See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/308313269

Model Order Reduction Methods for Explicit FEM

Conference Paper · May 2016

CITATION READS
1 1,327

4 authors, including:

Fabian Duddeck
Technische Universität München
108 PUBLICATIONS   566 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Wingbox optimisation View project

Cooperative Design View project

All content following this page was uploaded by Fabian Duddeck on 19 September 2016.

The user has requested enhancement of the downloaded file.

 Model Order Reduction Methods for Explicit FEM
Ali Cagatay Cobanoglua*, Simon Mößnera,b, Majid Hojjatb and Fabian Duddecka
a
Technische Universität München, Arcisstr. 21, 80333 Munich, Germany;
b
BMW Group, Knorrstraße 147, 80788 Munich, Germany

Abstract: Abaqus/Explicit is a well-established and widely used FEM solver for crash and
pedestrian safety simulations. However, due to the large number of degrees of freedom, simulation
time is still a limiting factor especially in context of structural optimization. In typical crash
simulation, a large portion of the model undergoes only elastic deformation. Hence, model order
reduction (MOR) methods can bring a significant decrease in the computational time. While
Abaqus/Standard already offers several reduction methods, MOR is currently not applicable for
Abaqus/Explicit. The purpose of this work is to enable MOR for explicit finite element models by
use of superelements. Reduced mass and stiffness matrices are generated by Abaqus/Standard and
transferred to the explicit solver through a VUEL subroutine. The method is applied successfully
to low speed vehicle crash simulations. The achieved results show a significant gain in
computational time and underline the potential of MOR for Abaqus/Explicit.

Keywords: explicit FEM, model order reduction, Guyan reduction, superelement,

crashworthiness, low-speed crash.

1. Introduction
Explicit FEM is an effective and efficient approach for solving complex three-dimensional
dynamic problems. Especially, it is appropriate for the simulation of impact loads with multiple
contacts thanks to its small time step size. In advanced commercial explicit FEM software, such as
Abaqus/Explicit, the computational time is dependent on both the number of elements and the
element size normally. When the critical element size is kept constant in the model, the
computational time scales almost linearly with the number of degrees of freedom, DOF. This
motivates the development of methods to simplify the numerical models decreasing simulation
time by, for example, reducing the number of DOF.

Model Order Reduction (MOR) has been widely used in implicit FEM codes for decades. There
have been various MOR methods proposed, which obtain an improvement in the computation time
by a slight loss in the accuracy of the results. Due to the increasing attention towards explicit FEM
in the last decade, a number of developed MOR techniques have been implemented to some
explicit FEM codes as well, some of which significantly decrease the computational cost (Faucher,
2003; Flídrová, 2010; Maker, 2002).

Impact simulation improving vehicular crashworthiness is a prominent application field for

explicit FEM. Being a highly nonlinear phenomenon, optimization of car crash is currently mainly
2016 Science in the Age of Experience 1
http://www.3ds.com/events/science-in-the-age-of-experience
done by use of zero order algorithms, which require multiple executions of the crash simulation,
e.g. (Duddeck, 2008). Typical FEM crash models have up to several million nodes and DOF;
hence, an enormous amount of data has to be handled during the simulation. In low-speed crash
simulations, only a small portion of the car models undergoes nonlinear plastic deformation. The
rest of the car body structure, which is away from the impactor, deforms only in the linear elastic
range. That part of the car model can be reduced to a limited number of DOF by use of MOR,
without losing accuracy. This can significantly reduce the computational cost roughly with a
comparable ratio as obtained between the number of DOF in the full and the reduced model.
Currently, MOR is not available for Abaqus/Explicit. On the contrary, Abaqus/Standard is capable
of creating the reduced models with different types of MOR methods. Here, the reduced part of the
model is called superelement. The mechanical properties of superelements created by
Abaqus/Standard, like mass, stiffness and damping matrices, can be used elsewhere and
independently of Abaqus/Standard. This helps to implement the superelements in Abaqus/Explicit
without the need to involve Abaqus/Standard after the reduction procedure. In this work, a VUEL
user subroutine (Abaqus, 2016) is developed, which enables the use of superelements in
Abaqus/Explicit. In the following section, the steps of this implementation are discussed in detail.

2. Superelement in Abaqus/Explicit
Using Abaqus/Standard for MOR, physical coordinates of a full-order model are replaced by
physical or modal coordinates in the reduced-order model. Here, we focus on reduction of the
model to the DOF of a small number of nodes, referred to as the interface nodes.

Figure 1. Superelement

In Figure 1, the full-order model (left) has m nodes and with each n degrees of freedom, hence in
total m×n DOF. In Equation (1), the equation of motion of the full model is represented as:
𝑴𝒖̈ + 𝑪𝒖̇ + 𝑲𝒖 = 𝒇𝒆𝒙𝒕 (1)
M, C, and K are the mass, damping and stiffness matrices, respectively. u is the displacement
vector and fext are the external forces. The reduced-order model (right) has k nodes, and therefore
k×n degrees of freedom. The structural matrices and the external load vector of the reduced model
are:
𝑴𝒓𝒆𝒅 = 𝑻 𝑴 𝑻𝑻 , (2)

2 2016 Science in the Age of Experience

http://www.3ds.com/events/science-in-the-age-of-experience
 𝑪𝒓𝒆𝒅 = 𝑻 𝑪 𝑻𝑻 , (3)
𝑲𝒓𝒆𝒅 = 𝑻 𝑲 𝑻𝑻 , (4)
𝒇𝒆𝒙𝒕
𝒓𝒆𝒅 = 𝑻𝒇 𝒆𝒙𝒕
. (5)
with 𝑻 being the transformation matrix defined through the reduction algorithm. The role of 𝑻 is to
reduce the full-model matrices into reduced ones and to recover the solution in the full model with
𝒖=𝑻𝒛 (6)
After the reduction, the equation of motion is:
𝑴𝒓𝒆𝒅 𝒛̈ + 𝑪𝒓𝒆𝒅 𝒛̇ + 𝑲𝒓𝒆𝒅 𝒛 = 𝒇𝒆𝒙𝒕
𝒓𝒆𝒅 (7)
Figure 2 illustrates the steps to generate a superelement for Abaqus/Explicit. Firstly, the portion of
the model, which undergoes only linear elastic deformation, is identified to be reduced as
superelement. This part of the model is then modeled in Abaqus/Standard and the corresponding
reduced matrices are exported. The obtained data is given to VUEL subroutine as superelement.
This superelement is defined as a normal user element inside the input deck of the main analysis.
In a final step, the created reduced-order model is run in Abaqus/Explicit: The implemented
VUEL subroutine is responsible for all element calculations of the superelement during the
analysis.

2016 Science in the Age of Experience 3

http://www.3ds.com/events/science-in-the-age-of-experience
 Figure 2. Flow chart of the superelement implementation to Abaqus/Explicit
(here for the undamped case)

3. Numerical Application
The method and the implementation were tested by small 2D examples with an ideal matching of
displacement, velocity and acceleration and then finally validated by a larger industrial test case.
The former examples are not presented here for the sake of briefness. Hence, the potential of the
method is demonstrated in this section via a body-in-white car model for an RCAR crash load case
(Research Council for Automobile Repairs, 1999).
4 2016 Science in the Age of Experience
http://www.3ds.com/events/science-in-the-age-of-experience
 Figure 3. Full-order model of body-in-white

In Figure 3, a full body-in-white model and the RCAR barrier placed in front of the car are shown.
Figure 4 shows on the right the part of the model which is assumed to deform nonlinearly and
plastically. The rest of the structure is considered to show linear elastic behavior throughout the
total crash simulation and can be reduced therefore by MOR. This separation is taken here as an
example to illustrate the method. An ideal approach should distinguish linear and nonlinear in a
more detailed and automated manner considering the number of interface nodes.
The linear-elastic part is modeled in Abaqus/Standard and reduced with the Guyan method
(Guyan, 1965). Degrees of freedom on the cut between linear and nonlinear areas, as shown in
Figure 4, are selected as retained DOF. In Table 1, the numbers of nodes and DOF in the full- and
reduced-order models are documented. Under the ideal assumption that the computation time
scales down linearly with the decreased number of DOF by MOR, the reduced model should run
ideally three times faster than the full model (4.5/1.5=3). Of course, this is an upper limit estimate
as the superelement operations and communications require a considerable amount of time.

Table 1. No. of nodes and DOF in full- and reduced-order model

No. of nodes in full-order model ~ 750,000

No. of nodes in reduced-order model ~ 240,000

No. of DOF in full-order model ~ 4.5 million

No. of DOF in reduced-order model ~ 1.5 million

2016 Science in the Age of Experience 5

http://www.3ds.com/events/science-in-the-age-of-experience
3.1 Results
The deformation of the full- and reduced-order model against the crash barrier is compared in
Figure 4. The general deformation patterns match very well. However, more plastic deformation is
observed in the crash box of the reduced model as seen in the detailed view of Figure 4.
Differences can be explained by the utilization of the simple Guyan reduction (Guyan, 1965) and
the selection of the interface nodes. Using Abaqus/Explicit version 6.13-5 with 1 CPU and
double=both precision, the reduced model decreased the computational cost by approximately
factor 2.1.

Figure 4. Comparison of full- and reduced-order model with detailed view of the
crash system; left: full model; right: reduced model

6 2016 Science in the Age of Experience

http://www.3ds.com/events/science-in-the-age-of-experience
 4. Conclusion
An implementation of MOR in Abaqus/Explicit was realized here to reduce the computational cost
of low-speed crash simulations. By taking superelement data outputs from Abaqus/Standard and
integrating it through a VUEL user-subroutine in Abaqus/Explicit, a significant decrease in the
computation time was obtained. These results show that superelements can improve the simulation
efficiency with very small loss in the accuracy and have great potential for low-speed-crash
simulation and other comparable impact situations simulated with Abaqus/Explicit.

5. References
1. Abaqus Users Subroutines Reference Guide, Version 6.14-1, Dassault Systèmes Simulia
Corp., Providence, RI.
2. Duddeck, F., “Multidisciplinary Optimization of Car Bodies”Structural and
Multidisciplinary Optimization, 2008, 35.4: 375-389.
3. Faucher, V., and Combescure, A., “A Time and Space Mortar Method for Coupling
Linear Modal Subdomains and Non-linear Subdomains in Explicit Structural Dynamics”.
Computer Methods in Applied Mechanics and Engineering, 2003, 192.5: 509-533.
4. Flídrová, K., Lenoir, D., Vasseur, N., and Jézéquel, L., “Modelization by Superelements
with Contact Management in Explicit Car Crash Simulations”1st Joint Int. Conf. on
Multibody System Dynamics IMSD 2010, Lappeenranta, Finland ,2010.
5. Guyan, R. J., “Reduction of Stiffness and Mass Matrices” AIAA journal, 1965, 3.2:
380-380.
6. Maker, B. N., and Benson, D. J., “Modal Methods for Transient Dynamics Analysis in
LS-Dyna” 7th Int. LS-Dyna Conf., Detroit, USA, 2002.
7. Research Council for Automobile Repairs, “The Procedure for Conduction a Low Speed
Crash 15 km/h Offset Insurance Crash Test to Determine the Damageability and
Reparability Features of Motor Vehicles”, 1999.

2016 Science in the Age of Experience 7

http://www.3ds.com/events/science-in-the-age-of-experience

View publication stats