a density
kg kg
of  D 1000Πm 3  and a dynamic viscosity of  D 0:1Πms . The kinematic viscosity
2
thus amounts to  D 0:0001Πms .
The characteristic length in the Reynolds number is the height of the channel
H D 5Œm. The plate was discretized by 10  1  18 p-elements with the Ansatz
p ;p ;p
space Sps   .˝sth / and p D 3. The maximum element aspect ratio was 1 W 160.
The fluid domain was computed with a uniform grid and a local refinement around
the plate as given in Fig. 16. The Mach number was M a D 0:05.
In order to obtain a stationary fluid solution, a low Reynolds number of Re D 10
was chosen. Generally, the force acting on the structure can be estimated as given
in [5]:
1
Fd D Cd Av2 : (28)
2
This is an approximation based on Stokes law. Cd is the drag coefficient, A is the
area perpendicular to the flow,  is the density of the fluid and v2 is the free field
velocity of the fluid. The crucial point here is that Cd is not only dependent on the
shape of the object but also on the Reynolds number. It is tabulated for standard
cases e.g. in [60]. For a sphere at Re D 1 the drag coefficient is Cd  140 while at
Re  1000 it decreases to around 0.47. For a plate in cross flow at Re  1000,
Cd  1. To the author’s knowledge, the case of a clamped plate subject to a
cross flow at Re D 10 is not tabulated. It was therefore determined by a fully
three-dimensional computation via CFX where the discretization was adapted to the
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 309

Fig. 16 Grid refinement around the plate domain.

pressure field in three steps. This resulted in a discretization of 2 016 708 tetrahe-
dral elements with 399 399 nodes. A view of the pressure field and its discretization
as computed by CFX is depicted in Fig. 17(a). The pressure field as computed by
LBM is depicted in Fig. 17(b).
From the steady state computation with CFX, the drag coefficient was determined
to be Cd D 9:46. A rough estimate of the expected deflection at the top of the plate
can be computed via application of the force given by Eq. (28) to a clamped beam.
Under the assumption of small displacements and a constant load, the analytical
solution simply is:
qL4
uy .A/ D I (29)
8EI
Fd
where the load is q D , the height of the beam is L D 3Œm. The stiffness EI is
A
the plate stiffness and can be computed by:

Et 3
EI D : (30)
12.1   2 /
310 S. Geller et al.

Pressure
1.01

1.00

1.00

1.00

Z
0.99
[Pa] Y
X
(a) CFX

pressure (Pa)
1.01

1.01

1.00

0.995
Z

Y
X
0.990
(b) V IRTUAL F LUIDS

Fig. 17 Pressure field of plate in cross flow computed by two different methods: (a) CFX:
Visualization with CFX-postprocessor. (b) V IRTUAL F LUIDS.

The first study is to evaluate the simplified stress computation and the correct detec-
tion of the neighbouring nodes next to the structure according to its normal. In order
to achieve a range of different deflections at Re D 10, the benchmark was computed
with three different inflow velocities v1 D 2:0  104 Πms , v2 D 2:0  103Πms  and
v3 D 2:0  102 Πms  where the fluid viscosity was adjusted accordingly. The deflec-
tion is expected to be in the range of the prediction via Eq. (28) and Eq. (29) and
scale quadratically with the inflow velocity as long as the deflection remains within
the limits of geometrical linearity.
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 311

To reach a steady state quickly, the mass of the structure was chosen to be s D
kg
1:0  106 Πm 3  and the first eigenfrequency of the structure was damped out with a
stiffness proportional damping. The Mach number was chosen to be M a D 0:05.
The results lie within the expected range of accuracy and are given in Table 6. The
transient behaviour over time is depicted in Fig. 18.
In a second step, the Reynolds number was raised to Re D 500 by means of
increasing the inflow velocity to v D 1:0 102 Πms , while all other parameters were
kept constant. Unlike at the Reynolds number of Re D 10, the Reynolds number
of Re D 500 leads to transient behaviour in the fluid. Vortices detach at the tip of
the plate and disperse into the domain. A snapshot of the streamlines is depicted
in Fig. 20 clearly demonstrating these vortices. However, they only have a minor
effect on the structure which settles at a quasi stationary deflection of uy .A/ D
8:97  102 Œm ˙ 5:64  104 Œm on a uniform grid. With a refinement of one level

Table 6 Displacement of the rectangular plate for Re D 10.

2
domain size u [m=s]  [ m= s] displacement cantilever [m] displacement [m]
96  144  60 0:0002 0:0001 1:52e 4 1:656e 4
96  144  60 0:002 0:001 1:52e 2 1:645e 2
96  144  60 0:02 0:01 1:52 1:235
192  288  120 0:0002 0:0001 1:52e 4 1:690e 4
192  288  120 0:002 0:001 1:52e 2 1:683e 2
192  288  120 0:02 0:01 1:52 1:249

10
v = 2e – 4[m/s]
v = 2e – 3[m/s]
1 v = 2e – 2[m/s]
beam model

0.1
displacement [m]

0.01

0.001

0.0001

1e–005

1e–006
0 50 100 150 200 250 300 350 400
time [s]

Fig. 18 Displacement of the plate for Re D 10.

312 S. Geller et al.

and two levels the mean deflections are 8:91m  102 Œm and 8:93  102 Œm. The
grid refinement is depicted in Fig. 16
By contrast, Bathe and Ledezma [1] compute the deflection at Point A to be
uy .A/ D 6:6102 Œm. In order to clarify the deviation, the benchmark was recom-
puted with Ansys-CFX Multiphysics Version 12. The explicit coupling algorithms
in Ansys-CFX Multiphysics lead to instabilities for this benchmark which is why
a two-way Fluid-Structure interaction with an implicit coupling needed to be run.
The implicit coupling required an underrelaxation in the displacements with a factor
of ! D 0:15. The time step was chosen to be t D 2:5Œs for fluid and structure
alike. The structure was discretized with 1536 hexahedral elements. It was inte-
grated with a Newmark [59] method with a numerical damping where ˇ D 0:49
and D 0:9. The fluid was discretized with 569 402 tetrahedral Finite Volume ele-
ments with 102 384 nodes. The complete time history of the deflection of point A as
computed by V IRTUAL F LUIDS -A DHO C (LBM-p-FEM) is directly compared to the
time history of Point A obtained by the computation with Ansys-CFX Multiphysics
in Fig. 19.
A direct comparison as shown in Fig. 19 is very strict. Different starting values,
for example, will almost always lead to different time-histories before the system
has locked in. This difference diminishes as time procedes. The deflection at point
A is thus compared in its last oscillation only and amounts to uy .A/ D 9:43 
102 Œm ˙ 1:0  103 Œm for the Ansys-CFX Multiphysics computation. It is thus

0.16
VirtualFluids-AdhoC
CFX-ANSYS
0.14

0.12

0.1
displacement

0.08

0.06

0.04

0.02

0
0 500 1000 1500 2000
time

Fig. 19 Displacement of the plate for Re D 500.

Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 313

Fig. 20 Screen shots of the rectangular plate for Re D 500 (streamlines, vorticity).

4:9% larger in the mean and a little less than twice as large with respect to the
oscillation amplitude as compared to the solution computed by V IRTUAL F LUIDS -
A DHO C (LBM-p-FEM) .
These differences might be due to the fact that LBM discretization is still too
compressible considering the intermediate Mach number of M a D 0:05 as com-
pared to the incompressible nature of the fluid discretization utilized in CFX. These
lower deflections might also partly be attributed to the simpler and less exact next
neighbor force transfer method chosen in the LBM-solver. However, the results
agree quite well considering the complicated physics involved and the completely
different approaches used for this comparison. The reason why both computations,
(V IRTUAL F LUIDS -A DHO C) as well as Ansys-CFX Multiphysics deviate from the
results published in [1, 2] could not be detected and the authors are in contact with
Prof. Bathe in an effort to gain more insight into this issue.
In a third test the Reynolds number was raised to Re D 2500 by increas-
ing the fluid velocity at the inflow to v D 5:0  102Πms  while all other fluid
and structural discretization parameters were kept constant w.r.t. the test case
Re D 500. The resulting deflection at point A is depicted in Fig. 21(a) and the
streamlines are depicted in Fig. 21(b). The mean deflection for this case amounts to
314 S. Geller et al.

(a) Streamlines

(b) Deflection at point A

Fig. 21 Plate in cross flow at Re D 2500.

Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 315

uy .A/ D 1:36Œm. As in the previous case where Re D 500, Bathe and Ledezma
[1] again arrive at a smaller deflection and give uy .A/  1:0Œm3 .
It is pointed out that the explicit coupling setup V IRTUAL F LUIDS -A DHO C
proved to be very efficient also in three dimensions. A typical computation for
this plate in a cross flow benchmark (e.g. the computation of the time history as
shown in Fig. 19) with A NSYS M ULTIPHYSICS took approximately four days when
computed in parallel with four processes on a modern 2GHz Intel Xenon E5405
four processor CPU’s. The same task took only one day with the serial versions
of V IRTUAL F LUIDS -A DHO C. The proposed setup is therefore by more than one
magnitude faster.

10 Beyond benchmarking

10.1 Very large structural deflection

As demonstrated by benchmark of a falling sphere in Sect. 8, it comes as no surprise

that the ability to compute very large displacements without the necessity to remesh
can also be taken advantage of when coupling to flexible structures. As shown in
Fig. 22 it is possible, for example, to let the structural component of the experimen-
tal benchmark described in Sect. 7 fall freely through the fluid domain. This can be
realized via releasing the Dirichlet boundary conditions at the center of the cylin-
der. It is stressed that, once the boundary conditions are correctly imposed and the
coupling setup is validated and verified, these large deflections impose virtually no
additional difficulty to this type of coupling.

10.2 A free surface example

This example is to point out a typical benefit of a partitioned as opposed to a

monolithic solution approach. The computation of other types of surface coupled
problems are possible with only minor changes. Figure 23 shows the coupling
between a tower modeled by high-order FEM hit by a free surface wave. Again,
the tower motion was computed by A DHO C and the fluid by V IRTUAL F LUIDS.

3
Part of the problem here is that such comparative studies are painfully time consuming, even on
the most modern machines. The last run with Ansys-CFX Multiphysics for the case Re D 500,
for example, took 4 days even though it was computed in parallel using one Intel Xeon E5405
processors with four cores at 2.0[GHz]. Many runs were required to tune the simulation parameters
such as a good underrelaxation factor, the proper time step, tight but not too tight convergence
values, a fine but computable spatial resolution of the domain, etc..
316 S. Geller et al.

Fig. 22 Very large displacement FSI: free falling experimental benchmark5 (Re D 140).

10.3 Towards realistic engineering problems

Large parts of this work dealt with the verification and validation of examples which
are interesting for research purposes in order to gain confidence in the involved
codes and the newly developed methods. However, having achieved this it is of inter-
est as well to attempt to compute examples of industrial scale. The Millau bridge
fits into this class of problems. This example is to serve as a motivation for future
development and to discuss the principle applicability of the setup to problems of
this size.
The Millau bridge is the world largest cable-stayed road-bridge spanning the val-
ley of the river Tan in southern France [84]. A picture of the bridge in construction
is depicted in Fig. 24. It was constructed using a timed shifting method. It is prone
to wind induced vibrations especially in the construction phase when the cantilever
arm of the bridge is not yet supported by the subsequent bridge pier.
The computational model of this most critical situation is depicted in Fig. 25. The
structure was discretized by 239 Hexahedral elements of order 4 in all directions
while 2:85  106 nodes were used for the LBM computation in V IRTUAL F LUIDS.
Figure 25 depicts this computational setup, the geometry of the bridge and the valley.
The slices show the magnitude of the velocity vector. The Reynolds number was
set to Re D 1:0  106 and a Smagorinsky turbulence model [75] was applied.
For examples of such size, three main problems arise: (a) The spatial resolution
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 317

Fig. 23 Tower hit by a wave.

Fig. 24 Millau Bridge in construction [84].

318 S. Geller et al.

Fig. 25 Millau Bridge modeled by A DHO C coupled to V IRTUAL F LUIDS.

necessary to resolve the eddies leads to enormous numbers of degrees of freedoms.

Although 2:85  106 nodes for the model depicted in Fig. 25 seems a lot but this is
not enough. With the computational domain having a hight of 400Œm the resolution
along this height is only about 200 nodes resulting in approximately one node for
every two meters or about 2.5 over the height of the cross section of the bridge. One
way to reach an adequate resolution around the bridge is to utilize local refinement.
Another possibility to model eddies which stem from obstacles with the same size
as the grid resolution is to apply a cascaded LBM model; see for example [25] in
which a vortex shedding behind a cylinder with the radius equal to the grid spacing is
accurately modeled. (b) The structural model might yet be too simple. For a realistic
modeling, structural cable elements would be required. Generally, depending on
the level of detail required, the structural solver, too, reaches its limits in terms of
computational complexity. In an effort to reduce the structural complexity of the
model, the bridge deck and pillars are currently modeled as one solid body. This can
lead to an overestimated rotational stiffness due to an overestimated shear modulus.
The computational models are currently refined step by step in order to be able to
tackle this class of engineering problems.
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 319

11 Conclusions

This contribution presents an explicit coupling of a Lattice Boltzmann Method to

a high-order structural discretization for bi-directional Fluid-Structure interaction
with very large displacements. Based on the work of Krafczyk [46,47], it was jointly
developed in the course of the work for the DFG-Research Unit 493 ‘Fluid-Structure
Interaction: Modeling, Simulation, Optimization’ Previous contributions in to this
project were presented in [27, 29, 44, 45, 71].
In all examples, the presented setup turned out to be well suited for this type of
Fluid-Structure interaction. Within this setup it was possible to:

 Use an efficient, explicit overall coupling approach in a straight forward manner,

 verify and validate the results against purely computational and experimental
benchmarks, respectively,
 compute examples where the structure undergoes very large deflections, only
limited by the size of the computational domain,
 extend the approach to three dimensions and apply the methodology to examples
of large scale.
Moreover, this work has applied the proposed methods to realistic engineering
problems. In this context, benchmarks with turbulent flow are subject of current
investigations.

Acknowledgements Financial support of the first authors of each research group (S.Geller and S.
Kollmannsberger) by the German Research Foundation in the framework of the Research Unit 493
Fluid-Struktur-Wechselwirkung is greatfully acknowledged. The authors would also like to thank
K. Iglberger for providing the PE -P HYSICS E NGINE.
320 S. Geller et al.

Appendix

Transformation matrix M

2 3
1 .1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1/
6 7
6 c 2  .1 0 0 0 0 0 1/ 7
6 0 1 1 1 1 1 1 1 1 1 7 1 1
6 7
6 c 4  .1 2 2 2 2 2 2 1 1 1 1 1 1 1 1 1/ 7
6 1 1 1 7
6 7
6 7
6 c  .0 1 1 0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0/ 7
6 7
6 3 7
6 c  .0 2 2 0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0/ 7
6 7
6 7
6 c  .0 0 0 1 1 0 0 1 1 1 1 0 0 0 0 1 1 1 1/ 7
6 7
6 7
6 c 3  .0 0 0 2 2 0 0 1 1 1 1 0 0 0 0 1 1 1 1/ 7
6 7
6 7
6 c  .0 0 0 0 0 1 1 0 1 1 1 1 1 1 1 1/ 7
6 0 0 0 7
6 7
6 3 7
6 c  .0 0 0 0 0 2 2 0 0 0 0 1 1 1 1 1 1 1 1/ 7
6 7
6 2 7
6 c  .0 2 2 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2/ 7
6 7
6 4 7
6 c  .0 2 2 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2/ 7
6 7
6 2 7
6 c  .0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0/ 7
6 7
6 7
6 c 4  .0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0/ 7
6 7
6 7
6 c 2  .0 0 0 0 0 0 1 1 1 0 0 0 0 0 0 0 0/ 7
6 0 1 7
6 7
6 2 7
6 c  .0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1/ 7
6 7
6 2 7
6 c  .0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 0 0 0 0/ 7
6 7
6 3 7
6 c  .0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0/ 7
6 7
6 3 7
6 c  .0 0 0 0 0 0 0 1 1 1 1 0 0 0 0 1 1 1 1/ 7
4 5
c3  .0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1/

Orthogonal basis vectors f˚ i ; i D 0; : : : ; b  1g

˚0;˛ D 1; ˚1;˛ D e2˛  c 2 ; ˚2;˛ D 3.e2˛ /2  6e2˛ c 2 C c 4 ; (31)

˚3;˛ D e˛x ; ˚5;˛ D e˛y ; ˚7;˛ D e˛z ; (32)
˚4;˛ D .3e2˛  5 c 2 /e˛x ; ˚6;˛ D .3e2˛  5 c 2 /e˛y ;
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 321

˚8;˛ D .3e2˛  5 c 2 /e˛z ; (33)

˚9;˛ D 2
3e˛x  e2˛ ; ˚11;˛ D 2
e˛y  2
e˛z ; (34)
˚13;˛ D e˛x e˛y ; ˚14;˛ D e˛y e˛z ; ˚15;˛ D e˛x e˛z ; (35)
˚10;˛ D .2e2˛  3 c 2 /.3e˛x
2
 e2˛ /; ˚12;˛ D .2e2˛  3c 2 2
/.e˛y  2
e˛z /; (36)
˚16;˛ D 2
.e˛y  e˛z
2
/e˛x ; ˚17;˛ D 2
.e˛z  2
e˛x /e˛y ;
˚18;˛ D .e˛x  e˛y /e˛z :
2 2
(37)

References

1. K.J. Bathe and G.A. Ledezma. Benchmark problems for incompressible fluid flows with
structural interactions. Computers and Structures, 85:628–644, 2007.
2. K.J. Bathe and H. Zhang. A mesh adaptivity procedure for CFD & fluid-structure interactions.
Computers and Structures, 87:604–617, 2009.
3. M. Bouzidi, M. Firdaouss, and P. Lallemand. Momentum transfer of a Boltzmann-Lattice fluid
with boundaries. Physics of Fluids, 13(11):3452–3459, 2001.
4. M. Brenk, H.-J. Bungartz, M. Mehl, and T. Neckel. Fluid-Structure Interaction on Cartesian
Grids: Flow Simulation and Coupling Environment. In Fluid-Structure Interaction, Modelling,
Simulation and Optimisation, volume 53 of Lecture Notes in Computational Science and
Engineering, pages 294–335. Springer, 2006.
5. Encyclopedia Britannica. fluid mechanics. Encyclopedia Britannica, http:// www.britannica.
com/ EBchecked/ topic/ 211272/ fluid-mechanics, 2009.
6. S. Chapman and T.G. Cowling. The Mathematical Theory of Non-Uniform Gases. Cambridge
University Press, New York, 1990.
7. J. Chung and G.M. Hulbert. A time integration algorithm for structural dynamics with
improved numerical dissipation: the generalized-˛ method. Journal of Applied Mechanics,
Transactions ASME, 60:371–375, 1993.
8. B. Crouse. Lattice-Boltzmann Strömungssimulation auf Baumdatenstrukturen. PhD thesis,
Lehrstuhl für Bauinformatik, Fakultät für Bauingenieur- und Vermessungswesen, Technische
Universität München, 2003.
9. B. Crouse, E. Rank, M. Krafczyk, and J. Tölke. A LB-based approach for adaptive flow
simulations. International Journal of Modern Physics B, 17:109–112, 2003.
10. D. d’Humières. Generalized lattice-Boltzmann equations. In B. D. Shizgal and D. P. Weave,
editors, Rarefied Gas Dynamics: Theory and Simulations, volume 159 of Prog. Astronaut.
Aeronaut., pages 450–458, Washington DC, 1992. AIAA.
11. D. d’Humières, I. Ginzburg, M. Krafczyk, P. Lallemand, and L. Luo. Multiple-relaxation-time
lattice Boltzmann models in three dimensions. Philosophical Transactions: Mathematical,
Physical and Engineering Sciences, 360(1792):437–451, 2002.
12. A. Düster, H. Bröker, H. Heidkamp, U. Heißerer, S. Kollmannsberger, Z. Wassouf, R. Krause,
A. Muthler, A. Niggl, V. Nübel, M. Rücker, and D. Scholz. AdhoC 4 – User’s Guide. Lehrstuhl
für Bauinformatik, Technische Universität München, 2004.
13. A. Düster, H. Bröker, and E. Rank. The p-version of the finite element method for three-
dimensional curved thin walled structures. International Journal for Numerical Methods in
Engineering, 52:673–703, 2001.
14. A. Düster, L. Demkowicz, and E. Rank. High order finite elements applied to the discrete
Boltzmann equation. International Journal for Numerical Methods in Engineering, 67:1094–
1121, 2006.
15. A. Düster, S. Hartmann, and E. Rank. p-fem applied to finite isotropic hyperelastic bodies.
Computer Methods in Applied Mechanics and Engineering, 192:5147–5166, 2003.
322 S. Geller et al.

16. A. Düster, A. Niggl, and E. Rank. Applying the hp-d version of the FEM to locally enhance
dimensionally reduced models. Computer Methods in Applied Mechanics and Engineering,
196:3524–3533, 2007.
17. A. Düster, D. Scholz, and E. Rank. pq-Adaptive solid finite elements for three-dimensional
plates and shells. Computer Methods in Applied Mechanics and Engineering, 197:243–254,
2007.
18. A. Fayon and J. Happel. Effect of a cylindrical boundary on fixed rigid sphere in a moving
fluid. AIChE journal, 6:55–58, 1960.
19. C. Felippa, K. Park, and J. DeRuntz. Stabilization of staggered solution procedures for fluid-
structure interaction analysis. In T. Blytschko and T. Geers, editors, Computational Methods
for Fluid-Structure Interaction Problems, volume 26, pages 95–124. American Society of
Mechanical Engineers, New York, 1977.
20. C. Felippa, K. Park, and C. Farhat. Partitioned analysis of coupled mechanical systems.
Computer Methods in Applied Mechanics and Engineering, 190:3247–3270, 2001.
21. Z.G. Feng and E.E. Michaelides. The immersed boundary-lattice Boltzmann method for solv-
ing fluid-particles interaction problems. Journal of Computational Physics, 195:602–628,
2004.
22. O. Filippova and D. Hänel. Boundary-fitting and local grid refinement for LBGK models.
International Journal of Modern Physics C, 8:1271–1279, 1998.
23. Message Passing Interface Forum. MPI: A Message-Passing Interface Standard. International
Journal of Supercomputer Applications, 8(3/4), 1994.
24. S. Freudiger. Entwicklung eines parallelen, adaptiven, komponentenbasierten Strmungskerns
fr hierarchische Gitter auf Basis des Lattice Boltzmann Verfahrens. PhD thesis, Technische
Universität Braunschweig, 2009.
25. M. Geier. De-aliasing and stabilization formalism of the cascaded lattice Boltzmann automaton
for under-resolved high reynolds number flow. International Journal for Numerical Methods
in Engineering, 56:1249–1254, 2008.
26. S. Geller, M. Krafczyk, J. Tölke, S. Turek, and J. Hron. Benchmark computations based on
Lattice-Boltzmann, Finite Element and Finite Volume Methods for laminar Flows. Computers
& Fluids, 35:888–897, 2006.
27. S. Geller, J. Tölke, and M. Krafczyk. Lattice Boltzmann Methods on Quadree-Type Grids for
Fluid-Structure Interaction. In H.J. Bungartz and M. Schäfer, editors, Fluid-Structure Interac-
tion, Modelling, Simulation and Optimisation, volume 53 of Lecture Notes in Computational
Science and Engineering, pages 270–293. Springer, 2006.
28. S. Geller, J. Tölke, M. Krafczyk, S. Kollmannsberger, A. Düster, and E. Rank. A coupling
algorithm for high order solids and lattice Boltzmann fluid solvers. In P. Wesseling, E. Onate,
and J. Periaux, editors, Proceedings of the European Conference on Computational Fluid
Dynamics, ECCOMAS CFD 2006, TU Delft, The Netherlands, 2006.
29. S. Geller, J. Tölke, M. Krafczyk, D. Scholz, A. Düster, and E. Rank. Simulation of bidirec-
tional fluid-structure interaction based on explicit coupling approaches of Lattice Boltzmann
and p-FEM solvers. In Proceedings of the Int. Conf. on Computational Methods for Coupled
Problems in Science and Engineering, Santorini, Greece, 2005.
30. I. Ginzburg and D. d’Humières. Multi-reflection boundary conditions for lattice Boltzmann
models. Physical Review E, 68:066614, 2003.
31. J.P. Gomes and H. Lienhart. Experimental study in a fluid-structure interaction reference test
case. In H.J. Bungartz and M. Schäfer, editors, Fluid-Structure Interaction, Modelling, Simula-
tion and Optimisation, volume 53 of Lecture Notes in Computational Science and Engineering,
pages 294–335. Springer, 2006.
32. J.P. Gomes and H. Lienhart. Experimental study in a fluid-structure interaction reference test
case. In H.J. Bungartz and M. Schäfer, editors, Fluid-Structure Interaction, Lecture Notes in
Computational Science and Engineering. Springer, 2010. to appear.
33. A. Halfmann. Ein geometrisches Modell zur numerischen Simulation der Fluid-Struktur-
Interaktion windbelasteter, leichter Flächentragwerke. PhD thesis, Lehrstuhl für Bauinfor-
matik, Fakultät für Bauingenieur- und Vermessungswesen, Technische Universität München,
2002.
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 323

34. K. Han, Y.T. Feng, and D.R.J. Owen. Coupled lattice Boltzmann and discrete element
modelling of fluid-particle interaction problems. Computers & Structures, 85:1080–1088,
2007.
35. K. Han, Y.T. Feng, and D.R.J. Owen. Coupled lattice Boltzmann method and discrete element
modelling of particle transport in turbulent fluid flows. International Journal for Numerical
Methods in Engineering, 72:1111–1134, 2007.
36. U. Heißerer, S. Hartmann, A. Düster, W. Bier, Z. Yosibash, and E. Rank. p-fem for finite
deformation powder compaction. Computer Methods in Applied Mechanics and Engineering,
197:727–740, 2008.
37. U. Heißerer, S. Hartmann, A. Düster, and Z. Yosibash. On volumetric locking-free behavior of
p-version finite elements under finite deformations. Communications in Numerical Methods in
Engineering, 24:1019–1032, 2008.
38. J. Hron and S. Turek. A monolithic FEM solver for ALE formulation of fluid structure inter-
action with configurations for numerical benchmarking. In Proceedings of the International
Conference on Computational Methods for Coupled Problems in Science and Engineering,
Santorini, 2005.
39. B. Hübner, E. Walhorn, and D. Dinkler. A monolithic approach to fluid-structure interaction
using space-time finite elements. Computer Methods in Applied Mechanics and Engineering,
193:2087–2104, 2004.
40. K. Iglberger. pe – physics engine. http://www10.informatik.uni-erlangen.de/klaus/, 2009.
41. K. Iglberger, N. Thürey, and U. Rüde. Simulation of moving particles in 3d with the lattice
Boltzmann method. Computers & Mathematics with Applications, 55:1461–1468, 2008.
42. M. Junk, A. Klar, and L. S. Luo. Asymptotic analysis of the lattice Boltzmann equation. Journal
of Computational Physics, 210:676, 2005.
43. S. Kollmannsberger, A. Düster, and E. Rank. FSI based on bidirectional coupling of high order
solids to a Lattice Boltzmann Method. In Proceedings of PVP 2006, 11th International Sympo-
sium on Emerging Technologies in Fluids, Structures, and Fluid/Structure Interactions, within
the ASME Pressure Vessel and Piping Conference, Vancouver, B.C, Canada, 2006.
44. S. Kollmannsberger, S. Geller, A. Düster, J. Tölke, M. Krafczyk, and E. Rank. Fluid-Structure
Interaction based on Lattice Boltzmann and p-FEM: Verification and Validation. In Proceed-
ings of the International Conference on Computational Methods for Coupled Problems in
Science and Engineering, Ischia Island, Italy, 2009.
45. S. Kollmannsberger, S. Geller, A. Düster, J. Tölke, C. Sorger, M. Krafczyk, and E. Rank. Fixed-
grid Fluid-Structure interaction in two dimensions based on a partitioned Lattice Boltzmann
and p-FEM approach. International Journal for Numerical Methods in Engineering, 79:817–
845, 2009.
46. M. Krafczyk. Gitter-Boltzmann-Methoden: Von der Theorie zur Anwendung. Postdoctoral
thesis, Lehrstuhl für Bauinformatik, Fakultät für Bauingenieur- und Vermessungswesen,
Technische Universität München, 2001.
47. M. Krafczyk, J. Toelke, E. Rank, and M. Schulz. Two-dimensional simulation of fluid-structure
interaction using lattice-boltzmann methods. Computers and Structures, 79:2031–2037, 2001.
48. U. Küttler and W. Wall. Vector extrapolation for strong coupling fluid-structure interaction
solvers. Journal of Applied Mechanics, 2(76), 2009.
49. A. Ladd. Numerical simulations of partiuclate suspensions via a discretized Boltzmann
equation. Part 1: theoretical foundations. Journal of Fluid Mechanics, 271:285, 1994.
50. A. Ladd. Numerical simulations of partiuclate suspensions via a discretized Boltzmann
equation. Part 2: numerical results. Journal of Fluid Mechanics, 271:311, 1994.
51. P. Lallemand and Luo L.-S. Lattice boltzmann method for moving boundaries. Journal of
Computational Physics, 184:406–421, 2003.
52. P. Lallemand and L. Luo. Theory of the lattice Boltzmann method: Dispersion, dissipation,
isotropy, Galilean invariance, and stability. Physical Review E, 61(6):6546–6562, 2000.
53. P. LeTallec and J. Mouro. Fluid structure interaction with large structural displacements.
Computer Methods in Applied Mechanics and Engineering, 190:3039–3067, 2001.
324 S. Geller et al.

54. R. Löhner, J.R. Cebral, F.F. Camelli, J.D. Baum, E.L. Mestreau, and O.A. Soto. Adaptive
embedded/immersed unstructured grid techniques. Archives Of Computational Methods In
Engineering, 14:279–301, 2007.
55. W.E. Lorensen and H.E. Cline. A Discontinuous Galerkin Method for the Navier-Stokes
Equations. ACM SIGGRAPH Computer Graphics archive, 21:163–169, 1987.
56. X.J. Luo. An automatic adaptive directional variable p-version method in 3D curved domains.
PhD thesis, Rensselaer Polytechnic Institute, Troy, New York, 2005.
57. R. Mei, L.S. Luo, P. Lallemand, and D. d’Humières. Consistent Initial Conditions for LBE
Simulations. Computers & Fluids, 35:855–862, 2006.
58. R. Mei, D. Yu, W. Shyy, and L Lou. Force evaluation in the lattice Boltzmann method involving
curved gemoetry. Physical Review E, 65:041203, 2002.
59. N.M. Newmark. A numerical method for structural dynamics. Journal of Engineering Mechan-
ics (ASCE), 85:67–94, 1959.
60. H. Oertel, Böhle M., and U. Dohrmann. Strömnungsmechanik. Vierweg+Teubner, 5. edition,
2009.
61. K. Park and C. Felippa. Partitioned analysis of coupled systems. In T. Blytschko and
T. Hughes, editors, Computational Methods for Transient Analysis, chapter 3, pages 157–219.
North-Holland, Amsterdam-New York, 1984.
62. S. Piperno and C. Farhat. Partitioned procedures for the transient solution of coupled aroelastic
problems – part II: Energy transfer analysis and three dimensional applications. Computer
Methods in Applied Mechanics and Engineering, 190:3147–3170, 2001.
63. S. Piperno, C. Farhat, and B. Larrouturou. Partitioned procedures for the transient solution of
coupled aroelastic problems – part I: Model problem, theory and two-dimensional application.
Computer Methods in Applied Mechanics and Engineering, 124:79–112, 1995.
64. Y.H. Qian, D. d’Humières, and P. Lallemand. Lattice BGK models for Navier-Stokes equations.
Europhysics Letters, 17(6):479–484, 1992.
65. E. Rank, A. Düster, V. Nübel, K. Preusch, and O.T. Bruhns. High order finite elements for
shells. Computer Methods in Applied Mechanics and Engineering, 194:2494–2512, 2005.
66. M. Rheinländer. A consistent grid coupling method for lattice-boltzmann schemes. Journal of
Statistical Physics, 121:49–74, 2005.
67. L. Schiller and A.Z. Naumann. Über die grundlegenden berechnungen bei der schwerkraftauf-
bereitung. Journal of Verein Deutscher Ingenieure, 77:318–320, 1933.
68. D. Scholz. An anisotropic p-adaptive method for linear elastostatic and elastodynamic analysis
of thin-walled and massive structures. Dissertation, Lehrstuhl für Bauinformatik, Fakultät für
Bauingenieur- und Vermessungswesen, Technische Universität München, 2006.
69. D. Scholz, A. Düster, and E. Rank. Model-adaptive fluid-structure interaction using high order
structural elements. In Proceedings of the Int. Conf. on Computational Methods for Coupled
Problems in Science and Engineering, Santorini, Greece, 2005.
70. D. Scholz, A. Düster, and E. Rank. Model-adaptive structural FEM computations for fluid-
structure interaction. In Proceedings of the Third M.I.T. Conference on Computational Fluid
and Solid Mechanics, Cambridge, USA, 2005.
71. D. Scholz, S. Kollmannsberger, A. Düster, and E. Rank. Thin Solids for Fluid-Structure
Interaction. In H.J. Bungartz and M. Schäfer, editors, Fluid-Structure Interaction, Modelling,
Simulation and Optimisation, volume 53 of Lecture Notes in Computational Science and
Engineering, pages 294–335. Springer, 2006.
72. X. Shi and S.P. Lim. A LBM-DLM/FD method for 3D fluid-structure interactions. Journal of
Computational Physics, 226:2028–2043, 2007.
73. X. Shi and N. Phan-Thien. Distributed Lagrange multiplier/fictitious domain method in the
framework of lattice Boltzmann method for fluid-structure interactions. Journal of Computa-
tional Physics, 206:81–94, 2005.
74. Xing Shi and Siak Piang Lim. A LBM-DLM/FD method for 3D fluid-structure interactions.
Journal of Computational Physics, 226:2028–2043, 2007.
75. J. Smagorinsky. General circulation experiments with the primitive equations – i the basic
experiment. Monthly Weather Review, 91:99–164, 1963.
Fluid-Structure Interaction with Lattice Boltzmann and p-FEM 325

76. Y. Sui, Y.-T. Chew, P. Roy, and H.-T. Low. A hybrid immersed-boundary and multi-block lat-
tice Boltzmann method for simulating fluid and moving-boundaries interactions. International
Journal for Numerical Methods in Engineering, 53:1727–1754, 2007.
77. B.A. Szabó and I. Babuška. Finite element analysis. John Wiley & Sons, 1991.
78. B.A. Szabó, A. Düster, and E. Rank. The p-version of the Finite Element Method. In E. Stein,
R. de Borst, and T. J. R. Hughes, editors, Encyclopedia of Computational Mechanics, volume 1,
chapter 5, pages 119–139. John Wiley & Sons, 2004.
79. N. Thürey. A single-phase free-surface Lattice-Boltzmann Method. PhD thesis, Lehrstuhl f”ur
Systemsimulation (Informatik 10), Universität Erlangen-Nürnberg, 2003.
80. S. Turek and J. Hron. Proposal for Numerical Benchmarks for Fluid-Structure Interaction
between an Elastic Object and Laminar Incompressible Flow. In H.J. Bungartz and M. Schäfer,
editors, Fluid-Structure Interaction, Modelling, Simulation and Optimisation, volume 53 of
Lecture Notes in Computational Science and Engineering, pages 371–385. Springer, 2006.
81. S. Turek and J. Hron. Numerical Benchmarks for Fluid-Structure Interaction between an
Elastic Object and Laminar Incompressible Flow. In H.J. Bungartz and M. Schäfer, edi-
tors, Fluid-Structure Interaction, Lecture Notes in Computational Science and Engineering.
Springer, 2010. to appear.
82. E. Walhorn, A. Kölke, B. Hübner, and Dinkler D. Fluidstructure coupling within a monolithic
model involving free surface flows. Computers & Structures, 83:2100–2111, 2005.
83. W. Wall, P. Gammnitzer, and A. Gerstenberger. A strong coupling partitioned approach for
fluid-structure interaction with free surfaces. Computers & Fluids, 36:169–183, 2007.
84. Wikipedia. Viaduc de millau. http://de.wikipedia.org/wiki/Viaduc de Millau.
85. D. Yu, R. Mei, and W. Shyy. A multi-block lattice Boltzmann method for viscous fluid flows.
International Journal for Numerical Methods in Fluids, 39:99–120, 2002.
•
An XFEM Based Fixed-Grid Approach
for 3D Fluid-Structure Interaction

W.A. Wall, A. Gerstenberger, U. Küttler, and U.M. Mayer

Abstract This paper gives an overview on our recent research activities on a fixed
grid fluid-structure interaction scheme that can be applied to the interaction of most
general structures with incompressible flow. The developed approach is based on an
eXtended Finite Element Method (XFEM) based strategy to allow moving interfaces
on fixed Eulerian fluid grids. The enriched Eulerian fluid field and the Lagrangian
structural field are partitioned and iteratively coupled using Lagrange multiplier
techniques for non-matching grids. The approach allows the simulation of the
interaction of thin and bulky structures exhibiting large deformations. Extensions
towards automatic adaptivity and fluid boundary layer meshes are sketched and the
principle applicability to contact simulations of submerged structures are presented.

1 Introduction

Fluid-structure interaction is of great relevance in many fields of engineering as well

is in the applied sciences. Hence, the development and application of respective sim-
ulation approaches has gained great attention over the past decades. Some current
endeavors in this field are: the advancement from special purpose or special problem
to quite general approaches; the desire to capture very general and complex systems;
and the exigent need of robust high quality approaches even for such complex cases,
i.e. approaches that have the potential to turn over from being a challenging and fas-
cinating research topic to a development tool with real predictive capabilities [63].
Often, when interaction effects are essential this comes along with large structural
deformations.
A sketch of the general problem of the interaction of a flow field and a flexible
structure is shown in Fig. 1. The interface  i separates the structural domain ˝ s

W.A. Wall, A. Gerstenberger, U. Küttler, and U.M. Mayer

Institute for Computational Mechanics, Technische Universität München, Boltzmannstr. 15, 85747
Garching, Germany
e-mail: [email protected]

H.-J. Bungartz et al. (eds.), Fluid Structure Interaction II, Lecture Notes 327
in Computational Science and Engineering 73, DOI 10.1007/978-3-642-14206-2 12,
c Springer-Verlag Berlin Heidelberg 2010
328 W.A. Wall et al.

u, p ds

Ωf Γi Ωs

Fig. 1 FSI setup: Interface  i separates fluid field ˝ f and structural field ˝ s .

from the fluid domain ˝ f . Most research and commercial codes that are available
for simulations of the interaction of flows and flexible, often thin-walled, structures
are based on the Arbitrary Lagrangian Eulerian (ALE) method. These approaches
go back to early works like [7, 8, 16, 30, 32, 52]. The essential feature of ALE based
methods is that the fluid field is formulated and solved on a deforming grid. This
fluid grid deforms with the structure at the interface and then the fluid grid is
smoothed within the fluid domain. Within the last fifteen years, our research group
contributed a number of different ALE-based schemes and has applied them to a
variety of problems, see e. g. [19–21, 35–37, 48, 64, 65].
But even the most advanced and best cultured ALE based scheme comes to its
limits where only re-meshing helps to prevent severe fluid mesh distortion. In such
situations, one might be tempted to turn over to approaches that work with a fixed
grid. Here, the interface is described either explicitly, using some kind of Lagrangian
interface markers or a Lagrangian structural discretization, or implicitly, using e. g.
level-set functions on a fixed fluid grid. Changing properties and discontinuities in
the fluid solution have to be taken care of with modifications on the fluid equations
and/or fluid discretization.
Prominent fixed grid methods for incompressible flow include the Immersed
Boundary (IB) method [54, 55] and its many derivations [38, 40, 46]. It is capa-
ble to simulate thin and deformable boundaries and fully fledged, deformable 3d
structures submersed in incompressible flow [66, 68]. An approach with many simi-
larities to the IB method is the so called Distributed Lagrange Multiplier / Fictitious
Domain (DLM/FD) method [26, 27], which has since been extended to simulate
thin, deformable structural surfaces [3,14,42] as well as to flexible and fully fledged
structures [67]. Both methods have in common a Lagrangian mesh for the structure
moving through the fluid mesh and forcing the fluid material inside the structure
to deform as the structure. Other fixed-grid or Cartesian-grid methods are, among
others, [11], [17] and [43].
In [23], we listed a number of requirements that fixed grid methods should fulfill
to be applicable to general structures interacting with incompressible flow:
 the physical and fictitious flow region are completely decoupled such that no
artificial energy gain or loss occurs across the interface,
 the fluid-structure coupling takes place only along the interface,
 there is no mesh size dependency between fluid and structure mesh, and
 it should be possible to turn off most parts of the fictitious domain in order to
reduce memory and performance consumption.
An XFEM Based Fixed-Grid Approach for 3D Fluid-Structure Interaction 329

Until a few years ago, many fixed grid methods in literature did not fulfill one or sev-
eral of above listed requirements. For instance, in most DLM/FD and IB approaches
to date, if used with volume occupying structures, the coupling takes place in the
entire structure domain between the fictitious flow field and the structure domain.
Since the structure movement is tied to the fictitious flow, it is forced to deform in an
incompressible way and artificial viscosity may modify the real physics of the struc-
tural domain. Of course the error made by such coupling might not be prohibitive
high in specific situations, for example, if incompressible biophysical materials are
modelled, however, it still generates unwanted dependencies between the fluid and
the structure mesh and adds additional sources of errors for an already complex
physical problem.
In an attempt to meet all of the above requirements, we proposed a partitioned
iterative coupling scheme between a standard Lagrangian structural description and
an Eulerian formulation for the fluid in [23] that uses features of the eXtended
Finite Element Method (XFEM) [6, 50] and Lagrange multiplier/Mortar techniques
[9,10,53]. With such an XFEM based approach for moving boundaries on fixed fluid
grids, in principal all of the listed requirements can be addressed, which is demon-
strated within this review. Because of the attractive features of XFEM, XFEM-based
techniques have become popular for fixed-grid FSI methods and are meanwhile an
very active research area, see e. g. [39, 61, 69].
The paper is structured as follows: The general FSI problem is stated in Sect. 2
and the intermediate interface field as an external reference is defined. The introduc-
tion of the intermediate interface field as an external reference allows us to derive the
XFEM fluid problem and its coupling to the interface in Sect. 3 separately from the
coupling between structure and interface. The interface-structure coupling and the
overall solution approach is then reviewed in Sect. 4. Extensions to improve bound-
ary layer resultion by means of boundary layer meshes combined with adaptivity
and an outlook towards contact of submerged structures is shown in Sect. 5.

2 Statement of Coupled Fluid-Structure Problem

A general fluid-structure interaction problem statement consists of the description of

fluid and solid fields, appropriate fluid-structure interface conditions at the common
interface and conditions for the remaining boundaries, respectively.

2.1 Single Fields

For the fluid domain ˝ f with the position vector x, the Navier-Stokes equations for
incompressible flow in Eulerian form is stated as

@u
f D f u  r u C r   f C bf in ˝ f ; (1)
@t
r u D 0 in ˝ f : (2)
330 W.A. Wall et al.

Here, the partial time derivative of the fluid velocity u times the fluid density f
is balanced by the convection term, the divergence of the Cauchy stress tensor  f
and external, velocity independent volumetric forces bf . For brevity, the volumet-
ric forces are omitted in the subsequent derivation, however they are included in
the actual implementation and their derivation requires no additional difficulties.
For incompressible flow, the Cauchy stress tensor  f is split into pressure p and a
deviatoric viscous stress tensor  f

 f D pI C  f : (3)

For simplicity of this presentation, we use the Newtonian material law, which
defines the viscous stress as a product of the dynamic viscosity denoted as  and
the strain rate tensor "

 f D 2" D .ruf C .ruf /T /: (4)

Dirichlet and Neumann boundary conditions at  f;D and  f;N are given as
f
uf D uO f in  f;D and  f  nf D hO in  f;N ; (5)

and are applied as usual. Variables with superscript O denote prescribed conditions for
the fluid computation. Without consideration of a moving interface, the weak form
of the Navier-Stokes equation in Eulerian form, after integration by parts, becomes
 f @u   f   
f C v;  u  r u ˝ f C r  v; 
f
v;  ˝ ˝f
@t
   f
C q; r  u ˝ f  v; hO  f;N D 0: (6)

For our implementation, we use linear or quadratic equal-order shape functions for
velocity and pressure. Such a formulation is known to show instabilities for two
reasons – dominating convection and (inf-sup) unstable pairs of velocity and pres-
sure shape functions. The stabilization parameters used throughout this work are
discussed and defined in [23].
In most applications, the structure is described using a Lagrangian description,
where the material derivative becomes a simple partial derivative with respect to
time, such that the momentum equation becomes
@2 d s
s D r   s C s bs : (7)
@t 2

The displacement d s is defined as the difference between the current position x and
the initial position X and bs is an external acceleration field acting on the structural
domain. Structural velocity us and acceleration as are defined as

@us @2 d s
or as D uP s D dR :
s
as D D
@t @t 2
An XFEM Based Fixed-Grid Approach for 3D Fluid-Structure Interaction 331

In the large deformation case it is common to describe the constitutive equation

using a stress-strain relation based on the Green-Lagrange strain tensor E and the
2. Piola-Kirchhoff stress tensor S .E / as a function of E . The 2. Piola-Kirchhoff
stress can be obtained from the Cauchy stress  as

1 T @x
S D J F 1    F T with ED .F  F  I/; F D ;
2 @X
where J denotes the determinant of the deformation gradient tensor F . Dirichlet
and Neumann conditions are defined as

d s D dO
s
and  s  ns D hO
s
in  s;D in  s;N : (8)

For transient simulations, initial conditions are required for structural displace-
ments. The weak from after integration by parts without consideration of the
fluid-structure coupling is
s
.ıd s ; s dR /˝ s C .r ıd s ;  s /˝ s  .ıd s ; hO / s;N D 0:
s
(9)

As demonstrated in Sect. 4, the structural equation is not affected by the fixed-

grid fluid treatment and complex material and geometric behavior can be treated
as usual. In particular, the equation is discretized using different (hybrid-mixed)
element techniques in space and direct time integration schemes. For the numerical
results presented in this article we use the St.-Venant-Kirchhoff material law, i.e. a
linear relation between S and E , for simplicity. All non-linearities are handled via a
Newton-Raphson scheme. As it will become clear in the next section, the structural
formulation is by no means restricted to non-linear elasticity. Any material model
described in the Lagrangean formulation can be used without modification of the
fixed-grid FSI scheme presented in this article.

2.2 Fluid-Structure Interface Conditions

Fluid-structure interaction is a surface-coupled problem with dynamic and kine-

matic coupling conditions only along the interface. For our applications, we assume
no mass flow normal to the interface. Consequently, the normal velocities at the
interface have to match as

uf  nf D dP  ns
s
in  i : (10)

Note the opposite sign due to the different normal vector nf and ns for fluid and
structural domain, respectively. If viscous fluids are considered, there is usually also
a matching condition for the tangential velocities, which can be combined with the
equation above to obtain no-slip coupling conditions as

uf D dP
s
in  i : (11)
332 W.A. Wall et al.

ni
u, p di ds

Ωf Γ fFSI Γi Γ sFSI Ωs

Fig. 2 3-field setup: fluid field ˝ f , interface  i and structural field ˝ s along with respective
domain normals and variables.

The force equilibrium requires the surface traction to be equal as

 f  nf D  s  ns in  i : (12)

A three-field setup as introduced in [23], where fluid and structural fields and
an independent interface field are treated separately, has proven to be very valuable
in separating formulation and implementation of the fixed-grid FSI approach. This
setup – along with the respective variables living on these fields or interfaces – is
shown in Fig. 2. Similar ideas for coupling 2 fields using an intermediate reference
surface (“frame”), named as localized Lagrange multiplier method, can be found
e. g. in [53] and references therein. For the 3-field problem, the interface condition
Eq. (11) is modified as

uf D dP ;
i
(13)
d Dd :
i s
(14)

In other words, if we constrain the velocity of both fields independently to be the

same as the interface velocity, then the original matching condition (11) is fulfilled.
Let  and  be two traction fields on the interface  i such that

 f  nf D ; (15)
 s  ns D : (16)

Then the interface traction balance (12) can be stated as

 D : (17)

Since we generally assume that neither fluid nor structural discretization match
the interface mesh, we need to couple three non-fitting meshes. The single fields can
be described separately using the coupling of each field to the interface. During this
derivation, the interface serves as a reference field with given position and velocity
in time. In the final FSI system, depending on the Lagrange multiplier technique
An XFEM Based Fixed-Grid Approach for 3D Fluid-Structure Interaction 333

and the monolithic or partitioned solution approach, the interface unknowns remain
in the system of equations or can be condensed out.

3 Moving fluid boundaries and interfaces

on fixed Eulerian grids

3.1 From Explicit to Implicit Surfaces

The key element of a fixed-grid FSI approach is to properly define moving inter-
faces on a fixed Eulerian fluid grid. For that purpose, we define a domain ˝ that
contains the fluid domain ˝ f completely and extends into the structural domain ˝ s .
The interface between fluid and structure now becomes an internal interface that
separates ˝ into two subdomains ˝ C and ˝  , where ˝ C corresponds to the phys-
ical fluid domain ˝ f and ˝  is the remaining domain filling ˝. Hence, the flow
field in ˝  is entirely fictitious with no physical meaning to the FSI problem and
should be removed from the system of equations.
As requested in the introduction, the main tasks are to decouple the physical flow
in ˝ C from the fictitious flow in ˝  and, using the 3-field setup, to employ the
interface conditions between the interface and the physical flow field in ˝ C . The
re-formulation of the explicit fluid surface as an embedded interface is depicted in
Fig. 3.
The jump in the velocities u between the physical values uC and the void (u )
can be expressed as

u D uC  „ƒ‚…
u D uO i in  i ; (18)
D0

where the jump height equals the interface velocity uO i from the kinematic fluid-
interface coupling condition Eq. (13). The interface velocity uO i is assumed to be

nf Γ + = Γ fFSI
ni
u, p di u, p di

Ω+ = Ωf Ω–

Ωf Γ fFSI Γi Ω Γi
f;FSI C
(a) Explicit fluid surface  (b) Implicit fluid surface 

Fig. 3 Fluid part of the FSI problem:  i divides the computational fluid domain ˝ into a physical
fluid domain ˝ C and the fictitious domain ˝  .
334 W.A. Wall et al.

given for the moment. Likewise, one can identify a jump in the stress field, which
results in a surface traction vector field  as

  n D  C  n  „ƒ‚…
 n D  in  i : (19)
D0

Initial conditions and Dirichlet and Neumann boundary conditions are only
required for ˝ C as no flow is modelled in ˝  . Likewise, fluid body forces b, if
present, are only applied to ˝ C .
For simplicity of presentation, the one-step- time discretization is employed in
the following. The time-discrete weak form without interface condition is developed
by testing with velocity and pressure test functions v and q, respectively
 f  n    o
v;  u ˝ C t v; f u  r u  r   f ˝ C q; r  u ˝
 
D v; f un C t.1  /f an ˝ : (20)

The acceleration of the old time step is denoted as an . The stabilization terms are
omitted to better clarify the basic principles of this approach.

3.2 Moving Interfaces by the XFEM

As should be clear by now, in the proposed fixed grid scheme, the fluid-structure
interface is generally not aligned with fluid element surfaces. Consequently, we
need to find a way to represent a jump in the primary variables, namely the veloc-
ity and pressure, within the elements. Furthermore, also the derivatives have to be
discontinuous, since the stress field is discontinuous, too.
The eXtended Finite Element Method has been proposed in [6, 50] to represent
discontinuities in the Finite Element approximation. For that purpose, the finite
element shape functions are extended or enriched by using additional degrees of
freedom combined with known solution or special enrichment functions. Applied to
the velocity field, an enriched velocity approximation can be defined as
X X
uh .x; t/ D NI .x/uQ I C NJ .x/ .x; t/uL J C : : : : (21)
I J

Here, uQ I represent the standard nodal degrees of freedom at node I , while additional
degrees of freedom uL I multiplied by a properly chosen enrichment function .x; t/
are used to enhance the solution. The superscript h indicates the discretized field
function.
The choice of the enrichment function depends on the character of the interface
conditions. In the fixed-grid fluid setup, there is a jump from the physical field in ˝ C
to zero velocity and pressure in ˝  . Likewise, derived variables like stress tensors
An XFEM Based Fixed-Grid Approach for 3D Fluid-Structure Interaction 335

Ω+ Ω1+

Ω− Ω−
Ω2+

(a) bulky structures (b) thin structures

Fig. 4 Enrichment for bulky and thin structures: For thin structures, two overlapping enrichments
(Þ and ) represent the discontinuity, while away from the surface, standard degrees of freedom
(ı) are used in ˝ C . Uncut elements within the fictitious domain ˝  get no degrees of freedom.

resulting from velocity derivatives and pressure are only present in the physical
flow field, whereas in ˝  they are zero as no flow occurs. Hence, for our purpose
the enrichment function is defined as a step function L .x; t/ [13].

L .x; t/ D C1 in ˝ C
: (22)
0 in ˝ 

Along  i , both velocity and pressure are discontinuous and enriched with L .
Several interfaces within one fluid element can be handled without problems,
which allows to treat thin structures within an FSI approach on “large” fluid
elements. If several interfaces intersect the local support of unknowns, multiple
enrichments can be used, using one enrichment for each interface as depicted
in Fig. 4. Higher order shape functions are constructed with no additional effort
making this approach capable for consistent low and high order finite element
approximations. Elements divided by discontinuities are subdivided into subdo-
mains and numerical integration of the weak form is performed over sub-domains
[50]. The subdivision is performed using a constrained Delaunay algorithm. The
search to find candidates of intersecting background fluid elements and interface ele-
ments is accelerated using parallel search tree techniques. For a detailed description
of this process for higher order elements with curved interfaces, see [45].
If the interface moves over time, the enriched approximation functions P .x; t/ D
NI .x/ .x; t/ at n and n C 1 are generally different near the interface. Mesh topol-
ogy and fluid node positions do not change. Nodes that happen to be in the structural
domain at time step n and are in the fluid domain at the new time level n C 1 have to
be treated properly as no old solution is available. Our current approach is inspired
by the Ghost fluid methods [1, 17, 31], which are popular for Finite Difference and
Finite Volume methods. Here, values at the old time level are constructed by extrap-
olation of each sides interface velocity, which in our case is given by the movement
of the Lagrangian interface mesh. After extrapolation, a projection step generates
a discrete incompressible velocity field at the old time level. This has proven to be
a robust method, however, the temporal accuracy for moving interfaces has to be
336 W.A. Wall et al.

investigated in more detail. Other ways to deal with this situation are presented in
[12] and [69].

3.3 Fluid-Interface Coupling

Due to the non-fitting fluid and interface meshes, it is necessary to weakly enforce
the fluid-interface conditions Eq. (13). In our work, two different ways of weakly
enforcing the interface constraints on the background grid have been developed
and implemented. In the first approach that has been used in the initial 2D imple-
mentation [23], the fluid-interface condition is enforced weakly at t nC1 at the t nC1
location of the interface  i by testing the condition with a test function ı.x/ along
the interface as
 f  n     
v;  u ˝ C t v; f u  r u ˝ C r  v; pI C  f ˝ C q; r  u ˝
    o
 v;   C  ı; u  ui;nC1  C
 
D vf ; un C t.1  /an ˝ : (23)

The Lagrange multiplier field and the corresponding test functions are discretized
along the interface  i , which establishes an interface mesh along  i . Similarly, we
can define discrete interface velocities ui;h and displacements d i;h . The choice of
appropriate approximations for  can not be made without consideration of the
underlying fluid discretization. Inappropriate choices can lead to leakage, oscil-
lations (zero-energy modes) or locking due to violation of the inf-sup condition,
see e. g. [34, 39, 49]. For the 2D implementation in [23], we used bi-quadratic
ansatz functions NI for fluid velocity and pressure and linear ansatz function NIi
for interface velocities and the Lagrange multiplier field. The Lagrange multiplier
nodes where placed at the intersection of fluid element edges and structural surface
element edges.
This approach has been used successfully for 2D problems and its accuracy
has been shown in [22, 23]. However, it is quite hard to construct stable inter-
face Lagrange multiplier spaces for 3D models. Furthermore, when thinking about
large-scale computations as typically encountered for 3D fluid problems, the parallel
interface mesh generation can become costly and solving the resulting saddle-point
problem in Eq. (23) poses additional problems for the parallel iterative solution of
the linearized fluid system. Hence, we proposed an alternative approach in [24] that
is simpler to implement in parallel and does not result in a saddle point problem like
the Lagrange multiplier approach. It also does not require user-provided stabiliza-
tion parameters for the interface constraint like Nitsches method [4,15,18,28,29,51],
an alternative emerging approach without saddle point structure.
The key point of the new method is the replacement of the vector field  along
the interface with an additional stress field N as independent unknown within inter-
sected fluid elements. Hence, we have three primary unknown fields, an elementwise
An XFEM Based Fixed-Grid Approach for 3D Fluid-Structure Interaction 337

continuous velocity u and pressure p and an elementwise discontinuous Cauchy

stress N and their corresponding test functions v, q and , N respectively. An extra
strain rate balance equation defines the 6 unknowns of the symmetric stress tensor
field uniquely. The resulting task is therefore: find u, p and N such that
 f  ˚     
v;  u ˝ C t v; f u  r u ˝ C r v; pI C  ˝ C q; r  u ˝
   
 ;N "N  "/˝  N  nf ; u  ui / i  v; N  nf / i
  
D vf ; un C t.1  /uP n ˝ C t v; h/ O  N: (24)

Dirichlet and Neumann conditions away from the interface are employed as usual.
Note that we intentionally do not replace  in the fluid momentum equation with
the primary stress unknowns. This way, the (stabilized) fluid formulation remains
unchanged. Such a replacement, however, might become an option, if complex fluid
materials are modelled as, e. g. in [5]. The two strain rates "N and " and the velocity-
dependent stress tensor  are computed from the primary unknowns

1
"N D .N C pI /; (25)
2
1
" D .r u C .r u/T /; (26)
2
 D 2": (27)

The stress is approximated with element-wise discontinuous shape functions and

can be condensed out on the element level! For uncut elements, the stress equation
decouples from the momentum and mass continuity equations and can be omit-
ted, altogether. The resulting global system contains only velocity and pressure
unknowns and has no saddle-point characteristic. For an in-depth discussion of this
approach including benchmark computations and numerical convergence analysis,
see [24]. Exemplarily, two 3D benchmark computations from [24] are shown in
Fig. 5. Computations showed optimal performance and convergence rates for all

(a) Pressure field for case 3D-1Q. (b) Velocity field for case 3D-1Z.

Fig. 5 3D benchmark cases 3D-1Q and 3D-1Z.

338 W.A. Wall et al.

common element shapes, namely equal order, linear and quadratic hexahedral and
tetrahedral elements.
An important property of the new embedded Dirichlet approach with beneficial
consequences for the FSI coupling is that the interface discretization can be chosen
arbitrarily. This allows to choose the interface mesh to be identical to the struc-
ture surface mesh. Hence, no weak coupling between interface and structure mesh
is required and the weak interface-structure coupling as presented in [23] can be
replaced by matching nodes algorithms.
Both fluid-interface coupling approaches in combination with the XFEM formu-
lation allow to decouple the physical domain from the fictitious domain ˝  and a
weak coupling between fluid flow and interface movement. They are now used in
an FSI approach as presented in the next section.

4 Fluid-Structure Coupling

In this section, we combine both fields in one coupled system. Before the dis-
crete system can be established, the interface velocity ui has to be discretized in
time, since interface displacement d i and velocity ui are connected by the simple
differential equation dP D ui . For simplicity of presentation, we chose the one-
i

step- method for the interface movement where the time-discrete form of dP D ui
i

becomes

d i;nC1  d i;n
D  i ui;nC1 C .1   i /ui;n : (28)
t

For second order accuracy,  i has to be chosen as  i D 0:5 resulting in the second
order trapezoidal rule. However, it is noted in [19] that the trapezoidal rule might
lead to oscillations, since no damping is included.
The fully coupled system using three distinct meshes in ˝, ˝ s and  i , respec-
tively, is shown in Fig. 6. The three meshes are required, if the original fluid-
interface coupling with Lagrange multiplier  as described in the previous section is
employed. The structure domain is coupled via a Mortar formulation to the interface

master — slave slave — master

λ μ

u, p di ds

Ω+ = Ω f Ω–

Ω Γi Ωs

Fig. 6 Coupled problem with 3 fields: fluid-interface coupling with Lagrange multiplier approach.
An XFEM Based Fixed-Grid Approach for 3D Fluid-Structure Interaction 339
 
using ı; d s  d i  i . The interface hosts both Lagrange multiplier fields, which
makes it the slave side for both couplings — the fluid-interface coupling and the
structure-interface coupling.

 @u     
ıP f D vf ; C u  r u ˝ C r v; pI C  ˝ C q; r  u ˝
@t
 
 v; / i  ı; u  dP / i ;
i
(29a)
 
ıW i D ıd i ;  C   i (29b)
2
@ d s
ıW s D.ıd s ; s /˝ s ; C.r ıd s ;  s /˝ s  .ıd s ; / i  .ı; d s  d i / i : (29c)
@t 2
Correct scaling of coupling matrices using Eq. 28 allows to solve for velocities in
the fluid field and displacements on the interface mesh and the structural domain in
one monolithic system. After linearization, the final monolithic system at Newton-
Raphson iteration step k is of the form
2 3 2 3 2 3
Fuu Fup MT 0 0 0 u ru
6F 0 7 6 7 6r 7
6 pu Fpp 0 0 0 7 6 p 7 6 p 7
6 7 6 7 6 7
6 M 0 0 D 0 0 7 6 l 7 6 r 7
6 7 6 7 D  6 7 : (30)
6 0 0 DT 0 DT 0 7 6 di 7 6 rd i 7
6 7 6 7 6 7
4 0 0 0 D 0 M 5 4 m 5 4 r 5
0 0 0 0 MT Sd s d s k ds rd s k

Here, Sd s d s represents the linearized structural system matrix and M and D the
standard interface-structure surface Mortar matrices. M and D are the previously
defined Mortar matrices between intersected fluid elements and the interface mesh.
The Lagrange multiplier increments are denoted as l and m. Note that we did
not separate structure and fluid unknowns into degrees of freedom in the interior
of each domain and surface degrees of freedom. Of course, the coupling matrices
influence only surface degrees of freedom of the structure and degrees of freedom of
intersected fluid elements. Note further that the necessary scaling factors to couple
displacements and velocities at the interface are assumed to be contained within the
coupling matrices and the interface residual terms. However, the scaling terms are
exactly the same as in ALE coupling approaches and can essentially be copied from
the appropriate publications, see e. g. [37].
In the monolithic case, that system is solved as given in Eq. (30). For a partitioned
approach, the lower right corner describing the interface-structure coupling can be
replaced by a partitioned Dirichlet-Neumann scheme, where the structural displace-
ment at the surface forms a Dirichlet condition for the interface-fluid system, while
the resulting fluid surface traction from  D  is applied as Neumann condition
on the structure. The details of this approach can be found in [23]. For the dynamic
FSI problems involving incompressible fluid flow and lightweight structures con-
sidered here, we employ an iterative staggered scheme based on [47, 48, 64], where
each field is solved implicitly and an iterative procedure over the fields ensures
340 W.A. Wall et al.

slave — master slave — master

¯ · nf
σ μ
˜

u, p di ds

Ω+ = Ω f Ω–

Ω Γi Ωs

Fig. 7 Coupled problem with 3 fields: fluid-interface coupling with hybrid stress approach.

convergence for the interface conditions at the new time step level n C 1. In [20] it
has been shown that such strong coupling schemes are necessary for these cases. As
proposed in [48,64], we use Aitken’s acceleration scheme for vector sequences [33]
to obtain the relaxation parameter !i .
The monolithic system has two times a saddle point structure due to the two
Lagrange multiplier fields  and . However, even for the partitioned FSI cou-
pling, the fluid-interface system still has a saddle point structure, which hamperes
the intended solution of large fluid systems on parallel machines.
With the new embedded Dirichlet approach, the Lagrange multiplier field 
is replaced by the hybrid stress formulation. Since the stress field is defined on
the background grid, the background fluid could be called the slave side of the
fluid-interface coupling. As mentioned in the previous section, the master field for
the fluid-interface coupling can be chosen arbitrarily and we choose the interface
mesh to be the same as the structure surface, which results in a node-matching FSI
approach between interface mesh and structure surface mesh. This improved setup
is shown in Fig. 7. The corresponding weak form can be written as

 @u     
ıP f D vf ; C u  r u ˝ C r v; pI C  ˝ C q; r  u ˝
@t
  
N "N  "/˝  v; N  nf / i  N  nf ; u  dP / i ;
i
 ; (31a)
 
ıW i D ıd i ; N  nf C   i ; (31b)
2
@ d s
ıW s D.ıd s ; s /˝ s C .r ıd s ;  s /˝ s  .ıd s ; / i  .ı; d s  d i / i : (31c)
@t 2
The new hybrid fluid-interface coupling does not suffer from the saddle point
structure. Here, the monolithic system with condensed element stresses for inter-
sected elements can be stated as
2 3 2 3 2 3
Fuu C Cuu Fup C Cup Cud i 0 0 u ru C rcu
6 F
6 pu Fpp 0 0 0 7 6
7 6 p 7
7 6r
6 p
7
7
6 7 6 i7 6 c 7
6 Cd i u Cd i p Cd i d i DT
0 7 6 d 7 D  6 rd i C rd i 7 :
6 7 6 7 6 7
4 0 0 D 0 M 5 4 m 5 4 r 5
0 0 0 M Sd s d s k d
T s
rd s k
(32)
An XFEM Based Fixed-Grid Approach for 3D Fluid-Structure Interaction 341

The matrices C and force terms rc stem from the stress condensation process,
which is described in detail in [24]. Again the time scaling terms are included in
coupling matrices and residuals.
The advantage of being able to chose the interface discretization identical to the
structure surface discretization is two-fold: For a partitioned scheme, where the
linear system in Eq. (32) is split around the Lagrange multiplier  as usual,
the structure-interface coupling is a node-matching scheme, because for matching
interface-structure surfaces with matching interface-structure shape functions we
have

di D D1
 M d D Id D d ;
s s s
(33)

with I being the identity matrix. This essentially replaces the weak interface-
structure coupling with a matching-nodes coupling approach. For the monolithic
scheme, the identical discretizations remove the need for an extra interface mesh.
It can be removed from the global system, which leaves the following system to be
solved

2 3 2 3 2 3
Fuu C Cuu Fup C Cup Cud i u ru C rcu
4 Fpu Fpp 0 5 4 p 5 D  4 rp 5 : (34)
Cd i u Cd i p Sd d C Cd i d i i d
s s
s
rd C rd i i
s
c

As said before, the coupling terms influence only surface degrees of freedom of the
structure and degrees of freedom of intersected fluid elements.
Two-dimensional examples for the fluid-interface coupling via Lagrange multi-
plier  are given in [22, 23, 62]. The new hybrid approach allows now fully coupled
three-dimensional stationary and fully transient simulations. In this presentation, we
constrain ourself to an illustrative example result of a steady-state 3D FSI simula-
tion. In this example, a flexible structure (Poison ratio  D 0:48, Young’s modulus
E D 90 N=m2 ) is deformed due to a flow through a channel with dimensions
0:5  1:0  3:0m. The fully coupled FSI simulation is performed until steady-state is
reached. Top and bottom channel walls have zero (no-slip) velocity prescribed, the
inflow from the right is prescribed by a parabolic velocity condition. The Reynolds
number based on the mean inflow velocity and the channel height is 16. The station-
ary equilibrium solution is shown in Fig. 8. A transient simulation is given together
with a contact simulation in Sect. 5.2. A detailed description of the 3D FSI approach
using the new embedded Dirichlet formulation will be the topic of an upcoming
paper [25].
342 W.A. Wall et al.

Fig. 8 Stationary flow field through a channel with a flexible structure. The parabolic inflow (right)
and the wall boundaries are standard Dirichlet boundary conditions. The outflow boundary (left)
is of Neumann (zero traction) type. The structure is fixed at the lower wall. The fluid-structure
interface is modeled using the proposed embedded Dirichlet approach.

5 Enhancements towards complex FSI problems

5.1 Adaptivity vs. Boundary Layer Meshes

‘Pure’ fixed-grid methods in principle allow for unlimited deformation of the struc-
ture. In addition, no extra computational costs for mesh movement and mesh
smoothing is required. However, unlike in ALE methods, an adequate (boundary
layer) mesh can rarely be constructed a priori, since the movement of the fluid-
structure interface may vary much more than ALE methods would allow. Hence, for
pure fixed-grid methods special care is necessary to create an appropriate – meaning
an sufficiently fine – mesh that allows for reliable simulation of complex problems.
In [22], we propose and discuss two different techniques to improve this situa-
tion. The first approach is a rather straightforward usage of adaptivity. It is based
on local, adaptive mesh refinement and coarsening scheme combined with an error-
estimator and/or mesh refinement heuristics like the distance to structure surfaces.
Our second proposal to improve the accuracy/efficiency of fixed-grid methods is
to use a hybrid approach combining fixed-grid and ALE techniques. It essentially
adds a surface layer of deformable fluid elements with an ALE formulation to the
structural surface. Such a fluid patch would capture the near-surface flow efficiently
with an appropriate boundary layer mesh, which is then coupled to the fixed Eule-
rian background mesh. A sketch of this setup using the original Lagrange multiplier
for the fluid-interface coupling is shown in Fig. 9. The coupled fixed fluid - moving
fluid system can then efficiently be solved monolithically as demonstrated in [22].
An XFEM Based Fixed-Grid Approach for 3D Fluid-Structure Interaction 343

master — slave slave — master

λ μ κ

u, p ui uALE , pALE ds

Ω+ Ω−

i
Ω Γ ΩALE Ωs

Fig. 9 Boundary layer mesh approach: a fixed fluid grid and a deforming ALE fluid grid are
coupled using the XFEM/Lagrange Multiplier approach. The moving ALE grid can in turn be
coupled to the structure field. Shown are also variables living on each domain and the Lagrange
multiplier fields.

(a) Maximal refinement used in computation

(b) x-component of the velocity field

Fig. 10 CFD benchmark: Finest, adapted mesh used to calculate lift and drag values on the
cylinder. It consists of 284 linear Lagrange multiplier elements and 2740 quadratic 9-node fluid
elements. Fig. 10(b) shows the x-component of the velocity field. The structure is not displayed.

Our new hybrid fluid-interface coupling would simplify the fixed-fluid moving-fluid
coupling, as the interface mesh is not required anymore.
The following example shall further clarify the principle difference between
automatic adaptivity and the boundary layer mesh approach. In Fig. 10, the flow
around a cylinder has been resolved by using Cartesian subdivision of elements
with hanging nodes as described in [22]. Due to the circular shape, it is not possible
to generate a mesh that optimally resolves only the strong gradient normal to the
cylinder surface. Figure 10 shows the finest mesh that we used to study this prob-
lem. With this, we could already achieve very good agreement with the averaged
results given in [59].
In contrast, with a boundary layer mesh, instead of refining the mesh near the
cylinder surface as before, we used the proposed hybrid approach. The final solution
can be seen in Fig. 11. The velocity component ux of the fluid solution is shown over
the whole fluid domain in Fig. 11(a) and near the interface in Fig. 11(b). The surface
344 W.A. Wall et al.

(a) x-component of the velocity field (b) close up view

Fig. 11 CFD benchmark using the boundary layer mesh approach: The initial mesh was refined
in the inflow region and in addition, a fluid mesh patch surrounding the structure is applied.

fitted mesh resembles the boundary layer near the cylinder surface much closer and
the boundary layer can be resolved much more efficiently. Of course, the boundary
layer mesh approach can be combined with the aforementioned h-adaptivity. For
high Reynolds number flows, the boundary layer mesh approach should outperform
adaptivity by element subdivision, especially for three-dimensional problems.

5.2 Contact of Submerged Structures

Contact of submerged structures can be observed in many FSI problems. For

instance, in bio-physical applications, closing valves, collapsing veins, interaction
of blood cells are a few examples, where numerical tools could provide valuable
insight. With ALE-methods, the deforming mesh essentially does not allow two
objects to come into full contact. With special techniques, however, even with
ALE fluid methods simplified FSI contact is possible, see e.g. [58, 60]. Fixed-grid
methods, on the other hand, have no deforming mesh between such objects and
potentially open up an entire new field that can be treated by numerical tools. Recent
examples for fixed-grid FSI simulations including contact of submerged structures
can be found in [2, 41, 57].
In [44], the presented XFEM-FSI approach has been combined with a new con-
tact formulation based on dual-Lagrange multipliers and a so-called primal-dual
active set strategy for contact constraint enforcement as described in [56]. From an
algorithmic point of view, the key element of the contact formulation is the complete
condensation of the contact Lagrange multipliers within the structural formula-
tion. Consequently, the structure block can be extended with a contact formulation
without interfering with the FSI formulation. To our knowledge, [44] presents the
first fluid-structure-interaction approach including advanced approaches for FSI and
contact.
A preliminary example of such FSI contact is shown in Fig. 12. A body force
pulls a little, soft block towards a large stiff block, that has a fixed bottom surface.
At the same time, a growing inflow at the top of the fluid domain pushes the little
An XFEM Based Fixed-Grid Approach for 3D Fluid-Structure Interaction 345

(a) setup (b) close up views

Fig. 12 Setup of contact example and 3 snapshots of the first contact between a soft body and a
relatively stiff block.

block additionally in x-direction towards the wall. Two exits at the lower end of
the channel allow the flow to exit the domain. The three little pictures in Fig. 12(b)
show (from top to bottom): how the fluid is pushed out of the closing gap between
the little and the large block, the moment of largest deformation of the little block
at first contact, and the bouncing back of the small block. The same simulation
without fluid flow (not shown here) allows the block to bounce back many times to
approximately the same height from where it started. Including the FSI simulation,
the movement is damped by the fluid viscosity, such that it comes to rest on the
lower block quickly. Since no frictional contact between both objects is employed,
eventually, the little block is pushed sideways out of the fluid domain. Note that in
this simulation we have not employed any special wet or dry contact physics and that
this simulation is not properly resolved to be quantitatively accurate. The example
shall just demonstrate the increased flexibility that comes with fixed grid methods
for FSI.
346 W.A. Wall et al.

6 Conclusion

Fixed grid methods for fluid-structure interaction are subject of a growing number of
current research undertaken. The ultimate goal is to remove the burden of fluid mesh
movement and, if deformation of the structure becomes excessive, remeshing. For
that purpose, we worked out a list of minimal requirements that have to be fulfilled
before fixed grid methods can match or surpass ALE based methods.
In an attempt to meet these requirements, we proposed a new XFEM based
approach. Main improvements compared to most IB and DLM/FD methods are the
complete removal of the fictitious domain and a sharp interface description. Phys-
ical and fictitious domains are completely decoupled and the fluid unknowns from
within the fictitious domain could be removed from the fluid system of equation.
Due to the use of a three-field approach, the treatment of the structure is indepen-
dent of any internals of the fluid XFEM discretization. From the structural point of
view, the coupling exists solely between structural surface and a matching interface
mesh. The newly developed embedded Dirichlet technique for the fluid-interface
allows to apply parallel, iterative solution techniques by avoiding the saddle point
characteristics of non-matching mesh coupling.

Acknowledgements The present study is supported by grant WA1521/1 of the “Deutsche Forsch-
ungsgemeinschaft” (DFG) within the DFG’s Forschergruppe 493 “FSI: Modelling, Simulation,
and Optimization”. This support is gratefully acknowledged.

References

1. Arienti, M., Hung, P., Morano, E., Shepherd, J.E.: A level set approach to Eulerian-Lagrangian
coupling. Journal of Computational Physics 185(1), 213–251 (2003)
2. Astorino, M., Gerbeau, J.F., Pantz, O., Traoré, K.F.: Fluid-structure interaction and multi-
body contact: Application to aortic valves. Computer Methods in Applied Mechanics and
Engineering 198(45-46), 3603–3612 (2009)
3. Baaijens, F.P.T.: A fictitious domain/mortar element method for fluid-structure interaction.
International Journal for Numerical Methods in Fluids 35(7), 743–761 (2001)
4. Becker, R., Burman, E., Hansbo, P.: A nitsche extended finite element method for incom-
pressible elasticity with discontinuous modulus of elasticity. Computer Methods in Applied
Mechanics and Engineering 198(41-44), 3352–3360 (2009)
5. Behr, M.A., Franca, L.P., Tezduyar, T.E.: Stabilized finite element methods for the velocity-
pressure-stress formulation of incompressible flows. Computer Methods in Applied Mechanics
and Engineering 104(1), 31–48 (1993)
6. Belytschko, T., Black, T.: Elastic crack growth in finite elements with minimal remeshing.
International Journal for Numerical Methods in Engineering 45(5), 601–620 (1999)
7. Belytschko, T., Kennedy, J.M.: Computer models for subassembly simulation. Journal of
Nuclear Engineering and Design 49, 17–38 (1978)
8. Belytschko, T., Kennedy, J.M., Schoeberle, D.: Quasi-Eulerian finite element formulation for
fluid structure interaction. Journal of Pressure Vessel Technology 102, 62–69 (1980)
9. Bernardi, C., Maday, Y., Patera, A.T.: Asymptotic and numerical methods for partial differen-
tial equations with critical parameters, vol. 384, chap. Domain decomposition by the mortar
element method, pp. 269–286. Reidel, Dordrecht (1993)
An XFEM Based Fixed-Grid Approach for 3D Fluid-Structure Interaction 347

10. Bernardi, C., Maday, Y., Patera, A.T.: A new nonconforming approach to domain decomposi-
tion: the mortar element method. Nonlinear partial differential equations and their applications
299, 13–51 (1994)
11. Cirak, F., Radovitzky, R.: A Lagrangian-Eulerian shell-fluid coupling algorithm based on level
sets. Computers & Structures 83(6-7), 491–498 (2005)
12. Codina, R., Houzeaux, G., Coppola-Owen, H., Baiges, J.: The fixed-mesh ale approach for
the numerical approximation of flows in moving domains. Journal of Computational Physics
228(5), 1591–1611 (2009)
13. Daux, C., Moës, N., Dolbow, J., Sukumar, N., Belytschko, T.: Arbitrary cracks and holes
with the extended finite element method. International Journal for Numerical Methods in
Engineering 48(12), 1741–1760 (2000)
14. De Hart, J., Peters, G.W., Schreurs, P.J., Baaijens, F.P.: A two-dimensional fluid-structure
interaction model of the aortic value. Journal of Biomechanics 33(9), 1079–1088 (2000)
15. Dolbow, J., Harari, I.: An efficient finite element method for embedded interface problems.
International Journal for Numerical Methods in Engineering 78(2), 229–252 (2009)
16. Donéa, J., Fasoli-Stella, P., Giuliani, S.: Lagrangian and Eulerian finite element techniques for
transient fluid-structure interaction problems. In: Trans. 4th Int. Conf. on Structural Mechanics
in Reactor Technology (1977)
17. Fedkiw, R.P., Aslam, T., Merriman, B., Osher, S.: A non-oscillatory Eulerian approach to inter-
faces in multimaterial flows (the ghost fluid method). Journal of Computational Physics 152(2),
457–492 (1999)
18. Fernández-Méndez, S., Huerta, A.: Imposing essential boundary conditions in mesh-free
methods. Computer Methods in Applied Mechanics and Engineering 193(12-14), 1257–1275
(2004)
19. Förster, C., Wall, W.A., Ramm, E.: On the geometric conservation law in transient flow calcu-
lations on deforming domains. International Journal for Numerical Methods in Fluids 50(12),
1369–1379 (2006)
20. Förster, C., Wall, W.A., Ramm, E.: Artificial added mass instabilities in sequential staggered
coupling of nonlinear structures and incompressible viscous flows. Computer Methods in
Applied Mechanics and Engineering 196(7), 1278–1293 (2007)
21. Gee, M.W., Küttler, U., Wall, W.A.: Truly monolithic algebraic multigrid for fluid–structure
interaction. International Journal for Numerical Methods in Engineering submitted (2009)
22. Gerstenberger, A., Wall, W.A.: Enhancement of fixed-grid methods towards complex fluid-
structure interaction applications. International Journal for Numerical Methods in Fluids 57(9),
1227–1248 (2008)
23. Gerstenberger, A., Wall, W.A.: An extended finite element method / Lagrange multiplier
based approach for fluid-structure interaction. Computer Methods in Applied Mechanics and
Engineering 197(19-20), 1699–1714 (2008)
24. Gerstenberger, A., Wall, W.A.: An embedded Dirichlet formulation for 3d continua. Interna-
tional Journal for Numerical Methods in Engineering DOI: 10.1002/nme.2755 (2009)
25. Gerstenberger, A., Wall, W.A.: An fixed-grid approach for 3d fluid-structure interaction in
preparation (2010)
26. Glowinski, R., Pan, T.W., Hesla, T.I., Joseph, D.D.: A distributed Lagrange multiplier / fic-
titious domain method for particulate flows. International Journal of Multiphase Flow 25(5),
755–794 (1999)
27. Glowinski, R., Pan, T.W., Periaux, J.: A fictitious domain method for external incompressible
viscous flow modeled by Navier-Stokes equations. Computer Methods in Applied Mechanics
and Engineering 112(1-4), 133–148 (1994)
28. Hansbo, A., Hansbo, P.: An unfitted finite element method, based on nitsche’s method, for
elliptic interface problems. Computer Methods in Applied Mechanics and Engineering 191(47-
48), 5537–5552 (2002)
29. Hansbo, A., Hansbo, P.: A finite element method for the simulation of strong and weak dis-
continuities in solid mechanics. Computer Methods in Applied Mechanics and Engineering
193(33-35), 3523–3540 (2004)
348 W.A. Wall et al.

30. Hirth, C., Amsden, A.A., Cook, J.: An Arbitrary Lagrangian-Eulerian computing method for
all flow speeds. Journal of Computational Physics 14, 227–253 (1974)
31. Hong, J.M., Shinar, T., Kang, M., Fedkiw, R.: On boundary condition capturing for multiphase
interfaces. Journal of Scientific Computing 31(1), 99–125 (2007)
32. Hughes, T.J., Liu, W.K., Zimmermann, T.: Lagrangian-Eulerian finite element formulation for
incompressible viscous flows. Computer Methods in Applied Mechanics and Engineering 29,
329–349 (1981)
33. Irons, B.M., Tuck, R.C.: A version of the Aitken accelerator for computer iteration. Interna-
tional Journal for Numerical Methods in Engineering 1(3), 275–277 (1969)
34. Ji, H., Dolbow, J.E.: On strategies for enforcing interfacial constraints and evaluating jump con-
ditions with the extended finite element method. International Journal for Numerical Methods
in Engineering 61(14), 2508–2535 (2004)
35. Küttler, U., Förster, C., Wall, W.A.: A solution for the incompressibility dilemma in partitioned
fluidstructure interaction with pure Dirichlet fluid domains. Computational Mechanics 38(4–5),
417–429 (2006)
36. Küttler, U., Wall, W.A.: Fixed-point fluid-structure interaction solvers with dynamic relaxation.
Computational Mechanics 43(1), 61–72 (2008)
37. Küttler, U., Wall, W.A.: Vector extrapolation for strong coupling fluid-structure interaction
solvers. Journal of Applied Mechanics 76(2) (2009)
38. Lee, L., LeVeque, R.J.: An immersed interface method for incompressible Navier-Stokes
equations. SIAM Journal on Scientific Computing 25(3), 832–856 (2003)
39. Legay, A., Chessa, J., Belytschko, T.: An Eulerian-Lagrangian method for fluid-structure
interaction based on level sets. Computer Methods in Applied Mechanics and Engineering
195(17-18), 2070–2087 (2006)
40. LeVeque, R.J., Calhoun, D.: Cartesian grid methods for fluid flow in complex geometries. In:
L.J. Fauci, S. Gueron (eds.) Computational Modeling in Biological Fluid Dynamics, vol. 124,
pp. 117–143. IMA Volumes in Mathematics and its Applications, Springer-Verlag (2001)
41. van Loon, R., Anderson, P., van de Vosse, F.: A fluid-structure interaction method with solid-
rigid contact for heart valve dynamics. Journal of Computational Physics 217(2), 806–823
(2006)
42. van Loon, R., Anderson, P.D., Baaijens, F.P., van de Vosse, F.N.: A three-dimensional fluid-
structure interaction method for heart valve modelling. Comptes Rendus Mecanique 333(12),
856–866 (2005)
43. Löhner, R., Baum, J.D., Mestreau, E., Sharov, D., Charman, C., Pelessone, D.: Adap-
tive embedded unstructured grid methods. International Journal for Numerical Methods in
Engineering 60(3), 641–660 (2004)
44. Mayer, U., Popp, A., Gerstenberger, A., Wall, W.A.: 3d fluid-structure interaction based on a
combined XFEM/FSI and dual mortar contact approach. Computational Mechanics accepted
(2009)
45. Mayer, U.M., Gerstenberger, A., Wall, W.A.: Interface handling for three-dimensional higher-
order XFEM computations in fluid-structure interaction. International Journal for Numerical
Methods in Engineering 47(7), 846–869 (2009)
46. Mittal, R., Iaccarino, G.: Immersed boundary methods. Annual Review of Fluid Mechanics
37(1), 239–261 (2005)
47. Mok, D.P.: Partitionierte Lösungsansätze in der Strukturdynamik und der Fluid-Struktur-
Interaktion. Tech. Rep. PhD Thesis, Report No. 36, Institute of Structural Mechanics,
University of Stuttgart (2001)
48. Mok, D.P., Wall, W.A.: Partitioned analysis schemes for the transient interaction of incom-
pressible flows and nonlinear flexible structures. In: W. Wall, K.U. Bletzinger, K. Schweizerhof
(eds.) Trends in Computational Structural Mechanics, pp. 689–698. CIMNE: Barcelona (2001)
49. Moës, N., Béchet, E., Tourbier, M.: Imposing Dirichlet boundary conditions in the extended
finite element method. International Journal for Numerical Methods in Engineering 67(12),
1641–1669 (2006)
50. Moës, N., Dolbow, J., Belytschko, T.: A finite element method for crack growth without
remeshing. International Journal for Numerical Methods in Engineering 46(1), 131–150 (1999)
An XFEM Based Fixed-Grid Approach for 3D Fluid-Structure Interaction 349

51. Nitsche, J.: Über ein Variationsprinzip zur Lösung von Dirichlet-Problemen bei Verwen-
dung von Teilräumen, die keinen Randbedingungen unterworfen sind. Abh. Math. Sem. Univ.
Hamburg 36, 915 (1971)
52. Noh, W.: CEL: A time-dependent two-space-dimensional coupled Eulerian-Lagrangian code.
In: B. Alder, S. Fernbach, M. Rotenberg (eds.) Methods in Computational Physics, vol. 3, pp.
117–179. Academic Press: New York (1964)
53. Park, K.C., Felippa, C.A., Ohayon, R.: Partitioned formulation of internal fluid-structure inter-
action problems by localized Lagrange multipliers. Computer Methods in Applied Mechanics
and Engineering 190(24-25), 2989–3007 (2001)
54. Peskin, C.S.: Numerical analysis of blood flow in the heart. Journal of Computational Physics
25(3), 220–252 (1977)
55. Peskin, C.S.: The immersed boundary method. Acta Numerica 11(1), 479–517 (2002)
56. Popp, A., Gee, M.W., Wall, W.A.: A finite deformation mortar contact formulation using a
primal-dual active set strategy. International Journal for Numerical Methods in Engineering
79(11), 1354–1391 (2009)
57. Diniz dos Santos, N., Gerbeau, J.F., Bourgat, J.F.: A partitioned fluid-structure algorithm for
elastic thin valves with contact. Computer Methods in Applied Mechanics and Engineering
197(19-20), 1750–1761 (2008)
58. Sathe, S., Tezduyar, T.: Modeling of fluidstructure interactions with the spacetime finite
elements: contact problems. Computational Mechanics 43(1), 51–60 (2008)
59. Schäfer, M., Turek, S.: Flow Simulation with High-Performance Computers II, Notes on
Numerical Fluid Mechanics, vol. 52, article Benchmark Computations of Laminar Flow
Around a Cylinder, pp. 547–566. Vieweg (1996)
60. Tezduyar, T.E., Sathe, S.: Modelling of fluid-structure interactions with the space-time finite
elements: Solution techniques. International Journal for Numerical Methods in Fluids 54(6-8),
855–900 (2007)
61. Wagner, G.J., Ghosal, S., Liu, W.K.: Particulate flow simulations using lubrication the-
ory solution enrichment. International Journal for Numerical Methods in Engineering 56(9),
1261–1289 (2003)
62. Wall, W.A., Gamnitzer, P., Gerstenberger, A.: Fluidstructure interaction approaches on fixed
grids based on two different domain decomposition ideas. International Journal of Computa-
tional Fluid Dynamics 22(6), 411–427 (2008)
63. Wall, W.A., Gerstenberger, A., Gamnitzer, P., Förster, C., Ramm, E.: Large deformation fluid-
structure interaction – advances in ALE methods and new fixed grid approaches. In: H.J.
Bungartz, M. Schäfer (eds.) Fluid-Structure Interaction: Modelling, Simulation, Optimisation,
LNCSE, vol. 53, pp. 195–232. Springer Verlag (2006)
64. Wall, W.A., Mok, D.P., Ramm, E.: Partitioned analysis approach of the transient coupled
response of viscous fluids and flexible structures. In: W. Wunderlich (ed.) Solids, Structures
and Coupled Problems in Engineering, Proc. ECCM ’99. Munich (1999)
65. Wall, W.A., Ramm, E.: Fluid-structure interaction based upon a stabilized (ALE) finite element
method. In: S. Idelsohn, E. Oñate, E. Dvorkin (eds.) Computational Mechanics New Trends
and Applications, Proc. 4th World Congress on Computational Mechanics - Buenos Aires.
CIMNE, Barcelona (1998)
66. Wang, X., Liu, W.K.: Extended immersed boundary method using FEM and RKPM. Computer
Methods in Applied Mechanics and Engineering 193(12-14), 1305–1321 (2004)
67. Yu, Z.: A DLM/FD method for fluid/flexible-body interactions. Journal of Computational
Physics 207(1), 1–27 (2005)
68. Zhang, L., Gerstenberger, A., Wang, X., Liu, W.K.: Immersed finite element method. Computer
Methods in Applied Mechanics and Engineering 193(21-22), 2051–2067 (2004)
69. Zilian, A., Legay, A.: The enriched space-time finite element method (EST) for simultane-
ous solution of fluid-structure interaction. International Journal for Numerical Methods in
Engineering 75(3), 305–334 (2008)
•
Fluid-Structure Interaction in the Context
of Shape Optimization and Computational
Wind Engineering

M. Hojjat, E. Stavropoulou, T. Gallinger, U. Israel, R. Wüchner,

and K.-U. Bletzinger

Abstract Within this contribution, an integrated concept for the shape optimal
design of light-weight and thin-walled structures like shells and membranes subject
to fluid flow is presented. The Nested Analysis and Design approach is followed and
a partitioned FSI simulation for the state analysis is embedded. The gained modular-
ity allows for the adaption of the single ingredients to various technical applications
by choosing appropriate coupling algorithms for the solution of the coupled prob-
lem and the sensitivity analysis as well as different strategies to describe the shapes
to be optimized. A non-matching grid capability at the coupling interface supports
this flexibility. The focus here is on problems of aeroelasticity in the field of Compu-
tational Wind Engineering. To ensure reliable results, investigations on the correct
modeling as well as goal-oriented benchmarking are carried out. Moreover, special
emphasis is given to the appropriate combination of different approaches for shape
description in establishing the closed design cycle. Finally, the success of the over-
all solution and optimization strategy is demonstrated with an example of a hybrid,
light-weight structure, subject to turbulent wind flow.

1 Introduction

In modern structural design, there is a strong tendency to build light structures

with very efficient load-carrying behavior. This is motivated by the optimal use of
material and their intended use (retractable roofs, airplanes, etc.). The design of
light-weight structures is a challenging task in structural engineering and demands
for additional investigations and in-depth analysis of effects which are of minor
importance in the design of standard structures. This is due to the fact that in case
of structures subject to loads caused by surrounding fluid flow, lightness and slen-
derness make structures more sensitive to flow-induced deformations and vibrations.

M. Hojjat, E. Stavropoulou, T. Gallinger, U. Israel, R. Wüchner, and K.-U. Bletzinger

Lehrstuhl für Statik, Technische Universität München, Germany
e-mail: fhojjat,stavropoulou,gallinger,israel,wuechner,[email protected]

H.-J. Bungartz et al. (eds.), Fluid Structure Interaction II, Lecture Notes 351
in Computational Science and Engineering 73, DOI 10.1007/978-3-642-14206-2 13,
c Springer-Verlag Berlin Heidelberg 2010
352 M. Hojjat et al.

These deformations may change the flow field considerably which results in an inter-
action between the structure and the flow field. Typical examples in civil engineering
are wind-induced effects on thin shells like chimneys, extremely light membrane
structures like wide-span roofs and severe oscillations of slender bridges like the
famous Tacoma Narrows bridge disaster. In these cases, the isolated consideration
of the structure, i.e. reducing the complex physical problem with aeroelastic effects
by finding appropriate assumptions about the fluid load, involves the risk of ignoring
effects which are relevant for the final design. Therefore, the reliable and verified
numerical simulation of the surface-coupled problem including the fluid-structure
interactions (FSI) is necessary and the modeling of the coupling, as well as the sin-
gle fields is a crucial part of the design procedure. Hence, the goal-oriented and
detailed benchmarking of all parts of the numerical simulations is indispensable.
Due to the interaction of the single fields, determination of a good and “opti-
mal” structure is far from trivial and hardly possible only on the basis of experience.
Therefore there is a strong need for an overall methodology to improve and optimize
the chosen designs. Since the constituents of the numerical wind tunnel, i.e. light
thin-walled structures undergoing large displacements and highly turbulent wind
flows, are very complex problems, a partitioned solution scheme is advantageous
because it allows to use the best-suited simulation technology for each problem.
This modular and flexible software concept for the solution of FSI problems needs
to be integrated in a shape optimization approach which leads to a Nested Anal-
ysis and Design (NAND) concept. In particular, the strong coupling between the
physical fields necessitates a coupled sensitivity analysis which is solved by a stag-
gered scheme. Finally, this integrated design process enables the simulation and
optimization of technically relevant free-form structures.

2 Design and optimization

The final goal of every structural design process is to create a structure which can
reliably withstand the external loadings, being at the same time efficient in terms of
material use. To determine the final structural layout, methods of structural opti-
mization can be applied. Typically, optimized structures are free-form shells or
membranes since they are very light with regard to their load bearing capacity. In
cases of structures interacting with surrounding fluid flows, the complete coupled
system must be considered in the optimization procedure which leads to a multi-
physics optimization approach [2, 12, 16, 25, 26]. In general, shape optimization in
the respective single fields fluid and structure [1,5,7,29,36] constitute valuable parts
and form a basis for the optimization in fluid-structure interaction problems.

2.1 Design workflow and optimization loop

From a methodological point of view, it is important to distinguish between par-

titioned (or staggered) and monolithic solution strategies for the simulation of
FSI in the context of Shape Optimization and Computational Wind Engineering 353

fluid-structure interactions and the optimization problem. In contrast to a monolithic

approach, a partitioned solution scheme in the analysis of the FSI problem sepa-
rates the two coupled physical fields and the whole system is solved by subsequent
single field solutions and appropriately tailored coupling strategies to represent
the complete system [13]. In the development of optimization strategies a similar
situation exists: In the Simultaneous Analysis and Design (SAND) approach all
physical equations are solved together with the optimization problem at once (i.e.
it is solved for the state and optimization variables at the same time), whereas in
Nested Analysis and Design (NAND), the complete Multidisciplinary Design Opti-
mization problem is decomposed into the ingredients of optimizer, state analysis
and sensitivity analysis.
In this contribution, a completely modular software framework for gradient based
shape optimization considering FSI, which guarantees the flexibility for various
technically relevant applications, is presented. The resulting workflow is depicted
in Fig. 1. Besides the links between the different modules, the corresponding vari-
ables are introduced: The general chain of shape parameterizations is visualized and
the dependency of the finite element nodal positions x from the form finding param-
eters y and the design variables which are typically the CAD parameters s can be
seen. Therefore, during the sensitivity analysis, the gradient of the objective func-
tion f with respect to the shape design variables s is computed by expanding it to
the sensitivities of the individual modules. Different possibilities of shape definition
are introduced in Sect. 2.2.

2.2 Shape representation

In optimization, a meaningful definition of design variables is of utmost importance.

In the case of shape optimization, there are several alternatives to represent the shape
changes which differ significantly in the numerical effort and the design freedom.

Fig. 1 Design optimization: workflow and respective shape variables.

354 M. Hojjat et al.

Fig. 2 Representation of shape, CAGD based (left), parameter-free based (middle) and form-
finding (right).

According to the desired task, the best suited method must be applied: in preliminary
design stages, it is typically needed to have a big design space, whereas in a nearly
finished structural design process a more detailed shape optimization with only a
very limited change in the concept is desired. This has a huge impact in the shape
description to be used and also on the technical realization. In Fig. 2 the different
possibilities are displayed.
CAGD based The current standard in structural shape optimization is the separa-
tion of the geometric model and the analysis model, which is the case in all CAGD
(Computer Aided Geometric Design) based shape optimizations. Here, the design
variables are selected geometry parameters of the CAGD-model which leads usu-
ally to a low number of optimization variables and renders this approach rather
inexpensive with regard to the optimization. The linking of the locations of the FE
nodes to the design elements is established by means of mathematical relations (e.g.
B-Splines, Bézier curves or Coons patches).
Finite element based The typical restriction in shape variations caused by the use
of CAGD is canceled by directly taking the nodal positions of the finite element
mesh as design variables. This so-called parameter-free or FE-based geometric mod-
eling approach has the maximum design freedom but is linked to a huge numerical
effort and therefore demands for suitable adjoint formulations in the sensitivity anal-
ysis. Moreover, some further numerical problems occur which demand for filtering
and regularization strategies [7].
Numerical form finding Within thin-walled structures, the ones in membrane
action are the most efficient structures which can be built. To reach this desirable
load-carrying behavior mechanical design criteria can be exploited via numeri-
cal form finding methods. As an example, the Updated Reference Strategy (URS)
[4, 5, 42] can be applied as an effective and efficient way to compute free form
shapes of membranes and (in combination with general shape optimization proce-
dures) of shells within the design process. Physically, the task of form finding is to
find an equilibrium shape for a given stress distribution and boundary conditions.
Mathematically, it represents an inverse problem which results in singular expres-
sions during the numerical solution of the governing equations for equilibrium. This
basic relationship is the principle of virtual work which is modified in the URS by
FSI in the context of Shape Optimization and Computational Wind Engineering 355

a homotopy mapping to overcome the singularity in the original and singular vir-
tual work expression (ıW D 0). The stabilizing term ıWS is formulated in terms
of PK2 stresses S rather than in Cauchy stresses  and fades out as the solution is
approached and therefore the original, unmodified solution is received. The weak
form of this stabilized form finding method which has to be discretized by finite
elements states as:

ıW D  ıW C .1  / ıWS ;

2 3 2 3
Z   Z
D  4h detF   FT W ıF dA5 C .1  / 4h .F  S/ W ıF dA5 D 0:
A A
(1)

As can be observed in this equation, the design variables under consideration are
the anisotropy and distribution of pre-stress together with the edge cable forces and
the positions of the supports. These (usually few) parameters define naturally real
free-form surfaces by enforcing the equilibrium condition which is also related to
the determination of minimal surfaces. Hence, a good reduction of design parame-
ters compared to CAGD-free methods can be realized by at the same time using a
non-parametric shape definition which in turn allows to obtain mechanically moti-
vated, real free-form shapes. Remarkably, the form finding process includes the
solution of nonlinear equations which has impact on the evaluation of sensitivities.

2.3 Mesh regularization

The quality of the solution in finite element and finite volume method strongly
depends on the domain discretization and therefore having a good mesh is strongly
desired. However, many of the mesh generation techniques are not able to provide
a discretization with good element shapes in the whole domain, especially in the
case of complex geometries. Moreover, even if the quality of the generated mesh is
acceptable, in several finite element applications which deal with varying geome-
tries, during computation, the shape of elements might get distorted. For instance,
in shape optimization problems and large deformation fluid-structure interaction
simulations the retaining of the initial discretization in not guaranteed.
In this work, a global method which regularizes the finite element mesh to a
desired condition is presented. In this method, an artificial stress is applied on the
surface or on the volume mesh and a global linear system of equilibrium equations
is solved. The applied stress adapts each element towards a desired predefined tem-
plate geometry and at the end a globally smooth mesh is achieved. In this way, both
shape and size of each element is controlled.
This method is closely related to the form finding method in terms of solving
Eq. (1). In the case of regularization, there are some additional aspects that should be
considered. Firstly, the reference geometry has to be changed such that the applied
356 M. Hojjat et al.

Fig. 3 Local refinement in case of singularity.

Fig. 4 Regularization of a curved distorted grid.

prestress brings the elements to their desired shapes. Secondly, the applied prestress
has to control the size of each element and at the end the resulting discretization has
to preserve the domain geometry.
In order to control the shape of elements their reference geometry is assumed to
be the desired one, hence the applied prestress modifies the elements towards the
given shape. When the control of the size of the elements is of interest, as it is in the
case of singularities, the applied prestress should be adapted accordingly. In Fig. 3 a
square template is used for all elements whereas the prestress is adjusted according
to the distance to the singular point.
Applying regularization on a mesh, the shape of the surface or volume has to
be preserved. For parameter-free geometries, this can be done in an approximation
manner, since the exact geometry is not known. However, in shape optimization this
approximation is of no harm, since the optimization is the driving process which
determines the final shape. In this work, the shape is preserved by allowing modifi-
cations of surface nodes only in the in-plane direction. Hence, a reduced system of
equations is solved. In Fig. 4 the distorted mesh of a curved surface on the right is
regularized while preserving the geometry.

2.4 Physical state

During optimization computation of the objective function involves the evaluation

of the actual physical state. In the present work, due to the coupling between the
FSI in the context of Shape Optimization and Computational Wind Engineering 357

structure and the fluid fields, a fluid-structure interaction analysis is required and a
three field analysis including structure, fluid and fluid mesh is performed.
In this chapter, the mesh equation is included in the fluid equation and for the
solution of the two remaining equations -fluid and structural-, a partitioned solution
strategy is applied. A detailed description of the two fields is given in the following
sections and the treatment of the interface in Sect. 2.4.3.
Within this work, the interaction between the fluid and the mesh fields is rep-
resented using an ALE approach in which the deformation of the structures wetter
surface is used to update the boundaries of the computational domain of the fluid
and the nodes of the corresponding mesh are fixed to the material points.

2.4.1 Structure - Shells and membranes

In the context of large deformations, where the geometrical nonlinearity is not

negligible, the following boundary value problem has to be considered:

@2 d
 D r.F  S/ C f in ˝S ; (2)
@t 2

with initial and boundary conditions, d .x; t/ D d .t/ ;  .x; t/ D   .t/ on  ,

where F, S and f are the material deformation gradient, the second Piola-Kirchhoff
stress tensor and the volume forces on ˝S , respectively.
The second Piola-Kirchhoff stress tensor S is given via an appropriate consti-
tutive relation, e.g. for St. Venant-Kirchhoff material with C being the constitutive
tensor and E representing the Green-Lagrange strain:

1 T
S D C W E; ED .F  F  I/: (3)
2
The computational domain is discretized and the above equations are solved with
the finite element method, in a Lagrangian description, using the isoparametric ele-
ment concept. For the time integration the generalized-˛ method is used [10] and
in the case of stationary problems the inertia term vanishes and Eq. (2) reduces to a
static equilibrium equation.
In the modeling of thin-walled structures, membrane or shell elements are used
for discretization of the finite element model. Within this work, for the shell appli-
cations a 7-parameter shell element is used. Furthermore, special care is given to
prevent locking phenomena by applying the well-known ANS and EAS methods
[3]. For membranes, the thickness is assumed relatively small and three translational
degrees of freedom are assumed per node. Shape of these structures is obtained with
form-finding, described in Sect. 2.2 [4, 42].
Both, shell and membrane elements produce reduced structural models with two
dimensional representation which can describe the three-dimensional physical prop-
erties by introducing mechanical assumptions for the thickness direction. With this
358 M. Hojjat et al.

way, more efficient computations are performed and the accuracy of the model is
significantly improved. Due to this reduced order modeling of the structure special
care has to be given to the interface and an additional operation has to be included
in the structural part to transfer information between the 2D structure and the 3D
fluid model [6].
More precisely, in the case of shells, the interface surface is found by moving
the 2 dimensional surface of the structure half of the thickness normal to the surface
in both sides. On these two moved surfaces, the exchange of data is performed
consistently with respect to shell theory: The displacements and rotation calculated
in the structure are transformed to displacements for the fluid and the normal and
shear forces from the fluid are transformed to normal forces and moments applied
on the shell. In membranes, the thickness is considered to be negligible. So, in both
fields the structure is represented with two surfaces which coincide.

2.4.2 Fluid - Atmospheric wind

In the case of structural design, since almost all civil engineering structures are
exposed to the atmospheric wind flow, a prediction of the wind loads applied on
them is needed. Specially, in the case of light-weight structure design, considera-
tion of the wind effects on the structure has a significant importance. In order to
achieve a good prediction of the wind loads, one should take into account main
characteristics of the atmospheric boundary layer (ABL). ABL has a complex form
in terms of velocity and turbulence intensity profiles, and their variations in time,
which are dependent on many parameters such as geographic location and type of
surrounding urban. A robust wind model should characterize these physical issues
in a set of few easy-to-handle model parameters. Furthermore, large structure size,
low air viscosity and high wind velocity lead to very big Reynolds numbers (several
millions) and subsequently large computation times. Therefore, computational wind
engineering (CWE) simulation has to be efficient with respect to the computation
time. In this section, governing equations and some of ABL modeling aspects with
focus on fluid-structure interaction application is presented.
State equations and turbulence models Considering physical properties of air
and the typical Mach number in wind engineering problems, simulation of wind
flow can be considered as an incompressible Newtonian problem. Therefore, one
can find state variable fields (pressure and velocity) by solving the momentum and
continuity equations (Navier-Stokes) on a computational domain with appropriate
boundary conditions. Using finite volume method, momentum equations can be dis-
cretized and linearized. In the case of transient simulation, this can be done by
using an iterative algorithm such as PISO (Pressure Implicit Split-Operation) [19].
For steady-state models, SIMPLE algorithm follows a similar approach to PISO,
neglecting the time-derivative term in the momentum equation.
Despite the fact that LES is known to be more accurate than RANS, taking into
account the required computational effort, RANS is rather a more realistic technique
FSI in the context of Shape Optimization and Computational Wind Engineering 359

for solving this specific type of FSI-CWE problems. However, reviewing the liter-
ature [24, 32] shows that a relatively high order of accuracy can be achieved using
RANS models in wind flow simulations. Particularly, since within this work, the
optimization is realized with steady-state assumption, time-averaged pressures cal-
culated by RANS are totally sufficient and quantities such as the peak pressure,
which can be captured only by more flexible models like LES are not of interest.
Reynolds averaged models are based on splitting field variables to the sum of the
mean value and the fluctuating part. This leads to an additional momentum term in
Navier-Stokes equations:
u D uN C u; (4)

@Nui @Nui 1 @pN @2 uN i @u0i u0j @uNi

C uN j D C  ; D 0: (5)
@t @xj  @xi @xk 2 @xj @xi
The aim of Reynolds averaged methods is to model (in contrast to resolve) this
momentum term. Turbulent viscosity models propose dividing it into an isotropic
part which will be added to the pressure field and an anisotropic part which is
considered as an additional viscosity, added to the molecular viscosity.

@Nui @Nui 1 @.pN C 23 ke / @  

C uN j D C 2. C T /SNij : (6)
@t @xj  @xi @xj

RANS models are classified by the number of extra transport equations that have
to be solved in order to obtain the turbulence viscosity value. Two widely-used two-
equation RANS models are k" and k!, in which T is modeled solving transport
equations of k (turbulence kinetic energy) and " (dissipation rate) for k  ", and k
and !  k" for k  !.
k  ! shows a better performance comparing to k  " in large pressure gradients
and near-wall flows [41], whereas k" gives a more realistic estimation of the turbu-
lence viscosity for free-stream flows and small pressure gradients. k! S S T model
[27] combines strengths of both models by using a linear combination of " and !
transport equations. Coefficients of this mixed transport equation are calculated as
follows:
Xk!SST D Xk! F1 C Xk .1  F1 /: (7)
Blending function (F1 ) gives the emphasis to k  ! results near the wall, and directs
the solution more to the k  " results far from the wall.
The nature of wind is time varying and a transient modeling technique is nec-
essary to simulate wind behavior. Although RANS is based on time averaging, it
is possible to use it for transient flows. The so-called URANS (Unsteady RANS)
model divides the time domain into time-steps and within each time-step, steady
RANS assumptions are applied. Time-dependent variables are estimated by com-
mon time-integration techniques. Validity of URANS and the proper choice of time
scales in it for different applications are still open questions in turbulence modeling.
Wind simulation In CWE, the geometry of the structure is built as an obstacle in
a discretized computational domain. Solving the turbulent flow around it, pressure
360 M. Hojjat et al.

values on the surface of the structure are calculated. In this type of problems, there
are some modeling aspects to be considered, few of which are as follows:
 Boundary conditions: Assuming a straight wind flow, the following boundary
conditions need to be specified. Top and side surfaces are treated as free-slip or
symmetric or periodic boundaries [17]. Outlet pressure can be set to a constant
value. The challenging part is to prescribe the fields (pressure, velocity and tur-
bulence intensity) at the inlet, as similar as possible to the real wind conditions.
Furthermore, at the bottom of the computational domain boundary, a rough wall
boundary shall be used, rather than a fully smooth no-slip wall [8]. This is due to
the natural roughness objects existing on the surface of the earth such as grass,
trees, other buildings, etc. In the following, the proper choice of inlet fields is
described considering a rough wall boundary layer.
 The domain size: Size of the computational domain should be much larger than
the structure. This prevents the disturbance of the fluid boundaries on the obsta-
cle. For different fields of application, in literature, there exist suggestions for the
proper CWE domain size. Franke et al. [14] proposes a distance of 2-8, 5, 6 and
15 (normalized by the structure size) for the front, side, top and back between
the obstacle and the computational domain, based on blockage percentage.
Atmospheric boundary layer Basically, a turbulent wall boundary layer consists
of a linear near-wall (viscous) sublayer and a logarithmic sublayer. Due to the large
Reynolds number, in ABL the viscous sublayer is very thin and therefore almost all
structures are located in the logarithmic region. The mean velocity profile in loga-
rithmic region of a smooth wall boundary is described as uN D u ln y C C B, where
u is the friction velocity,  D 0:41 is von Karm J constant, y C is the distance
J an
to the wall in inner coordinates and B D 5:2 is a universal constant. Existence
of roughness at the wall surface does not change the general form of the veloc-
ity profile, but just the multiplier of y C in the mentioned formula [37]. Therefore,
Inserting B and the effect of roughness inside the ”ln”
 expression, the velocity pro-
file in a rough boundary layer would be uN D  ln yy0 , where y0 is the roughness
u

length and is a function of viscosity, u and ks (sand roughness height). Assuming

this velocity profile, Richards and Hoxey [30] suggest the following inlet conditions
for atmospheric wind velocity, kinetic energy and dissipation rate profiles:
 
u z C z0 u 2 u 3
uN .z/ D ln ; kDp ; "D ; (8)
 z0 C k.z C z0 /

   
z
where u D  uref div ln ref z0
and C D 0:09. uN ref is a known wind velocity
at the height zref , and z0 is the surface roughness length.
Moving boundaries In case of moving boundaries, the mesh field has to be
deformed and the mesh fluxes have to be corrected, respectively. There are different
algorithms of calculating this mesh motion. For instance, H. Jassak and Z. Tukovic
[20] determine the mesh field motion by solving the Laplace equation with variable
diffusion on elements. One of the strengths of this method is that less distortion
FSI in the context of Shape Optimization and Computational Wind Engineering 361

in the mesh close to the moving boundary occurs. In the present work, this mesh
motion algorithm is applied to calculate the movement of the boundaries due to the
structural displacement in fluid structure interaction analysis as well as in update of
the design in the optimization loop.

2.4.3 Coupling

In order to establish equilibrium between the fluid and the structure field at the
common interface  F S , certain coupling conditions have to be fulfilled. These are
the continuity of displacements and surface tractions, respectively:
FS FS FS FS FS FS
dS D dF ; n
S  S D n
F  F : (9)

The equilibrium state of a coupled system is reached, if for a certain interface

FS
displacement d  the field and interface residua tend to zero:
FS FS FS
RS .t; d  / D 0; RF .t; d  / D 0; R .t; d  / D 0: (10a)

Coupled problem in operator form The solution of the coupled problem is

based on a partitioned Dirichlet-Neumann approach. This means that the fluid and
structure fields are solved as separate partitions and the interaction is captured by
influencing the specific boundary conditions. The fluid field is solved due to a given
interface displacement and is therefore the Dirichlet partition. The structure field
is solved due to given boundary forces and is therefore the Neumann partition. To
develop algorithms for the implicit solution of the coupled problem, the problem is
denoted in its operator form. Because the used algorithms rely on black-box field
solutions, the corresponding field operators are introduced. The fluid field is denoted
by the field operator F and the structure field by the field operator S . Due to
the nonlinearity of the field and the coupled problem, the solution process involves
subiterations within as well as between the fields. Typically, each iteration consists
of the following steps:
1. Apply a displacement d onto the fluid interface  . The fluid domain ˝ F is
deformed and the mesh fluxes are calculated.
2. Solve the fluid problem on the deformed state using an ALE-approach, leading
to velocity v and pressure p. Calculate the surface forces f ext at the interface:
f ext D F .d /:
3. Apply the surface forces f ext onto the structure interface. Solve the struc-
ture problem for the corresponding displacements dQ and evaluate them at the
interface: dQ  D S .f ext / D S .F .d // D S ı F .d /:
The nonlinear coupled problem is therefore given by dQ  D S ı F .d / , and its
solution is found if dQ  D d . The residuum at the interface R .d / is defined as:

R .d / D S ı F .d /  d D dQ   d : (11)

362 M. Hojjat et al.

Solution algorithms for coupled problems There exists a huge variety of algo-
rithms for the partitioned solution of the coupled problem, e.g. fixed-point [22, 43],
vector extrapolation [23, 33], quasi-Newton [11, 39], inexact Newton-Krylov [28]
or Multigrid/Multilevel methods. One of the greatest advantages of a partitioned
approach is its modularity. Depending on the objective of a coupled simulation, the
computational framework can be easily adapted for the specific needs of the single-
field solvers, what leads to a great flexibility and a broad application range. But
if a coupled solution algorithm is chosen, which necessitates complicated changes
and enhancements to the single-field solution method, this approach is not suit-
able. Therefore, within this project the emphasis is placed on methods, which rely
on black-box field solvers. There exist mainly two approaches, which fulfill this
requirement, namely fixed-point iteration in combination with Aitken’s 2 -method
[22, 43] and a quasi-Newton method with reduced order modeling of the field
behavior [11].
Fixed-point iteration: The basic form of a fixed-point method is given by:

x kC1 D ˚.x k /; k D 1; :::; n: (12)

Operator ˚ evaluates the solution of step number k C1, being given the last iteration
solution, x k . Considering a coupled problem at time n C 1, the iteration directive
has the form:  
;k ;k
˚ d nC1 D dnC1 C dnC1 : (13)

Using a relaxation technique for the update of the interface displacement leads to:

k Q ;k
d;kC1
nC1 D d;k ;k ;k k ;k k ;k
nC1 C dnC1 D dnC1 C ! RnC1 D .1  ! /dnC1 C ! dnC1

where ! k is the relaxation factor in iteration k. There exist several possibilities for
the choice of the relaxation factor. A constant factor can be used, which is simple,
but stability problems are faced for large ! k and inefficiency and small convergence
rates for small ! k . Choosing a dynamic relaxation factor in every iteration gives
improvements w.r.t. efficiency and stability. Therefore, Aitken’s 2 -method is used
to determine the relaxation factor. As a basis, the Aitken factor has to be determined
by:
 T
;k1
  dnC1  d;knC1  d;k
nC1
nC1 D nC1 C nC1  1
k k1 k1
 2 (14)
;k1
dnC1  d;k
nC1

with
;k1
dnC1 D dQ nC1
;k1 ;k1
 dnC1 ; d;k Q ;k ;k
nC1 D dnC1  dnC1

and the corresponding relaxation parameter is given by ! k D 1  knC1 .

FSI in the context of Shape Optimization and Computational Wind Engineering 363

Quasi-Newton method: The method used within this approach is called a quasi-
Newton method, because it is similar to the classical Newton approach in the sense
of using history information of the interface residual to minimize the interface resid-
ual and determine incremental interface updates, but the Jacobian is not directly
calculated within the solution process. Basically, the method consists of two main
parts within every iteration: 1. Approximation of the new interface residual based
on a linear combination of the previous residuum increments. This leads to a set
of coefficient factors. 2. Determination of the new interface displacement incre-
ment as a linear combination of those known increments, applying the coefficients
determined in the first step.
The algorithm needs a starting procedure, to determine the first residuum and
displacement increments. Therefore, a predictor and a relaxation step have to be
performed in advance. After this, the quasi-Newton iterations are entered. In every
step, the desired residuum of the next iteration is zero, so the desired residuum incre-
ment is given by Rk D 0  R;k . This is approximated by a linear combination
of the known residuum increments from previous iterations, leading to:

X
k1
Rk D 0  R;k  ˛ik R;i D V k ˛ k ; (15)
i D1

which gives the linear coefficients ˛ik and V k being a matrix that stores the residuum
increments from past steps. This is in general an  overdetermined equation system,
which is solved using a least squares method: V k ˛ k  R;k  ! mi n. V k is a
non-symmetric matrix and so it can be decomposed into an orthogonal Q k and a
upper-triangular matrix R k using e.g. Householder transformation. This gives great
numerical benefit in the solution of the problem.
The coefficients, determined for the combination of the residuum increments, are
now applied to the corresponding relaxed interface displacements:

X
k1
dQ ;k D ˛ik dQ ;i D W k ˛ k : (16)
i D1

The correlation between d and dQ  is given by R D dQ   d , and therefore:

d D dQ   R . For the corresponding incremental state, this leads to

d D dQ   R D W k ˛ k C R;k ; (17)

as the new interface displacement increment.

A detailed description of the algorithm is presented in algorithm 1.
364 M. Hojjat et al.

Algorithm 1 Quasi-Newton coupling algorithm [11]

1: for t D 0 to t D ttot do
2: k=1
3: Predictor step:
;1
4: dnC1 D 5=2dn  2dn1 C 1=2dn2
5: dQ ;1
nC1 D S ı F .dnC1 /
;1

6: R;1 Q ;1
nC1 D dnC1  dnC1
;1

7: relaxation step:
8: d;2 ;1
nC1 D dnC1 C !RnC1
;1

9: quasi-Newton iterations:
10: while (!converged) do
dQ nC1 D S ı F .dnC1 /
;k ;k
11:
RnC1 D dQ nC1  dnC1
;k ;k ;k
12:
ˇ ˇ
ˇ ;k ˇ
13: if ˇRnC1 ˇ <
then
L2
14: dnC1 D d;k
nC1
15: converged
16: else
17: R;k D R;k ;k1
nC1  RnC1
18: dQ D dQ nC1  dQ nC1
;k ;k ;k1

19: V k D R;k R;k1 : : : R;1

20: W k D dQ ;k dQ ;k1 : : : dQ ;1

21: Householder transformation: V k D Qk R k
T  
22: Rk ˛ k D Qk R;k ! ˛k
23: d D W k ˛ k C R;k
24: d;kC1 ;k
nC1 D dnC1 C d


25: end if
26: k k C1
27: end while
28: n nC1
29: end for

2.5 Optimization

In the optimization module, decisions about the update of the design as well as the
convergence of the overall procedure are taken. In general, the optimization problem
can be stated as follows:

f .s/ ! min : gj .s/  0 , j 2 Œ1; m and hj .s/ D 0 , j 2 Œm C 1; m C N ; (18)

where sl  sl  sl , s 2 Rn and f , gj , hj are the objective function, the

inequality and the equality constraints, respectively. They are all functions of s D
.s1 ;    ; sn /T , which is the vector of the optimization variables. The existence of
constraints changes totally the way to treat the problem. Consequently, constrained
and unconstrained problems form two separate classes of optimization problems.
Unconstrained gradient based optimization problems, in which the gradient of the
objective function rf is required, are discussed in the sequel.
FSI in the context of Shape Optimization and Computational Wind Engineering 365

Structure Coupling Fluid

Update design
Update variable Si Update design
Ψ(s, u, v) min

Si

Struct. Analysis u, v Fluid analysis

State S(s, u, v) = 0 F(s, u, v) = 0

si , ui , vi

Sensitivity analysis du dv Fluid Sensitivity

Sensitivity d Ψ dS dF ds ds analysis
ds ds ds

Fig. 5 Three levels of optimization using NAND approach.

Within this project, the physical state (Sect. 2.4) involves a fluid-structure inter-
action analysis and the final goal will be to obtain a design which improves the
behavior of the structure. The objective function f will be a function of the
structural displacement u, the fluid state variable v and the design variable s, i.e.
f D f .u; v; s/. Design variables of structural shape optimization are the geometric
parameters representing the shape of the structure.
Furthermore, the physical state is described by the structural and the fluid par-
tial differential equations which have to be satisfied at the optimum. As a result, in
the optimization problem, two PDE constraints have to be added. These constraints
are not treated as any other constraint, as in the case of Simultaneous Analysis and
Design (SAND) but their first order consistency is enforced in each iteration step.
This strategy of implementing the optimization problem is known as Nested Anal-
ysis and Design (NAND) [1]. For the following, let us assume that the discretized
equations describing the two fields are S.s; u; v/, F.s; u; v/, respectively.
NAND consists of three main stages: the optimizer, the analysis of the physical
state and the sensitivity analysis (Fig. 5). The analysis of the physical state is already
described in Sect. 2.4 whereas the optimizer and the sensitivity analysis are going
to be discussed in the sequence. Furthermore, within this work, the physical state is
a steady state fluid-structure interaction analysis and thus time is omitted from both
fluid and structural equation.

2.5.1 Optimizer

The optimizer is the driving module of the overall optimization procedure. Here, the
decision about the size and the direction of the update vector of the structural design
is taken based on the gradient information obtained by the sensitivity analysis.
In general, gradient-based methods start with an initial guess for the design
variable s0 and in each iteration the solution is updated as follows:

si C1 D si C ˛i Di ; (19)
366 M. Hojjat et al.

where Di D Di .rf .si /I r 2 f .si // 2 Rn is the search direction and ˛i 2 R is the

step size which is usually calculated with a line-search algorithm. As for the most of
engineering application the Hessian matrix is not availabe, simple gradient methods
prove to be more robust and reliable alternatives, in particular for a large number of
design variables. Standard methods are “Steepest descent” and “Conjugate gradient”
for unconstrained and the method of “Feasible directions” for constrained problems.
In the steepest descent algorithm, the steepest descent direction (the negative gra-
dient), is assumed as the search direction, therefore, D D rf .s/. This method
specifically is very efficient when the initial guess is comparatively far from the
solution but closer to the minimum it converges rather slowly [15]. The description
of the algorithm is presented in algorithm 2.

Algorithm 2 Steepest descent algorithm

1: Initialize i D 0 W s0
2: while (!converged) do
3: D i D rsi f
4: siC1 D si C ˛i Di
5: i i C1
6: end while

Conjugate gradient is a method which converges faster than steepest descent,

without increasing complexity and computational cost. In this method, the search
direction is:

rsi f T  rsi f
Di D rsi f C ˇi  Di 1 , where ˇi D : (20)
rsi 1 f T  rsi 1 f

In the case that the objective function is quadratic, the method converges in max-
imum n steps, where n is the number of design variables. This is not the case for
non-quadratic functions. For such functions, intelligent techniques which take into
account the nonlinearity of the function in order to decide about the restarting of the
algorithm are applied [15].

2.5.2 Sensitivity analysis

During sensitivity analysis the gradient of the objective function with respect to the
design variables has to be computed. As already described in Sect. 2.5, the objective
function is assumed to be f D f.s; u; v/, where u and v represent the structural
displacements and the fluid forces, respectively.
Applying the chain rule, the gradient of the objective function can be written as:

df @f @f d u @f d v
D C C : (21)
ds @s @u d s @v d s
FSI in the context of Shape Optimization and Computational Wind Engineering 367

In order to compute the gradient of f from the above formula, all the terms in the
right hand side have to be computed. The derivatives of the objective function with
respect to the state variables ( @f , @f ) can be easily computed since the function f
@u @v
is usually given in analytical form. The derivatives of the state variables with respect
to the design variable s are computed from the sensitivity equations. To obtain this
set of equations, first-order consistence of the PDE of the fluid and the structure has
to be enforced:

dS @S @S d u @S d v
D C C D 0;
ds @s @u d s @v d s (22)
dF @F @F d u @F d v
D C C D 0:
ds @s @u d s @v d s
The above equations form a linear system of equations with derivatives of the
state variables with respect to the design variables as the unknowns:
0 1 0 1 0 1
@S @S du @S
B @u C B C B  C
B @v C B d s C B @s C:
@ @F  D (23)
@F A @ d v A @ @F A

@u @v ds @s

The system described by Eq. (23) is solved in an iterative manner, described in

Sect. 3.2, and the solution procedure is similar to a fluid structure interaction
analysis [2, 35].

3 Solution Strategy and Realization

In this section the strategy followed for the realization of the overall framework is
presented. In 3.1 the software environment is introduced and in 3.2 the algorith-
mic implementation of the shape optimization technique within this environment is
explained.

3.1 Computational Framework

A robust computational framework should be able to treat a broad range of appli-

cations, from small-sized principle to large industrial examples, from laminar to
turbulent, from pure FSI-simulations to optimization of coupled problems. There-
fore, a highly modular environment is needed, so that for each specific problem the
best-suited approach is possible. To fulfill these requirements, a software realization
based on three different codes is chosen. Two of these codes are the single-field
solvers for fluid and structure. The third code, called coupling code, is located
368 M. Hojjat et al.

between the field solvers and operates as a master process, controlling the whole
simulation, and working as an interface between fluid and structure field. It has to
be noted that all codes within this environment are capable of performing parallel
computations and completely object-oriented. These three software are:
Carat: The in-house structure field solver CARAT++ is based on Finite Element
method. Different sorts of analysis, such as static and dynamic, linear and nonlin-
ear are available in this code. For the dynamic analysis, different time-integration
schemes such as Generalzed-˛ are used.
OpenFOAM: For the fluid field, a solver based on OpenFOAM is used. Open-
FOAM is a freely available set of object-oriented libraries for Finite Volume method.
Various numerical schemes for interpolation and integration are available.
CoMA: CoMA (Coupling for Multiphysics Analysis) is an independent software
developed specifically for the scope of this work. This program is responsible for
the simulation control of surface-coupled simulations and is located as a central
process between single-field solvers. It offers a wide variety of features, making
it a powerful tool for all kind of surface coupled simulations. The most important
features of CoMA are:
 Parallelization and communication: In coupled computations, the field solvers
typically work in parallel due to high computational effort. This means that the
coupled surface of one field is spread over multiple field processes. To handle
this, CoMA uses the so-called process group concept. The subset of all field
processes which take part in the coupled computation is identified and grouped
into a process group. The communication with processes can be based on MPI
or files, depending on the possibilities offered by the field solvers.
 Data transfer: One of the key ingredients of CoMA is to handle data transfer
between non-matching surface meshes. For flexibility and due to the paralleliza-
tion of the field solvers, the surface mesh consists of different partitions, each
of which is connected to a process, with the possibility of multiple partitions
per process. Both flux and field quantities can be transferred. The data trans-
fer is based on a nearest-neighbor search and linear interpolation. This transfer
method is conservative with respect to field quantity fluxes. More details can be
found in [21].
 Coupled simulation control: CoMA serves as a master process controlling
black-box single field solvers in coupled computations. It offers the possibil-
ity of different coupling concepts (one-way and two-way coupling, explicit and
implicit coupling) and for two-way coupling different coupling algorithms. The
key ingredient is the so-called exchange point concept. Exchange point is an
entity, which transfers a quantity, e.g. force, from a sender process group to
a receiver process group. The exchange point can have additional features to
influence and adapt the quantity. This can be e.g. a coupling algorithm apply-
ing underrelaxation or a convergence proof. Exchange points are arranged into
an exchange point group, what resembles one cycle/iteration within a coupled
simulation. The chosen coupling algorithm influences the way, how the group is
FSI in the context of Shape Optimization and Computational Wind Engineering 369

called, i.e. once per timestep in explicit coupling, multiple times in implicit cou-
pling. This concept offers great flexibility and also new coupling strategies can
be easily realized.

3.2 Optimization strategy

As described in Sects. 2.1 and 2.5, the steady-state shape optimization method pre-
sented in this work consists of three main steps, the geometry update, physical
state evaluation and the sensitivity analysis. Accordingly, in software implemen-
tation phase, these steps are introduced to all three modules of the coupled problem,
within a loop and in the same order. Therefore, in each design step, all three codes
go through this loop and perform these steps one after each other simultaneously. In
this section, the structure of the design loop as well as the communication and syn-
chronization of field solvers within the optimization procedure is briefly introduced,
which is indeed the challenge in software development of such complex class of
analysis.
Optimization loop Each design step starts by updating the shape with the new
set of design variables proposed by the optimizer. After performing the structural
geometry update, the new geometry is sent to the fluid solver via the coupling code.
This new geometry is applied to the fluid by moving its boundaries toward the new
design and after that the CFD solver computes a converged solution for the new
boundary. Then the fluid-structure interaction analysis is performed by exchanging
displacement-pressure values at the interface between the fields until convergence.
Resulting pressures and displacements are stored as the “state-variables” of the cur-
rent time-step, i. Afterwards, the sensitivity analysis is performed for all the design
variables (2.5.2). Within the sensitivity loop and after evaluation of sensitivity for
each design variable, it is necessary to reset all the fields back to their state values.
The result of this step is the gradient vector of the objective function with respect to
all design variables, which is subsequently given to the optimizer for calculation of
the new set of design variables. This is done according to the particular optimization
algorithm used in the problem (see Sect. 2.5.1).
The optimization loop is repeated unless the optimizer reports the design con-
vergence. This could be due to reaching a minimum in the objective function or
because some technical constraint at the design variables is reached.
Evaluation of sensitivities The aim is to solve Eq. (23) for ddus and dd vs which is
then used in Eq. (21) in order to calculate df
ds
. Direct solution of this system of equa-
tion is possible just if both fluid and structure derivatives of Eq. (23) are explicitly
known. In this project, structural derivatives ( @S@u
and @S
@s
) are calculated directly at
the structural code, Carat, whereas the fluid solver is treated as a black-box module.
Therefore, an iterative mixed strategy is applied, which uses directly the sensitivity
equation of the structure and finite-difference evaluation of derivatives for the fluid.
The main steps of this method are explained in the following.
370 M. Hojjat et al.

At the beginning of each design variable sensitivity analysis, a perturbation of that

particular design variable is applied on the design. Then, the coupled sensitivity
loop starts. In this loop, first, the structure sensitivity equation is solved using the
fluid-derivative information from the previous iteration:
   1      
du @S @S dv @S
D C : (24)
ds i @u i @v i ds i 1 @s i

Then, using the resulting structure-derivatives, ddus , a new perturbation on the fluid
is calculated:
 
du
ui D u0 C s D u0 C u; (25)
ds i
where u0 is the state displacement. Response of the fluid to the new perturbation
could be written as:
vi D v.u0 C u; s0 C s/: (26)
s0 is the state design variable and s is the perturbation of it. This response (vi ) is
used to produce the fluid derivative, dd vs by finite differencing:
 
dv vi  v0
 : (27)
ds i s

v0 is the state force vector. These steps are repeated until convergence of sensitivity
values.
This method is more efficient than external finite difference evaluation of the sen-
sitivity on the coupled problem, since the structural equation is reduced to a linear
equation. It should also be mentioned that unlike the evaluation of the state vari-
ables, a high accuracy for the calculation of sensitivities is not required since their
values are used only for trend information. Therefore, a lower order approximation
of the sensitivities is sufficient to perform gradient-based optimization. Thus, by
setting a larger tolerance for the sensitivity analysis in this method, one can reduce
the computational effort.

4 Verification and Validation

In this section, the solution technique introduced in Sect. 3 is examined by two

benchmark examples. The first one, which is a pure CWE problem, is used for
validation of the wind simulation strategy. The second example investigates the cor-
rectness of the whole FSI solution with the focus on the coupling algorithms and
their computational efficiency.
FSI in the context of Shape Optimization and Computational Wind Engineering 371

4.1 Numerical simulation of the Silsoe cube

The “Silsoe” cube is a 6m  6m  6m cube constructed by Silsoe research insti-

tute in England [31]. The cube is equipped with different sensors for measurements
of wind engineering experiments. The flat field around the cube allows the wind to
have a close form to the ideal atmospheric boundary layer. Both size of the cube and
the wind velocity gives to this experiment the characteristics of a real wind engineer-
ing problem. Since accurate measured data from the Silsoe experiment is published,
numerical simulation of this experiment is chosen as a “reference” problem for val-
idation of the wind simulation technique. Furthermore, different modeling aspects
are studied based on this problem. For instance, it is observed that k  ! SST is
a more proper model than other common RANS turbulence viscosity models w.r.t.
accuracy. Furthermore, similarity of the Silsoe experiment ABL data to the inlet
consitions for velocity, k and " confirms the validity of those logarithmic profiles.
Results In Fig. 6 the pressure coefficient on the shown paths of the cube can
be seen. These curves correspond to the Silsoe cube experiment and wind tunnel
results reported by Hoelscher and Niemann from Windtechnologische Gesellschaft
[18], and Castro and Robins [9] and this project. Diversity of the curves shows the
complexity of wind engineering simulations. However, it can be seen that in both
diagrams, almost on the entire path, the numerical results of this work have the clos-
est value to the Silsoe experiment. It is seen that the obtained pressure coefficients
on cube’s surfaces are also very close to the values provided in known literature,
e.g. ”Wind Effects on Structures” [34].
The Silsoe cube problem was solved also using unstructured tetrahedral mesh
(instead of cartesian structured mesh) and the difference in the solution was neg-
ligible. This shows that this modeling approach is applicable for more complex
geometries which cannot be meshed with structured grids. Furthermore, some
robustness studies on different modeling parameters such as the mesh size, the
time step, Reynolds number etc. were performed. The mentioned achieved accu-
rate results for the wind loads on structures is a required basis for performing wind
engineering FSI simulations.

Fig. 6 Pressure Coefficient values on centerlines of the cube [9, 18, 31].
372 M. Hojjat et al.

4.2 Coupled Computation in laminar Regime

In this example, a coupled computation is carried out, which scope is twofold: First,
it is used to validate the coupled algorithms within the developed environment.
Second, the efficiency of different coupling algorithms is examined. The testcases
developed within the DFG research group 493 by [38] are chosen, because several
groups have performed those simulations and therefore a broad range of verified
results exists. Additionally, the testcases cover a broad range of coupling.
The testcase is described as follows: A fixed, rigid cylinder is placed slightly
unsymmetric into a wall-bounded channel. An elastic beam is attached at the back
of the cylinder. The geometric representation of the computational domain and a
full description of the problem setup can be found in [38]. If a fluid flow is imposed
onto the system, the flow regime behind the cylinder influences the beam and leads
to deformations. Up to a certain Reynolds number, the flow is steady. In larger values
of the Reynolds number, the flow gets unsteady and Karmann vortices occur behind
the cylinder.
The cantilever beam is modeled using 30 4-node shell elements. The mid-surface
of the shell is located in the transverse direction of the domain. The geometric rep-
resentation of the interface between fluid and structure domain is evaluated by the
method described in Sect. 2.4.1. The fluid is modeled based on a block-structured
grid with 54000 elements. The values taken as reference are the results published
in [38]. Different mesh levels were examined on both fields. The results shown in
this contribution are the ones giving the best compromise between accuracy and
numerical effort.
Results Three different cases are examined, called FSI1 to FSI3, in which the inlet
velocity, the structure density and the elastic modulus are varied. FSI1 results in a
steady, whereas FSI2 and FSI3 in unsteady solutions, showing periodic oscillations
of the beam with differing frequency and amplitude. The results are given in Table 1,
in which the x- and y-displacement of point A located at the beam’s tip are given.
The numbers in brackets denote the frequency of the oscillations in the case of
transient behavior.
The difference in displacements is below 5:5%. Considering the results from
other groups differences with similar magnitude occur. Therefore, the verification
of the coupled computation is done.

Table 1 Benchmark results FSI1-3.

FSI1 FSI2 FSI3
ux .A/ uy .A/ ux .A/ uy .A/ ux .A/ uy .A/
own 2:264e 05 8:280e 04 1:489e02 ˙ 1:30e 03 ˙ 2:95e 03 ˙ 1:55e03 ˙
1:24e 02 Œ3:16 8:40e 02 Œ1:90 2:84e 03 Œ11:24 3:66e02 Œ5:49
reference 2:270e 05 8:208e 04 1:485e02 ˙ 1:30e 03 ˙ 2:88e 03 ˙ 1:47e03 ˙
1:27e 02 Œ3:86 8:17e 02 Œ1:93 2:72e 03 Œ10:93 3:49e 02 Œ5:46
difference C0:00% 0:88% 0:27%C C0:00% 2:43%C C5:44%
2:42%ŒC18:13% 2:82%ŒC1:55% 4:41%Œ2:83% 4:87%Œ0:55%
FSI in the context of Shape Optimization and Computational Wind Engineering 373

Table 2 Coupling algorithm comparisons - FSI2-3.

Type FSI2 FSI3
num average iterations reduction num average iterations reduction %
constant 35 C446:9% no convergence ––
Aitken 6:4 0:0% 11:2 0:0%
QNR 4:68 26:9% 8:54 23:8%
QNR-H 3:14 50:9% 5:34 52:3%
QNR-HRC 3:26 49:0% 5:06 54:8%

Coupling algorithm efficiency In this section, different coupling algorithms intro-

duced in Sect. 2.4.3, are compared with respect to their efficiency. As an example,
the cases FSI2 and FSI3 (4.2) are selected, because they cover different density
ratios and therefore different coupling strengths. The first 100 timesteps are exam-
ined and an average is made over the number of subiterations per timestep. A
Newmark type displacement predictor is used giving 2nd order accuracy. As a basis,
the fixed-point iteration in combination with Aitken’s 2 -method is used. It is com-
pared to the quasi-Newton method in 3 different variants: The basic version (QNR),
the version taking also history incremental values into account (QNR-H), and the
version taking not only history effects into account, but also replacing the constant
relaxation step in every timestep (QNR-HRC). For sake of completeness, also con-
stant relaxation method is taken into comparison, showing the inefficiency of this
approach. The relaxation parameter is set to 0:2. The results for FSI2 with density
ratio of FS D 10 and for FSI3 with FS D 1 are given in Table 2.
The difference in performance of the mentioned coupling algorithms is clearly
noticeable in these tables. As expected, the constant relaxation is extremely ineffi-
cient and its computation time is more than 5 times longer compared to Aitken’s
method. Aitken’s method performs well in both cases. The results of the quasi-
Newton method with reduced order modeling are very encouraging. Its basic version
(QNR) gives a speed-up of around 25 % compared to Aitken’s method. But a
real boost is achieved by taking history information into account, where the effi-
ciency can be doubled, and a speed-up of over 50 % compared to Aitken’s method
is reached. Replacing the constant relaxation step by a quasi-Newton step (QNR-
HRC), just gives minimal speed-up. It also has to be noted, that the speed-up
of the QNR and its derivatives is not only w.r.t. subiterations, but with nearly
the same factor in computing time. In comparison with fixed-point methods QNR
needs the additional solution of a minimization problem with number of unknowns
equal to the number of subiterations plus the number of history steps. These are
approximately 5  10 unknowns. This is absolutely negligible in comparison to
the unknowns in fluid and structure field. As summary, the quasi-Newton method
including history information has a significant performance improvement compared
to Aitken’s method and serves as an appropriate coupling algorithm for the solution
of strongly coupled problems.
374 M. Hojjat et al.

5 Numerical example - Shape optimization with FSI

In this section, practical application of the general optimization approach described

in Sect. 2 is reviewed by the help of a real size engineering example which combines
and relates all the several points described in the previous sections. More precisely,
the behavior of a 5m10m light-slender membrane roof surrounded by the turbulent
wind flow is optimized. Large deformations on the structure are expected and a
strong coupling treatment between the fields is required.
In this example, the description of the structural shape is a combination of the
parametric and non-parametric shape representations. The structure consists of a
membrane part which is described with form-finding and two rigid supporting
frames which are described with NURBS curves. The shape of these frames is
decisive for the final shape of the overall structure.

5.1 Structural design chain

The main body of the structure is a membrane supported by the rigid frames on sides
and by cables at the edges of the structure (Fig. 7). The shape of the membrane
is derived from a form-finding procedure with 4=3 prestress ratio in the principal
directions.
First, the initial shape is designed by “Rhinoceros 4.0” which is a CAD soft-
ware specialized for NURBS curves and surfaces and related operators. For the
membrane part, the Rhinoceros plug-in “Rhino-membrane” (developed based on
the URS method explained in Sect. 2.2) is used, which finds the membrane geom-
etry by applying form-finding. Rhino and similar software are frequently used by
architects for structural design. This initial design is exported to Carat. In Carat,
applying form-finding on the structure with prescribed prestresses in the membrane

Fig. 7 Geometry of the membrane structure and the wind direction.

FSI in the context of Shape Optimization and Computational Wind Engineering 375

Fig. 8 The design process.

and supporting cables, the initial equilibrium state is found. The geometry of the
membrane is sent to ICEM-CFD (Fig. 8).
Design update At the beginning of each design step, having the updated design
variables (here, the NURBS parameters) the new curve for the supporting frame is
calculated. The nodes on the membrane which are fixed by the frame are mapped to
the new curve and then assuming the new given boundaries the updated geometry
of the membrane is calculated by form-finding (Fig. 9).
Considering the mentioned steps, the position of finite element nodes on the
structure is a function of the design variable s through the following function chain:

x D x.y.s//; (28)

where the y operator returns the geometry of the supporting frame, being given the
design variables, and x, the form finding operator, gives the nodal position of the
surface mesh due the given boundary. As explained in 2.3, after few design steps
the mesh looses its initial quality, specially in the first and the last rows of mem-
brane elements. Applying “mesh regularization” after each update step remedies
this problem and keeps the elements in their best possible shapes (Fig. 10).

5.2 FSI Analysis

Since this optimization example is a steady-state (and not a transient) problem,

the physical state is solved also with stationary fluid and static structure assump-
tions. The fluid-structure interaction is performed over a pseudo time with arbitrary
time steps. In each step the load applied by the fluid to the structure is divided to a
376 M. Hojjat et al.

Fig. 9 Updating strategy.

Fig. 10 Regularization of the membrane mesh.

certain number of load steps for the structural static-nonlinear analysis. Similarly,
the structural displacement is given to the fluid within some fluid sub-steps. Addi-
tionally, after applying the full displacement on the fluid domain, CFD computation
is continued for some more sub-steps until the fluid fields are completely converged.
Around the tent, the fluid field has a much finer mesh comparing to the structure.
The front edge of the tent is locally refined in order the high gradient flow fields
around the separation point to be captured (Fig. 11). The turbulence model, the
computational domain and boundary conditions are as described in Sect. (2.4.2).
The wind velocity at the height of the tent has an approximate value of 8 ms . The
wind velocity profile as well as the mesh are shown in Fig. 12.
FSI in the context of Shape Optimization and Computational Wind Engineering 377

Fig. 11 Refinement at the leading edge.

Fig. 12 Computational wind engineering domain and the inlet velocity profile.

5.3 Optimization

The algorithm introduced in Sect. (3.2) is applied to this example. As in many engi-
neering design problems, a multi-objective target function is chosen, i.e. a linear
combination of lift and drag forces applied to the whole structure. Since the entire
shape is controlled by the NURBS parameters of the frames, the same parame-
ters are selected as design variables of the optimization problem. Therefore, the
optimization algorithm minimizes the overall force applied on the structure, by mod-
ifying the postion (and the weight) of the control points of the NURBS curve of
378 M. Hojjat et al.

the supporting frames. Due to the strong coupling and relatively complex doubled-
curved shape of the structure, it is not possible to predict the final optimized shape
and hence this example is definitely not trivial.
Sensitivity analysis According to Eq. (28), since the objective function,f , is
directly a function of nodal postions, its sensitivity w.r.t. the design variables can
@x @y
be obtained by applying the chain rule, @f @s
D @f@x @y @s
(Fig. 13).
The size of the perturbation for the sensitivity analysis shall be small enough due
to the linearity assumptions, but also large enough compared to the element size
scale, to be captured in the discretized solution. For this example, the calculated
sensitivities with the method described in Sect. (3.2) were compared to external
finite difference values, for the sake of verification. Having the same level of accu-
racy, the suggested method of this work shows a significant improvement in the
computational time, compared to external finite differences.
Results The optimization was performed with constant step size, based on steepest
decent algorithm and the computation was carried out until convergence. Figure 14
shows the evolution of the objective values toward the final design. An improvement
of 23% in the design has been achieved which is certainly a large amount in the field
of structural design.

Fig. 13 Distribution of pressure (left) and sensitivity of pressure to one of design variables (right).

700

650

600
Objective

550

500

450

400
1 2 3 4 5 6 7 8
Design step

Fig. 14 Evaluation of the objective function.

FSI in the context of Shape Optimization and Computational Wind Engineering 379

6 Summary and conclusion

An integrated concept for shape optimal design of light-weight and thin-walled

structures considering fluid-induced effects was developed. The Nested Analysis
and Design approach (i.e. separate optimizer, state analysis and sensitivity analysis)
was followed and a partitioned FSI simulation for the state analysis was embed-
ded. The gained modularity allows for the adaption of the single ingredients to a
wide spectrum of applications by choosing appropriate coupling algorithms for the
solution of the coupled problem and the sensitivity analysis as well as different
strategies to describe the shapes to be optimized. Furthermore, the non-matching
grid capability at the coupling interface, which enables the use of problem-specific
discretizations in each field, guarantees flexibility to several technical applications.
Within this contribution, the focus was on problems of aeroelasticity in the field
of wind engineering. To ensure reliable results, investigations on the correct mod-
eling of single fields and the coupling as well as goal-oriented benchmarking were
carried out: In the case of Computational Wind Engineering (CWE) the validation
was done with the Silsoe cube example and for verification of all numerical simula-
tions of single fields and in particular the coupled problem, a testcase was developed
together with the partners within the DFG research group 493 [38].
With respect to shape optimal design, various methods for the shape definition
and the design modifications during the optimization process were introduced and
their respective numerical effort and achievable design freedom were pointed out.
Moreover, special emphasis was put on the appropriate combination of different
approaches for shape description and determination and to establish the resulting
closed design cycle. The complete optimization workflow incorporates their con-
nection and the influence on the sensitivity analysis. For the sensitivity analysis of
the coupled system a newly developed iterative strategy to compute the coupled sen-
sitivities is applied. To maintain the mesh quality even for large design updates, a
new regularization technique based on the principles of numerical form finding was
developed and applied.
Finally, the success of the overall solution and optimization strategy was demon-
strated with an example of a hybrid structure consisting of supporting frames
described by NURBS (as typically used in CAD systems) and a connected mem-
brane roof which was treated by the URS as a stabilized scheme for numerical form
finding. The design of this lightweight free-form structure which is subject to tur-
bulent wind flow was analyzed and improved by the developed and implemented
methodology.
As an outlook, the presented modular multidisciplinary optimization framework
can be enhanced and augmented by more complex methods in the single parts, i.e.
the optimizer, the coupled state and sensitivity analysis. These modifications depend
on the type of problem to be solved: e.g. in case of shape optimization using an FE-
based shape description, an adjoint coupled sensitivity analysis would be beneficial
and for problems in Computational Wind Engineering, a still refined modeling of
the surrounding wind field is important. Moreover, a continued benchmarking for
the different applications is necessary.
380 M. Hojjat et al.

Acknowledgements The present research project is part of the DFG Forschergruppe 493 -
Fluid-Struktur-Wechselwirkung: Modellierung, Simulation, Optimierung. The authors gratefully
acknowledge the financial support for their research work through the German Science Foundation
(DFG) - Germany.

References

1. Arora JS (2007) Elements of structural optimization World Scientific Publishing Company

2. Barcelos M, Maute K (2008) Aeroelastic design optimization for laminar and turbulent flows.
Computer Methods in Applied Mechanics and Engineering 197:1813–1832
3. Bischoff M, Wall WA, Bletzinger K-U, and Ramm E (2004) Models and finite elements
for thin-walled structures. In: Erwin Stein, De Borst R, Hughes TJR (ed) Encyclopedia of
Computational Mechanics, pages 59–138, John Wiley & Sons, Ltd., Chichester
4. Bletzinger K-U, Ramm E (1999) A general finite element approach to the form finding of ten-
sile structures by the updated reference strategy. International Journal of Space and Structures
14(2):131–145
5. Bletzinger K-U, Wüchner R, Daoud F, Camprubi N (2005) Computational methods for form
finding and optimization of shells and membranes. Computer Methods in Applied Mechanics
and Engineering 194:3438–3452
6. Bletzinger K-U, Wüchner R, Kupzok A (2006) Algorithmic treatment of shells and free form-
membranes in FSI In: Bungartz HJ, Schäfer M (ed) Fluid-Structure Interaction: Modelling,
Simulation, Optimization pages 336–355, Springer
7. Bletzinger K-U, Firl M, Linhard J, Wüchner R (2008) Optimal shapes of mechan-
ically motivated surfaces. Computer Methods in Applied Mechanics and Engineering,
doi:10.1016/j.cma.2008.09.009
8. Blocken B, Stathopoulos T, Carmeliet J (2007) CFD simulation of the atmospheric boundary
layer: wall function problems. Atmospheric Environment 41:238–252
9. Castro IP, Robins AG (1977) The flow around a surface-mounted cube in uniform, turbulent
streams. J. Fluid Mech. 79(2):307
10. Chung J, Hulbert GM (1993) A time integration algorithm for structural dynamics with
improved numerical dissipation: the generalized-˛ method. Journal of Applied Mechanics
60:371–375
11. Degroote J, Bathe K-J and Vierendeels J (2009) Performance of a new partitioned procedure
versus a monolithic procedure in fluid-structure interaction. Computers & Structures 87:793–
801
12. Etienne S, Pelletier D (2005) A general approach to sensitivity analysis of fluid-structure
interactions. Journal of Fluids and Structures 21:169–186
13. Felippa CA, Park KC, Farhat C (2001) Partitioned analysis of coupled mechanical systems.
Computer Methods in Applied Mechanics and Engineering 190:3247–3270
14. Franke J, Hellsten A, Schünzen H, Carissimo B (2007) Best practice guideline for the
CFD simulation of flows in the urban environment. Cost Action 732, Quality assurance and
improvement of microscale meteorological models.
15. Haftka RT, Gürdal Z (1992) Elements of structural optimization. Springer, Dordrecht
16. Haftka RT, Sobieszczanski-Sobieski J, Padula SL (1992) On options for interdisciplinary
analysis and design optimization. Structural Optimization 4:65–74
17. Hargreaves DM, Wright NG (2007) On the use of the k " model in commercial CFD software
to model the neutral atmospheric boundary layer. Journal of Wind Engineering and Industrial
Aerodynamics 95:355–369
18. Hölscher N, Niemann HJ (1998) Towards quality assurance for wind tunnel tests: A compara-
tive testing program of the Windtechnologische Gesellschaft. Journal of Wind Engineering and
Industrial Aerodynamics 74-76:599–608
FSI in the context of Shape Optimization and Computational Wind Engineering 381

19. Issa RI (1985) Solution of the implicitly discretised fluid flow equations by operator-splitting.
Journal of Computational Physics 62:40–65
20. Jasak H, Tukovic Z (2007) Automatic mesh motion for the unstructured finite volume method.
Transactions of FAMENA 30:1–18
21. Kupzok A. (2009) Modeling the interaction of wind and membrane structures by numerical
simulation. Ph.D. Thesis, Lehrstuhl für Statik der TU München, München
22. Küttler U, Wall W A (2008) Fixed-point fluid-structure interaction solvers with dynamic
relaxation. Computational Mechanics 43(1):61–72
23. Küttler U. and Wall W. A. (2009) Vector extrapolation for strong coupling fluid-structure
interaction solvers. Journal of Applied Mechanics 76
24. Lübcke H, Schmidt St, Rung T, Thiele F (2001) Comparison of LES and RANS in bluff-body
flows. Journal of Wind Engineering and Industrial Aerodynamics 89:1471–1485
25. Lund E, Møller H, Jakobsen LA (2003) Shape design optimization of stationary fluid-structure
interaction problems with large displacements and turbulence. Struct Multidisc Optim 25:383–
392
26. Maute K, Nikbay M, Farhat C (2003) Sensitivity analysis and design optimization of three-
dimensional non-linear aeroelastic systems by the adjoint method. International Journal for
Numerical Methods in Engineering 56:911–933
27. Menter FR (1993) Zonal two equation   ! turbulence models for aerodynamic flows. AIAA
Paper 93-2906
28. Michler C, Van Brummelen EH and De Borst R (2005) An interface Newton-Krylov solver
for fluid-structure interaction. International Journal for Numerical Methods in Fluids 47:1189–
1195
29. Mohammadi B, Pironneau O (2004) Shape optimization in fluid mechanics. Annual Review of
Fluid Mechanics 36:255–279
30. Richards PJ, Hoxey RP (1993) Appropriate boundary conditions for computational wind
engineering models using the k  " model. Journal of Wind Engineering and Industrial
Aerodynamics 46 & 47:145–153
31. Richards PJ, Hoxey RP, Short LJ (2001) Wind pressure on a 6 m cube. Journal of Wind
Engineering and Industrial Aerodynamics 89:1553–1564
32. Schmidt S, Thiele F (2002) Comparison of numerical methods applied to flow over wall-
mounted cubes. International Journal of Heat and Fluid Flow 23:330–339
33. Sidi A (1991) Efficient implementation of minimal polynomial and reduced rank extrapolation
methods. Journal of Computational and Applied Mathematics 36:305–337
34. Simiu E, Scanlan RH (1996) Wind Effects on Structures. 3rd edn. Wiley-Interscience
35. Sobieszcanski-Sobieski J (1990) Sensitivity of complex, internally coupled systems. AIAA
Journal 28(1):153–160
36. Soto O, Löhner R (2001) CFD shape optimization using an incomplete-gradient adjoint
formulation. International Journal for Numerical Methods in Engineering 51:735–753
37. Suga K, Craft TJ, Iacovides H (2006) An analytical wall-function for turbulent flows and heat
transfer over rough walls International Journal of Heat and Fluid Flow 27:852–866
38. Turek S and Hron J (2006) Proposal for numerical benchmarking of fluid-structure interaction
between an elastic object and laminar incompressible flow. Lecture Notes in Computational
Science and Engineering 52:371–385
39. Vierendeels J, Lanoye L, Degroote J, Verdonck P (2007) Implicit coupling of partitioned
fluid-structure interaction problems with reduced order models. Computers & Structures 85:
970–976
40. Weller HG, Tabor G, Jasak H, Fureby C (1998) A tensorial approach to CFD using object
oriented techniques. Computers in Physics 12(6):620–631
41. Wilcox DC (2006) Turbulence Modeling for CFD. 3rd edn. DCW Industries
42. Wüchner R, Bletzinger K-U (2005) Stress-adapted numerical form finding of pre-stressed
surfaces by the updated reference strategy. International Journal for Numerical Methods in
Engineering 64(2):143–166
43. Wüchner R, Kupzok A, Bletzinger K-U (2007) A framework for stabilized partitioned analysis
of thin membrane-wind interaction. International Journal for Numerical Methods in Fluids
54:945–963
•
Experimental Benchmark: Self-Excited
Fluid-Structure Interaction Test Cases

J. Pereira Gomes and H. Lienhart

Abstract The swivelling motion of a flexible structure immersed in a flow can

become self-excited as a result of different fluid-structure interaction mechanisms.
The accurate simulation of these mechanisms still constitutes a challenge with
respect to mathematical modelling, numerical discretization, solution techniques,
and implementation as software tools on modern computer architectures. Thus, to
support the development of numerical codes for fluid structure interaction compu-
tations, in the present work an experimental investigation on the two-dimensional
self-excited periodic swivelling motion of flexible structures in both laminar and
turbulent uniform flows was performed. The investigated structural model consisted
of a stainless-steel flexible sheet attached to a cylindrical front body. At the trailing
edge of the flexible sheet, a rectangular mass was considered. The entire structure
model was free to rotate around an axle located in the central point of the front
body. During the experimental investigation, the general character of the elastic-
dynamic response of the structure model was studied first. The tests in laminar
flows were performed in a polyglycol syrup (dynamic viscosity: 1:64  104 m2 /s)
for a Reynolds number smaller than 270, whereas the tests in turbulent flows were
conducted in water for Reynolds numbers up to 44000. In both cases, the maxi-
mum incoming velocity tested was about 2 m/s. Subsequently, three specific test
cases were selected and characterized in more detail as far as the flow velocity field
and structure mechanical behavior are concerned. Thus, the present contribution
presents the detailed results obtained at 1:07 m/s and at 1:45 m/s in laminar and at
0:68 m/s in turbulent flows. It also compares the experimental data with numeri-
cal results obtained for the same conditions using different simulating approaches.
They revealed very good agreement in some of the fluid-structure interaction modes
whereas in others deficiencies were observed that need to be analyzed in more
detail.

J.P. Gomes and H. Lienhart

Institute of Fluid Mechanics and Erlangen Graduate School in Advanced Optical
Technologies, University of Erlangen-Nürnberg Cauerstr. 4, 91058 Erlangen, Germany
e-mail: [email protected], [email protected]

H.-J. Bungartz et al. (eds.), Fluid Structure Interaction II, Lecture Notes 383
in Computational Science and Engineering 73, DOI 10.1007/978-3-642-14206-2 14,
c Springer-Verlag Berlin Heidelberg 2010
384 J.P. Gomes and H. Lienhart

1 Introduction

The excitation, and posterior amplification, of a structure and fluid system oscilla-
tion can be described, in a first approach to the problem, as follows. If a structure
deforms, its orientation to the flow changes and a change on the fluid forces act-
ing upon the surface of the structure results. A new set of fluid forces determines
a new structure deformation. As soon as this coupled mechanism is initiated, as
a result of any initial flow or movement instability, the damping imposed by the
fluid can become negative as a result of different mechanisms by which energy is
transfered from the fluid to the structure. Whether the vibration is damped out or
amplified is just a matter of the sign of the net damping coefficient. For lightly
damped structures, the fluid damping becomes dominant and the coupled fluid and
structure movement may become self-excited. Depending on whether the fluctu-
ation of the flow plays a significant role in the initial excitation process or not,
the excitation is either called flow-induced or movement-induced excitation (MIE).
In the case of flow-induced excitation, one can further distinguish between the
extraneously-induced and instability-induced excitation (EIE and IIE) [5, 6].
In the case of EIE, the fluctuations are produced by an extraneous source. There-
fore the exciting frequency is independent of the structure movement and controls
the frequency of the resulting oscillation. Excitation mechanisms of this kind induce
the structure to undergo forced-oscillations. In opposition to EIE, IIE is caused by
an instability of the flow which gives rise to flow fluctuations if a certain threshold
value of flow velocity is reached. As a rule, this instability is intrinsic to the flow
created by the structure and exists even in those cases in which structure oscillations
do not occur. An example of flow instability is the alternate vortex shedding forma-
tion in the wake behind a bluff body. Depending on the control and amplification
mechanism affecting the instability, these fluctuations and the forces they generate
can become well correlated and close to a dominant frequency of the mechanical
oscillator, so they can lead to large-amplitude movements of the oscillating system.
Thus, this type of excitation is expected to occur in a finite small range of flow veloc-
ities in which the flow fluctuation is in resonance with the first or higher harmonics
of the natural frequencies of the structure [7]. This corresponds to the condition
U 1
 ; (1)
fN d nSt
where U is the flow velocity, fN and d are the natural frequency and the character-
istic length of the structure and St is the Strouhal number.
In the resonance range, the main feature of fluid-elastic and fluid-resonant con-
trol is the amplification of the exciting force (because of the negative damping)
and a ”locking-in” of the frequency of the fluid to that of the structure frequency.
The velocity range of locking-in and its extension is very sensitive to the instability
amplification and phase condition, which is intimately associated with the compat-
ibility between the flow and structure kinematics. It is also to be expected that in
the resonance range no change in the dynamics of the structure occurs because it is
Experimental Benchmark: Self-Excited Fluid-Structure Interaction Test Cases 385

reasonable to assume that the net effect of the fluid forces in-phase with the structure
acceleration is in average zero, i.e., the fluid stiffness balances the added mass.
Excitation of the type MIE is characterized by fluctuating forces that arise from
the movement of the structure. Whenever the structure is accelerated in the fluid,
an unsteady flow is induced that alters the fluid forces acting upon the structure. If
this alteration in the fluid load leads to a negative damping or to a transfer of energy
to the moving structure, a MIE process starts. In this process, the forces that are
responsible for the excitation are inherently linked to the structure movement and
disappear if the structure comes to rest. The sensitivity of the oscillating system to
MIE can be described by the equation of movement
   
m C A0 yR .t/ C B C B 0 yP .t/ C Cy .t/ D 0; (2)
where m, B and C are the mass, damping and stiffness of the structure, respectively.
A0 and B 0 are known as added, or fluid coefficients and account for the effects of the
fluid. The criterion for dynamic instability with respect to infinitesimal disturbances
can be stated in terms of the net damping of the system as

2m!N C B 0  0 : (3)
Regarding the behavior of the added mass, it can be seen from Eq. (2) that A0 is
not zero at the onset of MIE. Thus, in opposition to IIE, an alteration of the dynamics
of the structure in the case of MIE is expected.
Regardless the excitation mechanism, the self-excitation onset and resulting
oscillation movements are very difficult to predict whenever the structure has a com-
plex geometry or multiple degrees-of-freedom (such as flexible structures). From
the numerical research view point, the challenges with respect to mathematical
modeling, numerical discretization, solution techniques, and their implementation
as software tools on modern computer architectures are still huge, in particu-
lar if accuracy, flexibility and simulation efficiency are considered. Most of the
actual software packages for computational fluid dynamics (CFD) and computa-
tional structure dynamics (CSD) are already on a high level and allow, to some
extend, the simulations of certain classes of fluid-structure interaction (FSI) prob-
lems. However, although much research has been invested in the development of
models and coupling strategies for numerical simulations and in the creation of
coupling algorithms between CFD and CSD solvers, many of the key questions
regarding accuracy, robustness, flexibility and efficiency have not been yet com-
pletely answered. An overview of the present development of numerical solution
strategies and their applications is given, for instance, in [9] and [4].
The recent developments in new numerical methodologies triggered the present
work to perform an experimental research in the field of FSI investigating the insta-
bility and the resulting FSI induced swivelling motion of complex flexible structures
immersed in uniform laminar and turbulent flows. The object of the research were
two-dimensional structure models which combine the elastic behavior of a flex-
ible sheet in an axial flow with the dynamic of a rigid cylinder in cross-flow.
The selection of the experiments and the detailed measurements performed during
386 J.P. Gomes and H. Lienhart

the present work were used to compile a reliable and extensive experimental data
base on reference test cases. The data base created addressed the need for well-
defined experimental benchmarks to be used as a diagnostic and validation tool
for numerical models for FSI simulations. The data base on these reference test
cases included the time-phase resolved characterization of the flow velocity field and
the mechanical behavior of the structure such as its deflection, principal deflection
modes, periodic motion amplitude and frequency. The present contribution presents
three selected benchmarks extracted from the present work as well as comparisons
between the experimental results and numerical simulations.

2 Definition of the test cases

The project requirements for reproducibility and periodicity of the resulting flow and
structure motion, both in the laminar and turbulent regime, imposed stringent restric-
tions on the design of the experiments. Thus, the definition of the test case took
into account six principal requirements: (i) reproducibility of the resulting motion,
(ii) well controlled working and boundary conditions, (iii) two-dimensionality of
the structure deflection, (iv) moderate structure motion frequency, (v) significant
deformation of the structure, and (vi) linear material properties.

2.1 Structure definition

The structure found to serve the requirements of the project, see Fig. 1, consisted
of a 0:04 mm thick stainless-steel flexible sheet attached to an aluminum cylindri-
cal front body. At the rear end of the sheet, a rectangular stainless-steel mass was
located. Both the rear mass and the front body could be considered rigid. An extra
degree of freedom was taken into account as the structure was free to rotate around
an axle located in the center of the cylindrical front body.
The densities of the different materials used in the construction of the model as
well as all parameters of the different test cases discussed in the present paper are

Fig. 1 Structure model geometry (all dimensions in mm).

Experimental Benchmark: Self-Excited Fluid-Structure Interaction Test Cases 387

Table 1 Parameters of the test cases.

Variable 1st test 2nd test 3rd test
Laminar Laminar Turbulent
Reynolds number  140  190  15000
cylinder 2828 kg=m 3 2828 kg=m 3 2828 kg=m 3
flexible sheet 7855 kg=m 3 7855 kg=m 3 7855 kg=m 3
rear mass 7800 kg=m 3 7800 kg=m 3 7800 kg=m 3
Eflexible sheet 2  1011 N=m2 2  1011 N=m2 2  1011 N=m2
fluid 1050 kg=m 3 1050 kg=m 3 998 kg=m 3
fluid 1:64  104 m2 =s 1:64  104 m2 =s 0:97  106 m2 =s
gravity 9:81 m=s2 in x 9:81 m=s2 in x 9:81 m=s2 in x
incoming velocity 1:07 m=s 1:45 m=s 0:68 m=s

Table 2 Natural frequencies of the structure model.

mode 0 1st 2nd 3rd 4th
frequency 1.94 Hz 5.89 Hz 27.43 Hz 92.73 Hz 234.41 Hz

summarized in Table 1. For the range of forces acting on the structure during the
tests, the mechanical behavior of the flexible stainless-steel sheet could be consid-
ered to be linear and Young’s modulus was measured to be 200 kN=mm2 within this
range of strain. As it was already discussed, in fluid-structure interaction problems
the natural frequencies of the structure play an important role in the overall char-
acteristics of the elastic-dynamic response of the coupled system. To support this
statement, one can argue that the domain of relevance of each individual instability-
induced exciting mechanism is directly dependent on the natural frequencies of the
structure. Hence, for a better understanding and analysis of the results, the definition
of the model also include its natural frequencies. The natural frequencies were com-
puted in a finite element code and are listed in Table 2. The mode identified as the
zero mode corresponded to the rigid body motion mode, i.e. the structure swiveled
around its free axle without changing its original straight shape.

2.2 Flow definition

The tests were conducted in a vertical, closed-loop tunnel capable of operating with
liquids of different viscosity. The spatial dimensions of the tunnel test section are
given in Fig. 2. It had an overall length of 338 mm and a cross-sectional area of
180 mm  240 mm. The structure was mounted 55 mm downstream of the inlet
plane.
Opting for a vertical tunnel, the gravity force was aligned with the x-axis and so
it did not introduce any asymmetry. Special attention was given to the model support
on the test section. Low-friction bearings were used for this specific task to guaran-
tee a frictionless rotational degree of freedom of the front cylinder. The supporting
388 J.P. Gomes and H. Lienhart

338 mm

(-55 mm; 120 mm) (284 mm; 120 mm)

240 mm
0 x

55 mm

(-55 mm; -120 mm) (284 mm; -120 mm)

Fig. 2 Experimental domain.

system also considered a non-contacting magnetic position sensor to measure the

angular position of the structure front cylinder. The output of this sensor was used
during the tests to measure the time-dependent angle of the structure front body and
to determine the beginning of a new model swivelling motion period.
For the tests in the laminar regime, a polyethylene glycol PG-12000 (polyglycol)
syrup was used. The measurements were performed at 25o C within an uncertainty
of 0:5o C. For this temperature range, the properties of the syrup could be considered
constant and the kinematic viscosity was measured as 1:64104 m2 /s. For the same
conditions, the density was equal to 1050 kg/m3 . After an analysis of the elastic-
dynamic behavior of the structure for an incoming flow velocities up to 2 m/s, it
was decided to perform the laminar reference test case at two different velocities,
the first at 1:07 m/s and the second at 1:45 m/s. The tests inlet velocity profiles were
measured in the absence of the model at the location x D 55 mm (see Fig. 3) in
the center of the spanwise direction.
On the other hand, the tests in turbulent flows were conducted in water at the
C0ı C
temperature of 22ı C1 ı C (w D 998 kg/m ; w D 0:97  10
3 6
m2 /s). For this
regime, only one incoming velocity was selected for the reference tests, namely
0:68 m/s. Figure 4 shows the measured inlet velocity profile at 0:68 m/s. It was also
measured in the absence of the model at the location x D 55 mm in the center of
the spanwise direction.
For both flow regimes, the flow angularity fluctuation was measured to be less
than 0:5ı and the RMS of the velocity magnitude variation was less than 1%. As
a synopsis, Table 1 gives a complete overview of the values relevant for the three
referent test cases discussed in this paper. The Reynolds number of the tests was
defined based on the diameter of the structure front cylinder.
Experimental Benchmark: Self-Excited Fluid-Structure Interaction Test Cases 389

120 120

80 80

40 40
y [mm]

y [mm]
0 0

-40 - 40

-80 - 80

-1200 -120
0.25 0.5 0.75 1 1.25 1.5 1.75 2 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2
velocity [m/s] velocity [m/s]

Fig. 3 Measured inlet velocity profiles for the first (left) and second (right) laminar reference test
case.

120

80

40
y [mm]

- 40

- 80

- 120
0 0.25 0.5 0.75 1 1.25 1.5 1.75 2
velocity [m/s]

Fig. 4 Measured inlet velocity profile for the turbulent reference test case.

3 Measurement techniques

3.1 Flow velocity measurements

The task of sampling the flow velocity of the fluid surrounding the flexible struc-
ture was addressed to a two-component, two-dimensional PIV system. The PIV
system adopted for the measurements consisted of two 1280 pixel  1024 pixel
synchronized cameras and a pulsed double-head 120 mJ laser.
The cameras were used to acquire two simultaneous time-dependent pairs of
PIV images per measurement. For that purpose, they are mounted parallel to the
rotating axle of the flexible structure to visualize the flow in a plane perpendicular
to it. To assure the correct position of the two adjacent images, a special support
was designed to hold both cameras and to permit the adjustment of each individual
390 J.P. Gomes and H. Lienhart

camera or of both cameras simultaneously. The PIV images acquired by the two
cameras were then imported into a MATLAB-based post-processing software to
stitch the correspondent pairs of images before being cross-correlated. Opting for
the solution of two synchronized cameras, it was possible to achieve an extended
measuring area while keeping the spatial resolution as low as 2:1 mm  2:1 mm for
a 16 pixel  16 pixel interrogation area. In relation to Fig. 2, the flow velocity field
measuring area extended from 35 mm up to 225 mm in the x-direction and from
85 mm up to 85 mm in the y-direction.
Preliminary tests with the present system showed that reliable and continuous
data acquisition from the two synchronized cameras was possible for long time
periods using a maximum frequency of 1 Hz. Concerning the velocity accuracy,
the system proved to have a measurement uncertainty of less than 1:5 % for the
undisturbed flow.
The laser sheet for flow illumination is positioned perpendicular to the rotating
axis of the flexible structure at the center of the structure in the spanwise direc-
tion. The presence of the swivelling flexible structure in the measurement plane
imposes extra problems. The most important of them is related to the fact that the
flexible structure is a swivelling opaque body which creates a time-dependent dark
shadow region in the fluid. This behavior not only reduces the measuring area to
almost only one side of the flexible structure but also makes the masking of the PIV
images in post-processing difficult to perform. To cope with this problem, multiple
light sources were used to illuminate the structure from both sides and, therefore, to
cover the entire flow field avoiding the shade behind the model. As seeding particles,
10 m mean diameter hollow glass spheres were chosen to be used in water. They
provided a good match of density and enough scattering signal over all the measur-
ing area. During the laminar tests, in the polyglycol syrup, silver-coated hollow glass
spheres with the same diameter were adopted as seeding particles. They produce
higher signal levels in high light-absorbing media compared with non-coated hol-
low glass spheres. The major drawback of the silver-coated glass spheres is related
to their density; the relative density of this kind of particles is about 1:4. Neverthe-
less, this drawback was acceptable because of the high viscosity of the fluid and the
velocity of the flow during the tests.

3.2 Structure deflection measurements

For structure deflection measurements, the PIV system was modified to provide it
with structure deflection analysis capabilities. The idea behind this setup was to use
the PIV system to acquire images from the swivelling structure and to adopt spe-
cially developed software to analyze and reconstruct the time-dependent deflection
of the structure. The major advantage of this approach was that the same measuring
system as used for the velocity field measurements could be employed.
In order to generate coherent results, the relative position of the laser source was
maintained unchanged to illuminate the time-dependent deflection of the flexible
Experimental Benchmark: Self-Excited Fluid-Structure Interaction Test Cases 391

structure at the same plane as for the velocity field measurements. The cameras
were now mounted in a side-by-side arrangement in such a way as to acquire simul-
taneous images of the flexible structure illuminated by the laser sheet from each side
of the model.
The quantitative analysis was performed in MATLAB workspace by a script
developed for the specific task. The software analyzed and compared the images
illuminated from both sides of the model and reconstructed the time-dependent
image of the light sheet reflected by the structure. To achieve that purpose, it mapped
the pixel value in the gray-scale of the entire image and detected the line result-
ing from the intersection of the laser sheet and the structure in addition to the
edges of the rear mass. Then the same code, including camera calibration routines,
reconstructed the real scale and perspective-corrected position of that line. With
the information on the position of the flexible sheet and of the time-phase detector
module, the algorithm finally computed all the relevant data for the structure move-
ment such as time-phase resolved angle of the front body, structure deflection shape
and trailing edge coordinates. Based on these data, the modes present in the structure
were identified and characterized. Taking into account the accuracy of the measure-
ment technique together with the resolution of the time-phase reconstruction, one
could expect an uncertainty of the structure deflection measurements of the order of
0:3 mm.

3.3 Resolution of the results in the time-phase space

When investigating fluid-structure interaction problems, two additional difficulties

appear when it comes to resolving the measured data in time-phase space. First,
the periodicity of the structure motion is sensitive and therefore there are cycle-to-
cycle variations of the period time. Second, the velocity of the structure motion is
not predefined (as it is, for example, in crank shaft-driven set-ups), which makes it
impossible to reconstruct the time-phase resolved data from position resolved mea-
surements. For these reasons, a different approach was adopted. Instead of triggering
the data acquisition by the experiment, the measuring system was operated at a con-
stant acquisition rate and both events, the acquisition of a measurement and the start
of a new movement cycle, were recorded based upon an absolute clock. Using this
time information, the data were reorganized in a post-processing program in order to
provide the time-phase resolved values. The hardware module designed to perform
the event monitoring was based on an FPGA (Field Programmable Gate Array) and
a 1 MHz internal clock. In this way, it was able to record up to 250 events per second
with an accuracy of 2 s.
Two different kinds of events were recorded: the measurements (tj ) and the
beginning of the flexible structure motion period (ti; ). Depending on the type of
measurements performed (flow velocity measurements or structure deflection mea-
surements), the measurement events were detected using the first laser pulse trigger
signal or the camera trigger signal, respectively. After deciding which angular
392 J.P. Gomes and H. Lienhart

position of the structure front body corresponded to the beginning of the motion
period, this position was detected by a magnetic angular position sensor and the
small magnet attached to the axle of the model. This sensor was selected to perform
the task based on two criteria: non-contacting position angular sensor and direction
resolved output signal. An additional reason, namely the high sensitivity around a
predefined angular position, also contributed to the selection of this sensor.
From the recorded events time information, a software computed the measure-
ment time-phase angle tpaj within the structure motion period as

tj  ti;
tpaj D  360; (4)
Ti

where ti; and Ti corresponded to the starting instant and the period of the swivelling
cycle in which the measurements tj took place.
Finally, the time-phase resolved measurements were reorganized in the time-
phase space in a reference structure motion period, as time-phase angle, between
0ı and 360ı at a resolution of 2:5ı . Further details can be found elsewhere [2, 3].

4 Results

To assist the understanding of the results, they are sub-divided according to the
flow regime. Thus, in Sect. 4.1 the results obtained in the laminar regime are pre-
sented and in Sect. 4.2 the results in turbulent flows. For each regime, the structure
model was first tested at different incoming flow velocities up to 2 m/s. These
results defined the general character of the dynamic response of the structure model
as a function of the incoming flow velocity and showed the different swivelling-
modes exhibited by the structure. In the second stage, detailed measurements were
conducted at selected velocities to characterize each combined flow and structure
swivelling-mode. Section 4.1 includes detailed measurements obtained in laminar
flow at 1:07 m/s and 1:45 m/s. In turbulent flows, detailed measurements were per-
formed only for one approaching flow velocity, 0:68 m/s. These measurements are
presented in Sect. 4.2.

4.1 Results in laminar flows

Figure 5 shows the dynamic response of the structure in laminar flow. The Reynolds
number, based on the diameter of the front cylindrical body, reached the maximum
value of 270 at 2 m/s. At very low flow velocities, it was not possible to identify any
kind of motion. On increasing the flow speed, it was observed that the minimum
velocity needed to excite the movement of the structure varied slightly from test to
test. Nevertheless, in all cases it was possible to achieve a periodic cyclic swivelling
Experimental Benchmark: Self-Excited Fluid-Structure Interaction Test Cases 393

70 20
18
Peak-to-peak amplitude [deg]

60

swiveling frequency [Hz]

16
50 14
12
40
10
30 8
20 6
4
10
2
0 0
0 0.5 1 1.5 2 0 0.5 1 1.5 2
flow velocity [m/s] flow velocity [m/s]

Fig. 5 Structure front body peak-to-peak amplitude (left) and structure swivelling frequency
(right) versus incoming flow velocity (solid squares correspond to measurements acquired while
increasing and open squares while decreasing the flow velocity).

motion for velocities slightly smaller than 1 m/s. It should be mentioned that as
soon as the structure started to swivel, its motion frequency coincided with the line
in Fig. 5 independently of the velocity value at which the movement started. From
the instant that the structure started to swivel and for all the range of velocities
tested, the resulting motion proved to be symmetric and periodic. The RMS value
of the cycle-to-cycle structure motion period was measured to be less than 1%.
The most obvious aspect revealed by Fig. 5 is the existence of two distinc-
tive structure swivelling-modes separated by a pronounced, well-defined hysteretic
region. For both swivelling-modes, the frequency of the resulting motion increased
linearly with the velocity of the incoming flow. While the frequency increased
monotonically with the flow velocity, the amplitude of the structure motion showed
a maximum value for each swivelling motion mode. The first swivelling-mode, reg-
istered for incoming flow velocities up to approximately 1:1 m/s, was characterized
by the fact that the deflection of the structure model was strongly governed by the
first bending mode. In connection with this, the movement of the rear mass was in
concordance with the movement of the front body. The second swivelling-mode,
observed for incoming velocities higher than 1:3 m/s, was characterized by the fact
that the rear mass motion is in opposition to the movement of the structure front
body. At the same time, higher bending modes were present in the deflection of
the structure. The conclusion about the deflection modes exhibited by the structure
was supported by visualizations performed during the tests at different flow speeds.
Within the transition region, from 1:1 m/s to 1:3 m/s, the structure presented a hys-
teretic behavior where both swivelling-modes could be observed depending on the
previous frequency of the structure. In both cases, a delay of the movement of the
rear mass in relation to the front body was identified. This delay is a consequence
of the flexibility of the sheet and it is a function of the mechanical properties of
the structure. After identifying the two different self-excited swivelling-modes of
394 J.P. Gomes and H. Lienhart

the flow-structure system, each was characterized in more detail. To this end, two
flow velocities, 1:07 m/s and 1:45 m/s, were selected as representative of the two
swivelling-modes. Both values are located close to the velocity of the maximum
structure amplitude excitation (Fig. 5(a)).

Results at 1.07 m/s

In the following figures, the characterization of the structure movement for an

incoming flow velocity of 1:07 m/s is shown. At this velocity, the Reynolds number
is approximately 140, based on a front body diameter of 22 mm and a kinematic
viscosity of the polyglycol syrup of 1:64  104 m2 /s. The swivelling frequency
observed was equal to 6:38 Hz.
Figure 6 shows the evolution of the angle of the structure front body within the
swivelling motion averaged period and Figs. 7 and 8 present the successive positions
of the structure flexible sheet and the coordinates of the trailing edge during one
period. The time-phase resolution in Fig. 7 was set to 30ı whereas that used for
Fig. 8 was 5ı .
As far as the flow field surrounding of the structure model is concerned, Fig. 9
compiles the time-phase resolved combined flow field and structure deflection at
eight instants of the reference swivelling period. The successive results indicate
that the movement of the trailing edge is in-phase, but delayed, with respect to the
movement of the front body.
This delay could be quantified by comparing the time-phase resolved angle of the
front body with the y-coordinate of the structure trailing edge. Thus, at 1:07 m/s the
delay of the trailing edge with respect to the front body movement was computed to
be approximately 60ı of time-phase.

40

30

20
front body angle [deg]

10

-10

-20

-30

-40
0 60 120 180 240 300 360
time-phase angle [deg]

Fig. 6 Time-phase resolved front body angle within a period of motion at 1:07 m/s (Re  140).
Experimental Benchmark: Self-Excited Fluid-Structure Interaction Test Cases 395

20

10

y [mm]
0

-10

-20

0 10 20 30 40 50 60 70 80
x [mm]

Fig. 7 Flexible sheet deflection within a period of motion at 1:07 m/s (Re  140).

40

30

20

10
y [mm]

- 10

- 20

- 30

- 40
0 10 20 30 40 50 60 70 80
x [mm]

Fig. 8 Trailing edge coordinates within a period of motion at 1:07 m/s (Re  140).

Results at 1.45 m/s

A similar set of measurements were performed for an incoming flow velocity of

1:45 m/s, which corresponds to a Reynolds number close to 190. At 1:45 m/s,
the structure exhibited a more complex and faster swivelling motion. The resulting
motion frequency was measured to be 13:58 Hz and the maximum front body angu-
lar amplitude was ˙22ı . In Fig. 10 the angle of the front body within the averaged
period of motion is displayed. The time-phase delay of the trailing edge excursion
in relation to the front body movement increased to about 210ı.
Figures 11 and 12 show the time-phase resolved position of the structure and of
the structure model trailing edge within the averaged swivelling motion. Now the
collection of the deformations display a pronounced nodal region, indicating the
existence of higher bending modes in the structure deflection. In Figs. 11 and 12,
time-phase angle resolutions of 30ı and 5ı were used. Regarding the unsteady flow
396 J.P. Gomes and H. Lienhart

75 75

50 50

25 mag [-] 25 mag[-]

1.5
1.5
y [mm]

y [mm]
1.3 1.3
1.1 1.1
0 0.9
0.7
0 0.9
0.7
0.5 0.5
0.3 0.3

- 25 0.1
-25 0.1

- 50 -50
- 75 -75
- 25 0 25 50 75 100 125 150 175 200 225 -25 0 25 50 75 100 125 150 175 200 225
x [mm] x [mm]
ı ı
(a) tpa D 0 (b) tpa D 45
75 75

50 50

25 mag [-]
1.5
25 mag [-]
1.5
y [mm]

y [mm]
1.3 1.3
1.1 1.1
0 0.9
0.7 0 0.9
0.7
0.5 0.5
0.3 0.3
-2 0.1
-25 0.1

-5 -50
-7 -75
-25 0 25 50 75 100 125 150 175 200 225 -25 0 25 50 75 100 125 150 175 200 225
x [mm] x [mm]
ı ı
(c) tpa D 90 (d) tpa D 135
75 75

50 50

25 mag [-]
1.5
25 mag [-]
1.5
y [mm]

y [mm]

1.3 1.3
1.1 1.1
0 0.9
0.7 0 0.9
0.7
0.5 0.5
0.3 0.3
-25 0.1
-25 0.1

-50 -50
-75 -75
-25 0 25 50 75 100 125 150 175 200 225 -25 0 25 50 75 100 125 150 175 200 225
x [mm] x [mm]
ı ı
(e) tpa D 180 (f) tpa D 225
75 75

50 50

25 mag [-]
1.5
25 mag [-]
1.5
y [mm]

y[mm]

1.3 1.3
1.1 1.1
0 0.9
0.7 0 0.9
0.7
0.5 0.5
0.3 0.3
-25 0.1
-25 0.1

-50 -50
-75 -75
-25 0 25 50 75 100 125 150 175 200 225 -25 0 25 50 75 100 125 150 175 200 225
x [mm] x [mm]
(g) tpa D 270ı (h) tpa D 315ı

Fig. 9 Time-phase resolved combined flow field/structure deflection measurements at eight

different instants of the swivelling motion period at 1:07 m/s (Re  140).
Experimental Benchmark: Self-Excited Fluid-Structure Interaction Test Cases 397

40

30

20

front body angle [deg]

10

-10

-20

-30

-40
0 60 120 180 240 300 360
time-phase angle [deg]

Fig. 10 Time-phase resolved front body angle within a period of motion at 1:45 m/s (Re  190).

20

10
y[mm]

-10

-20

0 10 20 30 40 50 60 70 80
x[mm]

Fig. 11 Flexible sheet deflection within a period of motion at 1:45 m/s (Re  190).

40

30

20

10
y [mm]

-10
-20
-30
-40
0 10 20 30 40 50 60 70 80
x [mm]

Fig. 12 Trailing edge coordinates within a period of motion at 1:45 m/s (Re  190).
398 J.P. Gomes and H. Lienhart

field results, Fig. 13 presents the velocity field around the structure for eight, 45ı
equally distant, time-phase angles within the reference movement cycle.

4.2 Results in turbulent flows

In the turbulent regime, using water as the working fluid, the structure proved to
have the same well-defined multi swivelling-mode behavior as observed in laminar
flows. Figure 14 presents the dynamic response of the structure versus the incoming
water flow velocity up to 2 m/s (Re  44000).
Now the structure could be excited to a periodic swivelling motion at very low
flow velocities. Visualizations showed that this first excited mode corresponded
to the rigid body motion mode, i.e., the structure swiveled in the fluid around
its free rotating axle without changing its original and straight shape. Because it
corresponded to the rigid body motion, this mode was named the zero swivelling-
mode. As an example of this zero mode, Fig. 15 shows the structure behavior at
approximately 0:19 m/s in water.
A swivelling-mode transition was registered for a flow velocity close to 0:4 m/s.
The transition between modes is abrupt and it was not possible to observe either any
evolution of the structure motion during the transition or any hysteretic behavior. In
the new swivelling-mode, the structure deflection was dominated by the first bend-
ing mode. The behavior of the structure in this mode is in all respects similar to
the first swivelling-mode observed in laminar flows. For both swivelling-modes, the
amplitude of the structure movement was limited and exhibited a local maximum
value. The structure movement frequency increased approximately linearly with the
velocity of the incoming flow in both modes. The only exception occurred in the first
mode at approximately 0:6 m/s where a change in slope was registered because of
a lock-in occurrence. On further increasing the incoming flow velocity, an unusual
behavior was observed: as soon as the amplitude of the structure started to decrease,
after reaching the local maximum, the motion characteristics degraded very rapidly.
This effect was supported by the RMS value presented in Fig. 14. The coupled
movement became non-periodic and non-symmetric and led to a rapid destruction
of the structure. Therefore, no measurements could be obtained for flow velocities
higher than 0:9 m/s. This sequence of facts occurred when the structure swivelling
frequency was showing the first signs of transition to a new, second swivelling-
mode. In the range in which the structure movement is periodic and reproducible,
up to 0:9 m/s, the RMS value of the cycle-to-cycle motion period remained lower
than 1%.
Except for the trivial rigid body mode, the only self-exciting swivelling-mode
that could be characterized in detail using the present structure configuration was
the first one. Therefore, further investigations were performed in a water flow at
0:68 m/s. The decision regarding that velocity followed the same criteria as used for
the laminar investigations.
Experimental Benchmark: Self-Excited Fluid-Structure Interaction Test Cases 399

75 75

50 50

25 mag [-]
1.5
25 mag [-]
1.5
y [mm]

y[mm]
1.3 1.3
1.1 1.1
0 0.9
0.7
0 0.9
0.7
0.5 0.5
0.3 0.3
0.1
-25 0.1
-25

-50 -50

-75 -75
-25 0 25 50 75 100 125 150 175 200 225 -25 0 25 50 75 100 125 150 175 200 225
x [mm] x[mm]
ı
(a) tpa D 0 (b) tpa D 45ı
75 75

50 50

25 mag [-] 25 mag [-]

y [mm]

1.5 1.5

y [mm]
1.3 1.3
1.1 1.1
0 0.9
0.7
0 0.9
0.7
0.5 0.5
0.3 0.3
0.1 0.1
-25 -25

-50 -50

-75 -75
-25 0 25 50 75 100 125 150 175 200 225 -25 0 25 50 75 100 125 150 175 200 225
x [mm] x [mm]
(c) tpa D 90ı (d) tpa D 135 ı

75 75

50 50

25 mag[-]
1.5
25 mag [-]
1.5
y [mm]
y [mm]

1.3 1.3
1.1 1.1
0 0.9
0.7
0 0.9
0.7
0.5 0.5
0.3 0.3
0.1
-25 0.1
-25

-50 -50

-75 -75
-25 0 25 50 75 100 125 150 175 200 225 -25 0 25 50 75 100 125 150 175 200 225
x [mm] x [mm]
ı ı
(e) tpa D 180 (f) tpa D 225

75 75

50 50

25 mag [-]
1.5
25 mag [-]
1.5
y [mm]
y [mm]

1.3 1.3
1.1 1.1
0 0.9
0.7 0 0.9
0.7
0.5 0.5
0.3 0.3
-25 0.1
-25 0.1

-50 -50

-75 -75
-25 0 25 50 75 100 125 150 175 200 225 -25 0 25 50 75 100 125 150 175 200 225
x [mm] x [mm]
(g) tpa D 270ı (h) tpa D 315ı

Fig. 13 Time-phase resolved combined flow field/structure deflection measurements at eight

different instants of the swivelling motion period at 1:45 m/s (Re  190).
400 J.P. Gomes and H. Lienhart

70 20
Peak-to-peak amplitude [deg] 18
60
16

swiveling frequency [Hz]

50 14
12
40
10
30 8

20 6
4
10
2
0 0
0 0.5 1 1.5 2 0 0.5 1 1.5 2
flow velocity [m/s] flow velocity [m/s]

Fig. 14 Structure front body peak-to-peak amplitude (left) and structure swivelling frequency
(right) versus incoming flow velocity (solid squares correspond to measurements acquired while
increasing and open squares while decreasing the flow velocity).

40
20
30
front body angle [deg]

20
10
10
y [mm]

0 0

-10
-10
-20

-30 -20
-40
0 60 120 180 240 300 360 0 10 20 30 40 50 60 70 80
time-phase angle [deg] x [mm]

Fig. 15 Time-phase resolved front body angle (left) and flexible sheet deflection (right) within a
period of motion at 0:19 m/s.

Results at 0.68 m/s

At 0:68 m/s, the Reynolds number of the measurements performed in water was
about 15000, based on the diameter of the front cylinder and on the water properties.
Under such conditions, the structure exhibited a 4:45 Hz periodic swivelling motion
with the rear mass delayed, in time-phase angle, 95ı in relation to the front body.
Once again, the characterization of the resulting movement is based on the angle of
the structure front body, trailing edge position and structure deflection. Figures 16
to 18 show the time-phase resolved evolution of these quantities within the averaged
period of motion. The same resolutions as used to generate the laminar results were
used in the following figures.
Experimental Benchmark: Self-Excited Fluid-Structure Interaction Test Cases 401

40

30

front body angle [deg]

20

10

-10

-20

-30

-40
0 60 120 180 240 300 360
time-phase angle [deg]

Fig. 16 Time-phase resolved front body angle within a period of motion at 0:68 m/s (Re 
15000).

20

10
y [mm]

-10

-20

0 10 20 30 40 50 60 70 80
x [mm]

Fig. 17 Flexible sheet deflection within a period of motion at 0:68 m/s (Re  15000).

40

30

20

10
y [mm]

-10

-20

-30

-40
0 10 20 30 40 50 60 70 80
x [mm]

Fig. 18 Trailing edge coordinates within a period of motion at 0:68 m/s (Re  15000).
402 J.P. Gomes and H. Lienhart

Figure 19 shows the flow velocity field results obtained in water at 0:68 m/s for
eight successive time-phase angles measured within a period of the structure motion.
The time-phase angles presented correspond to those adopted in Sect. 4.1.

4.3 Discussion of the experimental results

The analysis of the structure model dynamic response proved the existence of dif-
ferent structure swivelling-modes. In the laminar regime, it was possible to observe
two swivelling-modes depending on the approaching flow velocity. For both modes,
the structure movement frequency increased linearly whereas the front body ampli-
tude presented a local maximum. The maximum excitation occurred at 1:1 m/s and
1:6 m/s for the first and second excitation mode, respectively. The corresponding
movement frequencies for these two instants of maximum excitation were measured
to be around 6:5 Hz and 15 Hz.
The first mode was excited for the first time at 0:8 m/s. For this flow velocity, the
corresponding Strouhal number (S t  0:175) and the first natural frequency of the
structure (N1 D 5:9 Hz) showed a strong interconnection between the excited move-
ment and the classical von Karman vortex shedding triggered by the structure front
cylinder. Despite a small delay, the trailing edge movement could be considered in
phase with the angular movement of the front body. Concerning the deflection of
the structure, this mode was characterized almost exclusively by the existence of
the first bending mode. At 1:07 m/s, the structure vibrated around its first natural
frequency; more precisely, the coupled fluid and structure unsteady motion was reg-
istered to occur at 6:38 Hz associated with a maximum excursion of the front body
and trailing edge of 19ı and 16 mm, respectively.
The transition to the second, more complex mode was observed between 1:12 m/s
and 1:3 m/s and it showed a strong hysteretic behavior. The second swivelling-mode
was characterized by a frequency much lower than the second natural frequency of
the structure (N2 D 27:4 Hz), indicating a self-exciting mechanism of a different
type than the first one. The trailing edge was now almost in phase opposition in
relation to the front body position and the structure deflection was mainly character-
ized by the second structure bending mode and higher. The presence of the second
bending mode justified the pronounced node observed in the structure deformation.
At 1:45 m/s, the front cylinder reached a maximum deflection of 26ı and the trailing
edge excursion was limited to 19 mm.
In the turbulent regime, one more mode was observed for very small inlet flow
velocities. The lowest mode to be excited in water was the rigid body mode (referred
to as the zero swivelling-mode). This mode started to be observed at a very small
approaching velocity and it was characterized by small structure deflections and
movement frequencies. The maximum rigid body excitation was registered for
0:2 m/s at the same time the structure swivelled at about 1:1 Hz. The transition
to the first self-excitation mode was registered at about 0:4 m/s and no hysteresis
was observed. In this mode, the maximum excitation of the structure was achieved
Experimental Benchmark: Self-Excited Fluid-Structure Interaction Test Cases 403

75 75
50 50
25 mag [-]
1.5
25 mag [-]
1.5
y [mm]

1.3

y[mm]
1.3
1.1 1.1
0 0.9
0.7 0 0.9
0.7
0.5 0.5
0.3 0.3
-25 0.1
-25 0.1

-50 -50

-75 -75
-25 0 25 50 75 100 125 150 175 200 225 -25 0 25 50 75 100 125 150 175 200 225
x [mm] x[mm]
ı ı
(a) tpa D 0 (b) tpa D 45
75 75

50 50

25 mag [-]
1.5
25 mag [-]
1.5

y [mm]
y [mm]

1.3 1.3
1.1 1.1
0 0.9
0.7
0 0.9
0.7
0.5 0.5
0.3 0.3
0.1
-25 0.1
-25

-50 -50

-75 -75
-25 0 25 50 75 100 125 150 175 200 225 -25 0 25 50 75 100 125 150 175 200 225
x [mm] x [mm]
(c) tpa D 90ı (d) tpa D 135ı

75 75
50 50
25 mag [-]
1.5 25 mag [-]
1.5
y [mm]
y [mm]

1.3 1.3
1.1 1.1
0 0.9
0.7 0 0.9
0.7
0.5 0.5
0.3 0.3
0.1
-25 -25 0.1

-50 -50
-75 -75
-25 0 25 50 75 100 125 150 175 200 225 -25 0 25 50 75 100 125 150 175 200 225
x [mm] x [mm]
ı ı
(e) tpa D 180 (f) tpa D 225
75 75

50 50

25 mag [-]
1.5
25 mag [-]
1.5
y [mm]

y [mm]

1.3 1.3
1.1 1.1
0 0.9
0.7 0 0.9
0.7
0.5 0.5
0.3 0.3
0.1
-25 -25 0.1

-50 -50

-75 -75
-25 0 25 50 75 100 125 150 175 200 225 -25 0 25 50 75 100 125 150 175 200 225
x [mm] x [mm]
ı ı
(g) tpa D 270 (h) tpa D 315

Fig. 19 Time-phase resolved combined flow field/structure deflection measurements at eight

different instants of the swivelling motion period at 0:68 m/s (Re  15000).
404 J.P. Gomes and H. Lienhart

for 0:8 m/s corresponding to a movement frequency of 5 Hz. Detailed measure-

ments at 0:68 m/s registered a 4:45 Hz self-exciting resulting motion associated
with a maximum excursion of the front body and the trailing edge of 29ı and
25 mm, respectively. Finally, at 0:9 m/s, the structure start to give indications of
a swivelling-mode transition. However, beyond this value it was not possible to reg-
ister the dynamic behavior of the structure model. On increasing the velocity, the
resulting movement of the model became unstable and non-reproducible, finally
leading to the failure of the structure.
In a similar way as for the first laminar swivelling-mode, a direct connection
between the movement excitation and the classical von Karman vortex shedding
behind the structure front cylinder was proved to exist. This is supported for the first
swivelling-mode observed in turbulent flows by the relation between the first natural
frequency of the structure (N1 ) and the turbulent Strouhal number (S t  0:21). The
same direct relation was observed for the rigid body swivelling-mode (or rigid body
mode) on comparing the Strouhal number and the rigid body natural frequency of
the structure (N0 D 0:19 Hz).
Based on the results, one can conclude that the first swivelling-mode is similar
in nature in both the laminar and turbulent regimes. The main differences between
the two are only connected with the fact that for turbulent flow the damping of
the coupled system was significantly reduced. Thus, for the turbulent case, higher
excursions from the trailing edge and front body and also higher deflection of the
flexible part of the structure were observed. Because of the lower damping imposed
by the fluid, the structure was exposed to higher accelerations during its swivelling
movement in the turbulent tests. Another difference appears when comparing the
movement of the structure trailing edge. At higher Reynolds number, the area cov-
ered by the ”figure-of-eight” shaped trajectory was considerably bigger than that
obtained in the laminar tests. This is related to the delay registered between the rear
mass and the front body movement. In both cases, the movement of the rear mass
could be considered to be in concordance with the front cylinder rotation, but the
delay increased with the Reynolds number: whereas at Re D 140 the delay was
measured to be close to 60ı , at Re D 15000 it reached 95ı .
Comparing both self-excited coupled motions, one may conclude that both are
triggered by the vortex shedding created around the front cylinder. Because the nat-
ural frequencies of the structure are constant and the Strouhal number is not so
sensitive to the Reynolds number in the range 140-15000, the resulting movements
have the same response as far as frequency of the movement versus approaching
flow velocity is concerned. Considering all evidences, it can be concluded that both
first swivelling-modes correspond to instability-induced excited (IIE) fluid-structure
interaction cases. The same applies to the rigid body mode observed in the turbulent
tests. In the laminar regime, a similar rigid body mode could not to be registered.
The excitation process responsible for the second mode observed in laminar flow is
more difficult to examine. However, the results indicate strongly that this mode can
be attributed to movement-induced excitation (MIE).
Experimental Benchmark: Self-Excited Fluid-Structure Interaction Test Cases 405

5 Comparison of numerical and experimental results

Within the DFG Research Unit 493 (Fluid-Structure Interaction: Modelling, Simu-
lation, Optimisation), two sets of numerical simulation data were produced. More
information about these two numerical studies including details on the numerical
schemes, modelling and coupling approaches adopted can be found in [8] and [1].
Both data sets covered the two swivelling-modes of the laminar test cases and are
discussed in the following subsections.

5.1 First swivelling-mode in laminar flows

At the inlet velocity of 1:07 m/s, a self excited cyclic swivelling of the structure
exposed to the uniform and constant flow was observed during the experiments
in which the movement of the rear mass was in concordance with the rotation
of the front cylinder. The numerical simulations reproduced the same kind of
movement for this first swivelling-mode. The experimentally determined swivel-
ling frequency was 6:38 Hz while the simulations yielded 6:72 Hz ([8]) and 7:47 Hz
([1]). Therefore, both the numerical values were somewhat high but still in fairly
good agreement.
A comparison of the time-phase resolved front cylinder angle for one swivelling
period is given in Fig. 20. The qualitative time traces turned out to be very similar, in
special for [8]. Quantitatively, the maximum values showed differences. The simu-
lated value was about 26% greater than the measured value. For [1], the peak values
better matched the experimental results but some non-symmetry in the positive and
negative excursion were displayed.
Figure 21 shows the trailing edge xy-coordinates for one swivelling period.
In this graph the two numerical results almost coincide. The qualitative behavior

30

Experiment
20 from[8]
from[1]
front body angle [deg]

10

-10

-20

-30
0 60 120 180 240 300 360
time-phase angle [deg]

Fig. 20 Comparison of the front body angle for one period at 1:07 m/s.
406 J.P. Gomes and H. Lienhart

30
Experiment
20 from [8]
from [1]
10

y [mm]
0

-10

-20

-30
20 30 40 50 60 70 80
x [mm]

Fig. 21 Comparison of the x-y trailing edge coordinates for one period at 1:07 m/s.

2
u-comp, exp
1.5 v-comp, exp
velocity components [m/s]

u-comp, from [8]

v-comp, from [8]
1

0.5

-0.5

-1
0 60 120 180 240 300 360
time-phase angle [deg]

Fig. 22 Comparison of the flow velocity at point 1 for one period at 1:07 m/s.

was again well captured but the absolute values compared to the experiment dif-
fered noticeably. Here again, the simulated excursions were larger than the mea-
sured ones.
For comparing the flow field two monitor points were chosen. The first one
(point 1) was positioned shortly downstream behind the structure on the symme-
try plane at the coordinates (82 mm, 0 mm) and the second one (point 2) at the same
x-coordinate but off-axis at (82 mm, 40 mm). Especially the first monitor point rep-
resented a very selective choice. Its position in the center of the structure near wake
resulted in the fact that all the fluctuating velocities caused by the passage of the
rear mass and the vortices emerging behind the swivelling structure were of direct
impact. The second monitor point was positioned at the boundary of the wake region
and turned out to be less hard to predict. Numerical velocity data was available only
in [8] and the u-, and v-components are plotted together with the experimental data
in Fig. 22 and Fig. 23 for the two monitor points, respectively. The numerical and
Experimental Benchmark: Self-Excited Fluid-Structure Interaction Test Cases 407

2
u-comp, exp
v-comp, exp
1.5 u-comp, from [8]

velocity components [m/s]

v-comp, from [8]

0.5

-0.5

-1
0 60 120 180 240 300 360
time-phase angle [deg]

Fig. 23 Comparison of the flow velocity at point 2 for one period at 1:07 m/s.

experimental results matched very well at the second monitor point. At the first
monitor point, the velocity time traces were very similar in shape and amplitude but
a phase shift between the two results was displayed. A cross check with the posi-
tion data of the trailing edge proved that this phase shift was not present in the
trailing edge movement and, therefore, could not be the cause for the shift in the
velocity data.

5.2 Second swivelling-mode in laminar flows

At the inlet velocity of 1:45 m/s both numerical approaches resulted in a movement
of the same character than that observed during the experiments; the rear mass now
moved in the opposite direction with reference to the front cylinder. The frequency
of the self excited motion was measured to be 13:58 Hz and the simulations resulted
in values of 14:42 Hz ([8]) and 16:78 Hz ([1]), respectively. Hence, a comparably
good agreement between the measured and computed frequencies was obtained as
observed for the first swivelling-mode. Again the computed frequencies were higher
than the measured values being [8] closer to the experiments than [1].
In Fig. 24, the time-phase resolved front body angle for one period is plotted.
A comparison between Figs. 20 and 24 clearly shows the differences in the dynamic
response of the structure for the two swivelling-modes. In the second swivelling-
mode, the two numerical data sets resulted in distinctively different time traces.
While the data from [8] almost coincided with the experimental one, indicating that
both the amplitude and the pronounced non-linear characteristic were accurately
reproduced by the simulation, the data from [1] exhibited a strong non-symmetry.
An examination to figure out the reasons for that somewhat odd behaviour did
not lead to convincing conclusions yet and had to be left over to future investi-
gations. Figure 25 displays the traces of the trailing edge for one period at 1:45 m/s.
408 J.P. Gomes and H. Lienhart

30
Experiment
from [8]
20 from [1]

front body angle [deg]

10

-10

-20

-30
0 60 120 180 240 300 360
time-phase angle [deg]

Fig. 24 Comparison of the front body angle for one period at 1:45 m/s.

30
Experiment
20 from [8]
from [1]

10
y [mm]

-10

-20

-30
20 30 40 50 60 70 80
x [mm]

Fig. 25 Comparison of the x-y trailing edge coordinates for one period at 1:45 m/s.

Comparing the two different numerical and the experimental results yielded in a
similar evaluation than for the front body angle but the differences were less pro-
nounced. The data from [8] and the experimental results virtually collapsed while
the data from [1] was noticeably off and showed some amount of non-symmetry.
But compared with Fig. 21, both numerical results were in closer agreement to the
experiment for the second swivelling-mode.
As was to be expected from the very good agreement of the numerical results
from [8] with the experimental data in the behaviour of the swivelling motion of
the structure the same good match was found in the flow velocities. In Figs. 26
and 27 the u-, and v-components are plotted for the two monitor points. At the
second monitor point, positioned off-axis, the graphs almost collapsed. For the more
critical monitor point positioned on the symmetry axis the agreement is still good,
although a slightly too large value for the u-component was predicted. The phase-
shift observed in the first swivelling-mode was no longer present in the second.
Experimental Benchmark: Self-Excited Fluid-Structure Interaction Test Cases 409

1.5

velocity components [m/s]

1

0.5

-0.5

-1
u-comp, exp
v-comp, exp
-1.5 u-comp, from [8]
v-comp, from [8]
-2
0 60 120 180 240 300 360
time-phase angle [deg]

Fig. 26 Comparison of the flow velocity at point 1 for one period at 1:45 m/s.

1.5
velocity components [m/s]

0.5

-0.5

-1
u-comp, exp
v-comp, exp
-1.5 u-comp, from [8]
v-comp, from [8]
-2
0 60 120 180 240 300 360
time-phase angle [deg]

Fig. 27 Comparison of the flow velocity at point 2 for one period at 1:45 m/s.

5.3 Conclusions of the comparison of numerical

and experimental results

The comparison of the outcome of the numerical and experimental studies described
above showed that for both tested flow velocities the models reached a very repro-
ducible steady-state oscillation within a short time (a few cycles). The comparison
of the results performed on the basis of the time-phase resolved front body angle, the
trailing edge coordinates, and the flow velocities at two monitor points proved that
all the trends observed in the experiments were correctly captured by the numerical
approaches except for some non-symmetric front body oscillation predicted in [1]
for the second mode. For both self-excited swivelling-modes, the outcomes of the
measurements and the simulations were, in general, in good agreement. The numer-
ical simulations always resulted in slightly higher oscillation frequencies compared
with the corresponding measurements and also the amplitudes tended to be on the
larger side.
410 J.P. Gomes and H. Lienhart

Nevertheless, there is one point which deserves further attention and which shall
be shortly discussed in the following, namely that for the first swivelling-mode the
simulation results matched the measured values much less closely compared with
the second mode. Whereas for the second mode the data for the simulations and
the experiments almost collapsed, for the first mode significant discrepancies were
observed. There are two facts that made a difference between the experimental and
the numerical models. First, the numerical model had perfectly two-dimensional
boundary conditions, while in the experiment the span of the model was limited, and
hence there were test section side walls with boundary layers and gaps between the
model and the side walls. Second, the rotational degree of freedom was considered
frictionless in the numerical simulations, whereas the bearings in the experiment had
extremely small but non-zero friction. Both differences introduced higher damp-
ing into the experimental set-up compared with the computations, which in turn
would result in the tendencies observed. Nevertheless, the differences between the
experimental and the numerical outcomes for the first swivelling-mode can hardly
be attributed to these minor differences in the damping only.
A possible explanation for the different degrees of agreement between the mea-
surements and numerical results for the two swivelling-modes can be related to
the different fluid-structure excitation mechanisms identified. In the first mode of
oscillation, the relevant frequency for the excitation process [Instability-Induced
Excitation (IIE); [6]] is the first natural frequency of the structure (the struc-
ture alone, i.e. surrounded by vacuum). This gives rise to the assumption that the
dynamic response of the system is very sensitive to the damping coefficient of
the structure associated with its first natural frequency, which is small because of
the material properties of the flexible sheet. Therefore, the reaction of the system
is more sensitive to deviations in the damping of the structure which affect the
experimental results but which are not considered in the numerical model. On the
other hand, in the second mode of vibration, the relevant frequency for the exciting
process [Movement-Induced Excitation (MIE); [6]] is the natural frequency of the
coupled system structure and surrounding fluid. For this fully coupled mode, the
added damping introduced by the viscous fluid exceeds the damping of the structure
by orders of magnitude. This additional damping term is included in the numerical
model because it takes into account the properties of the liquid, specially the viscos-
ity. Hence, for the second swivelling-mode, and because of the presence of the added
damping, the dynamic response of the system becomes less sensitive to deviations
of the structural damping. These arguments can justify the differences observed
between the measurements and the numerical results for the first swivelling-mode
while at the same time the agreement between the results is almost perfect for the
second swivelling-mode.

Acknowledgements The present research project is part of the DFG Forschergruppe 493 -
Fluid-Struktur-Wechselwirkung: Modellierung, Simulation, Optimierung. The authors gratefully
acknowledge the financial support for their research work through the German Science Foundation
(DFG) - Germany - and Fundação para a Ciência e a Tecnologia (FCT) - Portugal. In addi-
tion the authors acknowledge the funding of the Erlangen Graduate School in Advanced Optical
Technologies (SAOT) by DFG in the framework of the German excellence initiative.
Experimental Benchmark: Self-Excited Fluid-Structure Interaction Test Cases 411

References

1. Geller S., Bettah M., Krafczyk M., Kollmannsberger S., Scholz D., Düster A., Rank E. (2010)
An explicit model for three-dimensional fluid-structure interaction using LBM and p-FEM. In
Bungartz H. J., Schäfer M. (eds.) Fluid-Structure Interaction, Part II, Lecture Notes Comput.
Sci. & Eng., LNCSE. Springer
2. Gomes J. P., Lienhart H. (2006) Time-phase resolved PIV/DMI measurements on two-
dimensional fluid-structure interaction problems. International symposium 13th Int Symp on
Applications of Laser Techniques to Fluid Mechanics. Calouste Gulbenkian Fundation, Lisbon,
16. - 29. June
3. Gomes J. P., Lienhart H. (2006) Experimental Study on a Fluid–Structure Interaction Refer-
ence Test Case. In: Bungartz H J, Schäfer M (eds.) Fluid–Structure Interaction, Lecture Notes
Comput. Sci. & Eng., LNCSE. Springer
4. Hartmann S., Meister A., Schäfer M., Turek S. (eds.) (2009) Fluid-structure interaction: Theory,
numerics, applications. Kassel University Press
5. Naudascher E., Rockwell D. (1980) Oscillator-model approach to the identification and
assessment of flow-induced vibrations in a system. Journal Hydraulics Research 18:59-82
6. Naudascher E., Rockwell D. (1980) Flow-induced Vibrations, IAHR Hydrodynamic Structures
Design Manual, Vol.7. Balkena
7. Sarpkaya T. (1978) Fluid forces on oscillating cylinders. Journal of the Waterway Port Coastal
and Ocean Division WW4:275-290
8. Schäfer M., Sternel D.C., Becker G., Pironkov P. (2010) Efficient Numerical Simulation and
Optimization of Fluid-Structure Interaction. In: Bungartz H. J., Schäfer M. (eds.) Fluid-Structure
Interaction, Part II, Lecture Notes Comput. Sci. & Eng., LNCSE. Springer
9. Souli M., Hamdouni A. (eds.) (2007) Fluid structure interaction: Industrial and Academic
Applications. European Journal of Computational Mechanics 16:303-548
•
Numerical Benchmarking of Fluid-Structure
Interaction: A Comparison of Different
Discretization and Solution Approaches

S. Turek, J. Hron, M. Razzaq, H. Wobker, and M. Schäfer

Abstract Comparative benchmark results for different solution methods for fluid-
structure interaction problems are given which have been developed as collaborative
project in the DFG Research Unit 493. The configuration consists of a laminar
incompressible channel flow around an elastic object. Based on this benchmark con-
figuration the numerical behavior of different approaches is analyzed exemplarily.
The methods considered range from decoupled approaches which combine Lattice
Boltzmann methods with hp-FEM techniques, up to strongly coupled and even fully
monolithic approaches which treat the fluid and structure simultaneously.

1 Introduction

In [6] specific benchmark problems have been defined in order to have a well-
founded basis for validating and also comparing different numerical methods and
code implementations for fluid-structure interaction (FSI) problems. In the present
paper numerical results obtained with different FSI solution approaches are pre-
sented and discussed. The methods involve different discretization schemes and
coupling mechanisms.
The FSI benchmark is based on the widely accepted flow around cylinder con-
figuration developed in [7] for incompressible laminar flow and on the setup in [8].
Similar to these configurations we consider the fluid to be incompressible and in
the laminar regime. The structure is allowed to be compressible, and its deforma-
tions are significant. The overall setup of the interaction problem is such that the

S. Turek, J. Hron, M. Razzaq, and H. Wobker

Institut für Angewandte Mathematik, LS III, TU Dortmund, Vogelpothsweg 87,
44227 Dortmund, Germany
e-mail: [email protected]
M. Schäfer
Institut für Numerische Berechnungsverfahren im Maschinenbau, TU Darmstadt,
Dolivostr. 15, 64293 Darmstadt, Germany
e-mail: [email protected]

H.-J. Bungartz et al. (eds.), Fluid Structure Interaction II, Lecture Notes 413
in Computational Science and Engineering 73, DOI 10.1007/978-3-642-14206-2 15,
c Springer-Verlag Berlin Heidelberg 2010
414 S. Turek et al.

solid object with an elastic part is submerged in a channel flow. Two cases with a
steady and an unsteady solution, respectively, are considered. Results for charac-
teristic physical quantities are provided, facilitating comparisons without the use of
complicated graphical evaluations. The provided benchmark computations are the
result of collaborative simulations which have been part of the research projects
inside of the DFG Research Unit 493 and which are presented as separate contribu-
tions to this volume (which also contains a more detailed description of the different
used FSI codes). The parameter settings and results can be downloaded from http://
featflow.de/en/benchmarks/cfdbenchmarking/fsi benchmark.html.

2 Benchmark configuration

We recall the benchmark configuration defined in [6]. It concerns the flow of an

incompressible Newtonian fluid interacting with an elastic solid. We denote by ˝tf
and ˝ts the domains occupied by the fluid and the solid, resp., at the time t  0. Let
t0 D ˝N tf \ ˝N ts be the part of the boundary where the elastic solid interacts with
the fluid.

2.1 Fluid properties

The fluid is considered to be Newtonian, incompressible and its state is described by

the velocity and pressure fields vf ; p f . The balance equations are

@vf
f C f .rvf /vf D div  f ; f
@t in ˝t : (1)
div vf D 0

The material constitutive equation is

T
 f D p f I C f  f .rvf C rvf /: (2)

The constant density of the fluid is f and the viscosity is denoted by  f .

2.2 Structure properties

The structure is assumed to be elastic and compressible. Its deformation is described

s
by the displacement us , with velocity field vs D @u
@t
. The balance equations are
Numerical Benchmarking in FSI 415

@vs
s C s .rvs /vs D div. s / C s g in ˝ts : (3)
@t
Written in the more common Lagrangian description, i.e. with respect to some fixed
reference (initial) state ˝ s , we have

@2 us
s D div.J  s F T / C s g in ˝ s ; (4)
@t 2

where F D I C rus is the deformation gradient tensor. For further details see for
example [1].
The material is specified by the Cauchy stress tensor  s or by the 2nd Piola-
Kirchhoff stress tensor S s D J F 1  s F T via the St. Venant-Kirchhoff constitu-
tive law
1
s D F .s .tr E /I C 2s E / F T ; (5)
J
S s D s .tr E /I C 2s E ; (6)

where E D 12 .F T F  I/ is the Green-St. Venant strain tensor.

The density of the structure in the undeformed configuration is s . The elasticity
of the material is characterized by the Poisson ratio  s ( s < 0:5 for a compress-
ible structure) and by the Young modulus E s . The alternative characterization is
described by the Lame coefficients s and s (the shear modulus):

s s .3s C 2s /
s D Es D ; (7)
2.s C s / .s C s /
Es s E s
s D s D : (8)
2.1 C  s / .1 C  s /.1  2 s /

2.3 Interaction conditions

The boundary conditions on the fluid-solid interface are assumed to be

 f n D  sn
on t0 ; (9)
vf D vs

where n is a unit normal vector to the interface t0 . This implies the no-slip
condition for the flow, and that the forces on the interface are in balance.
416 S. Turek et al.

H h
l
(0,0)
L

c A
h
r
l

Fig. 1 Computational domain and details of the structure part.

2.4 Domain definition

The problem domain, which is based on the 2D version of the well-known CFD
benchmark in [7], is illustrated in Fig. 1.
The geometry parameters are given as follows (all values in meters):
 The domain has length L D 2:5 and height H D 0:41.
 The circle center is positioned at C D .0:2; 0:2/ (measured from the left bottom
corner of the channel) and the radius is r D 0:05.
 The elastic structure bar has length l D 0:35 and height h D 0:02, the right
bottom corner is positioned at .0:6; 0:19/, and the left end is fully attached to the
fixed cylinder.
 The control point is A.t/, attached to the structure and moving in time with
A.0/ D .0:6; 0:2/.
The setting is intentionally non-symmetric (see [7]) to prevent the dependence of
the onset of any possible oscillation on the precision of the computation.

2.5 Boundary conditions

The following boundary conditions are prescribed:

 A parabolic velocity profile is prescribed at the left channel inflow

y.H  y/ 4:0
vf .0; y/ D 1:5UN 2
D 1:5UN y.0:41  y/; (10)
.H=2/ 0:1681

such that the mean inflow velocity is UN and the maximum of the inflow velocity
profile is 1:5UN .
 The outflow condition can be chosen by the user, for example stress free or do
nothing conditions. The outflow condition effectively prescribes some reference
value for the pressure variable p. While this value could be arbitrarily set in the
incompressible case, in the case of compressible structure this will have influence
Numerical Benchmarking in FSI 417

Table 1 Parameter settings FSI1 FSI3

for the FSI benchmarks.
s [103 kg=m3 ] 1 1
s 0.4 0.4
s [106 kg=ms2 ] 0:5 2:0
f [103 kg=m3 ] 1 1
 f [103 m2 =s] 1 1
UN [m=s] 0.2 2

on the stress and consequently the deformation of the solid. In this proposal, we
set the reference pressure at the outflow to have zero mean value.
 The no-slip condition is prescribed for the fluid on the other boundary parts, i.e.
top and bottom wall, circle and fluid-structure interface t0 .

2.6 Initial conditions

The suggested starting procedure for the non-steady tests was to use a smooth
increase of the velocity profile in time as
(
f vf .0; y/Œ1  cos. t=2/=2 if t < 2:0;
v .t; 0; y/ D (11)
f
v .0; y/ otherwise ;

where vf .0; y/ is the velocity profile given in Eq. (10).

2.7 Physical parameters

We consider physical parameters for two different test cases as indicated in Table 1.
Defining the Reynolds number by Re D 2r UN = f the two cases correspond to
Re=20 and Re=200. FSI1 is resulting in a steady state solution, while FSI3 results
in a periodic solution (the numbering refers to the original benchmark definition
in [6]).

3 Quantities for comparison

Comparisons will be done for fully developed flow, and particularly for one full
period of the oscillation with respect to the position of the point A.t/. The quantities
of interest are:
1. The displacements u1 .t/ and u2 .t/ in x- and y-direction of the point A.t/ at the
end of the beam structure (see Fig. 1).
418 S. Turek et al.

S1
S2

Fig. 2 Integration path S D S1 [ S2 for the force calculation.

2. Forces exerted by the fluid on the whole submerged body, i.e. lift and drag forces
acting on the cylinder and the beam structure together
Z Z Z
f f
.FD ; FL / DT
 n dS D  n dS C  f n dS;
S S1 S2

where S D S1 [ S2 (see Fig. 2) denotes the part of the circle being in contact
with the fluid and n is the outer unit normal vector to the integration path with
respect to the fluid domain.
Time dependent values are represented by mean value, amplitude, and frequency.
The mean value and the amplitude are computed from the last period of the oscil-
lations by taking the maximum and minimum values, the mean value is taken as
average of the max/min values, and the amplitude is the difference of the max/min
from the mean:
1
mean D .max C min/;
2
1
amplitude D .max  min/:
2

The frequency of the oscillations are computed either as 1=T with the period time T
or by using Fourier analysis on the periodic data and taking the lowest significant
frequency present in the spectrum.

4 FSI solution approaches

In the following, a short description of the different approaches, which have been
applied in the subsequent benchmark calculations, will be provided. A more detailed
description of the underlying methodology and the corresponding software tools is
given as separate contributions to this volume.

Method 1 (Schäfer):

This method involves an implicit partitioned solution approach (see [4, 5] for
details). It is realized based on the finite-volume multigrid flow solver FASTEST
Numerical Benchmarking in FSI 419

involving an ALE formulation (hexahedral control volumes, second-order in space

and time), the finite-element structural solver FEAP (bilinear brick elements, sec-
ond-order Newmark time discretization), and the coupling interface MpCCI. For
each time step the implicit solution procedure consists in the application of different
nested iteration processes for linearization, pressure-velocity coupling, and linear
system solving. These are linked by an iterative fluid-structure coupling procedure
with structural underrelaxation and displacement prediction.

Method 2 (Rannacher):

Two different solvers have been used in this project, both based on monolithic
variational formulations of the FSI problem: (a) one based on a unified Eulerian
framework for describing the fluid as well as the structure deformation (“interface
capturing”) which allows for large structural deformations and topology changes of
the flow domain, and (b) for comparison one based on the standard ALE approach
(“interface fitting”). The discretization is fully implicit and uses in case (a) the
second-order fractional-step- scheme and in case (b) the stabilized Crank-Nicolson
scheme in time. The spatial discretization is by a conforming finite element Galerkin
method on quadrilateral meshes with mixed global and zonal refinement using in
case (a) the “equal-order” Q1c =Q1c Stokes-element for velocity and pressure with
pressure and transport stabilization by “local projection”, and in case (b) the inf-sup-
stable Q2c =P1d Stokes-element. For the resulting nonlinear algebraic problems in
each time step a Newton-like iteration is used (combined with pseudo-time stepping
in the stationary test case). The linear subproblems are solved in fully coupled form
in case (a) by a multigrid methods with GMRES acceleration, block-ILU smoothing
and canonical grid transfers, and in case (b) by a direct solver (“UMFPACK”) which
allows for up to 106 unknowns in acceptable time.

Method 3 (Turek/Hron):

The applied FSI solver is a fully implicit, monolithic ALE-FEM approach. The
discretization is based on the classical (nonparametric) Q2/P1 finite element pair,
together with the Crank-Nicolson scheme for time stepping. The resulting nonlin-
ear systems are solved by a discrete Newton method, while the linear subsystems
are treated via Krylov-multigrid solvers with smoothing operators of local pressure
Schur complement/Vanka-like type with canonical grid transfer routines.

Method 4 (Breuer):

The method relies on a partitioned solution approach. For the fluid flow a finite-
volume scheme (FASTEST-3D) is used to discretize the (filtered) Navier–Stokes
equations for an incompressible fluid. The discretization is done on a curvilinear,
420 S. Turek et al.

blockstructured body-fitted grid with colocated variable arrangement by applying

standard schemes. Linear interpolation of the flow variables to the cell faces and
a midpoint rule approximation for the integrals is used to obtain a second-order
accurate central scheme. In order to account for displacement or deformation of
the structure the Arbitrary Lagrangian-Eulerian formulation is applied. A new parti-
tioned coupling method based on the predictor-corrector scheme often used for LES
is prefered. A strongly coupled but nevertheless still explicit time-stepping algo-
rithm results, which is very efficient in the LES–FSI context. For the prediction of
the deformation and displacement of the structure, the forces are transferred to com-
putational structure dynamics code, here the finite-element solver Carat provided by
the Chair of Structural Analysis of TU Munich. The response of the structure, i.e. the
displacements are transferred to the fluid solver. The coupling interface CoMA also
developed at TUM is used for this purpose. This interface is based on the Message-
Passing-Interface (MPI) and thus runs in parallel to the fluid and structure solver.
Presently, the grid adjustment on the fluid side is performed based on a transfinite
interpolation.

Method 5 (Krafczyk/Rank):

This method involves an explicit partitioned solution approach. It is realized with the
Lattice-Boltzmann flow solver VirtualFluids (VF) and the structural p-FEM solver
AdhoC. VirtualFluids is based on an Eulerian grid (cubic, hierarchical, graded FD
grids, second-order in space and time). AdhoC discretizes the equations of struc-
tural dynamics in space with a Bubnov-Galerkin method utilizing hierarchic ansatz
functions while the time domain is discretized with a second-order finite difference
scheme (Newmark or generalized alpha). The coupling interface MshPI manages
the data transfer on an interface mesh common to all involved solvers. For each
time step, the partitioned solution procedure consists of nested substeps. VirtualFlu-
ids sends loads via the interface mesh to AdhoC which generates displacements.
These, in turn, are then mapped to the fluid grid. The exchange of tractions and dis-
placements is performed only once per structural time step. As the time step for the
flow solver is typically smaller than the structural one, VF proceeds for several time
steps with an interpolated geometry. It is interesting to note that this setup proved to
be stable without the need to perform interfield iterations.

Method 6 (Wall):

For this benchmark scenario we have used our rather “classic” approach selected
from the different FSI approaches that we have developed in recent years for solv-
ing incompressible fluid flow coupled with large structural deformations. A main
ingredient is a strongly coupled, iterative staggered scheme based on [2,9]. The fluid
field uses an ALE formulation, stabilized, quadratic Q2Q2 elements and BDF2 time-
discretiyation. The fluid mesh is deformed based on linear-elastic material behavior.
Numerical Benchmarking in FSI 421

The structure is discretized by bi-linear quadrilateral elements with EAS formu-

lation and the generalized-˛ method for time discretization. Each field is solved
implicitly and an iterative procedure over the fields using Aitken relaxation ensures
convergence for the interface conditions at the new time step level n C 1.

Method 7 (Bletzinger):

The coupled problem is solved by a partitioned approach. Three independent soft-

ware components are combined: in-house codes CARAT++ and CoMA (Computer
Aided Research Analysis Tool, Coupling for Multiphysics Analysis) for structural
analysis, coupling control and data transfer between non-matching grids, as well
as OpenFOAM, an open source finite volume solver. The single field solvers use
individual, at the interface non-matching, grids. Implicit coupling schemes based
on fixed-point iterations with Aitken relaxation or a quasi-Newton method are
used. Different to all other groups the structure has been modelled as a shell (with
mid-surface in x-z-plane). Two different shell theories have been applied: a “clas-
sical”, 5-parameter, Reissner-Mindlin shell model neglecting normal stresses in
thickness direction as well as a 7-parameter, 3D solid shell model with a straight
thickness director as only kinematic assumption. The difference becomes obvious
in FSI1 where a longitudinal Poisson effect due to cross thickness normal stress
is evident for the horizontal deformation u1 D 1:85  105 (classical shell) or
u1 D 2:26  105 (solid shell), respectively. All other data match well and show
good results in all benchmarks. It is concluded, that shell finite element formulations
can be used effectively in FSI analyses and give correct results for structures up to
a moderate thickness.

5 Numerical results

The results of the benchmark computations are summarized in Tables 2 and 3 (units
are omitted). Indicated are the displacements u1 .A/ and u2 .A/ in x- and y-direction
of the point A as well as the drag and lift forces FD and FL . For the unsteady case
also the frequencies f1 and f2 obtained for the displacements u1 .A/ and u2 .A/,
respectively, are given. The number in the first column refers to the methods given in
the previous section. The column ”Unknowns” refers to the total number (in space),
i.e., the sum of unknowns for all velocity components, pressure, and displacement
components.
As a first result for the FSI1 benchmark, which leads to stationary displacement
of the attached elastic beam, it is obvious that all applied methods and codes can
approximate the same results, at least with decreasing mesh width.
For FSI3, the evaluation of the results is a little bit more difficult: First of all,
all schemes show the tendency to converge towards the (more or less) same solu-
tion values, at least for increasing mesh level. Although the applied FSI techniques
422 S. Turek et al.

Table 2 Results for steady benchmark FSI1.

Unknowns u1 .A/ Œ105  u2 .A/ Œ104  FD FL
1 82722 – – 14.2770 0.77200
322338 – – 14.2890 0.76900
2a 11250 2.4800 7.7800 – –
2b 19488 2.2821 8.1957 14.2382 0.76481
29512 2.2793 8.2201 14.2263 0.76420
51016 2.2733 8.1867 14.2408 0.76400
93992 2.2710 8.1702 14.2500 0.76392
179912 2.2700 8.1609 14.2561 0.76389
351720 2.2695 8.1556 14.2603 0.76388
3 19488 2.2871 8.1930 14.2736 0.76175
76672 2.2774 8.2042 14.2918 0.76305
304128 2.2732 8.2071 14.2948 0.76356
1211392 2.2716 8.2081 14.2949 0.76370
4835328 2.2708 8.2085 14.2945 0.76374
19320832 2.2705 8.2088 14.2943 0.76375
5 884736 2.1270 11.0800 14.3179 0.85491
3538944 2.1990 8.3370 14.3127 0.75138
14155776 2.2160 8.2010 14.3815 0.75170
6 7059 2.5396 8.8691 14.2800 0.73690
19991 2.2630 8.2935 14.2970 0.76687
77643 2.2676 8.2347 14.2940 0.76545
164262 2.2680 8.2310 14.2940 0.76487
7 217500 2.2640 8.2800 14.3510 0.76351

are very different w.r.t. discretization, solver and coupling mechanisms, the FSI3
benchmark setting proves to be a very valuable tool for numerical FSI benchmark-
ing, leading to grid independent results for the prescribed geometrical and parameter
settings.
However, also clear differences between the different approaches with regard to
accuracy are visible. Particularly for the drag and lift values, which lead to differ-
ences of up to order 50%, and also for the displacement values which are in the range
of 10% errors. A more detailed evaluation and also more rigorous comparisons w.r.t.
the ratio ‘accuracy vs. efficiency’ are therefore planned for the future.

6 Summary

We have presented 2D benchmark results for different numerical approaches for

fluid-structure interaction problems. These benchmarks have been developed as a
collaborative project in the DFG Research Unit 493. The configurations have been
carefully chosen and validated via extensive numerical tests (see also [3, 6]) with
Numerical Benchmarking in FSI 423

Table 3 Results for unsteady benchmark FSI3.

Unknowns
t u1 .A/ Œ103  u2 .A/ Œ103  FD FL f1 f2
1 61318 1.0e3 2.54 ˙ 2.41 1.45 ˙ 32.80 450.3 ˙ 23.51 -0.10 ˙ 143.0 10.90 5.13
237286 2.0e3 2.88 ˙ 2.73 1.53 ˙ 34.94 458.6 ˙ 27.18 2.08 ˙ 153.1 10.60 5.30
237286 1.0e3 2.87 ˙ 2.73 1.54 ˙ 34.94 458.6 ˙ 27.31 2.00 ˙ 153.3 10.34 5.91
237286 5.0e4 2.86 ˙ 2.72 1.53 ˙ 34.90 458.6 ˙ 27.27 2.01 ˙ 153.4 12.16 6.08
941158 1.0e3 2.91 ˙ 2.77 1.47 ˙ 35.26 459.9 ˙ 27.92 1.84 ˙ 157.7 11.63 4.98
2a 11250 5.0e3 2.48 ˙ 2.24 1.27 ˙ 36.50 – – 10.10 5.10
2b 7176 5.0e3 2.44 ˙ 2.32 1.02 ˙ 31.82 473.5 ˙ 56.97 8.08 ˙ 283.8 11.07 5.29
7176 2.0e3 2.48 ˙ 2.39 0.92 ˙ 32.81 471.3 ˙ 62.28 6.11 ˙ 298.6 10.73 5.35
7176 1.0e3 2.58 ˙ 2.49 0.94 ˙ 33.19 470.4 ˙ 64.02 4.65 ˙ 300.3 10.69 5.36
27744 5.0e3 2.43 ˙ 2.27 1.41 ˙ 31.73 483.7 ˙ 22.31 2.21 ˙ 149.0 10.53 5.37
27744 2.0e3 2.63 ˙ 2.61 1.46 ˙ 33.46 483.3 ˙ 24.48 2.08 ˙ 161.2 10.66 5.43
27744 1.0e3 2.80 ˙ 2.64 1.45 ˙ 34.12 483.0 ˙ 25.67 2.21 ˙ 165.3 10.75 5.41
42024 2.5e3 2.40 ˙ 2.26 1.39 ˙ 31.71 448.7 ˙ 21.16 1.84 ˙ 141.3 10.72 5.42
42024 1.0e3 2.53 ˙ 2.38 1.40 ˙ 32.49 449.7 ˙ 22.24 1.61 ˙ 142.8 10.77 5.44
42024 5.0e4 2.57 ˙ 2.42 1.42 ˙ 32.81 450.1 ˙ 22.49 1.49 ˙ 143.7 10.79 5.42
72696 2.5e3 2.64 ˙ 2.48 1.38 ˙ 33.25 451.1 ˙ 24.57 2.04 ˙ 150.6 10.73 5.38
72696 1.0e3 2.79 ˙ 2.62 1.28 ˙ 34.61 452.0 ˙ 25.78 1.91 ˙ 152.7 10.78 5.42
72696 5.0e4 2.84 ˙ 2.67 1.28 ˙ 34.61 452.4 ˙ 26.19 2.36 ˙ 152.7 10.84 5.42
3 19488 1.0e3 3.02 ˙ 2.83 1.41 ˙ 35.47 458.2 ˙ 28.32 2.41 ˙ 145.6 10.75 5.37
19488 5.0e4 3.02 ˙ 2.85 1.42 ˙ 35.63 458.7 ˙ 28.78 2.23 ˙ 146.0 10.75 5.37
19488 2.5e4 3.02 ˙ 2.85 1.32 ˙ 35.73 458.7 ˙ 28.80 2.23 ˙ 146.0 10.74 5.33
76672 1.0e3 2.78 ˙ 2.62 1.44 ˙ 34.36 459.1 ˙ 26.63 2.41 ˙ 151.3 10.93 5.46
76672 5.0e4 2.78 ˙ 2.62 1.44 ˙ 34.35 459.1 ˙ 26.62 2.39 ˙ 150.7 10.92 5.46
76672 2.5e4 2.77 ˙ 2.61 1.43 ˙ 34.43 459.1 ˙ 26.50 2.36 ˙ 149.9 10.93 5.46
304128 1.0e3 2.86 ˙ 2.70 1.45 ˙ 34.93 460.2 ˙ 27.65 2.47 ˙ 154.9 10.95 5.47
304128 5.0e4 2.86 ˙ 2.70 1.45 ˙ 34.90 460.2 ˙ 27.47 2.37 ˙ 153.8 10.92 5.46
304128 2.5e4 2.88 ˙ 2.72 1.47 ˙ 34.99 460.5 ˙ 27.74 2.50 ˙ 153.9 10.93 5.46
4 81120 9.0e5 5.18 ˙ 5.04 1.12 ˙ 45.10 477.0 ˙ 48.00 7.00 ˙ 223.0 10.14 4.99
324480 2.0e5 4.54 ˙ 4.34 1.50 ˙ 42.50 467.5 ˙ 39.50 16.20 ˙ 188.7 10.12 5.05
5 2480814 5.1e5 2.88 ˙ 2.71 1.48 ˙ 35.10 463.0 ˙ 31.30 1.81 ˙ 154.0 11.00 5.50
6 7059 5.0e4 1.60 ˙ 1.60 1.50 ˙ 25.90 525.0 ˙ 22.50 -0.55 ˙ 106.0 10.90 5.45
27147 5.0e4 2.00 ˙ 1.89 1.45 ˙ 29.00 434.0 ˙ 17.50 2.53 ˙ 88.6 10.60 5.30
7 271740 5.0e4 3.04 ˙ 2.87 1.55 ˙ 36.63 474.9 ˙ 28.12 3.86 ˙ 165.9 10.99 5.51

various CFD codes so that, as a main result, characteristic flow quantities can be
provided which allow a quantitative validation and comparison of different numer-
ical methods and software tools. As an extension, corresponding 3D simulations
are planned as well as the embedding into outer optimization tools (see http://jucri.
jyu.fi/?q=node/14 for a first attempt towards optimal control on the basis of the
presented FSI1 configuration).
424 S. Turek et al.

Acknowledgements The described benchmarks were developed in collaboration with G. Becker,

M. Heck, S. Yigit, M. Krafczyk, J. Tölke, S. Geller, H.-J. Bungartz, M. Brenk, R. Rannacher,
T. Dunne, T. Wick, W. Wall, A. Gerstenberger, P. Gamnitzer, E. Rank, A. Düster, S. Kollmanns-
berger, D. Scholz, M. Breuer, M. Münsch, G. De Nayer, H. Lienhart, J. Gomes, K.-U. Bletzinger,
A. Kupzok, and R. Wüchner.
This work has been supported by German Research Association (DFG), Reasearch Unit 493.

References

1. P. G. Ciarlet. Mathematical Elasticity. Volume I, Three-Dimensional Elasticity, volume 20 of

Studies in Mathematics and its Applications. Elsevier Science Publishers B.V., Amsterdam,
1988.
2. Ulrich Küttler and Wolfgang A. Wall. Fixed-point fluid-structure interaction solvers with
dynamic relaxation. Computational Mechanics, 43(1):61–72, 2008.
3. M. Razzaq, S. Turek, J. Hron, and J. F. Acker. Numerical simulation and benchmarking of
fluid-structure interaction with application to hemodynamics. In Fundamental Trends in Fluid-
Structure Interaction. World Scientific Publishing Co. Pte Ltd, 2010.
4. M. Schäfer, M. Heck, and M. Schäfer (eds.) Fluid-Structure Interaction: Modelling Simula-
tion Optimization 53 pp. 171-194. Springer Berlin Heidelberg Yigit, S.: An implicit partitoned
method for the numerical simulation of fluid-structure interaction. In: H. J. Bungartz. 2006.
5. D.C. Sternel, M. Schäfer, M. Heck, and S. Yigit. Efficiency and accuracy of fluid-structure
interaction simulations using an implicit partitioned approach. Computational Mechanics,
43(1):103–113, 2008.
6. S. Turek and J. Hron. Proposal for numerical benchmarking of fluid-structure interaction
between an elastic object and laminar incompressible flow. In H.-J. Bungartz and M. Schäfer,
editors, Fluid-Structure Interaction: Modelling, Simulation, Optimisation, LNCSE-53. Springer,
2006.
7. S. Turek and M. Schäfer. Benchmark computations of laminar flow around cylinder. In E.H.
Hirschel, editor, Flow Simulation with High-Performance Computers II, volume 52 of Notes on
Numerical Fluid Mechanics. Vieweg, 1996. co. F. Durst, E. Krause, R. Rannacher.
8. W. A. Wall and E. Ramm. Fluid-structure interaction based upon a stabilized (ALE) finite
element method. In S. Idelsohn, E. Oñate, and E. Dvorkin, editors, 4th World Congress on
Computational Mechanics: New Trends and Applications, Barcelona, 1998. CIMNE.
9. Wolfgang A. Wall, Daniel P. Mok, and Ekkehard Ramm. Partitioned analysis approach of
the transient coupled response of viscous fluids and flexible structures. In W. Wunderlich,
editor, Solids, Structures and Coupled Problems in Engineering, Proc. ECCM ’99, Munich,
August/September 1999.
Editorial Policy
1. Volumes in the following three categories will be published in LNCSE:

i) Research monographs
ii) Tutorials
iii) Conference proceedings

Those considering a book which might be suitable for the series are strongly advised to contact
the publisher or the series editors at an early stage.
2. Categories i) and ii). Tutorials are lecture notes typically arising via summer schools or
similar events, which are used to teach graduate students. These categories will be emphasized
by Lecture Notes in Computational Science and Engineering. Submissions by interdisci-
plinary teams of authors are encouraged. The goal is to report new developments – quickly,
informally, and in a way that will make them accessible to non-specialists. In the evaluation of
submissions timeliness of the work is an important criterion. Texts should be well-rounded,
well-written and reasonably self-contained. In most cases the work will contain results of
others as well as those of the author(s). In each case the author(s) should provide sufficient
motivation, examples, and applications. In this respect, Ph.D. theses will usually be deemed
unsuitable for the Lecture Notes series. Proposals for volumes in these categories should be
submitted either to one of the series editors or to Springer-Verlag, Heidelberg, and will be
refereed. A provisional judgement on the acceptability of a project can be based on partial
information about the work: a detailed outline describing the contents of each chapter, the
estimated length, a bibliography, and one or two sample chapters – or a first draft. A final deci-
sion whether to accept will rest on an evaluation of the completed work which should include
– at least 100 pages of text;
– a table of contents;
– an informative introduction perhaps with some historical remarks which should be acces-
sible to readers unfamiliar with the topic treated;
– a subject index.
3. Category iii). Conference proceedings will be considered for publication provided that they
are both of exceptional interest and devoted to a single topic. One (or more) expert partici-
pants will act as the scientific editor(s) of the volume. They select the papers which are suitable
for inclusion and have them individually refereed as for a journal. Papers not closely related to
the central topic are to be excluded. Organizers should contact the Editor for CSE at Springer
at the planning stage, see Addresses below.
In exceptional cases some other multi-author-volumes may be considered in this category.
4. Only works in English will be considered. For evaluation purposes, manuscripts may
be submitted in print or electronic form, in the latter case, preferably as pdf- or zipped
ps-files. Authors are requested to use the LaTeX style files available from Springer at http://
www. springer.com/authors/book+authors?SGWID=0-154102-12-417900-0.
For categories ii) and iii) we strongly recommend that all contributions in a volume be writ-
ten in the same LaTeX version, preferably LaTeX2e. Electronic material can be included if
appropriate. Please contact the publisher.
Careful preparation of the manuscripts will help keep production time short besides ensuring
satisfactory appearance of the finished book in print and online.
5. The following terms and conditions hold. Categories i), ii) and iii):
Authors receive 50 free copies of their book. No royalty is paid.
Volume editors receive a total of 50 free copies of their volume to be shared with authors, but
no royalties.
Authors and volume editors are entitled to a discount of 33.3 % on the price of Springer books
purchased for their personal use, if ordering directly from Springer.
6. Commitment to publish is made by letter of intent rather than by signing a formal contract.
Springer-Verlag secures the copyright for each volume.

Addresses:

Timothy J. Barth Risto M. Nieminen

NASA Ames Research Center Department of Applied Physics
NAS Division Aalto University School of Science
Moffett Field, CA 94035, USA and Technology
[email protected] 00076 Aalto, Finland
[email protected]
Michael Griebel
Dirk Roose
Institut für Numerische Simulation
Department of Computer Science
der Universität Bonn
Katholieke Universiteit Leuven
Wegelerstr. 6
Celestijnenlaan 200A
53115 Bonn, Germany
3001 Leuven-Heverlee, Belgium
[email protected]
[email protected]
David E. Keyes Tamar Schlick
Mathematical and Computer Sciences Department of Chemistry
and Engineering and Courant Institute
King Abdullah University of Science of Mathematical Sciences
and Technology New York University
P.O. Box 55455 251 Mercer Street
Jeddah 21534, Saudi Arabia New York, NY 10012, USA
[email protected] [email protected]

and Editor for Computational Science

and Engineering at Springer:
Department of Applied Physics Martin Peters
and Applied Mathematics Springer-Verlag
Columbia University Mathematics Editorial IV
500 W. 120 th Street Tiergartenstrasse 17
New York, NY 10027, USA 69121 Heidelberg, Germany
[email protected] [email protected]
Lecture Notes
in Computational Science
and Engineering

1. D. Funaro, Spectral Elements for Transport-Dominated Equations.

2. H.P. Langtangen, Computational Partial Differential Equations. Numerical Methods and Diffpack
Programming.
3. W. Hackbusch, G. Wittum (eds.), Multigrid Methods V.

4. P. Deuflhard, J. Hermans, B. Leimkuhler, A.E. Mark, S. Reich, R.D. Skeel (eds.), Computational
Molecular Dynamics: Challenges, Methods, Ideas.

5. D. Kröner, M. Ohlberger, C. Rohde (eds.), An Introduction to Recent Developments in Theory and

Numerics for Conservation Laws.

6. S. Turek, Efficient Solvers for Incompressible Flow Problems. An Algorithmic and Computational
Approach.

7. R. von Schwerin, Multi Body System SIMulation. Numerical Methods, Algorithms, and Software.

8. H.-J. Bungartz, F. Durst, C. Zenger (eds.), High Performance Scientific and Engineering Computing.
9. T.J. Barth, H. Deconinck (eds.), High-Order Methods for Computational Physics.

10. H.P. Langtangen, A.M. Bruaset, E. Quak (eds.), Advances in Software Tools for Scientific Computing.

11. B. Cockburn, G.E. Karniadakis, C.-W. Shu (eds.), Discontinuous Galerkin Methods. Theory, Com-
putation and Applications.

12. U. van Rienen, Numerical Methods in Computational Electrodynamics. Linear Systems in Practical
Applications.

13. B. Engquist, L. Johnsson, M. Hammill, F. Short (eds.), Simulation and Visualization on the Grid.

14. E. Dick, K. Riemslagh, J. Vierendeels (eds.), Multigrid Methods VI.

15. A. Frommer, T. Lippert, B. Medeke, K. Schilling (eds.), Numerical Challenges in Lattice Quantum
Chromodynamics.
16. J. Lang, Adaptive Multilevel Solution of Nonlinear Parabolic PDE Systems. Theory, Algorithm, and
Applications.

17. B.I. Wohlmuth, Discretization Methods and Iterative Solvers Based on Domain Decomposition.

18. U. van Rienen, M. Günther, D. Hecht (eds.), Scientific Computing in Electrical Engineering.

19. I. Babuška, P.G. Ciarlet, T. Miyoshi (eds.), Mathematical Modeling and Numerical Simulation in
Continuum Mechanics.

20. T.J. Barth, T. Chan, R. Haimes (eds.), Multiscale and Multiresolution Methods. Theory and
Applications.

21. M. Breuer, F. Durst, C. Zenger (eds.), High Performance Scientific and Engineering Computing.
22. K. Urban, Wavelets in Numerical Simulation. Problem Adapted Construction and Applications.

23. L.F. Pavarino, A. Toselli (eds.), Recent Developments in Domain Decomposition Methods.
24. T. Schlick, H.H. Gan (eds.), Computational Methods for Macromolecules: Challenges and
Applications.

25. T.J. Barth, H. Deconinck (eds.), Error Estimation and Adaptive Discretization Methods in
Computational Fluid Dynamics.

26. M. Griebel, M.A. Schweitzer (eds.), Meshfree Methods for Partial Differential Equations.

27. S. Müller, Adaptive Multiscale Schemes for Conservation Laws.

28. C. Carstensen, S. Funken, W. Hackbusch, R.H.W. Hoppe, P. Monk (eds.), Computational

Electromagnetics.

29. M.A. Schweitzer, A Parallel Multilevel Partition of Unity Method for Elliptic Partial Differential
Equations.
30. T. Biegler, O. Ghattas, M. Heinkenschloss, B. van Bloemen Waanders (eds.), Large-Scale PDE-
Constrained Optimization.

31. M. Ainsworth, P. Davies, D. Duncan, P. Martin, B. Rynne (eds.), Topics in Computational Wave
Propagation. Direct and Inverse Problems.

32. H. Emmerich, B. Nestler, M. Schreckenberg (eds.), Interface and Transport Dynamics. Computa-
tional Modelling.

33. H.P. Langtangen, A. Tveito (eds.), Advanced Topics in Computational Partial Differential
Equations. Numerical Methods and Diffpack Programming.

34. V. John, Large Eddy Simulation of Turbulent Incompressible Flows. Analytical and Numerical
Results for a Class of LES Models.

35. E. Bänsch (ed.), Challenges in Scientific Computing - CISC 2002.

36. B.N. Khoromskij, G. Wittum, Numerical Solution of Elliptic Differential Equations by Reduction to
the Interface.
37. A. Iske, Multiresolution Methods in Scattered Data Modelling.

38. S.-I. Niculescu, K. Gu (eds.), Advances in Time-Delay Systems.

39. S. Attinger, P. Koumoutsakos (eds.), Multiscale Modelling and Simulation.

40. R. Kornhuber, R. Hoppe, J. Périaux, O. Pironneau, O. Wildlund, J. Xu (eds.), Domain Decomposition

Methods in Science and Engineering.

41. T. Plewa, T. Linde, V.G. Weirs (eds.), Adaptive Mesh Refinement – Theory and Applications.

42. A. Schmidt, K.G. Siebert, Design of Adaptive Finite Element Software. The Finite Element Toolbox
ALBERTA.
43. M. Griebel, M.A. Schweitzer (eds.), Meshfree Methods for Partial Differential Equations II.

44. B. Engquist, P. Lötstedt, O. Runborg (eds.), Multiscale Methods in Science and Engineering.

45. P. Benner, V. Mehrmann, D.C. Sorensen (eds.), Dimension Reduction of Large-Scale Systems.

46. D. Kressner, Numerical Methods for General and Structured Eigenvalue Problems.

47. A. Boriçi, A. Frommer, B. Joó, A. Kennedy, B. Pendleton (eds.), QCD and Numerical Analysis III.

48. F. Graziani (ed.), Computational Methods in Transport.

49. B. Leimkuhler, C. Chipot, R. Elber, A. Laaksonen, A. Mark, T. Schlick, C. Schütte, R. Skeel (eds.),
New Algorithms for Macromolecular Simulation.
50. M. Bücker, G. Corliss, P. Hovland, U. Naumann, B. Norris (eds.), Automatic Differentiation:
Applications, Theory, and Implementations.

51. A.M. Bruaset, A. Tveito (eds.), Numerical Solution of Partial Differential Equations on Parallel
Computers.

52. K.H. Hoffmann, A. Meyer (eds.), Parallel Algorithms and Cluster Computing.

53. H.-J. Bungartz, M. Schäfer (eds.), Fluid-Structure Interaction.

54. J. Behrens, Adaptive Atmospheric Modeling.

55. O. Widlund, D. Keyes (eds.), Domain Decomposition Methods in Science and Engineering XVI.

56. S. Kassinos, C. Langer, G. Iaccarino, P. Moin (eds.), Complex Effects in Large Eddy Simulations.

57. M. Griebel, M.A Schweitzer (eds.), Meshfree Methods for Partial Differential Equations III.

58. A.N. Gorban, B. Kégl, D.C. Wunsch, A. Zinovyev (eds.), Principal Manifolds for Data Visualization
and Dimension Reduction.

59. H. Ammari (ed.), Modeling and Computations in Electromagnetics: A Volume Dedicated to Jean-
Claude Nédélec.

60. U. Langer, M. Discacciati, D. Keyes, O. Widlund, W. Zulehner (eds.), Domain Decomposition

Methods in Science and Engineering XVII.

61. T. Mathew, Domain Decomposition Methods for the Numerical Solution of Partial Differential
Equations.
62. F. Graziani (ed.), Computational Methods in Transport: Verification and Validation.

63. M. Bebendorf, Hierarchical Matrices. A Means to Efficiently Solve Elliptic Boundary Value
Problems.

64. C.H. Bischof, H.M. Bücker, P. Hovland, U. Naumann, J. Utke (eds.), Advances in Automatic
Differentiation.

65. M. Griebel, M.A. Schweitzer (eds.), Meshfree Methods for Partial Differential Equations IV.

66. B. Engquist, P. Lötstedt, O. Runborg (eds.), Multiscale Modeling and Simulation in Science.

67. I.H. Tuncer, Ü. Gülcat, D.R. Emerson, K. Matsuno (eds.), Parallel Computational Fluid Dynamics
2007.

68. S. Yip, T. Diaz de la Rubia (eds.), Scientific Modeling and Simulations.

69. A. Hegarty, N. Kopteva, E. O’Riordan, M. Stynes (eds.), BAIL 2008 – Boundary and Interior Layers.

70. M. Bercovier, M.J. Gander, R. Kornhuber, O. Widlund (eds.), Domain Decomposition Methods in
Science and Engineering XVIII.

71. B. Koren, C. Vuik (eds.), Advanced Computational Methods in Science and Engineering.

72. M. Peters (ed.), Computational Fluid Dynamics for Sport Simulation.

73. H.-J. Bungartz, M. Mehl, M. Schäfer (eds.), Fluid Structure Interaction II - Modelling, Simulation,
Optimization.

For further information on these books please have a look at our mathematics catalogue at the following
URL: www.springer.com/series/3527
Monographs in Computational Science
and Engineering

1. J. Sundnes, G.T. Lines, X. Cai, B.F. Nielsen, K.-A. Mardal, A. Tveito, Computing the Electrical
Activity in the Heart.

For further information on this book, please have a look at our mathematics catalogue at the following
URL: www.springer.com/series/7417

Texts in Computational Science

and Engineering

1. H. P. Langtangen, Computational Partial Differential Equations. Numerical Methods and Diffpack

Programming. 2nd Edition

2. A. Quarteroni, F. Saleri, P. Gervasio, Scientific Computing with MATLAB and Octave. 3rd Edition

3. H. P. Langtangen, Python Scripting for Computational Science. 3rd Edition

4. H. Gardner, G. Manduchi, Design Patterns for e-Science.

5. M. Griebel, S. Knapek, G. Zumbusch, Numerical Simulation in Molecular Dynamics.

6. H. P. Langtangen, A Primer on Scientific Programming with Python.

7. A. Tveito, H. P. Langtangen, B. F. Nielsen, X. Cai, Elements of Scientific Computing.

For further information on these books please have a look at our mathematics catalogue at the following
URL: www.springer.com/series/5151