Journal of Computational Physics 319 (2016) 79–107

Contents lists available at ScienceDirect

Journal of Computational Physics

www.elsevier.com/locate/jcp

Phase field modeling and simulation of three-phase flow on

solid surfaces
Qian Zhang 1 , Xiao-Ping Wang ∗
Department of Mathematics, The Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong, China

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

Article history: Phase field models are widely used to describe the two-phase system. The evolution of
Received 19 April 2015 the phase field variables is usually driven by the gradient flow of a total free energy
Received in revised form 5 February 2016 functional. The generalization of the approach to an N phase (N ≥ 3) system requires
Accepted 6 May 2016
some extra consistency conditions on the free energy functional in order for the model
Available online 12 May 2016
to give physically relevant results. A projection approach is proposed for the derivation
Keywords: of a consistent free energy functional for the three-phase Cahn–Hilliard equations. The
Three-phase flow system is then coupled with the Navier–Stokes equations to describe the three-phase flow
Moving contact line problem on solid surfaces with moving contact line. An energy stable scheme is developed for the
Gradient projection three-phase flow system. The discrete energy law of the numerical scheme is proved which
Energy law ensures the stability of the scheme. We also show some numerical results for the dynamics
Cahn–Hilliard coupled Navier Stokes of triple junctions and four phase contact lines.
equation
© 2016 Elsevier Inc. All rights reserved.
Convex splitting
Pressure stabilization scheme

1. Introduction

Phase field method has been widely used for numerical simulations of two-phase systems. The basic idea of the phase
field method is to introduce an order parameter φ which takes two distinct values (+1 and −1, for instance) in each of
the phases, with a smooth transition between both values in the zone around the interface, which is then diffuse with a
finite width. The key advantage of the method is that the need to explicitly track the interface between different phases is
removed. The phase field concept can be generalized to systems with three phases. In such three-phase system, the scalar
variable φ is replaced by a vector c where the i-th element c i represents the volume fraction of i-th phase. These phases
are linked through the constraint

c 1 + c 2 + c 3 = 1. (1.1)
The evolution of the three-phase system is then driven by the gradient of a total free energy of the system subject to the
constraint Eq. (1.1). If the free energy of the solid boundary is taken into account, the total free energy is then a sum of
three terms: a bulk free energy which is usually taken as a multi-well potential describing the free energy density of the
bulk of each phase, an interface energy term depending on the gradient of c and a surface energy on the solid boundary.
Considering a three-phase system in domain  with a solid boundary ∂, the total free energy functional can be written as

* Corresponding author.
E-mail addresses: qzhangxd@ust.hk (Q. Zhang), mawang@ust.hk (X.-P. Wang).
1
Current address: Department of Engineering Sciences and Applied Mathematics, Northwestern University, 2145 Sheridan Road, Evanston, IL 60208-3125,
USA.

http://dx.doi.org/10.1016/j.jcp.2016.05.016
0021-9991/© 2016 Elsevier Inc. All rights reserved.
80 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107

 
F3 (c) = F (c)dx + γ3 (c)ds, where F (c) = F 3 (c) + G 3 (c, ∇ c).
 ∂

Here, G 3 (c, ∇ c) is the interfacial free energy density, and F 3 (c) and γ3 (c) are the free energy densities in the bulk and on
the solid boundary, respectively. The bulk free energy F 3 (c) should be a “triple-well” potential with minima at c i = 1 for all
i ∈ {1, 2, 3}. It is natural to require that the three-phase free energy functional is consistent with that of two-phase flow, i.e.,
given i ∈ {1, 2, 3}, if c i = 0 and j = i , k = i, then

F 3 (c j , ck , 0) = F 2 (c j , ck ), G 3 (c j , ck , 0) = G 2 (c j , ck ), γ3 (c j , ck , 0) = γ2 (c j , ck ),
where F 2 , G 2 and γ2 are the corresponding free energy densities in two-phase flow on the solid surfaces. If consider the
time evolution of the three-phase system driven by the minimization of the free energy functional, the dynamic equations
are

∂ ci
= −μi (1.2)
∂t
with a boundary condition on ∂

∂ ci
= −L i (1.3)
∂t
for i = 1, 2, 3, where μi and L i are defined by the variation of the total free energy functional F3 (c), that is
 
δF3 (c) = dx[μi δ c i ] + ds[ L i δ c i ]. (1.4)
 ∂

For (1.2) and (1.3), the following extra consistency conditions (P.1) and (P.2) are also needed in order for the system to give
physically relevant results, see also [3,9],

P.1 When the phase i does not present in the mixture at the initial time, the phase should not appear during the time
evolution of the system, i.e. given i ∈ [1, 3],

c i (0) = 0 =⇒ c i (t ) = 0, ∀t ≥ 0. (1.5)

P.2 When there are only two phases in the model, the model of three-phase problem should always degenerate into the
corresponding two-phase model.

Boyer et al. [3] shows that, for the three-phase problem, one must carefully choose the form of the bulk free energy
so that the model can be well-posed and satisfy the algebraically and dynamically consistency conditions. However, their
method is not easy to be generalized to general N-phase problems with N > 3 and to problems with general boundary
conditions such as (1.3).
When the multiphase fluid dynamics is considered, the multiphase system is then coupled with the Navier–Stokes equa-
tions. For two-phase flow on solid boundary, the generalized Navier boundary condition is proposed for the moving contact
line problem [12]. The two-phase flow model has been studied extensively [13,14] with many efficient numerical meth-
ods developed [6,8]. There are, however, very few results on the three-phase (or components) flow with solid boundary
modeling and simulations [3,11,16,17].
In this paper, a new approach is developed to derive the consistent free energy functional for three-phase flow with
moving contact line problem. The idea of the approach is to introduce a projection operator to enforce constraints Eq. (1.1),
(P.1) and (P.2). We then employ the operator to derive the three-phase Cahn–Hilliard Navier–Stokes equations with gener-
alized Navier boundary condition. In addition, we develop an energy stable scheme for the coupled Cahn–Hilliard Navier
Stokes equations with boundary conditions on solid surface. We also show that the scheme has the total energy decaying
property which ensures the stability of the scheme. Numerical examples are carried out to verify capability of our model
and numerical scheme.
The rest of this paper is organized as follows. In Section 2, the consistent bulk free energy functional and surface energy
functional is derived from the projection approach. The energy functional is then incorporated into the Cahn–Hilliard Navier
Stokes system for three-phase flow with the generalized Navier boundary condition for the motion of the contact line. In
Section 3, the energy law for the model of three-phase flows on solid surfaces is proved. The numerical scheme is developed
and discrete energy laws of the numerical scheme is discussed in Section 4. In Section 5, we give some numerical examples
for dynamics of triple junctions and four phase contact lines (points). The discrete energy law and mass conservation law is
verified through numerical examples. The paper concludes in Section 6.
 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107 81

Fig. 1. The intersection of the fluid 1–fluid 2 interface with the solid surface.

2. Modeling of three-phase flow on solid surfaces

2.1. The two-phase Cahn–Hilliard model

We give a brief overview of the Cahn–Hilliard model for two-phase system on a solid surface. Denoting the domain by
 and the boundary by ∂, we write the free energy of the system as
 
σ 2 3
F(c ) = 12 c (1 − c )2 + σ  | ∇ c |2 dx + γ w (c )ds. (2.1)
 4
 ∂

The free energy density on the solid boundary is given by

σ w1 + σ w2 σ w2 − σ w1
γ w (c ) = + (6c 2 − 4c 3 − 1) (2.2)
2 2
where  represents the interface thickness and κ is the surface tension between fluid 1 and fluid 2. In γ w (c ), γ w (0) and
γ w (1) give the fluid–solid interfacial tensions σ w1 and σ w2 for the two fluids respectively. The evolution of the system is
driven by minimization of the total free energy
⎧ ∂ δF 3 σ
⎪
⎨ ∂ t c = M δ c = M (− 2 σ  c + 24  c (1 − c )(1 − 2c )),
⎪ in 
(2.3)
⎪
⎪ ∂ 3 ∂ c ∂ γw
⎩ c = − ( σ + ), on ∂.
∂t 2 ∂n ∂c
The parameters Ld = M σ and νs = σ determine the relaxation time of the interface in the bulk and on the boundary,
respectively. The equilibrium interface structure is obtained by minimizing the total free energy functional F(c ), which
results in the following equation,
⎧ 3 σ
⎪
⎨ − 2 σ  c + 24  c (1 − c )(1 − 2c ) = 0,
⎪ in 
(2.4)
⎪
⎪ 3 ∂ c ∂ γw
⎩ σ + = 0, on ∂.
2 ∂n ∂c
For simplicity, considering two-dimensional case, we denote m as the unit normal vector to the fluid 1–fluid 2 interface
and n as the unit normal vector to the solid surface. Let m and n be the coordinates along the corresponding directions.
When the interface thickness is small, it is reasonable to assume that the phase function c is a one-dimensional function
in the direction m and c does not change in the direction parallel to the interface. When the interface of the two fluids is
exactly a planar surface, we could compute c 0 (m), which depends only on m, explicitly by solving first equation of Eq. (2.4)
under some boundary conditions, see Fig. 1.
That is,
⎧
⎪
⎪ 3 d2 c 0 σ
⎨ − σ 2
+ 24 c 0 (1 − c 0 )(1 − 2c 0 ) = 0,
2 dm  (2.5)
⎪
⎪
⎩ lim c 0 = 1, lim c 0 = 0,
m→+∞ m→−∞

whose solution is, up to a translation,

82 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107

1 2m 
c 0 (m) = 1 + tanh( ) . (2.6)
2 
Thus, in the equilibrium state, the total energy contained in the interfacial shape is

σ 2 3 dc 0 2 1 1
12 c 0 (1 − c 0 )2 + σ  dm = σ + σ = σ. (2.7)
 4 dm 2 2
R
Multiplying both sides of second equation in Eq. (2.4) by ∂ c /∂ x, and integrating across the fluid–fluid interface along the
solid boundary, we obtain,
+∞
3 ∂ c ∂ γw ∂ c
σ + dx = 0. (2.8)
2 ∂n ∂c ∂x
−∞

Notice that
+∞ 1
∂ γw ∂ c ∂ γw
dx = dc = γ w (1) − γ w (0) = σ w2 − σ w1 (2.9)
∂c ∂x ∂c
−∞ 0

and under the condition c (m) = c (x) with m = x sin θs , we have

+∞ +∞ +∞ 2
3 ∂c ∂c 3 ∂c ∂c 3 ∂c
σ dx = cos θs σ dx = cos θs σ dm. (2.10)
2 ∂n ∂ x 2 ∂m ∂ x 2 ∂m
−∞ −∞ −∞

Therefore, Eq. (2.8) implies the Young’s Equation, where θs is the static contact angle between the interface and the solid
boundary. In particular, for exactly planar surface, we have
+∞ 2
3 dc 0
cos θs σ dm = σ cos θs (2.11)
2 dm
−∞

which implies

σ cos θs = σ w1 − σ w2 . (2.12)

2.2. Enforcing the consistency conditions for a simple three-phase mode

The generalization of the two-phase model Eq. (2.4) to a three-phase model requires some consistency conditions. We
illustrate the issue using a simple example. Consider a two-phase free energy functional without the interfacial energy,

F2 (c) = F 2 (c 1 , c 2 )dx, where F 2 (c 1 , c 2 ) = σ12 c 12 c 22 , c1 + c2 = 1 (2.13)

and its generalization to a three-phase free energy functional

F̄3 (c) = F 3 (c 1 , c 2 , c 3 )dx, where F 3 (c 1 , c 2 , c 3 ) = σ12 c 12 c 22 + σ13 c 12 c 32 + σ23 c 22 c 32 , c 1 + c 2 + c 3 = 1. (2.14)

Consider a simple gradient flow for F2 , we substitute c 2 = 1 − c 1 into Eq. (2.13) and obtain

dc 1 δF2
= −M = − M (∂1 F 2 − ∂2 F 2 ) = −2M σ12 c 1 (1 − c 1 )(1 − 2c 1 ) (2.15)
dt δc1
or, equivalently

dc 2 δF2
= −M = − M (∂2 F 2 − ∂1 F 2 ) = −2M σ12 c 2 (1 − c 2 )(1 − 2c 2 ). (2.16)
dt δc2
When c 1 = 0 (or c 2 = 0), we have dc 1 /dt = 0 (or dc 2 /dt = 0) and (P.1) is automatically satisfied. However, the situation
is much more complicated for the three-phase functional. The contour plot of F 3 is displayed in Fig. 2 for σ12 = σ13 =
σ23 = 1, where points A (0, 0), B (1, 0), C (0, 1) are the three minimum points of F 3 . For constraint (P.1) to be satisfied,
we need to require that the straight line segments A B , BC , AC be invariant under the gradient flow. However, it is easy
to see from the contour plot that this is not true. The gradient directions on the straight line segments A B , BC , AC are
 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107 83

Fig. 2. The contour map of F 3 (c) = c 12 c 22 + c 12 c 32 + c 22 c 32 .

clearly pointing away from the line. The equations of the gradient flow are derived by substituting c 3 = 1 − c 1 − c 2 into
Eq. (2.14),

dc 1
= − M (∂1 F 3 − ∂3 F 3 ), (2.17)
dt
dc 2
= − M (∂2 F 3 − ∂3 F 3 ), (2.18)
dt
dc 3
= M (∂1 F 3 + ∂2 F 3 − 2∂3 F 3 ) (2.19)
dt
where ∂1 F 3 = 2c 1 (σ12 c 22 + σ13 c 32 ), ∂2 F 3 = 2c 2 (σ12 c 12 + σ23 c 32 ) and ∂3 F 3 = 2c 3 (σ13 c 12 + σ23 c 22 ). Direct calculations show that
dc 3 /dt = 2σ12 c 1 c 2 which is non-zero unless c 1 = 0 or c 2 = 0, when c 3 = 0. Therefore, BC is not invariant under the flow
and (P.1) is not satisfied.
Notice that the constraint (P.1) reveals the essential difference between two-phase (N = 2) problem and N-phase problem
(N > 2). For the two-phase problem, (P.1) is automatically satisfied. c 1 (0) = 0 implies c 2 (0) = 1. Since (0, 1) and (1, 0) are
two minimizers of the free energy functional. We have c 1 (t ) ≡ 0 and c 2 (t ) ≡ 1. However, for the three-phase problem, when
c 1 (0) = 0, (0, c 2 (0), c 3 (0)) may not be at one of the minimizers ((0, 0, 1) or (0, 1, 0)). Therefore, extra condition is needed
to ensure c 1 (t ) = 0 for all the time.

2.2.1. Gradient projection method

It is shown in [3] that one can modify the free energy functional in Eq. (2.14) so that (P.1) is satisfied. However, the
approach is ad-hoc and not easy to be generalized to N-phase (N > 3). Instead of modifying the free energy functional, our
approach is to modify the gradient flow of F̄3 first in Eq. (2.14) so that it enforces the constraints. The modified free energy
functional is then derived through projection. The procedure consists of the following three steps:
Step 1: Enforcing the constraint c 1 + c 2 + c 3 = 1.
Considering the evolution of the phases driven by energy minimization,

∂c ∂ F3 ∂ F3 ∂ F3 ∂ F3 ∂ F3
=− , where =( , , ). (2.20)
∂t ∂c ∂c ∂ c1 ∂ c2 ∂ c3
Let S = c 1 + c 2 + c 3 . In order to maintain S = 1 in time, a projection P 3 is introduced to project the vector ∂ F 3 /∂ c onto the
hypersurface

3
W = x = (x1 , x2 , x3 ) ∈ R3 | xi = 0 . (2.21)
i =1

Thus, after the projection, the continuous model for ensuring Eq. (1.1) is expressed as

∂c ∂ F3 1
= −P3 where P 3 = I − n1 n1T , n1 = √ (1, 1, 1) T ∈ R3 . (2.22)
∂t ∂c 3
84 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107

It is easy to check solution of Eq. (2.22) satisfies dS /dt = 0.

Step 2: Enforcing the constraint (P.1)
For solution of Eq. (2.22) to also satisfy (P.1), we require that, if c i (0) = 0 (i ∈ {1, 2, 3}), then c i (t ) = 0, for all t ≥ 0.
A projection Q 3i is applied so that ∂ F 3 /∂ c is projected onto the intersection line of two hypersurfaces

3 
Li = {x ∈ R3 | x i = 0} {x ∈ R3 | xi = 0}. (2.23)
i =1

We then have
∂c ∂ F3
= − Q 3i P 3 , where Q 3i = I − n2i (n2i ) T . (2.24)
∂t ∂c
√
Here n2i = (a1 , a2 , a3 )/ 6, ai = −2 and a j = 1 for all j = i.
Step 3: Enforcing the constraint (P.2)
According to the constraint (P.2), the degenerated model should be consistent with the two-phase model. Because Li is
a line, a stretching transformation R 3i applied to the vector − Q 3i P 3 (∂ F 3 /∂ c) is enough. With projection R 3i = 2I M kj , where
M kj is the mobility for system of phase k and phase j and k, j = i, the phase evolution equation system is

∂c ∂ F3
= −Wi3 , where W 3i = Ri3 Qi3 P3 (2.25)
∂t ∂c
It is easy to see that
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
0 0 0 1 0 −1 1 −1 0
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
W 31 = M 23 ⎝ 0 1 −1 ⎠ , W 32 = M 13 ⎝ 0 0 0 ⎠, W 33 = M 12 ⎝ −1 1 0⎠ .
0 −1 1 −1 0 1 0 0 0

To derive a uniform model for all i ∈ {1, 2, 3}, we introduce f i j (c) such that
⎧
⎨1 for c i + c j = 1, 0 < c i < 1, 0 < c j < 1, j = i
f i j (c) = 0 for c i = 0, or c j = 0, j = i (2.26)
⎩
>0 for 0 < ck < 1 for all k = i , j , and k ∈ [1, 3]

For example, in Fig. 2, we have

⎧
⎨1 on BC except at points B, C
f 12 (c) = 0on A B and AC (2.27)
⎩
>0 otherwise

Consider

W 3 = f 12 W 33 + f 13 W 32 + f 23 W 31 (2.28)

and the associated gradient flow

∂c ∂ F3
= −W 3 . (2.29)
∂t ∂c
It is easy to check that the solution of Eq. (2.29) satisfies (P.1). Assume for example c 3 = 0, then we have f 12 = 1, f 13 = 0,
f 23 = 0 and W 3 = M 12 W 33 . Therefore

dc 1
= − M 12 (∂1 F 3 − ∂2 F 3 )|c3 =0 = −2M 12 σ12 c 1 c 2 (c 2 − c 1 ) (2.30)
dt
dc 2
= − M 12 (∂2 F 3 − ∂1 F 3 )|c3 =0 = −2M 12 σ12 c 1 c 2 (c 1 − c 2 ) (2.31)
dt
dc 3
=0 (2.32)
dt
Taking M 12 = M, we have a system Eq. (2.29) that satisfies (P.1) and can be reduced to the same two-phase equation
Eq. (2.15) when c 3 = 0. Notice that the form of W 3 looks similar to the results obtained in [9]. However, our derivation
is in a totally different way than theirs. Moreover, our approach can cooperate the cases with different mobilities between
different phases.
 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107 85

2.2.2. Derivation of a consistent energy functional through the matrix obtained in the gradient projection
We now provide a new way of deriving a consistent energy functional based on the results  we obtained with projection
approach mentioned in last section. The idea is to find a revised energy functional F3 (c) =  F (c)dx such that the gradient
flow system
∂c ∂F
= − , (2.33)
∂t ∂c
satisfies Eq. (1.1), (P.1) and (P.2). Here
⎛ ⎞
1 0 0
⎜ ⎟
=⎝ 0 2 0 ⎠,
0 0 3
where 1 , 2 and 3 are constants remaining to be determined.
To satisfy constraints (P.1) and (P.2), we require
∂F ∂ F3
 |c =0 = M kj f kj W 3i |c =0 , for k, j = i , ∀i ∈ {1, 2, 3}. (2.34)
∂c i ∂c i
In the following, the revised free energy density function F (c) and 1 , 2 , 3 are derived according to Eq. (2.34). To
satisfy Eq. (1.1), it is easy to see that the projection operator P 3 is equivalent to introduce another extra term (Lagrange
multiplier) H (c) in the energy functional, i.e.,
3
i ∂( F 3 (c) + Q 1 (c))
F (c) = F 3 (c) + Q 1 (c) + (c 1 + c 2 + c 3 − 1) H (c), where H (c) = − (2.35)
T ∂ ci
i =1

where  T = 1 + 2 + 3 and Q 1 (c) is determined as follows.

Remark 2.1. According to Eq. (2.33),

∂ ci ∂( F 3 (c) + Q 1 (c)) ∂ H (c)
= i [ + H (c) + (c 1 + c 2 + c 2 − 1) ] (2.36)
∂t ∂ ci ∂ ci
for i = 1, 2, 3 together with c 1 + c 2 + c 3 = 1. Adding up Eq. (2.36) from 1 to 3, we have
3
i ∂( F 3 (c) + Q 1 (c))
H (c) = −
T ∂ ci
i =1

We also refer the derivation of Lagrangian multipliers to Refs. [3,4].

Inspired by Eq. (2.28), it is natural to assume Q 1 (c) has the following form

Q 1 (c 1 , c 2 , c 3 ) = f 23 (c) g 23 (c 1 , c 2 , c 3 ) + f 13 (c) g 13 (c 1 , c 2 , c 3 ) + f 12 (c) g 12 (c 1 , c 2 , c 3 ) (2.37)

where g i j (c 1 , c 2 , c 3 ) are determined as follows.
By consistency condition for free energy functional and the definition of f i j in Eq. (2.26), we have

Q 1 (0, c 2 , c 3 ) = Q 1 (c 1 , 0, c 3 ) = Q 1 (c 1 , c 2 , 0) = 0
which implies

g 23 (0, c 2 , c 3 ) = 0, g 13 (c 1 , 0, c 3 ) = 0, g 12 (c 1 , c 2 , 0) = 0.
To simplified the calculations, we further assume that,

g 23 (c 1 , 0, c 3 ) = g 23 (c 1 , c 2 , 0) = 0, (2.38)
g 13 (c 1 , c 2 , 0) = g 13 (0, c 1 , c 3 ) = 0, (2.39)
g 12 (c 1 , 0, c 3 ) = g 12 (0, c 2 , c 3 ) = 0. (2.40)
Therefore, we can take

g 23 (c 1 , c 2 , c 3 ) = c 1 c 2 c 3 h23 (c 1 , c 2 , c 3 ), (2.41)
g 12 (c 1 , c 2 , c 3 ) = c 1 c 2 c 3 h12 (c 1 , c 2 , c 3 ), (2.42)
g 13 (c 1 , c 2 , c 3 ) = c 1 c 2 c 3 h13 (c 1 , c 2 , c 3 ). (2.43)
86 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107

Taking c 1 = 0 and substituting Eq. (2.37) into Eq. (2.34), we have

2 + 3 2 3
1 ( ∂1 ( F 3 + Q 1 ) − ∂2 ( F 3 + Q 1 ) − ∂2 ( F 3 + Q 1 )) |c1 =0 = 0, (2.44)
T T T
1 1 + 3 3
2 (− ∂1 ( F 3 + Q 1 ) + ∂2 ( F 3 + Q 1 ) − ∂2 ( F 3 + Q 1 )) |c1 =0 = M 23 (∂2 F 3 − ∂3 F 3 ) |c1 =0 , (2.45)
T T T
1 2 1 + 2
3 (− ∂1 ( F 3 + Q 1 ) − ∂2 ( F 3 + Q 1 ) + ∂3 ( F 3 + Q 1 )) |c1 =0 = M 23 (−∂2 F 3 + ∂3 F 3 ) |c1 =0 . (2.46)
T T T
Therefore, from Eq. (2.44), we have
2 c 3 + 3 c 2
h23 (c 1 , c 2 , c 3 ) |c1 =0 = 2σ23 . (2.47)
2 + 3
In fact, we can also choose h23 (c 1 , c 2 , c 3 ) to be the same function for c 1 = 0. Therefore, we have
2 c 3 + 3 c 2
h23 (c 1 , c 2 , c 3 ) = 2σ23 . (2.48)
2 + 3
Eq. (2.45) and Eq. (2.46) then give
2 3
2 σ23 (c2 c32 − c22 c3 ) = M 23 (2σ23 c2 c32 − 2σ23 c22 c3 ), (2.49)
2 + 3
2 3
−2 σ23 (c2 c32 − c22 c3 ) = M 23 (−2σ23 c2 c32 + 2σ23 c22 c3 ) (2.50)
2 + 3
which gives
2 3
= M 23 . (2.51)
2 + 3
Similarly, we have
2 c 1 + 1 c 2 3 c 1 + 1 c 3
h12 (c 1 , c 2 , c 3 ) = 2σ12 , h13 (c 1 , c 2 , c 3 ) = 2σ13
2 + 1 3 + 1
Thus,
1 3 1 2
= M 13 , = M 12 . (2.52)
1 + 3 2 + 3
Solving 1 , 2 and 3 from Eqs. (2.51) and (2.52), we get
1 1 1 1 1
= ( + − ),
1 2 M 13 M 12 M 23
1 1 1 1 1
= ( + − ), (2.53)
2 2 M 12 M 23 M 13
1 1 1 1 1
= ( + − ).
3 2 M 13 M 23 M 12
In summary,
2 c 3 + 3 c 2 3 c 1 + 1 c 3 2 c 1 + 1 c 2
Q 1 (c) = f 23 (c)c 1 c 2 c 3 2σ23 + f 13 (c)c 1 c 2 c 3 2σ13 + f 12 (c)c 1 c 2 c 3 2σ12 .
2 + 3 3 + 1 2 + 1

Remark 2.2. Notice that there exist functions f 23 , f 13 , f 12 satisfying the conditions Eq. (2.26), such as,
− R 1c − R 1c − R 1c − R 1c − R 1c − R 1c
e 1 2 e 1 3 e 2 1 e 2 2 e 3 1 e 3 3
f 23 (c) = 1 1
, f 12 (c) = 1 1
, f 13 (c) = 1 1
− − R 1 (1−c 3 )
− R 2 (1−c 1 )
− R 2 (1−c 2 )
− R 3 (1−c 1 )
−R
e R 1 (1−c 2 )
e e e e e 3 (1−c 3 )

where R 1 , R 2 and R 3 are positive constants. There are three singular points (0, 0, 1), (0, 1, 0), (0, 0, 1) for free energy density
functional in the bulk. To remove the singular points, we define

0 for c = (1, 0, 0), (0, 1, 0), (0, 0, 1)
Q1 =  (2.54)
Q1 for otherwise
where
 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107 87

Fig. 3. Comparison behavior of F (c 1 , c 2 , 1 − c 1 − c 2 ) with σ12 = σ13 = σ23 = 1 and 1 = 2 = 3 = 1 for different values of R 1 , R 2 , R 3 .

 2 c 3 + 3 c 2 3 c 1 + 1 c 3 2 c 1 + 1 c 2
Q1 = f 23 (c)c 1 c 2 c 3 2σ23 + f 13 (c)c 1 c 2 c 3 2σ13 + f 12 (c)c 1 c 2 c 3 2σ12 .
2 + 3 3 + 1 2 + 1
It is obvious that function Q 1 ∈ C 2 (D ) where D = {(c 1 , c 2 , c 3 ) ∈ R3 |c 1 ≥ 0, c 2 ≥ 0, c 3 ≥ 0, c 1 + c 2 + c 3 = 1}. We comment
that the behavior of the energy functional can be controlled by adjusting the values of R 1 , R 2 and R 3 (see Fig. 3). It is
easy to see that when R 1 , R 2 , R 3 are large enough, the functional is convex. One can also introduce high order terms in Eq.
(2.54) to ensure this convexity property. That is, we take  Q1 as


Q1 = c 1 c 2 c 3 ( f 23 (c)(κ2 c 2 + κ3 c 3 ) + f 13 (c)(κ1 c 1 + κ3 c 3 ) + f 12 (c)(κ1 c 1 + κ2 c 2 ))
αβ τ β αβ τ β
+ f23 (c) c 23 c 1α c 2 c 3τ + f13 (c) c 13 c 1α c 2 c 3τ
α ≥2,β≥1,τ ≥1 α ≥1,β≥2,τ ≥1
αβ τ β
+ f12 (c) c 12 c 1α c 2 c 3τ ,
α ≥1,β≥1,τ ≥2
αβ τ αβ τ αβ τ
where c 23 ≥ 0, c 13 ≥ 0 and c 12 ≥ 0.

2.3. Enforcing the consistency conditions for three-phase flow on solid surfaces

We now consider the general three-phase system with the free energy functional
 3 
12 3
F̄3 (c 1 , c 2 , c 3 ) = F 3 (c) +  κi |∇ c i |2 dx + γ3 (c)ds (2.55)
 8
i =1
 ∂

where (κ1 , κ2 , κ3 ) are chosen such that F 3 and γ3 are consistent with the energy functional of two-phase flow both in the
bulk and on the solid surface respectively. Therefore we have

κ1 + κ2 κ1 + κ3 κ2 + κ3
= σ12 , = σ13 , = σ23 (2.56)
2 2 2
and

γ3 (c) = σ w1 (3c12 − 2c13 ) + σ w2 (3c22 − 2c23 ) + σ w3 (3c32 − 2c33 ).

We remark the different interpolations in the wall free energy can be used. Here we use polynomial interpolation which is
preferred as explain in [15] since in this case, the interface profile is consistent with the standard “tanh” profile near the
contact line.
Solving κ1 , κ2 and κ3 from Eq. (2.56), we have

κ1 = σ12 + σ13 − σ23 ,

κ2 = σ12 + σ23 − σ13 , (2.57)
κ3 = σ13 + σ13 − σ12 .
It is possible that κi is negative (in the so-called total spreading situation). However, for simplicity, we will limit our
discussions under the conditions κi ≥ 0, for i = 1, 2, 3.
The projection approach introduced in §2.2 can be applied directly to the above functional both in the bulk and on
the solid surface and we have the time evolution of three-phase flow on solid surface driven by energy minimization and
satisfying constrains Eq. (1.1), (P.1) and (P.2) is given by
88 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107

∂c
= − W 3 μ̄, in  (2.58)
∂t
∂c
= − M 3 L̄ , on ∂ (2.59)
∂t
where
 
δ F̄3 (c) = dx[μ̄i δ c i ] + ds[ L̄ i δ c i ]. (2.60)
 ∂

Here, W 3 is given by Eq. (2.28)

⎛ ⎞ ⎛ ⎞ ⎛ ⎞
0 0 0 1 0 −1 1 −1 0
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
W 3 = f 23 (c) M 23 ⎝ 0 1 −1 ⎠ + f 13 (c) M 13 ⎝ 0 0 0 ⎠ + f 12 (c) M 12 ⎝ −1 1 0⎠ ,
0 −1 1 −1 0 1 0 0 0

and
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
0 0 0 1 0 −1 1 −1 0
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
M 3 = f 23 (c) 23 ⎝ 0 1 −1 ⎠ + f 13 (c) 13 ⎝ 0 0 0 ⎠ + f 12 (c) 12 ⎝ −1 1 0⎠.
0 −1 1 −1 0 1 0 0 0

Here,
3 12 ∂ F 3
μ̄1 = κ1 c1 + ,
4  ∂ c1
3 12 ∂ F 3
μ̄2 = κ2 c2 + ,
4  ∂ c2
3 12 ∂ F 3
μ̄3 = κ3 c3 +
4  ∂ c3
together with

3
L̄ 1 = κ1 ∂n c1 + ∂1 γ (c),
4
3
L̄ 2 = κ2 ∂n c2 + ∂2 γ (c),
4
3
L̄ 3 = κ3 ∂n c3 + ∂3 γ (c).
4
Moreover,
   
d ∂c ∂c
F̄3 = μ̄ + L̄ =− μ̄ W 3 μ̄ − L̄M 3 L̄
dt ∂t ∂t
 ∂  ∂
3 j −1  3 j −1 
=− M i j f i j (μ̄i − μ̄ j )2 − i j f i j ( L̄ i − L̄ j )2 ≤ 0 (2.61)
j =2 i =1  j =2 i =1∂

which shows decaying of free energy in time.

Remark 2.3. As in two-phase system, the parameter Ld determines the relaxation time of the interface in the bulk. In a given
three-phase system, we assume Ld to be the same for all interfaces. Therefore, according to Eq. (2.3), if the three-phase
system degenerates into a two-phase system, the mobility of the two-phase system is given by M i j = Ld /σi j for i = j,
i , j ∈ {1, 2, 3}, i.e.

Ld = M 12 σ12 = M 13 σ13 = M 23 σ23 . (2.62)

Similarly, on the solid surface, we also take νs as a constant. Therefore, according to Eq. (2.3), when degenerated into a
two-phase system, the mobility on the solid boundary is given by i j = νs /σi j for i = j, i , j ∈ {1, 2, 3}, i.e.

νs = 12 12σ = σ =
13 13 σ
23 23 . (2.63)
 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107 89

Notice that Eqs. (2.58)–(2.59) are non-linear in the derivative terms (because of coefficients f i j ). This makes numerical
simulation less efficient. To keep the highest derivative terms linear, we will treat the interface energy term separately. We
want to find a revised energy functional F3

 3 
12 3 2
F3 (c) = F (c) +  κi |∇ c i | dx + γ (c)dx (2.64)
 8
i =1
 ∂

such that

∂ ci
= −i μi , in  (2.65)
∂t
∂ ci
= −i L i on ∂ (2.66)
∂t
where
 
δF3 (c) = dx[μi δ c i ] + ds[ L i δ c i ]. (2.67)
 ∂

According to the derivation in §2.2.2, same procedure can be applied directly to the derivation of γ (c) and ’s (shown in
Appendix B), we get

1 1 1 1 1
= ( + − ),
1 2 M 13 M 12 M 23
1 1 1 1 1
= ( + − ), (2.68)
2 2 M 12 M 23 M 13
1 1 1 1 1
= ( + − )
3 2 M 13 M 23 M 12
and
1 1 1 1 1
= ( + − ),
1 2 13 12 23

1 1 1 1 1
= ( + − ), (2.69)
2 2 12 23 13

1 1 1 1 1
= ( + − ).
3 2 13 23 12

Noticing that in Eq. (2.62), we have taken Ld = M 12 σ12 = M 13 σ13 = M 23 σ23 and νs = σ =
12 12 σ =
13 13 σ
23 23 . Therefore,
substituting Eq. (2.62) into Eq. (2.68), we have

2Ld 2Ld
1 = = ,
σ12 + σ13 − σ23 κ1
2Ld 2Ld
2 = = , (2.70)
σ12 + σ23 − σ13 κ2
2Ld 2Ld
3 = = .
σ13 + σ23 − σ12 κ3
Derivation of γ (c) is shown in Appendix B. Similarly, substituting Eq. (2.63) into Eq. (2.69)
2ν s 2ν s
1 = = ,
σ12 + σ13 − σ23 κ1
2ν s 2ν s
2 = = , (2.71)
σ12 + σ23 − σ13 κ2
2ν s 2ν s
3 = = .
σ13 + σ23 − σ12 κ3
90 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107

Thus, substituting Eq. (2.70) and Eq. (2.71) into Eq. (2.65) and Eq. (2.66) respectively, we get
⎛ ⎞ ∂F ⎞ ⎛
∂ c1 ⎛ ⎞ 1
⎜ ∂t ⎟ 1 κ1 c1 ⎜ ∂ c1 ⎟
⎜ ⎟ ⎜ ⎟ 12 ⎜ ⎟
⎜ ∂ c2 ⎟ 3 ⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ 2 ∂ F ⎟
⎜ ∂t ⎟ = − 4 ⎜ 2 κ2
⎝
c2 ⎟ −
⎠  ⎜ ∂ c2 ⎟
⎜ in  (2.72)
⎜ ⎟ ⎟
⎝ ∂c ⎠ 3 κ3 c3
⎝ ∂F ⎠
3
3
∂t ∂ c3
and
⎛ ⎞ ∂γ ⎞
⎛
∂ c1 ⎛ ⎞ 1
⎜ ∂t ⎟ 1 κ1 ∂n c 1 ⎜ ∂ c1 ⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ ∂ c2 ⎟ 3 ⎜ ⎟ ⎜ ∂γ ⎟
⎜ ⎟ ⎜ ⎟
⎜ ∂t ⎟ = − 4 ⎜ 2 κ2 ∂n c 2 ⎟ − ⎜ 2
⎝ ⎠ ⎜ ∂ c2 ⎟
on ∂ (2.73)
⎜ ⎟ ⎟
⎝ ∂c ⎠ 3 κ3 ∂n c 3 ⎝ ∂γ ⎠
3
3
∂t ∂ c3
together with

2Ld = 1 κ1 = 2 κ2 = 3 κ3 ,
2 ν s =  1 κ1 =  2 κ2 =  3 κ3 .
Here the revised free energy functional is
 3 
12 3
F3 (c 1 , c 2 , c 3 ) = F (c) +  κi |∇ c i |2 dx + γ (c)ds
 8
i =1
 ∂
where

F (c) = σ12 c 12 c 22 + σ13 c 12 c 32 + σ23 c 22 c 32 + Q1 (c) + (c 1 + c 2 + c 3 − 1) H (c), (2.74)

2
γ (c) = σ w1 (3c 1 − 2c 13 ) + σ 2
w2 (3c 2 − 2c 23 ) + σ 2
w3 (3c 3 − 2c 33 ) + Q2 (c) + (c 1 + c 2 + c 3 − 1) Q (c) (2.75)
and

Q1 (c) = c 1 c 2 c 3 (f23 (c)(κ2 c 2 + κ3 c 3 ) + f13 (c)(κ1 c 1 + κ3 c 3 ) + f12 (c)(κ1 c 1 + κ2 c 2 )),

σ w2 κ3 σ w3 κ2 σ w2 κ1 σ w1 κ2 σ w3 κ1 σ w1 κ3
Q2 (c) = 6c 1 c 2 c 3 (f23 (c)( + ) + f12 (c)( + ) + f13 (c)( + ))
κ2 + κ3 κ2 + κ3 κ1 + κ2 κ1 + κ2 κ1 + κ3 κ1 + κ3
and
3
i
H (c) = − ∂i (σ12 c 12 c 22 + σ13 c 12 c 32 + σ23 c 22 c 32 + Q1 (c)), (2.76)
T
i =1
3
i
Q (c) = − ∂i (σ w1 (3c 12 − 2c 13 ) + σ w2 (3c 22 − 2c 23 ) + σ w3 (3c 33 − 2c 33 ) + Q2 (c)). (2.77)
T
i =1

It is obviously that system (2.72) and (2.73) satisfies the constraints Eq. (1.1), (P.1) and (P.2).

2.4. The three-phase Cahn–Hilliard Navier–Stokes model with solid surfaces

As discussed above, the total free energy functional of the three-phase flow with solid surface is given by
 3 
12 3
F3 (c 1 , c 2 , c 3 ) = F (c) +  κi |∇ c i |2 dx + γ (c)ds
 8
i =1
 ∂
where c = (c 1 , c 2 , c 3 ) represents the volume fraction of each phase in the mixture which satisfies the constraint

c 1 + c 2 + c 3 = 1. (2.78)
 is a small parameter denoting the interface thickness and σ12 , σ13 and σ23 are surface tensions between two of the three
phases correspondingly. (κ1 , κ2 , κ3 ) are constants given in Eq. (2.57). F (c) and γ (c) denote the free energy densities in the
bulk region and on the solid surface, respectively.
 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107 91

Under the conditions

• Ld = M 12 σ12 = M 13 σ13 = M 23 σ23 ,

• νs = 12 σ12 = 13 σ13 = 23 σ23 ,
• κ1 > 0, κ2 > 0 and κ3 > 0,

where Ld and νs , as mentioned in last section, determines the relaxation time in the fluid–fluid interface and relaxation time
on the solid boundary in the three-phase system, respectively. The evolution of the system is driven by the minimization
of the total free energy under the constraints Eq. (1.1), (P.1) and (P.2). It is given by the following three-phase Cahn–Hilliard
equations with convection term

∂ ci 2Ld
+ u · ∇ ci = ∇ · ( ∇ μi ), in  (2.79)
∂t κi
12 3
μi = ∂i F (c) − κi c i , in  (2.80)
 4
for i = 1, 2 and 3. Here μ = (μ1 , μ2 , μ3 ) represents the chemical potential. In order to take into account the incompressible
hydrodynamics of the mixture, the Cahn–Hilliard systems are then coupled with the Navier–Stokes equations
3
∂u
ρ[ + u · ∇ u] = −∇ p + ∇ · (η Du) + μi ∇ c i + ρ gext , in  (2.81)
∂t
i =1

∇ · u = 0, in  (2.82)
3
Here u is the velocity and p is the pressure. D (u) = ∇ u + ∇ u T denotes the viscous part of the stress tensor. i =1 μi ∇ c i
represents the capillary force. ρ gext is the external body force density. ρ and η are the fluid mass density and viscosity of
the three-phase flow, which are obtained by the linear interpolations of densities and viscosities of the three phases, i.e.,
3 3
ρ= c i ρi , η= c i ηi . (2.83)
i =1 i =1

We apply the relaxation boundary condition on the phase field variables for the Cahn–Hilliard equations, the boundary
conditions at the solid surface is given by

∂ ci 2ν s
+ u τ · ∂τ c i = − L (c), on ∂ (2.84)
∂t κi i
3
Li (c) = κi ∂n c i + ∂i γ (c) on ∂ (2.85)
4
for i = 1, 2 and 3. Extending the idea of generalized Navier boundary condition (GNBC) from two-phase flow (Qian et al.
[12]) to three-phase flow, the boundary condition for velocity on the solid surface is given by
3
slip
β u τ = −η(∂n u τ + ∂τ un ) + [ Li (c)∂τ c i ] on ∂ (2.86)
i =1

together with the non-penetration boundary conditions

u n = 0, ∂n μi = 0 on ∂ (2.87)

for i = 1, 2, 3, where β is the slip coefficient. n, τ are unit vectors orthogonal and tangential to the solid boundaries,
u τ = u · τ , un = u · n.
To obtain the dimensionless equations, we scale length by some length scale L, velocity by some speed V , time by L / V ,
density by some liquid density ρ0 , viscosity by some liquid viscosity η0 and pressure by η0 V / L. Then

∂ ci 12 3
+ u · ∇ c i = Ld μi , μi = ∂i F (c) −  c i , in  (2.88)
∂t κi 4
3
∂u
Rρ [ + u · ∇ u] = −∇ p + ∇ · (η∇ u) + Bi μi ∇ c i + ρ fext , in  (2.89)
∂t
i =1

∇ · u = 0. in  (2.90)

Here μ = (μ1 , μ2 , μ3 ). μi (i=1,2,3) are the chemical potentials. Density and viscosity are given by
92 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107

3 3
ρ= c i λρ i , η= c i λη i (2.91)
i =1 i =1

where λρi = ρi /ρ0 , ληi = ηi /η0 denote the density ratio, viscosity ratio for each phase i to ρ0 and η0 . External force field
fext = Rgext G . The boundary conditions at the solid surface are

∂ ci 3 1
+ u τ · ∂τ c i = −Vs Li (c), Li (c) =  ∂n c i + ∂i γ (c), on ∂ (2.92)
∂t 4 κi
u n = 0, ∂n μi = 0, on ∂ (2.93)
3
(Ls l s )−1 u τ = −(∂n u τ + ∂τ un ) + [
slip
Bi Li (c)∂τ c i ]/η on ∂ (2.94)
i =1

for i = 1, 2 and 3, where l s is given by

3
ls = c i λls,i . (2.95)
i =1

Here λls,i = l s,i /l s,0 denoting the ratio of slip length for each phase i to l s,0 . Natural boundary conditions are assumed on the
other boundaries of the computational domain,

∂n c i = 0, ∂n μi = 0, ∂n u = 0 (2.96)

for i = 1, 2, 3.
For rectangular domain, if taking the solid boundary as top and bottom of the computational domain, we can also assume
periodic boundary condition in x-direction,

c i , μi , u are periodic in x direction. (2.97)

Nine sets of dimensionless parameters (except G ) appear in the above equations. They are (1) Ld = 2Ld / V L 2 ; (2) R =
ρ0 V L /η0 , which is a Reynolds number of liquid; (3) Bi = κi B0 , B0 = 1/η0 V ; (4) Vs = 2νs L / V ; (5)Ls = η0 β/ L; (6)  , which
is the ratio between interface thickness ξ and characteristic length L; (7) G = L / V 2 ; (8) λls , λρ , λη are ratio of slip length,
density, viscosity ratios of each phase to the characteristic parameters respectively.
Straightforward calculations show that the model proposed above satisfies constraints Eq. (1.1), (P.1) and (P.2). (For details,
see Appendix A.)

3. Energy law for the model of three-phase flow on solid surfaces

We now show that energy law can be derived from the above phase field model with the corresponding boundary
√
conditions. In [5], Guermond and Quatapelle introduced a new variable σ = ρ in place of ρ . Using mass conservation law,

ρt + ∇ · (ρ u) = 0, (3.1)

one can derive

1 1
σ (σ u)t = ρ ut + ρt u = ρ ut − ∇ · (ρ u)u. (3.2)
2 2
Therefore, it is physically consistent to replace ρ ut in Eq. (2.89) by σ (σ u)t + 12 ∇ · (ρ u)u, and we can get

3
1
R[σ (σ u)t + ρ (u · ∇)u + ∇ · (ρ u)u] = −∇ p + ∇ · [η D (u)] + Bi μi ∇ c i , (3.3)
2
i =1

neglecting external force. The main advantage of this form is that the following property holds:
 
1
ρ (u · ∇)w · w + ∇ · (ρ u)w · w = 0 if u · n|∂ = 0 and ρ , u, w are periodic in x-direction.
2
 

We then introduce the following notations for various energy terms:

 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107 93


1
F k (u) = R ρ |u|2 ,
2

 3
12 3
F b (c) = B0 F (c) +  κi |∇ c i |2 dx,
 8
i =1


F w f (c) = B0 γ (c)ds,
∂
F t = F k (u) + F b (c) + F w f (c). (3.4)
Here F k (u) is the bulk kinetic energy; F b (c) is the bulk free energy; F w f (c) represents the surface energy at the fluid–solid
boundary and F t is the total free energy.

Theorem 3.1. Suppose c i , u, p are the smooth solutions to the system Eq. (2.88), Eq. (2.89), with boundary conditions Eq. (2.92),
Eq. (2.93), Eq. (2.94), Eq. (2.97) then the following energy law holds:
3 3
d 1 √
Ft = − Bi Ld  ∇ μi 2 − Bi Vs  L i (c) 2∂ −  η D (u) 2
dt 2
i =1 i =1

η slip 2 1 η slip
− L−
s
1
 u τ ∂ − L−
s ( u τ , u w )∂ + (ρ fext , u) . (3.5)
ls ls
slip
Here u τ = u τ − u w , u w is the wall speed at the solid surface and  · , (··) represent L 2 norm and inner product on domain , or at
the solid surface ∂.

Proof. Multiplying equation Eq. (2.88) by Bi μi and summing from i = 1 to 3, dotting Eq. (2.89) by u, and adding them
together, we have, after integrating on the domain ,
3 3
∂ 1 1 ∂ ci
LHS = ( Rσ u2 ) + R(ρ (u · ∇)u + ∇ · (ρ u)u, u) + Bi ( , μi ) + Bi (u · ∇ c i , μi ).
∂t 2 2 ∂t
i =1 i =1

Since
 
∂ ci 3 ∂ ci 12 ∂ ci ∂ ci 3
( , B i μ i ) = B0 κi ∇ c i · ∇ + ∂i F (c) dx − B0 κi ∂n c i ds
∂t 4 ∂t  ∂t ∂t 4
 ∂
3
∂ ci
and = 0, we have
∂t
i =1

3 3 
∂ ∂ ci 3
LHS = ( F k (u) + F b (c)) + Bi (u · ∇ c i , μi ) − B0 κi ∂n c i ds.
∂t ∂t 4
i =1 i =1∂

By the boundary condition, we have

3  3  3 
∂ ci 3 ∂ ci ∂ ci
− B0 κi ∂n c i ds = − Bi Li (c)ds + B0 ∂i γ (c)ds
∂t 4 ∂t ∂t
i =1 ∂ i =1 ∂ i =1 ∂
3  
∂
= Vs Bi |Li (c)|2 ds + (Bi Li (c), u τ · ∂τ c i )∂ + B0 γ (c)ds.
∂t
i =1∂ ∂

Therefore,
3  3 3
∂
LHS = ( F k (u) + F b (c) + F w f (c)) + Vs Bi |Li (c)|2 ds + (Bi Li (c), u τ · ∂τ c i )∂ + Bi (u · ∇ c i , μi )
∂t
i =1∂ i =1 i =1

3  3
∂
= ( F k (u) + F b (c) + F w (c)) + Vs Bi |Li (c)|2 ds + Bi (u · ∇ c i , μi )
∂t
i =1∂ i =1
94 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107


−1 η slip 2 1 η slip
+ Ls  u τ ∂ + L−
s ( u τ , u w )∂ + (η ∂n u τ , u τ )∂ .
ls ls
For the right hand side,
3 3
RHS = (−∇ p + ∇ · (η(∇ u + ∇ u T )) + ρ fext , u) + Bi (μi ∇ c i , u) + Bi (Ld μi , μi )
i =1 i =1
3 3 3
= (∂ j (η(∂i u j + ∂ j u i )), u i ) + (ρ fext , u) + Bi (u · ∇ c i , μi ) − Bi Ld ∇ μi 2 + Bi Ld (μi , ∂n μi )∂
i =1 i =1 i =1
3 3
= −(η(∂i u j + ∂ j u i ), ∂ j u i ) + (n j η(∂i u j + ∂ j u i ), u i )∂ + (ρ fext , u) + Bi (u · ∇ c i , μi ) − Bi Ld ∇ μi 2
i =1 i =1
3 3
1 √
= −  η D (u)2 + (ρ fext , u) + (η∂n u τ , u τ )∂ + Bi (u · ∇ c i , μi ) − Bi Ld ∇ μi 2 .
2
i =1 i =1

Then, comparing LHS with RHS gives Eq. (3.5). 2

Remark 3.1. Notice that the first four terms on the right hand side in Eq. (3.5) are energy dissipations due to the corre-
sponding form as the four dissipation terms mentioned in [12], the fifth term is the work done per unit time by the flow to
the wall and the last term is the work done per unit time by the body force to the flow. Thus, Theorem 3.1 shows exactly
the energy conservation of the system.

4. Discrete energy law and numerical scheme

4.1. The Cahn–Hilliard equation

In this section, we design a numerical scheme that has discrete energy decaying property for three-phase model we
derived. Following the convex splitting ideas of Eyre [1,2], we first consider the pure Cahn–Hilliard equation,
∂ M0
ci = μi , in  (4.1)
∂t κi
12 3
μi = ∂i F (c) − κi c i , in  (4.2)
 4
together with the relaxation boundary condition for c i ,
∂ 0
c i = − Li (c), on ∂ (4.3)
∂t κi
3
Li (c) = ∂i γ (c) + κi ∂n c i , on ∂ (4.4)
4
and Neumann boundary condition for μi ,
∂n μi = 0. on ∂ (4.5)
For the gradient flow above, one can split the total free energy into two parts: the contractive part F bc (c), c
Fw f
(c) and the
expansive part F be (c), F w
e
f
(c).
 3 
12 3
F(c) = F (c) +  κi |∇ c i |2 dx + γ (c)ds = Fbc (c) − Fbe (c) + Fcw f (c) − Few f (c), (4.6)
 8
i =1
 ∂

where
3   3
3 3s √
Fbc (c) = κi |∇ c i |2 dx + κi c i2 dx = f bc ( κi c i ),
i =1 
8  i =1

 3  
12 3s
Fbe (c) = − F(c) + κi c i2 dx = f be (c)dx,
 i =1

  
 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107 95

3  3

α √
Fcw f (c) = κi c i2 ds = c
fw f( κi c i ),
2
i =1 ∂ i =1

3   

α
Few f (c) = κi c i2 ds − γ (c)ds = e
fw f (c)ds.
2
i =1 ∂ ∂ ∂

Here
 
3 3s α̃
f bc (φ) = (  |∇φ|2 + φ 2 )dx, c
fw f (φ) = φ 2 ds.
8  2
 ∂

The corresponding semi-implicit scheme is given by

cni +1 − cni M0
= μni +1 , in  (4.7)
δt κi
cni +1 − cni
Li (cn+1 )
0
=− on ∂ (4.8)
δt κi
where

δ F bc (cn+1 ) δ F be (cn )
μni +1 = − , (4.9)
δci δci
3
c
δFw (cn+1 ) e
δFw (cn )
Li (cn+1 ) = κi ∂n cni +1 +
f f
− (4.10)
4 δci δci
with the boundary condition

∂n μni +1 = 0 on ∂ (4.11)

Theorem 4.1. The energy law for the semi-implicit scheme designed above is as follows. Let cn , n = 0, 1, · · · , k satisfy equations and
their boundary conditions Eq. (4.7)–Eq. (4.11), and assume
3 1 ∂ γ
2 3 1 ∂ F 2
N= sup | j =1 κi ∂ c i ∂ c j (c (x)) | and M =
n
sup | j =1 κi ∂ c i ∂ c j (c (x)) |.
n
x∈∂,n=0,1,··· ,k,i =1,2,3 ¯ n=0,1,··· ,k,i =1,2,3
x∈,
Then, if s ≥ M, 
α ≥ N /2, we have
3 3
M0
F(cn+1 ) − F(cn ) ≤ −δ t ∇ μni +1 2 − δt  L i (cn+1 )2∂ .
0
(4.12)
i =1
κi i =1
κi

Proof.

F(cn+1 ) − F(cn ) = Fb (cn+1 ) − Fb (cn ) + F w f (cn+1 ) − F w f (cn ).

We consider
3
√ √ √
Fbc (cn ) − Fbc (cn+1 ) = f bc ( κi cni +1 + κi (cni − cni +1 )) − f bc ( κi cni +1 )
i =1
√ n +1 √ √ √
3
d f bc ( i ciκ + θ κi (cni − cni +1 )) 1
3
d2 f bc ( κi cni +1 + θ κi (cni − cni +1 ))
= |θ =0 + |θ =ξi
dθ 2 dθ 2
i =1 i =1
√ 
3
 δ f bc ( κi cni +1 )  3
3
= , cni − cni +1 + κi ∂n cni +1 (cni − cni +1 )ds
δci 4
i =1 i =1∂

3 
3 3s
+ κi |∇(cni − cni +1 )|2 + κi (cni − cni +1 )2 dx
i =1 
4 

and
96 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107

Fbe (cn+1 ) − Fbe (cn ) = Fbe (cn + (cn+1 − cn )) − Fbe (cn )


dF be (cn + θ(cn+1 − cn )) 1
= |θ = 0 + (cn+1 − cn )T ∇ 2 f be (cn + θ(cn+1 − cn ))(cn+1 − cn )dx |θ =ξ
dθ 2

3 
 δ F be (cn )  1 6s 12 ∂ 2 F
= , cn+1 − cn + [ κi − (cn + θ(cn+1 − cn ))](cni +1 − cni )2
δc 2
i =1, j =i 
  ∂ c i2
2
12 ∂ F
− (cn + θ(cn+1 − cn ))(cni +1 − cni )(cnj +1 − cnj )dx|θ =ξ .
 ∂ ci ∂ c j
Thus,

Fb (cn+1 ) − Fb (cn )

3
δ F c (cn+1 ) δ F be (cn ) n+1 3
3
=  b − , c i − cni  + κi ∂n cni +1 (cni +1 − cni )ds
δci δci 4
i =1 i =1∂

3 
3
− κi |∇(cni − cni +1 )|2 dx
4
i =1 

3 
6 1 ∂2 F
− κi [s(cni +1 − cni )2 − (cn + θ(cn+1 − cn ))(cni +1 − cni )2
i =1, j =i
 κi ∂ c i2

∂2 F
1
− (cn + θ(cn+1 − cn ))(cni +1 − cni )(cnj +1 − cnj )]dx|θ =ξ .
κi ∂ c i ∂ c j
If s ≥ M, and by the properties of
3 3
cni = cni +1 = 1, and ∂n μi = 0, for i = 1, 2, 3,
i =1 i =1

we have

Fb (cn+1 ) − Fb (cn )

3
 δ F bc (cn+1 ) δ F be (cn ) n+1  3
3
≤ − , c i − cni + κi ∂n cni +1 (cni +1 − cni )ds
δci δci 4
i =1 i =1∂

3 3 
  3
= μni +1 , cni +1 − cni + κi ∂n cni +1 (cni +1 − cni )ds
4
i =1 i =1∂

3 3 
  3
= μni +1 , cni +1 − cni + κi ∂n cni +1 (cni +1 − cni )ds
4
i =1 i =1∂

3 3 
 M0  3
= δt μni +1 , ∇ · ( ∇ μni +1 ) + κi ∂n cni +1 (cni +1 − cni )ds
i =1
κi i =1∂
4

3 3 
M0   3
= −δ t ∇ μni +1 , ∇ μni +1 + κi ∂n cni +1 (cni +1 − cni )ds.
i =1
κi i =1∂
4

Similarly,
3
√ √ √
Fcw f (cn ) − Fcw f (cn+1 ) = c
fw f( κi cni +1 + κi (cni − cni +1 )) − f wc f ( κi cni +1 )
i =1
√ √ √ √
3 c
d fw f
( i cni +1 κ + θ κi (cni − cni +1 )) 1
3 c
d2 f w f
( κi cni +1 + θ κi (cni − cni +1 ))
= |θ = 0 + |θ =ξi
dθ 2 dθ 2
i =1 i =1
 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107 97

3  3 
α̃
= α̃κi cni +1 (cni − cni +1 )ds + κi (cni − cni +1 )2 ds
2
i =1 ∂ i =1∂

and

Few f (cn+1 ) − Few f (cn ) = Few f (cn + (cn+1 − cn )) − Few f (cn )

e
dF w (cn + θ(cn+1 − cn )) 
1
(cn+1 − cn )T ∇ 2 f we f (cn + θ(cn+1 − cn ))(cn+1 − cn )ds |θ =ξ
f
= | θ =0 +
dθ 2
∂
3  3 
δ γ (cn ) n+1
= α̃ κi cni (cni +1 − cni )ds − (c i − cni )ds
δci
i =1 ∂ i =1∂

3 
1 1 ∂ 2γ
+ κi (α̃ (cni +1 − cni )2 − (cn + θ(cn+1 − cn ))(cni +1 − cni )2
2
i =1, j =i∂
κi ∂ c i2

1 ∂ 2γ
− (cn + θ(cn+1 − cn ))(cni +1 − cni )(cnj +1 − cnj ))ds|θ =ξ .
κi ∂ c i ∂ c j
Thus, if 
α ≥ N /2, we have
3 
δ γ (cn )
F w f (cn+1 ) − F w f (cn ) ≤ ( + α̃ (cni +1 − cni ))(cni +1 − cni )ds
δci
i =1∂

3 
3
= ( L i (cn+1 ) − κi ∂n cni +1 )(cni +1 − cni )ds
4
i =1∂

3  
3
= L i (cn+1 )(cni +1 − cni )ds − κi ∂n cni +1 (cni +1 − cni )ds
4
i =1∂ ∂
3  
3
| L i (cn+1 )|2 ds − κi ∂n cni +1 (cni +1 − cni )ds.
0
= −δt
i =1∂
κi 4
∂

Therefore,
3 3 
M0  
F(cn+1 ) − F(cn ) ≤ −δ t ∇ μni +1 , ∇ μni +1 − δt | L i (cn+1 )|2 ds.
0
2
i =1
κi i =1∂
κi

4.2. Schemes based on pressure stabilization for Navier–Stokes equation

To solve the Navier–Stokes equation, the pressure stabilization method has been employed as it is in [8]. It is a splitting
method for Navier–Stokes equations, which is based on pressure stabilization, involving only pressure Poisson equation with
constant coefficient proposed by Guermond and Salgado in [5]. In the scheme, the divergence free condition is replaced by

∇ · u − δ p t = 0. (4.13)
Combining the idea of convex splitting on Cahn–Hilliard equation and pressure stabilization for Navier–Stokes equation,
we propose the following numerical scheme for solving Eq. (2.88)–Eq. (2.97).
Given initial conditions c i0 , u0 , and set ρ = min ρ , computing cni +1 , un+1 , pn+1 for n ≥ 0 by,
⎧
⎪
⎪ cni +1 − cni
⎪
⎪
⎨ + un+1 · ∇ cni +1 = Ld μni +1 , x∈
δt
(4.14)
⎪
⎪
⎪
⎪ cn+1 − cni
⎩ i + u τ ∂τ cni +1 = −Vs Li (cn+1 ), ∂n μni +1 = 0 x ∈ ∂
δt
where
98 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107

3 6s 6s 12
μni +1 = −  cni +1 + cni +1 − cni + ∂i F (cn ), (4.15)
4   κi
3 1
Li (cn+1 ) =  ∂n cni +1 + α (cni +1 − cni )
∂i γ (cn ) +  (4.16)
4 κi
and
 
ρ n+1 , ηn+1 , lns +1 = (λρ 1 , λη1 , λls 1 )cn1+1 + (λρ 2 , λη2 , λls 2 )cn2+1 + (λρ 3 , λη3 , λls 3 )cn3+1 , (4.17)

⎧  1 n +1 
⎪
⎪ 2
(ρ + ρ n )un+1 − ρ n un n n n +1 1 n n n +1
⎪
⎪ R + ρ (u · ∇)u + (∇ · (ρ u ))u
⎪
⎪ δt 2
⎪
⎪
⎪
⎨ 3
= ∇ · [ηn+1 D (un+1 )] − ∇(2pn − pn−1 ) + Bi μi n+1 ∇ cni +1 , x ∈ , (4.18)
⎪
⎪
⎪
⎪ i =1
⎪
⎪ 3
⎪
⎪ (L ln+1 )−1 (u slip )n+1 =
⎪
⎩ ss τ Bi L i (cn+1 )∂τ cni +1 /ηn+1 − ∂n unτ+1 , unn+1 = 0, x ∈ ∂
i =1

⎧
⎨ ρ̄
( p n +1 − p n ) = R∇ · un+1 , x ∈ 
δt (4.19)
⎩ ∂ p n +1 = 0. x ∈ ∂
n

Here ρ̄ = min(1, λρ ). Similar to the continuous equations, the scheme we proposed could be proved to have the following
stability result.

Theorem 4.2. Let cn , pn , un , n = 0, 1, · · · , k satisfy equations and their boundary conditions Eq. (4.14)–Eq. (4.19). Assume N =
3 1 ∂ γ
2 3 2
κi ∂ c i ∂ c j (c (x)) |. Then, if s ≥ M, 
1 ∂ F
sup | j =1 κi ∂ c i ∂ c j (c (x)) | and M =
n
sup | j =1
n
α ≥ N /2, we
x∈∂,n=0,1,··· ,k,i =1,2,3 ¯ n=0,1,··· ,k,i =1,2,3
x∈,
have

δt 2
Ft (cn+1 ) − Ft (cn ) + (∇( pn − pn−1 )2 + ∇ pn+1 2 − ∇ pn 2 )
2ρ̄ R
3 3 
1
≤−δt Bi Ld  ∇ μin+1 2 −δt Bi Vs  L i (cn+1 ) 2∂ −δt  ηn+1 D (un+1 ) 2
2
i =1 i =1

η n +1 η
− δ t L− (u τ )n+1 2∂ − δt L− (u τ )n+1 , u w ))∂ .
1 slip 1 slip
s  s ( (4.20)
lns +1 lns +1

The proof is almost the same as Theorem 3.1 and Theorem 4.3.2 in [8].
As in [8], it is easy to construct second order accurate scheme in time as follows, (·)∗,n+1 = 2(·)n −(·)n−1 , ρ̄ = min(1, λρ ).
⎧
⎪
⎪ 3cni +1 − 4cni + cni −1
⎪
⎪
⎨ + un+1 · ∇ c i∗,n+1 = Ld μni +1 , x∈
2δ t
(4.21)
⎪
⎪
⎪
⎪ cn+1 − cni
⎩ i + u τ ∂τ cni +1 = −Vs Li (cn+1 ), ∂n μni +1 = 0 x ∈ ∂
δt
where
3 6s 6s ∗,n+1 12
μni +1 = −  cni +1 + cni +1 − ci + ∂i F (c∗,n+1 ),
4   κi
3 1
Li (cn+1 ) =  ∂n cni +1 + α (cni +1 − c i∗,n+1 )
∂i γ (c∗,n+1 ) + 
4 κi
and
 
ρ n+1 , ηn+1 , lns +1 = (λρ 1 , λη1 , λls 1 )cn1+1 + (λρ 2 , λη2 , λls 2 )cn2+1 + (λρ 3 , λη3 , λls 3 )cn3+1 , (4.22)
 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107 99

Fig. 4. (a) Schematic graph of the initial state of the four phases system. (b) The definitions of φ and θ .

⎧  
⎪
⎪ 3un+1 − 4un + un−1
⎪
⎪ R ρ n +1 + ρ n+1 (u∗,n+1 · ∇)un+1
⎪
⎪ 2δ t
⎪
⎪
⎪
⎨ 4 1
3
= ∇ · [ηn+1 D (un+1 )] − ∇( pn + ψ n − ψ n−1 ) + Bi μi n+1 ∇ cni +1 , x ∈ , (4.23)
⎪ 3 3
⎪
⎪ i = 1
⎪
⎪ 3
⎪
⎪
⎪
⎪ (L
⎩ ss l n+1 −1 slip n+1
) ( u τ ) = Bi L i (cn+1 )∂τ cni +1 /ηn+1 − ∂n unτ+1 , unn+1 = 0, x ∈ ∂
i =1

⎧
⎪
⎪ 3ρ̄
⎨ ψ n +1 = R∇ · un+1 , x∈
2δ t
∂n ψ n+1 = 0, x ∈ ∂ (4.24)
⎪
⎪
⎩ pn+1 = pn + ψ n+1 − ηn+1 ∇ · un+1 . x∈
∗,n+1 ∗,n+1
To implement the above numerical schemes, we treat the nonlinear terms explicitly, i.e., un+1 · ∇ c i as un · ∇ c i , which
will introduce a mild CFL-like constraint but will decouple the Cahn–Hilliard equation from the Navier Stokes equation. The
other numerical implementation details are similar to that in [7,8].

5. Numerical results

In this section, we present some numerical examples in which we study the three-phase steady state configuration and
the dynamics of the four phase contact line. We will always use the second order semi-implicit schemes described in the
last section and the following parameters in our simulations:

Ld = 5.0 × 10−4 , R = 10, B0 = 1, Vs = 500, Ls = 0.0038,  = 0.01, s = 6, α̃ = 10, λls,1 = λls,2 = λls,3 = 1.
The computational domain is (x, z) = [0, 0.8] × [0, 0.4]. The grid is (nx, nz) = (400, 200). The time step is δt = 0.1h where h
is the mesh size. The density ratios and the viscosity ratios will also be specified in different examples below.

5.1. Numerical result of a droplet spreading in a mixture of two-phase fluid between two stationary parallel planar solid surfaces

We first study the equilibrium configurations of a liquid lens 1 which is initially spherical and put on the bottom solid
surface and in between two other immiscible liquids, liquid 2 and liquid 3, and in a rectangular domain with length L and
height H as shown in Fig. 4(a). Denote σi j (i = j ), i = 1, ..., 4, j = 1, .., 4, the surface tension between the four phases, i.e.
liquid 1, liquid 2, liquid 3 and solid surface 4. For simplicity, we assume

σ24 = σ34 , σ12 = σ13 (5.1)

in the following discussions. It is easy to see that the equilibrium shape consists of spherical arcs and their intersections at
the triple junction and with the solid surface. The contact angles at the triple junction is determined by the three surface
tensions as following
σ23 σ12 σ13
= = (5.2)
sin(θ23 ) sin(θ12 ) sin(θ13 )
where θ23 + θ13 + θ12 = 2π . Because of the symmetry (σ12 = σ13 ), we can take θ23 = 2θ . Thus,
σ23 σ12 σ13 σ12 σ13
= = = = (5.3)
sin(2θ) sin(π − θ) sin(π − θ) sin(θ) sin(θ)
which implies
100 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107

Fig. 5. Equilibrium states for V 1 = 0.0113 and θ = π /6 and three different φ .

Fig. 6. Equilibrium states for θ = π /6 and three different φ = 3π /4 as the volume V 1 is increased.

σ23 = 2σ13 cos θ = 2σ12 cos θ. (5.4)

The static contact angle φ of fluid 1 and fluid 3 with solid surface 4 is determined from the Young’s equation,

σ14 + σ13 cos φ = σ34 = σ24 . (5.5)

In the numerical simulations in this subsection, we take density ratios and viscosity ratios to be 1. For small droplet
volume V 1 = 0.0113, given θ = π /6, the contact angle φ determines the equilibrium configuration of the liquid droplet 1.
In the hydrophilic cases φ = π /4, π /2, we have partial wetting states as shown in Fig. 5(a), 5(b). For φ = 3π /4, we have
a liquid lens in contact with the solid surface with a four phase contact line (Fig. 5(c)). It is worth to notice that, in this
case (Fig. 5(c)), the contact angles of the liquid-liquid interfaces at the four phase contact point are not equal to the Young’s
angle determined by Eq. (5.5). In addition, it is possible to show, from a energy argument, there is a critical φc for the
transition from partial wetting to the liquid lens with formation of the four phase contact line [10]. As the droplet volume
V 1 increases, the triple junction point will eventually touch the top boundary at about V 1 = 0.0804 (Fig. 6(a)) and form a
four phase contact line also. Further increase of the volume will eventually break up the four phase contact line as shown in
Fig. 6(b). It is possible to show that there is a range of volume ( V l , V h ) such that for V l < V 1 < V h , the four phase contact
line is stable [10].

5.2. Four phases moving contact line

As we have shown in the last section, a liquid lens 1, which is initially spherical putting at the interface of two other
immiscible liquids 2 and 3 between two parallel planar solid surfaces, can be in contact with the solid surfaces and form two
four phase contact lines when the volume of the lens is in certain range. We now consider a case with V 1 = 0.0616, θ = π /6
and φ = 3π /4. We are interested in the dynamics of the lens with four phase contact line under the Poiseuille flow and
study how the stability of the four phase contact line is affected by the density and viscosity ratios of the different phases.
We apply a body force in the right direction to the fluid in the channel with fext = (10, 0). Case (I): We assume that all
phases have equal density and viscosity, i.e., λη1 = λη2 = λη3 = 1, λρ 1 = λρ 2 = λρ 3 = 1. We observe a symmetric, deformed
liquid lens moving to the right with two moving contact points at the top and bottom boundaries as shown in Fig. 7. The
four phase contact line is stable in this case. Case (II): When the density ratios are changed to λρ 1 = 1, λρ 2 = λρ 3 = 0.1 (i.e.
the liquid 1 has much higher density), we see much weaker deformation of the liquid lens (Fig. 8). The four-phase contact
points are, however, maintained as the lens moves to the right, but with a slower velocity (due to higher density). Case (III):
When the viscosity ratios are changed to λη1 = 1, λη2 = λη3 = 0.1, the four phase contact point is maintained at the early
stage but becomes unstable at around t = 0.432 and eventually breaks up into two three-phase contact points, whereas the
whole system propagates to the right (see Fig. 9).
 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107 101

Fig. 7. Dynamics of liquid lens with four phase contact point for V 1 = 0.0616, θ = π /6, φ = 3π /4, λρ 1 = λρ 2 = λρ 3 = 1 and λη1 = λη2 = λη3 = 1.

Fig. 8. Dynamics of liquid lens with four phase contact point for V 1 = 0.0616, θ = π /6, φ = 3π /4, λρ 1 = 1, λρ 2 = λρ 3 = 0.1 and λη1 = λη2 = λη3 = 1.

Fig. 9. Dynamics of liquid lens with four phase contact point for V 1 = 0.0616, θ = π /6 and φ = 3π /4, λρ 1 = λρ 2 = λρ 3 = 1, λη1 = 1, λη2 = 0.1, λη3 = 0.1.
Four phase contact point is not stable in dynamics.

5.3. Discrete energy law and mass conservation

The discrete energy law proved in Theorem 4.2 of the numerical scheme is verified in the following. Two cases are
studied.
In the first case, the initial configuration is shown in Fig. 4(a) with V 1 = 0.0616. The wall speed is 0, the body force field
is also 0, λη1 = 1, λη2 = 1, λη3 = 1 and other parameters are the same as the case shown in Fig. 9. The evolution of various
energy terms is shown in Fig. 10(a), where F b , F w f , F k and F t are defined in (3.4). In Fig. 10(a), the kinetic energy is small
(∼ 10−4 ). The bulk energy (F b ) decreases fast at the beginning and is transferred to other energies and dissipations. The
surface energy at the fluid–solid boundary (F w f ) first increases due to σ41 > σ43 , σ42 , then decreases a little bit because
of the motion of the triple junction of liquid 1, 2 and 3 in the bulk. F w f keeps as a constant eventually when the system
approaches an equilibrium state. The total free energy (F t ) decreases in time as shown in Fig. 10(a), which is consistent
102 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107

Fig. 10. (a) Evolution of energies for the case without body force. The values in the figure are adjusted to zero at time t = 0 by subtracting their initial
values. (b) The absolute value of mass difference (|m(t ) − m(t = 0)|) as a function of time step in this case. The inset shows |m(t n+1 ) − m(t n )| at different
time steps.

Fig. 11. (a) Evolution of energies for the case with body force. The values in the figure are adjusted to zero at time t = 0 by subtracting their initial values.
(b) The absolute value of mass difference (|m(t ) − m(t = 0)|) as a function of time in this case. The inset shows |m(t n+1 ) − m(t n )| at different time steps.

with the conclusions of Theorem 3.1 and Theorem 4.2 when u w = 0 and fext = (0, 0) (here the pressure gradient term in
Theorem 4.2 is negligible because it is lower order of δt).
In the second case, we consider the simulation shown in Fig. 9. In this case, the wall speed is 0, the body force field is
fext = (10, 0) and λη1 = 1, λη2 = 0.1, λη3 = 0.1. The evolution of various energy terms is shown in Fig. 11(a), where
n n
ηk
δ t L−
slip
( F t + W )n = F tn + 1
s ( (u τ )k , u w ))∂ − δt (ρ k fext , uk )
k =0
lks k =0

is the total energy at time t = nδt plus work done to the fluid. In Fig. 11(a), when a body force is applied on the fluid, the
kinetic energy (F k ) increases in time. The surface energy F w f keeps as a constant before t = 0.4 because the liquid 1 has
not touched the solid boundaries (see Fig. 9(a)–(c)) and σ43 = σ42 , after liquid 1 touches the solid surface and gets wetting
on the surface (see Fig. 9(d)–(f)) F w f increases, after it reaches its static contact angle on the solid surface (see Fig. 9(g),
(h)) F w f keeps as a constant. The total free energy (F t ) increases in time, whereas F t + W decreases. This is consistent with
the conclusions of Theorem 3.1 and Theorem 4.2.
Mass conservation of the numerical scheme is also verified through these two cases (see Fig. 10(b), 11(b)). Since we only
calculate c 1 and c 2 in the numerical implementation and c 3 = 1 − c 1 − c 2 , the mass
 
m(t ) = c 1 (·, t )dx, or m(t ) = c 2 (·, t )dx
 
for phase 1 and phase 2, respectively. In both of the cases, the changing of mass comparing with the initial mass is in the
order 10−6 and this order is reached at the very beginning of the time evolution. Afterward, the changing of mass of c 1 and
c 2 comparing with the former time step (|m(t n+1 ) − m(t n )|) is kept in the order of 10−8 ∼ 10−10 .

6. Conclusions

In this paper, a new approach for deriving consistent free energy functional for three-phase system with solid boundaries
is proposed. Driven by minimization of the total free energy functional, a Cahn–Hilliard Navier Stokes equation system with
 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107 103

relaxation boundary condition for the phase parameters and generalized Navier boundary condition for the velocity field
on solid surfaces is derived by taking hydrodynamics of the mixture into consideration. The model is proved to be energy
decay. Moreover, a semi-implicit scheme is developed for the three-phase flow system with moving contact line problem
with variable density and viscosity. The scheme is shown to have energy decay property which ensures stability of the
scheme. We then show several numerical examples simulating the dynamics of triple junctions and four phase contact lines.
The discrete energy law and the mass conservation obtained in the numerical examples are consistent with the theoretical
conclusions.

Acknowledgements

We thank Tie-Zheng Qian for some helpful discussions. The research is supported in part by the Hong Kong RGC-GRF
grants 605513 and RGC-CRF C6004-14G-D. We are also grateful to the anonymous referees for their valuable comments and
suggestions.

Appendix A. Checking consistency with the model of two-phase flow on solid surfaces

In the following we are going to check the model proposed in §2.4 satisfy the constraints Eq. (1.1), (P.1) and (P.2).

A.1. Checking c 1 + c 2 + c 3 = 1

3
∂
Taking initial condition as c 1 (x, t 0 ) + c 2 (x, t 0 ) + c 3 (x, t 0 ) = 1, we only have to check that c i (x, t 0 ) = 0. It is
∂t
i =1
obviously that
3 3
∂ ci M0
= (∇ · ( ∇ μ i ) − u · ∇ c i ) = 0, (A.1)
i =1
∂t
i =1
κi
3 3
∂ ci 0
=− ( Li (c) + u τ · ∂τ c i ) = 0. (A.2)
i =1
∂t
i =1
κi

Then c 1 (x, t ) + c 2 (x, t ) + c 3 (x, t ) = 1 for any t ≥ t 0 .

A.2. Checking c i (x, t 0 ) = 0 =⇒ c i (x, t ) = 0, ∀t ≥ t 0

Taking i = 1 for example, i.e., c 1 (x, t 0 ) = 0, in the model we proposed

∂ c1 M0 12 3
+ u · ∇ c1 = ∇ · ( ∇ μ1 ), μ1 = ∂1 F (c) − κ1 c 1 , in 
∂t κ1  4
∂ c1 0 3
+ u τ · ∂τ c 1 = − L1 (c), L1 (c) = κ1 ∂n c1 + ∂1 γ (c), on ∂
∂t κ1 4
Thus, by the initial condition, it is obviously that
3 3
u · ∇ c 1 = 0, κ1 c1 = 0, u τ · ∂τ c1 = 0, κ1 ∂n c1 = 0 (A.3)
4 4
checking the other two terms, taking

F̄ (c) = σ12 c 12 c 22 + σ13 c 12 c 32 + σ23 c 22 c 32 + Q1 (c)

γ̄ (c) = σ w1 (3c12 − 2c13 ) + σ w2 (3c22 − 2c23 ) + σ w3 (3c33 − 2c33 ) + Q2 (c)
we have,
12 4κ T 1 1
∂1 F (c) = ( (∂1 F̄ (c) − ∂2 F̄ (c)) + (∂1 F̄ (c) − ∂3 F̄ (c)))
  κ2 κ3
4κ T 1 1 1 1
= (( + )∂1 F̄ (c) − ∂2 F̄ (c) − ∂3 F̄ (c))
 κ2 κ3 κ2 κ3
4κ T 4κ T
= (−2c 2 (1 − c 2 )(1 − 2c 2 ) − 2(1 − c 2 )c 2 (1 − 2(1 − c 2 ))) = (−2c 2 (1 − c 2 )(1 − 2c 2 − 1 + 2c 2 )) = 0
 
κT 1 1
∂1 γ (c) = ( (∂1 γ̄ (c) − ∂2 γ̄ (c)) + (∂1 γ̄ (c) − ∂3 γ̄ (c)))
3 κ2 κ3
104 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107

κT 1 1 σ w2 σ w3
= (− 6σ w2 c 2 (1 − c 2 ) − 6σ w3 c 3 (1 − c 3 ) + 6c 2 c 3 ( + ))|c2 +c3 =1 = 0
3 κ2 κ3 κ2 κ3
where 3/κ T = 1/κ1 + 1/κ2 + 1/κ3
Therefore,
∂
c i (x, t 0 ) = 0, c 1 (x, t 0 ) = 0 =⇒ c i (x, t ) = 0, ∀t ≥ t 0 (A.4)
∂t
A.3. Checking consistent with two-phase model in the bulk

Suppose c 1 ≡ 0, c 2 + c 3 = 1, in the model we proposed

∂ c2 M0
+ u · ∇ c2 = ∇ · ( ∇ μ2 ), in  (A.5)
∂t κ2
∂ c3 M0
+ u · ∇ c3 = ∇ · ( ∇ μ3 ), in  (A.6)
∂t κ3
where
12 3
μ2 = ∂2 F (c) − κ2 c 2
 4
12 3
μ3 = ∂3 F (c) − κ3 c 3
 4
Here
12 4κ T 1 1
∂2 F (c) = ( (∂2 F̄ (c) − ∂1 F̄ (c)) + (∂2 F̄ (c) − ∂3 F̄ (c)))
  κ1 κ3
4κ T 1 1 1 1
= (( + )∂2 F̄ (c) − ∂1 F̄ (c) − ∂3 F̄ (c))
 κ1 κ3 κ1 κ3
4κ T 1 1 1 1
= (( + )(2σ23 c 2 c 32 ) − (c 2 c 3 (κ2 c 2 + κ3 c 3 )) − (2σ23 c 3 c 22 ))
 κ1 κ3 κ1 κ3
4κ T 1 1
= c 2 c 3 ( (κ2 c 3 − κ2 c 2 ) + (κ2 + κ3 )(c 3 − c 2 ))
 κ1 κ3
Similarly,
12 4κ T 1 1
∂3 F (c) = ( (∂3 F̄ (c) − ∂1 F̄ (c)) + (∂3 F̄ (c) − ∂2 F̄ (c)))
  κ1 κ2
4κ T 1 1 1 1
= (( + )∂3 F̄ (c) − ∂1 F̄ (c) − ∂2 F̄ (c))
 κ1 κ2 κ1 κ2
4κ T 1 1 1 1
= (( + )(2κ23 c 3 c 22 ) − (c 2 c 3 (κ2 c 2 + κ3 c 3 )) − (2κ23 c 2 c 32 ))
 κ1 κ2 κ1 κ2
4κ T 1 1
= c 2 c 3 ( (κ3 c 2 − κ3 c 3 ) + (κ2 + κ3 )(c 2 − c 3 ))
 κ1 κ2
Therefore
1 12 4κ T 1 1 1 12
∂2 F (c) = c 2 (1 − c 2 )(1 − 2c 2 )( + + )= c 2 (1 − c 2 )(1 − 2c 2 ) (A.7)
κ2   κ2 κ1 κ3 
1 12 4κ T 1 1 1 12
∂3 F (c) = c 3 (1 − c 3 )(1 − 2c 3 )( + + )= c 3 (1 − c 3 )(1 − 2c 3 ) (A.8)
κ3   κ2 κ1 κ3 
Substitute Eq. (A.7) into Eq. (A.5),
∂ c2 3 12 3 24
+ u · ∇ c 2 = M 0 (−  c 2 + c 2 (1 − c 2 )(1 − 2c 2 )) = Ld (−  c 2 + c 2 (1 − c 2 )(1 − 2c 2 ))
∂t 4  2 
Similarly, substitute Eq. (A.8) into Eq. (A.6),
∂ c3 3 24
+ u · ∇ c 3 = Ld (−  c 3 + c 3 (1 − c 3 )(1 − 2c 3 ))
∂t 2 
which implies the three-phase model we proposed is consistent with two-phase model.
 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107 105

A.4. Checking coincidence with two-phase Young’s equation

We discuss the consistency with two-phase Young’s equation on the boundary condition in the following. On the solid
surface, the boundary condition is given by
∂ ci 0
+ u τ · ∂τ c i = − Li (c), on ∂ (A.9)
∂t κi
3
Li (c) = κi ∂n c i + ∂i γ (c), on ∂ (A.10)
4
Assume that there are only two-phases, i.e. c 1 = 0, at the equilibrium state, equations for c 2 and c 3 on the solid boundary
are
3
κ2 ∂n c2 + ∂2 γ (c) = 0 (A.11)
4
3
κ3 ∂n c3 + ∂3 γ (c) = 0 (A.12)
4
Due to the model we proposed,
κT 1 κT 1 1
∂2 γ (c) = (∂i γ̄ (c) − ∂ j γ̄ (c)) = ( (∂2 γ̄ (c) − ∂1 γ̄ (c)) + (∂2 γ̄ (c) − ∂3 γ̄ (c)))
3
j =2
κj 3 κ1 κ3
κT 1 σ w2 κ3 σ w3 κ2 1 1 1
= (− +
6c 2 c 3 ( ) + ( + )6σ w2 c 2 (1 − c 2 ) − 6σ w3 c 3 (1 − c 3 ))
3 κ1 κ2 + κ3 κ2 + κ3 κ1 κ3 κ3
κT 1 κ3 1 1 1 κ2
= (6σ w2 c 2 c 3 (− + + ) − 6σ w3 c 2 c 3 ( + ))
3 κ1 κ2 + κ3 κ1 κ3 κ3 κ1 (κ2 + κ3 )
κ2 κ2 κ2
=( ∂2 γ − ∂3 γ )|c1 =0 = 6c 2 c 3 (σ w2 − σ w3 )
κ2 + κ3 κ2 + κ3 κ2 + κ3
similarly,
κT 1 κT 1 1
∂3 γ (c) = (∂i γ̄ (c) − ∂ j γ̄ (c)) = ( (∂3 γ̄ (c) − ∂1 γ̄ (c)) + (∂3 γ̄ (c) − ∂2 γ̄ (c)))
3
j =3
κj 3 κ1 κ2
κ3 κ3 κ3
= (− ∂2 γ̄ + ∂3 γ̄ )|c1 =0 = − 6c 2 c 3 (σ w2 − σ w3 )
κ2 + κ3 κ2 + κ3 κ2 + κ3
Thus, under the condition of c 1 = 0, i.e. c 2 + c 3 = 1, Eq. (A.11) and Eq. (A.12) will reduce to
3 κ2
κ2 ∂n c2 + 6c 2 c 3 (σ w2 − σ w3 ) = 0 (A.13)
4 κ2 + κ3
3 κ3
κ3 ∂n c3 − 6c 2 c 3 (σ w2 − σ w3 ) = 0 (A.14)
4 κ2 + κ3
Eq. (A.13)–Eq. (A.14), we get
3 (κ2 + κ3 )
 ∂n c 2 + 6c 2 c 3 (σ w2 − σ w3 ) = 0 (A.15)
2 2
which implies
3
σ23 ∂n c2 + 6c2 (1 − c2 )(σ w2 − σ w3 ) = 0 (A.16)
2
Following the derivation mentioned in §2.1, we get

σ23 cos(θ23 ) = σ w3 − σ w2 (A.17)

Appendix B. Derivation of ’s and γ (c)


The idea is to find a revised solid surface free energy functional ∂ γ (c) such that the gradient flow system

∂c ∂γ
= − , on ∂ (B.1)
∂t ∂c
satisfies Eq. (1.1), (P.1) and (P.2). Here
106 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107

⎛ ⎞
1 0 0
⎜ ⎟
=⎝ 0 2 0 ⎠.
0 0 3
To satisfy constraints (P.1) and (P.2), we require
∂γ γ
i ∂ 3
 |c =0 = kj f kj W 3 |ci =0 , for k, j = i , ∀i ∈ {1, 2, 3}. (B.2)
∂c i ∂c
To satisfy Eq. (1.1), it is easy to see that the project operator P 3 is equivalent to introducing another extra term (Lagrange
multiplier) Q (c) in the energy functional, i.e.,
3
i ∂(γ3 (c) + Q 2 (c))
γ (c) = γ3 (c) + Q 2 (c) + (c1 + c2 + c3 − 1) Q (c), where Q (c) = − (B.3)
T ∂ ci
i =1

where  T = 1 + 2 + 3 .
Inspired by Eq. (B.2), it is nature to assume Q 2 (c) has the following form

Q 2 (c 1 , c 2 , c 3 ) = f 23 (c) p 23 (c 1 , c 2 , c 3 ) + f 13 (c) p 13 (c 1 , c 2 , c 3 ) + f 12 (c) p 12 (c 1 , c 2 , c 3 ) (B.4)

where h i j (c 1 , c 2 , c 3 ) are determined as follows. By consistency condition for free energy functional and the definition of f i j
in Eq. (2.26), we have

Q 2 (0, c 2 , c 3 ) = Q 2 (c 1 , 0, c 3 ) = Q 2 (c 1 , c 2 , 0) = 0
which implies

p 23 (0, c 2 , c 3 ) = 0, p 13 (c 1 , 0, c 3 ) = 0, p 12 (c 1 , c 2 , 0) = 0
To simplified the calculations, we further assume that,

p 23 (c 1 , 0, c 3 ) = p 23 (c 1 , c 2 , 0) = 0 (B.5)
p 13 (c 1 , c 2 , 0) = p 13 (0, c 1 , c 3 ) = 0 (B.6)
p 12 (c 1 , 0, c 3 ) = p 12 (0, c 2 , c 3 ) = 0 (B.7)
Therefore, we can take

p 23 (c 1 , c 2 , c 3 ) = c 1 c 2 c 3 q23 (c 1 , c 2 , c 3 ), (B.8)
p 12 (c 1 , c 2 , c 3 ) = c 1 c 2 c 3 q12 (c 1 , c 2 , c 3 ), (B.9)
p 13 (c 1 , c 2 , c 3 ) = c 1 c 2 c 3 q13 (c 1 , c 2 , c 3 ). (B.10)
Taking c 1 = 0, substitute Eq. (B.4) into Eq. (B.2), we have
2 + 3 2 3
1 ( ∂1 (γ3 + Q 2 ) − ∂2 (γ3 + Q 2 ) − ∂2 (γ3 + Q 2 )) |c1 =0 = 0 (B.11)
T T T
1 1 + 3 3
2 (− ∂1 (γ3 + Q 2 ) + ∂2 (γ3 + Q 2 ) − ∂2 (γ3 + Q 2 )) |c1 =0 = γ − ∂3 γ3 ) |c1 =0
23 (∂2 3 (B.12)
T T T
1 2 1 + 2
3 (− ∂1 (γ3 + Q 2 ) − ∂2 (γ3 + Q 2 ) + ∂3 (γ3 + Q 2 )) |c1 =0 = 23 (−∂2 3 γ + ∂3 γ3 ) |c1 =0 (B.13)
T T T
Therefore, from Eq. (B.11), we have
2 σ w2 + 3 σ w3
q23 (c 1 , c 2 , c 3 ) |c1 =0 = 6 (B.14)
2 +  3
In fact, we can also choose h23 (c 1 , c 2 , c 3 ) to be the same function for c 1 = 0. Therefore we have
2 σ w2 + 3 σ w3
q23 (c 1 , c 2 , c 3 ) = 6 . (B.15)
2 +  3
Eq. (B.12) and Eq. (B.13) then give
2 3
6 (σ w2 − σ w3 )c 2 c 3 = 6 23 (σ w2 − σ w3 )c 2 c 3 (B.16)
2 +  3
2 3
−6 (σ w2 − σ w3 )c 2 c 3 = 6 23 (−σ w2 + σ w3 )c 2 c 3 (B.17)
2 +  3
 Q. Zhang, X.-P. Wang / Journal of Computational Physics 319 (2016) 79–107 107

which gives
2 3
= 23 . (B.18)
2 +  3
Similarly, we have
1 σ w1 + 2 σ w2 3 σ w3 + 1 σ w1
q12 (c 1 , c 2 , c 3 ) = 6 , q13 (c 1 , c 2 , c 3 ) = 6
2 +  1 3 +  1
and
1 3  1 2
= 13 , = 12 . (B.19)
1 +  3 2 +  1
Solving 1 , 2 and 3 from Eq. (B.18) and (B.19), we have
1 1 1 1 1
= ( + − )
1 2 13 12 23
1 1 1 1 1
= ( + − )
2 2 12 23 13
1 1 1 1 1
= ( + − )
3 2 13 23 12

In summary,
2 σ w2 + 3 σ w3 3 σ w3 + 1 σ w1 2 σ w2 + 1 σ w1
Q 2 (c) = f 23 (c)6c 1 c 2 c 3 + f 13 (c)6c 1 c 2 c 3 + f 12 (c)6c 1 c 2 c 3
2 +  3 3 +  1 2 +  1

References

[1] D.J. Eyre, An unconditionally stable one-step scheme for gradient systems, Unpublished article, June 1998.
[2] D.J. Eyre, J.W. Bullard, in: Computational and Mathematical Models of Microstructural Evolution, vol. 529, The Materials Research Society, Warrendale,
PA, 1998, pp. 39–46.
[3] F. Boyer, C. Lapuerta, Study of a three component Cahn–Hilliard flow model, Math. Model. Numer. Anal. 40 (2006) 653–687.
[4] H. Garcke, B. Nestler, B. Stoth, On anisotropic order parameter models for multi-phase systems and their sharp interface limits, Physica D 115 (1998)
87–108.
[5] J.-L. Guermond, A. Salgado, A splitting method for incompressible flows with variable density based on a pressure Poisson equation, J. Comput. Phys.
228 (2009) 2834–2846.
[6] J. Shen, X. Yang, A phase-field model and its numerical approximation for two-phase incompressible flows with different densities and viscosities,
SIAM J. Sci. Comput. 32 (3) (2010) 1159–1179.
[7] M. Gao, X.-P. Wang, A gradient stable scheme for a phase field model for the moving contact line problem, J. Comput. Phys. 231 (2012) 1372–1386.
[8] M. Gao, X.-P. Wang, Accurate stable and efficient Scheme for a phase field model for the moving contact line problem with different density and
viscosity, J. Comput. Phys. 272 (2014) 704–718.
[9] P.C. Bollada, P.K. Jimack, A.M. Mullis, A new approach to multi-phase field for the solidification of alloys, IOP Conf. Ser., Mater. Sci. Eng. 33 (2012) 1–8.
[10] Q. Zhang, X.-P. Wang, Equilibrium state and stability of four phase contact line, preprint.
[11] S. Dong, An efficient algorithm for incompressible N-phase flows, J. Comput. Phys. 276 (2014) 691–728.
[12] T.-Z. Qian, X.-P. Wang, P. Sheng, Molecular scale contact line hydrodynamics of immiscible flows, Phys. Rev. E 68 (2003) 016306.
[13] T.-Z. Qian, X.-P. Wang, P. Sheng, Power-law slip profile of the moving contact line in two phase immiscible flows, Phys. Rev. Lett. 63 (2004) 094501.
[14] X.-P. Wang, T.-Z. Qian, P. Sheng, Moving contact line on chemically patterned surfaces, J. Fluid Mech. 605 (2008) 59–78.
[15] X. Xu, X.-P. Wang, Derivation of Wenzel’s and Cassie’s equations from a phase field model for two phase flow on rough surface, SIAM J. Appl. Math. 70
(2010) 2929–2941.
[16] Y. Shi, X.-P. Wang, Modeling and simulation of dynamics of three-component flows on solid surface, Jpn. J. Ind. Appl. Math. 31 (2014) 611–631.
[17] Z. Zhang, S.-X. Xu, W.-Q. Ren, Derivation of a continuum model and the energy law for moving contact lines with insoluble surfactant, Phys. Fluids 26
(2014) 062103.