Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

Available online at www.sciencedirect.

com
ScienceDirect

Mathematics and Computers in Simulation 182 (2021) 535–554


www.elsevier.com/locate/matcom

Original articles

Tailored finite point method for the approximation of diffusion


operators with non-symmetric diffusion tensor✩
Yihong Wang
School of Statistics and Mathematics, Shanghai Lixin University of Accounting and Finance, Shanghai, 201209, China
Received 19 December 2019; received in revised form 28 October 2020; accepted 14 November 2020
Available online 21 November 2020

Abstract
We present a tailored finite point method (TFPM) for anisotropic diffusion equations with a non-symmetric diffusion tensor
on Cartesian grids. The fluxes on each edge are discretized by using a linear combination of the local basis functions, which
come from the exact solution of the diffusion equation with constant coefficients on the local cell. In this way, the scheme
is fully consistent and the flux is naturally continuous across the interfaces between the subdomains with a non-symmetric
diffusion tensor. Additionally, it is convenient to handle the Neumann boundary condition or a variant of Neumann boundary
condition. Numerical results obtained from solving different anisotropic diffusion problems including problems with sharp
discontinuity near the interface and boundary, show that this approach is efficient. Second order convergence rate can be
obtained with the numerical examples. The new scheme is also tested on a time-dependent problem modeling the Hall effect
in the resistive magnetohydrodynamics, the results further show the robustness of this new method.
⃝c 2020 International Association for Mathematics and Computers in Simulation (IMACS). Published by Elsevier B.V. All rights
reserved.
Keywords: Non-symmetric diffusion tensor; Tailored finite point method; Discontinuous diffusivity; Interface layer

1. Introduction
Although the diffusion tensor is symmetric (even isotropic) in most diffusion problem, a non-symmetric
diffusion tensor has attracted increasing attention because it is frequently encountered in important models of
mathematical physics [6,13]. For example, it is used to represent eddy current induced transport phenomena in
ocean modeling [10,11] and in magnetohydrodynamics [25,30]. In this paper, we consider a steady state diffusion
problem,
−∇ · (K ∇u) = f, in Ω ,






u = u D, on Γ D (1)



n · K ∇u = u N , on Γ N .

where the following assumptions hold:


✩ Project supported by NSFC, China 11901393 and Natural Science Fund of Shanghai, China under the grant 19ZR1436300.
E-mail address: wyhlx2017@163.com.

https://doi.org/10.1016/j.matcom.2020.11.020
0378-4754/⃝ c 2020 International Association for Mathematics and Computers in Simulation (IMACS). Published by Elsevier B.V. All rights
reserved.
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

• Ω is an open subset of R 2 subject to the boundary conditions on ∂Ω = Γ D Γ N .



• The diffusion (or permeability) tensor K : Ω −→ R 2×2 may not be symmetric, yet it is assumed to be strongly
elliptic. There exist positive constants ζ1 and ζ2 satisfying ζ1 ∥ x ∥2 ≤ x ⊤ K x ≤ ζ2 ∥ x ∥2 , for any x ∈ R 2 .
• The function f is the source term and belongs to L 2 (Ω ).
The last decade has seen a development of a large number of discretization methods for (1). Finite volume
methods (FVM) [22,24,27,31,33] are the most popular ones, the most recent FVM are not only stable and accurate
for discontinuous and anisotropic diffusion tensors but also satisfying local mass conservation and discrete maximum
principle. There are three families of finite volume method schemes [1,2,4,5,7,8]: Multi Point Flux Approximation
methods, Hybrid Mimetic Mixed methods and Discrete Duality Finite Volume methods. The strengths of these
methods have been reviewed by Droniou in [4]. The recently developed other methods include finite difference
method [18], finite element method [26], discontinuous Galerkin method [3,9] and so on.
The traditional finite difference methods are not always valid for discontinuous diffusion problems and mimetic
finite difference method (MFDM) [12–14,20,21,23] can possess excellent robustness in the presence of severe jumps
in coefficients on general polygonal meshes and mimics important properties of the mathematical and physical
systems. Specifically, MFDM is constructed by starting from some differential operators (for example, divergence
or gradient), which can ensure some features of the operators at the continuous level. One of the main design
principles is the constructive use of the discrete dual relation.
In this paper, different from MFDM, we adopt the analytical properties of the local solutions of the considered
equation at the continuous level to develop an efficient TFPM. The main idea of TFPM is to firstly approximate
the diffusion coefficient near each grid point by a constant, then use solutions that satisfy exactly the homogeneous
equation with constant coefficients as local basis functions to construct a finite difference scheme. The weights of
a finite difference scheme are tailored to the exact solutions of the homogeneous governing equation. Since TFPM
makes full use of analytical properties of local solutions, it is capable of resolving high gradients near the layer re-
gions even with coarse meshes. Han, Huang and Kellogg first proposed TFPM for solving singular perturbed elliptic
equations in 2008 [16]. This method has been shown effectiveness in resolving convection–diffusion, convection–
diffusion–reaction, the non-homogeneous reaction–diffusion, transport systems, non-equilibrium radiation diffusion
equations and so on [15,17,19,28,29].
Our goal is to give a further study of the TFPM for solving diffusion operators with non-symmetric diffusion
tensor on Cartesian grids. The key point of the new scheme construction is the discretization of the flux across cell
edges. Let Γ is the common edge of two adjacent cell. On the common edge Γ , flux continuity can be written as
n · K + ∇u + = n · K − ∇u − ,
where u ± are localized solutions inside the two cells and n is the normal direction of Γ . Since u ± are approximated
by a linear combination of local solutions, so are n · K ± ∇u ± , thus a discrete analog of flux continuity equation can
be written down as well.
The main advantages of our proposed approach are that: (1) The normal flux is approximated by a linear
combination of local basis functions, no additional values to compute approximate gradients, no modifications to
meet with Neumann boundary condition and a variant of Neumann boundary condition. (2) The constructive use of
interface condition is one of the major design principle of this scheme. Since the normal flux is approximated by a
linear combination of local basis functions, it allows to meet with the discontinuity which is parallel to the common
edge Γ of two adjacent cell or aligned with the flux quadrature point. (3) The scheme can correctly reproduce the
fast change at the interface and boundary layer even the mesh size is much larger than the thickness of the boundary
layers or the interior layers.
This paper is organized as follows. The problem statement and the main idea of TFPM are given in Section 2.
The TFPM for diffusion operators with non-symmetric diffusion tensor is derived in Section 3. The discretization
of Neumann and a variant of Neumann boundary conditions are also provided. The scheme applying to parabolic
problem is presented briefly in Section 4. Some numerical experiments include problems with sharp discontinuity
near the interface and boundary and a time-dependent problem modeling the Hall effect in the resistive magneto-
hydrodynamics to illustrate the capabilities of the new method in Section 5. The magnetohydrodynamics problem
considered in this section, a variant of Neumann boundary condition is specified on the boundary. Finally, the
concluding remarks are drawn in Section 6.
536
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

2. Problem statement
2.1. Problem statement

Let Q is the unit diffusive flux vector. We can rewrite the original equation (1) in a mixed form:
∇ · Q = f, in Ω ,
{
(2)
−K ∇u = Q.
subject to the boundary conditions:
u = uD on ΓD , and n · K ∇u = u N on ΓN . (3)
We will also consider a variant of the Neumann boundary condition takes the following form:
n · ∇u = u N on ΓN . (4)

2.2. Antisymmetric diffusion as advection

Suppose the diffusion tensor K can be full tensor with non-symmetric matrix
( )
K 11 K 12
K = . (5)
K 21 K 22
The unit diffusive flux vector Q is naturally written as:
⎛ ⎞
K 11 ∂∂ux + K 12 ∂u
∂y ⎠
Q = −⎝ . (6)
∂u ∂u
K 21 ∂ x + K 22 ∂ y
Then Eq. (1) can be rewritten as
−K 11 u x x − K 11 u yy − (K 12 + K 21 )u x y − ((K 11 )x + (K 12 ) y )u x − ((K 21 )x + (K 22 ) y )u y = f. (7)
When K 21 = −K 12 , K is an antisymmetric matrix. A link between the antisymmetric part of the diffusion tensor
and the advection operator has been established in [13]. Let K α denote the antisymmetric part of K:
( )
1 0 K 12
K α := (K − K ⊤ ) = . (8)
2 −K 12 0
Consider the term
∇ · (K α ∇u) = K 12,x u y − K 12,y u x = curlK 12 · ∇u.
When K 12 is curl-free, the antisymmetric part of the diffusion is not functional to the elliptic operator. When K 12
is not curl-free the antisymmetric part K α leads to a term that can be taken as the advection operator. Then Eq. (1)
becomes
−K 11 u x x − K 22 u yy − ((K 11 )x + (K 12 ) y )u x − (−(K 12 )x + (K 22 ) y )u y = f. (9)
The basic steps of TFPM scheme used in this paper are as follows. First, we approximate the coefficient and
source term of Eq. (7) by piecewise constants on local cell. Then we find the solutions that satisfy exactly the
homogeneous differential equations with constant coefficients. Take the finite exact solutions as local basis functions
to approximate u and ∇u. Finally, a global discretization is obtained by imposing the normal flux continuity across
cell interface.

2.3. The main idea of the TFPM

As the base of development of TFPM for Eq. (1), a TFPM will be introduced for 1D state advection diffusion
equation in this section. The following 1D advection diffusion equation is considered,
∂ 2u ∂ K 11 ∂u

−K 11 2 + = g, x ∈ (0, a)

∂x ∂x ∂x (10)
u(x = 0) = u 0 , u(x = a) = u a .

537
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

Fig. 1. The stencil for 1D TFPM.

Divide the computation domain ⋃ [0, a] into I sub-intervals, and define h = a/I , xi = i h with i = 0, 1, . . . , I .
We consider a domain [0, a] = i=I i=1 L i with L i = [x i−1 , x i ]. Let the stencil be as in Fig. 1. x i−1/2 stands for the
center of L i . On L i , Eq. (10) can be approximated by
∂ 2ui ∂u i
− αi + βi = gi , (11)
∂x2 ∂x
where αi = K 11 (xi−1/2 ), βi = ∂ ∂Kx11 (xi−1/2 ), gi = g(xi−1/2 ).
x−xi−1/2
Let ψi (x) = βi
gi , u i (x) = ψi (x) + ϖi (x). Then ψi (x) and ϖi (x) satisfy
∂ 2 ψi ∂ψi
− αi + βi = gi , (12)
∂x 2 ∂x
and
∂ 2 ϖi ∂ϖi
− αi + βi = 0. (13)
∂x2 ∂x
The idea of TFPM is to use finite number of exact solutions of Eq. (13) as local basis functions to get the
βi (x−xi )
approximation solution in the local cell. Firstly we define ϖ1,i (x) = 1, ϖ2,i (x) = e αi
. Then we construct
the following function space
Wi = span{ϖ1,i (x), ϖ2,i (x)}. (14)
We can check that ϖ1,i and ϖ2,i are the exact solutions to the homogeneous Eq. (13). According to the linear space
(14) we can use the local∑basis functions to construct approximation solution on L i . Let ck,i k = 1, 2, i = 1, 2, . . . , I
are constants. ϖi (x) = 2k=1 ck,i ϖk,i (x), x ∈ L i . Then,
2

u i (x) = ck,i ϖk,i (x) + ψi (x),
k=1
2
(15)
∑ ∂ ∂
∂x u i (x) = ck,i ϖk,i (x) + ψi (x).
k=1
∂x ∂x

Suppose the diffusion coefficient is discontinuous at xi . Let u i± denote the solution on two sides of the xi , the
interface conditions are:
∂u i+ ∂u i−
u i+ = u i− , +
K 11 ∂x

= K 11 ∂x
. (16)
That is
⎧ 2 2
∑ ∑
ϖ ψ ck,i+1 ϖk,i+1 + ψi+1 )(xi ),

⎪(

⎪ ck,i k,i + i )(x i ) = (


⎨ k=1 k=1
(17)
2 2
∂ ∂ ∂ ∂


⎪ ∑ ∑
+
ϖ ψ −
ck,i ϖk,i+1 + ψi+1 )(xi ).

⎪ K ( ck,i k,i + i )(x i ) = K (
⎩ 11 ∂x ∂x 11
∂x ∂x

k=1 k=1

Combining the boundary conditions of Eqs. (10) and (17) this is a system of 2 × I linear algebraic equations and
the coefficients ck,i , can be determined. In addition, on the point xi−1/2 ,
2

u(x)|xi−1/2 = ck,i .
k=1
538
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

The main advantages of the approach (17) is that it allows to meet with the discontinuity since the scheme is
derived by the interface condition. We will arrive at a new TFPM for Eq. (1) with non-symmetric diffusion tensor
and discontinuity diffusion based on this approach.

Remark 1. The principle of the TFPM is to seek some constants κl , l = 1, 2, 3 to construct a finite difference
approximation for Eq. (13) at the grid point xi as
κ1 ϖi−1 + κ2 ϖi+1 + κ0 ϖi = 0. (18)
In order to determine the coefficients κl , l = 0, 1, 2 in Eq. (18), the two basis functions ϖ1,i , ϖ2,i are taken to
satisfy Eq. (18) exactly. We can get a finite difference scheme as
u i−1 − 2u i + u i+1 u i+1 − u i−1
− δi 2
+ βi = gi , (19)
h 2h
where δi = β2i h coth β2αi hi . The scheme (19) is a second order exponential finite difference scheme for the diffusion
reaction model problem (10), see [29]. The discrete scheme gives rise to a diagonally dominant tri-diagonal system
of equations and provides the exact solution for the 1D advection diffusion equation with constant diffusion and
advection coefficient in the case of g = 0.

3. TFPM for diffusion operators with non-symmetric diffusion tensor


3.1. Notations and assumptions

Assume that Ω = [0, a] × [0, b] in R 2 with uniform meshes {Ωi, j , i = 1, 2, . . . , I, j = 1, 2, . . . , J }. The


stencil is shown in Fig. 2. A cell Ωi, j = {(x, y)⏐ | x − xi+1/2 |≤ h x /2, | y − y j+1/2 |≤ h y /2} is built with

cell centered at z i+1/2, j+1/2 = (xi+1/2 , y j+1/2 ). h x = a/I , h y = b/J is the mesh size. The edge are denoted by
Γi, j,y = Ωi, j ∩ Ωi−1, j = {(xi , y)|y j ≤ y ≤ y j+1 }, Γi, j,x = Ωi, j ∩ Ωi, j−1 = {(x, y j )|xi ≤ x ≤ xi+1 } on Ωi, j . The four
edge centers are denoted by z i+1/2, j+1 , z i+1/2, j , z i+1, j+1/2 , z i, j+1/2 .
Considering Eq. (7) restricted on the cell Ωi, j and approximating coefficients by piecewise constants. We have

ai, j = −K 11 |zi+1/2, j+1/2 , ci, j = −K 22 |zi+1/2, j+1/2 , b1,i, j = −K 12 |zi+1/2, j+1/2 ,


b2,i, j = −K 21 |zi+1/2, j+1/2 , d1,i, j = −((K 11 )x + (K 12 ) y ) |zi+1/2, j+1/2 , (20)
d2,i, j = −((K 21 )x + (K 22 ) y ) |zi+1/2, j+1/2 .
We choose the average value f˜i, j = |Ω1i, j | Ωi, j f (x, y)dΩ to approximate f (x, y) in Ωi, j . Then Eq. (7) can be

approximated by a diffusion equation with constant coefficients on the cell Ωi, j


∂ 2u ∂ 2u ∂ 2u ∂u ∂u
ai, j + (b1,i, j + b2,i, j ) + ci, j + d1,i, j + d2,i, j = f˜i, j . (21)
∂x2 ∂ x∂ y ∂ y2 ∂x ∂y
It is easy to make detailed description of the principle of TFPM from Eq. (21). We will arrive at a TFPM for
solving this problem in the next Subsection.

3.2. Details of the TFPM

The idea of TFPM is to use finite number of exact solutions of homogeneous differential equations as local basis
functions and combine a particular solution of Eq. (21) to get the approximation solution on the local cell. Firstly,
we will find a particular solution of Eq. (21). Let
d1,i, j (x − xi+1/2 ) + d2,i, j (y − y j+1/2 )y ˜
ϕi, j (x, y) = 2 2
f i, j
d1,i, j + d2,i, j

j + d2,i, j ̸ = 0. We can check that ϕi, j (x, y) satisfies Eq. (21). Namely,
2 2
with d1,i,
∂ 2 ϕi, j ∂ 2 ϕi, j ∂ 2 ϕi, j ∂ϕi, j ∂ϕi, j
ai, j + (b1,i, j + b2,i, j ) + ci, j + d1,i, j + d2,i, j = f˜i, j . (22)
∂x 2 ∂ x∂ y ∂ y2 ∂x ∂y
539
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

Fig. 2. The four point stencil.

Decompose u(x, y) in Eq. (21) into

u i, j (x, y) = ϕi, j (x, y) + ωi, j (x, y). (23)

Then ωi, j (x, y) satisfies


∂ 2 ωi, j ∂ 2 ωi, j ∂ 2 ωi, j ∂ωi, j ∂ωi, j
ai, j + (b 1,i, j + b2,i, j ) + ci, j + + d2,i, j = 0. (24)
∂x2 ∂ x∂ y ∂ y2 ∂x ∂y
There are infinitely many linearly independent solutions to Eq. (24). We can find the exact solution of the type
{ }
1, r0 x + r1 y, s0 x 2 + s1 y 2 + s2 x y, . . . , eλx+µy , (w0 x + w1 y)eλx+µy , (25)

where rk , sk , wk , k = 0, 1, 2 are constants. For example, substituting ω(x, y) = eλx+µy into Eq. (24), then (λ, µ) is
the solution to

ai, j λ2 + ci, j µ2 + (b1,i, j + b2,i, j )λµ + d1,i, j λ + d2,i, j µ = 0. (26)


d d
We choose (λ, µ) = (0, 0), (− a1,i, i, j
j
, 0), (0, − c2,i,
i, j
j
). Besides, we can find solutions of the form ω(x, y) = r0 x + r1 y,
λx+µy
or (w0 x + w1 y)e .
If d1,i, j , d2,i, j ̸= 0, then we can get four simple local basis functions
q1,i, j (x, y) = 1, q2,i, j (x, y) = d2,i, j (x − xi+1/2 ) − d1,i, j (y − y j+1/2 ),
d
− a1,i, j (x−xi+1/2 )
d
− c2,i, j (y−y j+1/2 ) (27)
q3,i, j (x, y) = e i, j , q4,i, j (x, y) = e i, j .
If d1,i, j = 0, d2,i, j ̸= 0, then q3,i, j (x, y) = q1,i, j (x, y) = 1. We can find another solution q3,i, j (x, y) =
d2,i, j (y − y j+1/2 )2 − ai, j (x − xi+1/2 ). Similarly, if d1,i, j ̸= 0, d2,i, j = 0, we can find another solution q4,i, j (x, y) =
d1,i, j (x − xi+1/2 )2 − ci, j (y − y j+1/2 ).
Taking the four exact solutions of homogeneous differential equations as local basis functions, then we further
construct the following function space
{ 4
∑ }
Hi, j = ωi, j (x, y) | ωi, j = σk,i, j qk,i, j (x, y), σk,i, j ∈ R . (28)
k=1
540
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

The solution of Eq. (24) can be expressed as the linear combination of local basis functions in (28). According to
Eq. (23), we have the approximation solution in the Ωi, j
4

u i, j (x, y) = σk,i, j qk,i, j (x, y) + ϕi, j (x, y). (29)
k=1

And the gradient vector can be approximated


⎛ ⎞
∑4 ∂ ∂
σk,i, j ∂ x qk,i, j (x, y) + ∂ x ϕi, j (x, y)⎠
∇u i, j = ⎝∑k=1 . (30)
4 ∂ ∂
σ q
k=1 k,i, j ∂ y k,i, j (x, y) + ϕ
∂ y i, j
(x, y)
As described above, the approximation solution u(x, y) and ∇u(x, y) in Ωi, j are explicitly expressed in terms
of local basis functions of the homogeneous equation and a particular solution of the inhomogeneous equation.

Remark 2. We remark that the choice of exact solutions qk,i, j (x, y) to the homogeneous Eq. (24). In the case of
d1,i, j = d2,i, j = 0, we can get four linearly independent solutions as q1,i, j (x, y) = 1, q2,i, j (x, y) = x − xi+1/2 ,
q3,i, j (x, y) = y − y j+1/2 , q4,i, j (x, y) = ci, j (x − xi+1/2 )2 − ai, j (y − y j+1/2 )2 .

Remark 3. We remark that the choice of exact solutions qk,i, j (x, y) another way from [15] to the homogeneous
Eq. (24) in the case of b1,i, j + b2,i, j = 0. Decompose u(x, y) in (21) into u i, j (x, y) = ϕi, j (x, y) + ωi, j (x, y). Let
d (x−x ) d (y−y )
−( 1,i, j 2a i+1/2 + 2,i, j 2c j+1/2 )
ωi, j (x, y) = vi, j (x, y)e i, j i, j . Then vi, j (x, y) satisfy
∂ vi, j
2
∂ vi, j 2
ai, j + ci, j + d0,i, j vi, j = 0 (31)
∂x2 ∂ y2
2
d1,i, 2
d2,i,
where d0,i, j =
4ai,2
j
+ 2
4ci,
j
. The general solutions of Eq. (31) have the expression with w(x, y) =
j j
λ(x−xi+1/2 )+µ(y−y j+1/2 )
e . (λ, µ) satisfy
ai, j λ2 + ci, j µ2 + d0,i, j = 0. (32)
We can select four pairs
√ √
( ) ( d0,i, j ) ( ) ( d0,i, j )
λk,i, j , µk,i, j = ± ,0 , k = 1, 2, λk,i, j , µk,i, j = 0, ± , k = 3, 4. (33)
−ai, j −ci, j
d1,i, j d2,i, j
Let λ̃k,i, j = λk,i, j − 2ai, j
, µ̃k,i, j = µk,i, j − 2ci, j
. qk (x − xi+1/2 , y − y j+1/2 ) =eλ̃k,i, j (x−xi+1/2 )+µ̃k,i, j (y−y j+1/2 ) satisfy
∂ 2 ωi, j ∂ 2 ωi, j ∂ωi, j ∂ωi, j
ai, j ∂x2
+ ci, j ∂ y2
+ d1,i, j ∂x
+ d2,i, j ∂y
= 0 exactly.

3.3. The framework of the new TFPM scheme

In this Subsection, we derive the TFPM for Eq. (1). We follow the notations in Section 3.1. For each cell Ωi, j
with the boundary ∂Ωi, j , and ∂Ωi, j = Γi, j,y ∪ Γi+1, j,y ∪ Γi, j,x ∪ Γi, j+1,x . See Fig. 2. We can obtain an explicit flux
discretization on the edge Γi, j,x and Γi, j,y ,

FΓi, j,x = ni, j,x · Qds ≈ |Γi, j,x |ni, j,x · Q |Γi, j,x ,
Γ
∫ i, j,x (34)
FΓi, j,y = ni, j,y · Qds ≈ |Γi, j,y |ni, j,y · Q |Γi, j,y ,
Γi, j,y

where ni, j,x is the normal direction of Γi, j,x , ni, j,y is the normal direction of Γi, j,y , |Γi, j,x |, |Γi, j,y | denote the length
of Γi, j,x , Γi, j,y respectively.
Γi, j,y is the interface of Ωi, j and Ωi−1, j . The interface conditions on Γi, j,y are
u + |Γi, j,y = u − |Γi, j,y , ni,+j,y · Q+ |Γi, j,y = −ni,−j,y · Q− |Γi, j,y . (35)
541
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

We have ni,+j,y = −ni,−j,y . The signs ± indicate the two sides of Γi, j,y . From the approximation solution Eq. (29)
and the gradient vector Eq. (30), we can get two equations at the point z i, j+1/2 of Γi, j,y
⎧ ∑4 ∑4
⎨( k=1 σk,i, j qk,i, j + ϕi, j ) |zi, j+1/2 = ( k=1 σk,i−1, j qk,i−1, j + ϕi−1, j ) |zi, j+1/2
(36)
⎩ +
ni, j,y · (K ∇u i, j ) |zi, j+1/2 = −ni,−j,y · (K ∇u i−1, j ) |zi, j+1/2
with i = 2, . . . , I , j = 1, . . . , J .
The interface conditions on Γi, j,x are
u + |Γi, j,x = u − |Γi, j,x , ni,+j,x · Q+ |Γi, j,x = −ni,−j,x · Q− |Γi, j,x . (37)
Inserting Eqs. (29) and (30) into Eq. (37), the interface conditions can be applied to the point z i+1/2, j of Γi, j,x with
the following two requirements,
⎧ ∑4 ∑4
⎨( k=1 σk,i, j qk,i, j + ϕi, j ) |zi+1/2, j = ( k=1 σk,i, j−1 qk,i, j−1 + ϕi, j−1 ) |zi+1/2, j
(38)
⎩ + −
ni, j,x · (K ∇u i, j ) |zi+1/2, j = −ni, j,x · (K ∇u i, j−1 ) |zi+1/2, j
with i = 1, . . . , I , j = 2, . . . , J .
The boundary conditions are imposed on ∂Ω . Let z 1, j+1/2 = (x1 , y j+1/2 ), z I +1, j+1/2 = (x I +1 , y j+1/2 ) be the edge
center. The Dirichlet boundary conditions on Γ D are given
⎧ 4

σk,1, j qk,1, j (x, y) + ϕ1, j (x, y)) |z1, j+1/2 = u D |z1, j+1/2 ,




⎨ (
k=1
4 (39)
⎪ ∑
σk,I, j qk,I, j (x, y) + ϕ I, j (x, y)) |z I +1, j+1/2 = u D |z I +1, j+1/2 .

⎪(


k=1
Let z i+1/2,1 = (xi+1/2 , y1 ), z i+1/2,J +1 = (xi+1/2 , y J +1 ) be the edge center On Γ N . The Neumann boundary conditions
are obtained from
ni,1,x · (K ∇u i,1 ) |zi+1/2,1 = u N |zi+1/2,1 ,
{
(40)
ni,J +1,x · (K ∇u i,J ) |zi+1/2,J +1 = u N |zi+1/2,J +1 .

Remark 4. We also consider the a variant of the Neumann boundary condition of the form n · ∇u = u N on Γ N .
Let z i+1/2,1 = (xi+1/2 , y1 ) be the edge center. The non-standard boundary conditions take the following form
ni,1,x · ∇u i,1 |zi+1/2,1 = u N |zi+1/2,1 . (41)
Combining the interface condition equations Eqs. (36), (38) with the boundary conditions Eqs. (39), (40), we
can get a linear system for the unknown coefficients σk,i, j , k = 1, . . . , 4. Solving this global linear system yields
σk,i,

j s. From Eq. (29), we can get the approximation solutions at the grid point z i+1/2, j+1/2
4

u(x, y)|zi+1/2, j+1/2 = σk,i, j qk,i, j |zi+1/2, j+1/2 .
k=1
We have a system of 4 × I × J linear algebraic equations and the 4 × I × J coefficients σk,i, j can be determined.
But it is not efficient enough.
Another way of constructing the scheme is to take u(xi , y j+1/2 ), u(xi+1 , y j+1/2 ), u(xi+1/2 , y j ) and u(xi+1/2 , y j+1 )
as unknowns on Ωi, j to get the new TFPM scheme. Eq. (29) holds for any u(x, y) on Ωi, j can be determined by
the local basis functions and a particular solution. This observation provides
⎧ ∑4

⎪ ( k=1 σk,i, j qk,i, j + ϕi, j ) |zi, j+1/2 = u(xi , y j+1/2 ),




⎪(∑4 σ q
⎨ k=1 k,i, j k,i, j + ϕi, j ) |zi+1, j+1/2 = u(xi+1 , y j+1/2 ),


(42)
∑4
σ ϕ ,



⎪ ( q
k=1 k,i, j k,i, j + ) |
i, j z i+1/2, j+1 = u(x i+1/2 y j ),




⎪ ∑4
( k=1 σk,i, j qk,i, j + ϕi, j ) |zi+1/2, j+1 = u(xi+1/2 , y j+1 ).

542
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

This results in 4 × 4 matrix system. The coefficients σk,i, j can be determined by u(xi+1/2 , y j±1 ) and u(xi±1 , y j+1/2 ).
The interface condition Eqs. (35) and (37) of the continuity normal flux allows us to construct associated four
equations at the position of the flux quadrature point on Γi, j,y , Γi+1, j,y , Γi, j,x , Γi, j+1,x



⎪ ni,+j,y · Q+ |zi, j+1/2 = −ni,−j,y · Q− |zi, j+1/2 ,


i+1, j,y · Q |z i+1, j+1/2 = −ni+1, j,y · Q |z i+1, j+1/2 ,
⎨n−
⎪ − + +
(43)
n +
· Q +
| = −n −
· Q −
| ,

z z


⎪ i, j,x i+1/2, j i, j,x i+1/2, j


ni, j+1,x · Q− |zi+1/2, j+1 = −ni,+j+1,x · Q+ |zi, j+1 .
⎩ −

Combining Eq. (43) with the given boundary conditions on Γ D and Γ N , this is a system with 2×I × J +I + J linear
algebraic equations which is consistent to the unknowns on edges. So u(xi , y j+1/2 ), u(xi+1 , y j+1/2 ), u(xi+1/2 , y j )
and u(xi+1/2 , y j+1 ) can be determined.

4. The applications of the new TFPM


Consider the following time-dependent diffusion problem,

⎨u t + ∇ · Q = f, in Ω × [0, T ]
(44)
−K ∇u = Q,

subject to boundary conditions in Eq. (1) and the initial condition,


u(x, y, 0) = u 0 (x, y), in Ω , t = 0, (45)
where t is time, T is the total simulation time, u 0 is the initial state and f = f (x, y, t) is source term.
The time discretization takes the implicit Euler form,
∇ · ∆tQn+1 + u n+1 = ∆t f n+1 + u n , (46)
with time step size ∆t = T /N , n = 0, 1, 2, . . . , N . Let K̂ = ∆t K , fˆ = ∆t f n+1 n
+ u , then
−∇ · K̂ ∇u n+1 + u n+1 = fˆn+1 . (47)
where
( )
K̂ 11 K̂ 12
K̂ = . (48)
K̂ 21 K̂ 22 .
Note that Eq. (47) can be solved by using the similar approach in Section 3 for the u at the new time level, n + 1.
The details, see [32].

5. Numerical experiments
In this section, we present numerical experiments to demonstrate the performance of our new scheme. Some of
problems present here can be found in [9,13,18]. Moreover, these tests have analytical solutions except the last one.
We proceed by studying the convergence rate for non-symmetric diffusion tensor problems.

5.1. Example with smooth solution

Example 1. Consider the stationary elliptic problem with an exact solution in Ω


∇(K · ∇u) = f, (49)
subject to the Dirichlet boundary conditions u = u 0 on ∂Ω .
Test A This test is from a numerical experiment in [13]. The exact solution and non-symmetric diffusion tensor
in Ω = [−0.5, 0.5]2 are given by
1 + x2
( )
sin(π x)y
u = sin(π x)y 2 , K = . (50)
− sin(π x)y 1 + 21 y 2
543
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

Fig. 3. Test A of Example 1. I × J = 32 × 32. (a) u(x, y); (b) K 12 .

Fig. 4. Test A of Example 1. The L 2 and L ∞ norm of the errors for different grids of the TFPM. (a) L 2 norm; (b) L ∞ norm.

The numerical solutions are shown in Fig. 3. In addition, the errors with different mesh sizes are shown in Fig. 4.
Second order convergence can be obtained.
Test B This test is rebuilt from a numerical experiment in [31]. The exact solution and non-symmetric diffusion
tensor in Ω = [0, 1]2 are given by
( )
x 2 + 10−3 y 2 (1 − 10−3 )x y
u = sin π x sin π y, K = 1 . (51)
2
(1 − 10−3 )x y 10−3 x 2 + y 2
Fig. 5(a) gives the numerical errors of TFPM in L 2 norm and second order convergence rate can be obtained.
Fig. 5(b) gives the numerical errors of TFPM in L ∞ norm and the convergence rate is about 1.65.

5.2. Boundary layer

Example 2. Consider the non-symmetric diffusion tensor problem


∇(K · ∇u) = f, (52)
544
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

Fig. 5. Test B of Example 1. The L 2 and L ∞ norm of the errors for different grids of the TFPM. (a) L 2 norm; (b) L ∞ norm.

Fig. 6. Example 2. The L 2 and L ∞ norm of the errors for different grids and ϵ values of the TFPM. (a) L 2 norm; (b) L ∞ norm.

subject to the Dirichlet boundary conditions u = u 0 on ∂Ω and non-symmetric diffusion tensor is given by
− π1 y
( )
1
K = 1 . (53)
π
y ϵ
− y
The exact solution is given by u = sin(π x)+e ϵ 1/2 . The source term f is determined by the assumed exact solution.
When ϵ is small, the solution is almost constant in the y variable. The numerical errors of different ϵ are shown in
Fig. 6. We can see that the new TFPM has second order convergence. The numerical results are shown in Fig. 7.
The exact solution exhibits boundary layer in this example, TFPM is able to capture the boundary layer correctly
even for coarse grids.

5.3. Strongly anisotropic and discontinuous

Example 3. In this example we solve a strongly anisotropic and discontinuous problem. The computational domain
Ω is divided into two subdomains: Ω1 = [0, 0.5] × [0, 1] and Ω2 = [0.5, 1] × [0, 1].
545
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

Fig. 7. Example 2. The numerical solutions: I × J = 64×64. (a) u(x, y) with ϵ = 10−1 ; (b) Error with ϵ = 10−1 ; (c) u(x, y) with ϵ = 10−3 ;
(d) Error with ϵ = 10−3 ; (e) u(x, y) with ϵ = 10−9 ; (f) Error with ϵ = 10−9 .

546
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

The diffusion tensor varies as


⎧⎛ ⎞
⎪⎝ K 11 K 12 ⎠



⎨ for (x, y) ∈ Ω1 ,
K = K 21 K 22 (54)
( )
⎪ 106 0



⎩ for (x, y) ∈ Ω2 .
0 1
The exact solution is
2
⎧ K 12
1
⎨[1 + (x − 0.5)( K 11 +
⎪ K 11
40π (y − 0.5))]e−20π (y−0.5) x ≤ 0.5,
u(x, y) = (55)
⎩ x−0.5
⎪ 2
e 106 e−20π (y−0.5) x > 0.5.
Let θ = 25
180
π. The diffusion tensor in Ω1 is defined by
K 11 = 3000 cos2 θ + sin2 θ, K 12 = (3000 − 1) cos θ sin θ, K 22 = cos2 θ + 3000 sin2 θ.
Three different cases K 21 = −K 12 , K 21 = 0.1 × K 12 , K 21 = K 12 are tested.
Test A K 21 = −K 12 . Each subdomain has a different diffusion tensor such that
⎧⎛ ⎞


⎪⎝ 2464.360020 1148.683643 ⎠ for (x, y) ∈ Ω1 ,


K = −1148.683643 536.6399794 (56)
⎪ ( 6 )
⎪ 10 0
for (x, y) ∈ Ω2 .



0 1
Test B K 21 = 0.1 × K 12 . Each subdomain has a different diffusion tensor such that
⎧⎛ ⎞
⎪⎝2464.360020 1148.683643⎠



⎨ for (x, y) ∈ Ω1 ,
K = 114.8683643 536.6399794 (57)
⎪ ( 6 )
⎪ 10 0
for (x, y) ∈ Ω2 .



0 1
Test C K 21 = K 12 . Each subdomain has a different diffusion tensor such that
⎧⎛ ⎞


⎪ 2464.360020 1148.683643 ⎠ for (x, y) ∈ Ω1 ,
⎨⎝

K = 1148.683643 536.6399794 (58)
⎪ ( 6 )
⎪ 10 0
for (x, y) ∈ Ω2 .



0 1
The source term f and boundary conditions are chosen according to the exact solution. The solution and normal
flux are continuous but its tangential flux is discontinuous across the interface. The numerical solutions are shown
in Fig. 8. The L 2 , L ∞ norm of the errors for the three tests are presented in Fig. 9. Second order accuracy can be
observed.
The relative asymptotic errors for the flux are defined using the mean-square norm
h 2 h 2
∑ ∑
F
( i, j (FΓi, j,x − FΓi, j,x ) + i, j (FΓi, j,y − FΓi, j,y ) )1/2
El 2 = ∑ 2
∑ 2
(59)
i, j FΓi, j,x + i, j FΓi, j,y

where FΓhi, j,x , FΓhi, j,y is the numerical approximation flux in Eq. (34). Table 1 shows that second order convergence
rate can be obtained for the flux of the three tests.

5.4. Interface problem

Example 4. We consider the problem that exhibits interface from a numerical experiment in [9]. The domain
Ω = [0, 1] × [0, 1] is split into two subdomains: Ω1 = [0, 21 ] × [0, 1], Ω2 = [ 21 , 1] × [0, 1]. The diffusion tensor is
547
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

Fig. 8. Example 3. The numerical solutions: I × J = 64 × 64. (a) u(x, y) of Test A; (b) The error of Test A; (c) u(x, y) of Test B; (d)
The error of Test B; (e) u(x, y) of Test C; (f) The error of Test C.

548
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

Fig. 9. Example 3. The L 2 and L ∞ norm errors for the solutions of the three tests. (a) L 2 norm; (b) L ∞ norm.

Table 1
Example 3. The ElF2 errors and convergence order for the flux of the three tests.
I×J Test A Test B Test C
ElF2 Rate ElF2 Rate ElF2 Rate
8 × 8 3.9920E−01 – 2.7574E−01 – 1.9710E−01 -
16 × 16 1.0477E−01 1.9299 1.1076E−01 1.3158 1.2271E−01 0.6837
32 × 32 2.4752E−02 2.0816 2.5724E−02 2.1063 2.9249E−02 2.0688
64 × 64 6.1128E−03 2.0177 6.3138E−03 2.0265 7.2163E−03 2.0191
128 × 128 1.5238E−03 2.0041 1.5713E−03 2.0066 1.7980E−03 2.0049

Table 2
Example 4. I × J = 32 × 32. The discrete L 2 and L ∞ norm of the errors for different ϵ values in Ω1 .
ϵ 1 0.1 0.05 0.01 0.005 0.001
L2 2.1726E−16 1.6655E−16 1.4541E−16 1.5077E−16 1.7111E−16 2.7439E−16
L∞ 8.8818E−16 7.7716E−16 9.4369E−16 1.9984E−15 2.6645E−15 1.3878E−15

constant in each domain is defined by


ϵ
( ) ( )
y 1 y
K = for (x, y) ∈ Ω1 , K = for (x, y) ∈ Ω2 .
−y 1 −y 1
The exact solution in each subdomain

⎪u( 2 , y)1/2ϵ (x, y) ∈ Ω1 ∩ Ω2 ,
1

⎪ = g,
⎨g − e + (1 − g)e x/ϵ

u(x, y) = , (x, y) ∈ Ω1 , (60)
1 − e1/2ϵ
1/2 x−1/2

⎩ −ge + ge

, (x, y) ∈ Ω2 ,


1−e 1/2

e1/2ϵ 1/2ϵ
e 1 −1
with g = 1−e 1/2ϵ ( 1−e1/2ϵ + 1−e1/2 ) . Inserting the exact solution into −∇ · K ∇u = f , we can get the source term
f = 0. The numerical results are shown in Fig. 10. The TFPM is able to capture the sharp discontinuity even for
h x , h y ≫ ϵ. The numerical errors of different ϵ are reported in Table 2. We see that the errors from the method
is about O(10−16 ). This is due to the fact that the exact solution is independent of the y coordinate and changes
exponentially in the x direction. Table 2 shows that the TFPM preserves the exponential solution.
549
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

Fig. 10. Example 4. I × J = 32 × 32. (a) u(x, y) with ϵ = 0.05; (b) u(x, y) with ϵ = 0.01; (c) u(x, y) with ϵ = 0.005; (d) u(x, y) with
ϵ = 0.001.

5.5. Hall effect

Example 5. We deal with the resistive magnetohydrodynamics problem with the Hall effect presented in [13,18]. In
order to describe the problem clearly, let us repeat problem formulation. Let B be the z-component of the magnetic
field and σ is the electrical conductivity that satisfies
1 1 1 1
in (0, T ) × Ω ,

⎪ Bt − µ0 ∇ · ( σ ∇ B) − |qe |µ0 curl( n B∇ B) = 0,





⎨ B(0) = B0 , in Ω = [−0.5, 0.5]2 ,


(61)
⎪ B = Btop , onΓtop ,







onΓle f t ∪ Γbottom ∪ Γright ,

n · ∇ B = 0,
550
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

where Γle f t , Γright , Γtop , Γbottom indicate the left, right, top and bottom sides of the square domain Ω . Choose n r
as a unit of density and let λr , tr , Br , ω p be length, time, magnetic field units and the plasma frequency. such that,
λr = ctr ,
nr |qe |2 − 12
tr = ω−1
p = ( ϵ0 m e ) , (62)
Br = |qme |te r .
Let
α βu ϵ0 ω p
( )
B nr
u= , K = , α= , β= . (63)
Br −βu α σ n
Eq. (61) reduces to
in(0, T ) × Ω ,


⎪ u t − ∇ · (K (u)∇u) = 0



⎨u(0) = u 0 , in Ω = [−0.5, 0.5]2 ,



(64)
u = uN , onΓtop ,







onΓle f t ∪ Γbottom ∪ Γright .

n · ∇u = 0,

We consider the time interval from 0 to T = 0.5tr . The initial condition and the boundary conditions are
y < 0,
⎧ {
⎪u(t = 0) = u 0 = 0

−1 y > 0,

(65)
⎪ u = −1, on Γtop ,
on Γle f t ∪ Γbottom ∪ Γright .

∇u · n = 0,

Test A: The spatially dependent material parameters α and β in K are taken to be


f or x < 0, f or x < 0,
{ −2 { −1
10 10
α(x) = β(x) = (66)
10−3 f or x > 0, 10−5 f or x > 0.
In this test, the time discretization was done using the backward Euler scheme with time step △t = 0.01tr . Space
step h x = h y = 64 1
, 128
1
are employed in practical computation by using the proposed TFPM. The value of the
diffusion coefficient K (u) was computed using the last known time step solution. The numerical solutions are shown
in Fig. 11. In this case parameter α indicates stronger diffusion in the computational domain x < 0. The jump in
the antisymmetry parameter β ultimately lead to the formation of the spike along the interface in the solution. The
formation of the spike has been described as the Hall effect. The numerical solution captures this effect and is
visually identical to the reference solution obtained in [13,18].
Test B: In [13] the authors have pointed out that the spike would form even if the parameter α was constant
throughout the whole domain and did not experience discontinuity at the interface x = 0. In this case the α and β
in K are taken to be
f or x < 0
{ −1
10
α(x) = α1 in Ω = [−0.5, 0.5]2 , β(x) = (67)
10−5 f or x > 0
α is constant throughout the whole domain. Numerical simulation tests are carried out to study the influences of
parameters α1 . We choose two different values of α1 = 10−3 , 10−4 , the remaining parameters are used as the Test
A. The numerical solutions are shown in Fig. 12. We observe the Hall effect and the high gradient values along the
interface x = 0.

6. Conclusion
We develop an efficient TFPM for numerically solving anisotropic diffusion equations with a non-symmetric
diffusion tensor. Compared with the existing numerous possible choices, the discretization of the normal fluxes are
explicitly expressed by using a linear combination of exact solutions for the constant coefficient equation in each
551
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

Fig. 11. Example 5. (a) (b) I × J = 64 × 64. u(x, y); (c) (d) I × J = 128 × 128. u(x, y).

Fig. 12. Example 5. I × J = 64 × 64. (a) u(x, y) with α1 = 10−3 ; (b) u(x, y) with α1 = 10−4 .

552
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

cell. Such strategy can be conveniently used to construct numerical schemes according to interface conditions on
each cell edge, and handle the Neumann boundary condition or a variant of Neumann boundary condition.
The TFPM works well for anisotropic diffusion equations with a non-symmetric diffusion tensor as well as for
the strongly anisotropic and discontinuous problem. We adopt several numerical experiments, some are the same
problems in MFDM [13] to show the effectiveness of our new derived scheme. The TFPM can ensure second order
convergence.
In this paper, we only present the construction of the TFPM on Cartesian grids, demonstrate its capability of
capturing the interface and boundary layers numerically. However, the solution, as well as the gradient, can be
expressed in terms of the basis functions within each cell. To obtain such an expression of the TFPM, we just need
the geometry information and the diffusion tensor of the current cell. It is interesting for us to extend this method on
general meshes. A further study of the TFPM for solving anisotropic diffusion problem on unstructured polygonal
meshes is our on going work.

Acknowledgments
The author sincerely thanks the referees for many helpful recommendations for the improvement of the
manuscript.

References
[1] I. Aavatsmark, An introduction to multipoint flux approximations for quadrilateral grids, Comput. Geosci. (6) (2002) 405–432.
[2] I. Aavatsmark, G.T. Eigestad, B.T. Mallison, J.M. Nordbotten, A compact multipoint flux approximation method with improved
robustness, Numer. Methods Partial Differential Equations 24 (5) (2008) 1329–1360.
[3] D.N. Arnold, An interior penalty finite element method with discontinuous elements, SIAM J. Numer. Anal. 19 (4) (1982) 742–760.
[4] J. Droniou, Finite volume schemes for diffusion equations: Introduction to and review of modern methods, Math. Models Methods
Appl. Sci. 24 (08) (2014) 1575–1619.
[5] M. Edwards, M. Pal, Positive-definite q-families of continuous subcell darcy-flux cvd(mpfa) finite-volume schemes and the mixed finite
element method, Internat. J. Numer. Methods Fluids 57 (4) (2008) 355–387.
[6] M.G. Edwards, C.F. Rogers, Finite volume discretization with imposed flux continuity for the general tensor pressure equation, Comput.
Geosci. 2 (1998) 259–290.
[7] M. Edwards, H. Zheng, A quasi-positive family of continuous darcy-flux finite-volume schemes with full pressure support, J. Comput.
Phys. 227 (2008) 2152–2161.
[8] M. Edwards, H. Zheng, Double-families of quasi-positive darcy-flux approximations with highly anisotropic tensors on structured and
unstructured grids, J. Comput. Phys. 229 (3) (2010) 594–625.
[9] A. Ern, A. Stephansen, P. Zunino, A discontinuous galerkin method with weighted averages for advection-diffusion equations with
locally vanishing and anisotropic diffusivity, IMA J. Numer. Anal. 29 (2) (2009) 235–256.
[10] P.R. Gent, J.C. Mcwilliams, Isopycnal mixing in ocean circulation models, J. Phys. Oceanogr. 20 (1) (1990) 150–155.
[11] S. Griffies, The Gent-McWilliams skew flux, J. Phys. Oceanogr. 28 (1998) 831–841.
[12] V. Gyrya, K. Lipnikov, High-order mimetic finite difference method for diffusion problems on polygonal meshes, J. Comput. Phys.
227 (20) (2008) 8841–8854.
[13] V. Gyrya, K. Lipnikov, The arbitrary order mimetic finite difference method for a diffusion equation with a non-symmetric diffusion
tensor, J. Comput. Phys. 348 (899) (2017) 549–566.
[14] V. Gyrya, K. Lipnikov, G. Manzini, The arbitrary order mixed mimetic finite difference method for the diffusion equation, ESAIM
Math. Model. Numer. Anal. 50 (2016) 851–877.
[15] H. Han, Z.Y. Huang, Tailored finite point method based on exponential bases for convection–diffusion-reaction equation, Math. Comp.
82 (2013) 213–226.
[16] H. Han, Z.Y. Huang, B. Kellogg, A tailored finite point method for a singular perturbation problem on an unbounded domain, J. Sci.
Comput. 36 (2008) 243–261.
[17] H. Han, Z.Y. Huang, W.J. Ying, A semi-discrete tailored finite point method for a class of anisotropic diffusion problems, Comput.
Math. Appl. 65 (2013) 1760–1774.
[18] F. Hermeline, A finite volume method for the approximation of diffusion operators on distorted meshes, J. Comput. Phys. 160 (2000)
481–499.
[19] Z.Y. Huang, Y. Li, Monotone finite point method for non-equilibrium radiation diffusion equations, BIT 56 (2016) 659–679.
[20] Y. Kuznetsov, K. Lipnikov, M. Shashkov, The mimetic finite difference method on polygonal meshes for diffusion-type problems,
Comput. Geosci. 8 (4) (2004) 301–324.
[21] K. Lipnikov, G. Manzini, M. Shashkov, Mimetic finite difference method, J. Comput. Phys. 257 (2014) 1163–1227.
[22] K. Lipnikov, M. Shashkov, D. Svyatskiy, Y. Vassilevski, Monotone finite volume schemes for diffusion equations on unstructured
triangular and shape-regular polygonal meshes, J. Comput. Phys. 227 (1) (2007) 492–512.
[23] K. Lipnikov, M. Shashkov, I. Yotov, Local flux mimetic finite difference methods, Numer. Math. 112 (1) (2009) 115–152.
553
Y. Wang Mathematics and Computers in Simulation 182 (2021) 535–554

[24] K. Lipnikov, D. Svyatskiy, Y. Vassilevski, A monotone finite volume method for advection-diffusion equations on unstructured polygonal
meshes, J. Comput. Phys. 229 (11) (2010) 4017–4032.
[25] M. Medvedev, V.V. Medvedev, Asymmetric diffusion of cosmic rays, Phys. Plasmas 22 (2015) 091504.
[26] J. Pasdunkorale, I.W. Turner, A second order control-volume finite-element least-squares strategy for simulating diffusion in strongly
anisotropic media, J. Comput. Phys. 23 (1) (2005) 1–16.
[27] Z.Q. Sheng, J.Y. Yue, G.W. Yuan, Monotone finite volume schemes of nonequilibrium radiation diffusion equations on distorted meshes,
SIAM J. Sci. Comput. 31 (4) (2009) 2915–2934.
[28] Y. Shih, R.B. Kellogg, P. Tsai, A tailored finite point method for convection–diffusion-reaction problems, J. Sci. Comput. 44 (1) (2010)
108.
[29] M. Tang, Y.H. Wang, Uniform convergent tailored finite point method for advection-diffusion equation with discontinuous, anisotropic
and vanishing diffusivity, J. Sci. Comput. 70 (2017) 272–300.
[30] V.V. Vikhrev, O.Z. Zabaidullin, Magnetic field spreading along plasma interface due to the hall effect, Plasma Phys. Rep. 20 (1994)
867–871.
[31] J. Wu, Z. Gao, Interpolation-based second-order monotone finite volume schemes for anisotropic diffusion equations on general grids,
J. Comput. Phys. 275 (2014) 569–588.
[32] T.T. Yang, Y.H. Wang, A new tailored finite point method for strongly anisotropic diffusion equation on misaligned grids, Appl. Math.
Comput. 355 (2019) 85–95.
[33] G.W. Yuan, Z.Q. Sheng, Monotone finite volume schemes for diffusion equations on polygonal meshes, J. Comput. Phys. 227 (12)
(2008) 6288–6312.

554

You might also like