Engineering Analysis with Boundary Elements 35 (2011) 845–854

Contents lists available at ScienceDirect

Engineering Analysis with Boundary Elements

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

A new semi-analytical method with diagonal coefficient matrices for

potential problems
N. Khaji n, M.I. Khodakarami
Faculty of Civil and Environmental Engineering, Tarbiat Modares University, P.O. Box 14115-397, Tehran, Iran

a r t i c l e i n f o abstract

Article history: In this paper, a new semi-analytical method is proposed for solving boundary value problems of two-
Received 10 October 2010 dimensional (2D) potential problems. In this new method, the boundary of the problem domain is
Accepted 27 January 2011 discretized by a set of special non-isoparametric elements that are introduced for the first time in this
Available online 2 March 2011
paper. In these new elements, higher-order Chebyshev mapping functions and new special shape
Keywords: functions are used. The shape functions are formulated to provide Kronecker Delta property for the
Semi-analytical method potential function and its derivative. In addition, the first derivative of shape functions are assigned to zero
Diagonal coefficient matrices at any given control point. Finally, using weighted residual method and implementing Clenshaw–Curtis
Non-isoparametric element quadrature, the coefficient matrices of equations system become diagonal, which results in a set of
Clenshaw–Curtis quadrature
decoupled governing equations for the whole system. This means that the governing equation for each
Chebyshev polynomials
degree of freedom (DOF) is independent from other DOFs of the domain. Validity and accuracy of the
Decoupled differential equations
Potential problems present method are fully demonstrated through four benchmark problems.
& 2011 Elsevier Ltd. All rights reserved.

1. Introduction required to be stored, which leads to a saving in storage space and

A potential problem involves one of the most important partial
differential equations as it occurs in pollutant diffusion problems,
electrostatics and gravitational fields, and incompressible fluid
flow and steady-state heat conduction problems [1]. From math-
ematical point of view, well-known potential problems include
Laplace or Poisson equation in a specific domain with a number of
boundary conditions. From physical viewpoint, space coordinates
are the sole coordinates of the problem that is independent of
time variable.
Potential problems may be analytically solved when the
geometry of the problem is simple. In cases of complex geometry
and/or boundary conditions, these problems should be solved
using numerical methods.
Various types of numerical methods such as finite element
method (FEM), boundary element method (BEM), scaled bound-
ary finite element method (SBFEM), and meshless methods are
commonly used to model potential problems. The use of FEM is
advantageous as the procedures are versatile in nature and well-
established [2]. Alternatively, BEM requires substantially reduced
surface discretizations, and is appealing an alternative to FEM for
potential problems [3–11]. BEM does not require the domain of
the problem to be discretized. Consequently, fewer unknowns are
n
Corresponding author. Tel.: +98 21 82883319; fax: + 98 21 82883381.
E-mail address: nkhaji@modares.ac.ir (N. Khaji).
CPU time. BEM requires a fundamental solution to the governing
differential equation in order to obtain the boundary integral
equation. In other words, BEM needs fundamental solutions that
are dependent on the problem of interest. Although the coeffi-
cient matrices of BEM are much smaller than those of FEM, they
are usually non-symmetric, non-positive definite, and fully
populated.
Combining the advantages of FEM and BEM, SBFEM has been
successfully developed [12]. SBFEM discretizes only the boundary
of the domain of interest with surface finite elements by trans-
forming the governing partial differential equations to a system of
ordinary differential equations, which may be solved analytically.
SBFEM, which needs no fundamental solution as for BEM, have
also been used for the analysis of potential problems (see, for
example [13–15]).
During the last decade, researchers have paid attention to
meshless methods, which are the mesh reduction methods with
no mesh requirements and only boundary nodes are necessary.
Several meshless methods have been reported in the literature
among which, various element-free methods [16–18], various
boundary node methods [19–22], boundary face method [23],
boundary point method [24], local boundary integral equa-
tion [25], and coupled FEM and meshless local Petrov–Galerkin
method [26] have been employed in the analysis of potential
problems.
In this paper, two-dimensional (2D) potential problems are
investigated using a novel semi-analytical method. In this new

0955-7997/$ - see front matter & 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.enganabound.2011.01.011
846 N. Khaji, M.I. Khodakarami / Engineering Analysis with Boundary Elements 35 (2011) 845–854

method, only domain boundaries are discretized, and hence, the final governing equation of potential problems as
governing partial differential equations, Laplace, or Poisson equa-
rT krc þ Q b þQ s ¼ kr2 c þ Q b þQ s ¼ 0 ð6Þ
tions are analytically solved in the problem domain. The elements
of domain boundary are of special non-isoparametric ones, in that should be solved in domain O. On boundaries of such a
which higher-order Chebyshev mapping functions and new spe- domain, the specified boundary conditions should be satisfied.
cial shape functions are employed. Proposing a weak form of
weighted residual method and using Clenshaw–Curtis quadra-
ture, the coefficient matrices of equations system become diag- 3. The proposed numerical method
onal, which leads to decoupled ordinary differential equations for
the whole system. This means that the governing equation for In this section, we briefly outline how the proposed new
each degree of freedom (DOF) is independent from other DOFs of method may be used to solve the governing equation of potential
the domain. Efficiency and accuracy of the present method are problems, numerically. In the present method, only the bound-
demonstrated through some benchmark problems. aries of the domain are discretized, similar to what happens in the
SBFEM. Nevertheless, this is the only similarity between the
present method and the SBFEM. It means that the boundaries,
in the present method, are discretized using new higher-order
2. General statement of potential theory non-isoparametric elements. To calculate the coefficient matrices
of the present method, Clenshaw–Curtis quadrature [27] will be
As already discussed, governing equation of potential pro- implemented. Finally, using higher-order Chebyshev mapping
blems, such as diffusion, seepage, electrostatic, and torsion, is function and new special shape functions, the coefficient matrices
Laplace or Poisson equation. Let us consider a 2D domain O with of the set of equations will be diagonal. As a result, governing
Dirichlet (Gf) and Neumann (Gq) boundary conditions (see Fig. 1). equations of whole system will be decoupled. In other words, the
The flow per unit area {q} may be written in terms of its Cartesian governing equation for each DOF is independent from other DOFs
components as of the domain.
( ) To this end, a special approach has been proposed to produce
qx
fqg ¼ ð1Þ the weak form of governing equations, which yields much
qy
accurate solutions for the problems under study. Therefore, this
For a steady-state flow, the balance or continuity requirement method may be called as a semi-analytical solution method for
gives potential problems, which shows a numerical scheme on the
boundaries of the domain and presents an analytical resolution
@qx @qy inside the domain O.
þ þQb þQs ¼ 0 ð2Þ
@x @y
b b s s
where Q b ¼ ½ Qx Qy T and Q s ¼ ½ Qx Qy T represent the rate at 3.1. Mesh definition
which the relevant quantity is generated per unit volume and unit
surface, respectively. Introducing the gradient operator as fol- In order to analyze a problem using numerical methods,
lows: geometry discretization of the problem is a key step. In the
(@) present method, similar to the SBFEM [12], a local coordinates
@x origin (LCO) is selected from which all boundaries of the domain
r¼ @ ð3Þ
are visible (Fig. 2). For bounded domain, the LCO is chosen inside
@y
the domain or on the boundary. As a result, the total boundary of
one may rewrite Eq. (2) as the domain consists of two regions: the region that passes
through the LCO (e.g., G1 and G5 in Fig. 2a), and the other
rT fqg þ Q b þQ s ¼ 0 ð4Þ remaining region. In the present method, only the region that
Generally, the rates of flow may be related to gradients of does not pass through the LCO should be decomposed into ne one-
some potential quantity c. The potential quantity is temperature dimensional (1D) non-overlapping non-isoparametric elements
in the case of heat flow, for example. A general linear relationship Ge, e¼1, 2,y, ne.
may be given in the following form: In the present method, geometry transmission from global
8 9 Cartesian coordinates ð_ x,_
yÞ to local dimensionless coordinates
( )
  qx < @@xc = (x, Z) is obtained by Chebyshev polynomials as mapping functions.
q ¼ ¼ k @c ¼ krc ð5Þ In addition, potential functions are interpolated by special shape
qy : @y ;
functions that are introduced in this paper. Mapping functions and
where k is a constant coefficient, which is corresponding to the the special shape functions are explained in next sections.
problem of interest. Substituting Eq. (5) into Eq. (4) results in the
3.2. Mapping functions
q (q = q)
The next step after discretization of the boundaries is approx-
imation of geometry with mapping functions. Each element on
the boundary is analogous to a line; hence, we adopt a suitable
mapping between the line (master/reference element) and each
Ω
element Ge.
Two local coordinates x and Z may be defined for mesh
definition. The first coordinate x is a radial coordinate from the
LCO to the boundaries. The master line is defined in terms of the
 ( = )
second coordinate Z, 1 r Z r1, which is sometimes referred to
Fig. 1. A sample 2D domain (O) with Dirichlet (Gf) and Neumann (Gq) boundary as tangential coordinate on the boundaries. The radial coordinate
conditions for potential problems. x is equal to zero in the LCO and equal to 1 on the boundaries.
 N. Khaji, M.I. Khodakarami / Engineering Analysis with Boundary Elements 35 (2011) 845–854 847

of boundary in the following form:

y 3(x3, y3) _
xðx, ZÞ ¼ xxðZÞ ¼ x½FðZÞfxg ð9Þ

_
yðx, ZÞ ¼ xyðZÞ ¼ x½FðZÞfyg ð10Þ
Γ4 Γ3
The proposed mapping functions for an (nZ + 1)-node element
is defined as [28]
2 (x2 , y2)
4 (x4, y4) 2 X
nZ
1
ξ fi ðZÞ ¼ Tn ðZi ÞTn ðZÞ ð11Þ
nZ n ¼ 0 ci cn
Ω
in which Tn(Z)are the first-kind Chebyshev polynomials of the
order n. Moreover, cn ¼1 for 0 on onZ, and cn ¼2 for n ¼0, nZ. As a
result, the Chebyshev polynomials are equal to either zero or one

Γ2
Γ5

at any given control point (or node):

fa ðZb Þ ¼ dab ð12Þ
x
where d denotes the Kronecker Delta function.
O (0, 0) Γ1 1 (x1 , y1) For reasons discussed in the next paragraph, the control points
Zn, n ¼0, 1,y, nZ, needed in the definition (11) are placed at
special positions called Chebyshev–Lobatto–Legendre (CLL)
ξ
points, which are [28]
η 3 (ξ = 1, η = 1)  
np
Zn ¼ cos ; n ¼ 0,1,. . .,nZ ð13Þ
nZ
nξ
which means each element has (nZ + 1) CLL points for a poly-
Γ3 nomial of degree nZ.
nη ξ
As will be shown later, the choice of CLL points is motivated by
the fact that the combination of Clenshaw–Curtis integration
Ω3 2 (ξ = 1, η = −1) method and special shape functions exactly leads to diagonal
coefficient matrices and decoupled equations. For a three-node
element, the locations of control points and mapping functions
are plotted in Fig. 3.
y In what follows, 2D formulation of the potential problem is
presented and discussed. A differential element of area d_ x d_
y is
related to a differential element of area dx dZ by the following
equation:
x
_
dO ¼ d_ xd_y ¼ 9Jðx, ZÞ9dx dZ ¼ x9JðZÞ9dx dZ ð14Þ
O (ξ = 0)
_
where Jðx, ZÞ denotes the Jacobian matrix of transformation:
Fig. 2. Modeling of 2D bounded domain and local coordinates origin (LCO):
 ! !  (_ )
(a) in global coordinates system and (b) in scaled boundary local system.  @ _ @_ x  ! x
_  x _
Jðx, ZÞ ¼   ; x ¼ _ ð15Þ
 @x @Z  y
In the SBFEM, similar to isoparametric finite elements, map-
ping functions and shape functions are identical. Nevertheless, in On the boundary, the Jacobian matrix may be given as
the present method, the proposed elements are non-isopara- " #
metric with mapping functions [F(Z)] that are different from xðZÞ yðZÞ
JðZÞ ¼ ð16Þ
shape functions [N(Z)]. Each element Ge is defined in terms of a x, Z ðZÞ y, Z ðZÞ
set of nZ mapping functions fi(Z), i¼ 1, 2,y, nZ. The geometry of
elements in local coordinates is introduced as
8 9 1.2 φ1 φ2 φ3
>
> x1 >>
>
> > 1
< x2 >=
xðZÞ ¼ ½f1 ðZÞ, f2 ðZÞ,. . ., fnZ ðZÞ ¼ ½FðZÞfxg ð7Þ 0.8
>
> ^ >>
>
> >
: xnZ >
; 0.6
φ ( )

0.4
8 9 0.2
>
> y1 >>
>
> >
< y2 >= 0
yðZÞ ¼ ½f1 ðZÞ, f2 ðZÞ,. . ., fnZ ðZÞ ¼ ½FðZÞfyg ð8Þ
>
> ^ >>
-0.2
>
> >
: ynZ >
; -0.4
-1 -0.5 0 0.5 1

where x and y are global coordinates of boundary points.
In the present method, any point in the domain with _ x and _
y Fig. 3. Location of control points and mapping functions for a sample three-node
coordinates is related to the corresponding point on the elements element.
848 N. Khaji, M.I. Khodakarami / Engineering Analysis with Boundary Elements 35 (2011) 845–854

The spatial derivatives for two-coordinate systems are related unknown constant coefficients as follows:
as
8 9 Ni ðZÞ ¼ a0 þa1 Z þ a2 Z2 þa3 Z3 þ    þa2nZ 1 Z2nZ 1 ð29Þ
@ =
< @_
x   @  2 1 @
@ ¼ b1 ðZÞ þ b ðZÞ ð17Þ The constant coefficients of the polynomial are obtained using
: @_
y; @x x @Z Eqs. (27) and (28). In the present method, the control points are
those of mapping functions (i.e., CLL points of Eq. (13)).
where
( ) How to determine this new shape functions? To define the
  1 y, Z ðZÞ polynomial of order 2nZ  1, (nZ + 1) control points are needed,
b1 ðZÞ ¼ ð18Þ
9JðZÞ9 x, Z ðZÞ which are calculated using Eq. (13). Finally, by implementing
Eq. (29) in Eqs. (27) and (28), the constant coefficients ai may be
( )
  1 yðZÞ determined.
2
b ðZÞ ¼ ð19Þ For a three-node element as an example, nZ ¼ 2, and therefore
9JðZÞ9 xðZÞ
Z0 ¼  1, Z1 ¼0 and Z2 ¼ + 1. Using Eqs. (27)–(29), the shape
Along the radial lines passing through the LCO and a node on functions may be found as
the boundary, the nodal potential function on the boundary,
N1 ðZÞ ¼ Z2 54 Z3 12 Z4 þ 34Z5 ð30Þ
{c(x)}, is introduced. The potential {c(x, Z)} at point (x, Z) is
obtained by interpolation of the potential using shape functions
over each element. N2 ðZÞ ¼ 12Z2 þ Z4 ð31Þ
The unit normal vector {n} on the boundary is defined as
N3 ðZÞ ¼ Z2 þ 54 Z3 12 Z4 34Z5 ð32Þ
1 !
fng ¼ ! rx ð20Þ The shape functions for a three-node element are shown in
:r x :
Fig. 4. Moreover, constant coefficients of shape functions (i.e., ai in
Substituting Eqs. (7) and (8) into Eq. (20) results in the Eq. (29)) regarding to various common elements (with nZ r2) of
following relations: the present method are given in Table 1.
( ) As already discussed, the potential {c(x, Z)} at any point (x, Z)
n o 1 y, Z ðZÞ
nx ðZÞ ¼ ( ) ð21Þ may be obtained by interpolation of the potential using shape
 y, Z ðZÞ  x, Z ðZÞ
  functions. Therefore, the potential at any point in the domain may
 
 x, Z ðZÞ  be evaluated by {c(x,Z)} ¼[N(Z)] {c(x)}. Furthermore, the deriva-
( ) tive of potential at any given point (x, Z) may be calculated using
 Z  1 yðZÞ Eqs. (5) and (17) as follows:
n ðZÞ ¼ ( ) ð22Þ
 yðZÞ  xðZÞ  
      1 2   
  qðx, ZÞ ¼ k b1 ðZÞ NðZÞ fcðxÞg, x þ b ðZÞ ½NðZÞ, Z cðxÞ
 xðZÞ  x
Using Eqs. (21) and (22), Eqs. (18) and (19) may be written as ð33Þ
( ) Note that the derivatives of shape functions only at the control
 1  1  y, Z ðZÞ  x
b ðZÞ ¼    fn ðZÞg ð23Þ points are equal to zero, and are non-zero at other points of each
JðZÞ  x, Z ðZÞ 
element.
( )
  1  yðZÞ  Z 
b2 ðZÞ ¼   n ðZÞ ð24Þ 3.4. Weak formulation
9JðZÞ9  xðZÞ 

Eqs. (23) and (24) may be rewritten in the following forms: Instead of using the governing equation and associated BCs
 1  fnx ðZÞg qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi directly, strong form as in Eq. (6), one may use an integrated form
b ðZÞ ¼ y2, Z ðZÞ þ x2, Z ðZÞ ð25Þ (i.e., weak form such as weighted residual approach). This may be
9JðZÞ9
started by weighting Eq. (6) with an arbitrary test vector w and
  fnZ ðZÞg qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi integrating over the model volume O, which gives
b2 ðZÞ ¼ x2 ðZÞ þ y2 ðZÞ ð26Þ Z Z Z
9JðZÞ9
wkc, ii dO þ wQ b dO þ wQ s dO ¼ 0 ð34Þ
O O O

3.3. Shape functions

In the present method, special new shape functions [N(Z)] are 1.2
Ν1 Ν2 Ν3
proposed. To the authors’ knowledge, it is the first time that such
1
idea is raised and, as shown in next sections, diagonal coefficient
matrices will be obtained by implementing these shape functions 0.8
for potential problems. For this purpose, the potential function 0.6
N ()

and its derivative, across the element, are interpolated by poly- 0.4
nomials that have two specific properties: (a) these functions
0.2
have the property of Kronecker Delta function, and (b) their first
derivatives are equal to zero at any given control point: 0
Na ðZb Þ ¼ dab ð27Þ -0.2
-0.4
Na, Z ðZb Þ ¼ 0 ð28Þ -1 -0.5 0 0.5 1

For an (nZ + 1)-node element, these new shape functions are
expressed as a polynomial of degree (2nZ  1) that has 2nZ Fig. 4. Shape functions for a three-node element.
 N. Khaji, M.I. Khodakarami / Engineering Analysis with Boundary Elements 35 (2011) 845–854 849

Table 1
Constant coefficients ai, i¼ 0, y, 2nZ  1, for shape functions of various (nZ + 1)-node elements.

nZ + 1 2 3 4

Zi 1 1 1 0 1 1 1=2 1=2 1
Shape functions N1 N2 N1 N2 N3 N1 N2 N3 N4

Constant coefficient 1=2 1/2 0 1 0 11/54 8/27 8/27 11/54

 3/4 3/4 0 0 0  25/108  40/27 40/27 25/108
0 0 1 2 1  16/9 16/9 16/9  16/9
1=4  1/4  5/4 0 5/4 73/36 16/9  16/9  73/36
 1/2 1  1/2 40/9  40/9  40/9 40/9
3/4 0  3/4  46/9 8/9  8/9 46/9
 64/27 64/27 64/27  64/27
76/27  32/27 32/27  76/27

for a 2D domain (i.e., i¼ 1, 2). Using Eqs. (28) and (33), the first Employing Eqs. (21) and (22), Eq. (43) may be given as
term of Eq. (34) in local coordinates (i.e., i¼ x, Z) may be Z Z
expressed as wðxÞ½NðZÞT Q s dO ¼ wðxÞ½NðZÞT
O O0
Z Z j k j k1
T  T  y, Z ðZÞ x, Z ðZÞ yðZÞ xðZÞ
wkc, ii dO ¼ wðfb1 ðZÞg q , x þ 1xfb2 ðZÞg q , Z ÞdO ð35Þ B C
O O @ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi þ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi A
y2, Z ðZÞ þ x2, Z ðZÞ y2 ðZÞ þx2 ðZÞ
Using Eq. (5), the first derivative of flux {q} with respect to Z qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
can be written as fqs g y2, Z ðZÞ þ x2, Z ðZÞdx dZ ð44Þ

fqg, Z ¼ ðkfb1 ðZÞg½NðZÞfcðxÞg, x Þ, Z ð36Þ Substituting Eqs. (14) and (44) into Eq. (42) may lead to the
following expression:
Since k and {c(x)},x are not functions of Z, Eq. (36) yields Z 1 Z þ1
T
fqg, Z ¼ kðfb1 ðZÞg½NðZÞ, Z þ fb1 ðZÞg, Z ½NðZÞÞfcðxÞg, x ð37Þ wðxÞ ½NðZÞT fb1 ðZÞg ðkfb1 ðZÞg½NðZÞfcðxÞg, x Þ, x
0 1
!
As already discussed, the first derivatives of proposed shape 1 2
 1  
T   T b T s
þ xfb ðZÞg k b ðZÞ , Z NðZÞ cðxÞ , x þ½NðZÞ Q þ ½NðZÞ Q Þ
functions are equal to zero, N,Z(Z)¼0, at every control point.
Therefore, Eq. (37) may be simplified to the following expression:
 x9JðZÞ9dZ dx ¼ 0
fqg, Z ¼ kfb1 ðZÞg, Z ½NðZÞfcðxÞg, x ð38Þ ð45Þ
Now, substituting Eqs. (5) and (38) into Eq. (35) yields Eq. (45) is satisfied by setting the integrand of integrals over x
Z Z equal to zero. Therefore, using Eq. (44), the following relation may
T
wkc, ii dO ¼ w fb1 ðZÞg ðkfb1 ðZÞg½NðZÞfcðxÞg, x Þ, x be derived after some algebraic manipulations:
O O Z þ1
T      T
þ 1xfb2 ðZÞg k b1 ðZÞ , Z NðZÞ cðxÞ , x dO ð39Þ k½NðZÞT fb1 ðZÞg fb1 ðZÞg½NðZÞfcðxÞg, xx x9JðZÞ9dZ
1
Z þ1
Substituting Eq. (39) into Eq. (34), leads to þ
T
k½NðZÞT fb2 ðZÞg fb1 ðZÞg, Z ½NðZÞfcðxÞg, x 9JðZÞ9dZ
Z 1
T Z þ1
w fb1 ðZÞg ðkfb1 ðZÞg½NðZÞfcðxÞg, x Þ, x
O þ x½NðZÞT Q b 9JðZÞ9dZ
T      1 0 1
þ 1xfb2 ðZÞg k b1 ðZÞ , Z NðZÞ cðxÞ , x dO Z þ1
Z Z b y , Z ð Z Þ x , Z ð ZÞ c b yð Z Þ xð Z Þ c
þ x2 ½NðZÞT B C
@ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi þ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi A
þ wQ b dO þ wQ s dO ¼ 0 ð40Þ 1 y2, Z ðZÞ þ x2, Z ðZÞ y2 ðZÞ þ x2 ðZÞ
O O
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
In local coordinates (x, Z), the test vector w can be written as fqs g y2, Z ðZÞ þx2, Z ðZÞdZ ¼ 0 ð46Þ
wðx, ZÞ ¼ ½NðZÞwðxÞ ð41Þ Defining coefficient matrices and constant parameters as
Consequently, Eq. (40) may be rewritten as ½B ðZÞ ¼ fb1 ðZÞg½NðZÞ
1
ð47Þ
Z
T
wðxÞ½NðZÞT fb1 ðZÞg ðkfb1 ðZÞg½NðZÞfcðxÞg, x Þ, x ½B2 ðZÞ ¼ fb1 ðZÞg, Z ½NðZÞ ð48Þ
O
T     
þ 1xfb2 ðZÞg k b1 ðZÞ , Z NðZÞ cðxÞ , x dO ½B3 ðZÞ ¼ fb2 ðZÞg½NðZÞ ð49Þ
Z Z
Z þ1
þ wðxÞ½NðZÞT Q b dO þ wðxÞ½NðZÞT Q s dO ¼ 0 ð42Þ
O O ½C 1  ¼ k½B1 ðZÞT ½B1 ðZÞ9J9dZ ð50Þ
1
Substituting from Eq. (25) into the last term of Eq. (42) results
Z þ1
in the following expression:
½C 2  ¼ k½B3 ðZÞT ½B2 ðZÞ9J9dZ ð51Þ
Z Z qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1
wðxÞ½NðZÞT Q s dO ¼ wðxÞ½NðZÞT fnðZÞgT fqs gx y2, Z ðZÞ þ x2, Z ðZÞdx dZ Z
O O þ1

ð43Þ fQ b g ¼ ½NðZÞT Q b 9J9dZ ð52Þ

1
850 N. Khaji, M.I. Khodakarami / Engineering Analysis with Boundary Elements 35 (2011) 845–854

0 1 which leads to diagonal coefficient matrices. As a result, the

Z þ1 y ð Z Þ x ð ZÞ yð ZÞ xð Z Þ
 s
 B b , Z , Z c b c C system of differential Eq. (54) may be expressed as a single
Q ¼ ½NðZÞT @ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi þ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi A
1 y2, Z ðZÞ þ x2, Z ðZÞ y2 ðZÞ þ x2 ðZÞ differential equation corresponding to a given point i as
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
fqs g y2, Z ðZÞ þ x2, Z ðZÞdZ ð53Þ
xCii1 ci , xx þCii2 ci , x þ xQib þ xQis ¼ 0 ð66Þ

or
Eq. (46) may be written as the following form:
2 1
38 9 2 2
38 9
C11 0  0 >
> c1 >
> C11 0  0 >
> c1 >
>
x½C 1 fcðxÞg, xx þ ½C 2 fcðxÞg, x þ xfQ b g þ xfQ s g ¼ 0 ð54Þ 6 7>
> >
> 6 7 >
> >
>
6 0 1
C22  0 7< c2 = 6 0 2
C22  0 7< c2 =
As the coefficient matrices [C1] and [C2] and vectors {Qb} and x6
6
7 þ6 7
4 ^ ^ & ^ 75>
> ^ >>
6 ^
4 ^ & ^ 75>> ^ >>
{Qs} are independent of x, Eq. (54), as the governing equation of >
> > > >
0 0  1
Cnn :c > ; 0 0  2
Cnn :c >
> ;
n n
potential problems in the present method, is a set of ordinary , xx ,x
differential equations of radial coordinate x only. 8 b9 8 s9 8 9
>
> Q1 >
> >
> Q1 >
> 0
>
> > > > > > >>
< Qb >
= = >
< Qs >
> <0>=
3.5. Numerical integration 2 2
þx þx ¼ ð67Þ
> ^ >
>
> >
> > ^ >
>
> > >
> >^>
> >>
: Qb >
> ; > ; :0;
: Qs >
To calculate coefficient matrices [C1], [C2] and vectors {Qb} and n n
{Qs}, integrals need to be evaluated numerically over each
Note that Eq. (66) shows a system of ordinary differential
element. As control points are placed at CLL points, the
equations for a given potential problem with nDOFs. Each
Clenshaw–Curtis quadrature rule may be used, which leads to
differential equation in Eq. (66) depends only upon the potential
diagonal coefficient matrices. In 2D problems, integrations over
function of the ith DOF. In other words, the coupled system of
elements Ge may be approximated using the Clenshaw–Curtis
differential equations has been transformed into decoupled dif-
quadrature rule:
ferential equations using special mapping functions, special shape
Z Z þ1 Z þ1
functions, special numerical integration method, and special weak
f ðxðZÞ,yðZÞÞdG ¼ f ðxðZÞ,yðZÞÞJðZÞdZ ¼ gðZÞdZ ð55Þ
Ge 1 1 formulation procedure. This means that to evaluate the potential
function and its derivative at a given point, the governing
which is given by the following expression [27]:
equation that is related to the point should be solved, only.
Z b  
c2k þ 1 As may be observed in Section 4, the decoupled system of
gðZÞdZ ¼ ðbaÞ 12 c1 13 c3 15
1
c5      
a ð2k þ 1Þð2k1Þ differential equations proposed in this paper may also provide
¼ ðbaÞW ð56Þ higher rates of convergence by implementing a few numbers of
DOFs in comparison with other numerical methods.
where cj is defined as
( " For the ith DOF, the general solution of differential Eq. (66)
X
n þ1  #
2 1 pnZ k may be expressed in the following form:
cj ¼ gðZ0 Þ þ gðZnZ Þcos
nZ þ 1 2 nZ þ1 ðC 1 C 2 Þ=Cii1

k¼0
) Ai Cii1 x ii ii x2
XnZ
pðj1Þk ci ðxÞ ¼  ðQis þQib Þ þBi ð68Þ
þ gðZk Þcos ; j ¼ 1,3,5,. . .,nZ ð57Þ Cii1 Cii2 2ðCii1 þCii2 Þ
k¼1
nZ þ 1
where Ai and Bi denote the constant coefficients that are deter-
Moreover, Zk are CLL points. mined by imposing the boundary conditions at x ¼0, 1, corre-
sponding to the ith DOF.
3.6. Diagonal coefficient matrices Since Eq. (54) introduces the variation of potential function in
the domain of the problem along radial coordinate x related to the
Implementing Clenshaw–Curtis quadrature rule and special ith DOF, to calculate the potential at any point in the domain or on
shape functions results in diagonal coefficient matrices [C1] and the boundaries, except the control point, interpolating with the
[C2]. This diagonalization may be achieved by substituting the help of shape functions should be employed.
special shape functions (see Table 1) into Eqs. (50) and (51) and Obviously, the proposed method is a semi-analytical solution
then employing the result in Eq. (56). Consequently, the compo- that offers approximate numerical solutions on the boundaries,
nents of coefficient matrices [C1] and [C2] may be obtained as while provides exact analytical solutions throughout the domain
under study.
Cij1 ¼ 2dij wi ½B1 ðZi ÞT k½B1 ðZi Þ9JðZi Þ9 ð58Þ

Cij2 ¼ 2dij wi ½B3 ðZi ÞT k½B2 ðZi Þ9JðZi Þ9 ð59Þ

4. Numerical examples
where dij denotes the Kronecker Delta. According to the results
of Section 3.3, the special shape functions show the following The present new method has been implemented in a 2D code,
properties: in which new one-dimensional (1D) elements are provided. In
Ni ðZi ÞNi ðZi Þ ¼ 1 ð60Þ order to validate the nature and general behavior of the method,
four numerical examples have been considered. Numerical results
Ni ðZi ÞNj ðZi Þ ¼ 0 ð61Þ are compared with those obtained by exact analytical solutions
and/or by other numerical methods to assess the accuracy and the
Ni, Z ðZi ÞNj ðZi Þ ¼ 0 ð62Þ efficiency of the present method.

Ni, Z ðZi ÞNi ðZi Þ ¼ 0 ð63Þ

4.1. Steady-state heat conduction problem of a rectangular domain
Ni, Z ðZi ÞNi, Z ðZi Þ ¼ 0 ð64Þ
The first example is the temperature field T of a rectangular
Ni, Z ðZi ÞNj, Z ðZi Þ ¼ 0 ð65Þ plate, 0rx r5, 0ry r10 (see Fig. 5a), which is governed by 2D
 N. Khaji, M.I. Khodakarami / Engineering Analysis with Boundary Elements 35 (2011) 845–854 851

Fig. 5. The first numerical example: (a) geometry and boundary conditions in global Cartesian coordinates, (b) the LCO and proposed mesh including 5 three-node
elements, and (c) the three-node element.

Laplace equation: 80
2 2
Present Method
@ T @ T 70
r2 T ¼ þ 2 ¼0 ð69Þ Analytical solution
@x2 @y 60
with the following mixed boundary conditions: 50
T (2.5, )

Tðx,0Þ ¼ 0, 0 o x o5,
40
Tð0,yÞ ¼ 0, 0 o y o10,
px 30
Tðx,10Þ ¼ 100 sin , 0 o x o 5,
10 20
@Tð5,yÞ
¼ 0, 0 o yo 10 ð70Þ
@x 10
The analytical solution of this problem may be simply 0
obtained [1] as 0 0.2 0.4 0.6 0.8 1
ξ
100 px py
Tðx,yÞ ¼ sin sinh ð71Þ
sinhðpÞ 10 10 Fig. 6. The analytical solution and the numerical results for the temperature
distribution for 1st example along the Y-axis at X ¼2.5 (i.e., x ¼ Y/10).
In this problem, the LCO is selected as shown in Fig. 5b. In
addition, 5 three-node elements with total 11 DOFs are used may be observed from the figure, the numerical results agree very
(see Fig. 5b). To calculate the variation of potential function along well with the analytical solution, using only 11 DOFs.
the y-axis at x¼2.5, we need to solve the governing equation
related to the 6th DOF, only. As Qb ¼0 in this example, the general
4.2. Steady-state heat conduction problem of a circular plate
solution (see Eq. (68)) of the 6th DOF may be expressed as
ðC 1 C 2 Þ=C66
1  
1
A6 C66 x 66 66 x2 2 s b
In the second example, the temperature field of a circular
c6 ðxÞ ¼ 1 C 2
 1 þ C2 Þ 3
xQ 6 þ Q 6 þ B6 ð72Þ domain whose radius is 2 (see Fig. 7a) is studied. Laplace equation
C66 66 2ðC66 66
r2T ¼0 is considered with the following Dirichlet boundary
In this example, the potential function on the upper boundary conditions:
of the domain (i.e., between nodes 5 and 7) is a function of
Tð2, yÞ ¼ 4 sinðyÞcosðyÞ ð73Þ
x-coordinate. Therefore, Qis corresponding to variation of potential
function on boundaries is calculated using Eq. (53). Constant The analytical solution of this example for y ¼ p=4 is
parameters A6 and B6 of Eq. (72) may be simply determined by
Tðr,ðp=4ÞÞ ¼ 0:5r 2 ð74Þ
imposing the boundary conditions at the LCO and the 6th DOF.
The analytical solution and the numerical results using the The LCO (x ¼yE  1.41) is shown in Fig. 7b. Moreover, 4 three-
proposed new method along the y-axis are shown in Fig. 6. As node elements with total 7 DOFs are employed (see Fig. 7b).
852 N. Khaji, M.I. Khodakarami / Engineering Analysis with Boundary Elements 35 (2011) 845–854

In order to calculate potential function distribution along y ¼ p=4

direction, the governing equation of the 4th DOF should be solved.
As the body potential Qb is equal to zero in this example, the

2.5

2 Present Method
Analytical solution
1.5

T ()
1

0.5

0
0 0.2 0.4 0.6 0.8 1


Fig. 8. The analytical solution and the numerical results for the 2nd example from
the present method for the variation of temperature along y ¼ p=4 direction.

0
θ)= ∇2u − 4 = 0

u
(2
(2,

,θ
u

)=
0
0

u
)=

(1
,θ
,θ

)=
(1

r
u

0 x
θ 0
)=
u

,θ
(1

(1
,θ

u
)=
0

0
)=
u
(2

,θ
,θ

(2
)=

η
0

y
η
0)
=
(q
q
Γ

Γu

ξ
Γu

Fig. 7. The second numerical example: (a) geometry and boundary conditions in x
global Cartesian coordinates, (b) the LCO and proposed mesh that consists of
4 three-node elements, and (c) the three-node element. Γq (q= 0)

Fig. 9. The third numerical example: (a) geometry and boundary conditions in
global Cartesian coordinates, (b) the LCO and proposed mesh for one-sixth of the
domain including 4 three-node elements, and (c) the three-node element.
 N. Khaji, M.I. Khodakarami / Engineering Analysis with Boundary Elements 35 (2011) 845–854 853

general solution of the 4th DOF may be given as 0r y r2p). The boundary value problem is governed by the
1 C 2 Þ=C 1
ðC44 2
Poisson equation:
1
A4 C44 x 44 44
x
T4 ðxÞ ¼ 1 C 2
 1 þC 2 Þ
ðQ4s Þ þ B4 ð75Þ @2 u @2 u
C44 44 2ðC44 44 r2 u ¼ þ 4 ¼ 0 ð76Þ
@X 2 @Y 2
Following the same procedure as the first example, the
with the following boundary conditions:
analytical solution and the numerical results using the proposed
new method along y ¼ p=4 direction are depicted Fig. 8. As it is uð1, yÞ ¼ 0, uð2, yÞ ¼ 0 ð77Þ
evident from Fig. 8, the numerical results are identical to the As already discussed, the LCO should be selected in a location
analytical solution, using only 7 DOFs. from which all boundaries of the domain are visible. In this
problem, a point with the mentioned property is not available. As
4.3. Poisson equation—heat transfer in a torus this problem is symmetric, only one-sixth of the domain is taken
into account in numerical computations (Fig. 9b). For the selected
The third example is a torus (Fig. 9a) with prescribed tem- part of the domain, the LCO is located at X¼2 and Y ¼0. There are
peratures u(r, y) on the inner and outer boundaries (1rr r2, 5 three-node elements (with 8 DOFs) for discretization along the
boundaries.
0 The temperature field of this example for y ¼0 has a radial
distribution as the following analytical solution:
-0.1 uðr,0Þ ¼ 10 logr þ r 2 1 ð78Þ
Present Method
-0.2 Analytical solution To evaluate the variation of potential function along the
horizontal axis with y ¼0, it is sufficient to solve Eq. (66) for node
u ()

-0.3

-0.4

-0.5

-0.6
0 0.2 0.4 0.6 0.8 1


Fig. 10. The analytical solution and the numerical results for the potential
distribution for the 3rd example along the r-direction at y ¼ 0.

Fig. 11. The fourth numerical example: (a) geometry and boundary conditions in Fig. 12. The solution contours of the 4th example: (a) the present method using
global Cartesian coordinates and (b) the LCO and proposed mesh that consists of 4 three-node elements and 7 CLL points, (b) the present method using 12 three-node
4 three-node elements. elements and 22 CLL points, and (c) analytical solution results.
854 N. Khaji, M.I. Khodakarami / Engineering Analysis with Boundary Elements 35 (2011) 845–854
8 only. The solution results of this example with the present
method are shown in Fig. 10. Moreover, the analytical solution is
also presented in this figure. It is evident that the numerical
results are almost indistinguishable from the analytical solution.
4.4. Elliptic domain with Dirichlet boundary conditions
As the last example, we consider an elliptic domain with
governing Laplace equation in the problem domain (see Fig. 11a):
@2 j @2 j
r2 j ¼ þ ¼0 ð79Þ
@X 2 @Y 2
and Dirichlet boundary conditions as
j ¼ cosðyÞ þ 0:5 sinðyÞ: ð80Þ
The exact solution of this problem value problem may be
simply derived as j (X,Y)¼ X+ Y. In this example, the LCO is
located at X ¼  2 and Y ¼0. The problem has been solved using
two different meshes with 4 and 12 three-node elements. The
first mesh that consists of 4 three-node elements with 7 DOFs is
shown in Fig. 11b. The solution result of this example using the
new present method is depicted in Fig. 12a and b, for the first and
second meshes, respectively. It is obvious that the results of the
present method show excellent agreement with the analytical
solution result illustrated in Fig. 12c.
5. Conclusions
In this paper, a novel semi-analytical method with its detailed
formulation for modeling 2D potential problems was presented
for the first time. In the proposed method, the boundary of the
domain was discretized by new non-isoparametric elements with
higher-order Chebyshev polynomials as mapping functions.
Moreover, these new elements use new special shape functions
that were formulated in order to provide Kronecker Delta prop-
erty for the potential function and its derivative. In addition, the
first derivative of shape functions was assigned to zero at any
given control point. Lastly, using Clenshaw–Curtis quadrature, the
coefficient matrices of equations system became diagonal, which
resulted in a set of decoupled governing equations for the whole
system. Steady-state analyses of four examples were successfully
carried out using the new method. In these examples, various
potential problems and boundary conditions were selected to
illustrate the applicability and generality of the method. One may
note that all these examples were successfully modeled with very
small number of DOFs, preserving very high accuracy compared
with available analytical solutions.
Further development of the proposed method for the analysis
of three-dimensional problems is left as a future research topic,
which is currently being followed by the authors.
References
[1] Kreyszig E. Advanced engineering mathematics. John Wiley & Sons, Inc; 2006.
[2] Zienkiewicz OC, Taylor RL. The finite element method. Oxford: Butterworth
and Heinmann; 2000.
[3] Brebbia CA, Telles J, Wrobel L. Boundary element techniques—theory and
applications in engineering. Berlin and New York: Springer Verlag; 1984.
[4] Friedrich J. A linear analytical boundary element method (BEM) for 2D
homogeneous potential problems. Computers & Geosciences 2002;28:679–92.
[5] Zhou HL, Niu ZR, Wang XX. Regularization of nearly singular integrals in the
BEM of potential problems. Applied Mathematics and Mechanics 2003;24:
1208–14.
[6] Niu ZR, Zhou H. The natural boundary integral equation in potential problems
and regularization of the hypersingular integral. Computers and Structures
2004;82:315–23.
[7] Zhou HL, Niu ZR, Wang XX. Regularization of nearly singular integrals in the
boundary element method for 3-D potential problems. Chinese Journal of
Computational Physics 2005;22(6):501–6.
[8] Zhou H, Zh Niu, Ch Cheng, Zh Guan. Analytical integral algorithm in the BEM
for orthotropic potential problems of thin bodies. Engineering Analysis with
Boundary Elements 2007;31:739–48.
[9] Zhou H, Zh Niu, Ch Cheng, Zh Guan. Analytical integral algorithm applied to
boundary layer effect and thin body effect in BEM for anisotropic potential
problems. Computers and Structures 2008;86:1656–71.
[10] Zhang X, An X. Exact integration and its application in adaptive boundary
element analysis of two-dimensional potential problems. Communications in
Numerical Methods in Engineering 2008;24:1239–50.
[11] Shen L, Liu YJ. An adaptive fast multipole boundary element method for
three-dimensional potential problems. Computational Mechanics 2007;39:
681–91.
[12] Wolf JP. The scaled boundary finite element method. John Wiley & Sons Ltd;
2004.
[13] Deeks AJ, Cheng L. Potential flow around obstacles using the scaled boundary
finite-element method. International Journal for Numerical Methods in Fluids
2003;41:721–41.
[14] Teng B, Zhao M, He GH. Scaled boundary finite element analysis of the water
sloshing in 2D containers. International Journal for Numerical Methods in
Fluids 2006;52:659–78.
[15] Tao L, Song H, Chakrabarti S. Scaled boundary FEM solution of short-crested
wave diffraction by a vertical cylinder. Computer Methods in Applied
Mechanics and Engineering 2007;197:232–42.
[16] Zhang Z, Zhao P, Liew KM. Improved element-free Galerkin method for two-
dimensional potential problems. Engineering Analysis with Boundary
Elements 2009;33:547–54.
[17] Ren HP, Cheng YM, Zhang W. An improved boundary element-free method
(IBEFM) for two-dimensional potential problems. Chinese Physics B 2009;
18(10):4065–73.
[18] Peng M, Cheng Y. A boundary element-free method (BEFM) for two-dimen-
sional potential problems. Engineering Analysis with Boundary Elements
2009;33:77–82.
[19] Mukherjee YX, Mukherjee S. The boundary node method for potential
problems. International Journal for Numerical Methods in Engineering 1997;
40:797–815.
[20] Kulkarni SS, Telukunta S, Mukherjee S. Application of an accelerated bound-
ary-based mesh-free method to two-dimensional problems in potential
theory. Computational Mechanics 2003;32:240–9.
[21] Zhang JM, Tanaka M, Matsumoto T. Meshless analysis of potential problems
in three dimensions with the hybrid boundary node method. International
Journal for Numerical Methods in Engineering 2004;59:1147–60.
[22] Tan F, Wang YH. Multiple reciprocity hybrid boundary node method for
potential problems. Engineering Analysis with Boundary Elements 2010;34:
369–76.
[23] Zhang J, Qin X, Han X, Li G. A boundary face method for potential problems in
three dimensions. International Journal for Numerical Methods in Engineer-
ing 2009;80:320–37.
[24] Ma H, Qin QH. Solving potential problems by a boundary-type meshless
method—the boundary point method based on BIE. Engineering Analysis
with Boundary Elements 2007;31:749–61.
[25] Guo XF, Chen HB. Dual error indicators for the local boundary integral
equation method in 2D potential problems. Engineering Analysis with
Boundary Elements 2006;30:702–8.
[26] Chen T, Raju IS. A coupled finite element and meshless local Petrov–Galerkin
method for two-dimensional potential problems. Computer Methods in
Applied Mechanics and Engineering 2003;192:4533–50.
[27] Clenshaw CW, Curtis A. A method for numerical integration on an automatic
computer. Numerische Mathematik 1960;2:197–205.
[28] Lebedev NN. Special functions and their applications. Prentice-Hall; 1965.