Comput. Methods Appl. Mech. Engrg.

191 (2002) 2337–2370

www.elsevier.com/locate/cma

Boundary cloud method: a combined scattered

point/boundary integral approach for boundary-only analysis
Gang Li a, N.R. Aluru b,*
a
Department of Mechanical and Industrial Engineering, 3265 Beckman Institute, MC-251, Beckman Institute for Advanced Science
and Technology, University of Illinois at Urbana–Champaign, 405 N. Mathews Avenue, Urbana, IL 61801, USA
b
Department of General Engineering, 3265 Beckman Institute, MC-251, Beckman Institute for Advanced Science and Technology,
University of Illinois at Urbana–Champaign, 405 N. Mathews Avenue, Urbana, IL 61801, USA
Received 7 February 2001; received in revised form 2 October 2001

Abstract

A boundary cloud method (BCM), for boundary-only analysis of partial differential equations, is presented in this
paper for solving potential equations in two dimensions (2-D). The BCM combines a weighted least-squares approach
for construction of interpolation functions with a boundary integral formulation for the governing equations. Given a
set of scattered points on the surface of an arbitrary object, Hermite-type interpolation functions are constructed by
developing a weighted least-squares approach. We also introduce truncated Hermite-type interpolation functions. The
boundary integrals are evaluated by using a cell structure. We propose various schemes for evaluating the singular,
nearly singular and nonsingular integrals. We also propose a true meshless approach for boundary-only analysis of
potential equations. Numerical results, comparing the classical boundary element method and the BCM, are presented
for several 2-D potential problems. Ó 2002 Elsevier Science B.V. All rights reserved.

Keywords: Boundary cloud method; Meshless; Boundary integral method; Scattered point interpolation; Hermite interpolation;
Boundary-only analysis

1. Introduction

Boundary integral formulations and boundary element methods (BEMs) [1] are attractive computational
techniques for linear and exterior problems as they reduce the dimensionality of the original problem. For
example, for 3-D problems, the BEM requires discretization of the 2-D surface of the 3-D object and for
2-D problems, the BEM requires discretization of the 1-D boundary. For exterior problems, the use of
classical methods, such as finite difference [2] or finite element methods [3], requires discretization of the
entire exterior, whereas with a BEM only the surface needs to be discretized. Since only the surface is
discretized in BEMs, the mesh generation process is not as intensive compared to interior methods such as
finite element methods; however, for complex surfaces, mesh generation can still be a bottleneck in BEMs.

*
Corresponding author.
E-mail address: [email protected] (N.R. Aluru).
URL: http://www.staff.uiuc.edu/
aluru.

0045-7825/02/$ - see front matter Ó 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 0 4 5 - 7 8 2 5 ( 0 1 ) 0 0 4 1 5 - 7
2338 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370

Two key steps in finite element and boundary element methods are (i) the construction of interpolation
functions and (ii) discretization and integration. If these two steps can be implemented by using only a
scattered set of points instead of using elements (as commonly used in finite element methods) or panels (as
commonly used in boundary element methods), then the complex mesh generation process can be allevi-
ated. A popular approach for construction of interpolation functions over a scattered set of points is to
employ a moving least-squares technique [4]. Recently, several mesh-free methods have been proposed in
the literature (see e.g. [5] for an overview) combining moving least-squares approach with Galerkin [6] as
well as collocation techniques [7,8] for interior problems.
A boundary node method [9], for boundary-only analysis, combining the moving least-squares approach
with a boundary integral equation has been proposed. Recently, the boundary node method has been
extended to 3-D potential equations [10] and elasticity [11]. A key difficulty in the boundary node method is
the construction of interpolation functions using moving least-squares methods. For 2-D problems, where
the boundary is 1-D, ðx; yÞ coordinates cannot be used to construct interpolation functions (this issue is
explained in detail in Section 5). Instead, a cyclic coordinate (s) is used in the moving least-squares ap-
proach to construct interpolation functions. For 3-D problems, where the boundary is 2-D, curvilinear
coordinates ðs1 ; s2 Þ are used to construct interpolation functions. The definition of these coordinates is not
trivial for complex geometries. The boundary integrals in the boundary node method are integrated by
employing a cell structure. The concept of a cell is quite different from that of a panel in BEMs. Cells are
used to sprinkle integration points or Gauss points and the boundary integrals are evaluated using the
integration points. In the boundary node method, the integration is performed by using a regular Gauss
quadrature except for the cell where the integration is singular.
In this paper we introduce a boundary cloud method (BCM), which combines a scattered point approach
for constructing interpolation functions with boundary integral equations for governing partial differential
equations. The key ideas in the BCM are:

1. We introduce a least-squares approach to construct Hermite-type interpolation functions. This approach

overcomes the difficulty encountered with the classical weighted least-squares approach to construct in-
terpolation functions. Regular Cartesian coordinates ððx; yÞ for 2-D or ðx; y; zÞ for 3-DÞ can be employed
in the new approach without resorting to cyclic (s) or curvilinear ðs1 ; s2 Þ coordinates.
2. A cell structure is employed for numerical integration. Depending on the nature of the boundary inte-
grals (classified as singular, nearly singular and nonsingular), different techniques are employed to eval-
uate the boundary integrals: (1) direct analytical integration is employed when the boundary integrals
are singular, (2) a Nystr€ om scheme [12,13] combined with a singular value decomposition technique
[14] is employed to construct the quadrature rules when the boundary integrals are nearly singular
and (3) a regular Gauss quadrature is employed when the boundary integrals are nonsingular. The pro-
posed integration scheme achieves high accuracy and improves the convergence of the method.

In many emerging application areas, such as nano- and micro-electromechanical systems (MEMS) [15],
exterior electrostatic analysis is a common requirement. Currently, the exterior electrostatic analysis is
performed by BEMs [16]. Since MEMS are geometrically very complicated, meshing the surface for exterior
electrostatic analysis can be quite involved. A BCM is attractive for electrostatic and other exterior analysis
problems as it reduces the dimensionality of the problem and alleviates the meshing requirement. In ad-
dition, since we include the normal derivative of the unknown in the weighted least-squares minimization
approach, the BCM produces more accurate results in capturing discontinuities in the normal derivative of
the solution.
The rest of the paper is outlined as follows: Section 2 summarizes the governing equations, the boundary
integral formulations and the classical BEM, Section 3 describes the construction of interpolation functions
using a weighted least-squares approach, Section 4 introduces a truncated Hermite-type interpolation
 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370 2339

approach, Section 5 discusses the properties of the Hermite-type interpolation functions, Section 6 de-
scribes the discretization and numerical integration of the boundary integrals, Section 7 summarizes the
global matrix assembly and imposition of boundary conditions, Section 8 introduces a true meshless ap-
proach for boundary-only analysis of potential problems, numerical results are shown in Section 9 and
conclusions are given in Section 10.

2. Governing equations and boundary-integral formulation

We will focus on only 2-D problems in this paper, but the approach can be extended for 3-D problems.
The governing equations along with the boundary integral formulations are summarized below.
Consider an arbitrary domain X, as shown in Fig. 1, on which the potential equation along with the
boundary conditions is to be solved, i.e.
r2 u ¼ 0 in X; ð1Þ

u¼g on Cg ; ð2Þ

ou
q¼ ¼ h on Ch ; ð3Þ
on
where u is the unknown potential, Cg is the portion of the boundary where Dirichlet boundary conditions
are specified, Ch is the portion of the boundary where Neumann boundary conditions are specified, g and h
are specified Dirichlet and Neumann boundary conditions, respectively, q is the normal derivative of u and
n is the outward normal vector.
A boundary integral equation for the potential problem is given by (see e.g. [1] for details)
Z  
o ln rðP ; QÞ
ln rðP ; QÞqðQÞ  ½uðQÞ  uðP Þ dSQ ¼ 0; ð4Þ
dX onQ
where lnðrÞ is Green’s function for the 2-D potential equation, P and Q are source and field points, re-
spectively, r is the distance between P and Q and nQ is the outward normal vector at a field point Q on
boundary dX.
A classical approach to solve Eq. (4) is to use a BEM [1]. In a BEM, the surface is discretized into panels
and the potential and its derivative are assumed to be constant on each panel. The centroid of each panel is
taken as the collocation point and the value of the potential and its derivative at the collocation point
represent the value of the potential and its derivative on the panel. The boundary integral equation for a
source point P can be written as

Fig. 1. An arbitrary domain on which the potential equation is to be solved.

2340 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370

K Z
X K Z
X 
o ln rðP ; Qk Þ
ln rðP ; Qk ÞqðQk Þ dSQ ¼ ½uðQk Þ  uðP Þ dSQ ; ð5Þ
k¼1 dSk k¼1 dSk onQk
where K is the number of panels, dSk is the length of kth panel and Qk is the field point on the kth panel.
Eq. (5) can be rewritten in a matrix form as
Hbu ¼ G b q; ð6Þ
where Hb and G b are K K coefficient matrices, and u and q are K 1 potential and its normal derivative
vector, respectively. The entries of the coefficient matrices are given by
Z
Gb ði; jÞ ¼ ln rðPi ; Qj Þ dSQ ; ð7Þ
dSj

Z
b ði; jÞ ¼ o ln rðPi ; Qj Þ 1
H dSQ þ dij ; i; j ¼ 1; . . . ; K: ð8Þ
dSj onQj 2
By substituting boundary conditions into Eq. (6), the unknown potential or its normal derivative can be
computed. In Section 9, the convergence of the BEM is compared with the convergence of the BCM.

3. Hermite-type interpolation

In a BCM, the surface of the domain is discretized into scattered points (see Fig. 1). The points can be
sprinkled randomly covering the boundary of the domain. Interpolation functions are constructed by
centering a weighting function at each point or node. The unknown quantities, u and q, are approximated
by a Hermite-type interpolation. For a 2-D problem, as shown in Fig. 2, given a point t, the unknown and
its normal derivative in the vicinity of the point t are approximated by
uðx; yÞ ¼ pT ðx; yÞat ; ð9Þ

opT ðx; yÞ
qðx; yÞ ¼ at ; ð10Þ
on
where p is the base interpolating polynomial, at is the unknown coefficient vector for point t and n is the
direction of the outward normal to the boundary (note that n can be different at every point). For the
interpolation region shown in Fig. 2, at is constant. When the interpolation region changes, at can change.

Fig. 2. The interpolation region for point t. The weighting function is centered at point t and vanishes outside the shaded region.
 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370 2341

In this paper, we use a linear polynomial basis. The base interpolating polynomial and its normal deriv-
atives are given by
pT ðx; yÞ ¼ ½ 1 x y ; m ¼ 3; ð11Þ
 
opT ðx; yÞ ox oy
¼ 0 ¼ ½0 cosðc ny Þ ;
nx cosðc
nxÞ ny m ¼ 3; ð12Þ
on on on
where (cnx and (c
nx) ny ) are the angles between the outward normal direction and the positive x- and y-axis,
ny
respectively.
For a point t, the unknown coefficient vector at is computed by minimizing
XNP h i2 X
NP  T 2
T op ðxi ; yi Þ
Jt ¼ wi ðxt ; yt Þ p ðxi ; yi Þat  u^i þ wi ðxt ; yt Þ at  q^i ; ð13Þ
i¼1 i¼1
on
where NP is the number of nodes, and wi ðxt ; yt Þ is the weighting function centered at ðxt ; yt Þ and evaluated at
node i whose coordinates are ðxi ; yi Þ. u^i and q^i are nodal parameters. The weighting function is nonzero
when the location of node i is within a certain distance from point ðxt ; yt Þ. Thus, for a point ðxt ; yt Þ, the
weighting function is nonzero only for a few other nodes in the vicinity of it. The region where the
weighting function is nonzero is called a cloud. The weighting function is typically a cubic spline or a
Gaussian function. In this paper, a cubic spline is used whose definition is given by
1 xt  xi
wi ðxt Þ ¼ ^
w ; ð14Þ
dx dx
8
> 0; z < 2;
>
>
> ðz þ 2Þ3 =6;
> 2 6 z 6  1;
>
>
< 2=3  z2 ð1 þ z=2Þ; 1 6 z 6 0;
^ ðzÞ ¼
w ð15Þ
>
> 2=3  z2 ð1  z=2Þ; 0 6 z 6 1;
>
>
>
> ðz  2Þ3 =6; 1 6 z 6 2;
>
:
0; z > 2;
where z ¼ ðxt  xi Þ=dx and dx denotes the support size of the weighting function in the x-direction. A multi-
dimensional weighting function can be constructed as a product of 1-D weighting functions. In 2-D, the
weighting function is given by
1 xt  xi 1 yt  yi
wi ðxt ; yt Þ ¼ ^
w ^
w ; ð16Þ
dx dx dy dy
where dy is the support size in the y-direction. The stationary of Jt leads to
ðPT WP þ P0T WP0 Þat ¼ PT W^
u þ P0T W^
q: ð17Þ
Eq. (17) can be rewritten as
Ct at ¼ At ^
u þ Bt ^
q; ð18Þ

at ¼ C1 u þ C1
t At ^ t Bt ^
q; ð19Þ
where Ct is an m m matrix, At is an m NP matrix and Bt is an m NP matrix, whose definitions are
given by

Ct ¼ PT WP þ P0T WP0 ; ð20Þ

2342 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370

At ¼ PT W; ð21Þ

Bt ¼ P0T W; ð22Þ
0
P and P are NP m matrices, ^ u and ^
q are NP 1 vectors, W is a square NP NP diagonal matrix and their
definitions are given by
2 3 2 3
opT ðx1 ;y1 Þ
T
6 pT ðx1 ; y1 Þ 7 6 on1
6 opT ðx2 ;y2 Þ 7
7
6 p ðx2 ; y2 Þ 7 6 7
6 7
P¼6 . 7; P0 ¼ 6
6
on2
.
7;
7 ð23Þ
6 .
. 7 6 .. 7
4 5 4 5
pT ðxNP ; yNP Þ opT ðxNP ;yNP Þ
onNP

2 3 2 3
u^1 q^1
6 u^2 7 6 q^2 7
6 7 6 7
u¼6
^ .. 7; q¼6
^ .. 7; ð24Þ
4 . 5 4 . 5
u^NP q^NP
2 3
w1 ðxt ; yt Þ 0 0 0
6 0 w ðx t ; yt Þ 0 0 7
6 2 7
W¼6 .. 7; ð25Þ
4 0 0 . 0 5
0 0 0 wNP ðxt ; yt Þ
where n1 ; n2 ; . . . ; nNP are the outward normal directions of nodes 1; 2; . . . ; NP , respectively. The unknowns,
u and q, can be expressed as
uðx; yÞ ¼ pT ðx; yÞat ¼ pT ðx; yÞðC1 u þ C1
t At ^ qÞ;
t Bt ^ ð26Þ

opT ðx; yÞ opT ðx; yÞ 1

qðx; yÞ ¼ at ¼ u þ C1
ðCt At ^ qÞ:
t Bt ^ ð27Þ
on on
In short form,
uðx; yÞ ¼ Mðx; yÞ^
u þ Nðx; yÞ^
q; ð28Þ

qðx; yÞ ¼ Sðx; yÞ^

u þ Tðx; yÞ^
q; ð29Þ
where Mðx; yÞ, Nðx; yÞ, Sðx; yÞ and Tðx; yÞ are 1 NP interpolation vectors and they are defined by
Mðx; yÞ ¼ pT ðx; yÞC1
t At ; ð30Þ

Nðx; yÞ ¼ pT ðx; yÞC1

t Bt ; ð31Þ

opT ðx; yÞ 1
Sðx; yÞ ¼ Ct At ; ð32Þ
on

opT ðx; yÞ 1
Tðx; yÞ ¼ Ct Bt : ð33Þ
on
It can be shown that Mðx; yÞ, Nðx; yÞ, Sðx; yÞ and Tðx; yÞ are multivalued, i.e. Mðx; yÞ, Nðx; yÞ, Sðx; yÞ
and Tðx; yÞ change when the weighting function is centered at different points. To construct a useful
 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370 2343

approximation, the interpolation function should be limited to a single value. In this paper, we limit the
interpolation function to a single value by computing the interpolations at the point where the weighting
function is centered, i.e. when the weighting function is centered at ðxt ; yt Þ, the only useful values are
Mi ðxt ; yt Þ, Ni ðxt ; yt Þ, Si ðxt ; yt Þ and Ti ðxt ; yt Þ, i ¼ 1; 2; . . . ; NP . For the vector of unknowns, u and q, Eqs. (28)
and (29) can be rewritten in a matrix form as
u ¼ M^
u þ N^
q; ð34Þ

q ¼ S^
u þ T^
q; ð35Þ
where M, N, S and T are NP NP interpolation matrices. At any point ðx; yÞ, the unknowns are evaluated
by
X
NP X
NP
uðx; yÞ ¼ MI ðx; yÞ^
uI þ NI ðx; yÞ^
qI ; ð36Þ
I¼1 I¼1

X
NP X
NP
qðx; yÞ ¼ SI ðx; yÞ^
uI þ TI ðx; yÞ^
qI ; ð37Þ
I¼1 I¼1
where
MI ðx; yÞ ¼ pT ðx; yÞC1
t pðxI ; yI ÞwI ðxt ; yt Þ; ð38Þ

NI ðx; yÞ ¼ pT ðx; yÞC1 0

t p ðxI ; yI ÞwI ðxt ; yt Þ; ð39Þ

opT ðx; yÞ 1
SI ðx; yÞ ¼ Ct pðxI ; yI ÞwI ðxt ; yt Þ; ð40Þ
on

opT ðx; yÞ 1 0
TI ðx; yÞ ¼ Ct p ðxI ; yI ÞwI ðxt ; yt Þ: ð41Þ
on

Remarks

1. If the second term in Eq. (13) is neglected, the weighted least-squares approach is identical to the fixed
weighted least-squares approach introduced in [8]. A variant of the second term considered in Eq. (13)
has been added in previous works [17,18] to moving least-squares approximation to handle Neumann
boundary conditions.
2. In a fixed least-squares approach, Eq. (17) becomes
PT WPat ¼ PT W^
u: ð42Þ
T
As discussed in Section 5.2, P WP becomes singular when all the nodes lie along a straight line.

4. Truncated Hermite-type interpolation

In the Hermite-type interpolation introduced in Eqs. (34) and (35), unknowns u and q are approximated
in terms of both ^u and ^
q. Four coefficient matrices, M, N, S and T, need to be computed. As will be shown
in the following sections, when the Hermite-type interpolations are used in the boundary integral equation,
the final matrix form of the boundary integral equation is complicated and extra computational cost is
required to solve the final matrix problem. In this section, we introduce a truncated Hermite-type inter-
polation which has the same construction as the Hermite-type interpolation except that the definition of the
2344 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370

Fig. 3. Definition of weighting function in a (a) regular Hermite-type interpolation; (b) truncated Hermite-type interpolation.

weighting function is different at certain points. The truncated Hermite-type interpolation can significantly
reduce the computational cost.
From the definition given in Eq. (16), wi ðxt ; yt Þ is a weighting function centered at point ðxt ; yt Þ (referred to
as a star point) and evaluated at node i. wi ðxt ; yt Þ is zero when the distance from the star point ðxt ; yt Þ to node
i is larger than the support size of the weighting function. In a regular Hermite-type interpolation, the
weighting function extends onto another boundary edge if necessary. For example, as shown in Fig. 3(a),
when the weighting function is centered at point 7, the weighting function is nonzero for nodes 1–10
(referred to as a 10-point cloud). Nodes 1–4 lie on one boundary edge while nodes 5–10 lie on a different
boundary edge. In a truncated Hermite-type interpolation, nodes lying on a different boundary edge are not
included in the cloud. For example, when the weighting function is centered at node 7, the weighting
function includes only nodes 5–10 (six-point cloud) and does not include nodes 1–4 as they lie on a different
boundary edge.
In a truncated Hermite-type interpolation, Eq. (18) can be written as
Ct at ¼ At ~
u þ Bt ~
q; ð43Þ
u and ~
where ~ q are the nodal parameters, Ct is an m m matrix, At is an m NP matrix, and Bt is an m NP
matrix whose definitions are given by
Ct ¼ PT WP þ P0T WP0 ; ð44Þ

At ¼ PT W; ð45Þ

Bt ¼ P0T W: ð46Þ
W is an NP NP diagonal matrix
2 3
 1 ðxt ; yt Þ
w 0 0 0
6 0  2 ðxt ; yt Þ 0
w 0 7
6 7
W¼6 .. 7; ð47Þ
4 0 0 . 0 5
0 0 0 w  NP ðxt ; yt Þ

wi ðxt ; yt Þ; star point ðxt ; yt Þ and node i are on the same boundary edge;
 i ðxt ; yt Þ ¼
w ð48Þ
0; otherwise:
 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370 2345

In a truncated Hermite-type interpolation, Eqs. (28) and (29) can be rewritten as

uðx; yÞ ¼ Mðx; yÞ~
u þ Nðx; yÞ~
q; ð49Þ

qðx; yÞ ¼ Sðx; yÞ~

u þ Tðx; yÞ~
q; ð50Þ
where Mðx; yÞ, Nðx; yÞ, Sðx; yÞ and Tðx; yÞ are 1 NP vectors whose definitions are given by
1
Mðx; yÞ ¼ pT ðx; yÞCt At ; ð51Þ
1
Nðx; yÞ ¼ pT ðx; yÞCt Bt ; ð52Þ

opT ðx; yÞ 1
Sðx; yÞ ¼ Ct At ; ð53Þ
on

opT ðx; yÞ 1
Tðx; yÞ ¼ Ct Bt : ð54Þ
on
As described in Section 3, the interpolation functions are only computed at the point where the weighting
function is centered so that the interpolation functions are limited to a single value. It is shown in Appendix
A that Nðx; yÞ ¼ 0 and Sðx; yÞ ¼ 0 in a truncated Hermite-type interpolation. As a result, Eqs. (49) and (50)
can be rewritten as

u ¼ M~
u; ð55Þ

q ¼ T~
q: ð56Þ

At any point ðx; yÞ, the unknowns are evaluated by

X
NP
uðx; yÞ ¼ M I ðx; yÞ~
uI ; ð57Þ
I¼1

X
NP
qðx; yÞ ¼ T I ðx; yÞ~
qI ; ð58Þ
I¼1

where
1
M I ðx; yÞ ¼ pT ðx; yÞCt pðxI ; yI Þ
wI ðxt ; yt Þ; ð59Þ

opT ðx; yÞ 1 0
T I ðx; yÞ ¼ Ct p ðxI ; yI Þ
wI ðxt ; yt Þ: ð60Þ
on

5. Properties of Hermite-type and truncated Hermite-type interpolations

The Hermite-type interpolation functions presented in the previous section have several interesting
properties. First, unlike the construction of the interpolation functions in the classical BEM, the approach
presented here does not need to know the connectivity information among the points or nodes. Points can
be sprinkled randomly covering the boundary of the domain. This property greatly alleviates the task of
mesh generation, which can be difficult and time consuming for complex geometries. Second, the classical
weighted least-squares approach has difficulty computing interpolation functions when points are placed
2346 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370

Fig. 4. Illustration of boundary clouds. A circular and a truncated rectangular cloud are shown.

only along the boundary. For example, when all the points lie along a straight line, the moment matrix in a
weighted least-squares approach becomes singular. Hermite-type interpolation functions overcome this
disadvantage by including the normal derivatives of the base interpolating polynomial for the boundary
nodes. Third, Hermite-type interpolation functions satisfy the consistency conditions, i.e. they reproduces
any linear combination of the base interpolating polynomial exactly. Each of these three aspects is dis-
cussed below in detail.

5.1. Meshless property of interpolation functions

The Hermite-type interpolation functions can be constructed by simply sprinkling points without
knowing the connectivity information among the nodes or by creating elements which connect various
nodes. As shown in Fig. 4, the boundary of the domain is first discretized into points. Interpolation
functions are constructed by centering a weighting function at each point. When a weighting function is
centered at a point (called star point), there are a few other points in the vicinity of the star point for which
the weighting function does not vanish. The vicinity of the star point for which the weighting function does
not vanish is referred to as a cloud. The points outside the cloud do not have any influence on the inter-
polation function for the star point. The parameters in the weighting function (for example dx and dy in Eq.
(16)) determine the cloud size. The weighting functions used in defining clouds provide versatility in the
construction of interpolation functions. For a uniform point distribution, we can use a circular cloud with
the value of the weighting function determined only by Euclidean distance from the star node. For a
uniform point distribution with unequal spacing along each axis (for example for 3-D problems where the
boundary is 2-D), a rectangular-shaped cloud can be used with the value of the 2-D weighting function
determined by the product of two 1-D weighting functions aligned with the coordinate axis. For random
point distributions, the size and shape of each cloud can be constructed independent of each other.

5.2. Boundary-only property of Hermite-type interpolation

For boundary-only problems, where points are distributed only along the boundary of the domain,
neither the classical 2-D kernel approximations [7] nor the weighted least-squares approximations [4,8] can
be used due to the singularity of the moment matrix. Since only the boundary is considered in boundary
integral formulations, for 2-D problems the boundary is 1-D, and when all the points in a cloud lie along a
straight line, the moment matrix in weighted least-squares and kernel approximations loses its rank.
However, the Hermite-type weighted least-squares interpolation restores the full rank of the moment
 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370 2347

matrix and can be used to compute interpolation functions for any distribution of points without any
difficulty.
To illustrate the problem with the classical weighted least-squares approximation, consider a boundary
segment as shown in Fig. 4. Without losing generality, we will assume that the equation of the boundary
segment is of the form y ¼ kx þ b. The moment matrix with a classical weighted least-squares approxi-
mation is given by
2 PNP PNP PNP 3
j¼1 wj ðxt ; yt Þ j¼1 wj ðxt ; yt Þxj j¼1 wj ðxt ; yt Þyj
6 PNP PNP PNP 7
Ct ¼ PT WP ¼ 6 4 j¼1 wj ðxt ; yt Þxj j¼1 wj ðxt ; yt Þxj
2 7
j¼1 wj ðxt ; yt Þxj yj 5: ð61Þ
PNP PNP PNP 2
j¼1 wj ðxt ; yt Þyj j¼1 wj ðxt ; yt Þxj yj j¼1 wj ðxt ; yt Þyj

As shown in Fig. 4, if all the points in the cloud are on the same boundary segment, we have yj ¼ kxj þ b for
ðxj ; yj Þ 2 cloud. In this case, the moment matrix becomes
2 PNP PNP nP o nP o 3
NP NP
j¼1 wj ðxt ; yt Þ j¼1 wj ðxt ; yt Þxj k j¼1 wj ðxt ; yt Þxj þ b j¼1 wj ðxt ; yt Þ
6P PNP nP o nP o 7
6 NP 2 NP 2 NP 7
Ct ¼ 6 j¼1 wj ðxt ; yt Þxj j¼1 w j ðx t ; yt Þx j k j¼1 w j ðx t ; yt Þx j þ b j¼1 w j ðxt ; yt Þx j 7: ð62Þ
4P PNP n PNP o n PNP o 5
NP
j¼1 wj ðxt ; yt Þyj j¼1 wj ðxt ; yt Þxj yj k j¼1 wj ðxt ; yt Þxj yj þ b j¼1 wj ðxt ; yt Þyj

It is obvious that the third column of the moment matrix is the linear combination of the first two columns;
thus the rank of the above matrix is 2. If Ct is singular, at cannot be computed uniquely (see Eq. (42)).
Nevertheless, in Hermite-type weighted least-squares interpolation, the moment matrix for a star point t is
given by
2 PNP PNP PNP 3
j¼1 wj ðxt ; yt Þ j¼1 wj ðxt ; yt Þxj j¼1 wj ðxt ; yt Þyj
6 PNP PNP PNP 7
Ct ¼ PT WP þ P0T WP0 ¼ 6 4 j¼1 wj ðxt ; yt Þxj j¼1 wj ðxt ; yt Þxj
2 7
j¼1 wj ðxt ; yt Þxj yj 5
PNP PNP PNP 2
j¼1 wj ðxt ; yt Þyj j¼1 wj ðxt ; yt Þxj yj j¼1 wj ðxt ; yt Þyj
2 3
0 0 0
6 PNP 2
PNP 7
þ64
0 j¼1 wj ðxt ; yt Þ cos ðc nx
nxÞ j¼1 wj ðxt ; yt Þ cosðc nx cosðc
nxÞ ny Þ 7:
ny
5
PNP PNP 2
0 j¼1 wj ðxt ; yt Þ cosðc nx cosðc
nxÞ ny Þ
ny j¼1 wj ðxt ; yt Þ cos ðcny Þ
ny
ð63Þ
For the boundary segment under consideration,
cosðc
nx
nxÞ
¼ k: ð64Þ
ny Þ
cosðc
ny

Thus the moment matrix is given by

Ct ¼ PT WP þ P0T WP0
2 PNP PNP PNP 3
j¼1 wj ðxt ; yt Þ j¼1 wj ðxt ; yt Þxj j¼1 wj ðxt ; yt Þðkxj þ bÞ
6 PNP PNP PNP 7
¼64 j¼1 wj ðxt ; yt Þxj
2 2
j¼1 wj ðxt ; yt Þfxj þ k cos ðc
2
ny Þg
ny 2
j¼1 wj ðxt ; yt Þfkxj þ bxj  k cos ðc
2
ny Þg 7
ny 5:
PNP PNP 2 2
PNP 2 2
j¼1 wj ðxt ; yt Þðkxj þ bÞ j¼1 wj ðxt ; yt Þfkxj þ bxj  k cos ðc ny Þg
ny j¼1 wj ðxt ; yt Þfðkxj þ bÞ þ cos ðc ny Þg
ny
ð65Þ
From Eq. (65), it is clear that the rank deficiency in classical weighted least-squares is compensated by
the normal derivatives of the basis functions. The moment matrix in the new weighted least-squares
2348 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370

approximation has full rank. For special cases, such as vertical or horizontal segments, where k ¼ 1 or
k ¼ 0, respectively, it can be shown that the new weighted least-squares approximation provides full rank
for the moment matrix.
By following the steps outlined above, it can be shown that the truncated Hermite-type interpolation
functions can also be computed using Cartesian coordinates, i.e. the moment matrix Ct of the truncated
Hermite-type interpolation has full rank.

5.3. Consistency of Hermite-type boundary interpolation

In this section, we show that the interpolation functions can approximate at least all the functions
included in the definition of pðx; yÞ. To show this, consider the approximation given in Eqs. (26) and (27)
uðx; yÞ ¼ pT ðx; yÞðC1 u þ C1
t At ^ qÞ;
t Bt ^ ð66Þ

opT ðx; yÞ 1
qðx; yÞ ¼ u þ C1
ðCt At ^ qÞ;
t Bt ^ ð67Þ
on
where
T
u ¼ ½^
^ u1 ; u^2 ; . . . ; u^NP  ; ð68Þ
T
q ¼ ½^
^ q1 ; q^2 ; . . . ; q^NP  ; ð69Þ
u^i and q^i are nodal parameters corresponding to node i.
Assigning to each u^i and q^i the values of the basis functions vector gives
ui ¼ pT ðxi ; yi Þ;
^ ð70Þ

qi ¼ p0T ðxi ; yi Þ;
^ ð71Þ
i.e.
u ¼ P;
^ ð72Þ

q ¼ P0 :
^ ð73Þ
Substituting Eqs. (72) and (73) into Eqs. (66) and (67), respectively, gives
uðx; yÞ ¼ pT ðx; yÞðC1 u þ C1
t At ^ qÞ ¼ pT ðx; yÞðC1
t Bt ^
1 0
t At P þ C Bt P Þ

¼ pT ðx; yÞC1 T 0T 0 T 1 T
t ðP WP þ P WP Þ ¼ p ðx; yÞCt Ct ¼ p ðx; yÞ; ð74Þ
which shows that the approximation in Eq. (66) can interpolate any function exactly as part of the defi-
nition of pðx; yÞ. Similarly,
opT ðx; yÞ 1 opT ðx; yÞ 1
qðx; yÞ ¼ u þ C1
ðCt At ^ qÞ ¼
t Bt ^ ðCt At P þ C1 0
t Bt P Þ
on on
opT ðx; yÞ 1 T opT ðx; yÞ 1 opT ðx; yÞ
¼ Ct ðP WP þ P0T WP0 Þ ¼ Ct Ct ¼ ð75Þ
on on on
Eqs. (74) and (75) show that Hermite-type interpolation can interpolate exactly any member of the basis
polynomial and, consequently, any linear combination of the basis polynomial.
It can easily be shown that the truncated Hermite-type interpolation also satisfies the consistency con-
ditions. However, unlike the Hermite-type interpolation, in a truncated Hermite-type interpolation, u is
only interpolated by ^ u and q is only interpolated by ^q. The shape functions M and T have the property of
partion of unity which are given by
 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370 2349

X
NP
M I ðx; yÞ ¼ 1; ð76Þ
I¼1

X
NP
T I ðx; yÞ ¼ 1: ð77Þ
I¼1

5.4. Comparison of interpolation functions

Consider a portion of the boundary shown in Fig. 5. Points are uniformly distributed on the boundary.
The horizontal segment of the boundary has 20 evenly distributed points. Interpolation functions Mðx; yÞ,
N ðx; yÞ, Sðx; yÞ and T ðx; yÞ are computed for four points on the horizontal segment using a five-point cloud.
The coordinates of the four points are point 1 ¼ ð0:025; 0Þ, point 2 ¼ ð0:475; 0Þ, point 3 ¼ ð0:525; 0Þ, point
4 ¼ ð0:975; 0Þ. A comparison of all the interpolation functions for the four points is shown in Figs. 6–9.
In the truncated Hermite-type approach, the only nonzero interpolation functions are Mðx; yÞ and
T ðx; yÞ. These two interpolation functions are shown Figs. 10 and 11 for the four points shown in Fig. 5.

Remarks

1. For the regular Hermite-type interpolation functions shown in Figs. 6–9, note that the interpolation
functions N and S are zero for interior nodes. For nodes on the edge of the boundary or close to the
edge of the boundary, N and S are not zero.

Fig. 5. A portion of boundary where various interpolation functions are computed and compared.

Fig. 6. Interpolation function M for the four points shown in Fig. 5.

2350 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370

Fig. 7. Interpolation function N for the four points shown in Fig. 5.

Fig. 8. Interpolation function S for the four points shown in Fig. 5.

2. Note that the interpolation functions for the interior nodes are identical with both truncated and non-
truncated approaches. However, for nodes close to the edge of the boundary, both approaches give dif-
ferent interpolation functions.

6. Discretization and integration

The boundary of the domain is discretized into cells for integration purpose. As shown in Fig. 12, each
cell contains a certain number of nodes and the number of nodes can vary from cell to cell. The concept of
cell is quite different from that of an element or a panel in BEMs. The cell can be of any shape or size and
the only restriction is that the union of all the cells equal the boundary of the domain. Assuming that the
 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370 2351

Fig. 9. Interpolation function T for the four points shown in Fig. 5.

Fig. 10. Interpolation function M for the four points shown in Fig. 5.

boundary is discretized into NC cells, the boundary integral equation for the potential problem given in
Eq. (4) can be rewritten as
XNC Z  
o ln rðP ; QÞ
ln rðP ; QÞqðQÞ  ½uðQÞ  uðP Þ dSQ ¼ 0; ð78Þ
k¼1 dSk onQ
where dSk is the boundary length of cell k. The integral over a single cell k can be expressed as
Z Z Z
o ln rðP ; QÞ o ln rðP ; QÞ
Ik ¼ ln rðP ; QÞqðQÞ dSQ  uðQÞ dSQ þ uðP Þ dSQ : ð79Þ
dSk dSk onQ dSk onQ
Consider the first integral in Eq. (79) where the integrand is log singular. When the source point P and
the field point Q are far away from each other, the singularity is weak. On the contrary, when the source
2352 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370

Fig. 11. Interpolation function T for the four points shown in Fig. 5.

Fig. 12. For integration purpose, the boundary of the domain is broken into cells.

point P and the field point Q are close to each other or coincide, the integral becomes nearly singular or
singular, respectively. The use of regular Gauss quadrature would produce significant error in such cases
and special treatment needs to be applied. As shown in Fig. 12, given a source point P, cells are classified
into three categories: (1) cells which are far away from the source point; (2) cells which are near the source
point; (3) a cell which includes the source point. As shown in Fig. 12, given a source point and a distance,
defined as accurate integration distance in this paper, a circle centered at the source point and with the
radius of a given distance can be drawn. Any cell which is either completely or partially inside the circle is
considered to be a nearby cell for the source point. If a cell is completely outside of the circle, it is defined as
a far-away cell. Varying numerical integration schemes are applied depending on the classification of the
cell and these are described below.
 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370 2353

6.1. Cells are far away from the source point

When a cell is far away from the source point P, regular Gauss quadrature is sufficient for numerical
integration. For a given cell k, the boundary integral in Eq. (79) can be written as

X
NGk X
NGk
o ln rðP ; Qj Þ Xk o ln rðP ; Qj Þ
NG
I1k ¼ gj ln rðP ; Qj ÞqðQj Þ  gj uðQj Þ þ uðP Þ gj ; ð80Þ
j¼1 j¼1
onQj j¼1
onQj

where NGk is the number of Gauss points in cell k and gj is the weight of Gauss point j. Substituting the
approximations for u and q at the Gauss points into Eq. (80), we obtain
NP X
X NGk h i XNP XNGk
o ln rðP ; Qj Þ h i
I1k ¼ gj ln rðP ; Qj Þ Si ðQj Þ^
ui þ Ti ðQj Þ^
qi  gj Mi ðQj Þ^
ui þ Ni ðQj Þ^
qi
i¼1 j¼1 i¼1 j¼1
onQj
NP h
X iX
NGk
o ln rðP ; Qj Þ
þ Mi ðP Þ^
ui þ Ni ðP Þ^
qi gj : ð81Þ
i¼1 j¼1
onQj

6.2. Cells are near the source point

When a cell is near the source point, the first integral in Eq. (79) becomes nearly singular. The regular
Gauss quadrature becomes inaccurate. The integration is performed by employing a Nystr€ om scheme
[12,13] and a singular value decomposition (SVD) [14] technique to compute the quadrature weights di-
rectly for the nodes (not Gauss points) within each cell. Consider the integration in Eq. (79) over cell k.
Assuming that there are Mk nodes in cell k, Eq. (79) can be rewritten as

X
Mk X
Mk
o ln rðP ; Qj Þ X
Mk
o ln rðP ; Qj Þ
I2k ¼ nj ln rðP ; Qj ÞqðQj Þ  nj uðQj Þ þ uðP Þ nj ; ð82Þ
j¼1 j¼1
onQj j¼1
onQj

where nj is the quadrature weight for node j for the first integral in Eq. (79) and nj is the quadrature weight
for node j for the second and third integrals in Eq. (79). Note that the quadrature points are the nodes
themselves in each cell and no additional quadrature points (such as Gauss points) are used for integration.
The weights nj and nj are computed such that the integration is exact for a set of basis functions. For
example, for the nodes in cell k, the nj weights are computed by integrating the basis functions ½1; x; y along
with Green’s function, i.e.

X
Mk Z
 
nj ln rðP ; Qj Þ ¼ ln rðP ; QÞ dSQ ; ð83Þ
j¼1 oSk

X
Mk Z
 
nj ln rðP ; Qj Þxj ¼ ln rðP ; QÞx dSQ ; ð84Þ
j¼1 oSk

X
Mk Z
 
nj ln rðP ; Qj Þyj ¼ ln rðP ; QÞy dSQ : ð85Þ
j¼1 oSk

Similarly the nj weights can be computed by integrating the basis functions with the derivative of Green’s
function
2354 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370

Mk 
X  Z
 o ln rðP ; Qj Þ o ln rðP ; QÞ
nj ¼ dSQ ; ð86Þ
j¼1
onQj oSk onQ

Mk 
X  Z
o ln rðP ; Qj Þ o ln rðP ; QÞ
nj xj ¼ x dSQ ; ð87Þ
j¼1
on Qj oSk onQ

Mk 
X  Z
 o ln rðP ; Qj Þ o ln rðP ; QÞ
nj yj ¼ y dSQ : ð88Þ
j¼1
onQj oSk onQ

Eqs. (83)–(85) can be rewritten as

Kn ¼ R ð89Þ

and Eqs. (86)–(88) can be rewritten as

Hn ¼ P: ð90Þ

In Eqs. (89) and (90), K and H are matrices of size 3 Mk , n and n are unknown vectors of size Mk 1,
and R and P are known right-hand side vectors of size 3 1. The right-hand side vectors, R and P,
are computed analytically. Typically, the rank of K and H matrices is 2 as all the nodes in a cell lie along
a linear segment. If only two nodes are contained in a cell (i.e. Mk ¼ 2), then there is a unique solution
for the linear systems given in Eqs. (89) and (90). However, there is no constraint on the number of
nodes that can be in each cell. If the number of nodes in a cell exceeds 2, the linear system of equations
in Eqs. (89) and (90) are under-determined. Similarly, if the number of nodes in a cell is less than 2,
Eqs. (89) and (90) are over-determined. In such cases, SVD [14] can be used effectively to solve the
linear system of equations. Using the SVD technique, the coefficient matrices K and H are decomposed
as

K ¼ UK WK VTK ; ð91Þ

H ¼ UH WH VTH ; ð92Þ

and the weights are obtained by

n ¼ VK WK UTK R; ð93Þ

n ¼ VH WH UTH P; ð94Þ
where WK and WH are diagonal matrices whose diagonal entries are the reciprocal of the diagonal entries of
WK and WH , respectively.
Once the weights are computed, by substituting the approximations of u and q into Eq. (82), we
obtain

NP X
X Mk h i XNP XMk
o ln rðP ; Qj Þ h i
I2k ¼ nj ln rðP ; Qj Þ Si ðQj Þ^
ui þ Ti ðQj Þ^
qi  nj Mi ðQj Þ^
ui þ Ni ðQj Þ^
qi
i¼1 j¼1 i¼1 j¼1
onQj
NP h
X iX
Mk
o ln rðP ; Qj Þ
þ Mi ðP Þ^
ui þ Ni ðP Þ^
qi nj : ð95Þ
i¼1 j¼1
onQj
 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370 2355

6.3. Cell contains the source point

When the source node P and field node Q fall into the same cell, a small portion of the boundary in the
vicinity of P (defined as C in Fig. 12), on which the integration of Green’s function is singular, is taken out
of the cell and is integrated analytically. The rest of the cell is integrated numerically and the weights of the
nodes are determined by using the approach described in Section 6.2.
For example, for cell k, the first integral in Eq. (79) can be rewritten as
I3k ¼ I3kdSk C þ I3Ck ; ð96Þ

X X
NP Mk 1
 
I3kdSk C ¼ nj ln rðP ; Qj Þ ½ðSi ðQj Þ^
ui þ Ti ðQj Þ^
qi Þ
i¼1 j¼1

X X
NP Mk 1
 
o ln rðP ; Qj Þ
 nj ½Mi ðQj Þ^
ui þ Ni ðQj Þ^
qi 
i¼1 j¼1
onQj

X X
NP Mk 1
 
o ln rðP ; Qj Þ
þ nj ½ðMi ðP Þ^
ui þ Ni ðP Þ^
qi Þ; ð97Þ
i¼1 j¼1
onQj

NP  Z
X 
opðQÞ
I3Ck ¼ ln rðP ; QÞ dSQ ½C1
P ðAP u^i þ BP q^i Þ
i¼1 C on Q

XNP  Z 
o ln rðP ; QÞ
 pðQÞ dSQ ½C1P ðAP u^i þ BP q^i Þ
i¼1 C onQ
XNP  Z 
o ln rðP ; QÞ
þ dSQ ½ðMi ðP Þ^
ui þ Ni ðP Þ^
qi Þ: ð98Þ
i¼1 C onQ

The weights nj and nj ðj ¼ 1; . . . ; Mk  1Þ in Eq. (97) are computed by employing Eqs. (83)–(90). In Eq.
(98), the cloud is centered at the source point P and CP , AP and BP are evaluated at point P. The integrals
on C in Eq. (98) are computed analytically (see Appendix B for details).
In summary, for a source point P, the discretized form of Eq. (79) can be written as

X
N1 X
N2
I1k þ I2k þ I3k ¼ 0; ð99Þ
k¼1 k¼1

where N1 is the number of far-away cells, N2 is the number of nearby cells and N1 þ N2 þ 1 ¼ NC.
The discretization and integration of the boundary integral equation given in Eq. (79) using the truncated
Hermite-type interpolation can be carried out along the same lines as the approach described in Sections
6.1–6.3 for the regular Hermite-type interpolations. The only difference is the interpolation itself, which for
a truncated Hermite-type approach is given by
X
NP
uðX Þ ¼ M i ðX Þ^
ui ; X ¼ P or Q; ð100Þ
i¼1

X
NP
qðX Þ ¼ T i ðX Þ^
qi ; X ¼ P or Q: ð101Þ
i¼1
2356 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370

7. Global matrix assembly and imposition of boundary conditions

When the regular Hermite-type interpolation is used, by substituting Eqs. 81, (95)–(98) into Eq. (99) for
each node, the final matrix form in the BCM is given by
u ¼ G^
F^ q: ð102Þ

In a general mixed boundary value problem, either u or q is specified at each point. By rearranging u and q,
we put all the knowns given by the boundary conditions into a vector x and all the unknowns which are to
be computed into another vector y. By appropriately exchanging the rows of M, N, S and T, Eqs. (34) and
(35) can be rewritten as
x ¼ M0 ^
u þ N0 ^
q; ð103Þ

y ¼ S0 ^
u þ T0 ^
q: ð104Þ
By combining Eqs. (102)–(104), a 2NP 2NP linear system is obtained:
    
F G ^ u 0
¼ : ð105Þ
M 0 N0 ^
q x
^u and ^
q are computed by solving the above linear system. Once ^u and ^q are known, the unknown vector y is
computed from Eq. (104).
When a truncated Hermite-type interpolation is used, the matrix form of the BCM and the interpolations
are given by
u ¼ G^
F^ q; ð106Þ

u ¼ M^
u; ð107Þ

q ¼ T^
q: ð108Þ
A procedure, similar to the one described in Eqs. (102)–(105) can be employed to compute the unknowns ^u
and ^
q.

8. Towards true meshless boundary-only analysis

In the approach described in Section 7, a background cell structure is used to compute weights for the
scattered points. Since a background cell structure is used for integration, the technique described in
Section 7 may not be considered true meshless even though the interpolation functions are constructed by a
true meshless approach.
A trapezoidal rule, which employs the nodal volumes of the nodes as weights, can also be used to
evaluate Eq. (79). For 2-D problems, where the boundary is 1-D, the nodal volumes can be computed in a
straightforward manner. For example, the weighting function centered at each point can be used to
compute the nodal volumes. For 1-D boundaries, as shown in Fig. 13, the nodal volume of node 2 is half of
the curve length between node 1 and node 3. The curve length between two nodes is approximated by the
distance between the two nodes.
Given a source point P (see Fig. 13), the first integral in Eq. (79) is either singular or close to singular in
the vicinity of the source point P . The use of trapezoidal rule in a small region that contains P gives large
error. Thus, for any given source point P , the boundary is discretized into three parts: (i) C1 which contains
point P and a small neighborhood––in this region the integral is computed analytically (described in Section
 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370 2357

Fig. 13. Discretization scheme of the true meshless boundary-only analysis.

6.3 and Appendix B); (ii) C2 , the vicinity of point P , where the integral is close to singular––in this region,
the weights of the nodes are computed by the SVD scheme described in Section 6.2; and (iii) C3 , which is the
rest of the boundary––for nodes on C3 , the nodal volume of each node is taken as the weight.
Summarizing, the first integral in Eq. (79) can be rewritten as
Z Z
½ ln rðP ; QÞqðQÞ dSQ ¼ ½ ln rðP ; QÞqðQÞ dSQ
oX C1 þC2 þC3

NP  Z
X 
opðQÞ
¼ ln rðP ; QÞ dSQ ½C1
P ðAP u
^i þ BP q^i Þ
i¼1 C1 onQ
X
N2 X
NP
 
þ nj ln rðP ; Qj Þ ½ðSi ðQj Þ^
ui þ Ti ðQj Þ^
qi Þ
j¼1 i¼1

X
N3 X
NP
 
þ vj ln rðP ; Qj Þ ½ðSi ðQj Þ^
ui þ Ti ðQj Þ^
qi Þ; ð109Þ
j¼1 i¼1

where nj is the weight for nodes contained on C2 and vj is the nodal volume of nodes contained on C3 .
Similarly, the second integral in Eq. (79) can be rewritten as
Z   Z  
o ln rðP ; QÞ o ln rðP ; QÞ
uðQÞ dSQ ¼ uðQÞ dSQ
oSk onQ C1 þC2 þC3 onQ
XNP  Z 
o ln rðP ; QÞ
¼ pðQÞ dSQ ½C1 P ðAP u^i þ BP q^i Þ
i¼1 C 1
on Q

X N2 X NP  
 o ln rðP ; Qj Þ
þ nj ½Mi ðQj Þ^
ui þ Ni ðQj Þ^
qi 
j¼1 i¼1
onQj

X NP 
N3 X 
o ln rðP ; Qj Þ
þ vj ½Mi ðQj Þ^
ui þ Ni ðQj Þ^
qi ; ð110Þ
j¼1 i¼1
onQj

where nj is the weight for nodes contained on C2 .

2358 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370

Finally, the third integral in Eq. (79) can be rewritten as

XNP Z
o ln rðP ; QÞ
½Mi ðP Þ^
ui þ Ni ðP Þ^
qi  dSQ
i¼1 oSk onQ
( Z  X )
XNP N2   X N3 
o ln rðP ; QÞ  o ln rðP ; Qj Þ o ln rðP ; Qj Þ
¼ ½ðMi ðP Þ^
ui þ Ni ðP Þ^ qi Þ dSQ þ nj þ vj :
i¼1 C1 onQ j¼1
onQj j¼1
onQj
ð111Þ
In Eqs. (109)–(111), N2 is the number of points on C2 , N3 is the number of points on C3 and
1 þ N2 þ N3 ¼ NP . Since C1 and C2 are defined locally for each source point, there is no global geometrical
discretization. Thus this approach can be considered as a true meshless approach. Even though the as-
signment of nodal volumes is trivial for 2-D problems (where the boundary is 1-D), the assignment of nodal
volumes for randomly distributed points on 3-D geometries (where the boundary is 2-D) can be involved.
In Section 9, we compare the performance of the true meshless approach with that of the BCM.

9. Numerical examples

In this section we present results for several problems using the BCM. Unless otherwise mentioned we use
the 2-D cubic spline weighting function as given in Eq. (16). The convergence of the method is measured by
using a global error measure [7]
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u NP h i2
1 u X
 ¼ ðeÞ t1 ðeÞ
uI  uI
ðcÞ
ð112Þ
ju jmax NP I¼1

where  is the error in the solution and the superscripts ðeÞ and ðcÞ denote, respectively, the exact and the
computed solutions.
Consider the solution of a Laplace equation on various domains as shown in Fig. 14. Fig. 14(a) is a 1 1
square domain where Dirichlet or Neumann boundary conditions are specified along the four edges of the
boundary. Five examples with different boundary conditions are presented. In Fig. 14(b), the 1 1 square
domain shown in Fig. 14(a) is rotated by 45° and two cubic Dirichlet and Neumann problems are solved
over the rotated domain. For a uniform distribution of nodes with a linear basis, the cloud sizes are chosen
to be rx ¼ ry ¼ 1:17Ds, where Ds is the spacing between the points. A uniform cell structure is used and each
cell contains two Gauss points. Except for the example which uses a random point distribution, each cell
contains four nodes in all the other examples. The accurate integration distance, defined in Fig. 12, is set to
be 0.3. Fig. 14(c) is a circular domain with a radius of 1.0. Dirichlet and Neumann problems are solved over
the circular domain. Nodes are uniformly distributed along the boundary of the domain. A uniform cell
structure with two Gauss points per cell is considered. The cloud sizes are chosen to be rx ¼ ry ¼ 1:17Ds,
where Ds is the perimeter of the circle divided by the number of points. The accurate integration distance
for the circular domain is set to be 0.8. Fig. 14(d) is an ellipse domain and the boundary of the domain is
described by
x ¼ a cosðhÞ; y ¼ b sinðhÞ; ð113Þ
where a ¼ 1:0 and b ¼ 0:5. Dirichlet and Neumann problems are solved over the ellipse domain. The cell
structure, the cloud size and the accurate integration distance are the same as those used for the circular
domain.
The first example is a Dirichlet problem on the square domain with a quadratic solution for u. Dirichlet
boundary conditions are specified along the four edges of the boundary. The boundary conditions are
 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370 2359

Fig. 14. (a) 1 1 square domain on which the potential equation is solved, (b) a rotated square domain, (c) a circular domain, (d) an
ellipse domain.

u ¼ x2  y 2 for x ¼ 0; y ¼ 0; x ¼ 1; y ¼ 1: ð114Þ
The exact solution for q is given by
8
for y ¼ 0; x ¼ 0;
ou <
0
q¼ ¼ 2 for x ¼ 1; ð115Þ
on :
2 for y ¼ 1:

The exact solution of u is quadratic, which is outside the linear base interpolating polynomial. This problem
is solved by the BCM with a uniform distribution of 8, 16, 32 and 64 points per edge to study the con-
vergence behavior. A comparison of convergence of the classical BEM (described in Section 2, refer Eqs.
(6)–(8)) and BEM is shown in Fig. 15. The convergence rate obtained from BEM is found to be 0.48. The
convergence rate of the boundary cloud method is found to be 1.46 for both regular and truncated Hermite-
type interpolations.
The second example is a Neumann problem on the square domain with a quadratic solution for u.
Neumann boundary conditions are specified on the boundary. Dirichlet boundary condition is specified at
one point to make the problem well posed. The boundary conditions are
2360 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370

Fig. 15. Comparison of convergence of BEM and BCM for a Dirichlet problem with quadratic exact solution for u.

8
for y ¼ 0; x ¼ 0;
ou <
0
¼ 2 for x ¼ 1; ð116Þ
on :
2 for y ¼ 1;

u ¼ 0:8789 for x ¼ 0:9375; y ¼ 0: ð117Þ

The exact solution for u is given by
u ¼ x2  y 2 for x ¼ 0; y ¼ 0; x ¼ 1; y ¼ 1: ð118Þ
Comparison between the BEM and BCM is shown in Fig. 16 for a uniform distribution of 8, 16, 32 and 64
points per edge. The convergence rate of the BEM is found to be 1.04. The convergence rate of the BCM is
found to be 2.10 for both regular and truncated Hermite-type interpolations.

Fig. 16. Comparison of convergence of BEM and BCM for a Neumann problem with quadratic exact solution for u.
 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370 2361

The third example is a Dirichlet problem on the square domain with a cubic exact solution for u. The
specified Dirichlet boundary conditions are
u ¼ x3  y 3 þ 3x2 y þ 3xy 2 for x ¼ 0; y ¼ 0; x ¼ 1; y ¼ 1: ð119Þ
The exact solution for q is given by
8 2
> 3y
> for x ¼ 0;
ou < 3x2 for y ¼ 0;
q¼ ¼ ð120Þ
on > > 3y 2 þ 6y  3 for x ¼ 1;
: 2
3x þ 6x  3 for y ¼ 1:
The convergence rate of the BEM and the BCM is shown in Fig. 17 employing a uniform distribution of
8, 16, 32 and 64 points per edge. The convergence rate of the BEM is found to be 0.69 and the conver-
gence rate of the BCM is found to be 1.39 for both regular and truncated Hermite-type interpolations.
The fourth example is a Neumann problem on the square domain with a cubic exact solution for u. The
specified boundary conditions are
8 2
> 3y
> for x ¼ 0;
ou < 3x2 for y ¼ 0;
¼ ð121Þ
on > > 3y 2 þ 6y  3 for x ¼ 1;
: 2
3x þ 6x  3 for y ¼ 1;

u ¼ x3  y 3 þ 3x2 y þ 3xy 2 for x ¼ 0:97; y ¼ 0: ð122Þ

The exact solution for u is given by
u ¼ x3  y 3 þ 3x2 y þ 3xy 2 for x ¼ 0; y ¼ 0; x ¼ 1; y ¼ 1: ð123Þ
The convergence rate of the BEM and the BCM is shown in Fig. 18 employing a uniform distribution of 8,
16, 32 and 64 points per edge. The convergence rate of the BEM is found to be 1.21. The convergence rate is
found to be 2.06 for regular Hermite-type interpolation and 2.09 for truncated Hermite-type interpolation.
The fifth example is identical to the first example except that the nodes are distributed randomly on the
boundary. The convergence plot with a random point distribution of 8, 16, 32 and 64 nodes per edge is
shown in Fig. 19. The exact solution and the numerical solution obtained with a random distribution of 32

Fig. 17. Comparison of convergence of BEM and BCM for a Dirichlet problem with cubic exact solution for u.
2362 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370

Fig. 18. Comparison of convergence of BEM and BCM for a Neumann problem with cubic exact solution for u.

nodes per edge are compared in Fig. 20. The cloud size in this case is rx ¼ ry ¼ 1:67Ds, where Ds is the
spacing between the points. The convergence rate is found to be 1.73 for regular Hermite-type interpolation
and 1.82 for truncated Hermite-type interpolation.
The sixth and seventh examples are cubic Dirichlet and Neumann potential problems on the rotated
square domain shown in Fig. 14(b). For the Dirichlet problem, the boundary condition is

u ¼ x3  y 3 þ 3x2 y þ 3xy 2 on all the boundary: ð124Þ

The exact solution for q is

ou
¼ ð3x2  3y 2 Þ½ cosðc
ny Þ  cosðc
ny nx þ 6xy ½ cosðc
nxÞ nx þ cosðc
nxÞ ny Þ
ny on all the boundary: ð125Þ
on

Fig. 19. Convergence of BCM with a random point distribution for a Dirichlet problem with quadratic exact solution for u.
 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370 2363

Fig. 20. Comparison of results of BCM and exact solution for a Dirichlet problem with quadratic exact solution for u.

The convergence of the BCM is shown in Fig. 21 and the convergence rate is found to be 1.41 for both
regular and truncated Hermite-type interpolations. For the Neumann problem, the boundary conditions
are
ou
¼ ð3x2  3y 2 Þ½ cosðc
ny Þ  cosðc
ny nx þ 6xy ½ cosðc
nxÞ nx þ cosðc
nxÞ ny Þ
ny on all the boundary; ð126Þ
on

u ¼ x3  y 3 þ 3x2 y þ 3xy 2 at one point on the boundary: ð127Þ

The exact solution for u is
u ¼ x3  y 3 þ 3x2 y þ 3xy 2 on all the boundary: ð128Þ
The convergence of the BCM is shown in Fig. 22. The convergence rate is found to be 2.57 for regular
Hermite-type interpolation and 2.62 for truncated Hermite-type interpolation.

Fig. 21. Convergence of BCM for a Dirichlet problem with cubic exact solution for u on a rotated domain.
2364 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370

Fig. 22. Convergence of BCM for a Neumann problem with cubic exact solution for u on a rotated domain.

The next two examples are Dirichlet and Neumann potential problems on the circular domain shown in
Fig. 14(c). For the Dirichlet problem, the boundary condition is
u ¼ r cosðhÞ ¼ x on all the boundary: ð129Þ
The exact solution for q is
ou
¼ cosðhÞ on all the boundary: ð130Þ
on
The convergence of the BCM is shown in Fig. 23 and the convergence rate is found to be 3.61 for regular
Hermite-type interpolation. For the Neumann problem, the boundary conditions are
ou
¼ cosðhÞ on all the boundary; ð131Þ
on

Fig. 23. Convergence of BCM for a Dirichlet problem on a circular domain.

 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370 2365

Fig. 24. Convergence of BCM for a Neumann problem on a circular domain.

u¼x at two points on the boundary: ð132Þ

The exact solution for u is
u¼x on all the boundary: ð133Þ
The convergence of the BCM is shown in Fig. 24. The convergence rate is found to be 3.55 for regular
Hermite-type interpolation.
The next two examples are Dirichlet and Neumann potential problems on the ellipse domain shown in
Fig. 14(d). The boundary conditions and the exact solutions for this example are identical to those provided
for the circular domain examples. The convergence of the BCM is found to be 2.74 (Fig. 25) for the
Dirichlet problem and 2.61 (Fig. 26) for the Neumann problem.
Finally, the first example is solved by using the true meshless approach introduced in Section 8. In this
case, for each source point the size of C1 (refer Fig. 13) is taken as the nodal volume of the source point and

Fig. 25. Convergence of BCM for a Dirichlet problem on an ellipse domain.

2366 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370

Fig. 26. Convergence of BCM for a Neumann problem on an ellipse domain.

Fig. 27. Comparison of convergence of true meshless approach and BCM for a Dirichlet problem with quadratic exact solution for u.

the size of C2 is the portion of the boundary covered by a 0:335 0:335 square region which is centered at
the source point. The convergence rate of the BCM and the true meshless approach is shown in Fig. 27
employing a uniform distribution of 8, 16, 32 and 64 points per edge. The convergence rate of the BCM is
found to be 1.46. The convergence rate of the true meshless approach is found to be 1.15.

10. Conclusion

A BCM combining scattered points and boundary integral equations is presented in this paper. One of
the difficulties encountered in combining moving least-squares or other least-squares based approaches with
boundary integral equations is the development of interpolation functions using Cartesian coordinates. In
the boundary node method, the interpolation functions are constructed by employing cyclic or curvilinear
coordinates. In this paper, we introduce the construction of Hermite-type interpolation functions for use
 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370 2367

with boundary integral equations. A key advantage with Hermite-type interpolation functions is the use of
Cartesian coordinates. We have also introduced the construction of truncated Hermite-type interpolation
functions, which reduces the computational cost of the BCM.
Numerical results are presented for several potential problems comparing the classical BEM and the
BCM. The BCM exhibits superior convergence behavior. However, it is more expensive compared to the
BEM because of the cost involved in constructing meshless interpolation functions. We also observed that
both regular Hermite-type and truncated Hermite-type interpolation functions exhibit similar convergence
behavior. We have also attempted to introduce a true meshless approach using boundary integral for-
mulations. Preliminary results obtained with the true meshless approach are promising but several issues
still need to be resolved.

Acknowledgements

This work was supported by an NSF CAREER award to N.R. Aluru and by a grant from DARPA
under agreement number F30602-98-2-0178. This support is gratefully acknowledged.

Appendix A. Definition of N and S in truncated Hermite-type interpolation

Given a star point t, the shape functions N and S, defined in Eqs. (52) and (53), in a truncated Hermite-
type approach, are given by
1
Nðx; yÞ ¼ pðx; yÞCt Bt ; ðA:1Þ

opT ðx; yÞ 1
Sðx; yÞ ¼ Ct At : ðA:2Þ
on
At any point ðx; yÞ, the shape functions can be rewritten as
1
N I ðx; yÞ ¼ pT ðx; yÞCt p0 ðxI ; yI ÞwI ðxt ; yt Þ; ðA:3Þ
1
S I ðx; yÞ ¼ p0T ðx; yÞCt pðxI ; yI ÞwI ðxt ; yt Þ; ðA:4Þ
where
pT ðx; yÞ ¼ ½ 1 x y ; m ¼ 3; ðA:5Þ
 
ox oy
p0T ðx; yÞ ¼ 0 ¼ ½ 0 cosðc ny Þ ;
nx cosðc
nxÞ ny m ¼ 3: ðA:6Þ
on on
In a truncated Hermite-type interpolation, all nodes included in a cloud have the same normal direction
(assuming all boundary edges are straight lines). Denoting wj ðxt ; yt Þ as wtj , the moment matrix Ct can be
expressed as
2 PNP t PNP t PNP t 3
j¼1 wj j¼1 wj xj j¼1 wj ðkxj þ bÞ
6 PNP t PNP t 2 PNP t 7
Ct ¼ 6
4 j¼1 wj xj
2 2
j¼1 wj fxj þ k cos ðc ny Þg
ny 2
j¼1 wj fkxj þ bxj  k cos ðc
2
ny Þg 7
ny 5:
PNP t PNP t 2 2
P NP t 2 2
j¼1 wj ðkxj þ bÞ j¼1 wj fkxj þ bxj  k cos ðc ny Þg
ny j¼1 wj fðkxj þ bÞ þ cos ðc ny Þg
ny
ðA:7Þ
2368 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370

The inverse of Ct is given by

2 3
c11 c12 c13
1 1 4 c21 c22 c23 5;
Ct ¼ ðA:8Þ
det½Ct 
c31 c32 c33
where
X
NP X
NP X
NP X
NP X
NP X
NP
c11 ¼ b2 wtj wtj x2j  b2 wtj xj wtj xj þ wtj x2j wtj cos2 ðc
ny Þ
ny
j¼1 j¼1 j¼1 j¼1 j¼1 j¼1

X
NP
2
X
NP X
NP X
NP
þ wtj ðkxj þ bÞ wtj k 2 cos2 ðc
ny Þ þ 2
ny wtj xj ðkxj þ bÞ wtj k cos2 ðc
ny Þ;
ny ðA:9Þ
j¼1 j¼1 j¼1 j¼1

X
NP X
NP X
NP X
NP X
NP X
NP
c12 ¼ c21 ¼ kb wtj wtj x2j  kb wtj xj wtj xj  wtj wtj cos2 ðc
ny Þ
ny
j¼1 j¼1 j¼1 j¼1 j¼1 j¼1

X
NP X
NP
 wtj ðkxj þ bÞ wtj k cos2 ðc
ny Þ;
ny ðA:10Þ
j¼1 j¼1

X
NP X
NP X
NP X
NP X
NP X
NP
c13 ¼ c31 ¼ b wtj wtj x2j þ b wtj xj wtj xj  wtj xj wtj k cos2 ðc
ny Þ
ny
j¼1 j¼1 j¼1 j¼1 j¼1 j¼1

X
NP X
NP
þ wtj ðkxj þ bÞ wtj k 2 cos2 ðc
ny Þ;
ny ðA:11Þ
j¼1 j¼1

X
NP X
NP X
NP X
NP X
NP X
NP
c22 ¼ k 2 wtj wtj x2j  k 2 wtj xj wtj xj þ wtj wtj cos2 ðc
ny Þ;
ny ðA:12Þ
j¼1 j¼1 j¼1 j¼1 j¼1 j¼1

X
NP X
NP X
NP X
NP X
NP X
NP
c23 ¼ c32 ¼ k wtj wtj x2j þ k wtj xj wtj xj þ wtj wtj k cos2 ðc
ny Þ;
ny ðA:13Þ
j¼1 j¼1 j¼1 j¼1 j¼1 j¼1

X
NP X
NP X
NP X
NP X
NP X
NP
c33 ¼ wtj wtj x2j  wtj xj wtj xj þ wtj wtj k 2 cos2 ðc
ny Þ:
ny ðA:14Þ
j¼1 j¼1 j¼1 j¼1 j¼1 j¼1

Substituting Eqs. (A.8)–(A.14) into Eq. (A.3)

1
N I ðx; yÞ ¼ pT ðx; yÞCt p0 ðxI ; yI ÞwtI
2 32 3
c11 c12 c13 0
1 6 76 7
¼ ½ 1 x y 4 c21 c22 c23 54 cosðc nx 5wtI
nxÞ
det½Ct 
c31 c32 c33 cosðcny Þ
ny
2 3
nx þ c13 cosðc
c12 cosðc
nxÞ ny Þ
ny
1 6 7
¼ ½3 1 x y 4 c22 cosðc ny Þ 5wtI ¼ 0:
nx þ c23 cosðc
nxÞ ny ðA:15Þ
det½Ct 
nx þ c33 cosðc
c32 cosðc
nxÞ ny Þ
ny
1
Recalling that Ct is symmetric, S I ðx; yÞ is given by
 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370 2369

1 T T
S I ðx; yÞ ¼ p0T ðx; yÞCt pðxI ; yI ÞwtI ¼ ½pT ðx; yÞCt p0 ðxI ; yI Þ wtI ¼ 0: ðA:16Þ
Thus, the shape functions N and S are both zero in a truncated Hermite-type interpolation.

Appendix B. Analytical integration of singular integrals

As shown in Eq. (98), products of Green’s function and the basis polynomial need to be integrated over a
straight segment which contains the source point. Green’s function for 2-D potential problems is given by
GðP ; QÞ ¼ ln rðP ; QÞ; ðB:1Þ
where r is the distance from P to Q.
The integrals that need to be computed are
Z
I1 ¼ GðP ; QÞ dSQ ðB:2Þ
dS
Z
I2 ¼ GðP ; QÞx dSQ ðB:3Þ
dS
Z
I3 ¼ GðP ; QÞy dSQ ðB:4Þ
dS

As shown in Fig. B.1, the source point is denoted as P0 . When the domain of integration is from P1 to P2 , the
integrals in Eqs. (B.2)–(B.4) are singular. However, the Cauchy principle values [1] of the integrals exist. In
this case, numerical integration could introduce large errors. Hence, analytical integration is performed to
ensure accuracy.
We will assume that the segment is of the form y ¼ kx þ b. The coordinates of the points are shown in
Fig. B.1. The Cauchy principle values of the integrals are given by
Z
I1 ¼ GðP ; QÞ dSQ
dS
pffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2
¼ 1 þ k 2 jx1  x0 j ln ðx1  x0 Þ þ ðy1  y0 Þ  1
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 
þ jx2  x0 j ln ðx2  x0 Þ2 þ ðy2  y0 Þ2  1 ; ðB:5Þ

Fig. B.1. Singular integration along a straight segment.

2370 G. Li, N.R. Aluru / Comput. Methods Appl. Mech. Engrg. 191 (2002) 2337–2370

Z
I2 ¼ GðP ; QÞx dSQ
pdSffiffiffiffiffiffiffiffiffiffiffiffiffi  qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 þ k2
¼ jx1  x0 j2 ln ðx1  x0 Þ2 þ ðy1  y0 Þ2  1
4
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 
2 2 2
 jx2  x0 j ln ðx2  x0 Þ þ ðy2  y0 Þ  1
pffiffiffiffiffiffiffiffiffiffiffiffiffi  qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 þ k2 2 2
þ jx1  x0 jx0 ln ðx1  x0 Þ þ ðy1  y0 Þ  2
2
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 
 jx2  x0 jx0 ln ðx2  x0 Þ2 þ ðy2  y0 Þ2  2 ; ðB:6Þ

Z
I3 ¼ GðP ; QÞy dSQ ¼ kI2 þ bI1 : ðB:7Þ
dS

References

[1] J.H. Kane, Boundary Element Analysis in Engineering Continuum Mechanics, Prentice-Hall, Engelwood Cliffs, NJ, 1994.
[2] G.E. Forsythe, W.R. Wasow, Finite Difference Methods for Partial Differential Equations, Wiley, New York, 1960.
[3] T.J.R. Hughes, The Finite Element Method, Prentice-Hall, Engelwood Cliffs, NJ, 1987.
[4] P. Lancaster, K. Salkauskas, Surface generated by moving least squares methods, Math. Comput. 37 (1981) 141–158.
[5] T. Belytschko, Y. Krongauz, D. Organ, M. Fleming, P. Krysl, Meshless methods: an overview and recent developments, Comput.
Meth. Appl. Mech. Engrg. 139 (1996) 3–47.
[6] T. Belytschko, Y.Y. Lu, L. Gu, Element free Galerkin methods, Int. J. Numer. Meth. Engrg. 37 (1994) 229–256.
[7] N.R. Aluru, G. Li, Finite cloud method: a true meshless technique based on a fixed reproducing kernel approximation, Int.
J. Numer. Meth. Engrg. 50 (10) (2001) 2373–2410.
[8] E. Onate, S. Idelsohn, O.C. Zienkiewicz, R.L. Taylor, C. Sacco, A stabilized finite point method for analysis of fluid mechanics
problems, Comput. Meth. Appl. Mech. Engrg. 139 (1996) 315–346.
[9] Y.X. Mukherjee, S. Mukherjee, The boundary node method for potential problems, Int. J. Numer. Meth. Engrg. 40 (1997)
797–815.
[10] M.K. Chati, S. Mukherjee, The boundary node method for three-dimensional problems in potential theory, Int. J. Numer. Meth.
Engrg. 47 (2000) 1523–1547.
[11] M.K. Chati, S. Mukherjee, The boundary node method for three-dimensional linear elasticity, Int. J. Numer. Meth. Engrg. 46
(1999) 1163–1184.
[12] L.M. Delves, J.L. Mohamed, Computational Methods for Integral Equations, Cambridge University Press, Cambridge, 1985.
[13] S. Kapur, D.E. Long, High-order Nystr€ om schemes for efficient 3-d capacitance extraction, in: 38th International Conference on
Computer-Aided Design, San Jose, CA, USA, 1998, pp. 178–185.
[14] G.H. Golub, C.F. Van Loan, Matrix Computations, John Hopkins University Press, Baltimore, MD, 1989.
[15] G.T.A. Kovas, Micromachined Transducers Sourcebook, McGraw-Hill, New York, 1998.
[16] N.R. Aluru, J. White, An efficient numerical technique for electromechanical simulation of complicated microelectromechanical
structures, Sensors Actuators A 58 (1997) 1–11.
[17] T.J. Liszka, C.A.M. Duarte, W.W. Tworzydlo, hp-Meshless cloud method, Comput. Meth. Appl. Mech. Engrg. 139 (1996)
263–288.
[18] S.N. Atluri, J.Y. Cho, H.-G. Kim, Analysis of thin beams, using the meshless local Petrov–Galerkin method, with generalized
moving least squares interpolations, Comput. Mech. 24 (1999) 334–347.