See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/223201060

The method of fundamental solutions for solving incompressible Navier–

Stokes problems

Article in Engineering Analysis with Boundary Elements · August 2009

DOI: 10.1016/j.enganabound.2009.03.003

CITATIONS READS

45 315

4 authors, including:

Der-Liang Young Chia-Ming Fan

National Taiwan University National Taiwan Ocean University
259 PUBLICATIONS 6,585 CITATIONS 139 PUBLICATIONS 3,667 CITATIONS

SEE PROFILE SEE PROFILE

Chiu C.L.
Da-Nan Elementary school, Taichung City,Taiwan.
19 PUBLICATIONS 299 CITATIONS

SEE PROFILE

All content following this page was uploaded by Der-Liang Young on 24 November 2023.

The user has requested enhancement of the downloaded file.

 ARTICLE IN PRESS
Engineering Analysis with Boundary Elements 33 (2009) 1031–1044

Contents lists available at ScienceDirect

Engineering Analysis with Boundary Elements

journal homepage: www.elsevier.com/locate/enganabound

The method of fundamental solutions for solving incompressible

Navier–Stokes problems
D.L. Young , Y.C. Lin, C.M. Fan 1, C.L. Chiu
Department of Civil Engineering and Hydrotech Research Institute, National Taiwan University, Taipei 10617, Taiwan

a r t i c l e in fo abstract

Article history: A novel meshless numerical procedure based on the method of fundamental solutions (MFS) is
Received 21 November 2008 proposed to solve the primitive variables formulation of the Navier–Stokes equations. The MFS is a
Accepted 11 March 2009 meshless method since it is free from the mesh generation and numerical integration. We will
Available online 16 April 2009
transform the Navier–Stokes equations into simple advection–diffusion and Poisson differential
Keywords: operators via the operator-splitting scheme or the so-called projection method, instead of directly
Navier–Stokes equations using the more complicated fundamental solutions (Stokeslets) of the unsteady Stokes equations. The
Meshless numerical method resultant velocity advection–diffusion equations and the pressure Poisson equation are then calculated
Method of fundamental solutions by using the MFS together with the Eulerian–Lagrangian method (ELM) and the method of particular
Method of particular solutions
solutions (MPS). The proposed meshless numerical scheme is a first attempt to apply the MFS for
Eulerian–Lagrangian method
solving the Navier–Stokes equations in the moderate-Reynolds-number flow regimes. The lid-driven
Operator-splitting method
cavity flows at the Reynolds numbers up to 3200 for two-dimensional (2D) and 1000 for three-
dimensional (3D) are chosen to validate the present algorithm. Through further simulating the flows in
the 2D circular cavity with an eccentric rotating cylinder and in the 3D cube with a fixed sphere inside,
we are able to demonstrate the advantages and flexibility of the proposed meshless method in the
irregular geometry and multi-dimensional flows, even though very coarse node points are used in this
study as compared with other mesh-dependent numerical schemes.
& 2009 Elsevier Ltd. All rights reserved.

1. Introduction method of fundamental solutions (MFS) is currently the most

well-known method in the meshless numerical schemes.
In the computational fluid dynamics realm, many methodol- Some excellent literature reviews [1,2] have highlighted the
ogies have been developed to solve the Navier–Stokes equations. advantages of the MFS over other domain discretization numerical
Among them classical domain discretization numerical methods schemes. The features of boundary-type meshless methods and
have made tremendous progresses, however, in some cases the well-derived fundamental solutions have brought a timely
higher dimensional and complex-geometry problems still remain opportunity for the MFS as a powerful numerical method in
very hard and tedious to deal with. The mesh generation is a dealing with homogeneous partial differential equations. Ku-
major hurdle for the high dimensional and complex-geometry pradze and Aleksidze [3] first proposed the concept of the MFS,
flow simulations especially for three-dimensional (3D) problems and then Mathon and Johnston [4] used the MFS to numerically
by using mesh-dependent numerical methods, because the main solve the elliptic boundary value problems. Bogomolny [5]
effort of the simulation is spent at the stage of mesh generation. established the preliminary mathematical framework for the
The meshless (or the so-called mesh-free) methods are recently MFS and fixed the positions of the source distribution. The MFS
considered as a promising alternative to overcome the problem of was subsequently applied to some linear partial differential
mesh generation. The meshless methods require neither domain equations such as the eigenvalue problems [6], the biharmonic
nor surface grid generation, and only randomly distributed nodes equation [7]. The steady Stokes flow problems were also solved by
are necessary to describe the nodal distances between the source using the steady Stokeslets [8]. Golberg [9] and Golberg et al. [10]
and field points during the numerical implementation. The further extended the MFS to find a particular solution for the
inhomogeneous term of the Poisson equation, by using the
method of particular solutions (MPS). Young et al. [11] solved
 Corresponding author. Fax: +886 2 23626114.
the time-dependent diffusion equation by the diffusion funda-
E-mail address: [email protected] (D.L. Young).
mental solution without temporal finite-difference discretization.
1
Now at Department of Harbor and River Engineering, National Taiwan Ocean Young et al. [12,13] further combined the Eulerian–Lagrangian
University, Keelung 20224, Taiwan. method (ELM) with the MFS to successfully handle problems with

0955-7997/$ - see front matter & 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.enganabound.2009.03.003
 ARTICLE IN PRESS
1032 D.L. Young et al. / Engineering Analysis with Boundary Elements 33 (2009) 1031–1044

a convective term. The MFS has effectively solved the advection– 2. Mathematical model
diffusion [12] and quasi-linear Burgers’ equations [13] by the
extended ELMFS, since the ELM is able to convert the advection– 2.1. Governing equations
diffusion equation into the characteristic diffusion equation in the
Lagrangian coordinates. Thereafter, the convective term in the The motion of viscous incompressible flows is governed by the
advection–diffusion and Burgers’ equations can be obtained by following Navier–Stokes equations (nondimensional) in the
back-tracking scheme via the characteristics through the solution primitive variables form.
of the linear diffusion equation [11–13]. Continuity equation:
Although the MFS and the extended ELMFS have been widely
used in the linear and some quasi-linear partial differential r ~
uð~
x; tÞ ¼ 0 (1)
equations including the convective term, it is still a big challenge Momentum equation:
at present to use the MFS to solve the notorious Navier–Stokes
equations or even the unsteady Stokes equations [14,15]. The @~
uð~
x; tÞ 1 2
þ ð~ x; tÞ  rÞ~
uð~ uð~
x; tÞ ¼ rpð~
x; tÞ þ r~
uð~
x; tÞ (2)
fundamental solutions for the Navier–Stokes equations could not @t Re
be directly obtained due to the nonlinear convection term in the where ~x is the space vector, t is the time, ~
u ¼ ðu; v; wÞ, p are the
momentum equations. However, similar to the advection–diffu- dimensionless velocity and pressure, respectively, Re ¼ UL/n is the
sion equations, the ELM can transform the Navier–Stokes Reynolds number, U indicates a characteristic velocity, L is a
equations into unsteady Stokes equations in the Lagrangian characteristic length and n represents the kinematic viscosity of
coordinates, so that we are able to use the unsteady Stokeslets the fluid.
in the Lagrangian coordinates to numerically solve the Navier– The computational regime is defined in an unified space–time
Stokes equations [15]. The unsteady Stokeslets are the funda- domain, O with the initial and boundary conditions. The boundary
mental solutions of the unsteady Stokes operators [14]. Due to the conditions for the velocity and pressure are specified by the
constraint of incompressibility from the continuity equation Dirichlet part and Neumann part of the boundary GD and GN of G,
[14,15], the unsteady Stokeslets not only produce a complex respectively. In this study the boundary conditions for velocity are
coupled system of unsteady equations, but also introduce the non-slip boundaries, while the boundary conditions for the
the Dirac delta function which is rather hard for the method pressure consist of two parts, the Dirichlet part is the reference
of collocation [14,15]. The combination of unsteady Stokeslets pressure, pr and the Neumann parts are the momentum equations
and ELM therefore continued to be very difficult to apply specified at the boundaries. For high Reynolds number viscous
to the Navier–Stokes equations especially for the moderate or flows, however, the Neumann boundary condition for the pressure
high Reynolds number (Re) flow regimes, as reported by Young Poisson equation can be simply approximated by qp/qn ¼ 0 due to
et al. [15]. the nature of non-slip wall boundary and high Reynolds number.
The operator-splitting or projection method [16–18], however, Here n is the direction of the unit outward normal vector, ~ n on the
is another alternative to replace the unsteady Stokeslets for boundary GN.
advancing the Navier–Stokes equations in time, such that the The Reynolds number is the ratio of the inertial terms to the
resultant partial differential operators can be solved by using the viscous terms. When the Re becomes large, the inertial forces will
MFS without additional difficulty. The operator-splitting methods dominate the flow field. As a result, we will anticipate the facts
have different types depending on the methods of approximating that thin boundary layer is developed, and more difficulty to
the advection term and the schemes used for the time discretiza- compute these nonlinear and instable flow regimes. Since the
tion, and over the years they have undergone many modifications governing equations involve the convective terms in the momen-
[19–21]. The present work thus combines the operator-splitting tum equations, it is rather difficult to directly obtain the
method with the MFS for the solutions of the incompressible fundamental solutions for the MFS. In order to obtain the MFS
Navier–Stokes equations. solutions for the Navier–Stokes equations, some mathematical
The purpose of this paper is to develop a novel meshless treatment must be employed to simplify the original operators.
numerical scheme based on the operator-splitting method to Therefore a very popular formulation, which is named the
accurately and efficiently solve the Navier–Stokes problems with operator-splitting or projection method, has been adopted here.
moderate Reynolds numbers. The basic frameworks follow the In the following, we will briefly describe the procedures of the
operator-splitting or projection method originated by Chorin [16] operator-splitting or projection method.
and Temam [17] and adopt the ELMFS developed by Young et al.
[12,13] and the MFS–MPS proposed by Golberg and his collabora-
tors [9,10]. Meanwhile, the main features of the developed model 2.2. Operator-splitting or projection methods
include the following: (1) the temporal discretization employs the
operator-splitting method; (2) both spatial and temporal dis- Chorin [16] and Temam [17] independently proposed the
cretizations use the meshless numerical scheme; (3) the ELMFS is projection method to solve the Navier–Stokes equations,
introduced to solve the velocity advection–diffusion equations; which has become a classical tool for computational fluid
and (4) the MFS–MPS is employed to find the solution of the mechanics. The pressure gradient terms are dropped from the
pressure Poisson equation. To elucidate predominant features, the momentum equations in the first step. The resultant advection–
developed meshless numerical procedure is used to study some diffusion equations produced a provisional velocity u*. There
classical benchmark flows and also multi-dimensional irregular are various methods of approximating the advection term and
domain problems. The two-dimensional (2D) and 3D lid-driven the schemes used for the time discretization for the provi-
cavity flow problems are widely used to validate the accuracy of sional velocity. An explicit time discretization takes the following
new numerical methods. This study first examines 2D and 3D lid- simple form
driven cavity flow problems. After validating the 2D and 3D   n
benchmark cavity flows, the present algorithm concentrates on @~
u u ~
~ u n n 1 2 n
¼ ¼ ð~
u  rÞ~
u þ r~
u (3)
simulating the more complicated flow problems in irregular @t Dt Re
geometries, which are relatively difficult for mesh generation in On the other hand, an implicit scheme as used by Goldberg and
the conventional domain discretization methods. Ruas [19] which is governed by the advection–diffusion equations
 ARTICLE IN PRESS
D.L. Young et al. / Engineering Analysis with Boundary Elements 33 (2009) 1031–1044 1033

is as follows. This is the equation of provisional velocity adopted From the above procedures, the MFS, ELMFS and MPS are the
in this study. essential techniques to solve the Navier–Stokes equations. The
 following sections will depict the detailed meshless numerical
@~
u   1 2 
þ ð~
u  rÞ~
u ¼ r~
u (4) formulations.
@t Re
In the second step, the provisional velocity has been corrected
from the pressure gradient term and the continuity equation 3. Meshless numerical modeling
using the operator-splitting scheme by considering
nþ1  3.1. The ELMFS for the advection–diffusion equations
~
u ~
u
þ rpnþ1 ¼ 0 (5)
Dt
The MFS can solve the governing equations of the physical
By taking the divergence of Eq. (5) and subjecting to the problems including the convective terms with the coordination of
nþ1
continuity constraint r  ~
u ¼ 0, we are able to obtain the the ELM. In the present ELMFS method, Eq. (7) can be transformed
pressure Poisson equation as follows: into a diffusion equation in the Lagrangian coordinates and can be
r ~
u
 written as follows.
r2 pnþ1 ¼ (6)
Dt D~
u

1 2 
The solution procedure consists of first computing ~

u from ¼ r~
u (11)
Dt Re
Eq. (4). Therefore, the second step is to solve the pressure Poisson
where the material derivative is defined as [12,13]
equation, Eq. (6). When we have the provisional velocity and
pressure, the new velocity at the next time step can be D @ 
¼ þ ð~
u  rÞ (12)
determined from Eq. (5). Therefore, the operator-splitting or Dt @t
projection procedures are described in the following steps: Since the advection–diffusion equations are transferred to the
characteristic diffusion equations, the MFS is first adopted to solve
i. Provisional velocity the linear diffusion equations [11]. The MFS based on the time-
 dependent fundamental solutions can be used to solve the linear
@~
u   1 2 
þ ð~
u  rÞ~
u ¼ r~
u (7) homogeneous diffusion equation represented by Eq. (11) as
@t Re
follows:
ii. Pressure calculation

@~
u 1 2 
2 nþ1 1  ¼ r~
u (13)
r p ¼ r ~ u (8) @t Re
Dt
The fundamental solution of the linear diffusion equation is
iii. Velocity correction
governed by the following equation:
nþ1 
~
u u  Dtðrpnþ1 Þ
¼~ (9)
@Gð~x; t : ~
x; tÞ 1 2
Using the ELM, the advection–diffusion equations, represented ¼ r Gð~x; t : ~ x ~
x; tÞ þ dð~ xÞdðt  tÞ (14)
@t Re
by Eq. (4), can be transformed into characteristic diffusion
where Gð~ x; t : ~
x; tÞ is the fundamental solution of the linear
equations in the Lagrangian coordinates which form a decoupled
diffusion equation. ~ x ¼ ðx; y; zÞ and ~ x ¼ ðx; Z; zÞ are the spatial
system of differential equations. The diffusion solution can be
obtained as a combination of the diffusion fundamental solutions
based on the time-dependent MFS procedure [11].
The operator-splitting or projection method ensures the diver- Field Points ( x )
gence-free velocity field. The solutions for the pressure Poisson Source Points ( )
equation can be solved accurately by integrating the MFS with MPS. t
The value of the pressure can be expressed as a linear combination of
the homogeneous solution and particular solution as shown below
p ¼ ph þ pp (10)
Once the particular solution, pp, is obtained from the MPS, we can (n+1) Δt
get the modified boundary conditions for the homogeneous solution,
ph. According to Eq. (8), the MPS and MFS can acquire the desired
pressure field. The following steps describe the complete procedures: (n) Δt x
1. Input the initial conditions ~u0 and the boundary conditions ~ ub
for velocity.
2. Solve the unsteady advection–diffusion equations, Eq. (4), by (n-λ) Δt
employing the ELMFS to find the provisional velocity compo-

nents ~u . t
3. Calculate the particular solutions for pressure Poisson equa-
(n+1) Δt
tion, Eq. (6) by the MPS. C A
Solve the pressure Poisson equation, Eq. (6), by using the
MFS–MPS scheme to find the pn+1.
4. Find the new velocity ~ u
nþ1
by correcting the provisional (n) Δt
B x
velocity using Eq. (9).
5. Return to Step 2 if the solution does not approach steady or
terminal time. Output the solutions if it is already steady or has Fig. 1. (a) Schematic diagram for the location of source points and field points on
reached the terminal time. the 1D space–time domain and (b) schematic diagram for the characteristic AB.
 ARTICLE IN PRESS
1034 D.L. Young et al. / Engineering Analysis with Boundary Elements 33 (2009) 1031–1044

coordinates of the field and source nodes, t and t are the temporal source nodes for space should be promoted to next time level
coordinates of the field and source nodes, respectively. d() is the (n+1)Dt but different temporal positions for time as shown in
Dirac delta function. Fig. 1(a) through the time marching process. In Fig. 1(a), the
By employing the integral transform theory for Eq. (14), the parameter l is an arbitrary factor and Young et al. [13] have
Green’s function or the fundamental solution of the diffusion proposed an empirical formula to determine the temporal
equation can be obtained as follows [11]: location of the source nodes, which is chosen as a function of
the maximum distance of the spatial domain. Then by collocating
x~
ðj~ xjÞ2 =4ð1=ReÞðttÞ
e the initial and boundary conditions, the matrix equation can be
x; t : ~
Gð~ x; tÞ ¼  d=2 Hðt  tÞ (15)
4pð1=ReÞðt  tÞ formed by utilizing Eq. (16) and is listed below.

where d is the spatial dimension of the problem and H() is the ½Afag ¼ fgg (17)
Heaviside step function. The component of [A] is the representation of the funda-
The solution of the diffusion equation can be obtained by a mental solutions and {g} is the combination of initial and
linear combination of the diffusion fundamental solution, since boundary conditions. Once the unknown coefficients {a} are
the diffusion fundamental solution satisfies the homogeneous
diffusion equation everywhere except the source points. Solutions
in Eq. (11) thus can be expressed as:

X
N
1.00
u ð~
x; tÞ ¼ x; t; ~
aj Gð~ xj ; tj Þ (16)
j¼1

where N is the number of source nodes. aj is the unknown

coefficients which denote the source intensities of the corre- 0.80
sponding fundamental solution. Thereafter, the velocity of any
field nodes in the time–space domain can be solved accordingly
by using Eq. (16) with the imposition of non-slip boundary Ghia et al.
0.60
conditions. Present (CDN=6.4E06)
In our numerical experiments, Fig. 1(a) illustrates the locations Present (CDN=7.0E12)
Y

of the field and source nodes for space–time unified domain. For Present (CDN=4.9E16)
time-independent boundary conditions, the field and source 0.40 Present (CDN=1.4E17)
nodes for space are located at the same spatial positions and Present (CDN=8.7E18)
with the same number but different temporal positions for time.
However, for time-dependent boundary conditions, the field and
0.20

Field point
0.00
-0.40 0.00 0.40 0.80 1.20
U

Ω Γ 0.20
Ghia et al.
Present (CDN=6.4E06)
Present (CDN=7.0E12)
0.10 Present (CDN=4.9E16)
Present (CDN=1.4E17)
Present (CDN=8.7E18)
Source point
Field point 0.00
V

-0.10

-0.20
Ω Γ

-0.30
0.00 0.20 0.40 0.60 0.80 1.00
X

Fig. 2. Schematic diagram for the distribution of nodes on 2D computational Fig. 3. The velocity profile with different source location at Re ¼ 100 with global
domain. (a) The nodes distribution for the global grid and (b) the nodes grid 31  31 and boundary grid 31  4. (a) The u-velocity profiles along x ¼ 0.5 and
distribution for the boundary grid. (b) the v-velocity profiles along y ¼ 0.5.
 ARTICLE IN PRESS
D.L. Young et al. / Engineering Analysis with Boundary Elements 33 (2009) 1031–1044 1035

solved by inverting the above matrix equation, the function increments as

values of the diffusion equation inside the time–space box at
dx xnþ1  xn
t ¼ (n+1)Dt is acquired by Eq. (16). It is worth mentioning u¼ ¼ (18)
dt Dt
that we did not need to deal with the inhomogeneous term
due to the time discretization by using present time-space dy ynþ1  yn
unification scheme of the fundamental solution of the diffusion v¼ ¼ (19)
dt Dt
equation [11].
To obtain the solutions of advection–diffusion equations, we dz znþ1  zn
need to transfer the advection–diffusion equations into the w¼ ¼ (20)
dt Dt
characteristic diffusion equations in the Lagrangian coordinates
We thus can have the following trajectory relationships along
through the ELM and then by back-tracking the particles along the
the characteristics of the advection–diffusion equations for the
line of characteristics from the linear diffusion results. Through
back-tracking purposes.
the ELM, the convective or characteristic velocities in advection–
diffusion equations can be expressed in terms of the time xn ¼ xnþ1  un Dt (21)

1.00 1.00

0.80 0.80

Ghia et al.
44 nodes
0.60 0.60
84 nodes
Ghia et al. 124 nodes
Y

121 nodes 164 nodes

441 nodes
0.40 961 nodes 0.40

0.20 0.20

0.00 0.00
-0.40 0.00 0.40 0.80 1.20 -0.40 0.00 0.40 0.80 1.20
U U

0.20 0.20
Ghia et al.
121 nodes
Ghia et al.
441 nodes
44 nodes
0.10 961 nodes 0.10
84 nodes
124 nodes
164 nodes

0.00 0.00
V

-0.10 -0.10

-0.20 -0.20

-0.30 -0.30
0.00 0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00
X X

Fig. 4. The velocity profile at Re ¼ 100 with different number of global grid and Fig. 5. The velocity profile at Re ¼ 100 with different number of boundary grid and
fixed boundary grid 31  4. (a) The u-velocity profiles along x ¼ 0.5 and (b) the v- fixed global grid 31  31. (a) The u-velocity profiles along x ¼ 0.5 and (b) the v-
velocity profiles along y ¼ 0.5. velocity profiles along y ¼ 0.5.
 ARTICLE IN PRESS
1036 D.L. Young et al. / Engineering Analysis with Boundary Elements 33 (2009) 1031–1044

u compoment u compoment
-1.00 -0.50 0.00 0.50 1.00 -1.00 -0.50 0.00 0.50 1.00
1.00 1.00 1.00 1.00

Present Present
Ghia et al. Ghia et al.
0.80 0.80
0.50 0.50

v component

v component
0.60 0.60
y axis

y axis
0.00 0.00

0.40 0.40

-0.50 -0.50
0.20 0.20

0.00 -1.00 0.00 -1.00

0.00 0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00
x axis x axis

u compoment u compoment
-1.00 -0.50 0.00 0.50 1.00 -1.00 -0.50 0.00 0.50 1.00
1.00 1.00 1.00
1.00

Present Present
Ghia et al. Ghia et al.
0.80
0.80
0.50 0.50

v component
v component

0.60
0.60
y axis

y axis

0.00 0.00

0.40 0.40

-0.50 -0.50
0.20 0.20

0.00 -1.00 0.00 -1.00

0.00 0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00
x axis x axis

Fig. 6. Velocity profiles at horizontal and vertical line of 2D lid-driven cavity for (a) Re ¼ 100, (b) Re ¼ 400, (c) Re ¼ 1000, and (d) Re ¼ 3200.

yn ¼ ynþ1  vn Dt (22) the advection–diffusion equations at t ¼ (n+1)Dt can be obtained

[12,13]. This procedure is repeated until either the terminal time
or the steady-state solution is reached.
zn ¼ znþ1  wn Dt (23)

In Fig. 1(b), the line AB is the characteristic path on which the

transport of a scalar quantity can be better illustrated by a one- 3.2. The MFS–MPS for the Poisson equation
dimensional (1D) problem. If the velocity at event A is required,
we can trace the spatial location of the event B from The MFS had been extended to solve the Poisson equation without
Eqs. (21)–(23). Once the spatial location of event B is decided, boundary or domain discretization, since the introduction of the
the solution along the characteristic AB will follow the diffusion MFS–MPS for the Poisson equations was made by Golberg and his co-
equations. After calculating the diffusion process by the MFS, the workers [9,10]. The Poisson equation governs the pressure-correction
velocities at event C can be obtained to represent the velocities at step in the operator-splitting formulation, and it is necessary to
event A. Event B and C are at the same spatial position but located accurately solve the pressure Poisson equation to ensure the velocity
at different time level. The velocities at event A is properly field divergence-free. This section will introduce the combination of
replaced by the diffusion results at event C, such that the results of the MPS and MFS as a meshless method for the Poisson equation.
 ARTICLE IN PRESS
D.L. Young et al. / Engineering Analysis with Boundary Elements 33 (2009) 1031–1044 1037

0.5

Y
φ

-0.5

-1
-1 -0.5 0 0.5 1
X

1.50 1.50
Global grid Source points
Field points
1.00 1.00

0.50 0.50

0.00 0.00
Y

-0.50 -0.50

-1.00 -1.00

-1.50 -1.50
-1.50 -1.00 -0.50 0.00 0.50 1.00 1.50 -1.50 -1.00 -0.50 0.00 0.50 1.00 1.50
X X

Fig. 7. (a) Schematic diagram of the computational domain, (b) unstructured finite element mesh, (c) the distribution of global grid, and (d) the distribution of boundary
grid.

The Poisson equation takes the following form which will be where the particular solution,pp ð~
xÞ, satisfies
elucidated as follows:
r2 pp ð~
xÞ ¼ að~
xÞ (27)

2 r ~
u
r pð~
xÞ ¼ ¼ að~
xÞ in O (24)
Dt and the homogeneous solution, ph ð~
xÞ, satisfies

xÞ ¼ pr ¼ bð~
pð~ xÞ on GD r2 ph ð~
xÞ ¼ 0 in O (28)
~
@pðxÞ 1
¼ ðr 2 ~
uÞ  ~ xÞ on GN
n ¼ cð~ (25)
@n Re ph ð~
xÞ ¼ bð~ xÞ on GD
xÞ  pp ð~
In the MFS–MPS formulations, we decompose the solution pð~ xÞ @ph ð~
xÞ @pp ð~
xÞ
¼ cð~
xÞ  on GN (29)
into the particular solution, pp ð~
xÞ, and the homogeneous solution, @n @n
ph ð~
xÞ. That is,
Then the MPS is adopted to solve the particular solu-
pð~
xÞ ¼ ph ð~
xÞ þ pp ð~
xÞ (26) tions governed by Eq. (27). We may first interpolate the
 ARTICLE IN PRESS
1038 D.L. Young et al. / Engineering Analysis with Boundary Elements 33 (2009) 1031–1044

1 1

0.5 0.5
Y

Y
0 0

-0.5 -0.5

-1 -1
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
X X

0.5
Y

-0.5

-1
-1 -0.5 0 0.5 1
X
Fig. 8. Streamlines of the 2D circle with an eccentric rotation cylinder for (a) Re ¼ 1, (b) Re ¼ 100, and (c) Re ¼ 400.

right hand side of Eq. (27) by the scheme of the radial basis the following form:
functions (RBFs).
X
N
X
N pp ð~
xÞ ¼ bj Fðrj Þ (31)
að~
xi Þ ¼ bj f ðr ij Þ (30) j¼1
j¼1
where F(r) is the particular solution of the Laplacian operator with
where f(rij) is the RBF, r ij ¼ j~
xi  ~
xj j is the distance between a field inhomogeneous term of the RBF f(r)
point ~xi and j-th collocation point ~ xj , bj are the undetermined
r2 FðrÞ ¼ f ðrÞ (32)
coefficients. N is the number of collocation nodes.
The collocation points are distributed in the interior domain We can apply the MFS to solve the homogeneous solutions
for the particular solutions. The MPS procedures do not need governed by Eqs. (28) and (29), once the particular solutions for
boundary conditions to get the particular solutions, we only need the Poisson equation are found. Because the fundamental solution
to solve the equation with known functions að~ xÞ. After the bj have satisfies Eq. (28), we may assume the homogeneous solution is a
been solved, we are able to find the particular solution, which is in linear combination of the fundamental solution of the Laplacian
 ARTICLE IN PRESS
D.L. Young et al. / Engineering Analysis with Boundary Elements 33 (2009) 1031–1044 1039

operator, i.e.: determine the optimal source location for the MFS via tracking the
condition number (CDN) of the coefficient matrix. Because of the
X
M
ph ð~
xi Þ ¼ xi ; ~
gj hð~ xj Þ (33) finite-precision computation tool such as FORTRAN or C++,
j¼1 the precision is limited to double precision (16 digits). The
condition number of the resultant matrix is utilized to obtain the
where hð~ xi ; ~
xj Þ is the fundamental solution of the Laplace
better results based on the finite precision computation. Here, we
equation, and hð~ xi ; ~
xj Þ ¼ 1/(2p)Log(rij) in 2D and hð~xi ; ~
xj Þ ¼ 1/
~ define the condition number as follows.
(4prij) in 3D. rij ¼ j~ xi  xj j is the distance between a field point ~ xi
and the j-th source point ~ xj and M is the number of the source maximum singular value
nodes. gj are the undetermined coefficients of the source CDN ¼ (36)
minimum singular value
intensities. The source nodes are distributed away from boundary
to avoid the singularity due to the coincidence of ~ xi and ~
xj during In [27], the accurate solution is achieved when the condition
the collocation. The coefficients gj are solved by the method of number approaches to the limit of the equation solver as
collocation through the modified boundary conditions. Therefore, 1017–1018. Therefore, we assume that the results are accurate
we obtain the following matrix equation. and reasonable when the condition number of the coefficient
matrix [H] in Eq. (34) approaches to 1017.
½Hfgg ¼ flg (34)
The component of [H] represents the fundamental solutions and
{l} is the corresponding boundary conditions. The unknown 4.1. 2D lid-driven cavity flows
coefficients {g} could be found by inversing the matrix [H].
After acquiring gj, we can solve the homogeneous solutions
The present scheme is validated by applying to a square cavity,
through Eq. (33). The desired Poisson solutions can be obtained by
in which the top wall is assumed to move parallel to the x-axis
applying Eq. (26), since particular solution pp ð~
xÞ and the homo-
with a constant velocity of unity (dimensionless). The numerical
geneous solution ph ð~ xÞ have been solved.
experiments are performed for four different Reynolds numbers:
Many RBFs have been well described and derived in the
100, 400, 1000 and 3200, using uniform node distribution.
literature [22,23]. Here the thin plate splines (TPSs) [22,23] are
Figs. 3(a) and (b) show results with different source positions as
chosen to solve the particular solution of the Laplacian operator
the Reynolds number is equal to 100, and we clearly observe that
and the corresponding RBFs are shown below.
the results are accurate if the condition number is in the range of
( 2
r LogðrÞ for 2D 1012–1017. The following numerical experiments all follow this
f ðrÞ ¼ restricted condition to find the accurate numerical results.
r for 3D
8 4 4
Figs. 4(a) and (b) show the present results of u–y and x–v plots
>
> r þ 2r LogðrÞ for Re ¼ 100 obtained using different number of global grids by
>
< for 2D
32 the MPS procedures and the comparisons with those of Ghia et al.
FðrÞ ¼ (35)
>
> r3 [25] using multigrid FDM. Figs. 5(a) and (b) show the results for
>
: for 3D
12 u–y and x–v plots solved by different number of boundary grids by
The present study applies the MFS combining with the MPS to the present ELMFS. From those results, we can find that the
solve the pressure-correction equation for the incompressible present numerical scheme is stable and solutions with larger
Navier–Stokes flows. As previously mentioned, the TPSs have been number of grid points are closer to the results of Ghia et al. [25].
demonstrated efficiently and reliably to handle the particular These trends indicate that stable and convergent numerical
solution for the Laplacian operator. procedures are maintained for the Navier–Stokes equations in
this investigation.

4. Numerical results 1.00

We have developed a novel meshless framework to solve the

incompressible viscous flows in 2D and 3D problems. The present Present
method has two different node distributions. One is the global Unstructured FEM
nodes for the MPS, and the other is boundary nodes for the MFS 0.50
and ELMFS. The schematic diagrams for two different grid sets are
represented in Fig. 2. It is observed that the boundary nodes are
v component

the same as nodes from global grids on the boundary. Due to

the pioneer works by Burggraf [24], the lid-driven cavity flows are
0.00
always considered as classical benchmark problems to validate
the new numerical methods for the Navier–Stokes equations.
Therefore, the 2D and 3D lid-driven cavity flows are primarily
chosen to validate our numerical scheme and the numerical
solutions are compared with results in the literature [25,26].
-0.50
However, how to establish a well-conditioned resultant matrix is a
very crucial task in our algorithm. We must locate the source
points outside the computational domain by properly choosing
the positions of the source points to avoid the singularity problem
and obtain better results in solving the Poisson and -1.00
advection–diffusion equations. As claimed by Bogomolny [5], the -1.00 -0.50 0.00 0.50 1.00
results of the MFS could be improved by putting the source points x axis
in a circle outside of the domain with a selected radius R. On the
other hand, Tsai et al. [27] proposed a convenient method to Fig. 9. Velocity profiles along y ¼ 0 at Re ¼ 400.
 ARTICLE IN PRESS
1040 D.L. Young et al. / Engineering Analysis with Boundary Elements 33 (2009) 1031–1044

u compoment u compoment
-1.00 -0.50 0.00 0.50 1.00 -1.00 -0.50 0.00 0.50 1.00
1.00 1.00 1.00 1.00

Present Present
Jiang et al. Jiang et al.
0.80 0.80
0.50 0.50

w component

w component
0.60 0.60
z axis

z axis
0.00 0.00

0.40 0.40

-0.50 -0.50
0.20 0.20

0.00 -1.00 0.00 -1.00

0.00 0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00
x axis x axis

u compoment
-1.00 -0.50 0.00 0.50 1.00
1.00 1.00

Present
Jiang et al.
0.80
0.50

w component
0.60
z axis

0.00

0.40

-0.50
0.20

0.00 -1.00
0.00 0.20 0.40 0.60 0.80 1.00
x axis

Fig. 10. Velocity profiles at horizontal and vertical line of 3D lid-driven cavity for (a) Re ¼ 100, (b) Re ¼ 400, and (c) Re ¼ 1000.

Figs. 6(a)–(d) show the u and v velocity profiles along the vertical moving wall and the source nodes number must rise with the
and horizontal centerlines for cases of Re ¼ 100, 400, 1000 and 3200, increasing Reynolds number. We further note that the sharp
respectively, and four cases all use the uniform grid distribution. In u-velocity gradient still can be well captured near the up moving
the cases of Re ¼ 100 and 400, the global grids 31  31 and boundary wall at Reynolds number 3200.
grids 31  4 are adopted to acquire close solutions. At higher
Reynolds number flows, the increased inertial force results in the
development of the thin boundary layers. This forces us to use a finer 4.2. 2D circular cavity with an eccentric rotating cylinder
grid to completely capture the fine structures of flows. In the case of
Re ¼ 1000, the global grids 31  31 and boundary grids 41  4 are To validate the capacity of the proposed numerical scheme to deal
required. The global grids 41  41 and boundary grids 51  4 are able with irregular geometry and moving boundary, we consider a circular
to obtain good results for Re ¼ 3200. As shown on the above figures, cavity containing a rotating cylinder. Fig. 7(a) sketches the problem of
the present predictions are in close agreement with the classic interest and Figs. 7(c) and (d) show the schematic diagram of the
results of Ghia et al. [25]. They obtained the solutions with very fine nodes distribution. The outer circular cavity is kept stationary with
grids by the multigrid FDM. The increase in the Reynolds number radius R ¼ 1, while the inner rotating cylinder is located at (0.5, 0)
from 100 to 3200 results in the boundary-layer thinning near the with radius r ¼ 0.25. In this system, the tangent velocity of the
 ARTICLE IN PRESS
D.L. Young et al. / Engineering Analysis with Boundary Elements 33 (2009) 1031–1044 1041

Z rotating cylinder and the radius of the outer circle are used to define
Y the Reynolds number. The simulations are performed for three
different Reynolds numbers, Re ¼ 1, 100 and 400, and the results are
compared with the unstructured FEM solutions. Fig. 7(b) shows the
X
unstructured mesh for this problem. In these cases, the inner cylinder
rotates with a constant angular speed (f ¼ 4) and the outer circle is
imposed by the no-slip stationary boundary conditions. Comparing
H=1 with Figs. 8(a)–(c), the streamlines near the cylinder are very similar
for different Reynolds numbers, but one remarkable change in the
shape of the recirculation region is observed. With the increasing
R=0.125 Reynolds number, the recirculation eddy is skewed in the
counterclockwise direction. The numerical validation is chosen for
Re ¼ 400, since the flow condition is more complex as compared with
Re ¼ 1 and 100. Fig. 9 shows the comparison of the velocity profiles
with the results obtained from the unstructured FEM. The present
scheme shows good agreement with the unstructured FEM solutions.
The solutions based on the operator-splitting formulations for the
unstructured FEM have to use 9760 triangle elements to get the same
W=1 accuracy as the MFS algorithm.

4.3. 3D lid-driven cavity flows

L=1
We observe from Eq. (15) that the fundamental solutions of the
Fig. 11. Schematic diagram for 3D lid-driven cavity with a fixed sphere. diffusion equations can be easily extended to 3D problems, and

Fig. 12. u velocity contours at the y–z plane for (a) Re ¼ 1, (b) Re ¼ 100, and (c) Re ¼ 400.
 ARTICLE IN PRESS
1042 D.L. Young et al. / Engineering Analysis with Boundary Elements 33 (2009) 1031–1044

1 1

0.8 0.8

0.6 0.6
Z

Z
0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
X X

0.8

0.6
Z

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
X
Fig. 13. The streamline topologies at the x–z plane for (a) Re ¼ 1, (b) Re ¼ 100, and (c) Re ¼ 400.

the 3D fundamental solution for the Laplacian operator is also behavior, a finer grid is necessary in any type of numerical
widely used in other MFS and BEM studies. The present method scheme. In present method, the flow field for Re ¼ 1000 is
could thus be employed to solve 3D problems without additional obtained by employing 25  25  25 global grids and 21  21  6
difficulty which shows a very powerful algorithm in the numerical boundary grids. It is observed from the above figures that our
modeling of multi-dimensional flows. The flow behavior inside a meshless numerical predictions with very coarse grids still
lid-driven cubic cavity can be well compared by plotting the show good agreement with the results of the least-square FEM
various velocity profiles on the central planes of the principal with 50  52  50 trilinear elements based on the velocity–
axes. Figs. 10(a)–(c) show the u–z and x–w plots at y ¼ 0.5 planes pressure–vorticity formulation [26].
for Re ¼ 100, 400 and 1000 obtained by the present scheme and
comparisons with the results of Jiang et al. [26] using the least-
square FEM. In the cases for Re ¼ 100 and 400, the global grids 4.4. 3D lid-driven cavity flows with a fixed sphere
21  21  21 and boundary grids 21  21  6 are uniformly
distributed in the computational domain. Because of the high To demonstrate the capacity of the proposed method, a more
Reynolds number, the inertial effects are significant compared to complicated geometry is added to simulate the flow pattern.
the viscous damping forces. In order to accurately capture the flow Fig. 11 sketches the 3D lid-driven cavity with a fixed sphere inside,
 ARTICLE IN PRESS
D.L. Young et al. / Engineering Analysis with Boundary Elements 33 (2009) 1031–1044 1043

u compoment u compoment
-1.00 -0.50 0.00 0.50 1.00 -1.00 -0.50 0.00 0.50 1.00
1.00 1.00 1.00 1.00

Present Present
IBFEM, 81x81x81 IBFEM, 81x81x81
0.80 0.80
0.50 0.50

w component

w component
0.60 0.60
z axis

z axis
0.00 0.00

0.40 0.40

-0.50 -0.50
0.20 0.20

0.00 -1.00 0.00 -1.00

0.00 0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00
x axis x axis

u compoment
-1.00 -0.50 0.00 0.50 1.00
1.00 1.00

Present
IBFEM, 81x81x81
0.80
0.50

w component
0.60
z axis

0.00

0.40

-0.50
0.20

0.00 -1.00
0.00 0.20 0.40 0.60 0.80 1.00
x axis

Fig. 14. Velocity profiles at horizontal and vertical line of 3D lid-driven cavity with a fixed sphere for (a) Re ¼ 1, (b) Re ¼ 100, and (c) Re ¼ 400.

including the geometric parameter and boundary conditions. velocity contours will exhibit the asymmetric characteristics
The top wall of the cube is assumed to move parallel to the x-axis and the velocity gradient is sharper near the top wall, as shown
with a constant velocity of unity (dimensionless) and the sphere is in Figs. 12(b) and (c). Figs. 13(a)–(c) show the x–z plane at y ¼ 0.5.
located at the center of the cube with radius rs ¼ 0.125. Non-slip The streamline topology is symmetric and the vortex center is
stationary boundary conditions are specified on the surface located at the top of the sphere at Re ¼ 1. In Figs. 13(b) and (c),
of the sphere. The Reynolds numbers of 1, 100 and 400 are we can clearly notice the vortex center moves along the sphere
chosen for simulation and solutions are validated by results in the clockwise direction. Finally, we give comparisons in
obtained from the immersed boundary FEM (IBFEM) [28,29]. Figs. 14(a)–(c) of two different numerical schemes. The
The computational nodes are uniformly distributed in the domain, steady solutions of IBFEM [28,29] are obtained from grids
which use 21  21  21 global grids and 21  21  6 boundary 81 81 81, and the velocities along the centerlines of the
grids for these numerical experiments. The flow field is set as cube are chosen to compare with present solutions. Once
zero initial conditions in the interior domain. Fig. 12(a) represents again, we observe from Fig. 14(a)–(c) that the present results
the u velocity contour at the y–z plane, the result is very similar are in close agreement with results solved by the IBFEM
to a lid-driven cavity flow, except a fixed sphere is engulfed though very small amount of grids are used in this meshless
at the center. As the Reynolds number further increases, the u algorithm.
 ARTICLE IN PRESS
1044 D.L. Young et al. / Engineering Analysis with Boundary Elements 33 (2009) 1031–1044

5. Conclusion [4] Mathon R, Johnston RL. The approximate solution of elliptic boundary-
value problems by fundamental solutions. SIAM J Numer Anal 1977;
14:638–50.
Although the meshless MFS has become very popular and [5] Bogomolny A. Fundamental solutions method for elliptic boundary value
received much attention for decades, their applications to the problems. SIAM J Numer Anal 1985;22:644–69.
Navier–Stokes equations still remain challenging and difficult, and [6] Karageorghis A. The method of fundamental solutions for the calcula-
tion of the eigenvalues of the Helmholtz equation. Appl Math Let 2001;14:
become the focus of research subject in the scientific computing 837–42.
community. This paper presents for the first time a novel meshless [7] Karageorghis A, Fairweather G. The method of fundamental solutions for the
MFS procedure based on the operator-splitting method to solve the numerical solution of the biharmonic equation. J Comput Phys 1987;69:
434–59.
primitive variables form of the Navier–Stokes equations with the [8] Young DL, Jane SJ, Fan CM, Murugesan K, Tsai CC. The method of fundamental
moderate Reynolds numbers. The operator-splitting or projection solutions for 2D and 3D Stokes problems. J Comput Phys 2006;211:1–8.
method splits the solutions into two distinct steps, and omits the [9] Golberg MA. The method of fundamental solutions for Poisson’s equations.
Eng Anal Bound Elem 1995;16:205–13.
pressure gradient terms from the momentum equations. In the first [10] Golberg MA, Muleshkov AS, Chen CS, Cheng AH-D. Polynomial particular
step, the ELMFS is employed to obtain the provisional velocity. In the solutions for certain partial differential operators. Numer Methods Partial
second step, the MFS–MPS is used to handle the Poisson equation for Differential Equation 2002;19(1):112–33.
[11] Young DL, Tsai CC, Murugesan K, Fan CM, Chen CW. Time-dependent
pressure field. The present scheme is a meshless method, as it is free
fundamental solutions for homogeneous diffusion problems. Eng Anal
from mesh generation and numerical integration. Boundary Elem 2004;29:1463–73.
To validate the proposed meshless numerical scheme, four [12] Young DL, Fan CM, Tsai CC, Chen CW, Murugesan K. Solution of advection–
diffusion equation using the Eulerian–Lagrangian method of fundamental
numerical experiments of 2D and 3D lid-driven cavity flows with
solutions. Int Math Forum 2006;1(14):687–706.
moderate Reynolds numbers, a 2D circular cavity with an [13] Young DL, Fan CM, Hu SP, Atluri SN. The Eulerian–Lagrangian method of
eccentric rotating cylinder and a 3D cavity flow with a fixed fundamental solutions for two-dimensional unsteady Burgers’ equations. Eng
sphere are carried out. Those solutions are all well compared with Anal Boundary Elem 2008;32:395–412.
[14] Tsai CC, Young DL, Fan CM, Chen CW. MFS with time-dependent fundamental
the results from the domain discretization numerical methods, solutions for unsteady Stokes equations. Eng Anal Boundary Elem 2006;30:
such as the multigrid FDM [25], unstructured FEM, least-square 897–908.
FEM [26] and IBFEM [28,29] algorithms. The excellent capability [15] Young DL, Chen CW, Fan CM. The method of fundamental solutions for low
Reynolds number flows with moving rigid body. In: Chen CS, Karageorghis A,
of the present scheme is also observed, since we can adopt very Smyrlis YS, editors. The Method of Fundamental Solutions-A Meshless
coarse grids to get reasonable solutions and therefore alleviate the Method. Dynamic Publishers; 2008. p. 181–206.
excessive computations by using this meshless MFS algorithm. In [16] Chorin AJ. Numerical solution of the Navier–Stokes equations. Math Comput
1968;22:745–62.
addition, the multi-dimensional and irregular geometry flow [17] Temam R. Une méthode d’approximation de la solution des équations de
problems can also be solved without additional difficulty by Navier–Stokes. Bull Soc Math France 1968;98:115–52.
using this mesh-free MFS formulation. [18] Patankar SV, Spalding DB. A calculation procedure for heat, mass and
momentum transfer in three-dimensional parabolic flows. Int J Heat Mass
The computational efficiency and capacity are important issues
Transfer 1972;15:1787–806.
in the numerical simulations. In the present meshless method, we [19] Goldberg D, Ruas V. A numerical study of projection algorithms in the finite
can observe that the resultant dense matrix is the main reason to element simulation of three-dimensional viscous incompressible flow. Int J
Numer Methods Fluids 1999;30:233–56.
reduce the efficiency and capacity of the present method. The
[20] Brown D, Cortez R, Minion M. Accurate projection method for incompressible
research for releasing this constraint deserves further investiga- Navier–Stokes equations. J Comput Phys 2001;168:464–99.
tion in the future. [21] Chang W, Giraldo F, Perot B. Analysis of an exact fractional step method.
J Comput Phys 2002;180:183–99.
[22] Duchon J. Splines minimizing rotation invariant seminorms in Sobolev
spaces. In: Constructive theory of functions of several variables. Berlin:
Acknowledgements Springer; 1977. p. 85–110.
[23] Golberg MA, Chen CS. Discrete projection methods for integral equations.
Southampton: Computational Mechanics Publication; 1997.
The National Science Council of Taiwan is greatly appreciated
[24] Burggraf OR. Analytical and numerical studies of the structure of steady
for providing financial support of this research under the Grant separated flow. J Fluid Mech 1966;24:113–51.
no. NSC 94-2611-E-002-007. [25] Ghia U, Ghia KN, Shin CT. High-Re solutions for incompressible flow using the
Navier–Stokes equations and a multigrid method. J Comput Phys
1982;48:387–411.
References [26] Jiang BN, Lin TL, Provinelli LA. Large-scale computation of incompressible
viscous flow by least-square finite element method. Compu Methods Appl
Mech Eng 1994;114:213–31.
[1] Fairweather G, Karageorghis A. The method of fundamental solutions for [27] Tsai CC, Lin YC, Young DL, Atluri SN. Investigations on the accuracy and
elliptic value problems. Adv Comput Math 1998;9:69–95. condition number for the method of fundamental solutions. CMES
[2] Golberg MA, Chen CS. The method of fundamental solutions for potential, 2006;16(2):103–14.
Helmholtz and diffusion problems. In: Golberg MA, editor. Boundary integral [28] Young DL, Chiu CL, Fan CM. A hybrid Cartesian/immersed-boundary finite-
methods: numerical and mathematical aspects. Boston: Computational element method for simulating heat and flow patterns in a two-roll mill.
Mechanics Publications; 1998. p. 103–76. Numer Heat Transfer, B 2007;51:251–74.
[3] Kupradze VD, Aleksidze MA. The method of fundamental equations for the [29] Chiu CL, Fan CM, Young DL. 3D hybrid Cartesian/immersed-boundary finite-
approximate solution of certain boundary value problem. Zh Vychisl Mat element analysis of heat and flow patterns in a two-roll mill. Int J Heat Mass
1964;4(4):82–126. Transfer 2009;52(7–8):1677–89.

View publication stats