Arch. Comput. Meth. Engng.

Vol. 12, 2, 165-224 (2005) Archives of Computational

Methods in Engineering
State of the art reviews

Numerical Simulation of Two-Phase Free Surface

Flows
Alexandre Caboussat∗
Department of Mathematics
University of Houston
651 P. G. Hoffman Hall
Houston, Texas 77204-3008
[email protected]

Summary
Free surface flows are of most interest in many engineering or mathematical problems and many methods
have been developed for their numerical resolution in various fields of the physics or the engineering. In
this work, the volume-of-fluid method is used for the numerical simulation of two-phase free surface flows
involving an incompressible liquid and a compressible gas and taking into account the surface tension effects.
The incompressible Navier-Stokes equations are assumed to hold in the liquid domain, while the dynamical
effects in the ideal gas are disregarded. A time splitting scheme is used together with a two-grids method
for the space discretization. An original algorithm is introduced to track the bubbles of gas trapped in the
liquid. Numerical results are presented in the frame of mold filling and bubbles and droplets flows. Some
theoretical results concerning free boundary problems are also summarized.

1 INTRODUCTION

Free surface flows appear nowadays in many engineering or mathematical problems. In

the last years, many models and many numerical methods have been developed for the
simulation of free surface flows in different areas.
On one hand, processes such as casting, injection or mold filling involve complex free
surface phenomena with topology changes and turbulence effects that can be nowadays
solved numerically using commercial softwares. On the other hand, many processes such
as ink filling or vaporization involve the formation of small bubbles or droplets and require
a very accurate approximation of the interfaces.
Complex flows with liquid-gas free surfaces have already been considered in the litera-
ture. In most of the numerical models [17, 27, 34, 35, 42, 66, 94, 148, 149, 150, 156, 157], it
has been assumed that the behavior of the liquid-gas mixture was that of an incompressible
two-phase flow. Compressibility effects in two-phase flows have also been considered in
[2, 3, 85, 108, 138, 139], while methods mixing an incompressible liquid and a compressible
gas have been proposed in [23, 47].
Many numerical methods have been proposed to solve two-phase flows in various frame-
works. Lagrangian methods have been used mainly when the deformation of the liquid
domain is not too large and when this domain does not suffer from topology changes.
Arbitrary-Lagrangian-Eulerian methods are widely used for instance in fluid-structure prob-
lems, when the liquid domain slightly changes. In the general case, Eulerian methods are
the most used methods, since they permits to take into account large topology changes
and discontinuities. In Eulerian methods, two classes of methods can be distinguished: the
methods based on the interface tracking, such as the particles methods or MAC methods
and the methods based on the volume tracking. In this last category, the level sets meth-
ods and the volume-of-fluid methods can be cited, see e.g. [4, 74, 83, 130, 132, 145, 163].
Generally two-phase flow models are computationally expensive, especially in three space
∗
Supported by the Swiss National Science Foundation, Grant PBEL2-103152.

2005
c by CIMNE, Barcelona (Spain). ISSN: 1134–3060 Received: November 2004
166 Alexandre Caboussat

dimensions, since both velocity and pressure must be computed at each grid point of the
whole liquid-gas domain by solving either incompressible of compressible equations.
One of our goals in this paper is to present a numerical model in three space dimensions
which allows the velocity field to be computed only in the incompressible liquid, but without
neglecting compressibility effects of the gas onto the liquid.
The model is as follows. The velocity in the gas is disregarded and the compressibility
effects of the gas are taken into account by computing a constant pressure inside each
connected component of the gas domain. The gas is assumed to be an ideal gas and the
ideal gas law is used to compute the pressure. Given the gas pressure onto the liquid-gas
free surface, a volume-of-fluid method is used to track the liquid domain and to compute
the velocity and pressure fields in the liquid. Then the connected components of the gas
domain are found by using an original numbering algorithm. Surface tension effects are
also added. Finally, at given time t, the gas pressure is updated from the ideal gas law.
An implicit time splitting algorithm is applied to decouple all the physical phenomena,
see e.g. [59, 60, 97]. Advection phenomena (including the motion of the volume fraction of
liquid and the prediction of the liquid velocity) are solved first. Then, the bubbles of gas
are tracked and the pressure inside each bubble of gas is computed using the ideal gas law.
Finally a generalized Stokes problem is solved in order to update the velocity in the liquid.
Surface tension effects are computed on the liquid-gas interface.
Numerical results in two and three space dimensions are presented to illustrate the
influence of the gas on the liquid, as well as that of the surface tension effects.
This paper has various objectives. The first goal is to review some of the numerical
methods existing in the literature for the simulation of free boundary problems. Some
methods available in the literature to investigate free surface flows problems are mentioned
here, with special emphasis on the Eulerian methods, which are well adapted to mold filling
problems. This is described in Section 2. The second goal is to present our numerical
method to solve free boundary problems. Thus, in Section 3, our model of two-phase flows
is proposed and governing equations are given. In Section 4, the time splitting algorithm
is presented and a two-grids method for the space discretization is proposed in Section 5.
Finally, numerical results in two and three space dimensions are presented in Section 6 in
the frame of mold filling or bubbles and droplets simulations. A short overview of some
theoretical aspects of free boundary problems and some existence results are also mentioned
in Appendix A.

2 STATE OF THE ART

2.1 Various Models
Numerical methods for solving free surface problems are nowadays of great importance
in many engineering applications. But, historically, mathematical works concerning free
boundary problems have begun around 1800 with Lagrange and, then, with the works of
Stokes and St Venant. From this time, free boundaries problems have grown in various
directions, both for mathematical investigations and for numerical issues.
Free boundaries are very frequent in many different domains of the physics or engi-
neering. Problems with free surfaces appear in fluid-structure interactions [41, 64], blood
flows in moving arteries [121, 159], immiscible multi-fluids problems [81, 94, 153], motion
of glaciers [117], viscoelastic flows [15, 128, 136], mold filling [34, 86, 92, 100, 133], bubbles
and droplets simulations [120, 126, 165, 166], simulations for the (famous) America’s cup
[114] and many other domains. All these free surface problems come from various fields
and then have various characteristics which will require different numerical techniques. A
large number of methods have been introduced in the past thirty years and many of them
Numerical Simulation of Two-Phase Free Surface Flows 167

are reminded in the following. Nevertheless the field of free boundary problems is so broad
that methods are still emerging at a frequent pace.
From the numerical point of view, methods are aiming nowadays at a better accuracy,
specially around the free surface, to describe efficiently the interface between both media.
Another issue is obviously the computational cost of all these methods, since industrial com-
putations are becoming very demanding. Also the mathematical analysis of free boundary
problems still remains pretty much complicated, due to the deep nonlinear nature of these
problems. Mathematical results are then difficult to obtain in the general cases. Some of
the current works will be discussed in Appendix A.
In the following, we focus mainly on two-phase free surface flows. Without loss of
generality, a liquid and a gas are assumed to be enclosed in a bounded cavity denoted
by Λ. In mold filling problems, large modifications in the topology of the liquid domain
and complex geometries may be expected as well as large Reynolds numbers and turbulent
flows. But, on the other hand, in fluid-structure problems or when dealing with the motion
of glaciers, the deformations may usually be small and do not induce any breakup of the
domain. In both cases it is very important to have a very accurate approximation of the
position of the interface between the liquid and the gas, in particular for the approximation
of the surface tension effects or others physical forces applied at the interface. Nevertheless
the different natures of the physical problems clearly suggest different resolution methods.
In most of the numerical methods, both media are generally assumed to be either
incompressible, as in [7, 13, 148, 150] for instance, or compressible, see [1, 2, 3, 85, 138, 139].
In the incompressible case, the velocity v and pressure p satisfy the incompressible Navier-
Stokes equations in the whole liquid-gas domain and for all times in a given time interval,
that is:
∂v
ρ + ρ(v · ∇)v − 2div (µD(v)) + ∇p = f ,
∂t
div v = 0,
1
where D(v) = (∇v + ∇vT ) is the rate of deformation tensor, ρ the constant density and
2
f the external forces. The viscosity is denoted by µ. In the compressible case, the velocity
v, the pressure p and the total energy per unit volume E satisfy the compressible Euler
equations in the whole liquid-gas domain:
∂ρ
+ div (ρv) = 0,
∂t
∂ρv
+ div (ρvv) + ∇p = 0,
∂t
∂E
+ div (v(E + p)) = 0,
∂t

where the density is denoted again by ρ, but is variable. This approach permits to solve the
same equations in the whole domain but with different physical quantities in each phase.
Various relations may be enforced at the interface. The density and the pressure may
be discontinuous at the interface and continuity relations for the velocity or the normal
component of the velocity may be enforced. Initial and boundary conditions are added to
obtain a well-posed problem.
Even if they are less frequent, methods mixing an incompressible liquid and a compress-
ible gas have also been considered in the literature. For instance, a two-dimensional model
involving incompressible liquid and compressible gas has been presented in [23] where the
168 Alexandre Caboussat

pressure and normal velocity are imposed to be continuous across the interface (without
surface tension). In [47], a one-dimensional model for a liquid drop surrounded by high
speed gas is investigated. The Navier-Stokes equations are satisfied in the incompressible
liquid, while the Euler equations are satisfied in the compressible gas and chemical reac-
tions inside the gas are taken into account. However such models are expensive from a
computational point of view since a whole set of equations is solved in the whole liquid-gas
domain and a particular attention is required around the interface.
In the next sections, various methods are enumerated for the resolution of two-phase
free surface flows, independently of the compressible/incompressible nature of the media.

2.2 Various Methods

Several numerical procedures may be used for solving free surface flows, especially to de-
scribe with accuracy the motion of the free surface. Two classes of methodologies can
basically be distinguished: the Lagrangian methods and the Eulerian methods. The La-
grangian methods are based on the displacement of a system of coordinates at each point
of the free surface to track the displacement of the interface between the two phases, while
the Eulerian methods introduce a new variable in the model, denoted here by ϕ, to track
the presence or not of one of the two phases in the whole domain, see Figure 1.
Nowadays the methods proposed in the literature are difficult to put in one class or
the other, since most of them are a mixed compromise to use the advantages of several
approaches. Nevertheless, some very common methods and recent developments are high-
lighted in the following. The reader may also refer to [75, 86, 132] for reviews of various
techniques.
In this section, all the methods are not treated with the same emphasis. Some aspects
of the numerical simulation of free surface flows are discussed, in particular in the Eulerian
framework, while others, such as Lagrangian methods, are only sketched. Other aspects,
such as boundary integral methods or phase-field methods have been omitted. Note finally
that the content of this review reflects only the point of view of the author on the current
state of the art.

fixed mesh

reconstruction of the interface

interface
ϕ=0

ϕ=1
moving mesh

Figure 1. Two categories of models for the tracking of the free surface. Left: Lagrangian
methods, right: Eulerian methods
Numerical Simulation of Two-Phase Free Surface Flows 169

2.3 Lagrangian Methods

Lagrangian methods, described for instance in [68, 69, 78], include also front-tracking meth-
ods [57, 65] or Arbitrary-Lagrangian-Eulerian (ALE) methods [76, 104, 151]. A review of
Lagrangian techniques may also be found for instance in [48, 87]. They are mainly used if
the displacement of the liquid domain is small, for example in fluid-structure interactions or
small amplitude waves, so that a system of coordinates may be attached to each grid point
lying on the interface without suffering from large distortions, stretching or even breaks of
the interface and changes of topology of one of the domains.
In Lagrangian techniques, every point of the liquid domain (i.e. every particle of liquid)
is moved with the liquid velocity. The new position of the interface implies that the whole
liquid domain is stretched. Then both liquid and gas domains are remeshed at each time
step to take into account the displacement of the interface. If the deformation of the liquid
domain is large, for instance for high Reynolds numbers or for flows moving in cavities with
complex topological shapes, the remeshing of both domains can be difficult and leads to
possibly degenerated elements of the mesh. An example of stretching flow is illustrated in
Figure 2 for the case of a breaking wave. Distortion may then lead to a consequent loss of
accuracy and even a breakup of the simulation when the connectivity rules of the mesh are
no longer valid.

Figure 2. Lagrangian methods and moving meshes: between two time steps, the mesh
in the liquid domain may be stretched until degeneracy

On the other hand, such moving-mesh methods may offer a very good approximation of
the interface. The imposition of forces on the boundary is accurate since the mesh always
coincides with the approximation of the physical interface. For instance, the imposition of
the surface tension forces is more accurate when the real physical interface coincides with
the numerical approximation of the interface on the mesh.
This method has the main advantage to follow exactly the liquid domain and its bound-
ary without any changes in the physical model. The interface is then approximated by
a piecewise linear curve determined by the size of the mesh and the accuracy on the nu-
merical approximation of the liquid domain boundary depends then only on the mesh size.
Boundary conditions may be enforced easily at the grid points lying on the boundary.
Unfortunately, problems may be encountered when the deformation of the domains is
large since the elements of the mesh may degenerate. This requires a total or partial remesh-
170 Alexandre Caboussat

ing of the domain (rezoning) and that procedure can be computationally very expensive.
Moreover the method cannot deal with changes of topology of the liquid domain (and hence
of the triangulation), without a complete re-meshing of the domain, for instance when the
liquid breaks into several connected components or when two liquid fronts meet each other.
This drawback is crucial when dealing with mold filling problems.
2.4 Arbitrary Lagrangian Eulerian Methods
Also in order to deal with small deformations, the Arbitrary Lagrangian Eulerian method
has been introduced in [73] and developed since the beginning of the eighties [77]. Fluid-
structure interactions are considered for instance, see e.g. [49, 54]. Structure dynamics are
typically described in a Lagrangian frame of reference, while fluid equations can be written
in Eulerian coordinates. The moving domain is denoted by Ωt and given by the structure
dynamics. It is mapped at each time t into a reference domain, denoted by Ω0 , by an
arbitrary mapping

At : Ω0 → Ωt , where x = A(ξ, t), ξ ∈ Ω0 . (2.1)

∂At
Introducing the notion of domain velocity w = (also called mesh velocity), a generic
∂t
advection equation or conservation law in the moving frame of reference, for instance
∂u
+ ∇x F (u) = 0, (2.2)
∂t
defined on Ωt , can be written under an ALE form, that is


∂u


− w∇x u + ∇x F (u) = 0 (2.3)
∂t
ξ

thanks to the relation

∂u

∂u


= + w∇x u. (2.4)
∂t
ξ ∂t
x

The equations are then written on a fixed reference domain. The large deformations of
the moving domain are not taken completely into account and only a subset of the points
lying on the boundary of the liquid domain are moving. The mesh velocity is then given by
the domain velocity and is distinguished from the motion of the liquid particles. A careful
choice of this velocity may prevent the elements of the mesh to become singular. On the
other hand, the governing equations are modified since rewritten in the reference frame and
an advection term (2.4) is added to counterbalance this mapping. Figure 3 illustrates the
ALE method in the case of fluid-structure problems; the moving domain permits to take
into account the flexibility of the structure of the artery without following every particle of
liquid going through the section of it. The thickness of the artery varies but the inflow and
outflow sections are fixed.
ALE methods are widely used in the modeling of blood flows [49, 55, 159] or motions
of glacier [117], i.e. for problems with small or slow deformation. In these cases, the
triangulation is stretched and modified with its own velocity, independently of the velocity
of the liquid. For instance, in [117], the mesh is moved at each time step to be adjusted to
the height of the glacier, but the projection of the mesh on the horizontal base on which the
ice is lying is unchanged. Again changes of topology of the liquid domain cannot be dealt
with, but must be avoided by restraining the computation on a certain part of the domain.
Numerical Simulation of Two-Phase Free Surface Flows 171

Ω0

Ωt

Figure 3. Arbitrary Lagrangian Eulerian methods. In full line: deformed domain, in

dashed line: reference domain

ALE schemes have been widely used for compressible fluid dynamics and aero-elasticity.
Numerical instabilities and oscillations were noted due to computational errors on some
geometric quantities. For this reason, many stability analysis have been investigated, see
for instance [14, 49] and the ALE method has been linked to the so-called geometric con-
servation laws, see [45] for instance, which express theoretical conditions for the stability
of the method.
For ALE methods applied to compressible flows, one can refer to [140]. The coupling
of free surface flows with Maxwell equations may be found in [54], where the ALE method
is used to describe the surface of the fluid in an aluminum cell during the production of
aluminum. This magneto-hydrodynamic problem provides numerous difficulties such as the
computation of the surface tension effects or the formation of bubbles of gas in the aluminum
bath (see also [84, 131] for this last part). Notice finally that a space-time Galerkin method
has been introduced in [102], which is a generalization of both the Lagrangian and the ALE
approaches.
Nevertheless, the moving-mesh methods are not very well-fitted to mold filling problems
mainly because of the many possible changes of topology of the liquid domain during the
filling of a cavity. Then Eulerian approaches are preferred, but before describing them, let
us say a few words about mesh-free methods called particles methods.

2.5 Particles Methods

The particles methods are mesh-free methods to track the motion of the liquid domain.
The first occurrence of these methods is the particle in cell method (PIC) introduced in
the sixties [70]. Then the front-tracking methods (or surface tracking methods in two
dimensions, see for instance [91, 155]) have been derived from the original PIC method.
They aim at capturing the interface between the two media by using mass-free particles
which are moving with the liquid velocity, independently of any mesh. The volume markers
and the surface markers have to be distinguished here. The first class introduces markers
in the whole liquid domain, while the second one introduces markers only on the interface
between the liquid and the gas, as illustrated in Figure 4.
The volume markers method has been initiated in the sixties [71]. Mass-free particles in
the whole liquid domain are transported with the liquid velocity and track the presence or
not of liquid in each cell of the triangulation, see Figure 4 (left). The position xj of every
marker j satisfies a Lagrangian equation:

dxj
= v(xj ),
dt
172 Alexandre Caboussat

Figure 4. Markers methods. Left: markers method, the liquid domain is given by mass-
less particles; right: surface markers: massless particles are transported only
on the free surface

where v is the liquid velocity. A cell filled with volume markers is a liquid cell, while an cell
empty of markers belongs to the gas domain. The most famous volume markers method is
the markers and cells (MAC) method [71]. Several other methods have been inspired from
the MAC method, see [105] for a review.
On the other side, surface markers methods have been introduced, see for instance
[87, 132]. In order to describe the position of the domain at each time, these methods focus
on the description of the interface between the two media. The interface is defined by a
set of points (or particles) embedded within the computational grid and only these markers
lying on the interface are tracked, see Figure 4 (right). Examples may be found in [5, 28].
In [120], the surface markers are coupled with an accurate reconstruction of the interface
with emphasis on the computation of surface tension forces. For instance, the coupling of
markers and micro-cell method with rigid bodies has been done in [11] for the simulation of
Tsunami phenomena and the coupling with k − ε turbulence models may be found in [164].
Notice that these methods to track the interface are generally conjugated with Eulerian
methods, for instance volume-of-fluid methods, to have a more accurate approximations of
the interface, see e.g. [17, 18, 82, 83]. Repetitions of merging and breakup of the liquid
domain are highlighted in [137]. In particular, the coupling between particles and Eulerian
methods permits to capture details of the liquid front on scales finer than the size of the
fixed mesh, while keeping the advantages of the Eulerian approaches (see below). Markers
have been also thoroughly studied in [57, 65]. More recently an higher order method has
even been proposed in [58]. In general, surface markers are more used than volume markers
since they allow to track exactly the location of the interface with a smaller computational
cost.
On the other hand, Eulerian methods introduce an additional unknown in the whole
cavity in order to track the presence of liquid or gas. An additional equation for this addi-
tional unknown is introduced in order to guarantee the well-posedness of the problem. The
pseudo-concentration methods, see e.g. [35, 42, 93, 106, 152, 156], the level sets methods,
see [27, 34, 111, 112, 147], or the volume-of-fluid (VOF) methods, see [74, 101, 130, 163],
are the most-used Eulerian methods and are described in the following sections.
Let us introduce here some notations. Let T > 0 denote a finite time. The liquid
domain Ωt , t ∈ (0, T ), is assumed to be contained in a bounded cavity Λ. The gas domain
Numerical Simulation of Two-Phase Free Surface Flows 173

ϕ = −1

ϕ=0

Ωt ϕ=2 ϕ=1 Γt

Figure 5. Level sets methods. The interface between the liquid and the gas is the zero
level line of the function ϕ. The level lines corresponding to the negative values
of ϕ are in the gas domain

is defined by Λ\Ω̄t and the liquid-gas interface is denoted by Γt . The velocity and pressure
are denoted by v and p. The additional unknown used to track the presence or not of liquid
is defined on Λ × (0, T ) and is called ϕ.

2.6 Level Sets Methods

In the level sets methods, the free boundary is defined by the level line of a smooth function
ϕ, i.e.

Γt = ∂Ωt = {x : ϕ(x, t) = 0} , t ∈ (0, T ),

see Figure 5 for an illustration.

The level set function is assumed to be positive in the liquid, negative in the gas and at
least twice continuously differentiable. Therefore this function satisfies:

 > 0, if x ∈ Ωt
ϕ(x, t) = < 0, if x ∈
/ Ω̄t
 = 0, if x ∈ ∂Ωt = Γt

The motion is analyzed by advecting the values of ϕ with velocity field v, i.e. the variable
ϕ satisfies

∂ϕ
+ v · ∇ϕ = 0. (2.5)
∂t
This equation is deducted from the assumption that each particle of liquid moves with
the liquid velocity along the characteristic curves. In the level sets approach, (2.5) is
∇ϕ
transformed by setting vN = v to obtain an Hamilton-Jacobi equation:
||∇ϕ||

∂ϕ
+ vN ||∇ϕ|| = 0. (2.6)
∂t
174 Alexandre Caboussat

The quantity vN denotes the normal velocity along the gradient of ϕ. The resolution of
this equation, see for instance [8, 63, 79] needs only the normal component of the velocity
in a neighborhood of the interface.
The numerical resolution of the Hamilton-Jacobi equation is a real challenge for com-
putational mathematics, particularly to obtain stable numerical solutions. Many numerical
algorithms to solve (2.6) can be found in the literature, especially high order algorithms.
The ENO (essentially non-oscillatory) initially introduced by [113] and followed by the
WENO (weighted essentially non-oscillatory) schemes, see for instance [37, 79], are the
most-used algorithms in the level sets community. They are based on high order finite
differences approximations of each of the derivatives appearing in (2.6).
The resolution of the stationary Hamilton-Jacobi is also studied since it is naturally
obtained with a time discretization of (2.6). Notice that the stationary case is closely
related to the Eikonal equation; it consists in finding v satisfying

||∇v(x)|| = F (x), x ∈ Ω, (2.7)

with corresponding Dirichlet boundary conditions. In [135], numerical methods to solve

(2.7) called fast marching methods are extensively reviewed in the framework of the level
sets approach. These methods are based on upwind schemes and yield consistent and
accurate algorithms, but they are designed for problems in which the speed function never
changes sign, so that the front is always moving forward or backward.
A well-known drawback of the level sets approach is the degeneration of ϕ. When the
gradient of ϕ vanishes in the neighborhood of the free surface, that is the function ϕ becomes
flat, the accuracy on the approximation of the interface decreases dramatically. Then the
function may be rescaled, see [115] for instance, so that it remains a distance function, i.e.
the additional function ϕ(x, t) represents the distance between a point x of the domain
and the interface at each time step. Several techniques to re-build a function with such
properties may be found in the literature, see for instance [38, 63]. In most of the cases,
the stretched level sets function is updated when the distortion is too large and reinitialized
with a distance function.
Notice that, with the level sets approach, the conservation of the mass of liquid is not
automatically guaranteed, specially when ϕ is artificially rescaled.
The level sets method is widely used in the scientific community. In [29], the level
sets method is used and the finite element space is enriched with additional basis functions
around the interface to take into account the discontinuities appearing near the free surface.
Curvature-driven flows are treated with a level sets method in [30], while the computation
of dendritic growth (which is usually treated with phase field methods) is investigated with
the level sets approach in [56]. In this latter case, the temperature and normal velocity of
the interface are modeled, while the level sets are used to compute the normal vector and
curvature of the interface. Due to the sharp interface, emphasis is laid on the reinitialization
of the level sets function to keep a distance function. In [110], the level sets method is used
for the simulation of optical waves with reflections.
As a compromise between level sets methods and interface tracking methods, [154]
introduces a method called segment projection method which splits the interface into several
segments and moves each of them independently.
In [37], the level sets method is used to track the free surface in a parallelized Navier-
Stokes solver with free surfaces. The parallelization is based on classical domain decompo-
sition method and made with MPI. Note finally that the regularity of the level sets function
allows the level sets method to be used for digital imaging problems, see in particular [112],
because of its ability to compute the motion of smooth surfaces.
Numerical Simulation of Two-Phase Free Surface Flows 175

2.7 Pseudo-Concentrations Methods

In the pseudo-concentration method, the free boundary is also defined by a level set of a
smooth function in the neighborhood of the interface. Generally, the pseudo-concentration
function has one fixed value in one media and another in the other media and is smoothed
in a neighborhood of the interface. For instance, a continuous regularization may be:

 γ, if d(x) > γ
ϕ(x, t) = −γ, if d(x) < −γ , (2.8)
 d(x) if |d(x)| < γ

where d(x) is the signed distance between x and the interface and γ is a fixed threshold.
Various definitions of pseudo-concentration functions may be found in [35, 152, 156]. By
opposition to the level sets method, this tracking function generally satisfies the advec-
tion equation (2.5), see e.g. [106]. This formulation is easily compatible with geometric
considerations, see e.g. [25] when the pseudo-concentration function is nothing else than
(2.8).
In both the level sets methods and the pseudo-concentration methods, ϕ is smooth
around the interface. This is useful for instance for the computation of the surface tension
effects, since the normal vector n and the curvature κ of the interface may be expressed by
∇ϕ ∇ϕ
n=− , κ(x, t) = −∇ · , (2.9)
||∇ϕ|| ||∇ϕ||

see for instance [112, 134]. Notice that, for mold filling processes, the effects of surface
tension can generally be neglected since the Capillary number is much smaller than the
Reynolds number. If this is not the case, surface tension effects become dominant, for
instance in bubbly flows [65, 80] and viscoelastic flows [128, 136].

2.8 Volume-of-Fluid Methods

In the volume-of-fluid method (VOF), the fluid domain is tracked by its characteristic
function, that is

1, if x ∈ Ωt ,
ϕ(x, t) = (2.10)
0, else.

This function jumps over the interface. In most of the VOF methods, it satisfies the
advection equation (2.5), see for instance [88, 129, 130]. The mass of fluid is rigorously
conserved as long as the numerical scheme is a discrete form of a conservative advection
equation. For this reason, the volume-of-fluid method is also called volume tracking method,
since it is able to capture rather than follow the volume of liquid or its interface. Moreover,
like the level sets methods, the VOF method implicitly takes into account the possible
changes of topology of the liquid domain. Finally, the extension from two space dimensions
to the three-dimensional case is straightforward.
On the other hand, the computation of the curvature for instance is difficult since it
involves the derivatives at the interface of the non-smooth function ϕ. Methods are specially
dedicated to the improvement of the VOF method for the computation of the surface
tension terms. Most of the techniques tend to regularize the volume fraction of liquid in
order to estimate its derivatives. In [16, 20], the smoothing of the volume fraction of liquid
with different kernel functions is discussed. In [38], three approaches are compared: the
smoothing with kernel functions, the interpolation of the interface with a height function
and an estimation via the reconstruction of a distance function, similar to the level sets
176 Alexandre Caboussat

approach. Most of the time, an accurate reconstruction of the interface is needed, as we

will also see later on.
Keeping in mind that our interest is in mold filling problems, the volume-of-fluid method
permits to track the changes in the topology of the domain, see e.g. [66, 83], while conserving
the mass of liquid injected in the mold.
Naturally, many methods have been developed from the volume-of-fluid method and
numerous examples of volume-of-fluid methods may be found in the literature. In [124],
the case of two immiscible incompressible liquids is discussed, with emphasis when the two
liquids are moving past each other to form fingers or instabilities. The coupling of the
interface tracking with averaged equations over the two phases is discussed in [26]. The
volume-of-fluid method may also be mixed with some front tracking methods to obtain a
more accurate approximation of the interface, see for instance [106].
The mixing of VOF and level sets methods is useful to conjugate the advantages of both
methods, see e.g. [149, 157, 158]. It permits to conserve the mass of liquid and keep a smooth
approximation around the interface, but needs a method to let both approximated function
communicate. In [157, 158], a level sets function is reconstructed from the advected volume
fraction of liquid with a one-to-one relation. In [149], the axisymmetric case is treated and
the one-to-one mapping between the level sets function ϕ and the volume of the liquid
domain F is the natural mapping

1
F (C, t) = H(ϕ(x, t))dx,
|C| C
where C denotes any subdomain of Λ (for instance a cell of some mesh), H is the Heavyside
function (H(y) = 1 if y > 0 and 0 elsewhere) and |C| is the volume of the domain C.

2.9 Reconstruction of the Interface

The question of the reconstruction of the interface from the values of the volume fraction
of liquid is of great interest, especially for VOF methods. In general, the advection of the
characteristic function introduces numerical diffusion on the chosen mesh, see Figure 6.

1 1/2 1/2 1/4 1/2 1/4

1 1/2 1/2 1/4 1/2 1/4
1 1/2 1/2 1/4 1/2 1/4

Figure 6. Volume-of-Fluid method: Numerical diffusion of the volume fraction of liquid

for a simple case of the advection of ϕ without any reconstruction of the
interface

The numerical diffusion induces a significant loss of accuracy around the interface and
then some geometry-based techniques have been investigated to obtain a more precise ap-
proximation. The most simple algorithm to reduce the numerical diffusion is the SLIC
algorithm (Simple Line Interface Calculation), developed first in [109] and then adapted to
the combustion problems in [31]. If a structured grid of small cells is considered, the value
of the volume fraction of liquid in one cell is strictly between zero and one in a neighborhood
of the interface. In order to avoid the diffusion of the front, the SLIC algorithm reconstructs
the liquid front by defining simples lines inside the cells, as illustrated in Figure 7. These
straight lines are parallel to one of the coordinate directions and their direction and position
Numerical Simulation of Two-Phase Free Surface Flows 177

1 1/2 1/2
1 1/2 1/2
1 1/2 1/2

1 1 1
1 1 1
1 1 1

Figure 7. Volume-of-Fluid method: Numerical diffusion of the volume fraction of liquid

for a simple case of the advection of ϕ with SLIC method

are deduced from the values of the volume fraction of liquid in the cells in a certain neigh-
borhood of the cell considered, see e.g. [74, 98]. For instance in [98, 99], only the values of
ϕ in the adjacent cells may be used to define the straight line crossing one cell. The main
disadvantage is that the SLIC algorithm is only a first order algorithm with respect to the
mesh size.
The PLIC algorithm (Piecewise linear interface calculation) is also geometric in nature
and has been introduced to increase the order of convergence of SLIC for the reconstruction
algorithm of the interface, see [4, 39, 130, 132] and references therein. The PLIC methods are
second order algorithms. The principle is the following: instead of constructing the interface
by simple lines only along the coordinates directions, all the directions are allowed for one
line in one cell. The interface is not defined by a chain of joined piecewise linear segments,
but rather by a discontinuous chain of segments, with asymptotic small discontinuities,
see Figure 8. The method works well when the curvature of the interface is small with
respect to the mesh size and stays robust when the curvature is large compared to the size
of the mesh. The key point in the algorithm is the determination of the direction of each
segment of the reconstructed interface, which corresponds basically to the characterization
of its normal vector. Again the values of the volume fraction of liquid in the cells in a
neighborhood are taken into account (see for instance [132] for an example). High order
reconstruction algorithms are currently developed, see for instance [40] for the coupling of
interface tracking with finite volumes/finite differences schemes.
Others reconstructions of the interface are also detailed in the literature. For instance,
the aim of the methods presented in [120, 127] is to increase the accuracy on the computation
of the curvature; in [127], the interface is reconstructed locally with a smooth parabolic
function, result of a least-squares minimization. This method gives particular good results in
terms of spurious currents, which are parasitic velocities induced by the numerical algorithm
(see also Section 6). Such a reconstruction of the interface has been recently extended to
viscoelastic flows [128], since surface tension is very relevant for such phenomena, and to
droplets fragmentation [125]. In [120] cubic polynomials are used to interpolate smoothly
locally the interface and the computation of surface tension effects is highlighted.

2.10 Other Methods

Many other methods exist in the literature for the simulation of two-phase flows. Three of
them are mentioned here. The first class of methods we want to mention is the immersed
boundary methods, originally developed in [116] for cardiovascular flows. In this approach,
178 Alexandre Caboussat

Figure 8. Volume-of-Fluid method: reconstruction of the interface with SLIC and PLIC
algorithms. Left: the SLIC algorithm reconstructs the interface as a set of
horizontal or vertical segments. Right: the PLIC algorithm allows the seg-
ments to be oriented indifferently. In both cases, the reconstructed interface
is piecewise linear and discontinuous

the incompressible/compressible equations are solved in the whole domain in Eulerian vari-
ables, while a Lagrangian point of view is used for the simulation of an embedded flexible
structure. The points located on the flexible structure are assumed to move along the char-
acteristics. Several examples may be found in [13, 36, 51, 52, 90] for instance. Let us also
mention the ghost-fluid method, see e.g. [23, 46]. This method aims to increase the accu-
racy around the interface by introducing auxiliary cells, called ghost cells on each side of
the interface. A Riemann solver is used to take into account the discontinuities around the
free surface. Finally, the fictitious domain methods, see for instance [61, 62, 131], permits
to model the motion of bodies in an incompressible flow. The rigid or deformable bodies
are filled with liquid but their motion is imposed via a Lagrange multiplier. Finally, in
[131], the fictitious domain method is used for the simulation of incompressible bubbles of
air, immersed in a liquid.
In this article, the model involves an incompressible liquid and a compressible gas, as
it can be the case in mold filling problems. Given that we are interested in problems with
changes of topology, an Eulerian method has been chosen and the use of the characteristic
function of the liquid domain allows us to conserve a priori the mass of liquid. In the
remaining part of this paper, our method for the numerical simulation of free surface flows is
proposed and its ability to solve various free surface problems is demonstrated via numerical
results.

3 MATHEMATICAL MODEL
3.1 Governing Equations in the Liquid
The model presented in this section is extracted from [19, 20, 21]. Let Λ be a cavity of
Rd , d = 2, 3 in which the fluid must be confined, and let T > 0 be the final time of
simulation. For any given time t, let Ωt be the domain occupied by the fluid, let Γt be the
free surface defined by ∂Ωt \∂Λ and let QT be the space-time domain containing the liquid,
i.e. QT = {(x, t) : x ∈ Ωt , 0 < t < T }.
Some of the notations are reported in Figure 9 in the frame of a two-dimensional sit-
uation, namely the filling of an S-shaped channel. This situation corresponds to water
Numerical Simulation of Two-Phase Free Surface Flows 179

valve

Ωt t=T

t x2
Ω0 - 6 t
*
- x1
t=0

Figure 9. Computational domain for the filling of an S-shaped channel. At initial time,
the channel Λ is empty. Then water enters from the bottom and fills the
channel

entering a thin S-shaped channel lying between two horizontal planes thus gravity can be
neglected. A valve is located at the end of the channel so that gas may escape.
In the liquid region, the velocity field v : QT → Rd and the pressure field p : QT → R
are assumed to satisfy the time-dependent, incompressible Navier-Stokes equations, that is
∂v
ρ + ρ(v · ∇)v − 2div (µD(v)) + ∇p = f in QT , (3.1)
∂t
div v = 0 in QT . (3.2)
1
Here D(v) = (∇v + ∇vT ) is the rate of deformation tensor, ρ the constant density and f
2
the external forces. In order to take into account turbulence effects, a simplified algebraic
model is chosen [146]. The viscosity µ is defined by µ = µL + µT , where µL is the laminar,
constant, viscosity and µT = µT (v) is the additional turbulent viscosity, defined by

µT (v) = αT ρ 2D(v) : D(v) (3.3)
where αT is a parameter to be chosen, see Section 6.3 below. The use of a turbulent viscosity
is unavoidable in order to obtain numerical results that compare well to experiments, since
large Reynolds numbers and thin boundary layers are involved. However, the goal of this
work not being to discuss the influence of the turbulence model, a simple model is chosen
here.
Let ϕ : Λ × (0, T ) → R be the characteristic function of the liquid domain QT . The
function ϕ equals one if liquid is present, zero if it is not. In order to describe the kinematics
of the free surface, ϕ must satisfy (in a weak sense):
∂ϕ
+ v · ∇ϕ = 0 in QT . (3.4)
∂t
The initial conditions are the following. At initial time, the characteristic function of the
liquid domain ϕ is given, which defines the liquid region at initial time:
Ω0 = {x ∈ Λ : ϕ(x, 0) = 1} .
180 Alexandre Caboussat

The initial velocity field v is then prescribed in Ω0 .

The boundary conditions for the velocity field are the following. On the boundary of the
liquid region being in contact with the walls (that is to say the boundary of Λ, see Figure 9),
inflow, slip or Signorini boundary conditions are enforced, see [98, 99]. The reason for using
slip instead of noslip boundary conditions along the walls is due to the fact that, since large
Reynolds numbers are involved, noslip boundary conditions would induce strong boundary
layers along the walls, which would require fine layered meshes.
On the free surface Γt , tangential and capillary forces are neglected here, so that the
forces acting on the free surface are the normal forces due to the gas pressure and the
normal forces due to the surface tension effects. At given time t, let κ(x, t) be the local
curvature of the interface Γt at point x. The following equilibrium relation is then satisfied
on the liquid-gas interface:
−pn + 2µD(v)n = −P n + σκn on Γt , t ∈ (0, T ) , (3.5)
where n is the unit normal of the liquid-gas free surface oriented toward the gas, P is the
pressure in the gas, κ is the local curvature (or mean curvature in the three-dimensional
case) of the interface and σ is a constant surface tension coefficient which depends on both
media on each side of the interface (namely the liquid and the gas). The continuum surface
force (CSF) model (see e.g. [16, 43, 53, 130, 162]) is considered for the modeling of surface
tension effects. By definition, the curvature κ(x, t) is supposed to be negative if the center
of the osculating circle of the interface at point x is on the liquid side of the interface (see
for instance [32] for a similar convention). Note that the dynamical effects inside the gas
domain are neglected.
For example, consider the situation of Figure 10, namely the filling of a two-dimensional
S-shaped cavity (the numerical experiment is described in Section 6). When the cavity is
being filled with liquid, the gas between the valve and the liquid can escape, thus P = Patmo
is the atmospheric pressure on the upper part of the liquid-gas interface. However, a fraction
of gas is trapped by the liquid and cannot escape. A resulting force acts on the liquid-gas
interface which prevents the bubbles from vanishing during experiment. The contribution
to the normal force which comes from the surface tension effects and is also exerted on the
whole liquid-gas interface.
3.2 Governing Equations in the Gas
Consider again the case of Figure 10. Some gas is trapped by the fluid and is compressed.
In our model, the velocity in the gas is disregarded, since modeling the gas velocity would
require solving the Euler compressible equations, which is CPU time expensive.
The pressure P in the gas is assumed to be constant in each bubble of gas, that is to
say in each connected component of the gas domain. Let k(t) be the number of bubbles
of gas at time t and let Bi (t) denote the domain occupied by the bubble number i (the
i-th connected component). Let Pi (t) be the pressure in Bi(t). The pressure in the gas
P : Λ\Ωt → R is then defined by:
P (x, t) = Pi (t), if x ∈ Bi(t) .
Moreover, the gas is assumed to be an ideal gas. Let Vi (t) be the volume of Bi (t). At initial
time, all the gas bubbles have given pressure. At time t, the pressure in each bubble is
computed by using the ideal gas law:
Pi (t)Vi (t) = constant i = 1, . . . , k(t) , (3.6)
with constant temperature. It is assumed in the following that the total fraction number
of molecules inside the set of bubbles which are not in contact with a valve, see Figure 10,
Numerical Simulation of Two-Phase Free Surface Flows 181

valve
escaping gas

?

6 6
-
trapped gas

6filling

Figure 10. Filling of a S-shaped cavity. The gas in the upper part of the cavity is free
to escape through the valve. The gas trapped by the liquid may exert a force
on the liquid. Surface tension forces are exerted on the liquid-gas interface for
each bubble of gas

P (t), V (t)
P (t + τ )
V (t + τ )

t t+τ
Figure 11. One single bubble is floating in the liquid. The product P V remains constant
between time t and time t + τ , i.e. P (t + τ )V (t + τ ) = P (t)V (t)

is conserved between two time steps. Notice that this total fraction number of molecules
in one bubble is proportional to the product of the pressure of the bubble times its volume
since the temperature is assumed to be constant. Thus, the number of molecules of trapped
gas is conserved between time t and t + τ .
The case of a single bubble is first discussed. The situation of Figure 11 is considered.
Assume that the pressure P (t) in the bubble at time t and the volumes V (t) and V (t + τ )
are known. The fraction number of molecules inside the bubble is conserved, so that the
gas pressure at time t + τ is computed from the relation

P (t + τ )V (t + τ ) = P (t)V (t) .

The case when bubbles of gas are created is now discussed. The situations of Figures 12
and 13 are considered. In Figure 12, the broken dam problem in a confined domain is
described. A water column is kept in the left side of a cavity by a fictitious dam. The dam
is removed at time t = 0. At time t, the bubble number 2 is created with volume V2 (t) and
atmospheric pressure Patmo . If the volume V2 (t + τ ) of this bubble is known at time t + τ ,
182 Alexandre Caboussat

bubble 2 bubble 2
P2 (t) = Patmo P2 (t + τ )
valve V2 (t) V2 (t + τ )
bubble 1 bubble 1
bubble 1
P1 (t) = Patmo P1 (t + τ ) = Patmo
P1 (0) = Patmo
V1 (t) ? V1 (t + τ )
liquid ?
liquid liquid
t=0 t t+τ

Figure 12. Broken dam in a confined domain, creation of a bubble. At time t, gas is
trapped by the liquid and the pressure equals atmospheric pressure P1 (t) =
Patmo , P2 (t) = Patmo . At time t + τ , the pressure in bubble 2 is computed
from the relation P2 (t + τ )V2 (t + τ ) = P2 (t)V2 (t) = Patmo V2 (t)

bubble 1
bubble 1 bubble 2
P1 (t + τ ), V1 (t + τ )
P1 (t), V1 (t) P2 (t), V2 (t)

liquid liquid

t t+τ

Figure 13. Merging of two bubbles between time t and time t + τ . The pressure in
bubble 1 at time t + τ is computed from the relation P1 (t + τ )V1 (t + τ ) =
P1 (t)V1 (t) + P2 (t)V2 (t)

then the gas pressure at time t + τ is computed from the relation

P2 (t + τ )V2 (t + τ ) = P2 (t)V2 (t) = Patmo V2 (t) .

The situation of Figure 13 corresponds to the merging of two bubbles. The pressure at time
t + τ is computed by taking into account the conservation of number of molecules in the
bubbles which yields P1 (t + τ )V1 (t + τ ) = P1 (t)V1 (t) + P2 (t)V2 (t).
The case when one bubble splits into two bubbles is now discussed, see Figure 14. The
number of molecules inside the gas domain is conserved between time steps t and t + τ ,
that is

P1 (t + τ )V1 (t + τ ) + P2 (t + τ )V2 (t + τ ) = P1 (t)V1 (t) .

In this case, if the volumes V1 (t + τ ), V2 (t + τ ) and V1 (t) are known and if we know the
relative fraction of molecules in the bubble 1 at time t which is in bubble 1 (respectively
2) at time t + τ , the pressures P1 (t + τ ) and P2 (t + τ ) at time t + τ can be computed in
two steps. The relative fraction of molecules in the bubble 1 at time t which is in bubble 1
(respectively 2) at time t + τ is first determined from the computation of the sub-volumes
Numerical Simulation of Two-Phase Free Surface Flows 183

bubble 1
bubble 1 P1 (t + τ )
P1 (t), V1 (t) V1 (t + τ ) bubble 2
P2 (t + τ )
V2 (t + τ )
liquid liquid

t t+τ
Figure 14. Splitting of one bubble into two bubbles. Each molecule in the bubble number
1 at time t appears in one of the bubbles at time t + τ

bubble 1
bubble 1
P1 (t + τ )
P1 (t), V1 (t)
V1 (t + τ )
bubble 2
P2 (t + τ )
bubble 2 V2 (t + τ )
P2 (t), V2 (t)
liquid liquid

t t+τ
Figure 15. Splitting and merging of bubbles at the same time. Two bubbles at time t
lead to two other bubbles at time t + τ after merging and splitting

of the bubble 1 at the exact time of splitting, i.e. a time between t and t + τ . Then the
pressures P1 (t + τ ) and P2 (t + τ ) at time t + τ can be computed by taking into account the
agglomeration of the bubbles and the compression/decompression of each bubble.
In most cases and if the time step τ is small enough, one bubble either remains one
bubble as in Figure 11 or may be split into two bubbles, as in Figures 12 or 14, or two
bubbles may merge into one, as in Figure 13. Exceptionally, more complex situations may
happen as illustrated for instance in Figure 15. The numerical processing of all these cases
will be discussed in Section 4.3.
The mathematical description of our model is now completed. The model unknowns are
the characteristic function ϕ in the whole cavity, the velocity v and pressure p in the liquid
domain. Additional unknowns are the bubbles of gas, i.e. the connected components of the
gas domain, the constant pressure Pi in each bubble of gas and the curvature κ and the
normal vector n on the liquid-gas interface These unknowns satisfy equations (3.1), (3.2),
(3.4) and (3.6) with the boundary condition (3.5) on the free surface Γt .

4 TIME SPLITTING SCHEME

An implicit splitting algorithm is proposed to solve (3.1)-(3.5) when the pressure in the gas,
P , is computed with (3.6) and when the surface tension effects are taken into account.
Let 0 = t0 < t1 < t2 < . . . < tN = T be a subdivision of the time interval [0, T ], define
τ = tn − tn−1 the n-th time step, n = 1, 2, . . . , N , τ the largest time step.
n

Let ϕn , vn , Ωn , k n , P n , Bin , i = 1, 2, . . . , k n and κn , nn be approximations of ϕ, v, Ω,

184 Alexandre Caboussat

k, P , Bi , i = 1, 2, . . . , k (please remember that k is the number of bubbles in the gas at

given time) and κ, n respectively at time tn . Then the approximations ϕn+1 , vn+1 , Ωn+1 ,
k n+1 , P n+1 , Bin+1 , i = 1, 2, . . . , k n+1 and κn+1 , nn+1 at time tn+1 are computed by means
of an implicit splitting algorithm, as illustrated in Figure 16 for the case of a rising bubble
of gas in a domain filled with liquid.
First two advection problems are solved, leading to a prediction of the new velocity
vn+1/2 together with the new approximation of the characteristic function ϕn+1 at time
tn+1 , which allows to determine the new fluid domain Ωn+1 , the new gas domain Λ\Ω̄n+1
and the new liquid-gas interface Γn+1 . Then, the connected components of gas (bubbles)
Bin+1 , i = 1, . . . , k n+1 are tracked with a procedure explained in the following and the
pressure Pin+1 in each bubble Bin+1 is computed. Then an approximation of the curvature
κn+1 is obtained on the interface Γn+1 together with a normal vector nn+1 . Finally, a
generalized Stokes problem is solved on Ωn+1 with boundary condition (3.5) on the liquid-
gas interface, Dirichlet, slip or Signorini-type conditions on the boundary of the cavity Λ
and the velocity vn+1 and pressure pn+1 in the liquid are obtained.
This time splitting algorithm introduces an additional error on the velocities and pres-
sures which is of order O(τ 2 ) at each time step or equivalently of order O(τ ) on the whole
simulation, see e.g. [96]. On the other hand, the introduction of this splitting algorithm
permits to decouple the motion of the free surface from the diffusion step and to solve the
Stokes problem in a fixed domain. In the light of this remark, let us focus on the different
steps of the splitting algorithm in the following.
4.1 Advection Step
Solve between the times tn and tn+1 the two advection problems:
∂v
+ (v · ∇)v = 0 , (4.1)
∂t
∂ϕ
+ v · ∇ϕ = 0 , (4.2)
∂t
with initial conditions given by the values of the functions v and ϕ at time tn . This step
is solved exactly by the method of characteristics [103, 118, 119, 122, 161] and yields a
prediction of the velocity vn+1/2 and the approximation of the characteristic function of
the liquid domain ϕn+1 at time tn+1 , that is:

vn+1/2 (x + τ n vn (x)) = vn (x),

ϕn+1 (x + τ n vn (x)) = ϕn (x),

for all x belonging to Ωn . The domain Ωn+1 is then defined as the set of points such that
ϕn equals one Γn+1 = ∂Ωn+1 \∂Λ is the free surface.

4.2 Numbering of the Bubbles of gas

Given the new liquid domain Ωn+1 , the first task consists in finding the gas bubbles Bin+1 ,
i = 1, . . . , k n+1 . Then the pressure inside each bubble is computed.
The key point is to find the number of bubbles k n+1 (that is to say the number of
connected components) and the bubbles Bin+1 , i = 1, . . . , k n+1 . The algorithm for detecting
a connected component in the gas domain is the following. At each time step, given a point
P in the gas domain Λ\Ωn+1 , we first search for a function u such that −∆u = δP in
Λ\Ωn+1 , with u = 0 on Ωn+1 and u continuous. Since the solution u to this problem is
strictly positive in the connected component containing point P and vanishes outside, the
Numerical Simulation of Two-Phase Free Surface Flows 185

Advection Bubbles

v n , pn vn+1/2
ϕn , Ωn ϕn+1 , Ωn+1

κn
nn
Pin+1
Pin Bin+1
Bin

Time tn
Surface tension Diffusion

vn+1 , pn+1
κn+1
nn+1

Time tn+1

Figure 16. The splitting algorithm (from left to right, top to bottom). Two advection problems
are solved to determine the new approximation of the characteristic function ϕn+1 ,
the new liquid domain Ωn+1 and the predicted velocity vn+1/2. Then a constant
pressure Pin+1 is computed in each bubble Bin+1. The curvature κn+1 and the normal
vector nn+1 are then obtained on the liquid-gas interface. Finally, a generalized
Stokes problem is solved to obtain the velocity vn+1 and the pressure pn+1 in the
new liquid domain Ωn+1, taking into account the pressure P n+1 and curvature κn+1
on the liquid-gas interface

first bubble is found. The physical interpretation in two space dimensions is the following.
An elastic membrane is placed over the cavity Λ, deformation being impossible in the liquid
domain, a point force being applied at point P .
The above procedure is then repeated to recognize one connected component after the
other, see Figure 17. This procedure is called the numbering algorithm and is detailed
hereafter.
Recall that k(t) is the number of connected components of the gas domain at time t
and Bi(t) is the i-th connected component (i.e. bubble number i). Let ξ(t) be the bubble
numbering function, negative in the liquid region Ωt and equal to i in bubble Bi (t). At each
186 Alexandre Caboussat

11111111
00000000
00000000
11111111
00000000
11111111
×P
Θn+1
00000000
11111111
ξ n+1 = 1
00000000
11111111 000000000
111111111 111111111
000000000
00000
11111 000000000
111111111 111111
000000 000000000
111111111
Θn+1
00000
11111
Θn+1 000000000
111111111
Θn+1
000000000
111111111 000000
111111
Θn+1 000000000
111111111
000000000
111111111
00000
11111 000000
111111 ×P
Ωn+1 (liquid)111111111
000000000 n+1
Ω 000000000
111111111
Iteration 0 Iteration 1

ξ n+1 = 1 ξ n+1 = 1

11111
00000
00000
11111
00000
11111
×P
Θn+1 ξ n+1 = 2 ξ n+1 = 3 ξ n+1 = 2
00000
11111
Ωn+1 Ωn+1
Iteration 2 Iteration 3

Figure 17. Numbering algorithm of the gas bubbles. Initially the function ξ n+1 equals zero
everywhere in the gas domain. The domain Θn+1 corresponds to the set of points
in the gas region that have no bubble number (ξ n+1 (x) = 0, shaded region). At
each iteration of the algorithm a point P is chosen in Θn+1 . Problem (4.3) is solved
and a new bubble is numbered. Then, domain Θn+1 is updated and another point
P ∈ Θn+1 is chosen. The algorithm stops when Θn+1 = ∅

time step the approximations k n+1 , ξ n+1 , Bin+1 of k(tn+1 ), ξ(tn+1 ), Bi (tn+1 ) are computed
as follows.
The algorithm is initialized by setting the number of bubbles k n+1 to 0. Also, the
function ξ n+1 is set to 0 in the whole gas domain Λ\Ωn+1 and to −1 in the liquid domain
Ωn+1 . The goal is to assign to each point x in the gas an integer value ξ n+1 (x) = 0, the
so-called bubble number. The algorithm is illustrated in Figure 17 and is the following. Set

Θn+1 = x ∈ Λ : ξ n+1 (x) = 0 ;

While Θn+1 = ∅, do:

1. Choose a point P in Θn+1 ;
2. Solve the following problem: Find u : Λ → R which satisfies:

 −∆u = δP , in Θn+1 ,
u = 0, in Λ\Θn+1 , (4.3)

[u] = 0, on ∂Θn+1 ,

where δP is Dirac delta function at point P and [u] is the jump of u through
∂Θn+1 ;
3. Increase the number of bubbles k n+1 at time tn+1 , k n+1 = k n+1 + 1;

4. Define the bubble of gas number k n+1 : Bkn+1
n+1 = x ∈ Θn+1 : u(x) = 0 ;
5. Update the bubble numbering function ξ n+1 (x) = k n+1 , ∀x ∈ Bkn+1
n+1 ;
Numerical Simulation of Two-Phase Free Surface Flows 187

6. Update Θn+1 for the next iteration,


Θn+1 = x ∈ Λ : ξ n+1 (x) = 0 .

The cost of this original numbering algorithm is bounded by the cost of solving k n+1
times a Poisson problem in the gas domain. In the numerical experiments detailed hereafter
the corresponding CPU time used to solve the Poisson problems was always less than 10
percent of the total CPU time.

4.3 Computation of the Pressure in the Gas

Once the connected components of gas are numbered, an approximation Pin+1 of the pres-
sure in bubble i at time tn+1 is computed following the description of Section 3.2. The
pressure is constant inside each bubble of gas and is computed with the ideal gas law (3.6),
except for bubbles in contact with a valve which have atmospheric pressure, see Figure 10.
In the case of a single bubble traveling in the liquid, see Figure 11, the law of ideal gas
yields:

P n+1 V n+1 = P n V n ,

which means that the number of molecules inside the bubble is conserved between time tn
and tn+1 . In the case of the merging of two bubbles, see Figure 13, this relation becomes:

P1n+1 V1n+1 = P1n V1n + P2n V2n ,

which express again conservation of the number of molecules between tn and tn+1 . These
are the two simplest situations and more complex pictures can be seen in the frame of free
surface flows in complex geometries since bubbles may merge and divide at the same time
and the topology of the gas domain may change.
The general case of splitting and merging of bubbles (see Figures 14, 15) is described
in the following. Let Bin , Pin , Vin , i = 0, . . . , k n be the connected components of gas and
their related pressure and volume at time tn and Bin+1 , Pin+1 , Vin+1 i = 0, . . . , k n+1 the
same variables at time tn+1 . The bubble Bin may split in different parts between time tn
and time tn+1 . Each of these parts contributes to a bubble Bjn+1 at time tn+1 . The volume
n+1/2
fraction of bubble Bin which contributes to bubble Bjn+1 is noted Vi,j . The computation
of the pressure is then decomposed in two steps, as illustrated in Figure 18. First the
n+1/2
volume fraction contributions Vi,j are computed. Then the pressure in the bubble Bjn+1
is computed by addition of the contributions of the different bubbles at time tn :

kn
1 n+1/2
Pjn+1 = Pin Vi,j . (4.4)
Vjn+1 i=0

In practice, the most frequent cases encountered in the simulations are i) a single bubble
staying alone, see Figure 11, ii) the merging of two bubbles into one, see Figure 13, or iii)
the splitting of one bubble into two, see Figures 12 and 14. However, these splitting and/or
merging of bubbles may happen anywhere in the cavity, this being the case in the examples
of Section 6. The above procedure allows all possible cases to be considered while conserving
the number of gas molecules trapped by the liquid.
188 Alexandre Caboussat

B1n B2n
P1n , V1n P2n , V2n t = tn

n+1/2 n+1/2 n+1/2

P1n , V1,1 P2n , V2,1 P2n , V2,2

B1n+1 , P1n+1 , V1n+1 B2n+1 , P2n+1 , V2n+1

t = tn+1

Figure 18. At each time step, the merging/division of bubbles is split in two parts. First
n+1/2
Vi,j , the volume fraction of bubble Bin that contributes to bubble Bjn+1
is computed. Then the pressure Pin+1 is computed from conservation of the
number of molecules

4.4 Computation of the Curvature

Then the curvature of the free surface Γn+1 and the normal vector nn+1 are computed.
Since the characteristic function ϕn+1 is not regular, it is first mollified, see [162], in order
to obtain a smooth function ϕ̃n+1 such that the liquid-gas interface is given by the level line
{x ∈ Λ : ϕ̃n+1 (x) = 1/2}, with ϕ̃n+1 < 1/2 in the gas domain and ϕ̃n+1 > 1/2 in the liquid
domain. At each time step, the normal vector nn+1 , directed outside the liquid domain
towards the gas domain, is given by the normalized gradient of ϕ̃n+1 and the curvature
κn+1 on the liquid-gas interface is then given by (see e.g. [111, 134]):
∇ϕ̃n+1 ∇ϕ̃n+1
nn+1 = − , κn+1 = divnn+1 = −div , (4.5)
||∇ϕ̃n+1 || ||∇ϕ̃n+1 ||
where ||·|| denotes the Euclidean norm in Rd , d = 2, 3.
Let us turn to the smoothing of the characteristic function ϕn+1 . Let Kε (x) be a kernel
function with the following properties: the function Kε (x) has a compact support and is
radially-symmetric, monotonically decreasing with respect to r = ||x|| and is normalized.
The convolution of ϕn+1 with Kε leads to a smoothed volume fraction of liquid ϕ̃n+1 defined
by:

ϕ̃n+1 (x) = ϕn+1 (y)Kε (x − y)dy, ∀x ∈ Λ . (4.6)
Λ

The kernel used in our algorithm has been proposed in [162] and is given by:
  4

 ||x|| 2
Kε (x) = C 1− , if ||x|| ≤ ε , (4.7)
 ε

0, else
Numerical Simulation of Two-Phase Free Surface Flows 189

where C is a normalization coefficient. Hence the normal vector nn+1 and the curvature
κn+1 are approximated by (4.5).
Note that an exact expression for the first derivatives of ϕ̃n+1 is

∂ ϕ̃n+1 ∂Kε
(x) = ϕn+1 (y) (x − y)dy, i = 1, . . . , d . (4.8)
∂xi Λ ∂xi

4.5 Diffusion Step

The diffusion step consists in solving a generalized Stokes problem on the domain Ωn+1 using
the predicted velocity vn+1/2 and the boundary condition (3.5). The following backward
Euler scheme is used:

vn+1 − vn+1/2  
ρ n
− 2div µD(vn+1 ) + ∇pn+1 = f in Ωn+1 , (4.9)
τ
n+1
div v =0 in Ωn+1 , (4.10)

where vn+1/2 is the prediction of the velocity obtained after the advection step. The bound-
ary conditions on the free surface between the fluid and the bubble number i depend on the
gas pressure Pin+1 and the curvature κn+1 and are given by (3.5). The weak formulation
corresponding to (4.9) (4.10) and (3.5) therefore consists in finding vn+1 and pn+1 such
that vn+1 satisfies the essential boundary conditions on the boundary of the cavity Λ and
 
vn+1 − vn+1/2
· wdx + 2µ D(vn+1 ) : D(w)dx
n+1
Ω τn Ωn+1
 
− p n+1
div wdx − f · wdx − q div vn+1 dx (4.11)
Ωn+1 Ωn+1 Ωn+1
kn+1 
+ (Pin+1 − σκn+1 )nn+1 · wdS = 0 ,
i=1 ∂Ωn+1 ∩∂Bin+1

for all test functions (w, q) such that w vanishes on the boundary of the cavity where
essential boundary conditions are enforced.

5 SPACE DISCRETIZATION: A TWO-GRIDS METHOD

Advection and diffusion phenomena being now decoupled, Equations (4.1) (4.2) are first
solved using the method of characteristics on a structured mesh of small cells in order to
reduce numerical diffusion and have an accurate approximation of the liquid region, see
Figure 19.
Assume that the grid is made out of cubic cells of size h, each cell being labeled by
indices (ijk). Let ϕnijk and vijk
n
be the approximate value of ϕ and v at the center of cell
number (ijk) at time t . The unknown ϕnijk is the volume fraction of liquid in the cell ijk,
n

see [4], and is the numerical approximation of the characteristic function ϕ at time tn which
is piecewise constant on each cell of the structured grid. The advection step on cell number
(ijk) consists in advecting ϕnijk and vijk
n by τ n vn and then projecting the values on the
ijk
structured grid. An example of cell advection and projection is presented in Figure 20 in
two space dimensions.
In order to enhance the quality of the volume fraction of liquid, post-processing proce-
dures have been implemented. We refer to [98, 99] for a detailed description in two and
190 Alexandre Caboussat

Figure 19. Two-grids method, two-dimensional representation. Advection step is solved

on a structured mesh of small cubic cells (right), while diffusion step and
bubbles treatment are solved on a finite element unstructured mesh (right)

ϕn
ij ϕn
16 3 16ij

ϕn ϕn
3 16ij 9 16ij

τ n vij
n

index j
ϕnij

index i

Figure 20. An example of two dimensional advection of ϕn n n

ij by τ vij , and projection on
the grid. The advected cell is represented by the dashed lines. The four cells
containing the advected cell receive a fraction of ϕn
ij , according to the position
of the advected cell

three space dimensions. The first procedure reduces numerical diffusion and is a simplified
implementation of the SLIC (Simple Linear Interface Calculation) algorithm [109].
Consider cell number (ijk) being partially filled with liquid (this results from numerical
diffusion), let ϕnijk be the corresponding volume fraction of liquid, this value being less than
one. Instead of advecting ϕnijk and then projecting on the grid, the liquid is first pushed
on the sides of the cell, then it is advected and projected on the grid. A two-dimensional
example is reported in Figure 21.
The critical point is then to decide how to push the volume fraction of liquid in a given
Numerical Simulation of Two-Phase Free Surface Flows 191

3 1 1 1
16 4 16 4 0 0

9 1 3 1 1 0
16 4 16 4 41

1
ϕnij = 4
1

Figure 21. Effect of the SLIC algorithm on numerical diffusion. An example of two dimensional
advection and projection when the volume fraction of liquid in the cell is ϕn 1
ij = 4 .
Left: without SLIC, the volume fraction of liquid is advected and projected on
3 1
four cells, with contributions (from the top left cell to the bottom right cell) 16 4
,
1 1 9 1 3 1
, ,
16 4 16 4 16 4
. Right: with SLIC, the volume fraction of liquid is pushed at one
corner, then it is advected and projected on one cell only, with contribution 14

cell along the sides of this cell. For a given cell, the choice depends on the volume fraction
of liquid of the neighbor cells. Precise examples are given in [98, 99] for the two- and
three-dimensional cases.
The second procedure allows to guarantee the conservation of the mass of liquid. When
the computed values ϕn+1 ijk are greater than one, a fraction of the liquid contained in the
cavity is lost. The aim of the decompression algorithm is to produce new values ϕn+1 ijk
which are between zero and one. The algorithm is as follows. At each time step, all the
cells having values ϕn+1
ijk greater than one (strictly), or between zero and one (strictly) are
sorted according to their values ϕn+1 ijk . This can be done in an efficient way using quick
sort algorithms. The cells having values ϕn+1 ijk greater than one are called the dealer cells,
n+1
whereas the cells having values ϕijk between zero and one are called the receiver cells. The
algorithm then consists in moving the fraction of liquid in excess in the dealer cells to the
receiver cells.
n+1/2
Once values ϕn+1
ijk and vijk have been computed on the cells, values of the fraction
n+ 1
of liquid ϕn+1
P and of the velocity field vP 2 are computed at the nodes P of the finite
element mesh by multigrids restriction methods (see e.g. [67]): for any vertex P of the finite
element mesh let ψP be the corresponding basis function (i.e. the continuous, piecewise
linear function having value one at P , zero at the other vertices). All the tetrahedrons K
containing vertex P and all the cells (ijk) having center of mass Cijk contained in these
tetrahedrons are considered. Then, ϕn+1P , the volume fraction of liquid at vertex P and
time t n+1 is computed using the following formula:

ψP (Cijk )ϕn+1
ijk
K ijk
P ∈K Cijk ∈K
ϕn+1
P = .
ψP (Cijk )
K ijk
P ∈K Cijk ∈K
192 Alexandre Caboussat

1
The same kind of formula is used to obtain the predicted velocity vn+ 2 at the vertices of
the finite element mesh. When these values are available at the vertices of the finite element
mesh, the liquid region is defined as follows. An element of the mesh is said to be liquid if
(at least) one of its vertices P has a value ϕn+1
P > 0.5. The computational domain Ωn+1
used for solving (4.11) is then defined to be the union of all liquid elements.
Numerical experiments reported in [98, 99] have shown that choosing the size of the
cells of the structured mesh approximately 5 to 10 times smaller than the size of the finite
elements is a good choice to reduce numerical diffusion. Furthermore, since the character-
istics method is used, the time step is not restricted by the CFL number (which is the ratio
between the time step times the maximal velocity divided by the mesh size). Numerical re-
sults in [98, 99] have shown that a good choice generally consists in choosing CFL numbers
ranging from 1 to 5.
In number of industrial mold filling applications, the shape of the cavity containing
the liquid (the mold, an engine carter for instance) is complex. Therefore, a special data
structure has been implemented in the three-dimensional case in order to reduce the memory
requirements used to store the cell data, see [99]. An example is proposed in Figure 22.
The cavity containing the liquid is meshed into tetrahedrons. Without any particular cells
data structure, a great number of cells would be stored in the memory without being never
used. The data structure that has been adopted uses three levels to define the cells. At the
coarsest level, the so-called window level, the cavity is meshed into blocks, which are glued
together. Each window is then subdivided into cubes, this intermediate level is called the
block level. Finally, each block is cut into smaller cubes, namely the cells (ijk).
When a block is free of liquid (empty), it is switched off, that is to say the memory
corresponding to the cells is not allocated. When liquid enters a block, the block is switched
on, that is to say the memory corresponding to the cells is allocated. This data structure
was already used in industrial applications pertaining to dendritic solidification [123].



finite element mesh window level
(tetrahedrons)














block level k






















j


















































i
cell level

Figure 22. The hierarchical Window-block-Cell data structure used to implement the cells
advection

The numbering of the bubbles of gas requires to solving several Poisson problems (4.3).
An element of the mesh belongs to the gas domain if (at least) one of its vertices P has a
value ϕn+1
P < 0.5. The computational domain Θn+1 used for solving (4.11) is then defined
to be the union of all elements in the gas domain for the first iteration. These Poisson
Numerical Simulation of Two-Phase Free Surface Flows 193

B1n B1n+1 B1n B1n+1

liquid liquid

Figure 23. Computation of the pressure. Two cases can occur. Left: the bubble at time
tn+1 intersects one of the previous bubbles at time tn , the pressure can be
computed with (4.4). Right: the bubble at time tn+1 does not intersect any
bubble at previous time; the pressure is then computed with (5.1)

problems are solved on the finite element unstructured mesh, using piecewise linear finite
elements.
The pressure inside each bubble of gas is computed with (4.4) and the approximations
n+1/2
of the fractions of volumes Vi,j are computed on the finite element mesh. Two situations
may occur, as illustrated in Figure 23. When bubble j at time tn+1 intersects at least one
n+1/2
bubble of time tn , that is when there is at least one index i such that Vi,j = 0, then the
n+1
pressure Pj is computed with (4.4) because it is assumed that a part of bubble j at time
tn+1 is made by bubble i at time tn .
n+1/2
On the other hand, if the time step τ is too large (that is if Vi,j = 0, for all i =
n n+1 n
0, . . . , k ), the bubble Bj may not intersect any bubble of time t , see Figure 23 (right).
In this case, the origin of bubble number j is unknown. Relation (4.4) is useless and then
the pressure Pjn+1 is computed by dividing the remaining number of molecules at time tn
by the remaining volume at time tn+1 :
 n n+1/2 
P V
Pjn+1 = r
; (5.1)
(V n+1 )r

where
 
  kn kn+1
P n V n+1/2 = Pin V n+1 − n+1/2 
Pin Vi,j ;
i
r
i=0 j=0
kn+1 kn

n+1/2
(V n+1
)r = Vjn+1 − Vi,j .
j=0 i=0

This latter case appears generally when the time step is too large compared to the size
of the bubble. This procedure then permits to conserve the mass of gas in any situations.
An approximation of the curvature of the interface could be computed on the structured
grid of small cells with finite differences schemes for instance. Several methods may be found
in [80, 130] for instance. Nevertheless, these methods need a very accurate reconstruction
of the interface with sophisticated algorithms such as PLIC schemes for instance, see e.g.
[4]. On the other hand, the method proposed in the following uses a formulation on the
finite element unstructured mesh.
194 Alexandre Caboussat

Let Th be the triangulation of the cavity Λ and Ωn+1 h be the approximation of the liquid
domain composed by elements K in Th and Γh the approximation of Γn+1 . Let ψPj be
n+1

the basis functions of the piecewise linear finite element space associated to each node Pj ,
j = 1, . . . , N in the cavity. Finally, let the piecewise linear finite elements space be denoted
by Xh1 (Λ).
The approximation of (4.6) is obtained by numerical integration. Let ϕ̃n+1 n+1
h , ϕh be the
n+1 n+1 1
approximations of ϕ̃ ,ϕ in Xh (Λ). The numerical integration of (4.6) can either be
made on the regular grid of small cells or on the finite element mesh. For memory purposes,
the approximation of (4.6) is computed using the finite element mesh, that is:

1
ϕ̃n+1
h (Pi ) = |K| h (Pj )Kε (Pj − Pi ),
ϕn+1 i = 1, . . . , N, (5.2)
d+1
K∈Th Pj ∈K

where |K| denotes the surface (resp. volume) of element K. This permits to obtain an
approximation of the smoothed volume fraction of liquid for each grid point Pi of Th .
The approximation of the curvature (4.5) is computed on the finite element mesh grid
points. The procedure detailed hereafter gives a natural expression of the curvature in the
frame of finite elements and is not CPU time consuming.
The normal vector nn+1
h is given by (4.5) (4.8) (with ϕn+1 replaced by ϕn+1
h ) at each
grid point Pj . The curvature κn+1 in (4.5) is approximated by its L2 -projection on Xh1 (Λ)
with mass lumping, denoted by κn+1 h . Let us denote by Rh the Lagrange interpolant on
1
Xh (Λ). After integration by parts, this projection leads to the relation:
 
  ∇ϕ̃n+1
Rh κn+1
h ψPj dx = −div

n+1h


ψ dx, ∀j = 1, . . . , N . (5.3)
Λ Λ


∇ϕ̃

Pj
h

Relationship (5.3) leads to:

  
d+1  1 ∇ϕ̃n+1 (P )

κn+1
h (Pj ) = |K| 

h n+1
i
 ∇ψPj
K
|Ωj | d+1

∇ϕ̃


K∈Th Pi ∈K h

1 ∇ϕ̃n+1 (P ) 
−

h n+1
j
nΛ,h (Pj ) |∂K|
 , (5.4)
∂K⊂∂Λ
d

∇ϕ̃h


Pj ∈K


for all j in 1, . . . , N , where |Ωj | = K,Pj ∈K |K| and nΛ,h denotes the approximation of the
external normal vector to the cavity Λ.
One artificial value of the curvature is then given in the whole cavity by the piecewise
linear function κn+1 h defined by (5.4), that is for each level line of the smoothed volume
fraction of liquid ϕ̃n+1 n+1
h . The restriction of κh to the nodes Pj lying on Γn+1
h is then used
to compute (3.5).
Finally the diffusion step consists in solving the Stokes problem (4.11). Let vhn+1 (resp.
n+1
ph ) be the piecewise linear approximation of vn+1 (resp. pn+1 ). The Stokes problem is
solved with stabilized P1 − P1 finite elements (Galerkin Least Squares, see [50]) and consists
Numerical Simulation of Two-Phase Free Surface Flows 195

in finding the velocity vhn+1 and pressure pn+1

h such that:
 n+1/2 
vhn+1 − vh
wdx + 2µ D(vhn+1 ) : D(w)dx
Ωn+1 τn Ωn+1
h
  h

− fwdx − pn+1
h div wdx − div vhn+1 qdx (5.5)
Ωn+1 Ωn+1 Ωn+1
 h h h

+ (P n+1 − σκn+1 n+1

h )nh wdS
Γn+1
h
 n+1/2

vhn+1 − vh
− αK + ∇pn+1
h −f · ∇qdx = 0 ,
K τn
K⊂Ωn+1
h

for all w and q the velocity and pressure test functions, compatible with the boundary
conditions on the boundary of the cavity Λ.
The restriction of the continuous piecewise linear approximation vhn+1 at the center
n+1
of each cells Cijk permits to obtain a value of the velocity vijk on each cell ijk of the
structured grid for the next time step.

6 NUMERICAL RESULTS

Numerical results in two and three space dimensions are presented to validate our model,
both in the frame of mold filling and in the frame of bubbles and droplets simulations.
All the computations were performed on computers with single processor Pentium Xeon
2.8 GHz CPU, 3 Gb Memory and running under Linux operating system. The results are
post-processed with CalcosoftT M , EnsightT M or XD3DT M softwares.

6.1 Convergence of the Volume Fraction of Liquid Given a Prescribed Velocity

The goal of this section is to validate the computation of the volume fraction of liquid using
the algorithm described in Section 5. Standard two-dimensional test cases are taken from
[4, 130]. Surface tension effects are not taken into account in this section. The translation
and rotation of a mass of fluid are presented, as well as stretching flows examples.
The first situation is the translation of a circle of liquid, with given velocity and without
external forces. The cavity domain is the 0.1 × 0.1 square and the center of a circle of
radius 0.015 is initially located at (0.02, 0.05). The advection velocity is horizontal and
equals to 1. Three different meshes are used. The coarse finite element mesh is constituted
by 40 × 40 squares, each divided in 4 triangles, the middle mesh is divided in 80 × 80, while
the finer mesh is composed by 120 × 120 squares. The underlying regular grid is composed
respectively by 120 × 120 cells, 240 × 240 cells and 360 × 360 cells.
Figure 24 illustrates the position of the interface at times t = 0 and t = 0.06. The time
step is equal to τ = 0.002. The total volume of liquid is conserved as well as the mass of
liquid.
In the second situation,the same circle of liquid is rotated with given velocity, following
the situation described in [4]. The advection velocity is given by v(x, y) = 25(0.05 − y, x −
0.05). The meshes and time step are the same as the ones used for the translation test case.
Figure 25 illustrates the position of the interface at times t = 0 and t = 0.126, i.e. after a
half rotation.
The last test case deals with stretching flows. A classical test case widely treated in the
literature is considered, see e.g. [4, 130] which is the “vortex-in-a-box” test case. The initial
196 Alexandre Caboussat

Figure 24. Translation of a circle of liquid, representation of the computed interface at

initial time and after t = 0.06. Left: coarse mesh, middle: middle mesh, right:
fine mesh

Figure 25. Rotation of a circle of liquid, representation of the computed interface at initial
time and after t = 0.126. Left: coarse mesh, middle: middle mesh, right: fine
mesh

liquid domain is a circle of radius 0.015 with its center located in (0.05, 0.075). The given
velocity is given by the stream function:

ψ(x, y) = 0.01π sin2 (πx/0.1) sin2 (πy/0.1) cos(πt/2) ,

which stretches the initial circle of liquid with a given velocity. This velocity is periodic in
time, so that the initial liquid domain is reached after a time T = 2. Figure 26 illustrates
the liquid-gas interface for the three meshes. The interface with maximum deformation
and the interface after one period of time are represented and numerical results show the
efficiency of the method.
6.2 Linear Filling
Water enters a rectilinear three-dimensional channel and compresses the gas contained inside
the channel. Surface tension effects are not taken into account since we want to study more
precisely the computation of the gas pressure. The dimensions of the channel are 0.5 m
×0.08 m ×0.1 m and water is injected at horizontal speed 4.2 m/s. At time T = 0.5/4.2 s,
the channel is filled. Figure 27 shows a two-dimensional cut of the channel. Slip boundary
conditions are enforced on the lateral sides of the cavity. The mesh is a regular grid of 5049
nodes and 24000 tetrahedrons.
Numerical Simulation of Two-Phase Free Surface Flows 197

Figure 26. Single vortex test case, representation of the computed interface at times t = 1
(maximal deformation) and t = 2 (return to initial shape). Left: coarse mesh,
middle: middle mesh, right: fine mesh

- -
- -
- - trapped gas
liquid
- - pressure P
- - volume V
- -

Figure 27. Linear filling of a three-dimensional channel: two-dimensional cut. The inter-
face moves with constant velocity v m/s

For this simple test case, the exact volume of gas is given by 0.08 × 0.1 × (0.5 − 4.2 × t)
m3 , so that the exact volume and pressure can be computed with the ideal gas law. Three
different regular meshes are considered, a coarse mesh with 1380 nodes and 6000 elements,
a middle mesh with 9449 nodes and 48000 elements and a finer mesh with 69657 nodes and
384000 elements. Final time is T = 0.120 s and the time step is respectively τ = 0.002 s,
τ = 0.001 s and τ = 0.0005 s. The total CPU time is respectively 5 minutes, 92 minutes and
25 hours and is multiplied approximately by 24 = 16 each time that the mesh size and time
step are divided by 2. The number of operations is thus of order O(N 4 ), where N is the
number of discretization points along each axis. This is the same order of the computational
cost required to solving a Laplace problem with the conjugate gradient algorithm without
preconditioning (see for instance [6]). Figure 28 shows that the computation of the pressure
in the gas converges when the mesh size h tends to zero, the rate of convergence being
approximately O(h1/4 ).

6.3 S-shaped Channel

An S-shaped channel lying between two horizontal plates is filled. Results in two and three
space dimensions are compared with experiment [133]. The channel is contained in a 0.17
m×0.24 m rectangle. In the three-dimensional case, the distance between the two horizontal
plates is 0.008 m. Water is injected with constant velocity 8.7 m/s which corresponds to
the experimental value reported in [133]. Due to the large velocities involved, the surface
tension effects are not taken into account for this test case. A valve is located at the top
of the channel, as in Figure 10, allowing gas to escape. Density and viscosity are taken to
be respectively ρ = 1000 kg/m3 and µ = 0.01 kg/(ms) and initial pressure in the gas is
Patmo = 101300.0 Pa.
198 Alexandre Caboussat

theoretical
coarse mesh
600000 middle mesh
fine mesh

500000

400000

300000

200000

0.05 0.06 0.07 0.08 0.09 0.1

0.5

0
3
-0.5 3
3
-1

-1.5

Error 3
-2 slope 1
slope 1/2
slope 1/4
-2.5
-6.2 -6 -5.8 -5.6 -5.4 -5.2 -5 -4.8 -4.6

Figure 28. Filling of a rectilinear channel with compression of gas. Comparisons between
three different mesh sizes. Top: pressure in the gas function of time, bottom:
error function of the mesh size on a log-log scale

Several meshes are considered. In two dimensions, the coarser mesh has 3483 nodes and
6418 elements; the middle mesh has 8249 nodes and 15654 elements while the fine mesh is
made out of 14550 nodes and 27972 elements, see Figure 29. The three-dimensional meshes
are constructed using 5, respectively 8 and 10 layers of the 2D mesh.
When comparing numerical results to experimental ones, we have observed that the
liquid goes faster in the simulations than in the experiments. This is probably due to inexact
slip boundary conditions. On the other hand, due to large Reynolds numbers (Re  106 ),
noslip boundary conditions are not conceivable since they would require extremely fine
layered meshes along the boundary of the cavity. As a remedy, slip boundary conditions
Numerical Simulation of Two-Phase Free Surface Flows 199

Figure 29. Meshes used for the computations of the S-shape channel: left: 2D coarse
mesh, right: 3D mesh extracted with five layers of 2D coarse mesh

are enforced, but a turbulent viscosity is added, Equation (3.3), the coefficient αT being
proportional to h2 , as proposed in [146]. Since the ratio between Capillary number and
Reynolds number is very small (Ca  1.5), surface tension effects can be neglected.
Numerical results are first presented with the coarser mesh and αT = 4h2 . The final
time is T = 0.00532 s and the time step is τ = 0.0001 s. In Figure 30, the experiment is
compared to 2D and 3D computations when the influence of the surrounding gas is taken
into account and to 2D computations without the influence of gas. When the gas is not
taken into account, numerical results show that the bubble of trapped gas at the bottom
of the cavity vanishes rapidly. On the other side, when the effects of the surrounding gas
onto the liquid are taken into account, numerical results are much closer to experiment.
The CPU time for the simulations in two space dimensions is approximately 14 minutes
without the bubbles computations and 15 minutes with the bubbles computations. In three
space dimensions, these CPU times become 319 minutes without taking into account the
gas effect and 344 minutes with the bubbles computations. Most of the CPU time is spent
to solve Stokes problem.
The influence of the mesh size is reported in Figure 31 and 32. The time steps are
τ = 0.0001 s for the coarse mesh, τ = 0.00008 s for the middle mesh and τ = 0.00005 s
for the fine mesh. The size of the cells of the structured mesh used for advection step
is approximately 5 to 10 times smaller than the size of the finite elements, see [98, 99].
Numerical results show that the behavior of bubbles is well simulated even though the fluid
flow is slightly too fast. The total CPU time for 3D computations to reach final time is
approximately 29 hours for the middle mesh and 110 hours for the finer mesh.
Numerical results are presented for several coefficients αT in Figure 33 and show in
particular that the fluid velocity decreases when αT increases.
200 Alexandre Caboussat

Figure 30. S-shaped channel: influence of gas bubbles. Computations with coarse mesh
and αT = 4h2 in (3.3). Column one: 2D results without bubbles, column two:
2D results with bubbles, column three: 3D results with bubbles in the middle
plane, and column four: experimental results [133]. First row: time equals
7.15 ms, second row: 25.3 ms, third row: 39.3 ms and fourth row: 53.6 ms

6.4 Filling of a 3D Mold With Four Arms

A mold with four arms is considered, see Figure 34. Water enters from the top of the
mold with velocity 4.2 m/s. Two arms are closed, while there is a valve at the end of
Numerical Simulation of Two-Phase Free Surface Flows 201

coarse mesh middle mesh fine mesh

Figure 31. S-shaped channel: convergence with mesh size. Computations with gas bub-
bles, αT = 4h2 in (3.3), 2D results. Left: coarse mesh, middle: middle mesh,
right: fine mesh and extreme right: experimental results [133]. First row: time
equals 25.3 ms and second row: 39.3 ms

coarse mesh middle mesh fine mesh

Figure 32. S-shaped channel: convergence with mesh size. Computations with gas bub-
bles, αT = 4h2 in (3.3), 3D results. Left: coarse mesh, middle: middle mesh,
right: fine mesh and extreme right: experimental results [133]. First row: time
equals 25.3 ms and second row: 39.3 ms
202 Alexandre Caboussat

αT = 0 αT = 2h2 αT = 4h2 αT = 6h2

Figure 33. S-shaped channel: influence of the coefficient αT in (3.3). Computations with
gas bubbles, 3D results, middle mesh. First row: time equals 25.3 ms, second
row: time equals 39.3 ms

the other arms so that gas can escape. Viscosity is µ = 0.01 kg/(ms), while density is
ρ = 1000 kg/m3 . Initial pressure in the gas is Patmo = 101 300.0 Pa. Gravity forces and
surface tension effects are neglected and αT = 0 (no turbulence modeling). The mesh has
31961 vertices and 168 000 elements and the cells grid contains approximately 50 000 000
cells. The final time of simulation is 0.5 s with time step τ = 0.001 s. The CPU time is
approximately 24 hours. For this test case, the goal is to see the influence of gas on the
filling of each arm. Figures 35 and 36 show that if a valve is located at the end of an arm,
the arm is filled significantly faster.

6.5 Filling of a Circular 3D Mold

A circular mold is now considered. Water enters from the top of the mold with velocity
8.7 m/s. This is an axisymmetric case, but the computations are made for the three-
dimensional case. There is a valve at the bottom of the mold so that gas can escape.
Viscosity is µ = 0.01 kg/(ms), while density is ρ = 1000 kg/m3 . Initial pressure in the
gas is Patmo = 101300.0 Pa. Gravity forces and surface tension effects are neglected and
αT = 0 (no turbulence modeling). The finite element mesh is made out of 47137 nodes and
37152 tetrahedrons. The structured grid is made out of 1 200 000 cells. The final time of
simulation is 4.0 s with time step τ = 0.001 s. The CPU time is approximately 35 hours.
Figure 37 illustrates the position of the liquid domain at different times of the simulation.
Numerical results show that the symmetry is conserved, but numerical diffusion is large
due to the coarse mesh.
Numerical Simulation of Two-Phase Free Surface Flows 203

Figure 34. 3D mold filling: Finite element mesh

6.6 Convergence Results for the Computation of the Curvature

Test cases in two space dimensions are considered here to validate the computation of the
surface tension effects and convergence results are obtained.
The test case of a stationary circular droplet of water lying√in a square in the absence
of gravity field is considered. The radius of the droplet is 0.001 and the size of the
square is 0.1 × 0.1. Theoretical signed curvature is constant and equal to −31.62 on the
boundary of the liquid domain. Several finite element meshes are considered, namely regular
structured grids of squares, each square being divided in four triangles and unstructured
(isotropic) meshes. An optimal value of the smoothing parameter ε appearing in (4.7) is
first to determine. Table 1 shows the relative error (in percent) for several mesh sizes h and
smoothing parameters ε.
If ε → 0, the derivatives become difficult to approximate since the characteristic function
is discontinuous. On the other hand, taking a large value of ε leads to an inaccurate
approximation of the curvature (numerical diffusion). We look for a constant minimal
value of the smoothing parameter ε. According to Table 1, ε can be chosen approximately
by 0.05, independently of the mesh size. This value ensures approximately one percent of
error on the computation of the curvature for reasonable mesh sizes h.
For fixed smoothing parameter ε, the convergence of the approximation of the curvature
is discussed. Structured finite element meshes of squares are considered first from 40 × 40
squares to 400×400 squares, each square being divided in four triangles. The size of the cells
of the structured mesh used for advection step is approximately 5 to 10 times smaller than
the size of the finite elements. Figure 38 illustrates the convergence of the approximation
of the curvature when the mesh size tends to zero for ε = 0.05. The convergence order
204 Alexandre Caboussat

Figure 35. 3D mold filling. Liquid region: from top to bottom, left to right at times 0.05,
0.10, 0.15 and 0.20 s

is O(h). Similar results are obtained with unstructured (isotropic) meshes, with the same
order of convergence for the approximation of the curvature.
Since the time splitting algorithm presented in Section 4 is an order one algorithm (see
also [98, 99] for a numerical verification without taking into account the compressible gas)
an order one algorithm for the computation of the curvature is sufficient. The choice of the
smoothing parameter ε in function of h would permit to obtain a better convergence order,
see [162] for instance.
In the case of this stationary droplet, the numerical simulation introduces spurious
velocities around the interface due to the imposition of surface tension forces. In the physical
situation the velocity is identically zero. Hence these velocities are often called spurious
currents and their amplitude is widely discussed in the literature, see for instance [16, 127].
The amplitude depends numerically on the ratio between surface tension forces, i.e. σκ, and
the viscosity of the fluid, i.e. µ, see for instance [132]. In Table 2, the maximal amplitude of
Numerical Simulation of Two-Phase Free Surface Flows 205

Figure 36. 3D mold filling (continued). Volume fraction of liquid: from top to bottom,
left to right at times 0.25, 0.30, 0.35 and 0.40 s

the spurious currents is given after five time steps (τ = 0.001) for different mesh sizes. The
ratio between surface tension forces and the viscosity is σκ/µ = 31.6 · 0.0738/0.01  233.3.
Results compare well for instance with [162] but are not as good as methods which emphasize
a very accurate reconstruction of the interface, see [127].

6.7 Bubbles and Droplets Simulations

Deformed droplets have been widely treated in the literature [16, 107, 132, 155]. Here an
initially oval droplet of liquid in the absence of gravity forces is considered first. Since the
surface tension forces are more important in the extremities where the curvature is larger,
the shape of the droplet tends to become a circle. In our case, the viscosity is given by µ = 1
kg/(ms), while the density is ρ = 1000 kg/m3 . The surface tension coefficient is σ = 7.038
N/m. The external pressure of the surrounding gas is neglected. The initial liquid domain
206 Alexandre Caboussat

Figure 37. Circular 3D mold filling. Liquid region at times 0.5, 1.5, 2.5 and 3.5 (left to
right, top to bottom)

HH h
HH 0.0025 0.0016 0.00125 0.001 0.0008
ε H
0.5 3.85 1.51 1.45 1.10 0.85
0.1 3.85 1.42 1.49 1.10 0.85
0.05 3.82 1.42 1.42 1.10 0.85
0.01 4.14 1.48 1.45 1.12 0.89
0.005 8.03 1.67 1.49 1.15 1.04

Table 1. Circular droplet case: smoothing of volume fraction of liquid for projection
method. Values of the maximal relative error (in percent) on the curvature of
the circular droplet function of parameter ε and of mesh size h

is given by  
2 2
x−4 y−4
Ω0 = (x, y) ∈ R :
2
+ ≤1 .
2 3
A coarse structured mesh is made out of 20 × 20 squares, each divided in 4 triangles. A
middle mesh (40 × 40 squares) and a fine mesh (80 × 80 squares) are also considered. The
regular grid of small cells is constituted respectively by 500×500, 1000×1000 and 2000×2000
Numerical Simulation of Two-Phase Free Surface Flows 207

0
10
numerical
slope 1

error on the curvature

−1
10

−2
10
−4 −3 −2
10 10 10
h

Figure 38. Circular droplet test case: convergence error for the approximation of the
curvature

h |v|max
0.0025 0.0041
0.00125 0.0084

Table 2. Value of the maximal amplitude of spurious currents in function of the mesh size

cells. The time step is τ = 0.1, 0.05 and 0.025 s respectively. Since our advection scheme
uses the characteristics method, which is known to be dissipative, the oval droplet becomes
a circle after some oscillations. According to [89] and as reported in [155], the period ξ of
the first mode of oscillations of a droplet of water into the vacuum is given by:
6σ
ξ2 = , (6.1)
ρa3
where a is the radius of the circle in equilibrium. For a = 2.45, ρ = 1000 and σ = 7.038, the
theoretical value of the time required to make one complete oscillation is 2π/ξ = 0.117254
s, while the observed corresponding value in the simulation are respectively 0.13 s. for the
coarser mesh, 0.122 s. for the middle mesh and 0.115 s. for the finer mesh. This shows that
the diffusion of the algorithm decreases with the mesh size, with order one, as illustrated
in Figure 39.
 Same results may be obtained 
in the three dimensional case. For instance let Ω0 =
2 
 2 2
(x, y, z) ∈ R3 : x−0.05
0.8 + y−0.05
1.8 + z−0.05
0.8 ≤ (0.025)2 . Figure 40 illustrates the
position of the liquid-gas interface at different times for a coarse finite element mesh made
out of 96 000 tetrahedrons. The CPU time used for this computation is approximately 10
hours to achieve 400 time steps.
Let us consider finally the rising of a bubble. Consider a bubble of air initially at the
bottom of a cylinder filled with water. Under gravity forces, the bubble rises until reaching
208 Alexandre Caboussat

0
10
numerical
slope 1

relative error on time period

−1
10

−2
10
−2 −1 0
10 10 10
h

Figure 39. Oscillating droplet: Log-scale plot of the relative error on the time period for
the first oscillation of the droplet

Figure 40. Three-dimensional oval droplet. Representation of the liquid-gas interface at

various times (from left to right, top to bottom t = 0.15, 0.30, 0.44, 0.74, 1.2
and 1.5 s.)

the top of the cylinder. Consider first the two-dimensional case. Three structured finite
element meshes are considered. The coarse mesh is made out of 2576 nodes and 5000
elements, the middle mesh is made out of 10151 nodes and 20000 elements and the fine
mesh is made out of 22726 nodes and 45000 elements. The size of the cells of the structured
mesh used for advection step is approximately 5 to 10 times smaller than the size of the
finite elements and the time step is chosen such that the CFL number is approximately one.
Numerical Simulation of Two-Phase Free Surface Flows 209

Smoothing parameter is ε = 0.005. Figure 41 illustrates the liquid-gas interface at different

times for different mesh sizes. The convergence when the mesh size and time step tend to
zero appears clearly thanks to surface tension effects.

t = 0.20

t = 0.45

coarse mesh middle mesh fine mesh

Figure 41. Rising Bubble: results for σ = 0.0738 Nm−1 and several mesh sizes. Position
of the free surface at times t = 0.2 s (first row) and t = 0.45 s (second row).
Left: coarse mesh, middle: middle mesh and right: fine mesh

Similar results may be obtained in three space dimensions and are illustrated in Figure 42
for a mesh made out of 115 200 tetrahedrons. The CPU time for this computation is
approximately 20 hours to achieve 1000 time steps. Most of the CPU time is used to solve
the Stokes problem in the liquid domain.

7 CONCLUDING REMARKS
A numerical model has been presented for the simulation of two-phases free surface flows. In
order to simulate mold filling problems, an incompressible liquid and a compressible gas have
been considered. This problem highlights many difficulties that appears in the framework
of free surface flows such as topology changes of the liquid domain, complex geometries,
210 Alexandre Caboussat

Figure 42. Rising Bubble: three-dimensional results for σ = 0.0738 Nm−1 . Representa-
tion of the gas domain at times t = 0.0, 0.25, 0.5, 0.75 and 1 s. (left to right,
top to bottom)

turbulent flows or the approximation of the interface and its numerical reconstruction. The
algorithm is based on an order one time splitting scheme and a two-grids method. The time
splitting scheme allows us to consider only less complicated subproblems one after the other
and well-adapted methods have been chosen to the resolution of each of these steps. The
two-grids method permits to use a structured grid and the characteristic method for the
advection and a finite element mesh for the resolution of a Stokes problem. Compared to
the many methods existing in the literature, our approach does not require a very accurate
approximation of the interface since all the boundary terms on the interface are computed
with variational techniques. Moreover, the method has been proved to be computationally
efficient and robust, the main advantage being that the full Navier-Stokes equations are
solved only in the liquid domain. The computation method for the gas pressure and the
surface tension effects and the hierarchical data structure are all ingredients to obtain finally
a very efficient method. The robustness of this method allows us to extend its capabilities in
different ways. The simulation of viscoelastic flows has already been investigated, see [15].
The next step would naturally to extend it to mold filling with solidification, by coupling
the flow equations with the heat equation.
Numerical Simulation of Two-Phase Free Surface Flows 211

ACKNOWLEDGMENTS

The author wishes to thank Prof. Jacques Rappaz and Dr Marco Picasso, Institute of
Analysis and Scientific Computing, Ecole Polytechnique Fédérale de Lausanne, Switzerland
for their numerous advices and ideas concerning this work. Dr Vincent Maronnier, Calcom
SA-ESI Group, Parc Scientifique, Ecole Polytechnique Fédérale de Lausanne, Switzerland
is greatly acknowledged for his main contribution in this project and his implementation
support. Prof. Jiwen He and Prof. Roland Glowinski, University of Houston, Texas, USA
are kindly acknowledged for pertinent remarks and new original ideas. The author thanks
also Dr Doug Kothe and the members of his team, Los Alamos National Laboratory, New
Mexico, USA for fruitful discussions and helpful comments. The company Calcom SA-ESI
Group, Lausanne, Switzerland, is acknowledged for kindly providing the CalcosoftT M Pre-
and Post-Processors. Finally the Swiss National Science Foundation is acknowledged for
its financial support.

A SOME THEORETICAL CONSIDERATIONS ON FREE BOUNDARY

PROBLEMS
Introduction
Theoretical questions about free surfaces often appear in the frame of fluid flow problems,
see e.g. [141], or fluid-structure interaction problems, see [64]. The aim of this section is
to summarize the results obtained in [19] concerning two free boundary problems. First a
one-dimensional free surface problem is presented, which consists in the motion of a fluid
of particles on a space interval with one free extremity. It consists of viscous Burgers’
equation. The extremity of the interval is unknown and moves in time. A zero force
boundary condition is thus enforced on the free extremity of the interval. This problem
is an example of non-cylindrical space-time domain problems [95]. Similar problems have
already been treated in the literature, for instance for an Arbitrary-Lagrangian-Eulerian
formulation in [44], for a two-fluids problem in [10] or for a given moving boundary in [12].
The model involves the velocity u of the liquid and the position s of the free right end
side of the space interval. The velocity is assumed to satisfy the viscous Burgers equation
on the free space interval, while the position of the free extremity is determined by an
ordinary differential equation depending on u.
Results about this one-dimensional free surface problem are summarized here and ex-
tracted from [22]. The well-posedness of the problem is proved and error estimates for the
semi-discretization in space are obtained, following [24, 72] for the a priori error analysis
and [160] for a posteriori error estimates. We refer to [19, 22] for details and demonstrations.
Then the method is extended to the two-dimensional case of a Navier-Stokes problem
with a free surface. Neumann boundary conditions are imposed on the whole boundary
of the domain. The free surface is assumed to be the whole boundary. Existence and
uniqueness of a solution are proved for small times, with the same methodology as for the
one-dimensional problem. In [141, 143], the existence of the same problem is proved by
using Lagrange coordinates and a fixed point theorem for the velocity field. In [142, 144],
existence results are obtained for the same problem, but in different function spaces. In
[64], the problem with Dirichlet boundary conditions is considered and in [9] the domain is
assumed to be an infinite horizontal layer domain and mixed Dirichlet-Neumann boundary
conditions are enforced.
The methodology for the proof of the existence of a solution to the free surface problems
is the following. In both cases, the fluid flow is considered first. For each time t, the domain
is mapped into a reference domain by a given mapping which is first assumed to be known.
A modified problem is thus obtained in the reference domain. Two successive fixed point
212 Alexandre Caboussat

theorems are used, together with a Faedo-Galerkin method, to obtain the existence of a
solution to the fluid flow problem with given moving boundary. Finally, the equation for
the mapping is considered and another fixed point theorem is used to obtain the existence
and uniqueness of a solution to the free surface problem for small times.
Analysis of a One-Dimensional Problem
The problem consisting in the Burgers equation with an additional diffusion term on a
space interval with one free extremity is investigated here. Let T > 0 be the final time of
the simulation. Let s : (0, T ) → R be an unknown positive function defined on the interval
(0, T ) and let the space-time domain QT be defined by:

QT = {(y, t) : y ∈ (0, s(t)), t ∈ (0, T )} .

Given the real positive numbers s0 , α and ε and the functions f¯ : R × (0, T ) → R and
ū0 : (0, s0 ) → R, our problem is to find s : (0, T ) → R and ū : QT → R satisfying:

∂ ū ∂ ū ∂ 2 ū
+ αū − ε 2 = f¯, (y, t) ∈ QT , (A.1)
∂t ∂y ∂y
ū(y, 0) = ū0 (y), y ∈ (0, s0 ), (A.2)
ū(0, t) = 0, t ∈ (0, T ), (A.3)
∂ ū
(s(t), t) = 0, t ∈ (0, T ), (A.4)
∂y
ṡ(t) = ū(s(t), t), t ∈ (0, T ), (A.5)
s(0) = s0 , (A.6)

where ṡ(t) = ds(t)/dt. Equations (A.1)-(A.4) are the Burgers equation with mixed Dirichlet-
Neumann boundary conditions. Equations (A.5)-(A.6) are the coupling equations, neces-
sary to determine the free surface s(t) and to ensure the well-posedness of the mathematical
problem.
The problem (A.1)-(A.6) in the non-cylindrical domain QT is turned into a problem in a
cylindrical domain, namely UT = (0, 1) × (0, T ) by using the change of variables x = y/s(t)
or y = s(t)x. The functions f , u0 and u are defined respectively by f (x, t) = f¯(s(t)x, t),
u0 (x) = ū0 (s0 x) and u(x, t) = ū(s(t)x, t), 0 < x < 1, 0 < t < T . The problem of interest is
then to find s : (0, T ) → R and u : UT → R satisfying:

∂u α ∂u ε ∂ 2 u ṡ(t) ∂u
+ u − − x = f, (x, t) ∈ UT , (A.7)
∂t s(t) ∂x s(t)2 ∂x2 s(t) ∂x
u(x, 0) = u0 (x), x ∈ (0, 1), (A.8)
u(0, t) = 0, t ∈ (0, T ), (A.9)
∂u
(1, t) = 0, t ∈ (0, T ), (A.10)
∂x
ṡ(t) = u(1, t), t ∈ (0, T ), (A.11)
s(0) = s0 . (A.12)

The function s is first supposed to be given in the set S(T ) of acceptable boundary
functions defined by:

S(T ) = s ∈ W 1,∞ (0, T ) : β1 ≤ s(t) ≤ β2 , ∀t ∈ [0, T ],
s(0) = s0 and |ṡ(t)| ≤ K, a.e. t ∈ [0, T ]} . (A.13)
Numerical Simulation of Two-Phase Free Surface Flows 213

The first result is a result local in time concerning the given boundary problem, i.e. Equa-
tions (A.7)-(A.10) with given s.

Theorem A.1. There exists a time 0 < T̃ ≤ T such that, for all s ∈ S(T ), the problem
(A.7)-(A.10) admits one unique solution denoted by us . Furthermore us ∈ L2 (0, T̃ , H 2 (0, 1))∩
H 1 (0, T̃ , L2 (0, 1)) with us (0, t) = 0 and there exists a constant K̃ = K̃(T̃ ), independent of
s such that:
! "
||us ||L2 (0,T̃ ,H 2 (0,1))∩H 1 (0,T̃ ,L2 (0,1)) ≤ 2K̃ ||f ||L2 (UT ) + |u0 |V . (A.14)

The following function may be defined:

 t
s̃(t) = s0 + us (1, τ )dτ, t ∈ (0, T̃ ). (A.15)
0
 
Theorem A.1 and the inclusion L2 (0, T, H 2 (0, 1))∩H 1 (0, T, L2 (0, 1)) → C 0 UT imply that
(A.15) is well-defined. The application γ is then defined by

γ : S(T ) → S(T );
s → s̃ defined by (A.15).

If there exists us , solution to the problem (A.7)-(A.10) associated with s, such that ṡ(t) =
us (1, t), the couple (us , s) is the solution to (A.7)-(A.12). The existence of the couple (us , s)
is deduced from the next theorem:
Theorem A.2. There exists 0 < T̂ ≤ T such that γ : S(T̂ ) → S(T̂ ) possesses a unique
fixed point s.
A semi-discretization in space is introduced using a finite element method and the
existence and uniqueness of the solution to the semi-discrete problem is proved. In a second
step, a priori error estimates are obtained, as in [72] for instance. Finally an a posteriori
error analysis is investigated, as in [160] for instance.
Let V be defined by V = {v ∈ H 1 (0, 1) : v(0) = 0}. A weak form of the continuous
problem (A.7)-(A.12) consists in finding (u, s) satisfying:
 1  1 
∂u α ∂u ṡ(t) 1 ∂u
vdx + u vdx − x vdx (A.16)
0 ∂t s(t) 0 ∂x s(t) 0 ∂x
 1  1
ε ∂u ∂v
+ dx = f vdx, ∀v ∈ V, a.e. t ∈ (0, T ),
s(t)2 0 ∂x ∂x 0
u(x, 0) = u0 (x), x ∈ (0, 1),
ṡ(t) = u(1, t), a.e. t ∈ (0, T ) with s(0) = s0 .

A piecewise linear finite element semi-discretization in space of (A.16) is introduced.

Let N ∈ N be a given integer and let x0 = 0 < x1 < x2 < . . . < xN +1 = 1 be the
discretization points in the interval [0, 1] with hi = |xi+1 − xi | and h = maxi hi . The space
of continuous functions on [0, 1] which are vanishing into x = 0 and whose restrictions on
[xi , xi+1 ], i = 0, . . . , N belong to P1 is denoted by Vh . Let u0,h be the interpolant of u0 ∈ V
in Vh , i.e. with |u0,h |V ≤ |u0 |V , and, if u0 ∈ H 2 (0, 1), interpolation results (see for instance
[33]) lead to ||u0 − u0,h ||L2 (0,1) ≤ Ch2 |u0 |H 2 (0,1) where C is and will be in the following
a generic constant independent of h. Let sh ∈ S(T ) be an approximation of s ∈ S(T ).
214 Alexandre Caboussat

The semi-discretization in space of (A.16) consists in finding uh : (0, T ) → uh (t) ∈ Vh and

sh ∈ S(T ) satisfying (A.16) with s replaced by sh , u by uh , f by fh , v by vh and V by Vh .
The existence of a semi-discrete approximation (uh , sh ) is ensured, see [22], and there
exists 0 < T̂ ≤ T (T̂ independent
!



of h > "0) and a constant C independent of h such
that ||uh ||H 1 (0,T̂ ,Vh ) ≤ C




f L2 (Q ) + |u0 |V . A priori error estimates are obtained in the
¯
T̂
following theorem, whose proof is based on [72]:
Theorem A.3 (A priori error estimate). Assume that u0 ∈ H 2 (0, 1) with u0 ∈ V
and f¯ ∈ L2 (0, T, H 1 (0, β2 )). There exists 0 < T̃ ≤ T , independent of h, such that both
continuous and semi-discrete problems admit a unique solution (u, s), respectively (uh , sh ),
∀h > 0 on the time interval (0, T̃ ), and the following error estimate holds:
||u − uh ||L∞ (0,T̃ ,L2 (0,1))∩L2 (0,T̃ ,V ) + ||s − sh ||H 1 (0,T̃ ) ≤ Ch, (A.17)
where the constant C does not depend on h.
The a posteriori error analysis is summarized in the next theorem.
Theorem A.4 (A posteriori error estimate). Assume that u0 ∈ H 2 (0, 1) ∩ V and f¯ ∈
L2 (0, T, H 1 (0, β2 )) and set hi = |xi+1 − xi |, h = maxi hi and Ji = (xi , xi+1 ), i = 0, . . . , N .
Then there exists 0 < T̃ ≤ T and a constant C, all independent of h > 0, such that, for all
t in (0, T̃ ):
 t




ε ε

∂(u − uh )

2
||u − uh (t)||L2 (0,1) + 2 |s(t) − sh (t)| + 2
2 2



dτ
2β2 8β2 0

∂x


N ! "
≤C h2i ||F (uh )||2L2 (0,t,L2 (Ji )) + h2i ||u0 ||2H 2 (Ji ) ,
i=0

where
∂uh α ∂uh ṡh (t) ∂uh
F (uh) = fh − − uh + x . (A.18)
∂t sh (t) ∂x sh (t) ∂x

Extension to the Two-Dimensional Problem

Let T > 0 be a finite horizon of time. Consider a bounded moving domain Ωt ⊂ R2 ,
t ∈ (0, T ) with boundary ∂Ωt . Note QT the space-time domain QT = {(x, t) ∈ R2 ×
(0, T ) : x ∈ Ωt , t ∈ (0, T )}; its space-time free boundary is denoted by ΣT = {(x, t) ∈
R2 × (0, T ) : x ∈ ∂Ωt , t ∈ (0, T )}. The whole boundary is assumed to be the free surface.
Let η : Ω0 × (0, T ) → R2 be the mapping (a priori unknown) that transforms Ω0 into Ωt
for all times t ∈ (0, T ).
The velocity and the pressure in the fluid are denoted by u and p. The problem reads:
find u : QT → R2 , p : QT → R and η : Ω0 × (0, T ) → R2 satisfying:
∂u
− 2µ∇ · D(u) + (u · ∇) u + ∇p = f , in QT , (A.19)
∂t
∇ · u = 0, in QT , (A.20)
−pn + 2µD(u)n = 0, on ΣT , (A.21)
u(0) = u0 , in Ω0 , (A.22)
∂η
= u ◦ η, in Ω0 × (0, T ) , (A.23)
∂t
η(0) = Id, in Ω0 . (A.24)
Numerical Simulation of Two-Phase Free Surface Flows 215

Here µ > 0 is a given positive constant, ∇·u denotes the divergence

 operator of u and D(u)
denotes the rate of deformation tensor D(u) = 12 ∇u + ∇uT . The spatial unit normal
vector is denoted by n and Id denotes the identity operator Id : ξ ∈ Ω0 → ξ ∈ Ω0 . The
bounded domain Ω0 is given and is assumed to be convex. The boundary ∂Ω0 is assumed
to be sufficiently regular (say C ∞ , even if this assumption could be relaxed). Finally, the
functions f and u0 are given, respectively on R2 × (0, T ) and Ω0 .
Here the mapping η allows  to determine the domain Ωt for all times t ∈ (0, T ). The
domain Ωt is given by Ωt = x ∈ R2 : x = η(ξ, t), ξ ∈ Ω0 .
We are interested in finding a solution to (A.19)-(A.24) for small times. The principle
of the proof is the following. Let us first assume that the domain Ωt is known for all
times t ∈ (0, T ) and disregard the equations (A.23)-(A.24) for the moment. The problem
(A.19)-(A.22) is transformed in order to consider a cylindrical space-time domain. Let η be
a sufficiently regular application (not necessarily the solution of (A.23)-(A.24)) that maps
Ω0 into Ωt for all t ∈ (0, T ):

ηt : Ω0 → Ωt ,
ξ → x = ηt(ξ) . (A.25)

Here the notation ηt (ξ) is equivalent to η(ξ, t). The mapping η0 at time t = 0 is assumed to
be the identity operator and is denoted by Id. Let us assume that the inverse transformation
of ηt exists and is also sufficiently regular. It is defined ∀t ∈ (0, T ) by:

ξt : Ωt → Ω0 ,
x → ξ = ξt (x) . (A.26)

The functions v : Ω0 × (0, T ) → R2 and q : Ω0 × (0, T ) → R are respectively defined by

v(ξ, t) = u(η(ξ, t), t) = u(x, t) and q(ξ, t) = p(η(ξ, t), t) = p(x) when x = η(ξ, t). Let the
matrix A = A(x, t) be the Jacobian matrix given by (A)i,j = ∂(ξt )i /∂xj , 1 ≤ i, j ≤ 2, so
that ∇x = AT ∇ξ and ∇x u = (∇ξ v)A, where ∇x is the gradient with respect to the variable
x (respectively ∇ξ is the gradient with respect to ξ). The inverse matrix A−1 = A−1 (ξ, t)
is defined by (A−1 )i,j = ∂(ηt)i /∂ξj , 1 ≤ i, j ≤ 2. Let ∇ξ be denoted in the following by ∇
for simplicity. The problem (A.19)-(A.22) is transformed into a problem in the cylindrical
space-time domain Ω0 × (0, T ): find v : Ω0 × (0, T ) → R2 and q : Ω0 × (0, T ) → R satisfying:

∂v   ∂ξ
− µ(AT ∇) · ∇vA + (∇vA)T + ∇v
∂t ∂t
+(v · A ∇)v + A ∇q = f ◦ η,
T T
in Ω0 × (0, T ) , (A.27)
(A ∇) · v = 0, T
in Ω0 × (0, T ) , (A.28)
 
−qA N + µ ∇vA + (∇vA)T AT N = 0,
T
on ∂Ω0 × (0, T ) , (A.29)
v(0) = u0 ◦ η0 = u0 , in Ω0 , (A.30)

where N is the external normal vector to Ω0 and η0 is assumed to satisfy η0 = Id. The
functions f and u0 are given respectively on R2 × (0, T ) and Ω0 and are sufficiently regular.
Let F be defined by F = f ◦ η.
∂ξ
The function η, as well as the coefficients of A and , are assumed to be known for the
∂t
moment. Regularity of η will be precised later. The existence of the problem with given
mapping η is obtained in three steps that are detailed hereafter.
The following Stokes problem is first obtained when A is replaced by the identity matrix
in (A.27)-(A.30) and by disregarding the nonlinear term (v · AT ∇)v and the advection term
216 Alexandre Caboussat

∂ξ
∇v: find v : Ω0 × (0, T ) → R2 and q : Ω0 × (0, T ) → R satisfying:
∂t
∂v  
− µ∇ · ∇v + (∇v)T + ∇q = F, in Ω0 × (0, T ) , (A.31)
∂t
∇ · v = 0, in Ω0 × (0, T ) , (A.32)
 
−qN + µ ∇v + (∇v)T N = 0, on ∂Ω0 × (0, T ) , (A.33)
v(0) = u0 , in Ω0 . (A.34)

This problem is a Stokes problem with Neumann boundary conditions. The existence of
the velocity and pressure satisfying (A.31)-(A.34) is given in [141] for instance. Then the
mapping η is given in a well-chosen set that is now described. Set 1 < r < 3/2. Let S1 be a
sufficiently large constant. The mapping η is assumed to belong to the set S(T ) defined by


S(T ) = η : Ω0 × (0, T ) → R2 : (A.35)
2+ r2
η ∈ H (0, T, H (Ω0 ) ) ∩ H
1 r+2 2 2 2
(0, T, L (Ω0 ) ),
||η||H 1 (H r+2 )∩H 2+ r2 (L2 ) ≤ S1 , η0 = Id, det(∇ηt) = 1, ∀t ∈ (0, T ),
∂η0
= u0 , ηt invertible and ξt = ηt−1 ∈ C 1 (Ωt ), ∀t ∈ (0, T ) .
∂t

The following modified Stokes problem is considered: find v : Ω0 × (0, T ) → R2 and

q : Ω0 × (0, T ) → R satisfying:

∂v   ∂ξ
− µ(AT ∇) · ∇vA + (∇vA)T + ∇v + AT ∇q = F, in Ω0 × (0, T ) , (A.36)
∂t ∂t
(AT ∇) · v = 0, in Ω0 × (0, T ) , (A.37)
 
−qAT N + µ ∇vA + (∇vA)T AT N = 0, on ∂Ω0 × (0, T ),(A.38)
v(0) = u0 , in Ω0 . (A.39)

It consists of the original problem (A.27)-(A.30) without the nonlinear term. Problem
(A.36)-(A.39) is investigated by using a fixed point theorem to take into account the terms
involving the matrix A of ∂ξ/∂t.
In a third step, the problem (A.27)-(A.30) is then investigated. A rescaling procedure
and a fixed point theorem are used to take the nonlinear term into account.
Finally the free surface problem is considered. Equations (A.23)-(A.24) are taken into
account with a fixed point theorem. The mapping η is proved to be the fixed point of the
application γ defined by:
γ1 : S(T̃ ) → S(T̃ )
,
η → η̃
#t
where η̃(ξ, t) = ξ + 0 v(ξ, s)ds and v is the solution to (A.27)-(A.30) with given η.
In summary, two problems involving free surfaces have been presented in this section.
The solvability of both of them has been proved with the Faedo-Galerkin method, together
with fixed point theorems. Observe that the solvability of the corresponding problems with
given moving boundary has also been obtained. All the proofs that have been disregarded
here can be found in [22] for details concerning the one-dimensional problem and in [19] for
details about the two-dimensional problem.
Numerical Simulation of Two-Phase Free Surface Flows 217

REFERENCES
1 R. Abgrall and S. Karni (2001). Computations of Compressible Multifluids. J. Comput. Phys.,
169, 594–623.
2 R. Abgrall, B. Nkonga and R. Saurel (2003). Efficient Numerical Approximation of Compress-
ible Multi-Material Flow for Unstructured Meshes. Computers and Fluids, 32, 571–605.
3 R. Abgrall and R. Saurel (2003). Discrete Equations for Physical and Numerical Compressible
Multiphase Mixtures. J. Comput. Phys., 186, 361–396.
4 E. Aulisa, S. Manservisi and R. Scardovelli (2003). A Mixed Markers and Volume-of-Fluid
Method for the Reconstruction and Advection of Interfaces in Two-Phase and Free-Boundary
Flows. J. Comput. Phys., 188, 611–639.
5 E. Aulisa, S. Manservisi and R. Scardovelli (2004). A Surface Marker Algorithm Coupled to
an Area-preserving Marker Redistribution Method for Three-dimensional Interface Tracking.
J. Comput. Phys., 197 (2), 555–584.
6 O. Axelsson and V.A. Barker (1984). Finite Element Solution of Boundary Value Problems.
Computer Science and Applied Mathematics. Academic Press.
7 E. Bänsch (2001). Finite Element Discretization of the Navier-Stokes Equations with a Free
Capillary Surface. Numer. Math., 88 (2), 203–235.
8 T. J. Barth and J. A. Sethian (1998). Numerical Schemes for the Hamilton-Jacobi and Level
Set Equations on Triangulated Domains. J. Comput. Phys., 145 (1), 1–40.
9 J. T. Beale (1981). The Initial Value Problem for the Navier-Stokes Equations with a Free
Surface. Comm. Pure Appl. Math., 34, 359–392.
10 M. Bertsch, D. Hilhorst and Cl. Schmidt-Lainé (1994). The Well-Posedness of a Free Boundary
Problem for Burgers’ Equation. Nonlinear Anal., Theory Methods Appl., 23 (9), 1211–1224.
11 R. Bidoae, R. M. Ciobotaru and P. E. Raad (2003). An Eulerian-Lagrangian Marker and Micro
Cell Simulation Method for Fluid Interaction with Solid/Porous Bodies. In IUTAM - Sym-
posium on Integrated Modeling of Fully Coupled Fluid-Structure Interactions Using Analysis,
Computations and Experiments, June 1-6.
12 G. Biondini and S. De Lillo (2001). On the Burgers Equation with Moving Boundary. Phys.
Letters A, 279, 194–206.
13 D. Boffi and L. Gastaldi (2003). A Finite Element Approach for the Immersed Boundary
Method. Computer and Structures, 81, 491–501.
14 D. Boffi and L. Gastaldi (2004). Stability and Geometric Conservation Laws for ALE Formu-
lations. Comp. Meth. Appl. Mech. Eng., 193, 4717–4739.
15 A. Bonito, M. Picasso and M. Laso (2003). Numerical Simulation of 3D Viscoelastic Flows
with Free Surfaces. In J. Non-Newtonian Fluid Mechanics. XIIIth Workshop on Numerical
Methods for non-Newtonian Flows, submitted.
16 J. U. Brackbill, D. B. Kothe and C. Zemach (1992). A Continuum Method for Modeling Surface
Tension. J. Comput. Phys., 100, 335–354.
17 B. Bunner and G.Tryggvason (2002). Dynamics of Homogeneous Bubbly Flows Part 1. Rise
Velocity and Microstructure of the Bubbles. J. Fluid Mech., 466, 17–52.
18 B. Bunner and G.Tryggvason (2002). Dynamics of Homogeneous Bubbly Flows Part 2. Velocity
Fluctuations. J. Fluid Mech., 466, 53–84.
19 A. Caboussat (2003). Analysis and Numerical Simulation of Free Surface Flows. PhD
thesis, École Polytechnique Fédérale de Lausanne, Institute of Analysis and Scientific
Computing, 1015 Lausanne, Switzerland, December. Thesis number 2893, available on
http://library.epfl.ch/theses.
20 A. Caboussat (2004). A numerical method for the simulation of free surface flows with surface
tension. Computers and Fluids, submitted.
218 Alexandre Caboussat

21 A. Caboussat, M. Picasso and J. Rappaz (2005). Numerical Simulation of Free Surface Incom-
pressible Liquid Flows Surrounded by Compressible Gas. J. Comput. Phys., 203 (2), 626–649.
22 A. Caboussat and J. Rappaz (2005). Analysis of a One-Dimensional Free Surface Flow Problem.
Numer. Math., to appear.
23 R. Caiden, R. P. Fedkiw and C. Anderson (2001). A Numerical Method for Two-Phase Flow
Consisting of Separate Compressible and Incompressible Regions. J. Comput. Phys., 166, 1–27.
24 G. Caloz and J. Rappaz (1997). Numerical Analysis for Nonlinear and Bifurcation Problems,
volume 5 of Handbook of Numerical Analysis (P.G. Ciarlet, J.L. Lions eds), pages 487–637.
Elsevier, Amsterdam.
25 H.D. Ceniceros and A.M. Roma (2004). A Multi-Phase Flow Method with a Fast, Geometry-
Based Fluid Indicator. J. Comput. Phys., 205, 391–400.
26 G. Cerne, S. Petelin and Iztok Tiselj (2001). Coupling of the Interface Tracking and the Two-
Fluid Models for the Simulation of Incompressible Two-Phase Flow. J. Comput. Phys., 171,
776–804.
27 Y. C. Chang, T. Y. Hou, B. Merriman and S. Osher (1996). A Level Set Formulation of
Eulerian Interface Capturing Methods for Incompressible Fluid Flows. J. Comput. Phys., 124
(2), 449–464.
28 S. Chen, D.B. Johnson, P.E. Raad and D. Fadda (1997). The Surface Marker and Micro Cell
Method. Int. J. Numer. Meth. Fluids, 25, 749–778.
29 J. Chessa and T. Belytschko (2003). An Enriched Finite Element Method and Level Sets for
Axisymmetric Two-Phase Flow with Surface Tension. Int. J. Numer. Meth. Engrg, 58 (13),
2041–2064.
30 D. L. Chopp and J. A. Sethian (1999). Motion by Intrinsic Laplacian of Curvature. Interfaces
and Free Boundaries, 1, 107–123.
31 A. J. Chorin (1980). Flame Advection and Propagation Algorithms. J. Comput. Phys., 35,
1–11.
32 A. J. Chorin (1985). Curvature and Solidification. J. Comput. Phys., 58, 472–490.
33 P. G. Ciarlet and J.-L. Lions, Eds. (1991). Handbook of Numerical Analysis - Finite Element
Methods (Part I). North-Holland.
34 R. Codina and O. Soto (2002). A Numerical Model to Track Two-Fluid Interfaces Based on
a Stabilized Finite Element Method and a Level Set Technique. Int. J. Numer. Meth. Fluids,
40, 293–301.
35 R. Codina, U. Schäfer and E. Oñate (1994). Mould Filling Simulation using Finite Elements.
Int. J. Numer. Meth. Heat Fluid Flow, 4, 291–310.
36 G.-H. Cottet and E. Maitre (2004). A Level-Set Formulation of Immersed Boundary Methods
for Fluid-Structure Interaction Problems. C. R. Acad. Sci. Paris, Sér. I, 338, 581–586.
37 R. Croce, M. Griebel and M. A. Schweitzer (2004). A Parallel Level-Set Approach for Two-
Phase Flow Problems with Surface Tension in Three Space Dimensions. Technical Report 157,
Sonderforschungsbereich 611, Universität Bonn.
38 S. J. Cummins, M. M. Francois and D. B. Kothe (2005). Reconstructing Distance Functions
from Volume Fractions: A Better Way to Estimate Interface Topology? Computers and
Structures, 83, 425–434.
39 D. B. Kothe, W.J. Rider, S.J. Mosso, J.S. Brock and J.I. Hochstein (1996). Volume Tracking
of Interfaces Having Surface Tension in Two and Three Dimensions. Technical Report AIAA
96-0859, Los Alamos National Laboratory.
40 E. D. Dendy, N. T. Padial-Collins and W. B. VanderHeyden (2004). A General-Purpose,
Finite-Volume, Advection Scheme for Continuous and Discontinuous Fields on Unstructured
Grids. Technical Report LA-UR-01-1485, Los Alamos National Laboratory. also submitted to
J. Comput. Phys.
41 B. Desjardins, M. J. Esteban, C. Grandmont and P. Le Tallec (2001). Weak Solutions for a
Fluid-Elastic Structure Interaction Model. Rev. Mat. Comput., 14 (2), 523–538.
Numerical Simulation of Two-Phase Free Surface Flows 219

42 G. Dhatt, D. M. Gao and A. Ben Cheikh (1990). A Finite Element Simulation of Metal Flow
in Moulds. Int. J. Numer. Meth. Eng., 30, 821–831.
43 S. Dufour and D. Pelletier (1999). Computations of Multiphase Flows with Surface Tension
using Adaptive Finite Element Methods. In Proceedings of the 37th AIAA Aerospace Sciences
Meeting and Exhibit, Reno, number 99-0544 in AIAA Paper.
44 D. Errate, M. J. Esteban and Y. Maday (1994). Couplage Fluide-Structure. Un Modèle Simplifié
en Dimension 1. C.R. Acad. Sci. Paris, Sér 1, Math, 318, 275–281.
45 C. Farhat, Ph. Geuzaine and C. Grandmont (2001). Discrete Geometric Conservation Law and
the Nonlinear Stability of ALE Schemes for the Solution of Flow Problems on Moving Grids.
J. Comput. Phys., 174, 669–694.
46 R. Fedkiw (2002). Coupling an Eulerian Fluid Calculation to a Lagrangian Solid Calculation
with the Ghost Fluid Method. J. Comput. Phys., 175, 200–224.
47 R. P. Fedkiw, B. Merriman and S. Osher (1998). Numerical Methods for a One-Dimensional
Interface Separating Compressible and Incompressible Flows. In V. Venkatakrishnan, M. Salas,
and S. Chakravarthy, editors, Barriers and Challenges in Computational Fluid Dynamics, pages
155–194. Kluwer Academic Publishers.
48 J.M. Floryan and H. Rasmussen (1989). Numerical Methods for Viscous Flows with Moving
Boundaries. Appl. Mech. Rev., 42, 323–341.
49 L. Formaggia and F. Nobile (1999). A Stability Analysis for the Arbitrary Lagrangian Eulerian
Formulation with Finite Elements. East-West Journal of Numerical Mathematics, 7, 105–132.
50 L. P. Franca and S. L. Frey (1992). Stabilized Finite Element Method: II. The Incompressible
Navier-Stokes Equations. Comp. Meth. Appl. Mech. Engrg, 99, 209–233.
51 M. Francois and W. Shy (2003). Computations of Drop Dynamics with the Immersed Boundary
Method, Part 1: Numerical Algorithm and Buoyancy Effect. Numerical Heat Transfer Part B,
44, 101–118.
52 M. Francois and W. Shy (2003). Computations of Drop Dynamics with the Immersed Boundary
Method, Part 2: Drop Impact and Heat Transfer. Numerical Heat Transfer Part B, 44, 119–
143.
53 M. M. Francois, D. B. Kothe, E. D. Dendy, J. M. Sicilian and M. W. Williams (2003). Bal-
anced Force Implementation of the Continuum Surface Tension Force into a Pressure Correc-
tion Algorithm. In Proceedings of the FEDSM’03, 4th ASME JSME Joint Fluids Engineering
Conference, Honolulu, Hawai, USA.
54 J.-F. Gerbeau, T. Lelievre and C. Le Bris (2003). Simulations of MHD Flows with Moving
Interfaces. J. Comput. Phys., 184 (1), 163–191.
55 J.-F. Gerbeau and M. Vidrascu (2003). A Quasi-Newton Algorithm Based on a Reduced Model
for Fluid-Structure Interaction Problems in Blood Flows. Technical Report 4691, INRIA.
56 F. Gibou, R. Fedkiw, R Caflisch and S. Osher (2003). A Level Set Approach for the Numerical
Simulation of Dendritic Growth. J. Sci. Comput., 19 , 183–199.
57 J. Glimm, J. W. Grove, X. L. Li, K.-M. Shyue, Y. Zeng and Q. Zhang (1998). 3-Dimensional
Front Tracking. SIAM J. Sci. Computing, 19, 703–727.
58 J. Glimm, X. Li, Y. Liu, Z. Xu and N. Zhao (2003). Conservative Front Tracking with Improved
Accuracy. SIAM J. Numer. Anal., 41 (5), 1926–1947.
59 R. Glowinski, E.J. Dean, L.H. Juarez and T.W. Pan (2004). Applications of Operator-Splitting
Methods to the Direct Numerical Simulation of Particulate and Free-Surface Flows and to the
Numerical Solution of the Two-Dimensional Elliptic Monge-Ampère Equation. Von Karman
Prize Paper, SIAM Review, submitted.
60 R. Glowinski and L.H. Juarez (2003). Finite Element Method and Operator-Splitting for a
Time-Dependent Viscous Incompressible Free-Surface Flow. Comp. Fluid Dynam. J., 12 (3),
459–468.
61 R. Glowinski, T. W. Pan, T. I. Hesla, D. D. Joseph and J. Périaux (2001). A Fictitious
Domain Approach to the Direct Numerical Simulation of Incompressible Viscous Flow Past
Moving Rigid Bodies: Application to Particulate Flow. J. Comput. Phys., 169 (2), 363–426.
220 Alexandre Caboussat

62 R. Glowinski, T. W. Pan and J. Périaux (1994). A Fictitious Domain Method for Dirichlet
Problem and Applications. Comp. Meth. Appl. Mech. Engrg, 111 (3-4), 283–303.
63 J. Gomes and O. Faugeras (2000). Reconciling Distance Functions and Level Sets. Journal of
Vizualisation Communication and Image Representation, 11, 209–223.
64 C. Grandmont and Y. Maday (2000). Existence for an Unsteady Fluid-Structure Interaction
Problem. M2 AN, 34 (3), 609–636.
65 G.Tryggvason, B. Bunner, A. Esmaeeli, D. Juric, N. Al-Rawahi, W. Tauber, J. Han S. Nas
and Y.-J. Jan (2001). A Front Tracking Method for the Computations of Multiphase Flow. J.
Comput. Phys., 169, 708–759.
66 D. Gueyffier, J. Li, A. Nadim, R. Scardovelli and S. Zaleski (1999). Volume-of-Fluid Interface
Tracking with Smoothed Surface Stress Methods for Three-Dimensional Flows. J. Comput.
Phys., 152, 423–456.
67 W. Hackbusch (1985). Multi-Grid Methods and Applications. Springer-Verlag, Berlin.
68 P. Hansbo (1992). The Characteristic Streamline Diffusion Method for the Time-Dependent
Incompressible Navier-Stokes Equations. Comp. Meth. Appl. Mech. Engrg, 99, 171–186.
69 P. Hansbo (2000). A Free-Lagrange Finite Element Method using Space-Time Elements. Comp.
Meth. Appl. Mech. Engrg, 188, 347–361.
70 F. Harlow (1964). The Particle-in-Cell Computing Method for Fluid Dynamics. Methods of
Computational Physics, 3, 313–343.
71 F. H. Harlow and E. Welch (1965). Numerical Calculation of Time-Dependent Viscous Incom-
pressible Flow of Fluids with Free Surface. Phys. Fluids, 8, 2182.
72 J. G. Heywood and R. Rannacher (1982). Finite Element Approximation of the Nonstationary
Navier-Stokes Problem. Part I. Regularity of Solutions and Second-Order Error Estimates for
Spatial Discretization. SIAM J. Numer. Anal., 19 (2), 275–311.
73 C.W. Hirt, A.A. Amsden and J.L. Cook (1974). An Arbitrary Lagrangian-Eulerian Computing
Method for All Speeds. J. Comput. Phys., 14, 227–253.
74 C. W. Hirt and B. D. Nichols (1981). Volume of Fluid (VOF) Method for the Dynamics of
Free Boundaries. J. Comput. Phys., 39, 201–225.
75 T. Y. Hou (1995). Numerical Solutions to Free Boundary Problems. Acta Numerica, 4, 335–
415.
76 A. Huerta and W. K. Liu (1988). Viscous Flow with Large Free Surface Motion. Comp. Meth.
Appl. Mech. Engrg, 69, 277–324.
77 T.J.R. Hugues, W.K. Liu and T.K. Zimmermann (1981). Lagrangian Eulerian Finite Element
Formulation for Incompressible Viscous Flows. Comp. Meth. Appl. Mech. Engrg, 29, 329–349.
78 S. R. Idelsohn, M. A. Storti and E. O nate (2001). Lagrangian Formulations to Solve Free
Surface Incompressible Inviscid Fluid Flows. Comput. Meth. Appl. Mech. Engrg, 191, 583–593.
79 G.-S. Jiang and D. Peng (2000). Weighted ENO Schemes for Hamilton-Jacobi Equations. SIAM
J. Sci. Comp., 21, 2126.
80 Ch. Josserand and S. Zaleski (2003). Droplet Splashing on a Thin Liquid Film. Phys. Fluids,
15 (6), 1650–1657.
81 M. Kang, R. P. Fedkiw and X.-D. Liu (2000). A Boundary Condition Capturing Method for
Multiphase Incompressible Flow. Journal of Scientific Computing, 15, 323–360.
82 M. S. Kim and W. I. Lee (2003). A New VOF-Based Numerical Scheme for the Simulation of
Fluid Flow with Free Surface. Part I: New Free Surface-Tracking Algorithm and its Verification.
Int. J. Num. Meth. Fluids, 42, 765–790.
83 M. S. Kim, J. S. Park and W. I. Lee (2003). A New VOF-Based Numerical Scheme for the
Simulation of Fluid Flow with Free Surface. Part II: Application to the Cavity Filling and
Sloshing Problems. Int. J. Num. Meth. Fluids, 42, 791–812.
84 P. Klouček and M. V. Romerio (2002). The Detachment of Bubbles under a Porous Rigid
Surface during Aluminium Electrolysis. Math. Mod. Meth. App. Sci., 12 (11), 1617–1652.
Numerical Simulation of Two-Phase Free Surface Flows 221

85 B. Koren, M.R. Lewis, E.H. van Brummelen and B. van Leer (2001). Riemann-Problem and
Level-Set Approaches for Two-Fluid Flow Computations I. Linearized Godunov Scheme. Tech-
nical Report MAS-R0112, CWI, Amsterdam.
86 D. Kothe, D. Juric, K. Lam and B. Lally (1998). Numerical Recipes for Mold Filling Simulation.
In Proceedings of the Eighth International Conference on Modeling of Casting, Welding, and
Advanced Solidification Processes, San Diego, CA.
87 D.B. Kothe (1998). Perspective on Eulerian Finite Volume Methods for Incompressible Inter-
facial Flows, pages 267–331. Free Surface Flows (Udine, 1997). Springer, Vienna.
88 D.B. Kothe, M.W. Williams, K.L. Lam, D.R. Korzewka, P.K. Tubesing and E.G. Puckett
(1999). A Second-Order Accurate, Linearity-Preserving Volume Tracking Algorithm for Free
Surface Flows on 3-D Unstructured Meshes. In Proceedings of the 3rd ASME/JSME Joint
Fluids Engineering Conference, San Francisco, CA.
89 H. Lamb (1900). Hydrodynamics. Dover, New York.
90 L. Lee and R. J. LeVeque (2003). An Immersed Interface Method for Incompressible Navier-
Stokes Equations. SIAM J. Sci. Comput., 25 (3), 832–856.
91 R.J. LeVeque and K.-M. Shyue (1996). Two Dimensional Front Tracking Based on High
Resolution Wave Propagation Methods. J. Comput. Phys., 123, 354–368.
92 R. W. Lewis, S. E. Navti and C. Taylor (1997). A Mixed Lagrangian-Eulerian Approach to
Modelling Fluid Flow During Mould Filling. Int. J. Numer. Meth. Fluids, 25, 931–952.
93 R.W. Lewis, A.S. Usmani and J.T. Cross (1995). Efficient Mould Filling Simulation in Castings
by an Explicit Finite Element Method. Int. J. Numer. Meth. Fluids, 20, 493–506.
94 J. Li and Y. Renardy (2000). Numerical Study of Flows of Two Immiscible Liquids at Low
Reynolds Number. SIAM Review, 42 (3), 417–439.
95 J.-L. Lions (1957). Sur les Problèmes Mixtes pour Certains Systèmes Paraboliques dans des
Ouverts Non Cylindriques. Annales de l’Institut Fourier, VII, 143–182.
96 G. I. Marchuk (1990). Splitting and Alternating Direction Methods, volume 1 of Handbook of
Numerical Analysis (P.G. Ciarlet, J.L. Lions eds), pages 197–462. Elsevier.
97 G.I. Marchuk (1975). Methods of Numerical Mathematics. Springer, Applications of Mathe-
matics.
98 V. Maronnier, M. Picasso and J. Rappaz (1999). Numerical Simulation of Free Surface Flows.
J. Comput. Phys., 155, 439–455.
99 V. Maronnier, M. Picasso and J. Rappaz (2003). Numerical Simulation of Three Dimensional
Free Surface Flows. Int. J. Num. Meth. Fluids, 42 (7), 697–716.
100 J. C. Martin and W. J. Moyce (1952). An Experimental Study of the Collapse of Liquid
Collumns on a Rigid Horizontal Plate. Philos. Trans. Roy. Soc. London Sér., A244, 312–324.
101 F. Mashayek and N. Ashgriz (1995). A Hybrid Finite-Element-Volume-Of-Fluid Method for
Simulating Free Surface Flows and Interfaces. Int. J. Num. Meth. Fluids, 20, 1367–1380.
102 A. Masud and T.J.R. Hughes (1997). A Space-Time Galerkin/Least-Squares Finite Element
Formulation of the Navier-Stokes Equations for Moving Domains Problems. Comp. Meth. Appl.
Mech. Engrg, 146, 91–126.
103 B. Maury (1996). Characteristics ALE Method for the 3D Navier-Stokes Equations with a Free
Surface. Int. J. of Comp. Fluid Dynamics, 6, 175–188.
104 B. Maury (1999). Direct Simulations of 2D Fluid-Particle Flows in Biperiodic Domains. J.
Comput. Phys., 156, 325–351.
105 S. McKee, M. F. Tome, J. A. Cuminato, A. Castelo and V. G. Ferreira (2004). Recent Advances
in the Marker and Cell Method. Arch. Comput. Methods Engrg, 11 (2), 107–142.
106 M. Medale and M. Jaeger (1997). Numerical Simulations of Incompressible Flows with Moving
Interfaces. Int. J. Num. Meth. Fluids, 24, 615–638.
107 M. Meier, G. Yadigaroglu and B. L. Smith (2002). A Novel Technique for Including Surface
Tension in PLIC-VOF Methods. European Journal of Mechanics B - Fluids, 21, 61–73.
222 Alexandre Caboussat

108 W. Mulder, S. Osher and J.A. Sethian (1992). Computing Interface Motion in Compressible
Gas Dynamics. J. Comput. Phys., 100, 209–228.
109 W.F. Noh and P. Woodward (1976). SLIC (Simple Line Interface Calculation), volume 59 of
Lectures Notes in Physics, pages 330–340. Springer-Verlag.
110 S. Osher, L.-T. Cheng, M. Kang, H. Shim and Y.-H. Tsai (2001). Geometric Optics in a Phase
Space Based Level Set and Eulerian Framework. J. Comput. Phys., 179, 622–648.
111 S. Osher and R. P. Fedkiw (2001). Level Set Methods: An Overview and Some Recent Results.
J. Comput. Phys., 169, 463–502.
112 S. Osher and R. P. Fedkiw (2003). Level Set Methods and Dynamic Implicit Surfaces. Applied
Mathematical Sciences. Springer-Verlag.
113 S. J. Osher and J. Sethian (1988). Fronts Propagating with Curvature Dependent Speed:
Algorithms Based on Hamilton-Jacobi Formulations. J. Comput. Phys, 79, 12–49.
114 N. Parolini and A. Quarteroni (2004). Mathematical Models and Numerical Simulations for
the America’s Cup. Comp. Meth. Appl. Mech. Eng., 194, 1001–1026.
115 D. Peng, B. Merriman, S. Osher, H. Zhao and M. Kang (1999). A PDE-Based Fast Local Level
Set Method. J. Comput. Phys., 155, 410–438.
116 C. Peskin (1977). Numerical Analysis of Blood Flow in the Heart. J. Comput. Phys., 25,
220–252.
117 M. Picasso, J. Rappaz, A. Reist, M. Funk and H. Blatter (2004). Numerical Simulation of the
Motion of a Two Dimensional Glacier. Int. J. Num. Meth. Engrg, 60, 995–1009.
118 O. Pironneau (1989). Finite Element Methods for Fluids. Wiley, Chichester.
119 O. Pironneau, J. Liou and T. Tezduyar (1992). Characteristic-Galerkin and Galerkin/Least-
Squares Space-Time Formulations for the Advection-Diffusion Equation with Time-Dependent
Domain. Comput. Methods Appl. Mech. Eng., 100, 117–141.
120 S. Popinet and S. Zaleski (1999). A Front-Tracking Algorithm for Accurate Representation of
Surface Tension. Int. J. Numer. Meth. Fluids, 30, 777–793.
121 A. Quarteroni and L. Formaggia (2002). Mathematical Modelling and Numerical Simulation of
the Cardiovascular System, volume to appear of Handbook of Numerical Analysis (P.G. Ciarlet,
J.L. Lions eds), chapter Modelling of Living Systems. Elsevier.
122 A. Quarteroni and A. Valli (1994). Numerical Approximation of Partial Differential Equations.
Springer-Verlag Series in Computational Mathematics, no 23, second edition.
123 M. Rappaz, J.L. Desbiolles, C.A. Gandin, S. Henry, A. Semoroz and P. Thevoz (2000). Mod-
elling of Solidification Microstructures. Material Science Forum, 329 (3), 389–396.
124 M. Renardy, Y. Renardi and J. Li (2001). Numerical Simulation of Moving Contact Line
Problems Using a Volume-of-Fluid Method. J. Comput. Phys., 171, 243–263.
125 Y. Renardy (2003). Direct Simulation of Drop Fragmentation under Simple Shear. In Interfacial
Fluid Dynamics and Transport Processes, Lecture Notes in Physics, Springer Verlag Berlin
Heidelberg, pages 305–325.
126 Y. Renardy, S. Popinet, L. Duchemin, M. Renardy, S. Zaleski, C. Josserand, M. A. Drumright-
Clarke, D. Richard, C. Clanet and D. Quéré (2003). Pyramidal and Toroidal Water Drops
after Impact on a Solid Surface. J. Fluid. Mech, 484, 69–83.
127 Y. Renardy and M. Renardy (2002). PROST : A Parabolic Reconstruction of Surface Tension
for the Volume-Of-Fluid Method. J. Comput. Phys., 183, 400–421.
128 Y. Renardy, M. Renardy, T. Chinyoka, D.B. Khismatullin and J. Li (2004). A Viscoelastic
VOF-PROST Code for the Study of Drop Deformation. In Proceedings of the ASME Heat
Transfer/Fluids Engineering Summer Conference, Charlotte, North Carolina.
129 W.J. Rider, D.B. Kothe, E.G. Puckett and I.D. Aleinov (1998). Accurate and Robust Methods
for Variable Density Incompressible Flows with Discontinuities. In V. Venkatakrishnan, M.
D. Salas, and S. R. Chakravarthy, editors, Barriers and Challenges in Computational Fluid
Dynamics, pages 213–230.
Numerical Simulation of Two-Phase Free Surface Flows 223

130 W.J. Rider and D.B. Kothe (1998). Reconstructing Volume Tracking. J. Comput. Phys., 141,
112–152.
131 M.V. Romerio, A. Lozinski and J. Rappaz (2005). A New Modelling for Simulating Bubble
Motion in a Smelter. Light Metals, pages 547–552.
132 R. Scardovelli and S. Zaleski (1999). Direct Numerical Simulation of Free Surface and Interfacial
Flows. Annual Review of Fluid Mechanics, 31, 567–603.
133 M. Schmid and F. Klein (1996). Einflüss der Wandreibung auf das Füllverhalten Dünner
Platten. Preprint, Steinbeis Transferzentrum, Fachhochschule Aachen.
134 J.A. Sethian (1996). Level Set Methods, Evolving Interfaces in Geometry, Fluid Mechanics,
Computer Vision, and Material Science. Monographs on Applied and Computational Mathe-
matics. Cambridge University Press.
135 J.A. Sethian (1999). Fast Marching Methods. SIAM Review, 41 (2), 199–235.
136 M. J. Shelley, F.-R. Tian and K. Wlodarski (1997). Hele-Shaw Flow and Pattern Formation
in a Time-Dependent Gap. Nonlinearity, 10, 1471–1495.
137 S. Shin and D. Juric (2002). Modeling Three-Dimensional Multiphase Flow Using a Level
Contour Reconstruction Method for Front Tracking Without Connectivity. J. Comput. Phys.,
180, 427–470.
138 K.-M. Shyue (1999). A Fluid-Mixture Type Algorithm for Compressible Multicomponent Flow
with van der Waals Equation of State. J. Comput. Phys., 156, 43–88.
139 K.-M. Shyue (1999). A Volume-Of-Fluid type Algorithm for Compressible Two-Phase Flows.
Intl. Series of Numerical Mathematics, 130, 895–904.
140 R. W. Smith (1999). AUSM(ALE): A Geometric Conservative Arbitrary Lagrangian-Eulerian
Flux Splitting Scheme. J. Comput. Phys., 150, 268–286.
141 V. A. Solonnikov (1986). Solvability of the Problem of Evolution of an Isolated Volume of
Viscous, Incompressible Capillary Fluid. J. Soviet Math., 32 (2), 179–186.
142 V. A. Solonnikov (1988). On the Transient Motion of an Isolated Volume of Viscous Incom-
pressible Fluid. Math. USSR Izvestiya, 31(2), 381–405.
143 V. A. Solonnikov (1992). Solvability of the Problem of Evolution of a Viscous Incompressible
Fluid Bounded by a Free Surface on a Finite Time Interval. St Petersburg Math. J., 3 (1),
189–220.
144 V.A. Solonnikov (1978). The Solvability of the Second Initial Boundary-Value Problem for the
Linear, Time-dependent System of Navier-Stokes Equations. J. Soviet Math., 10 (1), 141–155.
145 O. Soto and R. Codina (2003). A Numerical Model for Mould Filling using a Stabilized Finite
Element Method and the VOF Technique. Int. J. Num. Meth. Fluids, submitted.
146 Ch. G. Speziale (1991). Analytical Methods for the Development of Reynolds-Stress Closures
in Turbulence. Annual Review of Fluid Mechanics, 23, 107–157.
147 M. Sussman, A. S. Almgren, J. B. Bell, Ph. Colella, L. H. Howell and M. L. Welcome (1999).
An Adaptive Level Set Approach for Incompressible Two-Phase Flows. J. Comput. Phys., 148,
81–124.
148 M. Sussman, E. Fatemi, P. Smereka and S. Osher (1998). An Improved Level Set Method for
Incompressible Two-Phase Flows. Computers and Fluids, 27 (5-6), 663–680.
149 M. Sussman and E. G. Puckett (2000). A Coupled Level Set and Volume-of-Fluid Method for
Computing 3D and Axisymmetric Incompressible Two-Phase Flows. J. Comput. Phys., 162,
301–337.
150 M. Sussman, P. Smereka and S. Osher (1994). A Level Set Approach for Computing Solutions
to Incompressible Two-Phase Flow. J. Comput. Phys., 114, 146–159.
151 T.E. Tezduyar, M. Behr, S. Mittal and J. Liou (1992). A New Strategy for Finite Element
Computations Involving Boundaries and Interfaces - The Deforming-Spatial-Domain/Space-
Time Procedure: II. Computations of Free Surface Flows, Two-Liquid Flows, and Flows with
Drifting Cylinders. Comp. Meth. Appl. Mech. Engrg, 94, 353–371.
224 Alexandre Caboussat

152 E. Thompson (1986). Use of Pseudo-Concentrations to Follow Creeping Viscous Flows During
Transient Analysis. Int. J. Numer. Meth. Fluids, 6, 749–761.
153 A.-K. Tornberg and B. Engquist (2000). Interface Tracking in Multiphase Flows. Multifield
Problems, State of the Art, pages 58–66.
154 A.-K. Tornberg and B. Engquist (2003). The Segment Projection Method for Interface Track-
ing. Commun. Pur. Appl. Math., 56, 47–79.
155 D.J. Torres and J.U.Brackbill (2000). The Point-Set Method: Front-Tracking without Con-
nectivity. J. Comput. Phys., 165, 620–644.
156 S.O. Unverdi and G. Tryggvason (1992). Computations of Multi-Fluid Flows. Physica D, 60,
70–83.
157 S.P. van der Pijl, A. Segal and C. Vuik (2003). A Mass-Conserving Level-Set (MCLS) Method
for Modeling of Multi-Phase Flows. Technical Report 03-03, Delft University of Technology.
158 S.P. van der Pijl, A. Segal and C. Vuik (2004). Modelling of Three-Dimensional Bubbly Flows
with a Mass-Conserving Level-Set Method. In Proceedings of the 4th European Congress on
Computational Methods in Applied Sciences and Engineering, ECCOMAS 2004, Jyväskylä,
Finland, CDROM, volume II.
159 S. Čanić, D. Lamponi, A. Mikelić and J. Tampača (2005). Self-Consistent Effective Equations
Modeling Blood Flow in Medium-to-Large Compliant Arteries. SIAM J. Multiscale Analysis
and Simulation, 3 (3), 559–606.
160 R. Verfürth (1996). A Review of A Posteriori Error Estimation and Adaptive Mesh-Refinement
Techniques. Wiley-Teubner.
161 H. Wang, H.K. Dahle, R.E. Ewing, M.S. Espedal, R.C. Sharpley and S. Man (1999). An ELLAM
Scheme for Advection-Diffusion Equations in Two Dimensions. SIAM J. Sci. Comput., 20 (6),
2160–2194.
162 M. W. Williams, D. B. Kothe and E. G. Puckett (1998). Accuracy and Convergence of Contin-
uum Surface Tension Models. In Cambridge Cambridge Univ. Press, editor, Fluid Dynamics
at Interfaces, pages 294–305. Fluid Dynamics at Interfaces, Gainesville, FL, 1999.
163 F. Xiao and A. Ikebata (2003). An Efficient Method for Capturing Free Boundaries in Multi-
Fluid Simulations. Int. J. Num. Meth. Fluids, 42, 187–210.
164 K. Zaidi, B. Abbes and C. Teodosiu (1997). Finite Element Simulation of Mold Filling using
Marker Particules and the k −  Model of Turbulence. Comp. Meth. Appl. Mech. Engrg, 144,
227–233.
165 Y.L. Zhang, K.S. Yeo, B.C. Khoo and W.K. Chong. Three-Dimensional Computation of Bub-
bles near a Free Surface. J. Comput. Phys., 146, 105–123.
166 H. Zhao, B. Merriman, S. Osher and L. Wang (1998). Capturing the Behavior of Bubbles and
Drops Using the Variational Level Set Approach. J. Comput. Phys., 143, 495–518.

Please address your comments or questions on this paper to:

International Center for Numerical Methods in Engineering
Edificio C-1, Campus Norte UPC
Grand Capitán s/n
08034 Barcelona, Spain
Phone: 34-93-4016035; Fax: 34-93-4016517
E-mail: [email protected]