The Method of Fundamental Solutions with

Eigenfunctions Expansion Method for 3D

Nonhomogeneous Diffusion Equations
D. L. Young, C. H. Chen, C. M. Fan, L. H. Shen
Department of Civil Engineering and Hydrotech Research Institute, National Taiwan
University, Taipei 10617, Taiwan

Received 24 October 2006; accepted 23 December 2007

Published online 18 March 2008 in Wiley InterScience (www.interscience.wiley.com).
DOI 10.1002/num.20336

After the successful applications of the combination of the method of fundamental solutions (MFS), the
method of particular solutions (MPS), and the eigenfunctions expansion method (EEM) to solve 2D homo-
geneous and nonhomogeneous diffusion equations by Young et al. (Young et al., Numer Meth Part Differ
Equat 22 (2006), 1173), this article intends to extend the same fundamental concepts to calculate more
challenging 3D nonhomogeneous diffusion equations. The nonhomogeneous diffusion equations with time-
independent source terms and boundary conditions are analyzed by the proposed meshless MFS-MPS-EEM
model. Nonhomogeneous diffusion equation in any complex domains can be decomposed into a Poisson
equation and a homogeneous diffusion equation by the principle of linear superposition. This approach is
proved to be far better off than solutions by using classic method of separation of variables with inefficient
multisummation of very sophisticated series expansion from special functions, which can only limit to treat
very simple 3D geometries such as cube, cylinder, or sphere. Poisson equation is solved by using the MPS-
MFS model, in which the source term in the Poisson equation is first handled by the MPS based on the
compactly-supported radial basis functions and the Laplace equation is solved by the MFS. On the other
hand, by utilizing the EEM, the homogeneous diffusion equation is first transformed into a Helmholtz equa-
tion, which is then solved by the MFS together with the technique of singular value decomposition (SVD)
to acquire the eigenvalues and eigenfunctions. After the eigenfunctions are obtained, we can synthesize the
diffusion solutions like the orthogonal Fourier series expansions but with only one summation for the series
even for multidimensional problems. Numerical results for four case studies of 3D homogeneous and nonho-
mogeneous diffusion problems show good agreement with the analytical and other numerical solutions, such
as finite element method (FEM). Thus, the present numerical scheme has provided a promising meshfree
numerical approach to solve 3D nonhomogeneous diffusion equations with time-independent source terms
and boundary conditions for very irregular domains. © 2008 Wiley Periodicals, Inc. Numer Methods Partial
Differential Eq 25: 195–211, 2009

Keywords: 3D nonhomogeneous diffusion equations; eigenfunctions expansion method; method of funda-

mental solutions; method of particular solutions; singular value decomposition

Correspondence to: Professor D. L. Young, Department of Civil Engineering and Hydrotech Research Institute, National
Taiwan University, Taipei 10617, Taiwan (e-mail: [email protected])
Contract grant sponsor: National Science Council of Taiwan; contract grant numbers: NSC 94-2611-E-002-007, NSC
94-2211-E-002-071
© 2008 Wiley Periodicals, Inc.
196 YOUNG ET AL.

I. INTRODUCTION

In recent years, the so-called meshless or meshfree methods have become very attractive numerical
alternatives and received a considerable attention as comparing to the classical mesh-dependent
numerical methods. The MFS is considered the simplest and most popular method in the category
of meshless or meshfree algorithms. The MFS like the boundary element method (BEM) is one
of the mesh-reduction methods. One great advantage of the mesh-reduction methods is to reduce
one dimension and to save a lot of CPU time and computer memory.
In the applications of the conventional BEM to diffusion problems, for examples, Zhu [1],
Zerroukat [2], and Sutradhar et al. [3] all applied the BEM to solve different diffusion equations.
Tremendous computational efforts are still required in BEM to calculate the domain integration
for the source terms especially in 3D nonhomogeneous diffusion problems. To circumvent those
drawbacks, the dual reciprocity method was initiated by Nardini and Brebbia [4] by transforming
the domain integral into a boundary integral by a series of radial basis functions. On the other hand,
Chen et al. [5] combined the MFS with MPS to deal with the diffusion equations. The combined
MFS-MPS model is free from the evaluation of singular integral for solving nonhomogeneous
linear operators, as is generally required by the BEM. Therefore the MFS-MPS model, a meshless
numerical algorithm, has been considered to be an attractive solver for homogeneous diffusion
problems [5–8].
In general, the MFS solutions of homogeneous diffusion equations either use the Laplace
transform [5, 6] or the finite difference [7, 8] to discretize the time derivative term. This is due
to the fact that the MFS is very capable to solve the spatial domain problems if discretization of
transient parts have already been taken care. Chen et al. [6] transformed the diffusion equation into
a nonhomogeneous modified Helmholtz equation using the Laplace transform and then used the
modified Helmholtz fundamental solution to solve the problems. When the Laplace transform is
adopted, the inverse Laplace transform will be needed and sometimes it leads to certain difficulties
in retrieving the forward diffusion solutions.
To overcome the drawbacks of the Laplace transform or finite difference discretization in
time state, Young et al. [9] used the time-dependent fundamental solutions of diffusion equations
directly to solve the multidimensional diffusion equations. Young et al. [10] further extended the
time-dependent MFS-MPS model to solve nonhomogeneous diffusion problems. By this time-
dependent MFS technique, they were able to solve the nonlinear advection-diffusion-reaction
equations if combining with the Eulerian-Lagrangian method (ELM) [11–13]. The only problem
in the time-dependent MFS formulation is to lose the advantage of mesh-reduction method since
the time dimension is retained. However, there is another way to avoid this disadvantage, which
is completely different from the above-mentioned methods to treat the time derivative. This is
the so-called EEM in this proposed algorithm. We will use the EEM to remove the time depen-
dence. Yao and Margrave [14] have already used the eigenfunctions transform method to solve
the wave equation. As far as time evolution is concerned, the EEM is more feasible and robust
for the transient calculations for diffusion and wave equations for any time without using the
time marching process. Recently, Young et al. [15] have successfully used this meshless scheme
of MFS-MPS-EEM model to solve 2D nonhomogeneous diffusion equations. Both regular and
irregular domains are investigated in their numerical tests and the excellent agreement of the
numerical results with the analytical solutions indicates the effectiveness of the 2D numerical
scheme. Once the numerical model is built, the numerical results at any time can be obtained
by a very simple summation procedure from the eigenfunctions like the solutions obtained by
the Fourier series expansions. In contrast to the numerical solutions obtained only in discrete
time levels by conventional methods, the proposed scheme can provide the continuous results

Numerical Methods for Partial Differential Equations DOI 10.1002/num

 3D NONHOMOGENEOUS DIFFUSION EQUATIONS 197

along the time axis. In addition, the numerical scheme is free from numerical stability mesh
generation, numerical quadrature, and singular integral. In this article, the numerical procedures
proposed by Young et al. [15] will be extended to 3D nonhomogeneous diffusion equations with
time-independent source terms and boundary conditions.
The MFS-MPS-EEM model with the SVD method is adopted to solve 3D nonhomogeneous
diffusion equations in both regular and irregular domains. At first the nonhomogeneous diffu-
sion equation is decomposed into a Poisson equation and a homogeneous diffusion equation
without the need of Laplace transform or finite difference or space-time collocation methods
to deal with time term as cited earlier. The Poisson equation is analyzed by the well-known
MFS-MPS model which is a mesh-free method. Meanwhile, the time-dependent solutions of
the homogeneous diffusion equation are directly solved by the superposition principle of the
eigenvalues and eigenfunctions obtained by the MFS with SVD. Looking into the physical
aspects, it is observed that only few eigenfunctions are needed to represent the solutions for
diffusion problems because of the diffusivity character. Moreover, initial condition is used to
determine the weighting coefficients of the orthogonal eigenfunctions. The model is applied to
3D diffusion problems with Dirichlet boundary conditions in both regular and irregular domains.
To validate the accuracy of the present method in 3D nonhomogeneous diffusion problems,
the analytical and FEM solutions are used. This article is organized as follows. In Section II,
the governing equations and the initial as well as boundary conditions (BCs) are elaborated.
The numerical discretization of the MFS, MPS, EEM, and SVD schemes are described in details
in Section III. In Section IV, we delineate the comparisons of the present results with analytical and
FEM computations. The conclusions based on this study and suggestions for further researches
are drawn in Section V.

II. GOVERNING EQUATION

Consider a nonhomogeneous diffusion equation with time-independent source term and BCs on
computational domain
with the corresponding boundary :
∂u(

x , t)
x , t) + A(

= k∇ 2 u(
x) (1)
∂t
in which x
= (x, y, z) is the 3D spatial coordinate, t is the time, k is the diffusion coefficient
(diffusivity), A(

x ) is the time-independent source function, and u(

x , t) is the scalar variable to be
determined. The initial condition (IC) of the diffusion equation at initial time t0 is:

x , t0 ) = B(

u(
x) in
(2)

with the Dirichlet and/or Neumann BCs:

x , t) = C(

u(
x) on 1
∂u
x , t) = D(

(
x) on 2 (3)
∂n
where  1 +  2 is equal to the boundary  and n is the outward normal direction. Moreover, the
BC is of the Dirichlet type if only  2 = 0, of the Neumann type if only  1 = 0, and of the Robin
type if both  1  = 0 and  2  = 0. The boundary conditions C(
x ) and D(
x ) are assumed to be
time-independent functions. The augmented data of the problem are A(
x ), B(
x ), C(

x ), and D(

x ),
which are all time-independent known functions.

Numerical Methods for Partial Differential Equations DOI 10.1002/num

198 YOUNG ET AL.

III. NUMERICAL METHOD

Based on the superposition principle of the linear system, the nonhomogeneous diffusion problem
with time-independent source terms and BCs can be decomposed into a Poisson equation and a
homogeneous diffusion equation. Hence, the solution of the nonhomogeneous diffusion equation
will be represented as follows:

x , t) = u1 (

u(
x ) + u2 (

x , t) in
(4)

where u1 (

x ) satisfies the Poisson equation with nonhomogeneous BCs.
1
x ) = − A(

∇ 2 u1 (
x)
k
x ) = C(

u1 (
x) on 1
∂
x ) = D(

u1 (
x) on 2 (5)
∂n
And u2 (

x , t) satisfies the homogeneous diffusion equation with homogeneous BCs and
inhomogeneous IC.

∂u2 (

x , t)
= k∇ 2 u2 (

x , t)
∂t
x , t0 ) = B(

u2 (
x ) − u1 (

x) in

x) = 0
u2 (
on 1
∂
x) = 0
u2 (
on 2 (6)
∂n
In Eq. (5), u1 (

x ) is a time-independent function that physically represents steady-state (or
quasi-static) solution [9, 15, 16]. Then from the superposition principle we can solve the Poisson
equation, Eq. (5), by decomposing the solution into the homogeneous and particular solutions as
follows [17]:
x ) = uh (

u1 (
x ) + up (

x ), (7)
where uh (

x ) is the homogeneous solution and up (
x ) is the particular solution.
The particular solution up (

x ) satisfies the governing equation without BCs.

1
∇ 2 up (

x ) = − A(

x) (8)
k
The particular solution corresponding to Eq. (8) can be approximated by the MPS for the
source term −(1/k)A(
x ).

1
p N
x) =
− A(
aj f (r)
k j =1
  
r 2
1− r≤α
f (r) = α (9)
0 r>α

Numerical Methods for Partial Differential Equations DOI 10.1002/num

 3D NONHOMOGENEOUS DIFFUSION EQUATIONS 199

where f (r) is the compactly-supported radial basis functions (CSRBFs) [18], α is the compact
radius, r = |
x − x
j | is the radial distance between the j th field point x
j and the collocation point
x
, and N is the number of collocation nodes. In this study, the collocation points are uniformly
distributed in the interior domain as well as on the boundary. After applying Eq. (9) at N collo-
p
cation points, the unknown coefficient aj can be solved [19]. Therefore, the particular solution
up (

x ) is determined by inverting the Laplace operator of Eq. (8) [20–22]:

N
p
x) =
up (
αj F (r)
j =1

r4 r3 r2
20α 2
− 6α + r≤α
F (r) = 6
(10)
α2 α3
12
− 30r r>α

where F (r) is the corresponding CSRBFs of f (r) in Eq. (9). On the other hand, the homogeneous
solution, uh (

x ), satisfies the Laplace equation as well as the modified BCs:

∇ 2 uh (

x) = 0 in

x ) = C(

uh (
x ) − up (

x) on  1
∂ ∂
x ) = D(

uh (
x) − up (

x) on  2 (11)
∂n ∂n

With the substitution of Eq. (10) into the modified boundary conditions of the homogeneous
equation, the results will be a well-posed Laplace problem. Therefore, the MFS is then applied
to solve the Laplace equation. The solution of Laplace equation is represented by the MFS form
as described by the following:

M
x) =
uh (
αjh G1 (r)
j =1

1
G1 (r) = (12)
4π r

where G1 (r) is the fundamental solution of the 3D Laplace equation, which can be obtained by
the potential theory. r = |

x − x
j | is the distance between the field point x
and the j th source point
x
j . M is the number of source points. So the solution u1 (
x ) of the Poisson equation is obtained
by the MFS-MPS model.
After applying the MFS-MPS model for the Poisson equation, the MFS-EEM is then utilized
to solve the homogeneous diffusion equation, Eq. (6). We now omit the summation of the infinite
2
series of the eigenfunctions expansion at this moment. Let u2 (
x )e−λ kt and substitute
x , t) = us (

into Eq. (6):

∇ 2 us (

x ) + λ2 us (

x) = 0
x) = 0
us (
on  1
∂
x) = 0
us (
on  2 (13)
∂n

Numerical Methods for Partial Differential Equations DOI 10.1002/num

200 YOUNG ET AL.

The homogeneous diffusion equation, Eq. (6), is transformed to the Helmholtz equation, Eq.
(13), with homogeneous boundary conditions. Since the homogeneous boundary conditions are
encountered, the above eigenvalue problem can be solved by the MFS with the SVD method
[23–25]. The solution of the Helmholtz equation by the MFS can be express as below:

Q
x) =
us (
βj G2 (r)
j =1

1 −iλr
G2 (r) = e (14)
4π r

where G2 (r) is the fundamental solution of the 3D Helmholtz equation, which can be obtained by
mathematical analysis. r = |
x − x
j | is the distance between the field point x
and the j th source
point x
j . Q is the number of source points. Since Eq. (13) has nontrivial solutions only for some
discrete eigenvalues, we use the MFS with SVD [23–25] to obtain the embedding eigenvalues, λ,
and corresponding eigenfunctions, us (
x ). Then the major advantage of the orthogonal eigenfunc-
tions expansion method is beneficial to this study. The modified IC of the homogeneous diffusion
equation is obtained by the following eigenfunction expansion formula:

∞
−λ2j kt0
x , t0 ) = B(

u2 (
x ) − u1 (

x) = γj usj (

x )e (15)
j =1

where usj (

x ) is the j th eigenfunctions with corresponding eigenvalue λj . γj is the weighting
coefficient of every eigenfunction which is determined by collocating the modified IC of Eq. (15)
at some points inside the domain. Depending on physical character of the diffusivity, only the
first few eigenfunctions are enough to represent the accurate solutions for the diffusion problems.
In our numerical tests, the finite terms are used to replace the infinite series in Eq. (15) and the
numerical results will justify the correctness of assumption. Thus we can obtain:

P
−λ2j kt
x , t) =
u2 (
γj usj (

x )e (16)
j =1

where P is the number of adopted eigenfunctions. Therefore, the EEM is capable to obtain time-
dependent solution without using the time-marching process, such as Laplace transform or finite
difference or space-time collocation methods. From the operational calculus of Eqs. (4)–(16), we
can obtain the solutions of Eq. (1) by the linear superposition:

x , t) = u1 (

u(
x ) + u2 (

x , t)
= uh (

x ) + up (

x ) + u2 (

x , t)


M

N

P
−λ2j kt
= αjh G1 (r) + αjP F (r) + γj usj (

x )e (17)
j =1 j =1 j =1

Thus, the solutions of the 3D nonhomogeneous diffusion equations with time-independent source
terms and boundary conditions can be obtained by the proposed MFS-MPS-EEM scheme. It

Numerical Methods for Partial Differential Equations DOI 10.1002/num

 3D NONHOMOGENEOUS DIFFUSION EQUATIONS 201

FIG. 1. Schematic diagram of (a) cubic cavity, (b) spherical cavity, (c) hollow spherical cavity, (d) spiral
cavity.

is very clear from the observation of Eq. (17) that steady and transient solutions are obtained
separately by using present meshless numerical method.

IV. RESULT AND DISCUSSION

Validation for the proposed numerical method is achieved by comparing the results with the ana-
lytical and FEM solutions for four 3D diffusion problems with Dirichlet BCs. In addition, the
effectiveness of the proposed method is verified by solving 3D homogeneous and nonhomoge-
neous diffusion problems in regular and irregular domains. In the numerical experiments (Fig. 1),
regular domains with analytical solutions for cubic cavity, spherical cavity, and irregular domains
without exact solutions for hollow spherical cavity, spiral cavity are considered. The numerical
results and comparisons will be shown and discussed in the following section.

Numerical Methods for Partial Differential Equations DOI 10.1002/num

202 YOUNG ET AL.

Example 1. Consider a cubic cavity of size [0, a] × [0, b] × [0, c] [Fig. 1(a)] and the following
diffusion equation (DE) with IC and BCs.
 
∂ 2u ∂ 2u ∂ 2u 1 ∂u
DE + 2+ 2 − 6x + 3 sin x sin y sin z =
∂x 2 ∂y ∂z k ∂t
x , 0) = xyz + x 3 + sin x sin y sin z
IC u(
in

BCs x , t) = x 3 + sin x sin y sin z
u(
on ∂ (18)

The analytical solution of the problem is given by

8abc


(−1)m+n+l
∞ ∞ ∞
mπ x nπ y lπ z −λ2 kt
x , t) = −
u(
3
sin sin sin e mnl
π l=1 n=1 m=1 mnl a b c

+ x 3 + sin x sin y sin z

  2
mπ 2 nπ 2 lπ
λ2mnl = + + . (19)
a b c

We set a = 0.9, b = 1.1, c = 1.3, and k = 1. Table I shows the numerical results obtained by
the MFS with the SVD. The eigenvalue presents the smallest singular value for various k, which
results in a spike when the solution of the Helmholtz equation is numerically singular. The first
8 calculated eigenvalues and eigenfunctions by MFS-EEM with SVD model, using 294 nodes in
MFS, and 1000 points to interpolate eigenfunctions are displayed in Fig. 2. These numerically
 mπ 2  nπ 2  lπ 2
obtained eigenvalues are almost the same as the exact solution of λmnl = a
+ b + c .
Comparison shows that the computed and analytical eigenvalues are different only after five sig-
nificant digits. And the time evolutions of the full field distribution are described in Fig. 3. Except
for the results at the beginning, the computed numerical results also show good agreement with
analytical solutions at different time stages. The fast decay of field variation also demonstrates
the physics underlying the diffusion process clearly. The results generally exhibit good agreement
with the analytical solutions at different time stages. Once the position is given, the numerical
results can be obtained by a simple summation procedure, Eq. (17). In this example, we also test
the maximum absolute error histograms using the present method for different collocating points.
And the comparison of the time history of the maximum absolute error for different eigenfunctions
is performed when we use 1000 points to interpolate eigenfunctions. As expected more points
generally will render better resolution. If the diffusivity is small, physically the first few eigen-
functions are almost qualified to represent the diffusion solutions. The advantage of the method
is the capability to obtain the solution for any time by the superposition principle of the first few
eigenfunctions, in which their weighting coefficients are dependent on IC.

TABLE I. The numerical results of the first 10 eigenvalues for Examples1, 2, 3, and 4.
k
e.g. 1 2 3 4 5 6 7 8 9 10
1 5.116 6.610 7.116 7.920 8.256 8.538 8.958 9.338 9.562 9.867
2 3.141 6.283 9.424 12.566 15.708 18.850 21.990 25.130 28.270 31.410
3 6.282 6.571 7.111 7.844 8.716 9.682 12.566 12.721 13.026 13.468
4 6.627 6.772 7.117 7.521 8.003 8.902 9.110 9.592 10.081 10.353

Numerical Methods for Partial Differential Equations DOI 10.1002/num

 3D NONHOMOGENEOUS DIFFUSION EQUATIONS 203

FIG. 2. The first 8 eigenvalues and eigenfunctions for cubic cavity problem for Examples 1 (1000 points
for interpolating eigenfunction, 8 eigenfunctions).

Numerical Methods for Partial Differential Equations DOI 10.1002/num

204 YOUNG ET AL.

FIG. 3. The full-field distribution of z = 0.65, x = 0.45, and y = 0.55 surfaces for Example 1.
(a) t = 0.05, (b) t = 0.08, (c) t = 0.1, (d) t = 0.15 (1000 points, 8 eigenfunctions). (—: analytical
solution; - - -: numerical solution).

Numerical Methods for Partial Differential Equations DOI 10.1002/num

 3D NONHOMOGENEOUS DIFFUSION EQUATIONS 205

Example 2. After simulating the cubic cavity, the proposed numerical scheme is utilized to
study the second example in a spherical cavity [Fig. 1(b)]. For simplicity, we choose the nonho-
mogeneous diffusion equation (DE) with radial and axial symmetry assumptions (independent of
θ and φ) to satisfy the following IC and BCs:

1 ∂u(r, t)
DE ∇ 2 u(r, t) =
k ∂t
0.8r(a 2 − r 2 )
IC u(r, t0 ) = in

a2
BCs u(r, t) = 0 on ∂ (20)

The analytical solution of the problem is given by

96a
(−1)n−1
∞
nπ r −kn2 π 2 t/a2
u(r, t) = sin e (21)
π 3 n=1 n3 a

We set t0 = 0, a = radius = 1 and k = 1. The comparison between the present results and
the analytical solutions of the time evolutions of the full-field distribution of z = 0 surfaces are
depicted in Fig. 4, when 6 eigenfunctions and 1062 points to interpolate each eigenfunction are
used. Almost identical results are obtained for the numerical and analytical solutions. Except for
the results at the beginning, the computed numerical results also show good agreement with ana-
lytical solution at different time stages. On the other hand, the further process has demonstrated
the maximum absolute error histogram for different numbers of points to interpolate each eigen-
functions, in which more points generally give better results. We have verified that the physical
results are generally acceptable for different eigenfunctions.

Example 3. The proposed numerical method is then extended to study a more complex domain
problem, the hollow spherical cavity based on the known eigenvalues and eigenmodes [Fig. 1(c)].
We choose a 3D hollow spherical cavity problem with the following IC and BCs:

∂u(r, θ , φ, t)
DE = k∇ 2 u(r, θ , φ, t) + f (r, θ , φ)
∂t
θ (1 − φ)
IC u(r, θ , φ, t0 ) = + 3e0.2r + sin θ sin2 φ in

2r
BCs u(r, θ , φ) = 3e0.2r + sin θ sin2 φ on ∂ (22)
 
where f (r, θ , φ) = sin θ[1−2 cos φ −2 cos 2φ]−1.2r 2 0.1 + 1r e0.2r is the source term function,
and t0 = 0. In this irregular domain, it is rather difficult to get an analytical solution, so, we choose
FEM with tetrahedron elements to obtain the numerical results for comparison. The distributions
of 19,429 structured tetrahedron elements of FEM meshes and 1150 meshless MFS nodes are
applied in this irregular example. Since the eigenvalues and eigenmodes have been obtained from a
previous study of solutions of wave or Helmholtz equations [24] up to the first six eigenvalues, we
can directly use them to construct the diffusion solutions by the EEM without repetitive works as
far as the eigenvalues calculations are concerned. However, we have repeated the works to check
the known first six (1–6) eigenvalues and to obtain another remaining four (7–10) eigenvalues and

Numerical Methods for Partial Differential Equations DOI 10.1002/num

206 YOUNG ET AL.

FIG. 4. The full-field distribution of z = 0 surfaces for Example 2. (a) t = 0.05, (b) t = 0.08, (c) t = 0.1,
(d) t = 0.15 (1062 points, 6 eigenfunctions). (— : analytical solution; - - - : numerical solution).

10 eigenmodes. This also demonstrates a big advantage for the proposed MFS-MPS-EEM model
to easily combine the wave and diffusion problems together for multiple-dimensional complex
domains. Table I shows the first 10 eigenvalues of numerical results of the MFS with the SVD.
The eigenvalue presents the smallest singular value for various k, which result in a spike when the
solution of the Helmholtz equation is numerically singular. Figure 5 portrays the time evolution
history of the present numerical code and FEM solutions at four arbitrary locations. It is observed
that the present numerical results are almost identical to FEM solutions as the time evolves. In
contrast to numerical solutions obtained at some discrete time levels by conventional methods,
the results by the MFS-MPS-EEM model always render continuous semi-analytic solutions along
the time axis.

Numerical Methods for Partial Differential Equations DOI 10.1002/num

 3D NONHOMOGENEOUS DIFFUSION EQUATIONS 207

FIG. 5. Comparison of time evolution of u at (a) (0.003, 0.199, 0.621), (b) (−0.176, 0.282, 0.598), (c)
(−0.191, −0.292, 0.5960), and (d) (0.278, −0.141, 0.559) for Example 3 (1150 points, 6 eigenfunctions).

Example 4. For the last problem, the proposed method is finally utilized to study the nonhomo-
geneous diffusion equation in more complicated irregular domain, the spiral cavity [Fig. 1(d)]. A
3D nonhomogeneous diffusion problem with nonhomogeneous source function and IC as well as
BCs is given below:
 
∂ 2u ∂ 2u ∂ 2u 1 ∂u
DE + 2+ 2 + 3ex − 2 sin y sin z =
∂x 2 ∂y ∂z k ∂t
IC x , 0) = xyz − 3ex − sin y sin z
u(
in

BCs x , t) = −3ex − sin y sin z
u(
on ∂ (23)

Numerical Methods for Partial Differential Equations DOI 10.1002/num

208 YOUNG ET AL.

TABLE II. Comparison of u with different eigenfunctions for Example 4 (t = 1 and 360 points).
Eigenfunctions
Positions 4 6 8 FEM solution
(0.3, 0.3, 0.6) −8.700457 −8.702140 −8.703516 −8.703501
(0.4, 0.4, 0.8) −6.364161 −6.365894 −6.366938 −6.366726
(0.5, 0.5, 1.0) −5.561102 −5.573606 −5.574195 −5.574472
(0.6, 0.6, 1.2) −3.845279 −3.878102 −3.889741 −3.889137
(0.66, 0.66, 1.33) −3.513890 −3.526476 −3.526705 −3.527706

In this irregular domain, it is difficult to find an analytical solution; so, we choose FEM with
tetrahedron elements to obtain the numerical results for comparison. The distributions of 10,300
structured linear tetrahedron elements for FEM meshes and only 232 meshless MFS nodes are
adopted in this example. Table I shows the numerical results of the first 10 eigenvalues with
MFS-SVD scheme. The eigenvalue will represent the smallest singular value for various k, which
results in a spike when the Helmholtz equation solution is numerically singular. And the first
8 eigenvalues and eigenfunctions are obtained by the MFS-EEM with SVD using 360 points to
interpolate eigenfunctions. The numerical solutions at t = 1 (near steady state solutions) obtained
by different eigenfunctions are shown in Table II. It can be observed that the solutions with more
eigenfunctions are closer to FEM results. Table III shows the near steady state solutions at t = 1
of the MFS-MPS-EEM model at some locations for different collocating points for interpolating
eigenfunctions. The calculated solutions and the FEM results are very close, which demonstrates
the capability of the present model to apply to different shapes of geometry. The results of the
MFS-MPS-EEM model with SVD method match very well with the FEM solutions, which use
10,300 linear tetrahedron elements as also depicted in Fig. 6 for some selected points of time evo-
lution as far as transient solutions are concerned. Hence, it is validated that the present numerical
method can be used appropriately for problems in irregular domain in a very simple and effective
computation.

V. CONCLUSIONS

Usages of the MFS-MPS-EEM model together with the SVD technique to solve the transient
inhomogeneous diffusion problems with time-independent source terms and boundary conditions
in three dimensions are described in this article. The nonhomogeneous diffusion problems are
decomposed into a Poisson equation and a homogeneous diffusion equation. As far as the solution

TABLE III. Comparison of u with different collocating points for interpolating eigenfunctions for Example
4 (t = 1 and 8 eigenfunctions).
Different collocating points for interpolating eigenfunctions
Positions 150 250 360 FEM solution
(0.3, 0.3, 0.6) −8.037162 −8.604939 −8.703516 −8.703501
(0.4, 0.4, 0.8) −6.079654 −6.201250 −6.366938 −6.366726
(0.5, 0.5, 1.0) −5.280098 −5.485165 −5.574195 −5.574472
(0.6, 0.6, 1.2) −3.440513 −3.814367 −3.889741 −3.889137
(0.66, 0.66, 1.33) −3.161461 −3.430545 −3.526705 −3.527706

Numerical Methods for Partial Differential Equations DOI 10.1002/num

 3D NONHOMOGENEOUS DIFFUSION EQUATIONS 209

FIG. 6. Comparison of time evolution of u at (a) (0.3, 0.3, 0.6), (b) (0.5, 0.5, 1.0), and (c) (0.66, 0.66, 1.33)
relative location points for Example 4 (360 points, 8 eigenfunctions)

Numerical Methods for Partial Differential Equations DOI 10.1002/num

210 YOUNG ET AL.

of Poisson equation is concerned, the MFS is adopted to obtain the homogeneous solution and
the MPS is utilized to solve the particular solution. On the other hand, the homogeneous diffusion
equation is first transformed by the EEM into a Helmholtz equation, which is then solved by the
MFS together with SVD scheme to obtain the corresponding eigenvalues and eigenfunctions.
The numerical scheme developed in the present work was validated by comparing with the ana-
lytical solutions and FEM results for 3D diffusion problems under Dirichlet boundary conditions.
Excellent agreements with the analytical and FEM results indicate the effectiveness of the present
method to solve 3D diffusion equations with time-independent source terms and boundary condi-
tions. Hence, it is concluded that the proposed method is capable to obtain reasonable results for
multidimensional nonhomogeneous diffusion equations in arbitrary domains if time-independent
source terms and boundary conditions are assumed. The extension of the MFS-MPS-EEM model
for diffusion problems with time-dependent sources terms will be another interesting topic for
future investigation.

The authors thank the reviewers for their constructive comments to help their for their improving
paper writing.

References
1. S. P. Zhu, Solving transient diffusion problems: time-dependent fundamental solution approaches versus
LTDRM approaches, Eng Anal Bound Elem 21 (1998), 87–90.
2. M. Zerroukat, A boundary element scheme for diffusion problems using compactly supported radial
basis functions, Eng Anal Bound Elem 23 (1999), 201–209.
3. A. Sutradhar, G. H. Paulino, and L. J. Gray, Transient heat conduction in homogeneous and nonhomo-
geneous materials by the Laplace transform Galerkin boundary element method, Eng Anal Bound Elem
26 (2002), 119–132.
4. D. Nardini and C. A. Brebbia, A new approach to free vibrations analysis using boundary elements.
Boundary element in engineering, C. A. Brebbia, editor, CMP, Southamptom and Springer, Berlin 1982,
pp. 312–326.
5. C. S. Chen, M. A. Golberg, and Y. C. Hon, The method of fundamental solutions and quasi-Monte-Carlo
method for diffusion equations, Int J Numer Methods Eng 43 (1998), 1421–1435.
6. C. S. Chen, M. A. Golberg, and Y. F. Rashed, A mesh free method for linear diffusion equations, Numer
Heat Transf B 33 (1998), 469–486.
7. M. A. Golberg and C. S. Chen, The method of fundamental solutions for potential, Helmholtz and
diffusion problems, Boundary integral methods: Numerical and mathematical aspects, Computational
Mechanics Publications, Boston, 1999, pp. 103-176.
8. K. Balakrishnan and P. A. Ramachandran, The method of fundamental solutions for linear diffusion-
reaction equations, Math Comput Model 31 (2002), 221–237.
9. D. L. Young, C. C. Tsai, K. Murugesan, C. M. Fan, and C. W. Chen, Time-dependent fundamental
solutions for homogeneous diffusion problems, Eng Anal Bound Elem 28 (2004), 1463–1473.
10. D. L. Young, C. C. Tsai, and C. M. Fan, Direct approach to solve nonhomogeneous diffusion problems
using fundamental solutions and dual reciprocity methods, J Chin Inst Eng 27 (2004), 597–609.
11. D. L. Young, C. M. Fan, C. C. Tsai, C. W. Chen, and K. Murugesan, Eulerian-Lagrangian method of
fundamental solutions for multi-dimensional advection-diffusion equation, Int Math Forum 1 (2006),
687–706.
12. D. L. Young, S. P. Hu, C. M. Fan, and C. W. Chen, Method of fundamental solutions and Cole-Hopf
transformation for Burgers’ equations, Comput Fluid Dyn J 14 (2006), 454–467.

Numerical Methods for Partial Differential Equations DOI 10.1002/num

 3D NONHOMOGENEOUS DIFFUSION EQUATIONS 211

13. D. L. Young, C. M. Fan, S. P. Hu, and S. N. Atluri, The Eulerian-Lagrangian method of fundamen-
tal solutions for two-dimensional unsteady Burgers’ equations, Eng Anal Bound Elem (doi: 10.1016/
j.enganabound.2007.08.011).
14. Z. Yao and G. F. Margrave, Wavefield extrapolation in laterally inhomogeneous media by generalized
eigenfunction transform method, CREWES Res Rep 11, 1999.
15. D. L. Young, C. W. Chen, C. M. Fan, and C. C. Tsai, The method of fundamental solutions with eigen-
function expansion method for nonhomogeneous diffusion equation, Numer Meth Part Differ Equat 22
(2006), 1173–1196.
16. J. T. Chen, H.-K. Hong, C. S. Yeh, and S. W. Chyuan, Integral representations and regularizations for a
divergent series solution of a beam subjected to support motions, Earthquake Eng Struct Dyn 25 (1996),
909–925.
17. M. A. Golberg, The method of fundamental solutions for Poisson’s equation, Eng Anal Bound Elem 16
(1995), 205–213.
18. C. S. Chen and R. A. Schaback, Recent developments of the dual reciprocity method using com-
pactly supported radial basis functions, Transformation of the domain effects to the boundary, editor,
Y. F. Rashed, WIT Press, 2003, pp. 183–225.
19. W. Madych, Miscellaneous error bounds for multiquadric and related interpolants, Comput Math Appl
24 (1992), 121–138.
20. C. S. Chen, C. A. Brebbia, and H. Power, Dual reciprocity method using compactly supported radial
basis functions, Commun Numer Methods Eng 15 (1999), 137–150.
21. A. H.-D. Cheng, D. L. Young, and C. C. Tsai, Solution of Poisson’s equation by iterative DRBEM
using compactly-supported, positive-definite radial basis function, Eng Anal Bound Elem 24 (2000),
549–557.
22. D. L. Young, C. C. Tsai, and A. H.-D. Cheng, Solution of Stokes flow using an iterative DRBEM based
on compactly-supported, positive-definite radial basis function, Comput Math Appl 43 (2002), 607–619.
23. J. T. Chen, I. L. Chen, and Y. T. Lee, Eigensolutions of multiply connected membranes using the method
of fundamental solutions, Eng Anal Bound Elem 29 (2005), 166–174.
24. C. C. Tsai, D. L. Young, C. W. Chen, and C. M. Fan, The method of fundamental solutions for
eigenproblems in domains with and without interior holes, Proc R Soc A 462 (2006), 1443–1466.
25. D. L. Young, S. P. Hu, C. W. Chen, C. M. Fan, and K. Murugesan, Analysis of elliptical waveguides by
the method of fundamental solutions, Microw Opt Technol Lett 44(6) (2005), 552–558.

Numerical Methods for Partial Differential Equations DOI 10.1002/num