Journal of Computational Physics 279 (2014) 127–144

Contents lists available at ScienceDirect

Journal of Computational Physics

www.elsevier.com/locate/jcp

Compressive VOF method with skewness correction to capture

sharp interfaces on arbitrary meshes
Fabian Denner ∗ , Berend G.M. van Wachem
Thermofluids Division, Department of Mechanical Engineering, Imperial College London, Exhibition Road, London, SW7 2AZ, United Kingdom

a r t i c l e i n f o a b s t r a c t

Article history: The accurate and efficient modelling of two-phase flows is at present mostly limited to
Received 11 March 2014 structured, unskewed meshes, due to the additional topological and numerical complexity
Received in revised form 1 September 2014 of arbitrary, unstructured meshes. Compressive VOF methods which discretize the interface
Accepted 3 September 2014
advection with algebraic differencing schemes are computationally efficient and inherently
Available online 9 September 2014
applicable to arbitrary meshes. However, compressive VOF methods evidently suffer
Keywords: severely from numerical diffusion on meshes with topological skewness. In this paper
Volume of fluid (VOF) method we present a compressive VOF method using a state-of-the-art donor–acceptor advection
Two-phase flows scheme which includes novel modifications to substantially reduce numerical diffusion
Interface capturing on arbitrary meshes without adding computational complexity. The new methodology
Advection scheme accurately captures evolving interfaces on any arbitrary, non-overlapping mesh and
Unstructured meshes conserves mass within the limits of the applied solver tolerance. A thorough validation
of the presented methods is conducted, examining the pure advection of the interface
indicator function as well as the application to evolving interfaces with surface tension.
Crucially, the results on equidistant Cartesian and arbitrary tetrahedral meshes are shown
to be comparable and accurate.
© 2014 Published by Elsevier Inc.

1. Introduction

The accurate representation and advection of the material interface between two fluids is essential for the predictive
quality of immiscible two-phase flow simulations. However, the interface is infinitesimally thin with respect to continuum
mechanics and, thus, represents a considerable challenge for finite volume and finite element frameworks. The Volume of
Fluid (VOF) [1] method is among the most widely used methods to represent two incompressible, immiscible fluids. Two
fundamental types of VOF methods may be distinguished: compressive methods and geometric methods. Both types of VOF
methods capture the discrete volume fraction of each phase and transport it based on the underlying fluid. Compressive
VOF methods discretize the partial differential equation describing the transport of the volume fraction of each phase using
algebraic differencing schemes [2–5]. However, the temporal and spatial discretisation requires special care in order to keep
the interface sharp and without distortion. Using a geometric method, an explicit representation of the interface is advected,
reconstructed from the VOF volume fraction field. The most notable state-of-the-art reconstruction methods are piecewise
linear (PLIC) methods [6–12] and parabolic reconstruction methods [13]. The explicit interface is geometrically fitted to the
VOF volume fraction field and advected in an Eulerian, Lagrangian or mixed Eulerian–Lagrangian fashion [14].

* Corresponding author. Tel.: +44 7500 060 053.

E-mail address: [email protected] (F. Denner).

http://dx.doi.org/10.1016/j.jcp.2014.09.002
0021-9991/© 2014 Published by Elsevier Inc.
128 F. Denner, B.G.M. van Wachem / Journal of Computational Physics 279 (2014) 127–144

The major advantages of compressive VOF methods compared to geometric VOF methods are the straightforward imple-
mentation on arbitrary meshes and the computational efficiency. Geometric methods advect the interface very accurately,
but the three-dimensional reconstruction of the interface requires significantly larger computational resources compared to
a two-dimensional reconstruction [12], and even more so on arbitrary meshes, due to the added complexity of the geo-
metric primitives used to reconstruct the interface. Hence, the implementation of geometric VOF methods is considerably
more complex than compressive VOF methods, the required computational effort is higher and, as a results, geometric VOF
methods are at present, apart from a few recently proposed exceptions [14–17], exclusively available on structured meshes.
Despite the greater flexibility and wider applicability of compressive VOF methods, geometric VOF methods such as PLIC
methods are in current literature widely preferred over compressive VOF methods, due to their claimed higher accuracy.
However, recent studies by Gopala and van Wachem [18], Park et al. [19] and Denner et al. [20] show that compressive VOF
methods are generally capable of advecting sharp, evolving interfaces with similar accuracy as state-of-the-art PLIC methods.
A major challenge for compressive VOF methods is to retain the shape and sharpness of the interface. The transient
advection of the interface is typically discretized using a second-order temporal discretisation scheme, such as the Crank–
Nicolson scheme or the Second-Order Backward Euler scheme [4,21], since first-order schemes are too diffusive to maintain
the sharpness of the interface. Studies by Ubbink [21], Jasak [22] and Darwish [23] comprehensively demonstrate that the
First-Order Backward Euler scheme as well as the First-Order Forward Euler scheme distort the shape of a circular interface
advected at a constant oblique velocity on an equidistant Cartesian mesh. The Crank–Nicolson scheme, on the other hand,
is able to preserve the shape of the circular interface of the same test case. In [24], Moukalled and Darwish propose a class
of temporal discretisation schemes, which switch between a compressive and a high-resolution scheme based on the angle
between the interface normal vector and the velocity vector. Moukalled and Darwish [24] use the Second-Order Backward
Euler scheme and a compressive Euler scheme, reporting better results at medium and high Courant numbers than with
other commonly used schemes, such as the Crank–Nicolson scheme.
Similarly, the careful spatial discretisation of the interface advection is essential to preserve a sharp interface. Low-order
advection schemes are not suitable, as they lead to significant smearing of the interface, whereas high-order schemes result
in numerical oscillations and wrinkling of the interface. Compressive VOF methods, therefore, use specifically designed
spatial advection schemes based on a donor–acceptor approach. A donor (upwind) and an acceptor (downwind) cell are
assigned for every cell face with respect to the underlying flow field. Based on the angle between the local interface
normal vector and the normal vector of the cell face, donor–acceptor schemes blend between compressive downwind and
diffusive upwind schemes. Following a donor–acceptor approach, a number of spatial discretisation schemes specifically
designed for compressive VOF methods have been proposed in recent years, most notably the SURFER scheme [2], the
CICSAM scheme [3], the STACS scheme [4], the HRIC scheme [5] and the HiRAC scheme [25]. The application of these
schemes to arbitrary meshes is typically straightforward but is also associated with discretisation errors induced by the
mesh orientation and arrangement as well as the construction of an artificial upwind node for each cell face. Zhang et al.
[26] examined the influence of the upwind node extrapolation and reported a profound impact on the accuracy of the tested
advection schemes. Denner and van Wachem [27] observed considerable numerical diffusion simulating the buoyancy-driven
rise of a bubble on a tetrahedral mesh, suggesting the numerical diffusion is a results of mesh skewness. A more detailed
examination of this issue by Denner [28] supports the assumption that mesh skewness is a major source of error with
respect to the interface advection. Ubbink [21] put forward two additional possible reasons for numerical diffusion of the
VOF volume fraction using donor–acceptor schemes. First, the extrapolation of the upwind node, required to determine
the advection of the colour function on unstructured meshes, and, second, the implicit assumption that an interface is
intersecting a cell face if both adjacent cells contain an interface. Numerical diffusion as a result of the applied differencing
schemes has previously been extensively discussed in the literature [4,21,29,30]. However, numerical diffusion induced by
mesh skewness has, so far, not been the focus of research in the context of interface advection. This is surprising, as
mesh skewness is the norm on arbitrary meshes, e.g. boundary-fitted hexahedral meshes or tetrahedral meshes, and leads
to a diffusion-like error that severely affects the accuracy of spatial interpolation, as for instance the linear interpolation of
face-centred values from adjacent cell-centred values required to discretize the spatial advection of the VOF volume fraction.
In this article we present a revised compressive VOF framework, based on the widely used CICSAM scheme, including
modifications aiming at mitigating numerical diffusion due to mesh skewness. In particular, an implicit skewness correction
is proposed which substantially decreases numerical diffusion on arbitrary meshes without notably affecting the compu-
tational complexity and mass-conservation of the VOF algorithm. It is worth mentioning that the proposed modifications
are not limited to the CICSAM scheme but can also be applied to any other donor–acceptor scheme (e.g. HRIC or STACS).
The comprehensive and thorough validation demonstrates the capabilities and accuracy of the presented methodology and
the results on Cartesian and tetrahedral meshes are shown to be comparable. Moreover, we show that our framework
is mass-conserving within the limits of the applied solver tolerance. The presented methodology including the proposed
modifications are of significant interest to the two-phase flow modelling community as it improves the predictive quality of
interfacial flow simulations considerably and, thus, makes it feasible for applications with complex geometries which cannot
be represented by general structured meshes.
The article is structured as follows. In Section 2 the governing equations are outlined and Section 3 briefly explains
the applied numerical framework. The discretisation of the applied compressive VOF method is discussed in Section 4.
Subsequently, Section 5 is concerned with numerical diffusion caused by mesh skewness and introduces the proposed
modifications to mitigate numerical diffusion. The presented compressive VOF methodology including the proposed modifi-
 F. Denner, B.G.M. van Wachem / Journal of Computational Physics 279 (2014) 127–144 129

cations are comprehensively tested and validated in Section 6. The findings are summarised and the article is concluded in
Section 7.

2. Governing equations

The Volume of Fluid (VOF) method [1] is used to determine the position and behaviour of the interface between two
incompressible, isothermal and immiscible fluids. The VOF method applies a volume fraction γ , also called colour function,
to every mesh cell, representing the local volume fraction as

γ (x, t ) = 0 fluid A (1)
1 fluid B .
Thus, every mesh cell where the colour function value is between 0 and 1 contains an interface. As the interface is trans-
ported with the underlying flow field, the colour function is advected by the hyperbolic equation
∂γ ∂γ
+ ui = 0, (2)
∂t ∂ xi
where u is the velocity of the flow and t represents time. The density ρ and viscosity μ are defined based on a linear
relationship with the colour function γ , given as

ρ = ρ A (1 − γ ) + ρ B γ (3)
μ = μ A (1 − γ ) + μ B γ , (4)

where subscripts A and B denote either of the fluids.

The motion of isothermal fluids is governed by the conservation of momentum and the conservation of mass. The
conservation of momentum for a Newtonian fluid is described by the momentum equations as
    
∂ ui ∂ u j ui ∂p ∂ ∂ ui ∂u j
ρ + =− + μ + + ρ g i + f s ,i , (5)
∂t ∂xj ∂ xi ∂xj ∂xj ∂ xi
where p represents pressure, g is the gravitational acceleration and f s is the volumetric force due to surface tension. The
conservation of mass is generally defined as

∂ ρ ∂(ρ u i ) ∂ ρ ∂ρ ∂ ui
+ = + ui +ρ = 0. (6)
∂t ∂ xi ∂t ∂ xi ∂ xi
Assuming the fluid is incompressible and isothermal, as considered in this study, the velocity field in each phase is
divergence-free, given by
∂ ui
= 0. (7)
∂ xi
Hence, Eq. (6) becomes
Dρ ∂ρ ∂ρ
= + ui = 0, (8)
Dt ∂t ∂ xi
which is equivalent to assuming the density of an infinitesimally small fluid particle does not change over time. As explained
in Section 3, continuity is implicitly enforced by the deployed numerical framework within the limits of the numerical solver
tolerance [28]. Hence, the spatial advection term of Eq. (8) is conservative [31] and, as a result, so is Eq. (8) as a whole. Thus,
given an isothermal, incompressible two-phase flow without phase changes and given the linear relationship of density and
colour function defined in Eq. (3), i.e. ρ = ρ (γ ), the mass of each phase is conserved if the VOF advection equation, Eq. (2),
is satisfied.

3. Numerical framework

The numerical framework follows a coupled implicit approach for the fluid flow with a segregated interface advection,
as presented by Denner and van Wachem [27]. A linear equation system describing the flow is constructed, containing the
discretized momentum equations as well as a fourth equation, which provides an additional relationship between pressure
and velocity to close the equation system. The fourth equation of the flow equation system is constructed based on the
continuity equation, given in Eq. (7), and is defined as

∂ u i  1  1  n
 ≈ (u f · n f ) A f ≈ u f A f = 0, (9)
∂ xi P VP VP
f f
130 F. Denner, B.G.M. van Wachem / Journal of Computational Physics 279 (2014) 127–144

where f denotes all bounding faces of a given mesh cell P , n f is the outward-pointing unit normal vector of face f and
A f is the area of face f . The advecting velocity un at cell faces is devised based on the momentum interpolation method as
detailed in [27]. The fully assembled equation system to describe the fluid flow then follows as
⎛ ⎞ ⎛ ⎞
A ux 0 0 A xp u
⎜ 0 Av
y
0 Ap ⎟
y
⎜ v ⎟
⎜ ⎟
⎝ 0 z ⎠ · ⎝ w ⎠ = b, (10)
0 A zw Ap
A cu A cv A cw A cp p
     
A φ

where A is the coefficient matrix, φ the solution vector and b the right-hand side vector. Inside matrix A, A ij represents
the coefficient submatrix for primitive variable j of the i-th equation, with superscripts x, y and z denoting the three
momentum equations and c denoting the fourth equation derived using the momentum interpolation method. Solution
vector φ is constituted by the solution subvectors of the three velocity components u, v and w, and pressure p. The
momentum equations, Eq. (5), are discretized using the Second-Order Backward Euler scheme for the transient term and
a central differencing scheme for the advection term. If applicable, the volumetric force due to surface tension f s acting at
the interface is discretized using the continuum surface force (CSF) model [32].
Before the new flow field is computed at any given time instant, the colour function is advected as defined in Eq. (2)
based on the flow field resulting from the previous time-step. Advecting the colour function prior to computing the flow
field at a given time instant ensures that the pressure field is calculated based on the new interface position. In contrast, up-
dating the interface position after the new flow field has been calculated would result in a lag between pressure distribution
and interface position. The discretized VOF advection equation is solved by means of a linear equation system,
A · γ = x, (11)
with a subsequent evaluation of the interface curvature if the simulated flow includes surface tension. The interface advec-
tion equation is discretized using algebraic discretisation schemes and advanced in a time-marching fashion. The general
discretisation of Eq. (2) is described in Section 4 and the proposed modifications to the discretisation are presented in
Section 5.
The time-step applied to advect the interface has a crucial influence on the ability to retain the sharpness of the interface.
Using the CICSAM scheme, Ubbink [21, Chap. 5] found the interface to stay reasonably sharp for Courant numbers Co =
|u | t /x ≤ 0.3. The case studies of Darwish and Moukalled [4] confirm the trend of significantly improved results for
decreasing Courant numbers as presented by Ubbink [21], suggesting viable results for Co ≤ 0.25. Studies performed by
Gopala and van Wachem [18] even suggest a Courant number limit of Co ≤ 0.01 in order to maintain a sharp interface. In
order to overcome the stringent Courant number constraints imposed by the interface advection, different time-steps are
applied to advect the interface and the fluid, as previously proposed by Gopala and van Wachem [18]. A similar approach
has also been used in a recent study by Heyns et al. [25]. Using a dual time-stepping approach, the Courant number criterion
imposed by the interface advection can be fulfilled and, at the same time, efficient use of the computational resources is
made while solving the flow equation system. For consistency, the time-step to solve the flow field has to be equal to or
a multiple of the interface advection time-step.

4. Compressive VOF method

In the following sections, the discretisation of the interface advection equation is presented and discussed. The numerical
schemes presented in this section represent only one specific choice, deemed to be best suited for the discretisation of the
interface advection on arbitrary meshes. Other schemes may be used in a similar fashion without affecting the applicability
of the proposed modifications presented in Section 5.

4.1. Temporal discretisation

In the presented compressive VOF method, the transient term of Eq. (2) is discretized using the Crank–Nicolson scheme.
The discretized interface advection equation is, therefore, given as

(γ Pt − γ Pt −t ) V P  γ ft F tf + γ ft −t F tf−t

=− , (12)
t 2
f

where F f represents the flux through face f . The discretisation presented in Eq. (12) requires the fluxes through the face
of two time levels. However, the flux of the new time level is not yet known. Ubbink [21] proposed that if the time-step is
sufficiently small, the right-hand side of Eq. (12) can be approximated as

 γ ft F tf + γ ft −t F tf−t  γ ft + γ ft −t

− ≈− F tf−t . (13)
2 2
f f
 F. Denner, B.G.M. van Wachem / Journal of Computational Physics 279 (2014) 127–144 131

Fig. 1. Example of the arrangement of the face under consideration f , donor cell D, acceptor cell A and the extrapolated upwind node U for an arbitrary,
unstructured mesh.

Hence, Eq. (12) simplifies to

(γ Pt − γ Pt −t ) V P  γ ft + γ ft −t
=− F tf−t . (14)
t 2
f

Analogous to the continuity constraint defined in Eq. (9), it is proposed to determine flux F f in Eq. (14) based on the
advecting velocity unf as

F f = unf A f . (15)

Thus, flux F f used to advect the interface satisfies continuity and the advection of the interface is defined consistently with
the flow advection.

4.2. Spatial discretisation

The CICSAM scheme of Ubbink and Issa [3] is implemented as part of the presented VOF methodology to determine the
value of colour function γ f at mesh face f in Eq. (14). The CICSAM scheme is founded on the Normalised Variable Diagram
(NVD) of Leonard [33]. By defining an acceptor cell A, a donor cell D and an upwind node U for cell face f , illustrated in
Fig. 1, the normalised colour function value is defined as
γ − γU
γ̃ = . (16)
γ A − γU
The Convective Boundedness Criterion (CBC) proposed by Gaskell and Lau [34] stipulates that in order to avoid unphysical
oscillations in the solution, γ D has to be bounded by γ A and γU , given as 0 ≤ γ̃ D ≤ 1. However, unlike on structured
meshes, the upwind value γU is not readily available on arbitrary, unstructured meshes. Jasak et al. [35] proposed a general
definition of the upwind node for arbitrary meshes following the CBC, defined as
γ U = γ A + 2 (x D − x A ) · ∇ γ | D , (17)
with x D and x A representing the position vector of donor cell D and acceptor cell A, respectively. Hence, the mesh ar-
rangement and cell shape are irrelevant. This upwind definition has also been used by Ubbink and Issa [3] in the original
proposal of the CICSAM scheme.
The CICSAM scheme is based on the Hyper-C (HC) scheme, which follows the upper bound of the CBC, and the ULTIMATE
QUICKEST (UQ) scheme of Leonard [33]. Using normalised values as defined in Eq. (16), the HC scheme is given as
 
γ̃ D 
when 0 ≤ γ̃ D ≤ 1
γ̃ fH C = min 1, Co D (18)
γ̃ D when γ̃ D > 0 and γ̃ D < 1,
and the UQ scheme is defined as
  8 Co D γ̃ D +(1−Co D )(6 γ̃ D +3) 
, γ̃ fHC when 0 ≤ γ̃ D ≤ 1
γ̃ fUQ = min 8 (19)
γ̃ D when γ̃ D > 0 and γ̃ D < 1,
where Co D is the Courant number with respect to the fluxes leaving the donor cell.
Having defined both discretisation schemes, a blending function ψ f is required to switch between the two discretisation
schemes, so that

γ̃ f = ψ f γ̃ fHC + (1 − ψ f ) γ̃ fUQ . (20)

The blending function ψ f proposed by Ubbink and Issa [3], is based on the angle Θ f between the interface normal vector
m f and the vector s f , depicted in Fig. 2, connecting the donor cell and acceptor cell, defined as
 
cos(2Θ f ) + 1
ψ f = min kψ ,1 , (21)
2
132 F. Denner, B.G.M. van Wachem / Journal of Computational Physics 279 (2014) 127–144

Fig. 2. Mesh element P with neighbour element Q and shared face f of a triangular two-dimensional mesh, including face normal vector n f , vector r f
from interpolation point f  to face centre f and vector s f connecting the adjacent cell centres. The dashed line depicts the interface with its normal
vector m.

with angle Θ f being

 
 sf 

Θ f = arccos m f ·  . (22)
|s f | 
The constant kψ in Eq. (21) controls the dominance of the HC scheme (the compressive scheme), where an increasing value
of kψ corresponds to a stronger dominance of the HC scheme. Ubbink and Issa [3] recommended kψ = 1, for which the
blending function becomes
cos(2Θ f ) + 1  2
ψf = = cos(Θ) . (23)
2
Defining a weighting coefficient β f using the normalised colour function γ̃ f , given as
γ̃ f − γ̃ D
βf = , (24)
1 − γ̃ D
the colour function value at face f follows as [3]
γ f = (1 − β f )γ D + β f γ A . (25)
As pointed out by Ubbink [21], inserting Eq. (25) in Eq. (14) does not guarantee a bounded solution on non-Cartesian
meshes. Because of the implicit implementation, the predicted value γ f cannot be corrected explicitly but instead the
weighting factor β f has to be corrected by means of a corrector step. For a detailed derivation of the corrector step the
interested reader may consult the thesis of Ubbink [21] or the paper of Ubbink and Issa [3]. Because the predictor–corrector
approach outlined above can only determine the colour function field within the applied solver tolerance, the colour function
value may not fulfill the criteria
0 ≤ γ ≤ 1. (26)
Thus, any colour function value lower than zero or larger than unity is explicitly set to the respective bound. As demon-
strated in Section 6, this does not notably affect the mass-conservation of the method.

5. Skewness correction

Skewness is a commonly encountered complication for non-Cartesian, arbitrary meshes, as for instance depicted in Fig. 2.
As a result of skewness, the geometric face centre does not coincide with the interpolation point at the face, as for instance
the point where the vector connecting the adjacent cell centres intersects the face, denoted in Fig. 2 as f  . Thus, the
linearly interpolated face value becomes inaccurate and the accuracy of the interpolation reduces formally to first order.
As demonstrated by Jasak [22], the numerical discretisation error caused by mesh skewness is diffusion-like. The advection
error E s resulting from mesh skewness for cell P is given as

E s, P = (r f · ∇ γ | f ) u f n f A f , (27)
f

where r f is the vector from interpolation point f  to face centre f , as shown in Fig. 2. Defining the diffusion coefficient of
the mesh skewness, following Jasak [22], as
Γs = u f · r f , (28)
the error induced by mesh skewness can be formulated as the diffusion term
E s = ∇ · (Γs ∇ γ ). (29)
 F. Denner, B.G.M. van Wachem / Journal of Computational Physics 279 (2014) 127–144 133

Hence, interpolating cell-centred values to face centres without correcting for skewness, as for instance by means of Eq. (25),
leads to significant numerical diffusion on meshes with appreciable skewness. Previously published simulation results [4,
21,27,28] indicate a considerable impact of mesh skewness on the interface advection on tetrahedral meshes compared to
Cartesian meshes. In what follows, three methods to mitigate the numerical diffusion caused my mesh skewness for the
compressive VOF method described in the previous section are proposed, discussed and validated.

5.1. Explicit skewness correction

The explicit correction of mesh skewness is widely applied to raise the formal accuracy of spatial interpolation on
arbitrary meshes to second order and to improve the accuracy of the results [36–38]. In order to correct the interpolation
for mesh skewness, the cell-centred values are interpolated to interpolation point f  , as depicted in Fig. 2. The interpolated
value at point f  is then corrected to face centre f using the first derivative at face centre f . Including an explicit skewness
correction, the interpolation of the colour function at face f presented in Eq. (25) becomes
γ f = (1 − β f ) γ D + β f γ A + ∇ γ | f · r f , (30)
where r f is the vector from interpolation point f  to face centre f . Including the explicit skewness correction, the dis-
cretized interface advection equation, Eq. (14), follows as

(γ Pt − γ Pt −t ) V P  γ ft + γ ft −t  
=− F tf−t − ∇ γ |tf−t · r f F tf−t , (31)
t 2
f f

and Eq. (2) becomes

 
∂γ ∂γ ∂ ∂γ
+ ui =− Γs . (32)
∂t ∂ xi ∂ xi ∂ xi
Essentially, the explicit skewness correction subtracts a diffusion term to Eq. (14), which counteracts the numerical diffusion
induced by mesh skewness as defined in Eq. (27). As a result, however, the conservation of the colour function as described
by Eq. (2) is violated and, hence, the mass of each phase is no longer conserved. The explicit skewness correction is,
therefore, expected to increase the mass conservation error.

5.2. Skewness-adapted blending function

One readily available measure to tackle numerical diffusion affecting the interface advection is the application of a more
compressive advection scheme. Increasing the coefficient kψ of the CICSAM scheme increases the dominance of the com-
pressive Hyper-C scheme over the more diffusive UQ scheme (see Section 4.2). Applying the CICSAM scheme, a notable
reduction in numerical diffusion has been observed by Ubbink [21] on triangular meshes and by Denner [28] on tetrahedral
meshes for increasing kψ . The findings reported by Ubbink [21] and Denner [28] suggest that applying a more compressive
advection scheme mitigates the adverse affects of numerical diffusion by overcompressing the interface.
It is, therefore, proposed to define the blending function ψ f based on the skewness ξ f of the given cell face, in addition
to the angle Θ f , with the skewness defined as [38]
|r f |
ξf = , (33)
|s f |
where r f is the vector from interpolation point f  to face centre f and s f is the vector connecting the donor cell and
the acceptor cell. The aim is to give the compressive scheme a higher priority for faces with high skewness. The proposed
blending function is defined as

1 when Θ f < d f
ψf = (34)
[cos((Θ f − d f )1+kξ f )]2 when Θ f ≥ d f ,
where k is a predefined coefficient defining the strength of the compression and d f is the phase shift of the blending
function. In order to assure ψ f = 0 for Θ f = 90◦ , the phase shift is given as
  1+1kξ
π π f
df = − . (35)
2 2
For zero skewness, e.g. on an equidistant Cartesian mesh, or compression k = 0, the blending function reduces to the original
blending function as defined in Eq. (23). Fig. 3 depicts examples of blending function ψ f for various magnitudes of skewness
ξ f and compression coefficients k as a function of angle Θ f . The blending function ψ f as defined in Eq. (34) ensures the
compressive scheme is also applied for larger angles Θ f with increasing skewness, thus, counteracting numerical diffusion.
Since no additional terms are added to the interface advection equation, the mass conservation in each phase is not
affected by the revised blending function. However, caution has to be exercised when overcompressing the interface by
134 F. Denner, B.G.M. van Wachem / Journal of Computational Physics 279 (2014) 127–144

Fig. 3. Examples of the skewness-adapted blending function ψ f as a function of skewness ξ f and angle Θ f for different compression coefficients k.

modifying the blending function. Adapting the blending function based on the mesh skewness in the proposed way does
neither account for the magnitude nor the direction of the skewness error E s as defined in Eq. (27). Furthermore, applying
the proposed blending function may lead to local wrinkling of the interface if the compression considerable exceeds the
numerical diffusion induced by mesh skewness.

5.3. Implicit skewness correction

As an alternative to the explicit skewness correction presented in Section 5.1 and the skewness-adapted blending function
proposed in the previous section, an implicit skewness correction is proposed. Instead of adding an explicit term that
accounts for the skewness error, the weighting factor β f is modified to correct the skewness error. This approach follows
a similar rational as the correction of unbounded values included in the CICSAM scheme [3], where the implicit weighting
factor β f is adapted. Applying the skewness correction given in Eq. (30), the colour function corrected for skewness γ fs at
face f is given as

γ fs = γ f + ∇ γ | f · r f , (36)

where γ f is the original value of the colour function calculated from Eq. (25). Following the definition of the face value
presented in Eq. (25), Eq. (36) can be written as
 
1 − β sf γ D + β sf γ A = (1 − β f )γ D + β f γ A + ∇ γ | f · r f , (37)

where β sf is the skewness-corrected weighting factor for γ fs . After rearranging, the skewness-corrected weighting factor β sf
follows as
∇γ | f · r f
β sf = β f + . (38)
γ A − γD
In order to obtain a bounded solution, β s is explicitly bounded to 0 ≤ β s ≤ 1. The revised weighting factor β s , therefore,
corrects the face value of the colour function implicitly if skewness is present, taking into account the values at the donor
cell and the acceptor cell as well as the magnitude and the direction of the colour function gradient. Applying the revised
weighting factor β sf instead of β f in Eq. (25) is expected to significantly reduce numerical diffusion as a result of mesh
skewness. Moreover, since no additional terms are added to the interface advection equation, Eq. (2), by applying the implicit
skewness correction, the discretisation retains its original mass-conserving properties, as demonstrated in Section 6.

6. Results

The compressive VOF methodology described in Section 4, as well as the remedies for mesh skewness proposed in Sec-
tion 5 are validated and compared using three representative test cases. The pure advection of a fluid particle in a constant
velocity field, see Section 6.1, and in a shear flow, see Section 6.2, is simulated. These test cases allow the assessment of the
proposed methods regarding accuracy and mass conservation without the influence of velocity gradients, pressure gradients
or surface tension. Subsequently, the rise of a bubble in an initially quiescent fluid is presented in Section 6.3, focusing
on the shape evolution of the bubble and the development of its rise velocity. Two bubbles with different properties are
considered. The accurate advection of the colour function is crucial in this test case, since the acting surface tension as
well as the buoyancy force and the viscous stresses are directly dependent on the colour function field. In a previous study,
 F. Denner, B.G.M. van Wachem / Journal of Computational Physics 279 (2014) 127–144 135

Fig. 4. Example of the spatial spreading of the colour function as a result of numerical diffusion. The depicted iso-contours represent γ = 0.2, γ = 0.5 and
γ = 0.9.

Denner and van Wachem [27] observed significant differences in the predicted behaviour between Cartesian and tetrahedral
meshes for this test case with respect to the bubble rise velocity and the numerical diffusion of the colour function.
Two reference values are defined to quantify the conservation of mass and the numerical diffusion of the tested methods.
Assuming a mass-conserving interface capturing/tracking method is deployed in a closed computational domain, the total
volume of each phase must be conserved. Hence, the volume of colour function in the domain,


N
Vγ = γP V P , (39)
P =1

where N is the total number of mesh cells, has to be conserved as long as no colour function γ = 0 is advected out of the
domain, regardless of the presence or extend of numerical diffusion. The conservation error of the colour function E ( V γ ) is
defined as
| V γ − V γ ,0 |
E (V γ ) = , (40)
V γ ,0
where V γ ,0 is the initial volume of the colour function. However, even given perfect conservation of the colour function,
the distribution of the colour function is affected by numerical diffusion. Numerical diffusion means that iso-contours of the
colour function are spreading out over time and, as a consequence, the volume enclosed by a given iso-contour, defined as
N 
 0 if γ ≤ a
V (γ ) = (41)
γ P V P if γ > a,
P =1

changes, as depicted in Fig. 4. Tracking the temporal evolution of the relative volume V ∗ for a representative colour function
value,
V (γ )
V ∗ (γ ) = , (42)
V 0 (γ )
with V 0 (γ ) being the initial volume enclosed by the iso-contour of γ , provides a means to quantify numerical diffusion.

6.1. Advection of a circular interface in a constant velocity field

The advection of a circular interface on a tetrahedral mesh is simulated to examine the ability of the presented methods
to preserve the shape of the interface. The domain has a size of 2 m × 1 m × 0.05 m and is resolved with a tetrahedral mesh
of approximately 4.0 × 105 cells. The circular interface has an initial diameter of 0.1 m and its centre is initially positioned
at x(t = 0 s) = (0.15, 0.15, 0) m, as depicted in Fig. 5. The velocity field is uniform with u = (1.7, 0.7, 0.0) m s−1 and the
time-step corresponds to a maximum Courant number of 0.1. Since the advection is constant and no forces are acting, the
interface should keep its original shape. Given the domain dimensions and the applied velocity, the interface centre should
be positioned at x(t = 1 s) = (1.85, 0.85, 0) m at the end of the simulation.
Fig. 6 shows the final colour function distribution for the original method as presented in Section 4 and the results
including the modifications proposed in Section 5. Even though the temporal and spatial velocity gradients are zero, this
simple convection test unveils considerable difficulties of state-of-the-art compressive VOF methods in preserving the shape
of the interface on the applied tetrahedral mesh. As observed in Fig. 6(a), the shape of the fluid particle is notably distorted if
no numerical treatment for mesh skewness is applied. The fluid particle disintegrates when the explicit skewness correction
presented in Section 5.1 is applied, shown in Fig. 6(b). The application of the skewness-adapted blending function, proposed
in Section 5.2, does not result in considerable differences compared to the original method. As observed in Figs. 6(c)–(e),
a higher compression coefficient k leads to a better preservation of the fluid particle shape. Applying the implicit skewness
correction proposed in Section 5.3, the interface is advected accurately and the final shape of the fluid particle is in very
good agreement with its initial shape, as shown in Fig. 6(f).
136 F. Denner, B.G.M. van Wachem / Journal of Computational Physics 279 (2014) 127–144

Fig. 5. Mesh and initial colour function distribution for the constant advection case at t = 0 s.

Fig. 6. Final distribution of the colour function for the constant advection case at t = 1 s.

Fig. 7. Conservation error of the total volume of the colour function E ( V γ ) as a function of time for the constant advection case.

With respect to mass conservation, given an applied solver tolerance of 10−6 and 7716 time-steps performed in the
simulation, the error in colour function volume of the order of 10−4 , as shown in Fig. 7, for the original methodology
without modifications as well as including the implicit skewness correction or the skewness-adapted blending function
is well within the solver tolerance. The explicit skewness correction, however, leads to an unacceptable error in volume
conservation larger than 10%. The sharp rise in conservation error at the end of the simulation observed in Fig. 7 for the
original method without modifications and including the modified blending function indicates that in these cases a small
amount of colour function is already advected out of the domain, indicating considerable numerical diffusion since the
interface should remain entirely in the domain. Considering the volume occupied by the iso-contour representing γ = 0.9
as a function of time in Fig. 8, a similar conclusion with regards to the impact of the modifications to the accuracy of
the results can be drawn as from the qualitative assessment of Fig. 6. Applying the implicit skewness correction clearly
results in the smallest diffusion of the colour function field, evident by the almost constant volume covered enclosed in
the iso-contour. The explicit skewness correction causes the largest errors, whereas the skewness-adapted blending function
and the original method are of similar accuracy.

6.2. Advection of a circular interface in shear flow

A circular interface in a shear flow is simulated to assess the accuracy of the proposed methods for shear-driven interface
advection. This case represents a more challenging test of the interface advection algorithm as the advection in a constant
velocity field presented in the previous section, and has been used in a number of similar studies (e.g. [3,18,39]). The
domain has a size of π m × π m × 0.1π m and the applied velocity field is given as
 F. Denner, B.G.M. van Wachem / Journal of Computational Physics 279 (2014) 127–144 137

Fig. 8. Relative volume V ∗ enclosed by the iso-contour of γ = 0.9 as a function of time for the constant advection case.

Fig. 9. Initial (t = 0 s), intermediate (t = 12.5 s) and final (t = 25 s) result of the shear advection test on two equidistant Cartesian meshes. The iso-contour
lines represent γ = 0.25 and γ = 0.75.

 
sin(x) cos( y )
u (x, y , z) = − cos(x) sin( y ) m s−1 . (43)
0
The applied tetrahedral mesh resolves the domain by approximately 2.38 × 105 cells. Reference results on Cartesian meshes
of 100 × 100 × 5 cells and 200 × 200 × 5 cells are also provided. Initially, the circular interface has a diameter of 0.2π m and
is positioned at (0.5π , 0.25π , 0) m. The interface is advected with the velocity given in Eq. (43) for 12.5 s with a Courant
number of 0.1. Subsequently, the flow field is inverted and the interface is advected for another 12.5 s. Hence, assuming
a perfect advection of the interface, the interface should assume its initial shape at the end of the simulation.
The results obtained on the Cartesian meshes, depicted in Fig. 9, show the initialised interface at t = 0 s as well as the
intermediate result at t = 12.5 s and the final result at t = 25 s. The tail of the stretched interface at the shown intermediate
stage (t = 12.5 s) is naturally more detailed on the finer mesh. The final interface shape at t = 25 s on both meshes is in
excellent agreement with the initial interface shape and only minor differences are noticeable. The major difference between
the two Cartesian meshes is the interface thickness, which directly depends on the resolution of the mesh.
Noticeable deficiencies are observed for the results obtained on the tetrahedral mesh, shown in Figs. 10 and 11, compared
with the results of the Cartesian meshes. Using the compressive VOF method in its original form without modifications, the
tail of the sheared interface at the intermediate stage is thicker and shorter than on the Cartesian meshes, as shown in
Fig. 10(a). The interface in the final result, given in Fig. 11(a), has some resemblance of the initially circular interface but
also shows significant distortion. The explicit correction is distorting the colour function field to an unrecognisable extent,
as seen in Figs. 10(b) and 11(b), and is clearly inapplicable in this particular test case. The results obtained using the
skewness-adapted blending function, shown in Figs. 10(c)–(e) and 11(c)–(e), retain the major interface features but show
an increased wrinkling and distortion of the interface with increased coefficient k. The wrinkling may be a result of the
overcompression of the interface. Applying the implicit skewness correction proposed in Section 5.3, the thickness of the
138 F. Denner, B.G.M. van Wachem / Journal of Computational Physics 279 (2014) 127–144

Fig. 10. Intermediate result at t = 12.5 s of the shear advection test on the tetrahedral mesh. The iso-contour lines represent γ = 0.25 and γ = 0.75.

Fig. 11. Final result at t = 25.0 s of the shear advection test on the tetrahedral mesh. The iso-contour lines represent γ = 0.25 and γ = 0.75.

tail in the intermediate result, see Fig. 10(f), is in better agreement with the result of the Cartesian meshes. The final result
obtained with the implicit correction, shown in Fig. 11(f), shows considerably less distortion than the results obtained with
the unmodified VOF method and the result is also in reasonable agreement with the initial interface shape.

6.3. Rising bubble

A bubble rising in a heavier fluid due to buoyancy is simulated, characterised by its Morton number Mo = g μo4 /ρo σ 3
and its Eötvös number Eod = ρo gd20 /σ . Two non-dimensional sets of properties are considered: a) Eod = 40, Mo = 0.056 and
b) Eod = 243, Mo = 266. The influence of surface tension and inertia is stronger for bubble a) than for bubble b), whereas
viscous stresses play a more pertinent role for bubble b).
For both bubbles the fluid of the continuous phase has a density of ρo = 1000 kg m−3 . The viscosity of the continuous
phase is μo = 0.2736 Pa s for the bubble with Eod = 40, Mo = 0.056 and μo = 0.5869 Pa s for the bubble with Eod = 243,
Mo = 266. The properties of the dispersed phase are set according to the chosen density and viscosity ratios. Following the
suggestions of Lebaigue et al. [40] and to make the results comparable to previously reported results of the chosen bubble
[20,41], the density and viscosity ratio of the bubble with Eod = 40 and Mo = 0.056 is ρi /ρo = μi /μo = 10−2 . The ratio of
density and viscosity of the bubble with Eod = 243 and Mo = 266 are ρi /ρo = 10−3 and μi /μo = 10−2 , in order to make
the results comparable to previously reported experimental [42] and numerical studies [41,43,44]. In both cases the bubble
is initially spherical with a diameter of d0 = 0.02 m and a surface tension coefficient defined by the non-dimensional
parameters given above. The gravitational acceleration of g = 10 m s−2 is acting in negative y-direction. Both fluids are
initially at rest and the motion of the bubble is induced by buoyancy only. The domain is of size 5d0 × 7d0 × 5d0 and is
resolved with a tetrahedral mesh of approximately 1.4 × 106 cells. Reference results are also provided using an equidistant
 F. Denner, B.G.M. van Wachem / Journal of Computational Physics 279 (2014) 127–144 139

Fig. 12. Rise velocity of the bubble with Eod = 40, Mo = 0.056 as a function of time on the equidistant Cartesian mesh and the tetrahedral mesh, including
the proposed modifications.

Cartesian mesh of 100 × 140 × 100 cells. For all cases, the applied time-step fulfills the capillary time-step constraint of
Brackbill et al. [32], given as

(ρi + ρo )x3
t ≤ , (44)
4πσ
where x represents the mesh spacing. For cases simulated on the tetrahedral mesh, the average distance between neigh-
bouring cell centres in the interfacial region is taken as the representative mesh spacing to determine the capillary time-step
constraint. This time-step corresponds to a Courant number of Co ≤ 0.83 in all cases. Using the dual time-stepping approach
described in Section 3, the time-step applied to advect the VOF colour function is dynamically adapted to satisfy Co = 0.01.
The boundary at the top of the domain is considered to be an outlet boundary, all other boundaries are free-slip walls. The
interface curvature is evaluated using the CELESTE method [27] and the fluid properties are convoluted at the interface as
described in [45].
The finite size of the fluid domain affects the rise velocity of the bubble, which is important for the correct interpretation
of the simulation results. Following the work of Harmathy [46], the expected rise velocity in a finite domain is
  2 
d0
u r ≈ u r∞ 1 − , (45)
L
where L is the domain size perpendicular to the rise axis and u r∞ is the rise velocity in a domain of infinite extend. Given
the domain width of 5d0 , the rise velocity in the deployed computational domain is expected to be 96% of the rise velocity
observed in a domain of infinite extend.

6.3.1. Eod = 40, Mo = 0.056

Empirical studies by Clift et al. [47, Fig 2.5] suggest a terminal Reynolds number of Red = |u |r ρo d0 /μ
o ≈ 20.5–21.0 for a
bubble with Eod = 40 and Mo = 0.056. This Reynolds number represents a Froude number of Frd = |u r |/ d0 g ≈ 0.626–0.642
with respect to the fluid properties given above. Hence, considering the finite extend of the computational domain, the
expected terminal Froude number is Frd ≈ 0.601–0.616. Numerical studies of Lubin et al. [48], using a computational domain
of unspecified extend, resulted in a terminal Froude number of 0.60. Denner et al. [20] reported a terminal Froude number
of 0.58 using a VOF-PLIC method in a domain of 5d0 width and Pivello et al. [41] obtained a terminal Froude number
Frd = 0.606 with a front-tracking method in a domain of 8d0 width.
Fig. 12 depicts the temporal evolution of the rise velocity of the bubble for the presented method and the proposed
modifications on the tetrahedral mesh in comparison with the results obtained on the equidistant Cartesian mesh. Applying
the presented compressive VOF method without modifications on the tetrahedral mesh results in a continuously decreasing
rise velocity after the initial acceleration and does not reach a terminal rise velocity in the simulated time frame, as observed
on the Cartesian mesh. Based on the similar evolution of the bubble shapes, shown in Fig. 13, on the Cartesian mesh and the
tetrahedral mesh applying the compressive VOF method without modifications, a difference in drag force can be disregarded
as a potential explanation for the difference in rise velocity. Denner and van Wachem [27] attributed the progressively
decreasing rise velocity to a gradual reduction of the effective density difference between the fluids as a result of numerical
diffusion of the colour function. Each of the three skewness correction methods improves the prediction of the rise velocity,
as observed in Fig. 12. However, the explicit skewness correction distorts the bubble considerably, as shown in Fig. 13,
whereas the implicit correction as well as the skewness-adapted blending function do not notably affect the bubble shape.
Interestingly, the skewness-adapted blending function with a compression coefficient k = 3 provides an accuracy very similar
to the implicit skewness correction for this test case. As mentioned in Section 5.2, adapting the blending function based on
the local mesh skewness compresses the interface regardless of the direction or magnitude of the skewness error. Thus, the
success of adapting the blending function strongly depends on the local mesh errors and the flow characteristics.
140 F. Denner, B.G.M. van Wachem / Journal of Computational Physics 279 (2014) 127–144

Fig. 13. Shape evolution of the rising bubble with Eod = 40, Mo = 0.056 on the Cartesian mesh and the tetrahedral mesh, including the proposed modifica-
tions. The bubble shapes are illustrated every 0.07 s, starting from the initial position, with the iso-contours representing γ = 0.5.

Fig. 14. Conservation error of the total volume of the colour function E ( V γ ) as a function of time for the rising bubble (Eod = 40, Mo = 0.056).

Fig. 15. Relative volume V ∗ enclosed by the iso-contour of γ = 0.9 as a function of time for the rising bubble (Eod = 40, Mo = 0.056).

The conservation error of the colour function, depicted in Fig. 14, presents a familiar picture. The conservation error of
the presented VOF method without modifications as well as with the implicit skewness correction or the skewness-adapted
blending function is negligible, whereas the explicit skewness correction leads to a substantial conservation error. The gain
in colour function observed for the explicit skewness correction increases the bubble volume and, as a result, increases the
acting buoyancy and the rise velocity of the bubble. According to the evolution of the volume enclosed by the iso-contour of
γ = 0.9, shown in Fig. 15, the implicit skewness correction and the skewness-adapted blending function reduce the numer-
 F. Denner, B.G.M. van Wachem / Journal of Computational Physics 279 (2014) 127–144 141

Fig. 16. Rise velocity of the bubble with Eod = 243, Mo = 266 as a function of time on the equidistant Cartesian mesh and the tetrahedral mesh, including
the proposed modifications.

ical diffusion considerably compared to the unmodified compressive VOF method and the explicit skewness correction. This
corresponds very well with the predicted rise velocities, especially the continuous reduction in rise velocity for the com-
pressive VOF method without modifications, and, more generally, further highlights the importance of mitigating numerical
diffusion on arbitrary, unstructured meshes.
The simulation of the rising bubble on the tetrahedral mesh without skewness correction requires on average 22.6% more
computational time to complete each time-step than on the Cartesian mesh. With respect to the three skewness correction
methods, the required computational time on the tetrahedral mesh increases by 32.2% using the explicit correction, by
26.5% using the skewness-adapted blending function and by 15.3% using the implicit correction, compared to the simulation
without skewness correction.

6.3.2. Eod = 243, Mo = 266

Experiments conducted by Bhaga and Weber [42] for a bubble with Eod = 243 and Mo = 266 rising freely due to buoy-
ancy resulted in a terminal Reynolds number of Red = 7.77, equivalent to Frd = 0.51, given the fluid properties applied in
the present study. The experimental apparatus used by Bhaga and Weber [42] can be assumed to represent a domain of
infinite size and the optical method used to determine the rise velocity of the bubble has a reported tolerance of 5–7%, sub-
ject to the shape of the bubble. The experimental results of Bhaga and Weber corresponds well with the empirical studies
of Clift et al. [47, Fig 2.5], suggesting a terminal Reynolds number of Red ≈ 8. Numerical simulations using a front-tracking
method reported by Hua et al. [43] predicted Red = 7.605 and Frd = 0.499 in a domain of 8d0 width with a resolution of
20 cells per bubble diameter. For the same domain width, with a mesh resolution of 32 cells per bubble diameter and also
using a front-tracking method, Pivello et al. [41] reported a terminal Reynolds number of Red = 7.67, representing a Froude
number of Frd = 0.503, for the density ratio applied in the present study. Considering the domain width of 8d0 , the re-
sults of Hua et al. [43] and Pivello et al. [41] are in very good agreement with the expected Froude number Frd = 0.984
Frd,∞ = 0.502 following Eq. (45), with Frd,∞ being the Froude number based on the rise velocity in a domain of infinite
size. Axisymmetric simulations using a coupled level-set/moment-of-fluid method published by Jemison et al. [44] predicted
a terminal Reynolds number of Red = 7.77 in a domain of 4d0 width and a mesh resolution of 64 cells per bubble radius,
precisely matching the experimental result. Jemison et al. [44] also reported consistently exact solutions of the rise velocity
for other bubbles with respect to the experiments of Bhaga and Weber [42], which can be considered a rather unexpected
outcome given the tolerance of the experiments and the impact of the finite width of the computational domain used by
Jemison and co-workers.
The rise velocity as a function of time, depicted in Fig. 16, shows a similar trend, albeit smaller differences between the
tested methods, as the rise velocity of the bubble with Eod = 40 and Mo = 0.056. As previously mentioned, the motion of
this bubble is less governed by inertia as the bubble discussed in the previous section, which may explain the different
impact of the discretisation of the spatial advection and the applied correction of mesh skewness. Apart from the Froude
number obtained by applying the explicit skewness correction on the tetrahedral mesh, the predicted Froude numbers on
the Cartesian mesh and the tetrahedral mesh are in very good agreement with the expected terminal Froude number of
0.4895, based on the experimental results and the size of the computational domain. The rise velocity recorded on the
tetrahedral mesh without correcting for skewness continuously decreases. A similar development can be observed when
the explicit skewness correction or the skewness-adapted blending function are applied. The implicit correction, on the
other hand, provides a stable and accurate prediction of the rise velocity. Fig. 17 shows the shape of the bubble on the
equidistant Cartesian mesh and the tetrahedral mesh. Apart from the result on the tetrahedral mesh with the explicit
skewness correction, the bubble shapes obtained on the tetrahedral mesh are in excellent agreement with the bubble
shapes predicted on the equidistant Cartesian mesh. Similar as for the rising bubble discussed in the previous section, the
explicit skewness correction distorts the bubble considerably.
The conservation of the colour function, shown in Fig. 18, is very similar to the conservation of the colour function for the
test cases presented above. An excellent conservation of the colour function is observed for the compressive VOF method
142 F. Denner, B.G.M. van Wachem / Journal of Computational Physics 279 (2014) 127–144

Fig. 17. Shape evolution of the rising bubble with Eod = 243, Mo = 266 on the Cartesian mesh and the tetrahedral mesh, including the proposed modifica-
tions. The bubble shapes are illustrated every 0.075 s, starting from the initial position, with the iso-contours representing γ = 0.5.

Fig. 18. Conservation error of the total volume of the colour function E ( V γ ) as a function of time for the rising bubble (Eod = 243, Mo = 266).

Fig. 19. Relative volume V ∗ enclosed by the iso-contour of γ = 0.9 as a function of time for the rising bubble (Eod = 243, Mo = 266).

without modifications as well as including the implicit skewness correction or the skewness-adapted blending function,
whereas the conservation of the colour function is clearly violated by the explicit skewness correction. The evolution of the
volume inside the iso-contour of γ = 0.9, presented in Fig. 19, is also very similar to the results obtained for the bubble
with different properties discussed in the previous section. The implicit correction as well as the skewness-adapted blending
function evidently reduce the numerical diffusion compared to the unmodified compressive VOF method. The differences
between the modifications are smaller than seen for the other bubble in the previous section, which corresponds well with
the smaller differences in rise velocity observed in Fig. 18.
 F. Denner, B.G.M. van Wachem / Journal of Computational Physics 279 (2014) 127–144 143

Contrary to the simulation of the bubble with Eo = 40 presented in the previous section, the simulation of the bubble
with Eo = 243 is faster on the tetrahedral mesh without skewness correction than on the Cartesian mesh, requiring on
average 18.3% less computation time per time-step. Applying the explicit skewness correction on the tetrahedral mesh
increases the simulation time of each time-step by 66.7%, by 83.8% using the skewness-adapted blending function and by
18.3% using the implicit skewness correction.

7. Conclusion

In this article, the discretisation and implementation of a compressive VOF method for three-dimensional arbitrary
meshes has been presented. The presented compressive VOF methodology is applicable to arbitrary, unstructured meshes
and conserves the mass of each phase within the limits of the applied solver tolerance. However, as reported in previous
studies [4,21,27,28], the predictive quality of compressive VOF methods on unstructured meshes is severely challenged by
numerical diffusion. Although numerical diffusion of the VOF colour function has been comprehensively studied with re-
spect to the applied differencing schemes, the numerical diffusion of the VOF colour function induced by an inferior mesh
quality, i.e. mesh skewness, has not been the subject of published research yet.
In order to mitigate the adverse effects of mesh skewness, three skewness correction methods have been proposed
and validated. The presented results demonstrate the profound and consistent improvement in accuracy achieved by the
proposed implicit skewness correction. The implicit skewness correction does not affect the mass-conserving properties
of compressive VOF methods and, in terms of accuracy, the implicit skewness correction brings the results obtained on
tetrahedral meshes in reach of simulations conducted on equidistant Cartesian meshes. The results also demonstrate that
the explicit correction of skewness is not a reliable remedy for the numerical diffusion of the VOF colour function and
that it compromises the accurate mass conservation originally provided by compressive VOF methods. The adaptation of the
blending function of the spatial discretisation schemes in response to the local mesh skewness reduces numerical diffusion,
but the indiscriminate compression of the interface, without accounting for the direction and magnitude of the skewness
error, affect the consistency of the solution accuracy.
The presented study also highlights the capabilities of compressive VOF methods in general as well as their applicability
to arbitrary meshes, and further supports previous studies [18–20] which suggest that compressive VOF methods are able to
predict two-phase flows with an accuracy comparable to interface reconstruction methods, such as PLIC methods. Evidently,
discretisation errors induced by the computational mesh have a crucial influence on the predictive quality of two-phase flow
simulations and, as demonstrated in this article, accounting for these errors is essential to make two-phase flow modelling
on tetrahedral and polyhedral meshes accurate and reliable.

Acknowledgement

The authors are grateful to the Engineering and Physical Sciences Research Council (EPSRC) for their financial support
(grant EP/K008595/1).

References

[1] C. Hirt, B. Nichols, Volume of fluid (VOF) method for the dynamics of free boundaries, J. Comput. Phys. 39 (1981) 201–225.
[2] B. Lafaurie, C. Nardone, R. Scardovelli, S. Zaleski, G. Zanetti, Modelling merging and fragmentation in multiphase flows with SURFER, J. Comput. Phys.
113 (1994) 134–147.
[3] O. Ubbink, R. Issa, A method for capturing sharp fluid interfaces on arbitrary meshes, J. Comput. Phys. 153 (1999) 26–50.
[4] M. Darwish, F. Moukalled, Convective schemes for capturing interfaces of free-surface flows on unstructured grids, Numer. Heat Transf., Part B, Fundam.
49 (2006) 19–42.
[5] S. Muzaferija, M. Perić, P. Sames, P. Schellin, A two-fluid Navier–Stokes solver to simulate water entry, in: Proceedings of the Twenty-Second Symposium
on Naval Hydrodynamics, Washington DC, USA, 1999, pp. 638–649.
[6] D.L. Youngs, Time-dependent multi-material flow with large fluid distortion, in: K. Morton, M. Baines (Eds.), Numerical Methods for Fluid Dynamics,
Academic Press, New York, 1982, p. 273.
[7] W.J. Rider, D.B. Kothe, Reconstructing volume tracking, J. Comput. Phys. 141 (1998) 112–152.
[8] B. van Wachem, J. Schouten, Experimental validation of 3-D Lagrangian VOF model: bubble shape and rise velocity, AIChE J. 48 (2002) 2744–2753.
[9] J.E. Pilliod, E.G. Puckett, Second-order accurate volume-of-fluid algorithms for tracking material interfaces, J. Comput. Phys. 199 (2004) 465–502.
[10] E. Aulisa, S. Manservisi, R. Scardovelli, S. Zaleski, Interface reconstruction with least-squares fit and split advection in three-dimensional Cartesian
geometry, J. Comput. Phys. 225 (2007) 2301–2319.
[11] J. López, J. Hernández, Analytical and geometrical tools for 3D volume of fluid methods in general grids, J. Comput. Phys. 227 (2008) 5939–5948.
[12] C. Wu, D. Young, H. Wu, Simulations of multidimensional interfacial flows by an improved volume-of-fluid method, Int. J. Heat Mass Transf. 60 (2013)
739–755.
[13] Y. Renardy, M. Renardy, PROST: a parabolic reconstruction of surface tension for the volume-of-fluid method, J. Comput. Phys. 183 (2002) 400–421.
[14] T. Maric, H. Marschall, D. Bothe, voFoam – a geometrical volume of fluid algorithm on arbitrary unstructured meshes with local dynamic adaptive
mesh refinement using OpenFOAM, arXiv:1305.3417, 2013, pp. 1–30.
[15] J. Mencinger, I. Žun, A PLIC–VOF method suited for adaptive moving grids, J. Comput. Phys. 230 (2011) 644–663.
[16] Z. Wang, J. Yang, F. Stern, A new volume-of-fluid method with a constructed distance function on general structured grids, J. Comput. Phys. 231 (2012)
3703–3722.
[17] M. Huang, L. Wu, B. Chen, A piecewise linear interface-capturing volume-of-fluid method based on unstructured grids, Numer. Heat Transf., Part B,
Fundam. 61 (2012) 412–437.
[18] V.R. Gopala, B.G.M. van Wachem, Volume of fluid methods for immiscible-fluid and free-surface flows, Chem. Eng. J. 141 (2008) 204–221.
144 F. Denner, B.G.M. van Wachem / Journal of Computational Physics 279 (2014) 127–144

[19] I. Park, K. Kim, J. Kim, A volume-of-fluid method for incompressible free surface flows, Int. J. Numer. Methods Fluids 61 (2009) 1331–1362.
[20] F. Denner, D. van der Heul, G. Oud, M. Villar, A. da Silveira Neto, B. van Wachem, Comparative study of mass-conserving interface capturing frameworks
for two-phase flows with surface tension, Int. J. Multiph. Flow 61 (2014) 37–47.
[21] O. Ubbink, Numerical prediction of two fluid systems with sharp interfaces, Ph.D. thesis, Imperial College London, 1997.
[22] H. Jasak, Error analysis and estimation for the finite volume method with applications to fluid flow, Ph.D. thesis, Imperial College London, 1996.
[23] M. Darwish, Development and testing of a robust free-surface finite volume method, Technical Report, Faculty of Engineering and Architecture, Ameri-
can University of Beirut, Beirut, Lebanon, 2003.
[24] F. Moukalled, M. Darwish, Transient schemes for capturing interfaces of free-surface flows, Numer. Heat Transf., Part B, Fundam. 61 (2012) 171–203.
[25] J. Heyns, A. Malan, T. Harms, O. Oxtoby, Development of a compressive surface capturing formulation for modelling free-surface flow by using the
volume-of-fluid approach, Int. J. Numer. Methods Fluids 71 (2013) 788–804.
[26] D. Zhang, C. Jiang, C. Yang, Y. Yang, Assessment of different reconstruction techniques for implementing the NVSF schemes on unstructured meshes,
Int. J. Numer. Methods Fluids 74 (2014) 189–221.
[27] F. Denner, B. van Wachem, Fully-coupled balanced-force VOF framework for arbitrary meshes with least-squares curvature evaluation from volume
fractions, Numer. Heat Transf., Part B, Fundam. 65 (2014) 218–255.
[28] F. Denner, Balanced-force two-phase flow modelling on unstructured and adaptive meshes, Ph.D. thesis, Imperial College London, 2013.
[29] H. Rusche, Computational fluid dynamics of dispersed two-phase flows at high phase fractions, Ph.D. thesis, Imperial College London, 2002.
[30] Y.-Y. Tsui, S.-W. Lin, T.-T. Cheng, T.-C. Wu, Flux-blending schemes for interface capture in two-fluid flows, Int. J. Heat Mass Transf. 52 (2009) 5547–5556.
[31] Y. Morinishi, T. Lund, O. Vasilyev, P. Moin, Fully conservative higher order finite difference schemes for incompressible flow, J. Comput. Phys. 143 (1998)
90–124.
[32] J.U. Brackbill, D. Kothe, C. Zemach, Continuum method for modeling surface tension, J. Comput. Phys. 100 (1992) 335–354.
[33] B.P. Leonard, The ULTIMATE conservative difference scheme applied to unsteady one-dimensional advection, Comput. Methods Appl. Mech. Eng. 88
(1991) 17–74.
[34] P. Gaskell, A. Lau, Curvature-compensated convective transport: SMART, a new boundedness-preserving transport algorithm, Int. J. Numer. Methods
Fluids 8 (1988) 617–641.
[35] H. Jasak, H. Weller, A. Gosman, High resolution NVD differencing scheme for arbitrarily unstructured meshes, Int. J. Numer. Methods Fluids 31 (1999)
431–449.
[36] I. Demirdžić, S. Muzaferija, Numerical method for coupled fluid flow, heat transfer and stress analysis using unstructured moving meshes with cells of
arbitrary topology, Comput. Methods Appl. Mech. Eng. 125 (1995) 235–255.
[37] S. Mathur, J. Murthy, A pressure-based method for unstructured meshes, Numer. Heat Transf., Part B, Fundam. 31 (1997) 195–215.
[38] F. Juretić, A. Gosman, Error analysis of the finite-volume method with respect to mesh type, Numer. Heat Transf., Part B, Fundam. 57 (2010) 414–439.
[39] M. Rudman, Volume-tracking methods for interfacial flows calculations, Int. J. Numer. Methods Fluids 24 (1997) 671–691.
[40] O. Lebaigue, C. Duquennoy, S. Vincent, Test-case No 1: rise of a spherical cap bubble in a stagnant liquid (PN), in: H. Lemonnier, D. Jamet, O. Lebaigue
(Eds.), Validation of Advanced Computational Methods for Multiphase Flow, Begell House Inc., 2005, p. 260.
[41] M. Pivello, M. Villar, R. Serfaty, A. Roma, A. Silveira-Neto, A fully adaptive front tracking method for the simulation of two phase flows, Int. J. Multiph.
Flow 58 (2014) 72–82.
[42] D. Bhaga, M. Weber, Bubbles in viscous liquids: shapes, wakes and velocities, J. Fluid Mech. 105 (1981) 61–85.
[43] J. Hua, J. Stene, P. Lin, Numerical simulation of 3D bubbles rising in viscous liquids using a front tracking method, J. Comput. Phys. 227 (2008)
3358–3382.
[44] M. Jemison, E. Loch, M. Sussman, M. Shashkov, M. Arienti, M. Ohta, Y. Wang, A coupled level set-moment of fluid method for incompressible two-phase
flows, J. Sci. Comput. 54 (2012) 454–491.
[45] F. Denner, B. van Wachem, On the convolution of fluid properties and surface force for interface capturing methods, Int. J. Multiph. Flow 54 (2013)
61–64.
[46] T.Z. Harmathy, Velocity of large drops and bubbles in media of infinite or restricted extent, AIChE J. 6 (1960) 281–288.
[47] R. Clift, J. Grace, M. Weber, Bubbles, Drops and Particles, Academic Press, New York, 1978.
[48] P. Lubin, S. Vincent, S. Abadie, J.-P. Caltagirone, Three-dimensional large eddy simulation of air entrainment under plunging breaking waves, Coast. Eng.
53 (2006) 631–655.