PHASE FIELD MODELING AND SIMULATION OF THREE-PHASE FLOWS

JUNSEOK KIM AND JOHN LOWENGRUB

Abstract. In this paper, we derive a thermodynamically consistent phase-field model for flows containing three
(or more) liquid components. The model is based on a Navier-Stokes (NS) and Cahn-Hilliard system (CH) which
accounts for surface tension among the different components and three-phase contact lines. We develop a stable
conservative, second order accurate fully implicit discretization of the NS and three-phase (ternary) CH system. We
use a nonlinear multigrid method to efficiently solve the discrete ternary CH system at the implicit time-level and
then couple it to a multigrid/projection method that is used to solve the NS equation. We demonstrate convergence
of our scheme numerically and perform numerical simulations to show the accuracy, flexibility, and robustness of
this approach. In particular, we simulate a three interface contact angle resulting from a spreading liquid lens on
an interface, a buoyancy-driven compound drop, and the Rayleigh-Taylor instability of a flow with three partially
miscible components.

Key words. ternary Cahn-Hilliard system, nonlinear multigrid, ternary fluid flow, interfacial tension, arbitrary
miscibility

1. Introduction. Many biomedical, chemical, and industrial processes involve mixtures

of three or more liquids components. In spite of the importance of three-phase flows, most
studies of three-phase systems do not consider hydrodynamical interactions (e.g. see [4, 5, 6,
7, 10, 15, 18, 20]). There have been few theoretical and numerical studies of flows containing
three or more liquid component compared to the large body of research for two-phase fluid
flows. This is partly due to the difficulties in dealing with hydrodynamics associated with
interfaces and triple junctions. In [41], a projection method is used for the motion of a triple
junction in level set framework. In this approach, a linear projection of the level set functions
onto a reduced set of variables is used to solve problems at triple lines and other multiple
junctions (i.e., quadruple points). In [28], a compound drop consisting three immiscible
fluids is simulated using the immersed boundary method [37]. The presence of the interface
and its effect on the flow is established via source terms in the governing equations.
There are a number of three-phase models that describe two immiscible fluids and sur-
factant. In low Reynolds number flows, boundary integral methods have been used to study
the effect of surfactants on drop dynamics (e.g. [33, 35, 42]) and tip-streaming [17]. In
[22, 24, 39], the effect of insoluble surfactants on drop deformation is studied in two and three
dimensions using volume-of-fluid methods. In [25] the effect of surfactants on the dynamics
of rising bubbles is investigated using an immersed boundary/front-tracking algorithm. In
[13], a hybrid level-set/ front-tracking algorithm was used to study the effect of surfactants
on capillary waves. Further, in [27], the effect of surfactants on the evolution of the shape
of an initially nonspherical drop translating in an otherwise quiescent fluid at low Reynolds
number is examined. A combination of the boundary-integral method and a finite-difference
scheme is used to solve the coupled fluid dynamics and surfactant transport problems.
In this paper, we build upon our results for two phase hydrodynamic systems [30] and
three-phase systems in the absence of hydrodynamic interactions [31] to model and simulate
general three-phase hydrodynamic systems. Advantages of this approach over level-set and
immersed boundary approaches described above, for immiscible three-phase systems, are:
(1) we do not need to perform any correction steps to multiple junctions; (2) it is easy to
incorporate other physical properties such as miscible and immiscible fluid components.

Department of Mathematics 103 Multipurpose Science and Technology Building University of California,

Irvine, CA 92697-3875 ([email protected]).

Corresponding author. Department of Mathematics 103 Multipurpose Science and Technology Building Uni-

versity of California, Irvine, CA 92697-3875 ([email protected]).

1
2 Junseok Kim and John Lowengrub

We derive a thermodynamically consistent system of governing equations based on a

phase-field approach. The system of equations couples the Navier-Stokes equations for the
fluid motion to a system of Cahn-Hilliard type (fourth order nonlinear advection-diffusion
equations) for the phase variables. Here the phase fields have a definite physical meaning:
they are the mass ratios of the fluid components.
We develop a conservative, second order accurate fully implicit discretization of the NS
and three-phase (ternary) CH system. It can be shown that if the time step is small enough,
the discretization has a discrete energy functional that is the natural discretization of the free
energy on the continuous level. In practice, we find that the discrete energy is always non-
increasing. We use a nonlinear multigrid method to efficiently solve the discrete ternary CH
system at the implicit time-level and then couple it to a multigrid/projection method that is
used to solve the NS equation.
We present examples of flow with miscible and immiscible components. We demon-
strate the convergence of our algorithm through a resolution study. In addition, we find good
agreement with the theory for an equilibrium liquid lens (lying atop an interface). We provide
demonstrations of liquid/liquid remediation. In the first example, a compound drop is simu-
lated, in which a light fluid encapsulated a heavy contaminant drop. The light fluid causes the
compound drop to rise and deposit the contaminant at an interface where it may be removed.
In the second example, we investigate the diffusional transfer of a preferentially miscible con-
taminant from one immiscible phase to another. In this example, the transfer is enhanced by
the flow and in particular the Rayleigh-Taylor instability.
The contents of this paper are as follows. In Section 2, the governing equations are de-
rived. In Section 3, we derive the discrete scheme and numerical solution. We also present
the approximate projection method used to solve the discrete generalized NS equations. Nu-
merical experiments are presented at Section 4. In Section 5, conclusions are drawn.

2. Derivation of the governing equations. We begin by deriving a thermodynamically

consistent system of governing equations for a general heterogeneous, isothermal mixture of
Nf fluids following the strategy developed for binary (two component) fluids by Lowengrub
and Truskinovksy [34]. Let the mass concentrations be ck = Mk /M for k = 1, . . . , Nf ,
where Mk are the masses of the components in a representative material volume V and M
PNf PNf
is the total mass of the mixture. Since M = k=1 Mk , we have k=1 ck = 1. Suppose
that each component moves with a velocity uk and has density k = Mk /Vk where Vk
is the volume of fluid k. Introducing volume fraction k = Vk /V , we have the relation
PNf
ck = k k where = k=1 k k is the density of the mixture. Herein, we will assume
that each component is incompressible, i.e. k is constant. Note that this does not mean that
the mixture density is constant since depends on k or, equivalently, on ck . The relation
between and ci is given by

Nf
1 X ck
= (2.1)
k
k=1

Such mixtures were termed quasi-incompressible by Lowengrub and Truskinovsky [34].

2.1. Balance equations. The balance of mass for each component is

(k k )
+ (k k uk ) = 0. (2.2)
t
 Phase field modeling and simulation of three phase flows 3

Define the mass-averaged (mixture) velocity field

Nf Nf
1X X
u= k k uk = ck uk , (2.3)

k=1 k=1

then summing Eq. (2.2) in k we obtain the balance of mass for the mixture

+ (u) = 0. (2.4)
t
From Eqs. (2.2) and (2.4), we obtain the mass concentration equation:
ck = Jk , (2.5)
where = t + u is the advective derivative with respect to the mixture velocity, and
Jk = ck wk , where wk = (uk u) , (2.6)
is the diffusion flux.
For each component, we have the balance of linear momentum
Dk uk
ck = Pk + ck g + k , (2.7)
Dt
where Dk /Dt = t + uk is the advective derivative with respect to the component
velocity, Pk is the stress tensor, g is the gravity force and k are the forces per unit volume
due to interactions with other phases. Note that as yet Pk and k are unspecified. Summing
PNf
over k, and requiring that k=1 k = 0, which is necessary for the conservation of linear
momentum of the mixture, we get the following linear momentum equation for the mixture:
u = P + g, (2.8)
PNf
where P = k=1 (Pk ck wk wk ) is the stress tensor of the mixture.
Following classical theory (e.g. see [21] and Lowengrub and Truskinovsky [34]), we
focus on the mixture equations and not on the detailed force interactions (e.g. i ). Therefore,
we consider an energy balance for the entire system and derive thermodynamically consistent
constitutive relations for P and Jk as follows. Let be an arbitrary domain that moves with
the mixture velocity u. Then, the integral form of the energy balance for the mixture is

  Nf
d 1
Z Z X Z
e + |u|2 d = Pn u + (tk n) ck d+ (r + g u) d,
dt (t) 2
k=1
(2.9)
where e is the internal energy, |u|2 /2 is the kinetic energy and n is the outward normal
vector to . The first term on the right hand side is the rate of work done on by the fluid
stress and the extra stresses due to concentration gradients, i.e. tk is a generalized force that
is as yet undetermined. This term is suggested by the variational analysis of Lowengrub and
Truskinovsky [34]. In the second term on the RHS, r is the density of heat sources necessary
to ensure that the temperature is constant and the remaining term is the rate of work done due
to gravity.
Using the mixture mass (2.4) and momentum balance equations (2.8), the local form of
the energy balance Eq. (2.9) is
Nf
X
e = P : u + (tk ck ) + r. (2.10)
k=1
4 Junseok Kim and John Lowengrub

2.2. Thermodynamics and constitutive relations. Because we are dealing with isother-
mal flow, it is useful to introduce the Helmholtz free energy F rather than the internal energy.
The relation between the two is

F = e T s. (2.11)

where T is the temperature. Thus, Eq. (2.10) becomes

Nf
T s = F + P : u +
X
(ti ci ) + r. (2.12)
i=1

Next, we make the constitutive assumption that the free energy

F = F c1 , . . . , cNf , c1 , . . . , cNf , (2.13)

such that
Nf  
F F
F =
X
ck + (ck ) . (2.14)
ck ck
k=1

Using the identity (ck ) = ck (u)T ck (e.g. see [34]) and plugging Eq. (2.14) in
Eq. (2.12) gives

Nf
X F
T s = P + ck : u
ck
k=1
Nf   Nf  
X F X F 1
+ tk ck tk ck + r (2.15)
ck ck
k=1 k=1

Next observe that because the fluid components are incompressible the velocity gradient u
and ck are not independent. They are related via Eqs. (2.1) and (2.5). That is, there is a
degeneracy in Eq. (2.15) since

Nf Nf
X 1 X 1
u = ck = Jk . (2.16)
ck 2 ck
k=1 k=1

Note that from Eq. (2.1) we have

1 1
2
= k , for k = 1, . . . Nf , (2.17)
ck k

is a constant.
To exploit the degeneracy introduce a scalar Lagrange multiplier p. This is the mixture
pressure. Then,

Nf
X
pI : u = k pck , (2.18)
k=1
 Phase field modeling and simulation of three phase flows 5

where I is the identity matrix. Using Eq. (2.18) in Eq. (2.15), we get

Nf
X F
T s = P + pI + ck : u
ck
k=1
Nf   Nf  
X F X F 1
+ tk ck + k p tk ck + r.
ck ck
k=1 k=1
(2.19)

According to the second law of thermodynamics, in the form of the Clausius-Duhem inequal-
ity, we have

0, where s + J r/T (2.20)

where is the internal dissipation and J is the entropy flux (see Truesdell and Noll [45]).
Now, from Eq. (2.19) together with Eq. (2.5), we obtain

Nf Nf  
1 X F 1 X F
1 Jk


= P + pI + ck : u + tk
T ck T ck
k=1 k=1

Nf Nf
1 X X k Jk
+ k Jk + J , (2.21)
T T
k=1 k=1

where k is the generalized chemical potential given by

F 1
k = + k p tk . (2.22)
ck
PNf
Next, because k=1 Jk = 0 this leads to additional constraints on the constitutive relations
PNf
for the stress tensor P, the forces tk and the fluxes Jk . Since k=1 ck = 1, the concentration
PNf 1
fields are not independent. We may thus define cNf = 1 k=1 ck . This leads to the
equivalent entropy form

Nf  
1 X F F
= P + pI + ck : u
T ck cNf
k=1
Nf  
1 X F F
1 Jk


+ tk
T ck cNf
k=1


Nf Nf
1 X
X k Nf Jk
+ k Nf Jk + J , (2.23)
T T
k=1 k=1

PNf

Taking the diffusion flux J = k=1 k Nf Jk /T we are now in a position to pose
thermodynamically consistent constitutive relations for P, tk and Jk . Following Coleman
and Noll [14] where u is varied independently from the other quantities leads to the consti-
6 Junseok Kim and John Lowengrub

tutive assumptions
Nf    
X F F 2
P = pI ck + D ( u) I , (2.24)
ck cNf 3
k=1
 
F F
tk = , (2.25)
ck cNf
Nf 1

X
Jk = k k Nf , for k < Nf , and JNf = Jk (2.26)
k=1

where D = u + uT is the rate of strain tensor and is the viscosity (note that the bulk

viscosity is assumed to be 0 for simplicity and that may be a function of c). We thus obtain
Nf
1 X
k | k Nf |2 0


= D : D + (2.27)
T T
k=1

and so the second law of thermodynamics is satisfied. Note that concentration gradients give
rise to extra fluid stresses. As will be discussed later, these mimic surface tension stresses.
Finally, although we have singled out one of the components, we demonstrate later that the
evolution of our system does not depend on the labeling of the components. We also note that
other constitutive relations may be taken (e.g. see Appendix A as well as [19]). Our choice
here is consistent with that considered previously by Morral & Cahn [36] for linearized three-
phase systems.
2.3. Summary of general equations. Putting together the results from the previous
section, the thermodynamically consistent system of equations governing a mixture of Nf
fluids is
Nf
X



u= k Nf k k Nf , (2.28)
k=1

Nf  
X F F
u = p ck
ck cNf
k=1
  
2
+ (c) D ( u) I + g,
3
(2.29)



ck = k k Nf , (2.30)

and
  
F 1 F F 1
k = + k p , k = (2.31)
ck ck cNf k
PNf 1
for k = 1, . . . , Nf 1 and cNf = 1 k=1 ck . This system couples a generalized Navier-
Stokes equation with a nonlinear advection-diffusion equation for the concentration.
2.4. Special choice of free energy, the Navier-Stokes-Cahn-Hilliard system and nondi-
mensionalization. To make further progress, we need to choose the form of the Helmholtz
 Phase field modeling and simulation of three phase flows 7

free energy. Following Cahn and Hilliard [12], we take

Nf 2
X k
|ck |2


F = F c1 , . . . , cNf + (2.32)
4
k=1

This gives
F 1
= 2k ck (2.33)
ck 2
Note that this makes the concentration equation (2.30) a fourth order nonlinear advection
diffusion equation and is a generalization of the classical Cahn-Hilliard equation used to
describe phase separation in binary mixtures [12]. Further, the extra stress in the Navier-
Stokes equation (2.29) can be written as
Nf   Nf 2
X F F X k
ck = ck ck , (2.34)
ck cNf 2
k=1 k=1

PNf
where we have used that k=1 ck
= 1. Further, we have

Nf 1
F F 1 X
+ k Nf p 2k ck + 2Nf


k Nf = cj , (2.35)
ck cNf 2 j=1

PNf
where we have again used that k=1 ck = 1. We may therefore define, for k = 1,. . . ,Nf 1,

Nf 1
F (c1 , . . . , cNf 1 ) = F c1 , . . . , cNf 1 , 1
X
ck , (2.36)
k=1

k = k Nk (2.37)

so that
F F F
= . (2.38)
ck ck cNf

Analogously, one may define

k = k Nf

Nf 1
F 1 X
= k p 2k ck + 2Nf
+ cj , (2.39)
ck 2 j=1

for k = 1,. . . ,Nf 1. The resulting system is a coupled Navier-Stokes and Cahn-Hilliard
system (see also [34]). We note that by equipping this system with natural boundary condi-
tions, i.e. u = 0 and ck n and k n = 0, this system (with gravity g = 0) has a discrete
(free) energy functional

Nf 2

Z
F (c1 , . . . , cNf 1 ) +
X
E(t) = k
|ck |2 + |u|2 d (2.40)
4 2
k=1
8 Junseok Kim and John Lowengrub

where is the (fixed) physical domain. In the remainder of the paper, we drop the tilde
notation for simplicity.
We nondimensionalize the system as follows (e.g. see also [34]). Let L and V denote
characteristic scales of length and velocity. Then introduce the dimensionless independent
variables x = x/L and t = V t/L and the natural scaling of the dependent variables u =
u/V , = / , = / , p = p/( V2 ),
k = k /k , etc...., where again the stars denote
characteristic quantities. We also assume that the Helmholtz free energy is given by the sum
Nf
k Fk .
X
F = (2.41)
k=1

The flow is then governed by the following nondimensional parameters:

k k / V L
Ck = p , Mk = , P ek = ,
L k L k k k
L V2 V L V
Ak = k , W ek = , Re = , Fr = p , (2.42)
k L |g|
where the top row are the nonclassical additional parameters introduced in the model and the
second row are the classical fluid dynamics parameters. In the top row, the first parameter
Ck is the Cahn number that is a nondimensional measure of the interface energy of the kth
component, the second Mk is a measure of the relative strength of the surface tension and
chemical energies and the third P ek is the diffusional Peclet number that measures the relative
strengths of (chemical) diffusion and advection. In the bottom row, the first parameter Ak is
a nondimensional measure of the density differences between components, the second W ek
is the Weber number where k is the phase specific surface tension (see [41] and below), the
third Re is the Reynolds number that measures the relative strength of inertial and viscous
forces, and finally the last F r is the Froude number that measures the relative strengths of the
inertial and gravitational forces.
Omitting the bar notation, the nondimensional Navier-Stokes-Cahn-Hilliard (NSCH)
system (2.28)-(2.31) is written as
Nf 1
X Ak
u= (k k ) , (2.43)
P ek
k=1

Nf 2
X C k
u = p ck ck
M k W ek
k=1
  
1 2 1
+ (c) D ( u) I + G,
Re 3 F r2
(2.44)
1
ck = (k k ) , (2.45)
P ek
where we have subtracted a linear term x g from the pressure and G = g/|g|. Further,
the chemical potential is

F
2 NX f 1
1 C 2 C Nf
k = + Ak W ek Mk p k ck + ci , (2.46)
ck 2 2 i=1
 Phase field modeling and simulation of three phase flows 9

for k = 1, . . . , Nf 1, where

F
N
X f
j Fj
= . (2.47)
ck ck
j=1 k

Finally, the nondimensional total free energy is

3
Ck2
 
2 1
Z Z

X
2
E(t) = |u| d + Fk + |ck | d. (2.48)
2 M k W ek 4
k=1

Hereafter, we drop the bar and double tilde notation.

2.5. Asymptotics and the sharp interface regime. The miscibility properties of the
flow components can be described through the free energy F (c1 , c2 ). For example, if all
the components are miscible the free energy is a convex function of its components (i.e.
Hessian matrix is positive definite). If the components are immiscible, the free energy is a
non-convex function of its components to reflect the coexistence of multiple phases. Various
mixtures of miscible and immiscible components can analogously be described. When the
flow components are immiscible, the NSCH system should reduce to the classical Navier-
Stokes equations together with the Laplace-Young surface tension jump conditions across
interfaces and multi-junctions. This requires assumptions on the nonclassical parameters Ck ,
Mk and P ek . Following the asymptotic analyses of Lowengrub and Truskinovsky [34] and
others (e.g. see the review [2]) leads to the scaling

Ck = 2 , Mk = / and Pek = O(1) or O(1/), (2.49)

where the parameter is a nondimensional measure of interface thickness. Then, it can be

shown that in the sharp interface limit 0, the classical Navier-Stokes system equations
and jump conditions are recovered. An interface separating two immiscible fluids has an
equilibrium profile ceq (z) where z is the coordinate in the normal direction to the interface.
The parameter is then given by
Z +  eq 2 !1
eq c
= (c ) dz. (2.50)
z

For example, with a free energy F (c) = 14 c21 c22 , where we assume that c3 = 0, then
  
ceq (z) = 1 tanh z/2 2 /2 (2.51)

and, if (c) = 1, then = 6 2.
For three-phase flows, following [41], the surface tension ij between immiscible flow
components i and j are decomposed into the phase-specific surface tensions i and j by

ij = i + j (2.52)

That is, given ij , a linear system of equations is solved for i and j . The phase-specific
surface tensions are used in the definition of the Weber number (2.42). For flows containing
four or more immiscible fluids, the decomposition above is overdetermined and it is possible
that no solution exists. Nevertheless, the model system (2.43)-(2.48) is still valid although an
alternate means may need to be used to determine the individual Weber numbers W ek . This
is currently under study.
10 Junseok Kim and John Lowengrub

2.6. The Boussinesq approximation and the ternary system. We next consider the
special case of a ternary system and use the Boussinesq approximation. In the Boussinesq
approximation, the densities of the flow components are nearly matched such that 1 but
the Froude number may be small enough such that ( 1) /F r2 is non-negligible. Thus, in
Eqs. (2.43)-(2.46), we take = 1 except in the gravitational term and we take Ak = 0. The
resulting system is the ternary version of model H in the nomenclature of Hohenberg and
Halperin [23].
The composition of a ternary mixture (A, B, and C) can be mapped onto an equilateral
triangle (the Gibbs triangle [38]) whose corners represent 100% concentration of A, B or C
as shown in Fig. 2.1(a). Mixtures with components lying on lines parallel to BC contain
the same percentage of A, those with lines parallel to AC have the same percentage of B
concentration, and analogously for the C concentration. In Fig. 2.1(a), the mixture at the
position marked contains 60% A, 10% B, and 30% C (The total percentage must sum to
100%).

C (0,0,1)

A B (1,0,0) (0,1,0)

(a) (b)
F IG . 2.1. (a) Gibbs triangle. (b) Contour plot of the free energy F (c)

Let c = (c1 , c2 ) be the phase variable (i.e. concentrations of component A and compo-
nent B). Since c1 + c2 + c3 = 1 we only need to solve the equations with c1 and c2 . Here,
for simplicity, we consider a constant mobility1 (k 1). F (c) is the Helmholtz free energy
which is defined on the Gibbs triangle. For three immiscible fluids, the free energy can be
modeled by

3
X
F (c) = Fk , Fk = c2k (c2rem(k,3)+1 + c2rem(k+1,3)+1 )/8 (2.53)
k=1

where rem(x, y) is the remainder of the ratio x/y and c3 = 1 c1 c2 . The contours of the
free energy F (c) projected onto the Gibbs triangle are shown in Fig. 2.1(b). Note the energy
minima are at the three vertices and the maximum is at the center.

1 The extension to more general k = k (c1 , c2 ) is straightforward.

 Phase field modeling and simulation of three phase flows 11

The non-dimensional Boussinesq ternary NSCH system is as follows:

u = 0, (2.54)
1
u = p + ((c)(u + uT ))
Re
3
X 1
(ck ck ) + g, (2.55)
W ek F r2
k=1
1
c = , (2.56)
Pe
= f (c) c, (2.57)

where f (c) = (f1 (c), f2 (c)) = (c1 F (c), c2 F (c)) and is the matrix
 2
2 /2


.
2 /2 2
The natural boundary and initial conditions for the ternary NSCH equation are
c
= = 0, and u = 0 on , c(x, 0) = c0 (x), u(x, t) = 0, (2.58)
n n
where n is the normal unit vector pointing out of . Finally, the nondimensional total free
energy for the Boussinesq system is
3
Ck2
 
1 2 1
Z X Z
2
E(t) = |u| d + Fk + |ck | d, (2.59)
2 M k W ek 4
k=1
P3
where F = k=1 Fk .
3. Numerical Solution. The numerical solution of the ternary NSCH system uses a
second-order accurate spatial discretization and a Crank-Nicholson type time stepping method.
For simplicity and clarity of exposition, we will present the numerical method in 2D, but the
extension to 3D is straightforward. The computational grid consists of square cells of a uni-
form size h; these cells ij are centered at (xi = (i 0.5)h, yj = (j 0.5)h), where
i = 1, , L and j = 1, , M . Given un1 , un , cn1 , cn , defined at cell centers and
1 1
pn 2 defined at cell corners, we want to find un+1 , cn+1 , and pn+ 2 . The outline of the
algorithm is as follows:

Step 1. Initialize c0 to be the locally equilibrated concentration profile and u0 to be the

divergence-free velocity field.

Step 2. Solve the CH system and update the concentration field cn to cn+1 . We use a
nonlinear Full Approximation Storage (FAS) multigrid method to solve the nonlinear discrete
system (3.1) and (3.2) given below at implicit time level. The nonlinearity is treated using
a nonlinear Gauss-Seidel relaxation. Details of this step are presented in our recent paper
[31]. Here, however, we have additional source term due to advection. That is, we solve the
following second-order accurate discrete system

cn+1
ij cnij 1 n+ 1 1
= d ij 2 (u d c)n+ 2 , (3.1)
t Pe
n+ 21
ij = (c nij , cn+1 ) 1 d (cnij + cn+1 ), (3.2)
ij ij
2
12 Junseok Kim and John Lowengrub

where d is the standard five point discretization of Laplacian operator in 2D. The advection
1
term (ud c)n+ 2 is approximated by using a fifth order weighted essentially non-oscillatory
(WENO) scheme [26] and is described in Sec. 3.1. The term = (1 , 2 ) and 1 (...) and
2 (...) denote Taylor series approximations to f1 and f2 up to second order, respectively:
1
1 (cn , cn+1 ) = f1 (cn+1 ) c1 f1 (cn+1 )(cn+11 cn1 )
2
1 1
c2 f1 (cn+1 )(cn+12 cn2 ) + c21 f1 (cn+1 )(cn+1
1 cn1 )2
2 3!
2 1
+ c2 c1 f1 (cn+1 )(cn+11 cn1 )(cn+1
2 cn2 ) + c22 f1 (cn+1 )(cn+1
2 cn2 )2
3! 3!
and
1
2 (cn , cn+1 ) = f2 (cn+1 ) c1 f2 (cn+1 )(cn+11 cn1 )
2
1 1
c2 f2 (cn+1 )(cn+12 cn2 ) + c21 f2 (cn+1 )(cn+1
1 cn1 )2
2 3!
2 1
+ c2 c1 f2 (cn+1 )(cn+11 cn1 )(cn+1
2 cn2 ) + c22 f2 (cn+1 )(cn+1
2 cn2 )2 .
3! 3!
As described in [31], this discretization ensures that in the absence of flow, a discrete
version of Eqs. (2.56)-(2.57) is non-increasing in time independent of the choice of t. This
gives enhanced stability.
1
Step 3. Update the velocity un to un+1 and the pressure pn+ 2 . We use the following
approximate projection method adapted from [8]. We solve
1
u un 1 1 (cn+ 2 ) 1
= d pn 2 + d (cn+1 )[d u + (d u )T ] + G
t 2Re F r2
1 n+ 1 1
+ d (cn )[d un + (d un )T ] + Fst 2 (u d u)n+ 2 (3.3)
2Re
using a multigrid method for the intermediate P3 velocity u without strictly enforcing the in-

compressibility constraint. We use Fst = k=1 W ek (|ck |2 I ck ck ) and the
P3
pressure field is replaced by p + k=1 Wek |ck |2 . This modification is performed for two
reasons. First, by taking into account analytic cancellation a simpler and more accurate dis-
cretization may be obtained. Second, when the interface is flat, this term is equal to zero
which is consistent with the fact that the interfacial force vanishes for flat interfaces.
The discretization of Fst is given in Sec. 3.2. The terms d p and d (c)[d u+d uT ]
are defined as following.
pi+ 21 ,j+ 21 + pi+ 12 ,j 21 pi 21 ,j+ 21 pi 12 ,j 21

(d p)ij = ,
2h
pi+ 12 ,j+ 21 pi+ 12 ,j 21 + pi 12 ,j+ 21 pi 21 ,j 21

.
2h

The first component of the viscous term d (c)[d u + d uT ] is discretized as

1 1 
d (c)[d u + d uT ] ij = 2 2(ci+ 12 ,j )(ui+1,j uij ) 2(ci 21 ,j )(uij ui1,j )
h
+ (ci,j+ 12 )(ui,j+1 uij + 0.25(vi+1,j+1 vi1,j+1 + vi+1,j vi1,j ))

(ci,j 21 )(uij ui,j1 + 0.25(vi+1,j vi1,j + vi+1,j1 vi1,j1 )) .
 Phase field modeling and simulation of three phase flows 13

The second component of the viscous term is discretized in a similar manner. The term
1
(u d u)n+ 2 is computed using a fifth order WENO scheme described in Sec. 3.1.
Then project u onto the space of approximately divergence-free vector fields and get

un
the velocity un+1 , i.e., u = un+1 + td , where satisfies d = d u t with the

Neumann boundary condition, n = 0, see [1] for more general boundary conditions. Note
1 1
d un+1 0 see [8]. Finally, update pressure by pn+ 2 = pn 2 + . This completes one
time step.
3.1. Approximation of the advection terms. In this section, we describe the discretiza-
n+ 1 n+ 1
tion of the advection terms. The values uij 2 and cij 2 are calculated using a second-
n+ 12 n1
order accurate extrapolation from previous values, i.e., uij = (3unij uij )/2 and
n+ 21 n1
cij = (3cnij cij )/2. From these cell centered values we obtain cell edged values
n+ 1 n+ 12 n+ 1 n+ 1 n+ 12 n+ 1
by ui+ 12,j = (uij + ui+1,j
2
)/2 and ui,j+2 1 = (uij + ui,j+1
2
)/2.
2 2
n+ 1 n+ 1
In general, the normal velocities ui+ 12,j and vi,j+21 at the edges are not divergence-free.
2 2
To reduce the overall error, we apply the MAC projection [43] before construction of the
convective derivatives. The equation
1
d = MAC un+ 2 (3.4)

is solved for a cell centered , where

n+ 1 n+ 1 n+ 1 n+ 1
ui+ 12,j ui 12,j vi,j+21 vi,j21
n+ 12
(MAC u )ij = 2 2
+ 2 2

h h

with the Neumann boundary condition, n = 0 The resulting linear system (3.4) is solved
using a multigrid method with Gauss-Seidel relaxation. Then the discrete divergence-free
cell-edge velocities u
and v are defined by

n+ 1 n+ 1 i+1,j ij n+ 1 n+ 1 i,j+1 ij
u
i+ 12,j = ui+ 12,j , vi,j+21 = vi,j+21 .
2 2 h 2 2 h
The convective terms are discretized as:

n+ 1 n+ 1
n+ 1
i+ 12,j + u
u i 12,j n+ 1 n+ 1
(u d u)ij 2 = 2 2
( i 12,j )
ui+ 12,j u
2h 2 2

n+ 1 n+ 1
vi,j+21 + vi,j2 1 n+ 1 n+ 1
+ 2 2
ui,j+2 1 u
( i,j2 1 ),
2h 2 2

n+ 1 n+ 1
n+ 12
u
i+ 12,j + u
i 12,j n+ 1 n+ 1
(u d c)ij = 2 2
(ci+ 12,j ci 12,j )
2h 2 2

n+ 1 n+ 1
vi,j+21 + vi,j2 1 n+ 1 n+ 1
+ 2 2
(ci,j+2 1 ci,j2 1 ),
2h 2 2

n+ 1 n+ 1 n+ 1 n+ 1
where the edge values ci 12,j , u i 12,j , ci,j2 1 , and u
i,j2 1 are computed using projected
2 2 2 2
velocity fields, u, v, and a fifth order WENO algorithm [26].
14 Junseok Kim and John Lowengrub

3.2. Discretization of surface tension terms. In this section we describe the finite dif-
ference approximation to the surface tension term, Fst . Let
3 3
X X
Fst = (fk , gk ) = (|ck |2 I ck ck )
W ek W ek
k=1 k=1
3
X
= (x ck yy ck y ck xy ck , y ck xx ck x ck xy ck ).
W ek
k=1

Then, the surface tension force components (f1 , g1 ) are discretized as

1
(f1 )ij = (c1i+1,j c1i1,j )(c1i,j+1 2c1ij + c1i,j1 ) (3.5)
2h3
1
3 (c1 i,j+1 c1 i,j1 )(c1 i+1,j+1 + c1 i1,j1 c1 i+1,j1 c1 i1,j+1 )
8h
1
(g1 )ij = 3 (c1 i,j+1 c1 i,j1 )(c1 i+1,j 2c1 ij + c1 i1,j ) (3.6)
2h
1
3 (c1 i+1,j c1 i1,j )(c1 i+1,j+1 + c1 i1,j1 c1 i+1,j1 c1 i1,j+1 )
8h
and the other components, (f2 , g2 ) and (f3 , g3 ), are similarly defined by replacing c1 in Eqs.
n+ 21
(3.5) and (3.6) by c2 and 1 c1 c2 , respectively. Further, Fst in Eq. (3.3) is evaluated
1
using cn+ 2 = (cn + cn+1 )/2.
3.3. The discrete energy functional. The discrete energy functional for the Boussinesq
approximation is
3
2k
 
1 X 1 2
Eh (t) = (u, u)h + (Fk , 1)h + |ck |1 (3.7)
2 M k W ek 4
k=1
PNx PNy PNx PNy
where (f, g)h = h2 i=1 2
j=1 fij gij and |g|1 = i=1
2
j=1 ((gi+1,j gij ) + (gi,j+1
2
gij ) ) taking into account the Neumann boundary conditions to handle the boundary terms.
Following the analysis in [29], it can be shown that if the time step t is small enough, the
discrete energy Eh is non-increasing. In practice, we find that Eh is nonincreasing for t
such that our multigrid methods converge.
We note that an alternate discretization may be developed following our work on phase-
field models of two-phase flows [30]. In that work, we formulated a version of the projection
method using a special discretization of the Navier-Stokes equations such that the coupled
discrete Navier-Stokes-Cahn-Hilliard system had a non-increasing discrete energy functional
for any value of t. We find that the scheme presented here results in slightly better accuracy
than that used in [30].
4. Numerical experiments. In this section, we demonstrate convergence of our scheme
numerically and simulate a three-phase contact angle, a buoyancy-driven compound drop,
and liquid/liquid remediation enhanced by the Rayleigh-Taylor instability. Unless otherwise
specified, we use the free energy function, F (c) by Eq. (2.53). From our choice of the
homogeneous free energy density Eq. (2.53) and an equilibrium profile Eq. (2.51), the con-
centration field varies from 0.05 to 0.95 over a distance of about 4 2 tanh1 (0.9). We have
4-8 grid points across interface in most of our simulations. Note that since we are taking a
polynomial form of the free energy, F (c), we may have non-physical value of concentra-
tion (i.e. less than zero and greater than one), but its deviation from the physical values are
 Phase field modeling and simulation of three phase flows 15

negligible (most of cases, the deviations are order 103 ). The time step is determined by
restrictions due to the CFL condition, gravity, viscosity, and surface tension [40]:
s s !
(1 + 2 + 3 )W ek W erem(k,3)+1 3 3Ren h2 h 2h
t < min min h2 , , n , ,
k=1,2,3 8(W ek + W erem(k,3)+1 ) 14 n |u | |F n |

where F n is the right hand side of Eq. (2.55) at time t = nt.

4.1. Convergence test. To obtain an estimate of the rate of convergence, we perform a
number of simulations for a sample initial problem on a set of increasingly finer grids. The
initial data is
y 31
  
1
c1 (x, y, 0) = 1 tanh , (4.1)
2 2 2
y 13 y 32
    
1
c2 (x, y, 0) = tanh tanh . (4.2)
2 2 2 2 2
That is, narrow transition layers separate three immiscible fluids. The initial velocity is a
swirling flow

(u(x, y, 0), v(x, y, 0)) = (0.25 sin2 (x) sin(2y), 0.25 sin2 (y) sin(2x)) (4.3)

on a domain, = (0, 1) (0, 1). No-slip boundary conditions are applied to top and bottom
planes and periodic ones are to the side walls. The numerical solutions are computed on the
uniform grids, h = 1/2n for n = 5, 6, 7, 8, and 9. For each case, the calculations are run to
time T = 0.2, the uniform
time steps, t = 0.1h, Re = 10, W e1 = W e2 = W e3 = 100,
P e = 2, and = 0.005 2, are used to establish the convergence rates. Fig. 4.1(a) and (b)
show the initial configuration and a snapshot of the solution at time T = 0.2, respectively
with h = 1/64 and Re = 100.
TABLE 4.1
Convergence Results with Re = 10 u, v, c1 , and c2 .

Case 32-64 64-128 rate 128-256 rate 256-512 rate

u 5.4530e-4 4.4986e-5 3.5995 1.0895e-5 2.0458 2.7798e-6 1.9706

v 6.8462e-4 2.9858e-5 4.5191 7.5194e-6 1.9894 1.8850e-6 1.9961
c1 2.9211e-2 3.1969e-3 3.1918 7.5863e-4 2.0752 1.8755e-4 2.0161
c2 4.1640e-2 4.2759e-3 3.2836 1.0037e-3 2.0908 2.4801e-4 2.0169

Since a cell centered grid is used, we define the error to be the difference between that
grid and the average of the next finer grid cells covering it:
 
def
eh/ h = chij c h + ch + ch + ch /4.
2 ij 2 2i,2j 2 2i1,2j 2 2i,2j1 2 2i1,2j1

The rate of convergence is defined as the ratio of successive errors in the discrete l2 -norm:

log2 (||eh/ h ||/||e h / h ||).

2 2 4

The errors and rates of convergence with Re = 10 and Re = 100 are given in Tables
4.1 and 4.2, respectively. The results suggest that the scheme is indeed second order accurate
when Re number is small. When Re = 100, the rate of convergence decreases.
16 Junseok Kim and John Lowengrub

TABLE 4.2
Convergence Results with Re = 100 u, v, c1 , and c2 .

Case 32-64 64-128 rate 128-256 rate 256-512 rate

u 7.9061e-4 2.1666e-4 1.8675 6.3568e-5 1.7691 1.8657e-5 1.7686

v 1.8242e-4 4.4618e-5 2.0316 1.4104e-5 1.6615 4.4661e-6 1.6590
c1 1.6055e-2 3.1327e-3 2.3576 7.6571e-4 2.0325 1.9041e-4 2.0077
c2 2.5432e-2 4.4223e-3 2.5238 1.0388e-3 2.0899 2.5764e-4 2.0114

(a) (b)
F IG . 4.1. Snapshots of the solutions at (a) T = 0 and (b) T = 0.2. The uppermost layer is fluid 3, the middle
layer is fluid 2, and the lower layer is fluid 1. The contour lines are from c2 = 0.1 to c2 = 0.9 with 0.1 step size.

We also examined the time evolution of the total energy (with F and Fk defined as in
Eqs. (3.7) and (2.53) respectively). The results are shown in Fig. 4.2 together with the inter-
facial (dotted) and kinetic energy (dot-dashed) components. The interfacial energy contains
both the gradient and Helmholtz free energy terms. The solid curve marks the total energy.
The insets show the layer morphology (c2 = 0.5 filled contour) at various times. The initial
data, boundary conditions and all the parameters are the same as considered previously except
that the Weber numbers are W e1 = W e2 = W e3 = 10 so that surface tension plays a more
important role in the evolution. Here the Reynolds number Re = 10. Observe that the total
energy is non-increasing throughout the evolution. The kinetic energy also monotonically de-
creases due to viscous dissipation. The interfacial energy is non-monotone, however. At early
times, the initially flat layer deforms and the surface energy increases. Around approximately
t = 0.4, the surface energy reaches a peak due to the increased deformation of the layer and
then decreases towards an equilibrium value corresponding to a flat layer (two interfaces).
When the layer is most deformed, the surface tension force induces a reversal of the velocity
field such that the deformed layer relaxes back to a flat configuration at later times.

4.2. Contact angles. Following [41], we next investigate the spreading of a liquid lens
consisting of an initially circular immiscible droplet of fluid located at an interface between
two other immiscible fluids. See Fig 4.3(a).
The initial condition is a circular droplet, 2 , (located at a free surface between 1 and
 Phase field modeling and simulation of three phase flows 17

total energy
interfacial energy
kinetic energy
0.5

0.4

0.3

0.2

0.1

0 0.5 1 1.5 2 2.5 3

time

F IG . 4.2. The time dependent energies of the numerical solutions with the initial data Eqs. (4.1)-(4.6) and with
Re = 10, W e1 = W e2 = W e3 = 10. Snapshots of the concentration field, c2 are shown with filled contour at
c2 = 0.5.

3 ) and the initial velocity is zero, i.e.,

    
y 0.5
c1 (x, y, 0) = max 0.5 1 + tanh c2 (x, y), 0 , (4.4)
2 2
p !!
0.15 (x 0.5)2 + (y 0.5)2
c2 (x, y, 0) = 0.5 1 + tanh , (4.5)
2 2
u(x, y, 0) = v(x, y, 0) = 0. (4.6)

The computational domain is = [0, 1] [0, 1]. The fluids are density and viscosity matched
( = 1, = 1) and

Re = 60, and W e2 = 60, W e1 = W e3 = 108, 60, 36.

In Figs. 4.3(b)-(d), the evolution of the c2 = 1/2 contour line is shown for three cases
with Re = 60, and W e2 = 60, W e1 = W e3 = 108, 60, 36, respectively. In all cases,
= 0.005 2, P e = 100/, h = 1/256, and t = 0.25h. As the droplet spreads, it reaches
an equilibrium shape. The most deformed curve in each figure is the numerical steady-state.
Theoretically, the shape of the steady-state drop is controlled by the drop-volume and the
three surface tensions (inverse Weber numbers). The equilibrium three-phase contact angle
is determined by
sin 1 sin 2 sin 3
1 = =
W e2 + W1e3 1
W e1 + W1e3 1
W e1 + W1e2

and the relation between the lens area A, its length d (the distance between triple junctions),
and the contact angles i of the i-th phase (Youngs law) is
   12
1 2( 1 ) sin(2( 1 )) 2( 3 ) sin(2( 3 ))
d= + .
8A sin2 ( 1 ) sin2 ( 3 )
18 Junseok Kim and John Lowengrub

0.9

0.8
,
1 1

0.7
12

0.6

0.5
2, 2

0.4
13 13

0.3

23

0.2
,
3 3

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

(a) (b)

(c) (d)
F IG . 4.3. (a) Initial configuration: the upper fluid is phase 1, the lower fluid is phase 3, and the droplet is
phase 2. (b), (c), and (d) are evolutions of initial circular drop for W e2 = 60 and W e1 = W e3 = 108, 60, 36,
respectively. The arrow shows the direction of the evolution and the most deformed lines in (b), (c), and (d) are
corresponding steady shapes and the enclosing box size is [0.23, 0.77] [0.23, 0.77].

TABLE 4.3
Equilibrium measurements

W e1 , W e2 , W e3 dexact dnumerical dnumerical , /2

108, 60, 108 0.3746 0.3982

60, 60, 60 0.4138 0.4368
36, 60, 36 0.4578 0.4622 0.4502

Thus, the accuracy of the steady lens shape can be measured by comparing the observed d
with the analytical value.
The evolution of d for the three cases is shown in Fig. 4.4. The numerical value of a d
is obtained from the c2 = 1/2 contour line. Note that in all cases, there is rapid increase in
d at early times followed by a slow approach to equilibrium. Note also there is an overshoot
in the early evolution of the W e3 = 36 and W e3 = 108 cases. In table 4.2, the equilibrium
values of d are shown for the three cases together with the corresponding theoretical values.
There is very good agreement between the theory and simulation.
 Phase field modeling and simulation of three phase flows 19

0.5

d, distance between triple junctions

0.45

0.4

0.35

We3=108
We3= 60
We3= 36
0.3

0.25

0.2
0 10 20 30 40 50
time
F IG . 4.4. Time evolution of d, distance between triple junctions. W e3 = 108 (), W e3 = 60 (), and
W e3 = 36 ().

In Fig. 4.5(a) and (b), we plot the contours of the concentration c2 at the five levels
c2 = 0.1, 0.3, 0.5, 0.7, and0.9 for the case with W e1 = W e3 = 36, and W e2 = 60.
In the left column = 0.005 2 while in the right column = 0.0025 2. Not only do the
concentration contours appear to converge, also note that when = 0.0025 2, we obtain
dnumerical = 0.4502, which is closer to the exact result than is the case with = 0.005 2.
Interestingly, the c2 = 0.1 contour appears to be elongated near the triple junctions a fact
that seems to persist under refinement of .
We next show that numerical results are not sensitive to the choice of component labels.
We take the same initial configuration but change the component labels, i.e.

    
y 0.5
c1 (x, y, 0) = max 0.5 1 + tanh c3 (x, y), 0 , (4.7)
2 2
p !!
0.15 (x 0.5)2 + (y 0.5)2
c3 (x, y, 0) = 0.5 1 + tanh . (4.8)
2 2

We then solve the equations for the concentrations, c1 and c2 = 1 c1 c3 with W e1 =

W e2 = W e3 = 60. In Fig. 4.6, the initial concentration profiles are shown: for the original
case (hereafter referred to as case 1) - (a1) c1 and (a2) c2 , the new case (referred to as case 2)
(b1) c1 and (b2) c2 . The time evolution of the distance d between the triple junctions for the
two cases (case 1 () and case 2 ()) are shown in (c). As can be seen from this figure, the
values of d from these two cases are essentially identical and thus the results are insensitive
to label switching.
In Fig. 4.7, the time evolution of total energy (with F and Fk defined as in Eqs. (3.7) and
(2.53) respectively) is shown together with the corresponding interfacial and kinetic energy
components of the numerical solutions. In addition, the filled contour c2 = 0.5 is shown
at various times indicating the corresponding morphology of the drop. The initial data are
Eqs. (4.4)-(4.6) and the parameters correspond to the case in which W e1 = W e3 = 36,
and W e2 = 60. In the figure, the solid line is the total energy, the dotted line is interfacial
20 Junseok Kim and John Lowengrub

0.7 0.7

0.65 0.65
0.5
0.3
0.6 0.6

0.55 0.55
0.9 0.7 0.9
0.5

1
0.7 0.9
0.3 0.7

0.
0.5 0.5
0.3 0.5
0.1 0.1
0.45 0.45

0.4 0.4

0.35 0.35

0.3 0.3
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

(a) (b)
0.7 0.7

0.65 0.65
0.1
0.3
0.6 0.6
0.5 0.9
0.7
0.55 0.9 0.55

0.5 0.5

0.45 0.9 0.45

0.1
0.4 0.30.70. 0.4
5 0.7
0.5
0.3
0.1
0.35 0.35

0.3 0.3
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

(a) (b)
0.7 0.7

0.65 0.65

0.6 0.6

0.55 0.55
1

0.1
0.

0.3 0.3 0.1 0.3 0.1

0.5 0.5 0.5 0.5 0.5
0.7
0.9
0.7
0.9 00..7
9
0.45 0.45

0.4 0.4

0.35 0.35

0.3 0.3
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

(a) (b)

c2 = 0.1, 0.3, 0.5, 0.7, and

F IG . 4.5. Contour lines of concentration c2 at the five levels 0.9 in the case of
W e1 = W e3 = 36, and W e2 = 60. Left column: = 0.005 2 and right column: = 0.0025 2, respectively.

0.5
case 1
d, distance between triple junctions case 2

0.45

0.4

0.35
(a1) (a2)
0.3

0.25

0.2
0 0.5 1 1.5 2 2.5
time
(b1) (b2) (c)
F IG . 4.6. Initial concentration profiles: case 1 - (a1) c1 and (a2) c2 , case 2 (b1) c1 and (b2) c2 . Time
evolution of d, distance between triple junctions with two cases (case 1 () and case 2 ()) with W e3 = 60.

energy, and the dash-dotted line is the kinetic energy. The inset shows a blow-up of the
kinetic energy. Since the initial velocity is zero and the evolution is primarily surface energy
driven, the kinetic energy is very small. At early times, a small amount of surface energy is
transferred to kinetic energy as the drop begins to elongate. At later times, the kinetic energy
rapidly decays as the drop approaches its equilibrium configuration. As expected, the total
energy monotonically decreases to its equilibrium value. Correspondingly, the interfacial
 Phase field modeling and simulation of three phase flows 21

total energy
interfacial energy
0.1
kinetic energy

0.08

0.06

4
x 10
0.04 2

1
0.02

0
0 1 2 3

0 1 2 3 4 5
time

F IG . 4.7. The time dependent energies of the numerical solutions with the initial data Eqs. (4.4)-(4.6) and
the case with W e1 = W e3 = 36, and W e2 = 60. Snapshots of the concentration field, c2 are shown with filled
contour at c2 = 0.5.

energy also decreases.

Next, we consider a similar problem in three-dimensions. We place a periodic array
of spheres on an interface between two immiscible fluids analogously to that shown in Fig.
4.3(a) in two dimensions. The computational domain is = [0, 1] [0, 1] [0, 1] and the
mesh size is 64 64 64 with time step, t = 0.001. = 0.008 2, and P e = 100/.
No-slip boundary conditions for the top and bottom planes and periodic boundary conditions
for the side walls are applied. Specifically, the initial data in a single period box are
    
z 0.5
c1 (x, y, z) = max 0.5 1 + tanh c2 (x, y), 0 ,
2 2
p !!
0.35 (x 0.5)2 + (y 0.5)2 + (z 0.5)2
c2 (x, y, z) = 0.5 1 + tanh ,
2 2
u(x, y, z) = v(x, y, z) = w(x, y, z) = 0.

We take the viscosities and densities of the components to be matched (1 = 2 = 3 = 1)

and (1 = 2 = 3 = 1) with the following parameters.

Re = 60, W e1 = W e3 = 36, W e2 = 60.

In Fig. 4.8, evolution is shown and only the spheres are visualized with a reference plane
passing through equator of the spheres. The actual interface deforms. As the drops spread
out and flatten due to surface tension forces, the drops interact with their periodic neighbors.
In this case, the distance between the neighbors is less than the equilibrium length of an
isolated drop. As seen in Fig. 4.8, this leads to merger with the periodic images and results
in a lattice-like microstructure of the second fluid on the interface between the two other
immiscible fluids.
4.3. Numerical simulation of a buoyancy-driven compound drop. In this section, a
buoyancy-driven evolution a 3-D compound drop is investigated. In Fig. 4.9, a schematic of
22 Junseok Kim and John Lowengrub

t=0.0 t=0.2 t=0.6

t=1.8 t=2.6 t=3.2

F IG . 4.8. Evolution of spheres under surface tension forces, the nondimensional times are shown below each
figure.

the initial configuration is shown. The three fluids are immiscible where a heavy droplet of
fluid I is encapsulated by a light fluid II. Fluid I is the heaviest component. This models a
flow in which a heavy fluid contains a dispersed contaminant. Releasing drops of the light
fluid II from the bottom of the container, provides the means to encapsulate the contaminants.
Restricting this to a single drop yields the initial condition we consider.

II 2 , 2

23
III 3 , 3

II
12
I
1 , 1

23

F IG . 4.9. Schematic of a compound drop.

Specifically, the initial data are

  
0.3 r
c1 (x, y, z) = 0.5 1 + tanh ,
2 2
    
0.5 r z 2.5
c2 (x, y, z) = 0.5 2 + tanh + tanh c1 (x, y, z),
2 2 2 2
u(x, y, z) = v(x, y, z) = w(x, y, z) = 0,
p
where r = (x 1)2 + (y 1)2 + (z 1.5)2 . We take the viscosities and surface tensions
 Phase field modeling and simulation of three phase flows 23

of the components to be matched (1 = 2 = 3 = 1) and (1 = 2 = 3 = 1) with the

following parameters.

1 = 1.044, 2 = 0.957, 3 = 1, Re = 36, W e1 = W e2 = W e3 = 1127, and F r = 1.

The computational domain is = [0, 2] [0, 2] [0, 4] and the mesh size is 32 32 64
with time step, t = 0.002. = 0.01 2, and P e = 10/. No-slip boundary conditions for
the top and bottom planes and periodic boundary conditions for the side walls are applied.

t=0.0 t=6.0 t=8.0 t=12.0

t=14.0 t=17.0 t=20.0 t=44.0

F IG . 4.10. Evolution of a compound drop, the nondimensional times are shown below each figures.

The evolution is presented in Fig. 4.10. An upper (flat) interface separates the heavy
ambient from same fluid that encapsulates the heavy drop. The compound drop is lighter
than the heavy ambient and so it rises and deforms. The encapsulating fluid rises faster than
the heavy inner drop but nevertheless the compound drop remains intact until it penetrates the
upper interface. The heavy inner drop is carried upwards as the encapsulated fluid is released.
The drop then falls back on the interface remaining trapped there by surface tension forces
even though if is heavier than the lower ambient. At this point, the drop could be removed
from the system by sucking it off the interface. Imagining that the heavy inner drop is a
contaminant in the lower ambient, this provides a mechanism of liquid/liquid extraction by
which fluid III may be cleansed.
4.4. Rayleigh-Taylor Instability of Ternary Fluid Flows. In this section, we exploit
the fact that our ternary NSCH system is capable of describing multicomponent fluid flows
containing immiscible, miscible and partially miscible components. The miscibility of the
components is modeled through the properties of the free energy F (c1 , c2 ). It is nontrivial to
construct free energies capable of describing partially miscible systems where, for example,
two components are immiscible and the third component is preferentially miscible in one of
the immiscible components. Nevertheless, we have been able to construct a class of such a
24 Junseok Kim and John Lowengrub

C
2

C1 C
3

F IG . 4.11. Contour plot of the free energy F (c1 , c2 ) F IG . 4.12. schematic of initial configuration
on the Gibbs triangle.

free energies and one example of which is given below:

F (c1 , c2 ) = 2c21 (1 c1 c2 )2 + (c1 + 0.2)(c2 0.2)2 + (1.2 c1 c2 )(c2 0.4)2 .

A contour plot of the free energy F (c1 , c2 ) on the Gibbs triangle is shown in Fig. 4.11. The
two minima of F (c1 , c2 ) are at (0.7779, 0.2330, 0.0109) and (0.0151, 0.3651, 0.6499).
These minima lie very slightly outside the Gibbs triangle. As a demonstration of the evo-
lution possible in partially miscible liquid systems, we present an example in which there
is a gravity-driven (Rayleigh-Taylor) instability that enhances the transfer of a preferentially
miscible contaminant from one immiscible fluid to another in 2D. The initial configuration is
shown in Fig. 4.12. The top half of the domain consists of a mixture of fluid I and fluid II,
and the bottom half consists of fluid III, which is immiscible with fluid I. Fluid II is prefer-
entially miscible with fluid III. Fluid I is assumed to be the lightest and fluid II the heaviest.
The density of the I/II mixture is heavier than that of fluid III, so the density gradient induces
Rayleigh-Taylor Instability.
In particular, the initial data are
  
y 0.5 0.1 cos(2x)
c1 (x, y) = c2 (x, y) = 0.25 1 + tanh
2 2
u(x, y) = v(x, y) = 0

and the simulation parameters are

1 = 1, 2 = 4, 3 = 2, F r = 1, Re = 313, W e1 = W e2 = W e3 = 9.8 106 .

The computational domain is = [0, 1] [0, 1] and the mesh size is 128 128 with time
step, t = 0.5h. = 0.01 2 and P e = 10/. Here, we assume that the surface tension is
small so that the interface undergoes the Rayleigh-Taylor instability (e.g. see paper [16] for
the effect of surface tension on the instability).
The evolution of the three phases is shown in Fig. 4.13. The top row shows the evolution
of fluid I, middle and bottom correspond to fluid II and fluid III, respectively. That is, the
contours of c1 , c2 , and c3 are visualized in gray-scale where darker regions denote larger
values of c1 , c2 , and c3 , respectively. As the simulation begins, the I/II mixture falls and
fluid II diffuses into fluid III. A characteristic Rayleigh-Taylor (inverted) mushroom forms,
the surface area of the I/III interface increases, and vorticity is generated and shed into the
 Phase field modeling and simulation of three phase flows 25

F IG . 4.13. Evolution of concentration of fluid I (top row), II (middle row), and III (bottom row). The con-
tours of c1 , c2 , and c3 are visualized in gray-scale where darker regions denote larger values of c1 , c2 , and c3 ,
respectively. Nondimensional times are t = 0, 3.91, 7.81, 15.63, and 195.31.

bulk. As fluid II is diffused from fluid I, the pure fluid I rises to the top as shown in Fig. 4.13.
Imagining that fluid II is a contaminant in fluid I, this configuration provides an efficient
means of cleansing fluid I since the buoyancy-driven flow enhances the diffusional transfer
of fluid II from fluid I to fluid III.
4.5. Example of Adaptive Mesh Refinement - binary spinodal decomposition. In
the typical setting, solutions to the Cahn-Hilliard equation are nearly constant in the so called
bulk regions. Between the bulk regions, solutions exhibit thin transition layers, through
which the solution takes on values intermediate to those in the bulk regions. Typically, the
bulk regions comprise the largest portion, by far, of the computational domain. Since in most
of the applications it is sufficient to finely resolve only the transition layers, fixed grid meshing
represents a waste of computational resources. Thus, efficient, adaptive-mesh solvers for the
CH equation, able to finely resolve only the diffuse interface, are highly desirable.
In preliminary work [32], we are currently adapting block-structured adaptive methods
originally designed by Berger and Oliger [9] for conservation laws and general hyperbolic
equations. The block-structured framework was adopted in the CHOMBO package [3], which
we utilize. However, we do not use the multi-level multigrid algorithm from CHOMBO (see
[3]), but rather we employ a nonlinear adaptive multigrid solver [44] following our work here
and in [30, 31].
To illustrate the potential of this approach, we present preliminary results on adaptive
mesh refinement for the binary CH equation from [32]. The governing equations are
c
= , (4.9)
t
= c3 1.5c2 + 0.5c 2 c. (4.10)

For the following test, we consider spinodal decomposition using the off-critical aver-
age composition c = 0.3. The results are shown in Fig. 4.14. The interface parameter is
= 0.0025. The initial condition has the composition being nearly uniform with random
26 Junseok Kim and John Lowengrub

(a) t=2 (b) t=5

(c) t=20 (d) t=25

F IG . 4.14. Evolution of the concentration c(x, y, t) with an average concentration cave = 0.3. The times are
shown below each figure.

perturbations of maximum magnitude 1 103 . The spatial domain is 1 1 and periodic

boundary conditions are applied to both directions. The overall coarse computational grid is
64 64 and there are two levels of adaptivity. The AMR libraries automatically generate
a grid structure that adapts to the locations of the phase interfaces. The effective fine grid
resolution is 256 256. As we would expect from Cahns theory [11], in Fig. 4.14 isolated
particles form relatively quickly as the two phase regions form, associated with c = 1 and
c = 0, respectively. Over a much longer time scale the system coarsens, and the mesh adapts
around the isolated particles. More detailed studies will be presented in a forthcoming work
[32].

5. Conclusion. In this paper we have presented a general model of three-phase flows

and developed an associated efficient, second-order accurate finite difference method to solve
the model equations numerically. The three-liquid phases may be fully miscible, partially
 Phase field modeling and simulation of three phase flows 27

miscible or immiscible. The miscibility of the phases is modeled thermodynamically through

the Helmholtz free energy. An additional advantage of the model is that triple interfaces are
handled without resorting to ad-hoc procedures.
We presented examples of flow with miscible and immiscible components. We demon-
strated the convergence of our algorithm through a resolution study. In addition, we found
good agreement with the theory for an equilibrium liquid lens (lying atop an interface). We
provided demonstrations of liquid/liquid remediation. In the first example, a compound drop
was simulated, in which a light fluid encapsulated a heavy contaminant drop. The light fluid
causes the compound drop to rise and deposit the contaminant at an interface where it may be
removed. In the second example, we investigated the diffusional transfer of a preferentially
miscible contaminant from one immiscible phase to another. The transfer is enhanced by the
flow and in particular the Rayleigh-Taylor instability.
In future work, we will perform more extensive studies of liquid/liquid remediation. In
addition, we will investigate the complex morphologies generated from the application of
chaotic mixing flow to three-phase dispersion. Under appropriate conditions, a coalescence
cascade ensures and the three-phases may become interpenetrating and continuous. We will
also investigate the limit in which one of the phases lies on the boundary between the other
phases, thus mimicking a surfactant. To perform these studies, we will incorporate adaptive
block-structured mesh refinement techniques [32].

Acknowledgments. The authors acknowledge the support of the Department of Energy,

Office of Basic Energy Sciences and the National Science Foundation, Division of Mathemat-
ical Sciences. The authors are also grateful for the support of the Minnesota Supercomputer
Institute and the Network & Academic Computing Services (NACS) at UCI.

6. Appendix A. In this appendix, we present an alternative set of constitutive assump-

tions based on an equivalent form of the entropy expression in Eq. (2.21) different from that
in Eq. (2.23). Rather than taking quantities relative to the contributions from the Nf com-
ponent, as done in Eq. (2.23), we instead may consider quantities relative to the mean of the
PNf
components. Accordingly, using that k=1 Jk = 0, re-write Eq. (2.21) as

Nf Nf
1 X F 1 X F
= P + pI + ck : u
T ck Nf j=1 cj
k=1

Nf Nf
1 X
tk F 1
X F
1 Jk


+
T ck Nf j=1 cj
k=1
 PNf 
Nf Nf Nf 1
1 X 1 X X k Nf j=1 j Jk
+ k j Jk + J ,
T Nf j=1 T
k=1 k=1

(6.1)

PNf  PNf 
Taking the diffusion flux J = k=1 k N1f j=1 j Jk /T we are now in a position
to pose thermodynamically consistent constitutive relations for P, tk and Jk . Arguing as in
28 Junseok Kim and John Lowengrub

section 2.1, we obtain the constitutive assumptions

Nf Nf  
X F 1 X F 2
P = pI ck + D ( u) I , (6.2)
ck Nf j=1 cj 3
k=1

Nf
F 1 X F
tk = , (6.3)
ck Nf j=1 cj

Nf
1 X
Jk = k k j , for k, . . . , Nf . (6.4)
Nf j=1

Using the form of the free energy from section 2.4, we obtain
Nf 2Nf
1 X F 1 X 2
= cj . (6.5)
Nf j=1 cj Nf j=1 j

Note that if j = , then the sum in Eq. (6.5) vanishes. Therefore the extra fluid stess induced
by concentration gradients is the same as obtained in section 2.4). However, the chemical
potential

Nf Nf Nf
1 X F 1 X F 1 X

k = k j = + k N p
Nf j=1 ck Nf j=1 cj Nf j=1 f

Nf
1 1 X
2k ck 2 cj , (6.6)
2 Nf j=1 j

is slightly different than that obtained in section 2.4. Thus, even though Eq. (6.1) is equivalent
to Eqs. (2.21) and (2.23), somewhat different constitutive relations may be taken.

REFERENCES

[1] Almgren A.S., Bell J.B., and Szymczak W.G., A numerical method for the incompressible Navier-Stokes
equations based on an approximate projection, SIAM J. Sci. Comput., 17 (1996) 358-369.
[2] Anderson D., McFadden G.B., and Wheeler A.A., Diffuse interface methods in fluid mechanics, Ann. Rev.
Fluid Mech., 30 (1998) 139-165.
[3] Applied numerical algorithms group, The chombo framework for block-structured adaptive mesh refinement.
Technical report, Lawrence Berkeley National Laboratory, http://seesar.lbl.gov/Chombo.
[4] Barrett J.W. and Blowey J.F., An improved error bound for a finite element approximation of a model for
phase separation of a multi-component alloy, IMA J. Numer. Anal., 19 (1999) 147-168.
[5] Barrett J.W. and Blowey J.F., and Garcke H., On fully practical finite element approximations of degenerate
Cahn-Hilliard systems, M2AN Math. Model. Numer. Anal., 35 (2001) 713-748.
[6] Barrett J.W. and Blowey J.F., An error bound for the finite element approximation of a model for phase
separation of a multi-component alloy, IMA J. Numer. Anal., 16 (1996) 257-287.
[7] Barrett J.W. and Blowey J.F., Finite element approximation of a model for phase separation of a multi-
component alloy with nonsmooth free energy and a concentration dependent mobility matrix, Math.
Models Methods Appl. Sci., 9 (1999) 627-663.
[8] Bell J., Collela P., and Glaz H., A second-order projection method for the incompressible Navier-Stokes
equations, J. Comp. Phys., 85 (1989) 257-283.
[9] Berger M.J. and Oliger J., Adaptive mesh refinement for hyperbolic partial differential equations, J. Comput.
Phys., 53 (1984) 484-512.
[10] Blowey J.F., Copetti M.I.M., and Elliott C.M., The numerical analysis of a model for phase separation of a
multi-component alloy, IMA J. Numer. Anal., 16 (1996) 111-139.
 Phase field modeling and simulation of three phase flows 29

[11] Cahn J.W., On spinodal decomposition., Acta Metall, 9 (1961) 795-801.

[12] Cahn J.W. and Hilliard J.E., Free energy of a non-uniform system. I. Interfacial free energy, J. Chem. Phys.,
28 (1958) 258-267.
[13] Ceniceros H.D., The effects of surfactants on the formation and evolution of capillary waves, Phys. Fluids, 15
(2003) 245-256.
[14] Coleman B.D. and Noll W., Thermodynamics of elastic materials with conduction and viscosity, Arch. Rat.
Mech. Anal., 13 (1963) 167-178.
[15] Copetti M., Numerical experiments of phase separation in ternary mixtures, Math. Comput. Simulation, 52
(2000) 41-51.
[16] Daly B.J., Numerical study of the effect of surface tension on interface instability, Phys. Fluids, 12 (1969)
1340-1354.
[17] Eggleton C.D., Tsai T.-M. and Stebe K.J., Tip streaming from a drop in the presence of surfactants, Phys. Rev.
Lett., 87 (2001) 048302.
[18] Eyre D.J., Systems for Cahn-Hilliard equations, SIAM J. Appl. Math., 53 (1993) 1686-1712.
[19] Garcke H., Nestler B., and Stinner B., A diffuse interface model for alloys with multiple components and
phases, SIAM J. Appl. Math., 64 (2004) 775-799
[20] Garcke H., Nestler B., and Stoth B., A multiphase field concept: numerical simulations of moving phase
boundaries and multiple junctions, SIAM J. Appl. Math., 60 (2000) 295315.
[21] De Groot S.R. and Mazur P., Non-equilibrium thermodynamics, Dover, 1984 (originally published by North-
Holland, Amsterdam, 1962).
[22] Drumright-Clarke M.A. and Renardy Y., The effect of insoluble surfactant at dilute concentration on drop
breakup under shear with inertia, Phys. Fluids, 16 (2004) 14-21.
[23] Hohenberg P.C. and Halperin B.I., Theory of dynamic critical phenomena, Rev. Mod. Phys., 49 (1977) 435-
479.
[24] James A.J. and Lowengrub J., A surfactant-conserving volume-of-fluid method for interfacial flows with
insoluble surfactant, J. Comp. Phys., 201 (2004) 685-722.
[25] Jan Y.-J. and Tryggvason G., Computational studies of contaminated bubbles, in: I. Sahin, G. Tryggvason
(Eds.), Proceedings of a Symposium on the Dynamics of Bubbles and Vorticity near Free Surfaces,
ASME, 119 (1991) 46-54.
[26] Jiang G.S. and Shu C.-W., Efficient implementation of weighted ENO schemes, J. Comput. Phys., 126 (1996)
202-228.
[27] Johnson R.A. and Borhan A., Stability of the shape of a surfactant-laden drop translating at low Reynolds
number, Phys. Fluids, 12 (2000) 773-784.
[28] Kan H.C., Udaykumar H.S., Shyy W., and Tran-Son-Tay R., Hydrodynamics of a compound drop with appli-
cation to leukocyte modeling, Phys. Fluids, 10 (1998) 760-774.
[29] Kim J.S., Ph.D. Thesis, University of Minnesota, 2002.
[30] Kim J.S., Kang K., and Lowengrub J.S., Conservative multigrid methods for Cahn-Hilliard fluids, J. Comput.
Phys., 193 (2004) 511-543.
[31] Kim J.S., Kang K., and Lowengrub J.S., Conservative multigrid methods for ternary Cahn-Hilliard systems,
Comm. Math. Sci., 2 (2004) 53-77.
[32] Kim J.S., Wise S.M., and Lowengrub J.S., Solving Cahn-Hilliard equation using adaptive mesh refinement,
(in preparation).
[33] Li X. and Pozrikidis C., The effect of surfactants on drop deformation and on the rheology of dilute emulsions
in Stokes flow, J. Fluid Mech., 341 (1997) 165-194.
[34] Lowengrub J.S. and Truskinovsky L., Quasi-incompressible Cahn-Hilliard fluids and topological transitions,
R. Soc. Lond. Proc. Ser. A Math. Phys. Eng. Sci., 454 (1998) 2617-2654.
[35] Milliken W.J. and Leal L.G., The influence of surfactant on the deformation and breakup of a viscous drop:
the effect of surfacant solubility, J. Coll. Int. Sci., 166 (1994) 879-902.
[36] Morral J.E. and Cahn J.W., Spinodal Decomposition in Ternary Systems, Acta Met., 19 (1971) 1037-1067.
[37] Peskin C.S., Numerical analysis of blood flow in the heart, J. Comput. Phys., 25 (1977) 220-252.
[38] Porter D.A. and Easterling K.E., Phase Transformations in Metals and Alloys, van Nostrand Reinhold, (1993).
[39] Renardy Y., Renardy M., and Cristini V., A new volume-of-fluid formulation for surfactants and simulations
of drop deformation under shear at a low viscosity ratio, Eur. J. Mech. B, 21 (2002) 49-59.
[40] Sussman M., Almgren A.S., Bell J.B., Colella P., Howell L.H., and Welcome M.L., An adaptive level set
approach for incompressible two-phase flows, J. Comput. Phys., 148 (1999) 81-124.
[41] Smith K.A., Solis F.J., and Chopp D.L., A projection method for motion of triple junctions by levels sets,
Interfaces Free Bound., 4 (2002) 263-276.
[42] Stone H.A. and Leal L.G., The effects of surfactants on drop deformation and breakup, J. Fluid Mech., 222
(1990) 161-186.
[43] Sussman M. and Puckett G.E., A coupled level set and volume-of-fluid method for computing 3D and ax-
isymmetric incompressible two-phase flows, J. Comput. Phys., 162 (2000) 301-337.
[44] Trottenberg U., Oosterlee C., and Schuller A., Multigrid, Academic Press, London, (2001).
30 Junseok Kim and John Lowengrub

[45] Truesdell C. and Noll W., The nonlinear field theories of mechanics, In The Encyclopedia of Physics. Ed. (S.
Flugge), vol III. Springer-Verlag. (1965).