INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING

Int. J. Numer. Meth. Engng. 47, 537–556 (2000)

New concepts in meshless methods

S. N. Atluri ∗ and Tulong Zhu
Center for Aerospace Research and Education; 7704 Boelter Hall; University of California at Los Angeles;
Los Angeles; CA 90095-1600; USA

SUMMARY

Meshless methods have been extensively popularized in literature in recent years, due to their exibility in
solving boundary value problems. Two kinds of truly meshless methods, the meshless local boundary integral
equation (MLBIE) method and the meshless local Petrov–Galerkin (MLPG) approach, are presented and
discussed. Both methods use the moving least-squares approximation to interpolate the solution variables,
while the MLBIE method uses a local boundary integral equation formulation, and the MLPG employs a
local symmetric weak form. The two methods are truly meshless ones as both of them do not need a ‘nite
element or boundary element mesh’, either for purposes of interpolation of the solution variables, or for the
integration of the ‘energy’. All integrals can be easily evaluated over regularly shaped domains (in general,
spheres in three-dimensional problems) and their boundaries. Numerical examples presented in the paper
show that high rates of convergence with mesh renement are achievable. In essence, the present meshless
method based on the LSWF is found to be a simple, ecient and attractive method with a great potential in
engineering applications. Copyright ? 2000 John Wiley & Sons, Ltd.

KEY WORDS: meshless methods; local boundary integral equation; local symmetric weak form; local Petrov–Galerkin
formulation; moving least-squares approximation

1. INTRODUCTION

Meshless methods in computational mechanics have attracted much attention in recent decades,
due to their exibility in solving practical engineering problems, especially those problems with
discontinuities or moving boundaries. The main objective of meshless methods is to get rid of or
at least alleviate the diculty of meshing and remeshing the entire structure, by only adding or
deleting nodes in the entire structure.
The initial idea of meshless methods dates back to the smooth particle hydrodynamics (SPH)
method for modelling astrophysical phenomena [1] while the research into meshless methods has
become very active only after the publication of the Diuse Element Method by Nayroles et al. [2].
Several meshless methods have also been reported in literature since then, such as the element-free
Galerkin method [3–8], the reproducing kernel particle method [9].

∗ Correspondence to: S. N. Atluri, Center for Aerospace Research and Education, 7704 Boelter Hall, University of California
at Los Angeles, Los Angeles, CA 90095-1600, U.S.A. E-mail: [email protected]

Contract=grant sponsor: Oce of Naval Research

Contract=grant sponsor: Federal Aviation Administration

CCC 0029-5981/2000/030537–20$17.50 Received 11 November 1998

Copyright ? 2000 John Wiley & Sons, Ltd.
538 S. N. ATLURI AND T. ZHU

Even though no mesh is required in these methods for the interpolation of the solution variables,
shadow elements are inevitable in these methods, for the integration of ‘energy’. Therefore, even
though called element-free methods, these methods are not truly meshless ones.
Recently, two truly meshless methods, the meshless local boundary equation (MLBIE) method,
and the meshless local Petrov–Galerkin approach have been successfully developed in [10 –15]
for solving linear and non-linear boundary problems. Both methods use local weak forms over
a local sub-domain, and shape functions from the moving least-squares (MLS) approximation.
The MLBIE method employs a local unsymmetric weak form, i.e. the local boundary integral
equation (LBIE), while the MLPG approach uses a local symmetric weak form. Both methods
are truly meshless ones, as no ‘nite element or boundary element mesh’ is required in these two
approaches, either for purposes of interpolation of the solution variables, or for the integration
of the ‘energy’. All integrals can be easily evaluated over regularly shaped domains (in general,
spheres in three-dimensional problems) and their boundaries.
These two methods are also more exible and easier in dealing with non-linear problems than
the conventional FEM, EFG and BEM, as domain integrals will not cause any diculty in imple-
menting these two methods.
In the present paper, by ‘the support of node xi ’ we mean a sub-domain (usually taken as
a circle of radius ri ) in which the weight function wi in the MLS approximation, associated
with node xi , is non-zero; by ‘the domain of denition’ of an MLS approximation for the trial
function at any point x (hereinafter simply called as the ‘domain of denition of point x’) we
mean a sub-domain which covers all the nodes whose weight functions do not vanish at x; by
‘the domain of in uence of node xi ’ we denote a sub-domain in which all the nodes have non-
zero couplings with the nodal values at xi , in the system stiness matrix; and by the ‘symmetric
weak form’ we mean the weighted residual form for the dierential equation, which is integrated
by parts enough times such that the dierentiability requirements for the trial and test functions
are the same. In our implementation, the domain of in uence of a node is the union of the
domains of denition of all points (in general, but the quadrature points in specic) on the local
boundary (for the MLBIE method) or in the local domain (for the MLPG approach) of the
node. We do not intend to mean these to be versatile denitions, but rather, explanations of our
terminology.
The following discussion begins with the brief description of the moving least-squares (MLS)
approximation in Section 2. The MLBIE method and its numerical implementation are presented
in Section 3. The MLPG approach and its numerical implementation are presented in Section 4.
Numerical examples for 2-D problems are given in Section 5. The paper ends with conclusions
and discussions in Section 6.

2. THE MLS APPROXIMATION SCHEME

This section gives a brief summary of the MLS approximation. For details of the MLS approxi-
mation, see [3; 14].
Consider a sub-domain
x , the neighbourhood of a point x and denoted as the domain of
denition of the MLS approximation for the trial function at x, which is located in the problem
domain
. To approximate the distribution of function u in
x , over a number of randomly
located nodes {xi }; i = 1; 2; : : : ; n; the moving least squares approximant u h (x) of u; ∀x ∈
x ; can be

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
 NEW CONCEPTS IN MESHLESS METHODS 539

Figure 1. The distinction betweeen ui and ûi

dened by
u h (x) = pT (x)a(x) ∀x ∈
x (1)
where p (x) = [p1 (x); p2 (x); : : : ; pm (x)] is a complete monomial basis of order m; and a(x) is a
T

vector containing coecients aj (x); j = 1; 2; : : : ; m; which are functions of the space co-ordinates
x = [x1 ; x2 ; x3 ]T .
The coecient vector a(x) is determined by minimizing a weighted discrete L2 norm, which
can be dened as
Pn
J (x) = wi (x)[ pT (xi )a(x) − ûi ]2 (2)
i=1

where wi (x) is the weight function associated with the node i; with wi (x)¿0 for all x in the
support of wi (x); xi denotes the value of x at node i; n is the number of nodes in
x for which
the weight functions wi (x)¿0.
Here it should be noted that ûi ; i = 1; 2; : : : ; n; in equation (2) are the ctitious nodal values,
and not the nodal values of the unknown trial function u h (x) in general (see Figure 1 for a simple
one-dimensional case for the distinction between ui and ûi ).
Solving for a(x) by minimizing J in equation (2) and substituting it into equation (1) give a
relation which may be written as the form of an interpolation function similar to that used in the
FEM, as
P
n
u h (x) = i (x)ûi ; u h (xi ) ≡ ui 6= ûi ; x ∈
x (3)
i=1

where
P
m
i (x) = pj (x)[A−1 (x)B(x)]ji (4)
j=1

with the matrices A(x) and B(x) being dened by

Pn
A(x) = wi (x)p(xi )pT (xi ) (5)
i=1
B(x) = [w1 (x)p(x1 ); w2 (x)p(x2 ); : : : ; wn (x)p(xn )] (6)
The MLS approximation is well dened only when the matrix A in equation (5) is non-singular.

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
540 S. N. ATLURI AND T. ZHU

i (x) is usually called the shape function of the MLS approximation corresponding to nodal
point xi . From equations (4) and (6), it may be seen that i (x) = 0 when wi (x) = 0. The fact that
i (x) vanishes, for x not in the support of nodal point xi preserves the local character of the
moving least squares approximation.
In implementing the MLS approximation for the present local symmetric weak form, the basis
functions and weight functions should be chosen at rst. Both Gaussian and spline weight functions
with compact supports can be considered in the present work. The Gaussian weight function
corresponding to node i may be written as

 exp[−(di =ci )2k ] − exp[−(ri =ci )2k ]
; 06di 6ri
wi (x) = 1 − exp[−(ri =ci )2k ] (7)

0; di ¿ri
where di = |x − xi | is the distance from node xi to point x; ci is a constant controlling the shape
of the weight function wi and therefore the relative weights and ri is the size of the support for
the weight function wi and determines the support of node xi . In the present computation, k = 1
was chosen.
A spline weight function is dened as
  2  3  4
 di di di
wi (x) = 1 − 6 ri +8
r
−3
ri
; 06di 6ri (8)
 i
0; di ¿ri
The size of support, ri ; of the weight function wi associated with node i should be chosen
such that ri should be large enough to have sucient number of nodes covered in the domain of
denition of every sample point (n¿m) to ensure the regularity of A. A very small ri may result
in a relatively large numerical error in using Gauss numerical quadrature to calculate the entries
in the system matrix. On the other hand, ri should also be small enough to maintain the local
character of the MLS approximation.
A generalization of the MLS interpolation scheme, suitable for 4th order problems of beam and
plate bending is given in Reference [14].

3. THE MESHLESS LOCAL BOUNDARY INTEGRAL EQUATION (MLBIE) METHOD

In this section, we use the following potential problem to illustrate the development of the MLBIE
approach:
∇2 u(x) = p(x); x∈
(9)
where p is a given source function and the domain
is enclosed by = u ∪ q; with boundary
conditions:
u = u on u (10a)
@u
≡ q = q on q (10b)
@n
where u and q are the prescribed potential and normal ux, respectively, on the essential boundary
u and on the ux boundary q , and n is the outward normal direction to the boundary .

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
 NEW CONCEPTS IN MESHLESS METHODS 541

In the present local boundary integral equation, a local weak form over a local sub-domain,

s (∈
), is employed. The local sub-domain
s is conveniently taken to be a sphere (in 3-D, or
a circle in 2-D) centred at a point x in question. The local weak form of the dierential equation
(9) and the boundary conditions (10), over a local sub-domain
s , can be written as
Z
u∗ (x; y)[ ∇2 u(x) − p(x)] d
= 0 (11)

s

where u is the trial function and u∗ is the test function which is chosen to be the solution, in
innite space, of the following equation

∇2 u∗ (x; y) + (x; y) = 0 (12)

with (x; y) being the Dirac delta function.

Using (∇2 u)u∗ = (u; i u∗ ); i − (uu∗ ); i + uu;∗ii , and the divergence theorem twice, in equation (11)
yields
Z Z
@u∗ (x; y)
u∗ (x; y)
@u(x)
u(y) = d − u(x) d
@
s @n @
s @n
Z
− u∗ (x; y)p(x) d
(13)

s

where @
s is the boundary of the local domain
s .
It should be noted that equation (13) holds irrespective of the size and shape of @
s . We can
deliberately choose a simple regular shape for
s and thus for @
s . The most regular shape of
a sub-domain should be an n-dimensional sphere for a boundary-value problem dened on an
n-dimensional space. Thus, an n-dimensional sphere (or a part of an n-dimensional sphere for a
boundary node), is chosen in our development (see Figure 2).
The above equation can be further simplied by using a modied test function ũ∗ (x; y), such
that it vanishes over Ls , where Ls is a part of @
s , on which no boundary conditions are specied
(see Figure 2). This can be easily accomplished by introducing a ‘companion solution’, ũ, which
satises
( 2
∇ ũ = 0 in
0s
(14)
ũ = u∗ (x; y) on @
0 s

where
0s ⊇
s such that
0s =
s for an interior source point y; and
0s is the extended whole
sphere which encloses @
s , a part of the sphere, for a boundary source point y (see Figure 2).
Using the modied test function ũ∗ (x; y) = u∗ − ũ in equation (13) and rearranging equation
(13), we obtain the following local boundary integral equation (LBIE):
Z Z
@ũ∗ (x; y) @u(x) ∗
(y)u(y) = − u(x) d + ũ (x; y) d
@
s @n s
@n
Z
− ũ∗ (x; y)p(x) d
(15)

s

where s = @
s ∩ , i.e. s is a part of @
s , over which boundary conditions are specied.

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
542 S. N. ATLURI AND T. ZHU

Figure 2. The local domains, the support of nodes, the domain of denition of a point, and the domain of in uence of a
node: (1) the domain of denition of any point x is the domain which covers all the nodes whose weight functions do not
vanish at x. (2) the domain of in uence for source point y is the union of all
x ; ∀x ∈ @
s (for the MLBIE method) or
∀x ∈
s (for MLPG approach)

Upon solving for the modied test function, we can solve the problem by using a numerical
discretization technique. For the 2-D harmonic operator, the modied test function ũ∗ can be easily
solved as
1 r0
ũ∗ = u∗ − ũ = ln (16)
2 r
since u∗ = − 1=(2)ln r and therefore ũ = − 1=(2) ln r0 , where r = |x − y| denotes the distance
from the source point to the generic point under consideration, and r0 is the radius of the local
sub-domain,
s .
Substituting equation (3) into equation (15), imposing boundary conditions on the right-hand
side for node i, and carrying out the integrals, we obtain the following linear equations:

P
N
i ui = Kij∗ ûj + fi∗ ; i = 1; 2; : : : ; N (17)
j

where N is the total number of nodes in the entire domain

,
Z Z
∗ ∗ @j (x) @ũ∗ (x; yi )
Kij = ũ (x; yi ) d − j (x) d
su
@n sq
@n
Z Z
@ũ∗ (x; yi )
− j (x) d + !2 j (x)ũ∗ (x; yi ) d
(18a)
Ls @n
s

and
Z Z Z
@ũ∗ (x; yi )
fi∗ = ũ∗ (x; yi )q d − u d − ũ∗ (x; yi )p(x) d
(18b)
sq su
@n
s

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
 NEW CONCEPTS IN MESHLESS METHODS 543

in which sq and su are the ux and essential boundary sections, respectively, of s with s = sq ∪
su , and u
 and q are the prescribed values of u and @u=@n on su and sq , respectively. For those
interior nodes located inside the domain
; Ls ≡ @
s , and the boundary integrals involving su
and sq vanish in equations (18a) and (18b).
From equation (18a), it is seen that no derivatives of the shape functions are needed in con-
structing the stiness matrix for the interior nodes and for those boundary nodes with no essential-
boundary-condition-prescribed sections on their local boundaries. This is attractive in engineering
applications as the calculation of derivatives of shape functions from the MLS approximation is
quite costly.
Here, it should be noted that since the unknown variable u at the source point y (or more
precisely, the nodal value of u h (x) itself) appears on the left-hand side of equation (17), it is very
convenient to impose the essential boundary conditions if any, at the global boundary . Upon
imposing the essential boundary condition for ui on the left-hand side of equation (17) for those
nodes where u is specied; or using equation (3) to represent u for those nodes with u unknown,
and rearranging equation (17), we have the following linear system of û:

K · û = f (19)

4. THE MESHLESS LOCAL PETROV–GALERKIN (MLPG) APPROACH

The MLPG approach is rst proposed in [12; 13], for solving linear and non-linear potential
problems. Unlike in the MLBIE method which uses a local unsymmetric weak form, i.e. the
LBIE, the MLPG approach uses a local symmetric weak form over a local sub-domain
s as a
formulation and uses the MLS approximation to develop a truly meshless method. As the MLS
approximation is used to construct shape functions, the essential boundary conditions in the MLPG
approach is enforced, a posteriori, by a penalty formulation.†
In the present paper, we use the following problem in linear elasticity to illustrate the formula-
tion:

ij; j + bj = 0 in
(20)

where
is the global domain bounded by ; ij is the stress tensor, which corresponds to the
displacement eld ui ; bi is the body force, and ( ); i denotes @( )=@xi . The corresponding boundary
conditions are given as follows:

ui = ui on u (21a)
ti ≡ ij nj = ti on t (21b)

where ui and ti are the prescribed displacements and tractions, respectively, on the displace-
ment boundary u and on the traction boundary t , and ni is the unit outward normal to the
boundary .

† A simple and elementary way to satisfy the essential b.c exactly, a priori, in a meshless method, is presented in Reference
[15].

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
544 S. N. ATLURI AND T. ZHU

A generalized local weak form of the dierential equation (20) and the boundary conditions
(21), over a local sub-domain
s ( ∈
), can be written as
Z Z
(ij; j + bi )vi d
−  (ui − ui )vi d = 0 (22)

s su

where ui and vi are the trial and test functions, respectively, su is a part of the boundary @
s of

s , over which the essential boundary conditions are specied, and /1 is a penalty parameter
used to impose the essential boundary conditions.
Using ij; j vi = (ij vi ); j −ij vi; j , and the divergence theorem, in equation (22) yields the following
expression:
Z Z Z
ij nj vi d − (ij vi; j − bi vi ) d
−  (ui − ui )vi d = 0 (23)
@
s
s su

in which @
s is the boundary of the sub-domain
s and ni is the outward unit normal to the
boundary @
s .
Imposing the natural boundary condition, ti ≡ ij nj = ti on st in equation (23) we obtain
Z Z Z Z Z
t i vi d + t i vi d + ti vi d − (ij vi; j − bi vi ) d
−  (ui − ui )vi d = 0 (24)
Ls su st
s su

in which st is a part of @
s , over which the natural boundary condition, ti = ti , is specied.
In the following development, the Petrov–Galerkin method is used. Unlike in the conventional
Galerkin method in which the trial and test functions are chosen from the same space, the Petrov–
Galerkin method uses the trial and the test functions from dierent spaces. In particular, the test
functions need not vanish on the boundary where the essential boundary conditions are specied.
In the present work, the trial functions ui are approximated by the MLS approximation, while the
test functions vi will be chosen from known functions.
As the test functions are chosen from known functions, the above equation can be further
simplied, by deliberately selecting the test functions vi such that they vanish over Ls , the circle
(for an internal node) or the circular arc (for a node on the global boundary ). This can be easily
accomplished by using the weight function in the MLS approximation as also a test function, with
the radius ri of the support of the weight function being replaced by the radius r0 of the local
domain
s , such that the test function vanishes on Ls .
Using such test functions and rearranging equation (24), we obtain the following local symmetric
weak form (LSWF) in linear elasticity, as
Z Z Z Z Z Z
ij vi; j d
+  u i vi d − t i vi d = 
t i vi d +  ui vi d + bi vi d
(25)

s su su st su
s

As the test functions vi is known, the discretization of equation (25) for one local domain will
only yield one linear algebraic equation. In order to obtain two independent linear equations, we
can apply two (for 2-D problems) or three (for 3-D problems) independent sets of test functions
in equation (25), as has been done in the boundary integral equation method, to give the following
local symmetric weak form (LSWF):
Z Z Z Z Z Z
ij vki; j d
+  ui vki d − ti vki d = ti vki d +  ui vki d + bi vki d
(26)

s su su st su
s

where vki denotes the ith component of the test function in the kth set.

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
 NEW CONCEPTS IN MESHLESS METHODS 545

For brevity, equation (26) can also be written in a matrix form as

Z Z Z Z Z Z
Uv b d
+  vu d − vt d = vt d + 
 vu d + vb d
(27)

s su su st su
s

where Uv denotes the strain matrix from the test functions, and b denotes the stress vector from
the trial functions. For 2-D elasticity
  " (1) (1) (1) #
 11  11 22 12
b = 22 ; Uv = (2) (2) (2) (28)
  11 22 12 v
12

in which the superscript (i) denotes the ith set of test functions, and
" # ( ) ( ) ( )
v11 v12 u1 t1 b1
v= ; u= ; t= ; b= (29)
v21 v22 u2 t2 b2

It should be noted that the two sets of test functions in v must be independent. In the present
work, v is chosen as

vij = vij or v = vI (30)

where ij is the Kronecker delta, I is the identity matrix, and v is taken as the weight function in
the MLS approximation, with the support ri being replaced by r0 , the size of the local sub-domain.
Substitution of equation (3) into the LSWF (27) for all nodes leads to the following discretized
system of linear equations:
P
N
Kij ûj = fi ; i = 1; 2; : : : ; N (31)
j=1

where N is the total number of nodes,

Z Z Z
Kij = Uv (x; xi )DBj d
+  v(x; xi )Sj d − v(x; xi )NDSBj d (32)

s su su

and
Z Z Z
fi = v(x; xi )t d +  v(x; xi )Su d + v(x; xi )b d
(33)
st su
s

with v(x; xi ) being the value of the test function matrix, corresponding to node i, evaluated at the
point x,
" #
n1 0 n2
N= (34)
0 n2 n1
 
j; 1 0
 
Bj =  0 j; 2  (35)
j; 2 j; 1

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
546 S. N. ATLURI AND T. ZHU

D is the stress–strain matrix:

 
1  0
E  
D=  1 0 for plane stress (36)
1 − 2
0 0 (1 − )=2
and S denotes the essential boundary condition index, dened as
 
S 0
S= 1 (37)
0 S2
with
(
1 if ui is prescribed on u
Si = (i = 1; 2) (38)
0 if ui is not prescribed on u

It can be easily seen that the system stiness matrix in the present method is banded but
unsymmetric. The locations of the non-zero entries in the system ‘stiness’ matrix depend upon
the nodes located inside the domain of in uence of the node.
A comprehensive discussion of the MLPG method, which may also lead to a symmetric stiness,
and may use other meshless interpolations such as partition of unity and Shephard functions, may
be found in Reference [15].

5. NUMERICAL EXAMPLES

In this section, some numerical results are presented to illustrate the implementation and con-
vergence of the present MLBIE and MLPG approaches. For the purpose of error estimation and
convergence studies, the displacement and energy norms, kuk and kUk (for the MLPG approach),
and the Sobolev norms kukk ; k = 0; 1 (for the MLBIE method) are calculated. These norms are
dened as
Z 1=2
kuk = u · u d

T
(39a)

 Z 1=2
1
kUk = UT · D · U d
(39b)
2

Z 1=2
kuk0 = u2 d
(39c)

and
Z 1=2
kuk1 = u2 + (|∇u|)2 d
(39d)

The relative errors for kuk; kUk and kukk are dened as
kunum − uexact k
ru = (40a)
kuexact k
kUnum − Uexact k
re = (40b)
kUexact k

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
 NEW CONCEPTS IN MESHLESS METHODS 547

Figure 3. Relative errors and convergence rates for the Dirichlet problem of Poisson’s equation: (a) for norm k · k0 ; and
(b) for norm k · k1

and
kunum − uexact kk
rk = ; k = 0; 1 (40c)
kuexact kk

5.1. Examples for the MLBIE approach

5.1.1. Poisson’s equation. The results from the present MLBIE formulation are studied for the
Poisson’s equation with a given source function p = x1 + x2 in the 2 × 2 domain, for which the
exact solution is taken to be

u = − 56 [(x1 )3 + (x2 )3 ] + 3(x1 )2 x2 + 3x1 (x2 )2 (41)

A Dirichlet problem is solved, for which the essential boundary condition is imposed on all
sides, and a mixed problem, for which the essential boundary condition is imposed on top and
bottom sides and the ux boundary condition is prescribed on left and right sides of the domain.
The MLS approximation with both linear and quadratic bases as well as Gaussian and spline
weight functions are employed in the computation. The size of support for both weight functions
are taken to be 2, and the parameter ci for Gauss weight function is ri =4.
Regular meshes of 9(3 × 3); 36(6 × 6) and 64(8 × 8) nodes are used to study the convergence
of the method. The local boundary integrals on @
s are evaluated by using 20 Gauss points on
each local boundary. The size (radius) of the local boundary for each node is taken as 0·005 in
the computation.
The convergence with mesh renement of the present method is studied for this problem. The
results of relative errors and convergence for the k · k0 and k · k1 norms are shown in Figure 3 for
the Dirichlet problem and in Figure 4 for the mixed problem, respectively. These gures show
that the present meshless method works quite well for the Poisson’s equation.
In this example, it can be seen that the quadratic basis yields somewhat of a better result than
the linear basis while both bases possess high accuracy. Also, the Gauss weight function works
better than the spline weight function. We should keep in mind that the appropriate parameters
ci in equation (7) need to be determined for all nodes for the Gauss weight function. The values
of these parameters will eect the numerical results considerably. With inappropriate ci used in

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
548 S. N. ATLURI AND T. ZHU

Figure 4. Relative errors and convergence rates for the mixed problem of Poisson’s equation: (a) for norm k · k0 ; and (b)
for norm k · k1

Figure 5. Flow around a cylinder in an innite eld and the model with boundary conditions

the Gaussian weight function, the results may become very unsatisfactory. The optimal choice of
these parameters is still an open research topic.

5.2. Potential ow

Consider the problem of a potential ow around a cylinder of radius a in an innite domain,

shown in Figure 5. u represents the stream function.
Due to symmetry, here only a part, 0 6 x1 6 4 and 0 6 x2 6 2, of the upper left quadrant of
the eld is modelled as shown in Figure 5. The exact solution for this problem is given by
 
x1
u=x 2
1− 2 2 (42)
(x ) + (x1 − L)2

Figure 5 shows the prescribed u and @u=@n values along all boundaries. The essential boundary
condition on the left and top edges is imposed according to the exact solution (42).
The initial mesh of 24 nodes is considered as shown in Figure 6(a). Subsequently, the number
of nodes is increased to 47 in Figure 6(b) and 74 in Figure 6(c) to study the convergence of the
present meshless method.

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
 NEW CONCEPTS IN MESHLESS METHODS 549

Figure 6. Flow around a cylinder: nodal arrangement: (a) 26 nodes; (b) 47 nodes; and (c) 74 nodes

Figure 7. Relative errors and convergence rates for the potential ow problem: (a) for norm k · k0 ; and (b) for norm k · k1

Both linear and quadratic bases as well as Gaussian and spline weight functions are considered.
We use ci = li and ri = 4ci in the calculation, where li is dened as

li = max kxj − xi k (43)

j ∈ Sj

where Sj is the minimum set of neighbouring nodes of xi which construct a polygon surroun-
ding xi .
The convergence for the Sobolev norms k · k0 and k · k1 is shown in Figure 7. The mesh size
h in this problem is dened as the average mesh sizes on the bottom edge. From Figure 7, we
notice that the quadratic basis is surprisingly not as good as the linear basis in this problem for
both Gaussian and spline weight functions.
The streamlines from the exact solution and from the numerical solution of the present meshless
formulation with a linear basis and Gaussian weight functions for the cases of 47 and 74 nodes are
also shown in Figure 8. It can be seen that the streamlines are well approximated by the present
method with 74 nodes as compared to the closed-form solution.

5.3. Examples for the MLPG approach

In the computation, we nd that the spline weight function works not as good as the Gaussian
weight function, therefore, only the Gaussian weight function is considered. Also, in the test
function v; r0 =ci is taken as 1.

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
550 S. N. ATLURI AND T. ZHU

Figure 8. The stream lines of the ow problem: (a) Linear Basis, Gaussian Weight Functions, 47 nodes; and (b) Linear
Basis, Gaussian Weight Functions, 74 nodes

Figure 9. Nodes for the cantilever beam with a parabolic-shear end load

5.3.1. Cantilever beam. The behaviour of the present MLPG formulation is studied in the canti-
lever beam problem (see Figure 9), for which the following exact solution is given [15] as
 
P D
u1 = − x −
2
[3x1 (2L − x1 ) + (2 + )x2 (x2 − D)] (44a)
6EI 2
"  2 #
P D 4 + 5 2 1
u2 = (x ) (3L − x ) + 3(L − x ) x −
1 2 1 1 2
+ D x (44b)
6EI 2 4
where
D3
I=
12
The stress corresponding to equations (44a) and (44b) are
 
P D
11 = − (L − x1 ) x2 − (45a)
I 2
22 = 0 (45b)
Px2 2
12 = − (x − D) (45c)
2I

The problem is solved for the plane stress case with P = 1; E = 1; D = 1; L = 8 and the penalty
parameter  = 106 . Both  = 0·25 and  = 0·4999 are considered, to study the locking property of
the present MLPG approach. Regular meshes of 52 (13 × 4); 72 (18 × 4) and 92 (23 × 4) nodes
are used.
In all cases, ri = 6h and ci = ri =4 are used, with h being the mesh size. In the formulation, the
size of each local sub-domain should be big enough such that the union of all local sub-domains

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
 NEW CONCEPTS IN MESHLESS METHODS 551

Figure 10. Relative errors and convergence rates for the displacement norm for the cantilever beam problem

Figure 11. Relative errors and convergence rates for the energy norm for the cantilever beam problem

covers as much as possible of the global domain. In this example, the size (radius) of the local
sub-domain of each internal node is taken as h2 , the mesh size in the x2 direction, and that of each
boundary node is taken as D=2. In the computation, nine Gauss points are used on each section
of s , and 6 × 9 points are used in each local domain
s for numerical quadratures.
The convergence with mesh renement of the present MLPG formulation is studied for this
problem. The results of relative errors and convergence rates are shown in Figures 10 and 11 for
displacements and strain energy, respectively. The mesh size h in these gures is dened as the
distance in x1 direction between two neighbouring nodes with same x2 co-ordinate. It can be seen
that the convergence rates for displacements and strain energy of the present formulation exceed
the corresponding convergence rates of the displacement-based FEM method, which are 2 and 1
for a linear basis, and 3 and 2 for a quadratic basis [16; 17].
Figures 10 and 11 also show that the present MLPG approach does not exhibit any volumetric
locking, even though no modication is made for nearly incompressible materials with v = 0·4999
in the plane stress case.
Figure 12 shows the normal stress 11 and the shear stress 12 for v = 0·25 at x1 = L=2 = 4 for
the present MLPG formulation with 52 and 72 nodes. The shear stress and normal stress are almost
as same as the exact solutions for 72 nodes, which are generally hard to obtain for the standard

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
552 S. N. ATLURI AND T. ZHU

Figure 12. (a) The normal stress 11 ; and (b) the shear stress 12 at x1 = L=2 for the cantilever beam problem

FEM. It is noted that the results for the present formulation satisfy the traction-free boundary
conditions at x2 = 0 and x2 = D = 1 almost exactly.

5.3.2. Innite plate with a circular hole. We consider an innite plate with a central hole:
(x1 )2 +(x2 )2 6a2 of radius a. The plate is subjected to a uniform tension,  = 1, in the x1 direction
at innity. The exact solution for stresses is
   
a2 3 3a4
11 =  1 − 2 cos 2 + cos 4 + 4 cos 4 (46a)
r 2 2r
 2  
a 1 3a4
12 =  − 2 sin 2 + sin 4 + 4 sin 4 (46b)
r 2 2r
 2  
a 1 3a4
22 =  − 2 cos 2 − cos 4 − 4 cos 4 (46c)
r 2 2r

where (r; ) are the polar co-ordinates and  is measured from the positive x1 -axis counter-
clockwise. The corresponding displacements, in plane strain case, are given by
 
1+ 1 2 a2 1 a2 1 a4
u1 =  r cos  + cos  + cos 3 − cos 3 (47a)
E 1+ 1+ r 2 r 2 r3
 
1+ − 1 −  a2 1 a2 1 a4
u2 =  r sin  − sin  + sin 3 − sin 3 (47b)
E 1+ 1+ r 2 r 2 r3

Due to symmetry, only a part, 0 6 x1 6 4 and 0 6 x2 6 4, of the upper right quadrant of the
plate is modelled as shown in Figure 13. Symmetry conditions are imposed on the left and bottom
edges, i.e. u1 = 0; t2 = 0 is prescribed on the left edge and u2 = 0; t1 = 0 on the bottom edge, and
the inner boundary at a = 1 is traction free. The traction boundary conditions given by the exact
solution (46a)–(46c) are imposed on the right (x1 = 4) and top (x2 = 4) edges (see Figure 13). In
the computation, a plane stress case with E = 1·0 and  = 0·25 is considered.
The initial mesh of 54 nodes, with six nodes in the r direction and nine nodes in the  direction, is
considered as shown in Figure 14(a). Subsequently, the number of nodes is increased to 99(11×9)

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
 NEW CONCEPTS IN MESHLESS METHODS 553

Figure 13. Plate with a central circular hole on fourfold symmetry; tractions at the outside boundaries are prescribed
according to the innite plate solution

Figure 14. Plate with a central circular hole: nodal arrangement: (a) 54 nodes; and (b) 99 nodes

and 135(15 × 9) to study the convergence of the present MLPG approach. The nodes are arranged
regularly in the  direction and irregularly in the r direction.
Both linear and quadratic bases are considered. We use ci = li and ri = 4ci in the calculation,
where li is dened as the jth smallest distance between node i and the other nodes. The size of a
local sub-domain is chosen as the minimum distance between the node in question and the other
nodes. In the computation, 11 Gauss points are used on each section of s , and 6 × 9 points are
used in each local domain
s for numerical quadratures.
The convergence for displacements and strain energy is shown in Figure 15. The mesh size
h in this problem is dened as the largest distance between neighboring nodes which form a
quadrilateral. For the present MLPG formulation, even though the convergence rates from the
quadratic basis are higher than those from the linear basis, the convergence rates for displacements
and strain energy with the linear basis exceed those of the conventional FEM, which are 2 and 1,
respectively [16; 17], while the convergence rates with the quadratic basis are not as good as those
of the FEM, which are 3 and 2, respectively. We note that the convergence rates for the present
formulation depend on the constants ci and ri in the weight functions. Optimal values of ci and
ri should be explored in order to obtain better results.

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
554 S. N. ATLURI AND T. ZHU

Figure 15. Relative errors and convergence rates for the problem of a plate with a circular hole: (a) for the displacement
norm; and (b) for the energy norm

Figure 16. Normal stress 11 at x1 = 0 for the problem of plate with a hole

The stress 11 at x1 = 0 obtained by the present MLPG method for linear elasticity with a
quadratic basis is also depicted in Figure 16. It can be seen that the steep stress 11 is well
approximated by the present method when 135 nodes are used.

6. CONCLUSIONS AND DISCUSSIONS

Two truly meshless methods, the MLBIE method based on a local boundary integral equation and
the MLPG approach based on the local symmetric weak form, are presented for solving boundary-
value problems, with the shape functions being from the MLS approximation. Convergence studies
in the numerical examples show that these two methods possess excellent rates of convergence for
both the unknown variables (displacements) and there derivatives (strain energy). Only a simple
numerical manipulation is needed for calculating the derivatives of the unknown variables as the
original approximated trial solution is smooth enough to yield reasonably accurate results for
derivatives. The numerical results show that using both linear and quadratic bases in the MLS
approximation can give quite accurate numerical results.

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
 NEW CONCEPTS IN MESHLESS METHODS 555

As is the case with other meshless methods based on a global Galerkin method, the present two
methods also possess the following advantages over the nite element method.
(a) The present methods are considerably more accurate for computing the values of the un-
known variables (displacements) and their derivatives (strains) than the nite element
method.
(b) No smoothing technique is required to compute the derivatives, as the original result is
smooth enough.
(c) No element connectivity is needed; and only randomly distributed nodes and the correspond-
ing sub-domains are constructed in the global domain.
Compared with the other meshless techniques, discussed in literature, based on a global Galerkin
formulation (for instance, the EFG method), the present approaches are found to have the following
advantages.
(i) Absolutely no elements are needed in the present two formulations, either for interpolation
purposes or for integration purposes, while shadow elements are required to evaluate volume
integrals, in the so-called meshless methods based on global weak forms.
(ii) No special integration scheme is needed to evaluate the volume and boundary integrals. The
integrals in the present methods can be easily evaluated over regularly shaped sub-domains
and their boundaries. The local boundary in general is the surface of a ‘sphere’ centred at the
node in question. This exibility in choosing the size and the shape of the local sub-domain
will lead to more convenient formulations in dealing with the non-linear problems.
Even though meshless methods have been reported in literature for about 20 years, meshless
methods based on local weak forms are very new. There are many aspects in these methods
which need to be explored, such as the selection of the parameters in the MLS approximation, the
selection of the sizes of the local sub-domains, the selection of test functions in the MLPG method,
etc. However, due to their exibility in solving practical problems, especially those problems with
discontinuities and moving boundaries, the meshless methods based on local weak forms have
shown a great promise in engineering applications.

ACKNOWLEDGEMENTS

This work was supported by research grants from the Oce of Naval Research, and Federal
Aviation Administration, with Y. D. S. Rajapakse, and C. C. Seher as cognizant program ocials.

REFERENCES
1. Lucy LB. A numerical approach to the testing of the ssion hypothesis. The Astronomical Journal
1997; 8:1013 –1024.
2. Nayroles B, Touzot G, Villon P. Generalizing the nite element method: diuse approximation and diuse elements.
Computational Mechanics 1992; 10:307–318.
3. Belytschko T, Lu YY, Gu L. Element-free Galerkin methods. International Journal for Numerical Methods in
Engineering 1994; 37:229–256.
4. Belytschko T, Organ D, Krongauz Y. A coupled nite element-element-free Galerkin method. Computational
Mechanics 1995; 17:186 –195.
5. Krysl P, Belytschko T. Analysis of thin plates by the element-free Galerkin methods. Computational Mechanics 1995;
17:26 –35.
6. Organ D, Fleming M, Terry T, Belytschko T. Continuous meshless approximations for nonconvex bodies by diraction
and transparency. Computational Mechanics 1996; 18:225 –235.

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)
556 S. N. ATLURI AND T. ZHU

7. Zhu T, Atluri SN. A modied collocation and a penalty formulation for enforcing the essential boundary conditions in
the element free Galerkin method. Computational Mechanics 1998; 21:211–222.
8. Mukherjee YX, Mukherjee S. On boundary conditions in the element-free Galerkin method. Computational Mechanics
1997; 19:264 –270.
9. Liu WK, Chen Y, Chang CT. Belytschko T. Advances in multiple scale kernel particle methods. Computational
Mechanics 1996; 18:73–111.
10. Zhu T, Zhang J-D, Atluri SN. A local boundary integral equation (LBIE) method in computational mechanics, and a
meshless discretization approach. Computational Mechanics 1998; 21:223–235.
11. Zhu T, Zhang J, Althuri SN. A meshless local boundary integral equation (LBIE) method for solving nonlinear problems
Computational Mechanics 1998; 22:174 –186.
12. Atluri T, Zhu T. A new meshless local Petrov–Galerkin (MLPG) approach in computational mechanics. Computational
Mechanics 1998; 22:117–127.
13. Atluri SN, Zhu T. A new meshless local Petrov–Galerkin (MLPG) approach for nonlinear problems in Computer
modeling and simulation. Computational Modelling Simulation Engineering (in press).
14. Atluri SN, Cho JY, Kim HG. Analysis of thin beams, using the meshless local Petrov–Galerkin method, with generalized
moving least squares interpolation. Computational Mechanics (In Press), 1999.
15. Atluri SN, Kim HG, Cho JY. A critical assessment of the truly meshless local Petrov–Galerkin (MLPG), and local
boundary integral equation (LBIE) methods. Computational Mechanics (In Press), 1999.
16. Lancaster P, Salkauskas K. Surface generated by moving least squares methods. Mathematics of Computation 1981;
37:141–158.
17. Timoshenko SP, Goodier JN. Theory of Elasticity, 3rd edn. McGraw-Hill: New York, 1970.
18. Hughes TJR. The nite Element Method. Prentice-Hall; Englewood Clis, NJ, 1987.
19. Strang G, Fix GJ. An Analysis of the Finite Element Method. Prentice-Hall; Englewood Clis, NJ, 1973.

Copyright ? 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 47, 537–556 (2000)