Received: 1 April 2019 Revised: 29 October 2019 Accepted: 30 December 2019

DOI: 10.1002/fld.4805

RESEARCH ARTICLE

A control volume finite element method for

three-dimensional three-phase flows

Zhihua Xie1,2,3 Dimitrios Pavlidis2 Pablo Salinas2 Christopher C. Pain2

Omar K. Matar3

1
Hydro-environmental Research Centre,
School of Engineering, Cardiff University, Summary
Cardiff, UK A novel control volume finite element method with adaptive anisotropic
2
Applied Modelling and Computation unstructured meshes is presented for three-dimensional three-phase flows with
Group, Department of Earth Science and
interfacial tension. The numerical framework consists of a mixed control volume
Engineering, Imperial College London,
London, UK and finite element formulation with a new P1 DG-P2 elements (linear discon-
3
Department of Chemical Engineering, tinuous velocity between elements and quadratic continuous pressure between
Imperial College London, London, UK elements). A “volume of fluid” type method is used for the interface captur-
Correspondence ing, which is based on compressive control volume advection and second-order
Zhihua Xie, Hydro-environmental finite element methods. A force-balanced continuum surface force model is
Research Centre, School of Engineering,
employed for the interfacial tension on unstructured meshes. The interfacial
Cardiff University, Cardiff CF24 3AA, UK.
Email: [email protected] tension coefficient decomposition method is also used to deal with interfacial
tension pairings between different phases. Numerical examples of benchmark
Funding information
tests and the dynamics of three-dimensional three-phase rising bubble, and
Engineering and Physical Sciences
Research Council, EP/K003976/1, droplet impact are presented. The results are compared with the analytical solu-
EP/M012794/1, EP/R022135/1, tions and previously published experimental data, demonstrating the capability
EP/S016376/1
of the present method.

KEYWORDS
adaptive unstructured mesh, control volume finite element method, discontinuous Galerkin,
interfacial tension, Navier-Stokes model, three-phase flows

1 I N T RO DU CT ION

Multiphase flows, where two or more fluids have interfacial surfaces, are often found in industrial engineering appli-
cations, such as oil-and-gas transportation, nuclear reactors, and microfluidics. These applications typically involve the
motion of bubbles, droplets, fluid films, and jets, featuring tremendous complexity in interfacial topology and a large
range of temporal and spatial scales.
Significant advances have been made in the theoretical,1 experimental,2 and numerical3 studies for better under-
standing of multiphase flow phenomena. With increase in computer power and developments in computational fluid
dynamics, a relatively large number of modeling approaches have been developed in the literature, which provide reliable
and accurate tools to investigate multiphase flows in detail.

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the
original work is properly cited.
© 2020 The Authors. International Journal for Numerical Methods in Fluids published by John Wiley & Sons, Ltd.

Int J Numer Meth Fluids. 2020;1–20. wileyonlinelibrary.com/journal/fld 1

2 XIE et al.

In the past, numerous methods have been proposed and used to simulate two-phase flows on a fixed mesh,4 such as
marker-and-cell,5 volume-of-fluid (VOF),4,6,7 front-tracking,8 level set,9,10 phase field,11 and moment-of-fluid12 methods.
In addition, particle13,14 methods have also been developed. An alternative is to consider the use of dynamically adaptive
mesh methods, where the mesh resolution can vary in time in response to the evolving solution fields. There are some
examples of the use of adaptive mesh refinement for structured meshes with VOF,15 hybrid-level set/front-tracking,16 and
moment-of-fluid17 methods. Unstructured meshes are very attractive when dealing with complex geometries in engineer-
ing applications and there are examples of adaptive unstructured meshes with the interface tracking method,18 level set
method,19 phase field method,20 and algebraic VOF method.21
Despite the fact that there are a number of numerical studies on two-phase flows, research on three-phase flows
(gas-liquid-liquid) is still limited, partly due to the complexity of interactions between phases and numerical treat-
ment of triple junction points. The moment-of-fluid method has been developed in dealing with more than two
materials.22,23 A few numerical examples with interfacial tension can be found in the literature for the level set,24,25 VOF,26
moment-of-fluid,27 phase field,28,29 smooth particle hydrodynamics (SPH),30,31 and moving mesh32 methods. When con-
sidering interfacial tension effects, most of the interface capturing methods use the continuum surface force (CSF)
approach33 in both level set and VOF methods. However, far less attention has been paid to level set and VOF methods
with fully unstructured meshes for three-dimensional (3D) three-phase flows, even without mesh adaptivity.
In this article, we build upon our novel interface-capturing method21 and balanced-force interfacial tension
algorithm,34 from two-dimensional (2D) two-phase flow modeling with interfacial tension and adaptive unstructured
meshes35,36 and 2D three-phase flow modeling without interfacial tension37 to 3D three-phase flow modeling with inter-
facial tension. The main novelty of the present article is the coupling of interface capturing and interfacial tension on
adaptive unstructured meshes for 3D three-phase flow problems, whereas most previous studies focused on two-phase
flows. An adaptive unstructured mesh modeling framework is developed here, which can modify and adapt unstructured
meshes to better represent the underlying physics of multiphase problems and reduce computational effort without sacri-
ficing accuracy. The numerical framework consists of a mixed control-volume and finite-element formulation, a VOF-type
method for the interface-capturing based on compressive control volume advection and second-order finite elements,
and a force-balanced algorithm for the interfacial tension that minimizes spurious velocities often found in such simula-
tions. The interfacial tension coefficient decomposition method has been employed to deal with tension pairings between
different phases via a compositional approach. Numerical examples of some benchmark tests and the dynamics of 3D
three-phase rising bubble and droplet impact are presented to demonstrate the capability of this method. To the best of our
knowledge, this is the first 3D control-volume and finite-element method adaptive-mesh simulation of 3D three-phase
rising bubble and droplet impact dynamics.
The remainder of this article is organized as follows. Description of the model and numerical methods is given in
section 2. 2D and 3D numerical results are presented and compared with analytical solutions or recently published
experimental data in section 3. Finally, some concluding remarks and future work are given in section 4.

2 MATHEMATICAL MODEL AND NUMERICAL METHODS

2.1 Governing equations

A three-phase flow modeling framework has been developed based on the multi-component modeling approach with
information on interfaces embedded into the continuity equations. In multi-component flows, a number of components
exist in one or more phases (one phase is assumed here but the method is easily generalized to an arbitrary number of
phases or fluids). Let 𝛼i be the volume fraction of component i, where i = 1, 2, … , Nc , and Nc denotes the number of
components. The density and dynamic viscosity of component i are 𝜌i and 𝜇i , respectively. A constraint on the system is

∑
Nc
𝛼i = 1. (1)
i=1

For each fluid component i, the conservation of mass may be defined as

𝜕
(𝛼i ) + ∇ ⋅ (𝛼i u) = 0, i = 1, 2, … , Nc , (2)
𝜕t
XIE et al. 3

and the equations of motion of an incompressible Newtonian fluid may be written as

𝜕(𝜌u)
+ ∇ ⋅ (𝜌u ⊗ u) = −∇p + ∇ ⋅ [𝜇(∇u+∇T u)] + 𝜌g + F𝜎 , (3)
𝜕t
∑Nc
where t is the time, u is velocity vector, p is the pressure, the bulk density is 𝜌 = i=1 𝛼i 𝜌i , the bulk dynamic viscosity is
∑Nc
𝜇 = i=1 𝛼i 𝜇i , g is the gravitational acceleration vector, and F𝜎 is the interfacial tension force. In the present study, we
focus on interfacial flows with three components, that is, Nc = 3.

2.2 Numerical methods

2.2.1 Computational grid

The numerical framework consists of a control-volume and finite-element formulation and also a discontinu-
ous/continuous finite-element pair. The domain is discretized into triangular or tetrahedral elements and, in this work,
they are P1 DG-P2 elements (linear discontinuous velocity between elements and quadratic continuous pressure between
elements).21 Figure 1 shows the locations of the degrees of freedom for the this element pair and the boundaries of the
control volumes in two dimensions.

2.2.2 Temporal discretization

A new time discretization scheme is employed here. When high-order discretization is sought, the method is based on
traditional Crank-Nicolson time stepping. This method is often used because it has the simplicity of a two-level time
stepping method, is unconditionally stable, and second-order accurate. However, for interface-capturing applications,
the time discretization scheme is based on the explicit forward Euler time stepping method. This introduces negative
dissipation and is thus a compressive scheme, which helps maintain sharp interfaces. The use of time steps of the order
of the grid Courant number and above can result in numerical oscillations and unphysical solutions. For this reason, an
adaptive 𝜃 parameter is introduced in Reference 21, in which the forward Euler time stepping method is obtained for
𝜃 = 0, the Crank-Nicolson method is obtained for 𝜃 = 0.5, and the backward Euler method is obtained for 𝜃 = 1. The 𝜃
parameter on each control volume face is chosen using a total variation diminishing criterion.21

2.2.3 Spatial discretization for the global continuity and momentum equations

In order to discretize the above governing Equations (2) and (3), a finite element representation for u and p is assumed,
expressed in terms of their finite element basis functions Qj and Pj , respectively, as

u 
∑ ∑p
u= Qj uj and p= Pj pj . (4)
j=1 j=1

Pressure
Velocity

FIGURE 1 A, Finite element used to

discretize the governing equations. The central
position of key solution variables are indicated here
for the P1 DG-P2 element pair.21 B, Diagram showing
the relationship between intersecting control
volumes (shaded area) and the finite elements (A) (B)
4 XIE et al.

Here, u and p are the total number of degrees of freedom for the velocity and pressure representations. Qj = Qj I,
where I is the identity matrix and Qj is the basis function.
Testing the global continuity equation with a quadratic continuous Galerkin basis function Pi and applying integration
by parts once, the discrete form of the global continuity equation can be obtained as

( )
Pi ∇ ⋅ un+1 dV ≈
∫V

− (∇Pi ) ⋅ un+1 dV + Pi n ⋅ un+1 dΓ + Pi n ⋅ ubc

n+1
dΓ = 0, (5)
∫V ∫Γ−Γbc ∫Γbc

where n is the current time level, n is the outward-pointing unit normal vector to the surface Γ of the domain V, and
subscript “bc” means the value on the boundary.
Testing Equation (3) with a linear discontinuous Galerkin (DG) basis function Qi and applying integration by parts
twice over each element e with the 𝜃 time stepping method, the discrete form of the momentum equation can be obtained
as
( )
𝜕(𝜌u)
Qi + ∇ ⋅ (𝜌u ⊗ u) + ∇p − ∇ ⋅ [𝜇(∇u+∇T u)] − 𝜌g − F𝜎 dV
∫VE 𝜕t
( )
𝜌(un+1 − un )
≈ Q + ∇p − 𝜌g − F𝜎 dV
∫V E i Δt
( )
+ 𝜃 Qi 𝜌u∗n+1 ∇ ⋅ un+1 − ∇Qi [𝜇(∇un+1 +∇T un+1 )] dV
∫V E
( )
1
− 𝜃Qi 𝜌u∗n+1 n ⋅ (un+1 − uin
n+1
) − n ⋅ [𝜇(∇un+1 +∇T un+1 ) + 𝜇(∇unb
n+1
+∇T un+1 )] dΓ
∫ΓE 2 nb

( )
+ (1 − 𝜃) Qi 𝜌un ∇ ⋅ un − ∇Qi [𝜇(∇un +∇T un )] dV
∫VE
( )
1
− (1 − 𝜃)Qi 𝜌un n ⋅ (un − unin ) − n ⋅ [𝜇(∇un +∇T un ) + 𝜇(∇unnb +∇T unnb )] dΓ
∫ΓE 2
n+1
+ Qi n(pbc − pn+1 ) dΓ = 0, (6)
∫Γbc

where n is the outward-pointing unit normal vector to the surface ΓE of the element VE , Γbc is a boundary with prescribed
pressure, unb means the value in the neighboring element along the face ΓE , 𝜃 ∈ {0, 1} is the implicitness parameter, and
Δt is the time step size.
The upwind DG method is used for the advection terms, where uin is the upwind velocity calculated from
the neighboring element or boundary and the subscript (∗ ) represents the latest value during the iteration in one
time step.
For the viscous terms 𝜏 = 𝜇(∇u + ∇T u), we use a high-order linear scheme, which results in a compact stencil with
an element coupling only to its surrounding elements. In order to evaluate the viscous stress tensor 𝜏 on the boundary of
element ΓE , we integrate over the volume of two neighboring elements in order to calculate the derivatives on the element
face between the two elements. For example, the derivative in x for the x component of u is obtained as

𝜕u
Qi ux dV = Qi dV − Q n (u − unb ) dΓ, (7)
∫VE1 ∫VE1 𝜕x ∫ΓE1∩E2 i x

𝜕u
Qi ux dV = Qi dV − Q n (u − unb ) dΓ, (8)
∫VE2 ∫VE2 𝜕x ∫ΓE1∩E2 i x

in which ΓE1∩E2 is the shared face between element 1 (E1) and element 2 (E2), u is the x component of velocity u, and
unb is the value of u in the neighboring element along the face ΓE1∩E2 . It is worth noting that this is not only validated for
discontinuous elements but also can be applied for continuous elements.
XIE et al. 5

2.2.4 Projection method

The discretized form of the global mass balance (Equation (5)) and the momentum (Equation (6)) equations is solved
using a pressure projection method. This effectively eliminates the unknown velocity and solves a system of equations for
pressure or pressure correction. The discretized momentum and continuity equations, at time level n + 1, can be written
in matrix form, respectively, as
(Mu + A)un+1 = Cpn+1 + sn+1
u
, (9)

BT un+1 = sn+1
p
, (10)

where un+1 and pn+1 are the FEM solution fields for velocity and pressure, respectively, and sn+1
u
and spn+1 are discretized
sources.
For generality, we have introduced a matrix A, containing advection and diffusion contributions, which may
be distributed and thus cannot be easily inverted. This allows the method to be applied to inertia-dominated or
viscous-dominated flows without modification. Since the velocity is discontinuous between elements, the mass matrix
Mu = ∫V Qi Qj dV is block-diagonal and thus can be easily inverted, each block being local to an element.
E
The solution method proceeds by first solving for an intermediate velocity un+1
∗
using a guessed pressure pn+1 . On the
∗
first iteration within a time step, one may use pn+1 = pn . Equation (9) becomes
∗

(Mu + A)un+1
∗
= Cpn+1 + sn+1
u
. (11)
∗

The matrix equation for velocity to be satisfied is

Mu un+1 + Au∗n+1 = Cpn+1 + sun+1 . (12)
Subtracting these two equations, the velocity correction equation is obtained

Mu (un+1 − un+1
∗
) = C(pn+1 − pn+1 ). (13)
∗

Multiplying this equation by BT M−1u (see Equation (10)) and using the global continuity equation to eliminate out
n+1
u , one obtains the pressure correction equation

BT M−1
u C(p
n+1
− pn+1 ) = −BT un+1
∗
+ sn+1
p
. (14)
∗

This equation is solved for pn+1 and the velocity is corrected using Equation (13).
This pressure equation is typically solved using the implementation of the GMRES Krylov subspace solver38 in
the PETSc framework.39 The Boomer algebraic multigrid preconditioner from the HYPRE library40 is used to accel-
erate the process. The velocity solutions are calculated using GMRES38 with symmetric successive overrelaxation
preconditioning.
It is worth mentioning that in the first time step after the mesh adaption, the interpolated velocity field on the new
mesh is projected to a continuity-satisfied space, otherwise it will not satisfy the divergence free condition. Thus, the
following calculation steps are performed: the interpolated velocity after a mesh adaption is placed into the right-hand
side (RHS) of the pressure equation (Equation (14)); this is then used to produce a pressure correction that is placed in
the RHS of the velocity correction equation (Equation (13)); and these velocity and pressure corrections are then added
to the interpolated velocity and pressures and we commence time stepping from these values. If this modification is not
made, then after a mesh adaption, there is typically a spike in pressure and velocity magnitude in isolated regions. This
can result in poor accuracy in these regions and this typically leads to the simulations becoming unstable.

2.2.5 Interface capturing method

The interface capturing method is based on a compressive advection method, which uses a mathematically rigorous
nonlinear Petrov-Galerkin method that attempts to keep interfaces between components sharp. The mass conservation
6 XIE et al.

for each component (Equation (2)) is solved using a control volume formulation, involving a high-order accurate finite
element projection to obtain fluxes on the control volume boundaries, where these fluxes are subject to flux-limiting using
a normalized variable diagram approach41 to obtain bounded and compressive solutions for the interface. More details
can be found in the work of Pavlidis et al.21 In this study, the interface capturing method has been further extended to solve
three-phase flows in 3D and coupled with interfacial tensions. In contrast to the interface reconstruction in geometrical
VOF methods, the interface is represented as the isosurface of the volume fraction at the value of 0.5 in our algebraic VOF
interface capturing method here.

2.2.6 Interfacial tension model

The interfacial tension force F𝜎 in Equation (3) is obtained commonly in two-phase flows using the CSF method33 as

̃
F𝜎 = 𝜎𝜅 n𝛿, (15)

where 𝜎 is the interfacial tension coefficient, 𝜅 = ∇ ⋅ ñ is the interfacial curvature, ñ is the interface unit normal, and 𝛿
is the Dirac delta function.
However, the above definition has to be further developed for three-phase flows as the fluid flow might be affected by
more than two interfaces at the same time near triple junction points. In addition, the interfacial tension coefficient might
be different for the interface between different phases. Here, the interfacial tension coefficient decomposition method
is employed to deal with tension pairings between different phases using a compositional approach.24,31,42 The physical
interfacial tension coefficients 𝜎ij between phase i and phase j are decomposed into phase-specific interfacial tension
coefficients as 𝜎ij = 𝜎i + 𝜎j , such that for three-phase flows, once the physical interfacial tension coefficients 𝜎ij are known,
the phase-specific interfacial tension coefficients can be obtained as

⎧𝜎1 = 0.5(𝜎12 + 𝜎13 − 𝜎23 ),

⎪
⎨𝜎2 = 0.5(𝜎12 + 𝜎23 − 𝜎13 ), (16)
⎪
⎩𝜎3 = 0.5(𝜎13 + 𝜎23 − 𝜎12 ).

In this approach, the resulting interfacial tension force can be rewritten as sum of each component force:

∑
3
F𝜎 = 𝜎i 𝜅i ñ i 𝛿i . (17)
i=1

∇𝛼i
Here, we use 𝛿i = |∇𝛼i | and ñ i = to reformulate the CSF based on the component volume fraction:
|∇𝛼i |

∑
3
F𝜎 = 𝜎i 𝜅i ∇𝛼i . (18)
i=1

A balanced-force algorithm for the CSF model on unstructured meshes is used here. In contrast to the previous work
on two-phase flows,34 the unit normal and the curvature have to be calculated three times here for each phase. The
diffused interface approach based on the original volume fraction 𝛼i is employed in order to estimate accurately the unit
normal and the curvature 𝜅i on unstructured meshes. More details can be found in Reference 34.

2.2.7 Mesh adaptivity algorithm

As the spatiotemporal dynamics of three-phase flows are highly complex, the dynamically adaptive mesh method has the
advantage to resolve small interfacial features during a typical simulation. Mesh adaptivity can also reduce the computa-
tional effort, as finer mesh is placed around interfaces during their development, while coarser mesh is used away from
these regions, as necessary.
XIE et al. 7

The present model adapts the mesh to the solution without sacrificing the integrity of the boundary (geometry)
or internal boundaries (regions) of the domain. It circumvents the complexities of boundary-conforming Delaunay
methods by operating on the existing mesh. The error measure employed to adapt the mesh is based on the solu-
tion of the phase volume fraction 𝛼i and provides a directional measure. The objective is to obtain a mesh that has
a uniform interpolation error in any direction. This is accomplished with use of a metric, which is related with
the Hessian of the solution field.43 Appropriate scaling of the metric enables the resolution of multiscale phenom-
ena as encountered in multiphase flows. The resulting metric is used to calculate element size and shape. The mesh
optimization method is based on a series of mesh connectivity and node position searches of the landscape, defin-
ing mesh quality that is gauged by a functional. The mesh modification thus fits the solution field(s) in an optimal
manner. The anisotropic mesh adaptivity technique developed by Pain et al.43 is used here. In this article, the pres-
sure and volume fraction projection uses consistent interpolation, which sets the node value on the new mesh based
on its physical position on the old mesh, while the velocity projection uses a conservative Galerkin interpolation
technique.44
The computational cost of the metric construction and the mesh optimization procedure are approximately equal
to the time spent on solving the governing equations in one time step. The frequency of mesh adaption is based on
our experience and we intend to use the Courant number × frequency < 2 for choosing the frequency of refinement
by considering the computational efficiency and accuracy. As we normally adapt the mesh every 10 or 20 time steps,
the overhead and extra computational cost introduced by the adaptive procedure would be around 10% of the total
CPU time.

3 NUMERICAL EXAMPLES

3.1 Spiral-shearing flow

Before we apply our method to coupled fluid flow problems, we first test our control volume-based interface-capturing
method for a pure advection problem, where the fluid interfaces move under a prescribed velocity field. The single vor-
tex problem,7 which is widely used as a benchmark test for two-phase flows, is extended here for three-phase flows.
Two semicircles (radius 0.15) are initially placed on the left and right of the center point at (0.50, 0.75) in a unit square
computational domain. The velocity field is defined by the stream function 𝜓 as
( ) 𝜕𝜓 𝜕𝜓
1 2 𝜋t
𝜓= sin (𝜋x)cos2 (𝜋y) cos , u=− , v= , (19)
𝜋 T 𝜕y 𝜕x

where u and v are the horizontal and vertical components of the velocity field, respectively, and T corresponds to the total
time in the simulation.
The initial interface shape is deformed by the velocity field and then returned to its initial state at t = T, where T = 4
is used in the simulation. In order to avoid the influence of time step size on the results, Δt = 2.5 × 10−4 is used in the
computations with minimum mesh size hmin = 1∕512 and the Courant number is 0.025. Figure 2 shows the interface
shape alongside the mesh during the simulation. It can be seen that the semicircles are stretched from t = 0 under the
specified velocity field, reaching its maximum deformation at t = 2. At this stage, they are spiral in shape with elongated
filament, which is very thin at the tail. Due to the shearing low, the left semicircle moves to the inner part of the spiral.
The fluid interfaces have been efficiently captured in the computation by using the adaptive unstructured mesh, which
provides fine resolution equivalent to that of a 512 × 512 uniform mesh (“equivalent” means two meshes have the same
minimum grid size in the present article). After t = 2, the velocity field is reversed and the interface shape is returned to
its initial shape.
In order to demonstrate the width of the calculated interface, the volume fraction distribution for three phases
near the largely deformed interface at t = 2 is shown in Figure 3. In addition, the enlarged areas in the vicinity
of the tip are also presented on the bottom of Figure 3 to show the relation between the adaptive mesh and the
interface. It can be seen that sharp interfaces are captured by the present compressive VOF method, where the
interfaces normally span one or two elements. It is worth noting that it is not easy to present three-phase
flow results compared with two-phase flows results; thus, we will only show the region for each phase
8 XIE et al.

FIGURE 2 Volume fraction fields (left) and

associated mesh (right) for the single-vortex shearing
flow test case (Equation (19)). Three time levels are
shown at T = 0 (top), T = 2 (middle), and T = 4
(bottom). The adaptive mesh with minimum length
hmin = 1∕512 provides a fine resolution equivalent to
that of a 512 × 512 uniform mesh [Color figure can be
viewed at wileyonlinelibrary.com]

(volume fraction from 0.5 to 1.0) for all the results hereinafter, rather than present three different volume fraction
distributions for each phase.
The comparison of the final deformation of the interface against the analytical solution is shown in Figure 4, where
the simulation result for the initial circle which is not artificially cut in half as carried out in Reference 34 is also presented
here for comparison. It can be seen that the three-phase result is in good agreement with the analytical solution and is
similar to the performance of two-phase flows. The normalized deviation of the outer fluid interface position after one
rotation can be calculated as
1 ∑N |√ |
1 || (x − 0.5) + (y − 0.75) − 0.15||
2 2
N
error = , (20)
L

where N is the number of interpolation points along the interface and L is a reference length scale, which is the radius here.
It is found that the errors for the simulations with minimum mesh size hmin = 1∕512 are 0.28% and 0.51% for two-phase
and three-phase flows, respectively.
Figure 5 shows the time history of the total number of elements during the simulations, Nadaptive , which has been
normalized by the number of elements on an equivalent 512 × 512 fixed mesh, Nadaptive . It can be seen that the ratio
increases gradually before reaching its maximum value when the interfaces are at their maximum deformation and
XIE et al. 9

F I G U R E 3 Detailed volume fraction distribution and associated mesh for phase 1 (left), phase 2 (middle), and phase 3 (right) for the
single-vortex shearing flow test at largest interface deformation T = 2 shown in Figure 2. Zoom in area in the vicinity of the tip is shown on
the bottom [Color figure can be viewed at wileyonlinelibrary.com]

FIGURE 4 The deformation of the final interface shape for

the single-vortex shearing flow test case with adaptive unstructured
mesh with minimum length hmin = 1∕512 (see Figure 2). The red
and blue solid lines are the interface for the three-phase case, the
black dashed line is the analytical solution, and the green solid line
is the two-phase case in Reference 34 [Color figure can be viewed at
wileyonlinelibrary.com]

then decreases gradually to its initial state. It also shows that the maximum number of total elements for the adap-
tive mesh simulation is less than 4% of total degrees of freedom for its equivalent fixed mesh. In order to assess the
mesh quality, the maximum aspect ratio and maximum angle within the elements during the simulation are shown
in Figure 6. It can be seen that the unstructured mesh is anisotropic with large angle and aspect ratio; however,
good results can still be obtained. This test demonstrates the power of the adaptive mesh approach, which can refine
the mesh in the vicinity of the interface or an area of interest, and reduce computational effort without sacrificing
accuracy.
10 XIE et al.

0.04 FIGURE 5 Temporal evolution of the total number of

elements (Nadaptive ) for the single-vortex shearing flow test case
0.035 (see Figure 2) using unstructured mesh adaptivity. The number
has been normalized by the number of cells of a fixed mesh
(Nfixed = 512 × 512), which has equivalent mesh resolution
0.03
[Color figure can be viewed at wileyonlinelibrary.com]
Nadaptive/Nfixed

0.025

0.02

0.015

0.01
0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1
t/T

45 178

40 176

35 174

30 172
max aspect ratio

max angle

25 170

20 168

15 166

10 164

5 162
0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1 0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1
t/T t/T

F I G U R E 6 Time history for the elements' quality: maximum aspect ratio (left) and maximum angle (right) for the single-vortex
shearing flow test case (see Figure 2) [Color figure can be viewed at wileyonlinelibrary.com]

3.2 3D reversible vortex

As in Reference 27, we test the reversible vortex in 3D. A sphere with radius 0.15 and center (0.35, 0.35, 0.35), which is
artificially cut in half along the vertical axis, is placed in the following flow field:

u = 2 cos(𝜋t∕3)sin2 (𝜋x) sin(2𝜋y) sin(2𝜋z),

v = − cos(𝜋t∕3) sin(2𝜋x)sin2 (𝜋y) sin(2𝜋z),
w = − cos(𝜋t∕3) sin(2𝜋x) sin(2𝜋y)sin2 (𝜋z), (21)

where u, v, and w are the components of the velocity field in the x, y, and z directions, respectively, and t corresponds to
the time in the simulation.
The initial interfaces are deformed by the velocity field for 0 < t < 3∕2 and then the flow is reversed for 3∕2 < t < 3,
and finally the interfaces return to their initial state at t = 3. Δt = 1.0 × 10−3 is used in the computations with minimum
mesh size hmin = 1∕100 and maximum mesh size hmax = 1∕4 and the Courant number is 0.1. Only fine mesh is used
in the vicinity of the interface and coarse mesh is used away from the interface in order to reduce the computational
effort. Figure 7 shows the interface shape alongside the mesh during the simulation, in which a reasonable agreement is
XIE et al. 11

FIGURE 7 Three-dimensional
three-phase deformation for the
reversible vortex test case (see
Equation 21) for (a) two-phase case and
(b) three-phase case. Three time levels
are shown at t = 0 (left), t = 1.5 (middle)
and t = 3 (right). The fully unstructured
mesh on the interface between different
phases is shown here [Color figure can
be viewed at wileyonlinelibrary.com]

FIGURE 8 Predicted error of the interfacial location as function of the 100

number of elements. The order of convergence is about 2 and the first-order two-phase formulation
three-phase formulation
slope is shown as black dash-dotted line and the second-order slope is shown
as black dashed line [Color figure can be viewed at wileyonlinelibrary.com]
first order
error

10-1

second order

10-2
103 104 105 105
number of elements

obtained for this 3D shear flow field. For comparison, we also performed a two-phase simulation in which the sphere is
not artificially cut in half and we can see that similar results are obtained.
As the location of the interface is known (which is represented as the isosurface of the outer fluid volume fraction at
𝛼 = 0.5), the normalized deviation of the interface position after the reversible vortex can be calculated as

1 ∑N |√ |
1 || (x − 0.35) + (y − 0.35) + (z − 0.35) − 0.15||
2 2 2
N
error = , (22)
L

where N is the number of interpolation points along the interface and L is a reference length scale, which is the radius
here. Figure 8 shows the convergence for the computations for two-phase and three-phase cases with three different
simulations with minimum mesh size hmin = 1∕25, 1∕50, 1∕100. It is shown that the three-phase case has slightly bigger
error than the two-phase case due to the additional calculation for the third phase. It can be seen that the present method
is close to second-order accurate, which is consistent with the quadratic polynomial function used for the P2 finite element
type. It is worth mentioning that the maximum total number of element for the three-phase case with hmin = 1∕100 is
61506, which is one order of magnitude less than the total number of cells for an equivalent fixed mesh (106 ) and can
reduce the computational effort for the simulations.
12 XIE et al.

FIGURE 9 The floating lens

problem with the initial setup (a) and
the equilibrium shapes for three
different cases for We2 = 60 and
We1 = We3 = 108, 60, 36 shown in (b),
(c) and (d) respectively. Here,
𝜌1 = 𝜌2 = 𝜌3 = 1 and μ1 = μ2 = μ3 = 1.
Fully unstructured adaptive meshes
with minimum length hmin = 1∕256 are
also shown [Color figure can be viewed
at wileyonlinelibrary.com]

3.3 Floating lens

In order to check that the present method produces the correct equilibrium contact angles with triple junction points,
we follow References 24 and 28 to investigate the floating lens problem, where a droplet of one liquid spreads between
two other immiscible liquids, eventually forming a lens shape at steady state (see Figure 9(A)). A circular droplet (fluid
2) is initialized between the interface of fluid 1 and fluid 3 with zero velocity as shown in Figure 9(A). The droplet (radius
0.15) is placed at the center point (0.5, 0.5) in a unit square computational domain, whereas no-slip boundary condition
is applied at the walls. The fluids have the same density and viscosity (𝜌1 = 𝜌2 = 𝜌3 = 1 and 𝜇1 = 𝜇2 = 𝜇3 = 1), and Re =
𝜌L∗ V∗ 𝜌 L V2
𝜇
= 60 and We2 = 2 𝜎∗ ∗ = 60, where L∗ = 60 and V∗ = 1 denote characteristic scales of length and velocity and 𝜎2 is
2
the phase-specific interfacial tension for phase 2. We vary the value of the phase-specific interfacial tension coefficient
𝜎 of phases 1 and 3 in order to consider three different cases when increasing the interfacial tension with We1 = We3 =
108, 60, 36.
In Figure 9(B-D), the equilibrium droplet shapes are shown for the three cases for We2 = 60 and We1 = We3 =
108, 60, 36, respectively. It can also be seen that the fully unstructured meshes with minimum length hmin = 1∕256 and
Courant number of 0.25 are adapted in order to capture the evolution of the interface shapes. The equilibrium contact
angle is known when specifying the interfacial tension coefficients, and there is an analytical solution for the lens area
and its length (the distance between two triple junction points).24,28 Thus, we compare the length predicted by the numer-
ical simulations and compare it against analytical solutions and a previous study in Table 1; this shows a good accuracy
of the present method.
A convergence study with three different fixed meshes and an adaptive mesh has also been carried out for case 2 (We1 =
We3 = 60). Figure 10 shows the comparison between the analytical solution and numerical simulations for the lens' length
XIE et al. 13

T A B L E 1 A comparison of the equilibrium

We1 We2 We3 Analytical Reference 28 Present
distance between triple junctions obtained from
analytical solutions and past (Reference 28) and 108 60 108 0.3746 0.3982 0.3701
present numerical simulations 60 60 60 0.4138 0.4368 0.4069
36 60 36 0.4578 0.4622 0.4485

F I G U R E 10 Evolution of the interfaces of a floating lens for 0.5

Case 2 in Figure 9C with different meshes: 64 × 64, 126 × 128,
256 × 256 uniform meshes and an adaptive mesh with minimum 0.45

distance between triple junctions (m)

length hmin = 1∕256 [Color figure can be viewed at
wileyonlinelibrary.com]
0.4

0.35
Analytical
64 × 64
0.3 128 × 128
256 × 256
Adaptive
0.25

0.2
0 0.5 1 1.5 2 2.5 3
time (s)

F I G U R E 11 CPU time for the simulations of a floating lens for 106

case 2 in Figure 9C with different meshes: 64 × 64, 126 × 128, 64 × 64
128 × 128
256 × 256 uniform meshes and an adaptive mesh with minimum 256 × 256
length hmin = 1∕256. The slope of h−3 is also included for reference Adaptive

[Color figure can be viewed at wileyonlinelibrary.com]

105
CPU time (s)

h -3
104

103
10-3 10-2 10-1
h

during the interface evolution. It can be seen that the result is converged and all the numerical errors for different meshes
are less than 2% when compared with the analytical solution. It is worth noting that the error for the adaptive mesh is
slightly larger when compared with the fixed mesh simulations, which might be due to the fully unstructured mesh. We
note, however, that the adaptive simulations require fewer degrees of freedom (the number of elements is less than 1300)
as the mesh is only refined near the interface.
In order to demonstrate the efficiency and the benefits of complexity introduced by adaptive scheme, Figure 11 shows
the CPU wall time running in a single core for the simulations with different meshes used to generate Figure 10. As
expected, the CPU time increases when refining the mesh. It follows the h−3 slope, which is consistent with the theory of
order of 3 as we double the resolution in each direction in 2D and also double the total time step when keeping the constant
Courant number. It is worth noting that the CPU time for adaptive mesh is less than all three fixed mesh simulations,
being only 0.55% of its CPU time on an equivalent fixed mesh.
14 XIE et al.

3.4 Drop levitation

In this section, we consider a droplet leaving an interface under surface tension forces.24,31 We followed Reference 24 and
the same computational setup for floating lens in Section 3.3 is used here, where a circular droplet (fluid 2) is initialized
between the interface of fluid 1 and fluid 3 with zero velocity as shown in Figure 9A. Here, we consider when wetting
occurs by varying the interfacial tension as 𝜎13 (interfacial tension between top and bottom fluids) and 𝜎23 (interfacial
tension between droplet and top fluid) are equal to unity, whereas 𝜎12 (interfacial tension between droplet and bottom
fluid) are set to 10. Figure 12 shows the predicted results for the evolution of the 2D droplet due to surface tension forces.
The present method is able to deal with large deformation of the interface, which is similar to the results in Reference 24.
From t = 0 to t = 0.3, the area for the 2D droplet for fluid 2 has lost only 0.11% during the simulation.
We also extend this case in a fully 3D simulation. Figure 13 shows the predicted evolution of the 3D droplet at the
initial state, during leaving and about to leave the interface. At t = 0.25, it can be seen that there is an attachment of the

F I G U R E 12 Evolution of a 2D
droplet leaving an interface under surface
tension forces. Here, 𝜌1 = 𝜌2 = 𝜌3 = 1,
μ1 = μ2 = μ3 = 1, 𝜎 13 = 𝜎 23 = 1, and
𝜎 12 = 10. Fully unstructured adaptive
meshes with minimum length hmin = 1∕256
are also shown [Color figure can be viewed
at wileyonlinelibrary.com]

F I G U R E 13 Evolution of
a 3D droplet leaving an interface
under surface tension forces
with the same parameters
shown in Figure 12 [Color figure
can be viewed at
wileyonlinelibrary.com]
XIE et al. 15

T A B L E 2 Physical parameters taken from26 used for the three-phase

Variables Value Units
bubble rising simulation, where results are shown in Figure 14
Bubble diameter d 0.011 m
Water density 𝜌1 1211 kg/m3
Air density 𝜌2 1.2 kg/m3
Oil density 𝜌3 961 kg/m3
Water viscosity 𝜇1 0.1026 kg/(m s)
Air viscosity 𝜇2 1.81 × 10−5 kg/(m s)
Oil viscosity 𝜇3 0.1138 kg/(m s)
Interfacial tension 𝜎12 0.0487 N/m
Interfacial tension 𝜎23 0.0207 N/m
Interfacial tension 𝜎13 0.03 N/m

red color fluid and the blue color fluid at the center and it will eventually leave the bottom fluid during the rising process.
From t = 0 to t = 0.25, the volume for the 3D droplet for fluid 2 has lost 0.25% during the simulation.

3.5 Air bubble crossing an oil-water interface

After validating the numerical predictions against analytical solutions involving an equilibrium three-phase problem and
a clinical test case, in this section, we study 3D three-phase flows of an air bubble crossing a totally stratified oil-water
interface. The experimental setup and results have been reported in Reference 26 and a spherical cap bubble case (com-
putational parameters are presented in Table 2) is considered here. The Bond number is Bo = 𝜌1 gd2 ∕𝜎12 = 29.5 and the
Archimedes number is Ar = 𝜌1 (gd3 )1/2 ∕𝜇1 = 42.6 for this case, where d denotes the bubble diameter. For this simulation,
we use a relative large computational domain 6d × 12d × 6d to avoid boundary effects; on the other hand, the no-slip con-
dition is applied; the center of bubble is initialized at y = 2d, and the oil-water interface is located at y = 5d in the vertical
direction. The minimum length hmin = 0.02d is set in the computation and the maximum Courant number is about 0.5
during the whole simulation.
Figure 14 shows the evolution of the three-phase flow of an air bubble crossing an oil-water interface. 2D pro-
files (Figure 14B) in a central plane associated with the adaptive unstructured mesh (Figure 14C) and 3D side-view
results (Figure 14D) are presented, along with experimental results in Reference 26 (Figure 14A) for comparison. It
can be seen that the initial spherical cap bubble deforms to a slightly dimpled shape when reaching the oil-water
interface. There is a thin film of water above the bubble (shown in the second column of Figure 14B) as it pene-
trates through the interface. The thin film ruptures (as shown in the third and fourth columns of Figure 14B) and
the bubble recovers its spherical cap shape, albeit with a smaller radius in the oil phase. The predicted results are
in agreement with the experimental observations at the early stage and the deviation in the late stage (as shown in
the fifth and last columns of Figure 14B) is mainly due to the difficulty to simulate the rupture of the thin film.
The adaptive meshes on a 2D plane are also shown in Figure 14C, which only refine the resolution in the vicin-
ity of the interfaces and demonstrate the efficiency of the present method. It is worth noting that although the
predicted results in Reference 26 are in better agreement with the experimental observations for this case, they stud-
ied this problem using a 2D axisymmetric computations with a finer grid, whereas a fully 3D simulation is carried
out here.

3.6 Water drop impact onto a deep oil bath

In this section, we further investigate 3D three-phase flows by considering the dynamics of a water droplet impacting
upon a deep oil bath. The experimental measurements have recently been reported in Reference 45, and we follow the
same setup and parameters in that experimental study (shown in Table 3). In the simulation, a water drop with diameter
d = 0.00292 m impacts perpendicularly onto the horizontal surface of a deep silicone oil bath with velocity V = 0.73 m/s.
16 XIE et al.

F I G U R E 14 Bubble rising
through a stably stratified oil-water
interface: evolution of the air bubble
inside oil and water system with
d = 0.011 m, Bo = 29.5, and
Ar = 42.6; the rest of the parameter
values are listed in Table 2. A, The
experimental results in Reference 26
obtained for the same parameters as
in Table 2. B, The numerical results
along the central plane for air (red),
water (blue), and oil (green). C, The
adaptive mesh corresponding to (B).
D, The numerical results in a 3D
side-view. The time period between
two consecutive images is
√
≈ 1.7 d∕g [Color figure can be
viewed at wileyonlinelibrary.com]

T A B L E 3 Physical parameters taken from Reference 45 used for the

Variables Value Units
three-phase droplet impact simulation whose results are shown in Figure 15
Drop diameter d 0.00292 m
Water density 𝜌1 998 kg/m3
Air density 𝜌2 1.2 kg/m3
Oil density 𝜌3 932 kg/m3
Water viscosity 𝜇1 0.001 kg/(m s)
Air viscosity 𝜇2 1.81 × 10 −5
kg/(m s)
Oil viscosity 𝜇3 0.00932 kg/(m s)
Interfacial tension 𝜎12 0.072 N/m
Interfacial tension 𝜎23 0.0201 N/m
Interfacial tension 𝜎13 0.033 N/m
XIE et al. 17

F I G U R E 15 Evolution of a
water drop (diameter d = 0.00292 m)
impacting upon a deep oil bath with
velocity V = 0.73 m/s; the rest of the
parameters are listed in Table 3. A,
The experimental results in Reference
45 obtained for the same parameters
as in Table 3. B, Computational results
from the present article in a 3D view.
C, The numerical results along the
central plane for air (red), water
(blue), and oil (green). D, The adaptive
mesh corresponding to (C). Results
are shown at t = 8.8, 18, 31.2 ms after
the first contact for both the
computational and experimental
results [Color figure can be viewed at
wileyonlinelibrary.com]

The Weber number (We = 𝜌1 V2 d∕𝜎12 ) based on the water drop is 21. A large computational domain 6d × 9d × 6d, with
the air-oil interface at y = 6d, is used to avoid boundary effects, whereas no-slip condition is applied at boundaries, and
the minimum length hmin = 0.02d is set in the computation. During the simulation, the maximum Courant number is
about 0.5.
Figure 15 shows the evolution of the three-phase flow of a water drop impacting upon a deep oil bath. 3D view results
(Figure 15B) and 2D profiles (Figure 15C) in a central plane associated with the adaptive unstructured mesh (Figure 15D)
are presented, along with experimental results in Reference 45 (Figure 14A) for comparison. It can be seen that the water
drop opens up a crater and sets the surrounding oil liquid into motion during impact. The drop becomes flat and spreads
below the crater interface. Afterward, the crater collapses upward due to buoyancy effects, and the water drop decelerates
and deforms mainly due to the interfacial tension force. There is some discrepancy between the simulation and experi-
mental measurements, which is mainly due to the mesh not being sufficiently fine to resolve triple region between the
air, water, and oil.
18 XIE et al.

1 0.5 F I G U R E 16 Evolution of
0
normalized 3D droplet volume
0.995 (A) and bottom of the droplet
-0.5 position (B), in which the
0.99
-1
volume is normalized by its
V/V0

initial volume and position is

y/d
-1.5
0.985 normalized by the droplet
-2 diameter
0.98
-2.5

0.975 -3
0 0.01 0.02 0.03 0.04 0 0.01 0.02 0.03 0.04
t (s) t (s)
(A) droplet volume (B) droplet position

× 10 -3 F I G U R E 17 Number of total elements (Nadaptive ) for the

5
numerical simulation of three-phase water drop impact onto a
4.5 deep oil bath using unstructured mesh adaptivity. The number
droplet spreading has been normalized by the number of cells of a fixed mesh
4
droplet recession (Nfixed = 300 × 450 × 300), which has equivalent mesh resolution
3.5 [Color figure can be viewed at wileyonlinelibrary.com]
Nadaptive/Nfixed

3
crater-formation
2.5

1.5

0.5

0
0 10 20 30 40 50
time (ms)

Figure 16 shows the time history of the droplet volume and its bottom position during the simulation. It can be seen
that the droplet volume has lost less than 2.5% during the simulation and the droplet velocity is higher during the initial
impact and gradually decreases afterward.
In order to demonstrate the efficiency of the adaptive unstructured mesh simulation, Figure 17 shows the time history
of the total number of elements during the simulations, Nadaptive , which has been normalized by the number of elements
on an equivalent fixed mesh, Nfixed . It can be seen that Nadaptive of the adaptive mesh gradually increases after the droplet
impact in order to resolve the complex interfacial structures. The mesh is finest during the crater-generation process, and
the number of elements gradually decreases during droplet spreading and reaches a nearly steady value after the droplet
penetrates the air-oil interface. It can be also seen that the maximum number of elements used in the adaptive simulation
is less than 0.5% of the fixed mesh case with equivalent mesh resolution; this reduces significantly the computational
efforts without sacrificing accuracy.

4 CO N C LUSION S

In this article, a novel control volume finite element method with adaptive anisotropic unstructured meshes has been
presented for 3D three-phase flows with interfacial tension. This method can modify and adapt unstructured meshes to
better represent the underlying physics of interfacial problems and reduce computational effort without sacrificing accu-
racy. The numerical framework consists of a mixed control volume and finite element formulation, a VOF-type method
for the interface capturing based on a compressive control volume advection method and second-order finite element
methods, and a force-balanced algorithm for the interfacial tension implementation. The interfacial tension coefficient
XIE et al. 19

decomposition method has been employed to deal with tension pairings between different phases using a composi-
tional approach. The numerical framework has been validated with several benchmark problems for interface advection,
interfacial tension pairings for a floating lens, and three-phase rising bubble and droplet impact problems.
The proposed method has the ability to handle a range of complex three-phase flow problems in 3D, involving different
phase-phase interactions and triple junctions, in an efficient way. All simulations here were performed in serial using a
single processor (Intel Xeon E5-2697v3 2.6 GHz), and typical runtimes are a couple of hours for 2D and approximately
2 weeks for 3D. Future work will include high-performance computing for large scale problems and the inclusion of
surfactant effects.

ACKNOWLEDGEMENTS
The authors would like to thank the EPSRC MEMPHIS Programme Grant (number EP/K003976/1) for funding this work.
D.P. would like to acknowledge the support from the EPSRC UK-Japan Civil Nuclear Energy project (EP/M012794/1)
and H2020 IVMR. Funding for P.S. from ExxonMobil is gratefully acknowledged. Z.X. would like to acknowledge the
support from the EPSRC projects (EP/S016376/1, EP/R022135/1) and fruitful discussion with Dr. Jie Li at the University
of Cambridge during a short research visit funded by the UK Fluids Network.

ORCID
Zhihua Xie https://orcid.org/0000-0002-5180-8427
Pablo Salinas https://orcid.org/0000-0002-6496-6299

REFERENCES
1. Dussaud AD, Matar OK, Troian SM. Spreading characteristics of an insoluble surfactant film on a thin liquid layer: comparison between
theory and experiment. J Fluid Mech. 2005;544:23-51.
2. Bachalo W. Experimental methods in multiphase flows. Int J Multiphase Flow. 1994;20:261-295.
3. Tryggvason G, Scardovelli R, Zaleski S. Direct Numerical Simulations of Gas-Liquid Multiphase Flows. Cambridge, UK: Cambridge
University Press; 2011.
4. Scardovelli R, Zaleski S. Direct numerical simulation of free-surface and interfacial flow. Annu Rev Fluid Mech. 1999;31:567-603.
5. Harlow FH, Welch JE. Numerical calculation of time-dependent viscous incompressible flow of fluid with free surface. Phys Fluids.
1965;8:2182-2189.
6. Hirt CW, Nichols BD. Volume of fluid (VOF) method for the dynamics of free boundaries. J Comput Phys. 1981;39:201-225.
7. Rider WJ, Kothe B. Reconstructing volume tracking. J Comput Phys. 1998;141:112-152.
8. Unverdi SO, Tryggvason G. A front-tracking method for viscous, incompressible, multi-fluid flows. J Comput Phys. 1992;100:25-37.
9. Osher SJ, Sethian JA. Fronts propagating with curvature dependent speed: algorithms based on Hamilton-Jacobi formulations. J Comput
Phys. 1988;79:12-49.
10. Sethian JA, Smereka P. Level set methods for fluid interfaces. Annu Rev Fluid Mech. 2003;35:341-372.
11. Anderson DM, McFadden GB, Wheeler AA. Diffuse-interface methods in fluid mechanics. Annu Rev Fluid Mech. 1998;30:139-165.
12. Ahn HT, Shashkov M, Christon MA. The moment-of-fluid method in action. Commun Numer Methods Eng. 2009;25:1009-1018.
13. Monaghan JJ. Smoothed particle hydrodynamics. Annu Rev Astron Astrophys. 1992;30:543-574.
14. Sun X, Sakai M, Yamada Y. Three-dimensional simulation of a solid-liquid flow by the DEM-SPH method. J Comput Phys.
2013;248:147-176.
15. Popinet S. An accurate adaptive solver for surface-tension-driven interfacial flows. J Comput Phys. 2009;228:5838-5866.
16. Ceniceros HD, Roma AM, Silveira-Neto A, Villar MM. A robust, fully adaptive hybrid level-set/front-tracking method for two-phase flows
with an accurate surface tension computation. Commun Comput Phys. 2010;8:51-94.
17. Ahn HT, Shashkov M. Adaptive moment-of-fluid method. J Comput Phys. 2009;228:2792-2821.
18. Menon S, Mooney KG, Stapf KG, Schmidt DP. Parallel adaptive simplical re-meshing for deforming domain cfd computations. J Comput
Phys. 2015;298:62-78.
19. Zheng X, Lowengrub J, Anderson A. Adaptive unstructured volume remeshing - II: application to two- and three-dimensional level-set
simulations of multiphase flow. J Comput Phys. 2005;208:626-650.
20. Yue P, Zhou C, Feng J, Ollivier-Gooch C, Hu H. Phase-field simulations of interfacial dynamics in viscoelastic fluids using finite elements
with adaptive meshing. J Comput Phys. 2006;219:47-67.
21. Pavlidis D, Gomes JLMA, Xie Z, Percival JR, Pain CC, Matar OK. Compressive advection and multi-component methods for
interface-capturing. Int J Numer Methods Fluids. 2016;80:256-282.
22. Ahn HT, Shashkov M. Multi-material interface reconstruction on generalized polyhedral meshes. J Comput Phys. 2007;226:2096-2132.
23. Dyadechko V, Shashkov M. Reconstruction of multi-material interfaces from moment data. J Comput Phys. 2008;227:5361-5384.
24. Smith K, Solis F, Chopp D. A projection method for motion of triple junctions by levels sets. Interfaces Free Bound. 2002;4:263-276.
25. Li H, Yap Y, Lou J, Shang Z. Numerical modelling of three-fluid flow using the level-set method. Chem Eng Sci. 2015;126:224-236.
20 XIE et al.

26. Bonhomme R, Magnaudet J, Duval F, Piar B. Inertial dynamics of air bubbles crossing a horizontal fluid-fluid interface. J Fluid Mech.
2012;707:405-443.
27. Li G, Lian Y, Guo Y, et al. Incompressible multiphase flow and encapsulation simulations using the moment-of-fluid method. Int J Numer
Methods Fluids. 2015;79:456-490.
28. Kim J, Lowengrub J. Phase field modeling and simulation of three-phase flows. Interfaces Free Bound. 2005;7:435-466.
29. Boyer F, Lapuerta C, Minjeaud S, Piar B, Quintard M. Cahn-Hilliard/Navier-Stokes model for the simulation of three-phase flows. Transp
Porous Media. 2010;82:463-483.
30. Hu X, Adams N. A multi-phase SPH method for macroscopic and mesoscopic flows. J Comput Phys. 2006;213:844-861.
31. Tofighi N, Yildiz M. Numerical simulation of single droplet dynamics in three-phase flows using ISPH. Comput Math Appl.
2013;66:525-536.
32. Li J. An arbitrary Lagrangian Eulerian method for three-phase flows with triple junction points. J Comput Phys. 2013;251:1-16.
33. Brackbill J, Kothe D, Zemach C. A continuum method for modeling surface tension. J Comput Phys. 1992;100:335-354.
34. Xie Z, Pavlidis D, Salinas P, Percival JR, Pain CC, Matar OK. A balanced-force control volume finite element method for interfacial flows
with surface tension using adaptive anisotropic unstructured meshes. Comput Fluids. 2016;138:38-50.
35. Xie Z, Pavlidis D, Percival JR, Gomes JLMA, Pain CC, Matar OK. Adaptive unstructured mesh modeling of multiphase flows. Int
J Multiphase Flow. 2014;67:104-110.
36. Lei Q, Xie Z, Pavlidis D, et al. The shape and motion of gas bubbles in a liquid flowing through a thin annulus. J Fluid Mech.
2018;855:1017-1039.
37. Pavlidis D, Xie Z, Percival JR, Gomes JLMA, Pain CC, Matar OK. Two- and three-phase horizontal slug flow modelling using an
interface-capturing compositional approach. Int J Multiphase Flow. 2014;67:85-91.
38. Saad Y, Schultz MH. GMRES: A generalized minimal residual algorithm for solving non-symmetric linear systems. SIAM J Sci Stat Comput.
1986;7:836-869.
39. Balay S, Gropp WD, McInnes LC, Smith BF. Efficient management of parallelism in object oriented numerical software libraries. In: Arge E,
Bruaset AM, Langtangen HP, eds. Modern Software Tools in Scientific Computing. Basel, Switzerland: Birkhäuser Press; 1997:163-202.
40. Henson VE, Yang UM. BoomerAMG: a parallel algebraic multigrid solver and preconditioner. Appl Numer Math. 2002;41(1):155-177.
41. Leonard BP. The ULTIMATE conservative difference scheme applied to unsteady one-dimensional advection. Comput Methods Appl Mech
Eng. 1991;88:17-74.
42. Kim J. Phase field computations for ternary fluid flows. Comput Methods Appl Mech Eng. 2007;196:4779-4788.
43. Pain CC, Umpleby AP, de Oliveira CRE, Goddard AJH. Tetrahedral mesh optimisation and adaptivity for steady-state and transient finite
element calculations. Comput Methods Appl Mech Eng. 2001;190:3771-3796.
44. Farrell PE, Maddison JR. Conservative interpolation between volume meshes by local Galerkin projection. Comput Methods Appl Mech
Eng. 2011;200:89-100.
45. Lhuissier H, Sun C, Prosperetti A, Lohse D. Drop fragmentation at impact onto a bath of an immiscible liquid. Phys Rev Lett.
2013;110:264-503.

SUPPORTING INFORMATION
Additional supporting information may be found online in the Supporting Information section at the end of this article.

How to cite this article: Xie Z, Pavlidis D, Salinas P, Pain CC, Matar OK. A control volume finite element
method for three-dimensional three-phase flows. Int J Numer Meth Fluids. 2020;1–20.
https://doi.org/10.1002/fld.4805