Printed in India
1. Introduction
The finite element method (FEM) has been established as a very powerful numerical technique
for the analysis of space domain problems having arbitrary shapes. It has been observed that in
the finite element method, mesh generation is a far more time-consuming and expensive task
than the assembly and solution of the finite element equations. Moreover, there are certain
classes of problems for which FEM is difficult, or even impossible to apply. For such problems,
it has become necessary to find the methods, which may be somewhat more expensive from the
viewpoint of computer time but require less time in the preparation of data. Recently, a class
of new methods, known as meshless methods, has been developed. Various meshless methods belonging to this family are: smooth particle hydrodynamics (Monaghan 1988, 1992),
diffuse element method (Nayroles et al 1992), reproducing kernel particle method (Liu et al
1995), the method of finite spheres (De & Bathe 2000), free mesh method (Yagawa & Yamada
1996), local boundary integral equation (LBIE) method (Zhu et al 1998), the partition of unity
A list of symbols is given at the end of the paper
285
286
I V Singh
method (Babuska & Melenk 1997), element-free Galerkin (EFG) method (Belytschko et al
1994a, 1996; Lu et al 1994; Dolbow & Belytschko 1998; Liu & Gu 2000), meshless local
PetrovGalerkin (MLPG) method (Atluri & Zhu 1998), natural element method (Sukumar
et al 1998) and natural neighbour Galerkin method (Sukumar et al 2001).
Among all these meshless methods, the element-free Galerkin method (EFG) method has
been successfully used to solve the problems of plates and shells (Krysl & Belytschko 1996a,
1996b), fracture mechanics and wave propagation (Belytschko et al 1994b; Belytschko &
Tabbara 1996, Lu et al 1995; Sukumar et al 1997), NDT (Xuan et al 2001), electromagnetic
field (Cingoski et al 1998) and heat transfer (Singh et al 2002, 2003a, 2003b).
In this paper, we have used the meshless element-free Galerkin (EFG) method for solving
two-dimensional fluid flow problems and have proposed a new exponential weight function.
In the EFG method, function over the solution domain requires only a set of nodes, and
does not require element connectivity. Integration over the solution domain requires only a
simple integration cell to obtain the solution. Variational technique has been used for the
discretization of the governing equations for two-dimensional fluid flow problems. Lagrange
multiplier method is used to enforce the essential boundary conditions. FORTRAN programs
have been developed for two-dimensional transient and steady state analysis of fluid flow
problems. The results are obtained for a model problem and compared with the results of
exact (Appel et al 1959) and finite element (Reddy 1993) methods.
2. The element-free Galerkin method
The discretization of the governing equations by element-free Galerkin method requires moving least square interpolation functions which are made up of three components: a weight
function associated with each node, a monomial basis function and a set of non-constant coefficients. The weight function is non-zero over a small neighbourhood at a particular node,
called support of the node.
2.1 Moving least square approximation
Using moving least square approximation, the unknown function u(y, z) is approximated by
uh (y, z) over the two-dimensional domain (Belytschko et al 1994a)
uh (y, z) =
m
X
(1)
j =1
where,
m = number of terms in the basis,
pj (y, z) = monomial basis function,
aj (y, z) = non-constant coefficients,
xT = [y
z], pT (x) = [1
z].
The coefficients aj (x) are found by minimizing the quadratic functional J (x) given by
(
)2
n
m
X
X
w(x xI )
pj (xI )aj (x) u1 ,
J (x) =
I =1
j =1
(2)
287
where w(x xI ) is a weight function which is non-zero over a small domain, called domain
of influence. n is the number of nodes in the domain of influence.
The minimization of J (x) w.r.t. a(x) leads to the following set of equations
a(x) = A1 (x)B(x)T,
(3)
n
X
I =1
1 y1 z1
1 yn
zn
= w(x x1 ) y1 y12 y1 z1 + . . . w(x xn ) yn yn2 yn zn ,
zn yn zn zn2
z1 y1 z1 z12
B(x) = [w(x x1 )p(x1 ), . . . w(x xn )p(xn )]
1
1
= w(x x1 ) y1 , . . . w(x xn ) yn ,
z1
zn
uT = [u1 , u2 , . . . un ].
(4)
(5)
(6)
n
X
81 (x)uI = 8(x)u,
(7)
I =1
m
X
pj (x)(A1 (x)B(x))j I = pT A1 BI .
(8)
j =0
(9)
288
I V Singh
1
,
100r
(10b)
0r1
,
r>1
(10c)
where
ky yI k
kz zI k
and (rz )I =
,
dmzI
dmyI
= dmax cyI and dmzI = dmax czI , dmax = scaling parameter,
(ry )I =
dmyI
cyI and czI at node I , are the distances to the nearest neighbours. dmyI and dmzI are chosen
such that the matrix is non-singular everywhere in the domain.
The weight function at any given point can be calculated as
w(x xI ) = w(rY )w(rz ) = wy wz ,
(11a)
dwy
dwz
wz , w,z =
wy .
dy
dz
(11b)
= .
(12a)
+ 2
2
y
z
dx
t
The Initial conditions are
at t = 0,
u = uini ,
in .
(12b)
u = uS1
u = u S2
u = uS3
u = u S4
= 0,
= 0,
= 0,
= 0.
(12c)
= 0.
(13)
289
Mud.
5(u) =
2 ,y
(14)
Enforcing essential boundary conditions using Lagrange multipliers, the functional 5 (u) is
obtained as
Z
Z
Z
Z
2
u,y + u2,z d + uud
03
04
Mud
5 (u) = u,y u,y + u,z u,z d + uud
Z
(u uS1 )d0 +
+
01
ud0 +
01
(u uS3 )d0 +
03
(u uS2 )d0 +
02
ud0 +
03
ud0
02
(u uS4 )d0 +
04
ud0.
04
(16)
Since u and are arbitrary in preceding equation, the following relations are obtained using
(7),
+ [G]{} = {f},
[K]{u} + [C]{u}
(17a)
= {q},
[G ]{u}
where
Z
KI J =
CI J =
8I,y
8I,z
8TI
T
0
0
(18a)
0
8J d,
0
(18b)
Z
8I NK d01 +
01
Z
8I NK d02 +
02
fI =
8J,y
d,
8J,z
GI K =
(17b)
Z
8I NK d03 +
03
8I NK d04 ,
04
M8I d,
qK =
(18d)
Z
uS1 NK d01 +
01
(18c)
Z
uS2 NK d02 +
02
Z
uS3 NK d03 +
03
uS4 NK d04 .
04
(18e)
290
I V Singh
=
,
GT
0 q
(19)
where
Rn0 = ([C] (1 )1t[K]){u}n0 1 + 1t{f}n0 + (1 )1t{f}n0 1 ,
K = 1t[K].
(20a)
(20b)
Value of parameters
010 m
010 m
40000 (N/m2 /m)
1000 kg/m3
25 Ns/m2
00 m/s
00 m/s
001 s
291
Quartic spline
Gaussian
Exponential
FEM
000
010
020
030
040
050
060
070
080
090
100
00000
03256
04895
05859
06442
06795
07009
07139
07217
07265
07294
00000
03235
04855
05805
06375
06719
06925
07050
07125
07170
07197
00000
03257
04896
05861
06444
06797
07012
07141
07220
07268
07297
00000
02994
04702
05711
06332
06718
06958
07107
07200
07258
07293
velocity has also been shown in figure 2 at the centre (y = 0, z = 0) of the duct cross-section
for the same number of nodes.
The steady state analysis has been carried out for a model problem by using different EFG weight functions. Table 3 shows a comparison of the results obtained using EFG
1.2
Velocity (m /s )
0.8
0.6
0.4
FEM
Quartic spline
Gaussian
E x ponential
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Time (t), s
Figure 2. Comparison of transient EFG results with FEM at the centre (y = 0, z = 0) of the duct
cross-section.
292
I V Singh
Table 3. Comparison of steady state EFG results for different weight
functions with FEM and exact methods at the centre (y = 0, z = 0) of
the duct cross-section.
Velocity (m/s)
EFG
No. of
nodes
Quartic spline
Gaussian
Exponential
FEM
Exact
25
81
121
441
12338
11916
11869
11808
11753
11776
11780
11786
12358
11920
11872
11808
09839
11619
11696
11764
11802
11802
11802
11802
method with FEM and exact methods. The L2 -error has been calculated for different EFG
weight functions and FEM with exact solution and is given in table 4. The L2 -error has
also been plotted in figure 3 for the calculation of rate of convergence. The rate of convergence for quarticspline, Gaussian, exponential and FEM are obtained as 10521, 09305,
10630 and 13709 respectively. From table 4 and figure 3, it is observed that the convergence of exponential weight function is better than those obtained using quartic spline
and Gaussian weight functions. It can be also be noted that the accuracy in EFG results is
more as compared with FEM results but the convergence of FEM results is better than EFG
results.
Table 5 shows the effect of scaling parameter (dmax ) on EFG results at the location (y =
D/4, z = H /4) of the duct cross-section. A similar comparison of results has also been shown
in figure 4 at the centre (y = 0, z = 0) of the duct cross-section. From these results, it is clear
that the variation in EFG results with scaling parameter for exponential weight function is
less as compared quartic spline and Gaussian weight functions. The scaling parameter (dmax )
gives acceptable results for the range 1 < dmax 15.
It can be noted from the tables and figures that the EFG results are well converged and in
good agreement with the results of finite element and exact methods. From the analysis, it
is also clear that the proposed exponential weight function is more promising as compared
to quatric spline and Gaussian weight functions. For higher accuracy, large number of nodes
Quartic spline
(x104 )
Gaussian
(x104 )
Exponential
(x104 )
0000625
0000156
0000100
0000025
3077780
656262
382998
106004
3220820
221061
207301
176356
3195243
681253
397280
106089
FEM
(x104 )
21493360
3913477
680464
288473
293
1.5
L2-error
2.5
3.5
4.5
5
FEM
Quartic spline
Gaussian
Exponential
4.8
4.6
4.4
4.2
3.8
3.6
3.4
3.2
Log (a)
Figure 3. A plot of L2 -error for different weight functions of EFG and FEM.
1.28
Exact
FEM
Quartic spline
Gaussian
Exponential
1.26
Velocity (m/s)
1.24
1.22
1.2
1.18
1.16
1.14
1
1.5
2
Scaling parameter (dmax)
2.5
Figure 4. Effect of scaling parameter (dmax ) on EFG results at the centre (y = 0, z = 0) of the duct
cross-section.
294
I V Singh
Table 5. Effect of scaling parameter (dmax ) on EFG results for different weight functions at the location (y = D/4, z = H /4) of the duct
cross-section.
Velocity (m/s)
EFG
Scaling
parameter
Quartic spline
Gaussian
Exponential
FEM
Exact
101
11
12
13
14
15
16
17
18
19
20
07338
07333
07369
07428
07461
07499
07517
07516
07520
07517
07498
07299
07334
07368
07395
07432
07457
07480
07502
07512
07508
07485
07341
07412
07417
07420
07336
07389
07402
07456
07473
07489
07549
07139
07139
07139
07139
07139
07139
07139
07139
07139
07139
07139
07260
07260
07260
07260
07260
07260
07260
07260
07260
07260
07260
can be easily generated with least effort. This work can be extended for the three-dimensional
analysis of complex fluid flow problems using EFG method.
List of symbols
a
aj (y, z)
dmyI &dmzI
dmax
M
n
n0
pj (y, z)
S
t
uh (y, z)
u,y , u,z
u,yy , u,zz
w(x xI )
NK
8(x)
295
References
Appel D W, Hubbard P G, Landweber L, Laursen E M, McNown J S, Rouse H, Siao T T, Touch A,
Yih C S 1959 Advanced mechanics of fluids (Iowa: Institute of Hydraulic Research)
Atluri S N, Zhu T 1998 A new meshless local PetrovGalerkin (MLPG) approach in computational
mechanics. Comput. Mech. 22: 117127
Babuska I, Melenk J M 1997 The partition of unity method. Int. J. Numer. Methods Eng. 40:
727758
Belytschko T, Tabbara M 1996 Dynamic fracture using element-free Galerkin methods. Int. J. Numer.
Methods Eng. 39: 923938
Belytschko T, Lu Y Y, Gu L 1994a Element-free Galerkin methods. Int. J. Numer. Methods Eng. 37:
229256
Belytschko T, Gu L, Lu Y Y 1994b Fracture and crack growth by element-free Galerkin methods.
Model. Simul. Mater. Sci. Eng. 2: 519534
Belytschko T, Organ D, Krongauz Y 1995 A coupled finite element-element-free Galerkin method.
Comput. Mech. 17: 186195
Belytschko T, Krongauz Y, Organ D, Fleming M, Krysl P 1996 Meshless methods: an overview and
recent developments. Comput. Methods Appl. Mech. Eng. 139: 347
Cingoski V, Miyamoto N, Yamashita H 1998 Element-free Galerkin method for electromagnetic field
computations. IEEE Trans. Magn. 34: 32363239
De S, Bathe K J 2000 The method of finite spheres. Comput. Mech. 25: 329-345
Dolbow J, Belytschko T 1998 An introduction to programming the meshless element-free Galerkin
method. Arch. Comput. Mech. 5: 207241
Krysl P, Belytschko T 1996a Analysis of thin plates by element-free Galerkin method. Comput. Mech.
17: 2635
Krysl P, Belytschko T 1996b Analysis of thin shells by element-free Galerkin method. Int. J. Solids
Struct. 33: 30573080
Liu W K, Jun S, Zhang Y F 1995 Reproducing kernel particle methods. Int. J. Numer. Methods Eng.
20: 10811106
Liu G R, Gu Y T 2000 Coupling of element-free Galerkin method and hybrid boundary element
methods using modified variational formulation. Comput. Mech. 26: 166173
Lu Y Y, Belytschko T, Gu L 1994 A new implementation of the element-free Galerkin method.
Comput. Methods Appl. Mech. Eng. 113: 397414
Lu Y Y, Belytschko T, Tabbara M 1995 Element-free Galerkin methods for wave propagation and
dynamic fracture. Comput. Methods Appl. Mech. Eng. 126: 131153
Monaghan J J 1988 An introduction to SPH. Comput. Phys. Commun. 48: 8996
Monaghan J J 1992 Smoothed particle hydrodynamics. Annu. Rev. Astron. Astrophys. 30: 543574
Nayroles B, Touzot G, Villon P 1992 Generalizing the finite element method: diffuse approximation
and diffuse elements. Comput. Mech. 10: 307318
Reddy J N 1993 An introduction to the finite element method (McGraw-Hill: Singapore)
Singh I V, Sandeep K, Prakash R 2002 The element-free Galerkin method in three-dimensional steady
state heat conduction. Int. J. Comput. Eng. Sci. 3: 291303
Singh I V, Sandeep K, Prakash R 2003a Heat transfer analysis of two-dimensional fins using meshless
element-free Galerkin method. Numer. Heat Transfer A44: 7384
Singh I V, Sandeep K, Prakash R 2003b Meshless EFG method in transient heat conduction problems.
Int. J. Heat Technol. 21: 99105
Sukumar N, Moran B, Black T, Belytschko T 1997 An element-free Galerkin method for threedimensional fracture mechanics. Comput. Mech. 20: 170175
Sukumar N, Moran B, Belytschko T 1998 The natural element method in solid mechanics. Int. J.
Numer. Methods Eng. 43: 839887
Sukumar N, Moran B, Yu Semenov A, Belikov V V 2001 Natural neighbour Galerkin methods. Int.
J. Numer. Methods Eng. 50: 127
296
I V Singh
Xuan L, Zeng Z, Shanker B, Udpa L 2001 Meshless element-free Galerkin method in NDT applications. 28th Annu. Rev. of Progress in Quantitative and Nondestructive Evaluation, Maine, USA
(AIP Conf. Proc.)
Yagawa G, Yamada T 1996 Free mesh method, a new meshless finite element method. Comput. Mech.
18: 383386
Zhu T, Zhang J D, Atluri S N 1998 A meshless local boundary integral equation (LBIE) method for
solving nonlinear problems. Comput. Mech. 22: 174186