Journal of Computational Physics - Korea...
Transcript of Journal of Computational Physics - Korea...
Journal of Computational Physics 230 (2011) 7441–7455
Contents lists available at ScienceDirect
Journal of Computational Physics
journal homepage: www.elsevier .com/locate / jcp
A conservative numerical method for the Cahn–Hilliard equation incomplex domains
Jaemin Shin, Darae Jeong, Junseok Kim ⇑Department of Mathematics, Korea University, Seoul 136-701, Republic of Korea
a r t i c l e i n f o a b s t r a c t
Article history:Received 13 January 2011Received in revised form 28 April 2011Accepted 7 June 2011Available online 17 June 2011
Keywords:Cahn–Hilliard equationDegenerate mobilityMultigrid methodPhase separationComplex domain
0021-9991/$ - see front matter � 2011 Elsevier Incdoi:10.1016/j.jcp.2011.06.009
⇑ Corresponding author. Tel.: +82 2 3290 3077; faE-mail addresses: [email protected], cfdkim@dURL: http://math.korea.ac.kr/~cfdkim/ (J. Kim).
We propose an efficient finite difference scheme for solving the Cahn–Hilliard equationwith a variable mobility in complex domains. Our method employs a type of uncondition-ally gradient stable splitting discretization. We also extend the scheme to compute theCahn–Hilliard equation in arbitrarily shaped domains. We prove the mass conservationproperty of the proposed discrete scheme for complex domains. The resulting discretizedequations are solved using a multigrid method. Numerical simulations are presented todemonstrate that the proposed scheme can deal with complex geometries robustly. Fur-thermore, the multigrid efficiency is retained even if the embedded domain is present.
� 2011 Elsevier Inc. All rights reserved.
1. Introduction
In this paper, we consider an efficient finite difference scheme for the Cahn–Hilliard (CH) equation with a variable mobil-ity in complex domains.
@/ðx; tÞ@t
¼ r � ½Mð/ðx; tÞÞrlð/ðx; tÞÞ�; x 2 X; 0 < t 6 T; ð1Þ
lð/ðx; tÞÞ ¼ F 0ð/ðx; tÞÞ � �2D/ðx; tÞ; ð2Þ
where X � Rdðd ¼ 1;2;3Þ is a domain, / = mA �mB is a difference of molar fractions mA and mB of species A and B, and M(/) isa nonnegative diffusional mobility. F(/) is the free energy density of a homogeneous fluid and � is a positive constant. The CHequation has the total free energy functional:
Eð/Þ ¼Z
XFð/Þ þ �
2
2jr/j2
� �dx: ð3Þ
The CH equation was originally introduced to describe the complicated phase separation and coarsening phenomena calledthe spinodal decomposition in binary alloys [5,6]. We also define the quantity c as the molar fraction mA, i.e., c = mA. Note that/ = 2c � 1. Typical forms of the free energy are given as [11]
Fð/Þ ¼ ð1� /2Þ2
4and FðcÞ ¼ c2ð1� cÞ2
4: ð4Þ
. All rights reserved.
x: +82 2 929 8562.ongguk.edu (J. Kim).
−1.5 −1 −0.5 0 0.5 1 1.50
0.05
0.1
0.15
0.2
0.25
φ : concentration
F(φ)
(a)
−1 −0.5 0 0.5 10
0.2
0.4
0.6
0.8
1
φ : concentration
M(φ)
(b)
Fig. 1. (a) Free energy F(/). (b) Variable diffusional mobility M(/).
7442 J. Shin et al. / Journal of Computational Physics 230 (2011) 7441–7455
The free energy has a double-well potential that has local minima at the minimum and maximum of the concentration asshown in Fig. 1(a). And we take variable mobilities of the form
Mð/Þ ¼ 1� /2 and MðcÞ ¼ cð1� cÞ ð5Þ
which are degenerated at the minimum and maximum of the concentration as shown in Fig. 1(b) [7]. This mobility signif-icantly lowers the long-range diffusion across bulk regions with the enhanced diffusion in the interfacial region. The systemis completed by taking an initial condition and the natural and no-flux boundary conditions n�r/ = n�rl = 0 on @X, where nis a unit normal vector to @X.
In the original derivation of the CH equation, a concentration dependent mobility appeared [4–6]. Various numericalmethods have been applied to solve the CH equation with a constant mobility [1,9,12,14,15,17,22,25,27,32]. However, rel-atively few authors have considered the CH equation with a concentration dependent mobility [2,3,16,19–21,26,30,31,33].
In this paper, we propose an efficient finite difference scheme for the CH equation with a variable mobility. Furthermore,we also extend the scheme to compute the CH equation in arbitrarily shaped domains. The application of Cartesian mesh forsolving problems with complex geometry has been used in the past decade. Many authors have studied the Cartesian gridmethod and the finite volume method for solving the Poisson equation with an irregular domain [8,10,18,23,29]. Cartesiangrid embedded boundary methods have advantages over unstructured grid methods because of simpler grid generation.
This paper is organized as follows. In Section 2, we describe the numerical solution of the CH equation with a variablemobility and the details of the treatment of arbitrarily shaped domains. The numerical results showing the robustnessand superiority of the proposed scheme with a variable mobility are described in Section 3. Conclusions are presented inSection 4.
2. Numerical methods
2.1. Numerical discretization
2.1.1. The Cahn–Hilliard equation in a rectangular domainIn this section, we present fully discrete schemes for the CH equation in a rectangular domain. For simplicity of exposi-
tion, we shall discretize the CH Eqs. (1) and (2) in two-dimensional space X = (a,b) � (c,d). Three-dimensional discretizationis analogously defined. Let Nx and Ny be positive even integers, Dx = (b � a)/Nx and D y = (d � c)/Ny be the uniform mesh sizes.We assume h = Dx = D y. Also let Xh = {(xi,yj): xi = a + (i � 0.5)h, yj = c + (j � 0.5)h, 1 6 i 6 Nx,1 6 j 6 Ny} be the set of cell-cen-ters. Let /n
ij and lnij be approximations of /(xi,yj, tn) and l(xi,yj, tn), where tn = nDt and Dt is the time step. We use zero Neu-
mann boundary condition for / and l at the domain boundary, for example
Dx/12;j¼ Dx/Nxþ1
2;j¼ Dy/i;12
¼ Dy/i;Nyþ12¼ 0;
where the discrete differentiation operators are
Dx/iþ12;j¼
/iþ1;j � /ij
hand Dy/i;jþ1
2¼
/i;jþ1 � /ij
h:
And we use the notationrd/ij ¼ Dx/iþ12;j;Dy/i;jþ1
2
� �to represent the discrete gradient of / at cell-edges. Correspondingly, the
divergence at cell-centers, using values from cell-edges, is rd � Dx/iþ12;j;Dy/i;jþ1
2
� �ij¼ Dx/iþ1
2;j� Dx/i�1
2;jþ Dy/i;jþ1
2�
�Dy/i;j�1
2Þ=h.
Then we define the discrete Laplacian by Dd/ij:¼rd�rd/ij and the discrete l2 inner products by
J. Shin et al. / Journal of Computational Physics 230 (2011) 7441–7455 7443
ðc;dÞh :¼ h2XNx
i¼1
XNy
j¼1
cijdij;
rdc;rddð Þe :¼ h2XNx
i¼0
XNy
j¼1
Dxciþ12;j
Dxdiþ12;jþXNx
i¼1
XNy
j¼0
Dyci;jþ12Dydi;jþ1
2
!:
We also define discrete norms as jcj22 = (c,c)h and jcj21 ¼ ðrdc;rdcÞe.In Ref. [20], a Crank–Nicolson (CN) method for the CH equation was proposed as follows:
cnþ1ij � cn
ij
Dt¼ rd � MðcÞnþ
12
ij rdlnþ1
2ij
� �; ð6Þ
lnþ12
ij ¼ 12
F 0ðcnþ1ij Þ þ F 0ðcn
ij� �
� �2
2Ddðcnþ1
ij þ cnijÞ; ð7Þ
where MðcÞnþ12
iþ12;j
:¼ M ðcnij þ cn
iþ1;j þ cnþ1ij þ cnþ1
iþ1;jÞ=4� �
. This CN method has second-order accuracy in time and space. However, it
has a severe time step restriction.In this paper, we present a semi-implicit time and centered difference space discretization of Eqs. (1) and (2). For the
phase-field variable c, we have the following discretization:
cnþ1ij � cn
ij
Dt¼ rd � MðcÞnijrdl
nþ12
ij
� �; ð8Þ
lnþ12
ij ¼ 12
cnþ1ij � �2Ddcnþ1
ij þ F 0ðcnijÞ �
12
cnij ð9Þ
and for the phase-field variable /, we have the following one:
/nþ1ij � /n
ij
Dt¼ rd � Mð/Þnijrdl
nþ12
ij
� �; ð10Þ
lnþ12
ij ¼ 2/nþ1ij � �2Dd/
nþ1ij þ F 0ð/n
ijÞ � 2/nij; ð11Þ
where
rd � Mð/Þnijrdlnþ1
2ij
� �:¼
Mniþ1
2;jlnþ1
2iþ1;j � lnþ1
2ij
� ��Mn
i�12;j
lnþ12
ij � lnþ12
i�1;j
� �h2 þ
Mni;jþ1
2lnþ1
2i;jþ1 � lnþ1
2ij
� ��Mn
i;j�12
lnþ12
ij � lnþ12
i;j�1
� �h2 : ð12Þ
Here Mniþ1
2;j¼ M ð/n
iþ1;j þ /nijÞ=2
� �and the other terms are similarly defined. We note that if the mobility M(/) is constant, then
the discrete schemes (8)–(11) become unconditionally gradient stable linear splitting schemes [13].During computation, the values of the concentration c can be negative or greater than one and also / has a similar case.
For physical relevance of the solution, we want the non-negative mobility (M(�) P 0), therefore we use M(c) = jc(1 � c)j andM(/) = j1 � /2j for numerical evaluations with the concentration dependent mobility.
Fig. 2. Shaded area Xin is embedded in X and Xout is defined as X nXin ¼ X \Xcin .
7444 J. Shin et al. / Journal of Computational Physics 230 (2011) 7441–7455
2.1.2. The Cahn–Hilliard equation in a non-rectangular domainNow, we give a numerical scheme for the CH equation in a non-rectangular domain with homogeneous Neumann bound-
ary conditions. Let Xin be an arbitrarily shaped open domain which is embedded in X. Also, let C = oXin be the arbitrarilyshaped domain boundary (see Fig. 2).
To solve Eqs. (10) and (11) in an arbitrarily shaped domain Xin, we propose the following numerical scheme:
/nþ1ij � /n
ij
Dt¼ rd � GijMð/n
ijÞrdlnþ1
2ij
� �; ð13Þ
lnþ12
ij ¼ 2/nþ1ij � �2rd � Gijrd/
nþ1ij
� �þ F 0ð/n
ijÞ � 2/nij: ð14Þ
Here the boundary control function G is defined in Xh by
Gðx; yÞ ¼1 if ðx; yÞ 2 Xh
in;
0 if ðx; yÞ 2 Xhout ¼ X nXh
in;
(ð15Þ
where Xhin is an open region that consists of interior points of the union of the closed cells whose centers are inside the inter-
face C. For example, in Fig. 3 the value of function G at points (�,�) in Xhout is zero. The value of function G at point (�, �) in Xh
in
is one.As we refine the mesh, i.e., h ? 0, the arbitrarily shaped domain Xin is approximated by Xh
in (see Fig. 4).
2.2. Conservation property of the total mass
In this section, we show that our scheme inherits the conservation of total mass in a complex domain. We want to showthat
Xij2Xin/nþ1
ij ¼X
ij2Xin/n
ij for all n ¼ 0;1; . . . : ð16Þ
We start by summing Eq. (13) in the whole numerical domain X, i.e.,
Xij2X
/nþ1ij � /n
ij
Dt¼ 1
h2
Xij2X
Giþ12;j
Mniþ1
2;jlnþ1
2iþ1;j � lnþ1
2ij
� �� Gi�1
2;jMn
i�12;j
lnþ12
ij � lnþ12
i�1;j
� �h
þGi;jþ12Mn
i;jþ12
lnþ12
i;jþ1 � lnþ12
ij
� �� Gi;j�1
2Mn
i;j�12
lnþ12
ij � lnþ12
i;j�1
� �i
¼ 1
h2
XNy
j¼1GNxþ1
2;jMn
Nxþ12;j
lnþ12
Nxþ1;j � lnþ12
Nx ;j
� �� G1
2;jMn
12;j
lnþ12
1j � lnþ12
0;j
� �h i
þ 1
h2
XNx
i¼1Gi;Nyþ1
2Mn
i;Nyþ12
lnþ12
i;Nyþ1 � lnþ12
i;Ny
� �� Gi;12
Mni;12
lnþ12
i1 � lnþ12
i0
� �h i: ð17Þ
Then, by applying zero Neumann boundary condition of lnþ12
ij on oX, we obtain
Xij2X
/nþ1ij � /n
ij
Dt¼ 0; i:e:;
Xij2X
/nþ1ij ¼
Xij2X
/nij: ð18Þ
Fig. 3. Shaded area Xhin is embedded in X and Xh
out is defined as X \Xhin
c. Note that the value of function G at points (�,�) in Xh
out is zero.
Fig. 4. Convergence of Xhin to Xin as we refine the mesh. Mesh sizes are shown below each figure.
J. Shin et al. / Journal of Computational Physics 230 (2011) 7441–7455 7445
Next, we consider summation of Eq. (13) in the outside numerical domain Xout, i.e.,
Fig. 5.
Xij2Xout
/nþ1ij � /n
ij
Dt¼ 1
h2
Xij2Xout
Giþ12;j
Mniþ1
2;jlnþ1
2iþ1;j � lnþ1
2ij
� �� Gi�1
2;jMn
i�12;j
lnþ12
ij � lnþ12
i�1;j
� �þ Gi;jþ1
2Mn
i;jþ12
lnþ12
i;jþ1 � lnþ12
ij
� �h�Gi;j�1
2Mn
i;j�12
lnþ12
ij � lnþ12
i;j�1
� �i¼ 0; ð19Þ
where we have used the assumption (15) that G is zero in Xout. By using the above property (18), we obtain
Xij2Xin
/nþ1ij ¼
Xij2X
/nij �
Xij2Xout
/nþ1ij ¼
Xij2Xin
/nij þ
Xij2Xout
/nij �
Xij2Xout
/nþ1ij ¼
Xij2Xin
/nij � Dt
Xij2Xout
/nþ1ij � /n
ij
Dt¼X
ij2Xin/n
ij;
where we have used Eq. (19) in the last equality. Therefore, we proved the conservation property of mass, Eq. (16), in a com-plex domain Xin.
2.3. Numerical algorithm
In this section, we show the way to control a matrix in finer and coarser grids with the boundary control function G. Andwe describe the multigrid method and implementation in detail to solve the resulting coupled system of equations.
First we set the boundary control matrix function G as defined in Eq. (15) to represent a complex domain. In Figs. 3 and 5,the cell-edge (�,�) values are used in the smoothing step and the cell-center (�) values are used in the restriction and inter-polation operators. Denote that Gh is a discrete boundary control function defined in Xh. In the coarser grid XH, the controlfunction GH is different from Gh because XH is defined as shown in Fig. 5. The details about the coarser domain XH are ex-plained by the restriction and interpolation operators. Note that as a set we have the following inclusion: Xh
in � XHin.
Next, we describe a full approximation storage (FAS) multigrid method to solve the discrete system (13) and (14) at theimplicit time level. A pointwise Gauss–Seidel relaxation is used as the smoother in the multigrid method. See the referencetext [28] for additional details and backgrounds. The algorithm of the multigrid method for solving the discrete CH system is:First, let us rewrite Eqs. (13) and (14) as follows.
Illustration of the coarsening process for the multigrid method. Shaded area XHin is also embedded in X but is not generally same to Xh
in , i.e., Xhin � XH
in .
7446 J. Shin et al. / Journal of Computational Physics 230 (2011) 7441–7455
Lð/nþ1;lnþ12Þ¼ ðnn;wnÞ; where Lð/nþ1;lnþ1
2Þ
¼ /nþ1ij =Dt�rd � GijMð/n
ijÞrdlnþ1
2ij
� �; �2rd � Gijrd/
nþ1ij
� ��2/nþ1
ij þlnþ12
ij
� �and the source term is ðnn;wnÞ¼ /n
ij=Dt;F 0ð/nijÞ�2/n
ij
� �:
In the following description of one FAS cycle, XH is coarser than Xh by factor 2. Given the number m1 and m2 of pre-and post-smoothing relaxation sweeps, an iteration step for the multigrid method using the V-cycle is formally written as in Fig. 6[28]. In Fig. 6, f/nþ1;m;lnþ1
2;mg and f/̂nþ1;mþ1; l̂nþ12;mþ1g are the approximations of /n+1(xi,yj) and lnþ1
2ðxi; yjÞ before and afteran FAS cycle.
The FAS multigrid algorithm in Fig. 6 consists of the following components.
� Smoothing:
f�/nþ1;m; �lnþ12;mg ¼ SMOOTHm1 ð/nþ1;m;lnþ1
2;m; Lh; nn;wnÞ;
which means performing m1-smoothing steps with the initial approximations /nþ1;m;lnþ12;m, source terms nn, wn, and SMOOTH
relaxation operator to get the approximations �/nþ1;m and �lnþ12;m. One SMOOTH relaxation operator step consists of solving the
system (22) and (23) given below by 2 � 2 matrix inversion for each i and j. Here, we derive the smoothing operator in twodimensions. Rewriting Eq. (13), we get
/nþ1;mij
Dtþ
Giþ12;j
Mniþ1
2;jþ Gi�1
2;jMn
i�12;jþ Gi;jþ1
2Mn
i;jþ12þ Gi;j�1
2Mn
i;j�12
h2 lnþ12;m
ij
¼ nnij þ
Giþ12;j
Mniþ1
2;jlnþ1
2;miþ1;j þ Gi�1
2;jMn
i�12;jlnþ1
2;mi�1;j
h2 þGi;jþ1
2Mn
i;jþ12lnþ1
2;mi;jþ1 þ Gi;j�1
2Mn
i;j�12lnþ1
2;mi;j�1
h2 : ð20Þ
Also, by rewriting Eq. (14), we obtain
� �2
h2 Giþ12;jþGi�1
2;jþGi;jþ1
2þGi;j�1
2
� �þ2
� �/nþ1;m
ij þlnþ12;m
ij ¼wnij��2
h2 Giþ12;j
/nþ1;miþ1;j þGi�1
2;j/nþ1;m
i�1;j þGi;jþ12/nþ1;m
i;jþ1 þGi;j�12/nþ1;m
i;j�1
� �:
ð21Þ
Fig. 6. The FAS multigrid cycle (Xh,XH) two-grid method.
J. Shin et al. / Journal of Computational Physics 230 (2011) 7441–7455 7447
Next, we replace /nþ1;mkl and lnþ1
2;mkl in Eqs. (20) and (21) with �/nþ1;m
kl and �lnþ12;m
kl if k 6 i and l 6 j, otherwise with /nþ1;mkl and
lnþ12;m
kl , i.e.,
�/nþ1;mij
Dtþ
Giþ12;j
Mniþ1
2;jþ Gi�1
2;jMn
i�12;jþ Gi;jþ1
2Mn
i;jþ12þ Gi;j�1
2Mn
i;j�12
h2�lnþ1
2;mij
¼ nnij þ
Giþ12;j
Mniþ1
2;jlnþ1
2;miþ1;j þ Gi�1
2;jMn
i�12;j
�lnþ12;m
i�1;j
h2 þGi;jþ1
2Mn
i;jþ12lnþ1
2;mi;jþ1 þ Gi;j�1
2Mn
i;j�12�lnþ1
2;mi;j�1
h2 : ð22Þ
� �2
h2 Giþ12;jþGi�1
2;jþGi;jþ1
2þGi;j�1
2
� �þ2
� ��/nþ1;m
ij þ �lnþ12;m
ij ¼wnij��2
h2 Giþ12;j
/nþ1;miþ1;j þGi�1
2;j�/nþ1;m
i�1;j þGi;jþ12/nþ1;m
i;jþ1 þGi;j�12
�/nþ1;mi;j�1
� �:
ð23Þ
H
� Restriction: A restriction operator Ih is defined asdHij ¼ IH
h dhij ¼
1Rij
Gh2i�1;2j�1dh
2i�1;2j�1 þ Gh2i�1;2jd
h2i�1;2j þ Gh
2i;2j�1dh2i;2j�1 þ Gh
2i;2jdh2i;2j
� �;
where Rij ¼ Gh2i�1;2j�1 þ Gh
2i�1;2j þ Gh2i;2j�1 þ Gh
2i;2j and if Rij = 0, then we define dHij ¼ 0 (see Fig. 7(a)). For example, Fig. 7(c) shows
notations and symbols on fine and coarse grids. Restriction from level-h to level-H is done as follows:
dH11 ¼
1R11
Gh11dh
11 þ Gh12dh
12 þ Gh21dh
21 þ Gh22dh
22
� �¼ 1
4dh
11 þ dh12 þ dh
21 þ dh22
� �;
dH12 ¼ 0; dH
21 ¼13
dh31 þ dh
32 þ dh34
� �; and dH
22 ¼12
dh33 þ dh
43
� �:
� Interpolation: The interpolation operator IhH (see Fig. 7(b)) represents the prolongation of corrections from level-H to
level-h and is defined as
dh2i�1;2j�1
dh2i�1;2j
dh2i;2j�1
dh2i;2j
8>>>>>><>>>>>>:
9>>>>>>=>>>>>>;¼ Ih
HdHij ¼
Gh2i�1;2j�1dH
ij for }
Gh2i�1;2jd
Hij for
Gh2i;2j�1dH
ij for M
Gh2i;2jd
Hij for � :
8>>>>><>>>>>:
ð24Þ
For example, interpolation from level-H and level-h is done as follows (see Fig. 7(c)):
dh11
dh12
dh21
dh22
8>>>>>><>>>>>>:
9>>>>>>=>>>>>>;¼
Gh11dH
11
Gh12dH
11
Gh21dH
11
Gh22dH
11
8>>>>>><>>>>>>:
9>>>>>>=>>>>>>;¼
dH11
dH11
dH11
dH11
8>>>>>><>>>>>>:
9>>>>>>=>>>>>>;;
dh13
dh14
dh23
dh24
8>>>>>><>>>>>>:
9>>>>>>=>>>>>>;¼
Gh13dH
12
Gh14dH
12
Gh23dH
12
Gh24dH
12
8>>>>>><>>>>>>:
9>>>>>>=>>>>>>;¼
0
0
0
0
8>>>><>>>>:
9>>>>=>>>>;;
dh31
dh32
dh41
dh42
8>>>>>><>>>>>>:
9>>>>>>=>>>>>>;¼
Gh31dH
21
Gh32dH
21
Gh41dH
21
Gh42dH
21
8>>>>>><>>>>>>:
9>>>>>>=>>>>>>;¼
dH21
dH21
0
dH21
8>>>>><>>>>>:
9>>>>>=>>>>>;;
dh33
dh34
dh43
dh44
8>>>>>><>>>>>>:
9>>>>>>=>>>>>>;¼
Gh33dH
22
Gh34dH
22
Gh43dH
22
Gh44dH
22
8>>>>>><>>>>>>:
9>>>>>>=>>>>>>;¼
dH22
0
0
dH22
8>>>>><>>>>>:
9>>>>>=>>>>>;:
By the above restriction and interpolation, the proposed discretization has an effect that none of grid points in Xout partic-ipates in the evolution of values in Xin.
3. Numerical experiments
In this section, we perform the following tests: showing the relation between the � value and the width of the transitionlayer, calculating the accuracy of the proposed method, demonstrating the multigrid efficiency, comparing stabilities of theCrank–Nicolson and the proposed scheme, investigating the effect of constant and variable mobilities in 2D and 3D, validat-ing the total mass conservation and the total energy dissipation, and spinodal decomposition on complex domains such as a
Fig. 7. (a) Restriction. (b) Interpolation. (c) Gh = 1 on gray region and Gh = 0 on white region.
7448 J. Shin et al. / Journal of Computational Physics 230 (2011) 7441–7455
wavy curved channel, sequential T-junctions, a sphere, and a wavy curved tube. Note that we will use either / or c in ournumerical experiments because two equations with different phase-field variables are equivalent.
3.1. The relation between the � value and the width of the transition layer
In the first numerical experiment, we consider the relation between the � value and the width of the transition layer forthe CH equation. Across the interfacial region, the concentration field varies from 0.05 to 0.95 over a distance of approxi-mately 4
ffiffiffi2p
tanh�1ð0:9Þ for c and varies from �0.9 to 0.9 over a distance of approximately 2ffiffiffi2p
tanh�1ð0:9Þ for /. Therefore,if we want this value to be approximately m grid points, the � value needs to be taken as follows, respectively:
�m ¼mh
4ffiffiffi2p
tanh�1ð0:9Þfor c and �m ¼
mh
2ffiffiffi2p
tanh�1ð0:9Þfor /:
To confirm those formula, we run simulations with the initial data /(x,y,0) = �0.4 + 0.001rand (x,y) in a domain X = (0,64)� (0,64) with a 64 � 64 mesh and h = Dt = 1. Here, the random number, rand (x,y), is uniformly distributed between �1 and1. Shaded region indicates the embedded domain. Fig. 8(a) and (b) show contours from / = �0.9 to / = 0.9 increased by 0.2with �3 and �10 at t = 5000, respectively. From the two figures, we can see that the interfacial transition layer is about 3 and10 grid points, respectively.
3.2. Convergence test
We should demonstrate the convergence of the proposed scheme. The initial state for the test is /(x,y,0) = 0.8cos(2px)-cos(2p y) in a domain Xin = (�1,1) � (�1,1) � (�0.5,0.5) � (�0.5,0.5) (see Fig. 9). To calculate the accuracy of the proposedmethod, we perform a number of simulations on a set of increasingly finer grids. The numerical solutions are computed onthe grids h = 1/2n+1 for n = 5, 6, 7, 8, and 9. For each case, the calculation is run up to time t = 0.01 with the time stepD t = 0.001/4n�5 and a fixed � = 0.04. We define the error of a grid to be the discrete l2-norm of the difference between thatgrid and the average of the next finer grid cells covering it as follows. eh=h
2 ij:¼ /hij � ð/h
22i;2jþ /h
22i�1;2jþ /h
22i;2j�1þ /h
22i�1;2j�1Þ=4.
Fig. 8. The concentration contours at the nine levels from / = �0.9 to / = 0.9 increased by 0.2 with (a) �3 and (b) �10.
Fig. 9. The initial condition /(x,y,0) with a small obstacle on X = (�1,1) � (�1,1).
Table 1The errors and rates of convergence for concentration /.
Case 32–64 Rate 64–128 Rate 128–256 Rate 256–512
l2-error 1.44E�2 1.89 3.89 E�3 2.12 8.94E�4 2.05 2.16E�4
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
10−6
10−4
10−2
100
102
fine grid iterations
resi
dual
1 level2 levels3 levels4 levels
Fig. 10. The number of fine grid iterations needed to converge to a residual norm of 1.0E-6.
J. Shin et al. / Journal of Computational Physics 230 (2011) 7441–7455 7449
Table 2The number of the fine grid iterations(FGI) for each level.
Level 1 2 3 4 5 6 7 8
FGI 9407 3289 809 221 65 29 29 29
Table 3Stability constraint of Dt for the CN scheme and the proposed scheme.
Mesh size Scheme
Crank–Nicolson scheme Proposed scheme
h = 1/64 Dt 6 0.001120 Dt 6 0.777h = 1/128 Dt 6 0.000270 Dt 6 0.914h = 1/256 Dt 6 0.000672 Dt 6 1.140h = 1/512 Dt 6 0.000166 Dt 6 0.879
7450 J. Shin et al. / Journal of Computational Physics 230 (2011) 7441–7455
And the rate of convergence is defined as the ratio of successive errors as log2ðkeh=h2k2=keh
2=h4k2Þ. Then Table 1 shows the errors
and rates of convergence. The results suggest that the scheme is first-order accurate in time and second-order in space as weexpect from the discretization.
3.3. Multigrid efficiency
To demonstrate the efficiency of the multigrid method, we perform the similar simulation as in [34]. The initial state andcomputational domain are same to the convergence test case. The parameters are �3, h = 1/256, Dt = 1.0h, and 512 � 512mesh grid. All available vcycle levels are considered. Fig. 10 shows the maximum value of residual versus the number of finegrid iterations needed to converge to a residual norm of 1.0E-6 (we only plot four different levels for clarity). And in Table 2the number of the fine grid iterations is listed for all levels. Significant improvement in the convergence is observed as weincrease the vcycle level.
3.4. Stability comparison of two schemes
We compare the stability constraints for the CN scheme (6) and (7) and the proposed scheme (8) and (9). For eachscheme, we calculate the maximum Dt so that stable solutions can be computed up to T = 100Dt. The initial concentrationis taken to be c(x,y,0) = 0.5 + 0.5rand (x,y) in the computational domain X = (0,1) � (0,1). Table 3 shows the results of thestability constraints of Dt with two schemes. We can see that the proposed scheme can use larger time steps. If the timestep is reasonably small, then the stability of the proposed scheme seems independent of space step size.
3.5. The effect of constant and variable mobilities
3.5.1. Two-dimensional caseWe examine the evolution of a random perturbation with the small magnitude about a mean composition and also con-
sider numerical experiments highlighting the difference between the variable mobility M(c) = c(1 � c) and the constantmobility M(c) 0.25 which is the maximum value of the variable mobility M(c). The initial condition is taken to bec(x,y,0) = 0.25 + 0.001rand (x,y) in the computational domain X = (0,1) � (0,1). And we take the simulation parameters:�4, h = 1/128, and Dt = 0.5h. We stop the numerical computations when the discrete l2 -norm of the difference between(n + 1)th andnth time step solutions becomes less than 10�6, that is, kcn+1 � cnk2 6 10�6. In the constant mobility case,we start the simulation with c(x,y,7.813) from the variable mobility case.
In Fig. 11, the first row and the second row represent time evolutions of phase-fields with the constant and the variablemobilities, respectively. The final numerical solutions are stationary numerical solutions according to the stopping criteria.In the bottom row of figure, dark regions are nucleated. As the remaining regions grow, a numerical equilibrium state isreached. The variable mobility generally reduces the bulk diffusion, which corresponds to the interface-diffusion-controlleddynamics, i.e., diffusion process takes place along the interface between the two phases. This fact is made clear by comparingthe results with the constant mobility. In the case of the constant mobility, the evolution leads to a single disk.
3.5.2. Three-dimensional caseNext, we investigate the effect of constant and variable mobilities on the three-dimensional domain. We simulate the
evolution on a 643 mesh and use the following parameters: � = 0.01, h = 1/64, and Dt = h. The initial condition is taken tobe c(x,y,z,0) = 0.25 + 0.2rand (x,y,z) in the computational domain X = (0,1) � (0,1) � (0,1). In the constant mobility case,we start the simulation with c(x,y,z,62.5) from the variable mobility case. In Fig. 12, evolutions of the concentration c(x,y,z, t)
Fig. 11. Evolution of the concentration c(x,y, t) with a constant mobility M(c) 0.25 (the top row) and a variable mobility M(c) = c(1 � c) (the bottom row).The times are shown below each figure.
Fig. 12. Evolution of the concentration c(x,y,z, t) with a constant mobility M(c) 0.25 (the top row) and a variable mobility M(c) = c(1 � c) (the bottom row).The times are shown below each figure.
J. Shin et al. / Journal of Computational Physics 230 (2011) 7441–7455 7451
with a constant mobility M(c) 0.25 (the top row) and a variable mobility M(c) = c(1 � c) (the bottom row) are shown. In thecase of the variable mobility, two nucleated components are the numerical steady state solution. However, in the case of theconstant mobility (the top row), only one component is shown.
3.6. The decrease of the total energy and the conservation of the average concentration
3.6.1. Rectangular domain caseIn order to demonstrate that the numerical scheme inherits the energy decreasing property, we consider the evolution of
the discrete total energy. We define a discrete energy functional by
EhðcnÞ ¼ ðFðcnÞ;1Þh þ�2
2jcnj21: ð25Þ
The initial state is taken to be c(x,y,0) = 0.25 + 0.001rand (x,y) in the computational domain X = (0,1) � (0,1). Parametersh = 1/128, �4, and Dt = 0.1 are taken. In Fig. 13, the time evolution of the non-dimensional discrete total energyEhðcnÞ=Ehðc0Þ (solid line) is shown. We can see that the total discrete energy is non-increasing and the average concentrationis preserved. And this numerical result agrees well with the total energy dissipation property. Also, the inscribed small fig-ures are the concentration fields at the indicated times.
0 1000 2000 3000 40000
0.2
0.4
0.6
0.8
1
Time
total energymass
Fig. 13. The temporal evolution of non-dimensional discrete total energy EhðcnÞ=Ehðc0Þ (solid line) of the numerical solutions with the initial condition,c(x,y,0) = 0.25 + 0.001 rand (x,y).
7452 J. Shin et al. / Journal of Computational Physics 230 (2011) 7441–7455
3.6.2. Non-rectangular domain caseNow, we apply the numerical scheme for arbitrarily domains. The initial state is taken to be /(x,y,0) = 0.25 + 0.001rand
(x,y) in the computational domain X = (0,1) � (0,1) with the mesh size 256 � 256 and the temporal step size Dt = 0.1/256.The interface C is a circle whose radius is 0.45 and center is located at (0.5,0.5). Xin is the enclosed domain by the interfaceC. To demonstrate the discrete energy dissipation property in an arbitrary domain, we define a discrete energy functional by
Fig. 14(x,y,0)
Ehð/nÞ ¼ ðGFð/nÞ;1Þh þ�2
2jG/nj21: ð26Þ
Fig. 14 shows that the discrete total energy is monotonically decreasing and the average concentration is conserved. The in-scribed small figures are the evolution of the concentration field.
3.7. Evolution in complex domains
3.7.1. Spinodal decomposition in a wavy curved channelWe also apply our proposed scheme to a more complicated domain. First, we define the interface C which has the upper
bound u(x) and the lower bound l(x) in the computational domain X = (0,4) � (0,1). That is,
uðxÞ ¼ 0:05ffiffiffixp
sinð3pxÞ þ 0:75; ð27ÞlðxÞ ¼ 0:01x2 cosð2pxÞ � 0:04 sinð6pxÞ þ 0:3: ð28Þ
We simulate the evolution on a 1024 � 256 mesh size. The spatial size h = 1/256, the temporal size Dt = 0.5h, and �4 are used.The initial condition is taken to be /(x,y,0) = �0.35 + 0.1rand (x,y). Fig. 15 shows the evolution of the concentration /(x,y, t)
0 20 40 60 80 1000
0.2
0.4
0.6
0.8
1
Time
total energymass
. The time dependent non-dimensional discrete total energy Ehð/nÞ=Ehð/0Þ (solid line) of the numerical solutions with the initial condition, /= 0.25 + 0.001 rand (x,y).
0 0.5 1 1.5 2 2.5 3 3.5 40
0.5
1
(a)
0 0.5 1 1.5 2 2.5 3 3.5 40
0.5
1
(b)
0 0.5 1 1.5 2 2.5 3 3.5 40
0.5
1
(c)
Fig. 15. Evolution of the concentration /(x,y, t) with a variable mobility M(/) = 1 � /2 in two-dimensional domain with the boundary control function G.The times of each figure are (a) t = 0.98, (b) t = 3.91, and (c) t = 9.77.
Fig. 16. Spinodal decomposition in sequential T-junctions.
J. Shin et al. / Journal of Computational Physics 230 (2011) 7441–7455 7453
with a variable mobility M(/) = 1 � /2 and the boundary control function G. The times of each figure are (a) t = 0.98, (b)t = 3.91, and (c) t = 9.77.
3.7.2. Spinodal decomposition in sequential T-junctionsMicrofluidic technology offers capabilities for the precise handling of small fluid volumes dispersed as droplets. Sequen-
tial application of the geometrically mediated T-junction is used to facilitate the precise conversion of large initial slugs ofthe dispersed phase into small droplets comparable in size to the channel [24]. We test our scheme on sequential T-junc-
Fig. 17. Evolution of the concentration c(x,y,z, t) with an initial condition c(x,y,z,0) = 0.25 + 0.2rand (x,y,z). The times are shown below each figure.
Fig. 18. Evolution of the concentration c(x,y,z, t) with a variable mobility M(c) = c(1 � c) in a three-dimensional domain. The times are shown below eachfigure.
7454 J. Shin et al. / Journal of Computational Physics 230 (2011) 7441–7455
tions. The numerical parameters are h = 1/512, Dt = 0.1h, and �4. The initial state is /(x,y,0) = �0.4 + 0.001rand (x,y) in thecomputational domain X = (0,3) � (0,2). In Fig. 16, we see spinodal decomposition in sequential T-junctions at timet = 0.0098.
3.7.3. Spinodal decomposition in a sphereWe extend the numerical experiment to the three-dimensional space. The surface C is a sphere whose radius is 0.45 and
center is (0.5, 0.5, 0.5). In this computation, we choose �4, h = 1/64, and Dt = h. The initial condition is taken to bec(x,y,z,0) = 0.25 + 0.2rand (x,y,z) on 643 mesh grid. In Fig. 17, Xin is inside mesh. For visibility, we plot only the longitudeand latitude of the boundary of the domain Xin. This simulation also represents that the separation and small nucleationoccur at an early stage (t = 2.34). After this initial phase separation, the coarsening process is developed slowly(t = 96.88,312.5).
3.7.4. Spinodal decomposition in a wavy curved tubeNext, for a more complex domain simulation in 3D, we use the following surface:
rðh; zÞ ¼ ð0:01z2 cosð2pzÞ � 0:04 sinð6pzÞ þ 0:3Þð0:9� 0:3 cosðhþ pzÞÞ for 0 6 h 6 2p; 2 6 z 6 4;
where r, h, and z are distance from, azimuthal angle about, and distance along axis, respectively (see Fig. 18 (a)). The numer-ical parameters are 128 � 128 � 256 mesh size, h = 1/128, Dt = h, and �3. The initial condition is c(x,y,z,0) = 0.25 + 0.002rand(x,y,z) in the computational domain X = (�0.5,0.5) � (�0.5,0.5) � (2,4). Fig. 18(b)–(d) show the evolution of the concentra-tion c(x,y,z, t) with a variable mobility M(c) = 1 � c2. The times are shown below each figure.
J. Shin et al. / Journal of Computational Physics 230 (2011) 7441–7455 7455
4. Conclusions
We considered an efficient finite difference scheme for the Cahn–Hilliard equation with a variable mobility of a model forphase separation in a binary mixture. The numerical method is based on a type of the unconditionally gradient stable split-ting discretization. We also extended the scheme to compute the Cahn–Hilliard equation in complex domains. And weproved the mass conservation property of the proposed discrete scheme for complex domains. The resulting discretizedequations are solved by a multigrid method. Numerical simulations were presented to show the robustness and the supe-riority of the proposed scheme compared to other Crank–Nicolson scheme. The most salient feature is that although thealgorithm is simple, it is applicable to many complex domains. Also, the algorithm is well suited to the multigrid method.The method described here is a specific application of a general formulism for constructing conservative finite differencemethods for problems with complex embedded domains. Therefore, it is applicable to other partial differential equationsin irregular domains.
Acknowledgments
This research was supported by Basic Science Research Program through the National Research Foundation of Korea(NRF)funded by the Ministry of Education, Science and Technology (No. 2010-0003989). The authors also wish to thank thereviewers for the constructive and helpful comments on the revision of this article.
References
[1] F.S. Bai, A. Spence, A.M. Stuart, Numerical computations of coarsening in the one-dimensional Cahn–Hilliard model of phase separation, Physica D 78(1994) 155–165.
[2] J.W. Barrett, J.F. Blowey, Finite element approximation of the Cahn–Hilliard equation with concentration dependent mobility, Math. Comput. 68 (1999)487–517.
[3] J.W. Barrett, J.F. Blowey, H. Garcke, Finite element approximation of the Cahn–Hilliard equation with degenerate mobility, SIAM J. Numer. Anal. 37(1999) 286–318.
[4] J.W. Cahn, C.M. Elliott, A. Novick-Cohen, The Cahn–Hilliard equation with a concentration dependent mobility: Motion by minus the Laplacian of themean curvature, Euro. J. Appl. Math. 7 (1996) 287–301.
[5] J.W. Cahn, J.E. Hilliard, Free energy of a non-uniform system. I. Interfacial free energy, J. Chem. Phys. 28 (1958) 258–267.[6] J.W. Cahn, J.E. Hilliard, Spinodal decomposition: a reprise, Acta Metall. 19 (1971) 151–161.[7] J.W. Cahn, J.E. Taylor, Surface motion by surface diffusion, Acta Metall. 42 (1994) 1045–1063.[8] D. Calhoun, A cartesian grid method for solving the two-dimensional streamfunction-vorticity equations in irregular regions, J. Comput. Phys. 176 (2)
(2002) 231–275.[9] M. Copetti, C.M. Elliott, Kinetics of phase decomposition processes: numerical solutions to the Cahn–Hilliard equation, Mater. Sci. Technol. 6 (1990)
273–283.[10] M.S. Day, P. Colella, M.J. Lijewski, C.A. Rendleman, D.L. Marcus, Embedded boundary algorithms for solving the Poisson equation on complex domains,
Technical Report LBNL-41811, Lawrence Berkeley National Laboratory, May 1998.[11] J.J. Eggleston, G.B. McFadden, P.W. Voorhees, A phase-field model for highly anisotropic interfacial energy, Physica D 150 (2001) 91–103.[12] C.M. Elliott, D.A. French, Numerical studies of the Cahn–Hilliard equation for phase separation, IMA J. Appl. Math. 38 (1987) 97–128.[13] D.J. Eyre, An unconditionally stable one-step scheme for gradient systems, Preprint (1997), <http://www.math.utah.edu/�eyre/research/methods/
stable.ps>.[14] D. Furihata, A stable and conservative finite difference scheme for the Cahn–Hilliard Equation, Numer. Math. 87 (2001) 675–699.[15] D. Furihata, T. Matsuo, A stable, convergent, conservative and linear finite difference scheme for the Cahn–Hilliard equation, Jpn. J. Indust. Appl. Math.
20 (2003) 65–85.[16] H. Gomez, V. Calo, Y. Bazilevs, T.J.R. Hughes, Isogeometric analysis of the Cahn–Hilliard phase field model, Comput. Methods Appl. Mech. Eng. 197
(2008) 4333–4352.[17] Y. He, Y. Liu, T. Tang, On large time-stepping methods for the Cahn–Hilliard equation, Appl. Numer. Math. 57 (2007) 616–628.[18] H. Johansen, P. Colella, A cartesian grid embedded boundary method for poissons equation on irregular domains, J. Comput. Phys. 147 1 (1998) 60–85.[19] D. Kay, R. Welford, A multigrid finite element solver for the Cahn–Hilliard equation, J. Comput. Phys. 212 (2006) 288–304.[20] J.S. Kim, A numerical method for the Cahn–Hilliard equation with a variable mobility, Commun. Nonlinear Sci. Numer. Simul. 12 (2007) 1560–1571.[21] J.S. Kim, K. Kang, A numerical method for the ternary Cahn–Hilliard system with a degenerate mobility, Appl. Numer. Math. 59 (2009) 1029–1042.[22] J.S. Kim, K. Kang, J.S. Lowengrub, Conservative multigrid methods for Cahn–Hilliard fluids, J. Comput. Phys. 193 (2004) 511–543.[23] R.J. LeVeque, Z. Li, The immersed interface method for elliptic equations with discontinuous coefficients and singular sources, SIAM J. Numer. Anal. 31
(1994) 1019–1044.[24] D.R. Link, S.L. Anna, D.A. Weitz, H.A. Stone, Geometrically mediated breakup of drops in microfluidic devices, Phys. Rev. Lett. 92 (2004) 054503.[25] E.V.L. Mello, O.T.S. Filho, Numerical study of the Cahn–Hilliard equation in one, two and three dimensions, Physica A 347 (2005) 429–443.[26] G. Sheng, T. Wang, Q. Du, K.G. Wang, Z.K. Liu, L.Q. Chen, Coarsening kinetics of a two phase mixture with highly disparate diffusion mobility, Comput.
Phys. 8 (2010) 249–264.[27] Z.Z. Sun, A second order accurate linearized difference scheme for the two-dimensional Cahn–Hilliard equation, Math. Comput. 64 (1995) 1463–1471.[28] U. Trottenberg, C. Oosterlee, A. Schüller, Multigrid, Academic press, London, 2001.[29] H.S. Udaykumar, R. Mittal, P. Rampunggoon, A. Khanna, A sharp interface Cartesian grid method for simulating flows with complex moving boundaries,
J. Comput. Phys. 174 (2001) 345–380.[30] G.N. Wells, E. Kuhl, K. Garikipati, A discontinuous Galerkin method for the Cahn–Hilliard equation, J. Comput. Phys. 218 (2006) 860–877.[31] Y. Xia, Y. Xu, C.W. Shu, Local discontinuous Galerkin methods for the Cahn–Hilliard type equations, J. Comput. Phys. 227 (2007) 472–491.[32] X. Ye, The Legendre collocation method for the Cahn–Hilliard equation, J. Comput. Appl. Math. 150 (2003) 87–108.[33] J. Zhu, L.Q. Chen, J. Shen, V. Tikare, Coarsening kinetics from a variable mobility Cahn–Hilliard equation – application of semi-implicit Fourier spectral
method, Phys. Rev. E 60 (1999) 3564–3572.[34] M. Francois, W. Shyy, Computations of drop dynamics with the immersed boundary method, Numer. Heat Transf. B 44 (2003) 101–118.