Академический Документы
Профессиональный Документы
Культура Документы
1.1 Introduction
The Boundary Element Method (BEM) has become an important tool for solving
problems in applied science and engineering. Its origins are quite recent compared
with those of the Finite Difference Method (FDM) and the Finite Element Method
(FEM). The BEM had its beginnings in the early 1960s with the research work of
Massonet (1956), Rizzo (1967) and Cruse (1969) in elasticity, Jaswon (1963) and
Symm (1963) in potential theory, Hess and Smith (1964) in aerodynamics, Shaw
(1970) in acoustics and Rizzo and Shippy (1970) in heat conduction, in which the
underlying physical processes are expressed as integral equations posed on the
boundary of the region of interest. The term 'boundary element method' was first used
by Brebbia and Dominguez (1997) who realised the analogy between the
discretisation process for the boundary integral method and that for the already
established finite element method.
Presently, the BEM is a very active field of study especially within the engineering
community, and it is experiencing very rapid development in research and
applications worldwide. The following are some points of view about the features and
the capability of BEM from researchers. One of the most interesting features of the
BEM is the much smaller system of equations and the considerable reduction of the
data required to run a program. Moreover, the numerical accuracy of the method is
generally greater than that of FEM (Brebbia, 1978). The BEM has emerged as a
powerful alternative to the FEM, particularly in case where better accuracy is required
due to difficulties such as stress concentration or where the domain extends to infinity
(Ciskowski and Brebbia, 1991). The BEM is well established by now as an accurate
and powerful numerical technique in continuum mechanics (Becker, 1992). The
most important feature of the BEM is that it requires only the discretisation of the
boundary rather than the domain. It is now gaining very considerable popularity and
being incorporated into high-speed computer algorithms immediately useful to the
practicing analyst (Banerjee, 1994). The BEM occupies a less prominent role in
engineering than its features and capability due to the complexity of the mathematical
background (Paris and Canas, 1997). More introductory discussion on this subject can
be founded in Brebbia and Dominguez (1989).
This chapter is an attempt to give the mathematical foundations of the method for the
reader who is approaching the BEM as a new user. It contains the preliminary
mathematics and the formulation of the boundary element method for the Laplace
equation. We discuss the Poisson equation in chapter 2. At the end of the section, we
mention the LINBEM code implemented using the method by Mushtaq (1995). The
code is based on the formulation in Brebbia (1978). Finally we show an example of
Laplace's problem executed by the LINBEM program.
1
1.2 Preliminary mathematics
f (x ) f (x ) f (t )
f (t )
transform
a b x -1 1 t
ba ba
x( )( )t (1.2.1)
2 2
Hence
(b a ) 1
b 1
I f ( x)dx f (b a ) (b a )t dt (1.2.2)
a
2 1 2
f (t )dt wi f (t i )
1
(1.2.3)
1
i 1
2
t i are the zeros of the Legendre polynomial
1 dn 2
Pn (t ) n n
(t 1) n (1.2.4)
2 n! dt
2(1 t i2 )
wi . (1.2.5)
n 2 Pn 1 (t i )
2
The proof of these formulae can be found in Buchanan and Turner (1992). The proof
confirms that the Gauss n-point integration will compute integrals of polynomials of
degree no more than 2n-1 exactly. Programming formula (1.2.3) is straightforward
once the points, t i , and the weights, wi , are known. Finally we obtain the
approximation to the integral (1.2.2). The first four points and Gaussian weights are
shown in Table 1.2.1.
2 0 (1.2.6)
on the domain .
2u 0 in (1.2.7)
3
u u on 1
and the Neumann condition
u
qq on 2
n
n
1 2
u u 2u 0 u
qq
n
Figure 1.2.2 Laplace's equation posed on the domain with the boundary
The combination of the partial differential equation with these boundary conditions is
called a potential problem. It has become traditional in the boundary element
u
literature to write the flux term, , as q. Problems governed by the Laplace equation
n
appear in different fields of applied science and engineering such as fluid flow,
electric and magnetic potential fields, stress and strain analysis of elastic solids,
groundwater transport, etc.
Let be a regular domain with boundary as shown in Figure 1.2 and let F be a
vector function continuously differentiable at every point of . The divergence or the
Gauss Theorem states
.FdA F.nds
(1.2.8)
4
Suppose u and v are two scalar functions defined in bounded by the closed curve
as shown in Figure 1.2.2. Assume that they are continuous and admit continuous
partial derivatives. From the basic property of the gradient, we have
Applying the Gauss Theorem (1.2.8) to the left hand side then gives
Using the property of the gradient on the right hand side gives
v u
uv vu dA u n v n ds
(1.2.13)
v u
u v v u dA u n v n ds
2 2
(1.2.14)
f ( x) ( x x0 ) f ( x0 ) (1.2.15)
for every continuous function f.
5
A complete discussion and more properties can be found in Humi and Miller (1992).
Suppose that r is the position vector of a point, Q, relative to the point, P, inside .
Surround P by a small disc, center P radius , as shown in Figure 1.2.3. The point
P will be called the source point and the point Q will be called the field point.
Suppose that r0 is the position vector of P and r1 is the position vector of Q.
qq
r1 1
Q
r
r0
u u
n P
2
Let r r1 r0 with r r .
2 u * (r ) (r ) 0 (1.2.16)
We see that
2 u * (r ) 0 (1.2.17)
To find the fundamental solution we first find the solution of the homogeneous
equation (1.2.17). In cylindrical co-ordinates,
u 2 1 u * 1 2 u *
2 u * (r ) (1.2.18)
r 2 r r r 2 2
6
d 2 u * 1 du *
0 (1.2.19)
dr 2 r dr
This ordinary differential equation can be solved analytically and then we get
u * C ln r D (1.2.20)
Let be the disc bounded by the curve with center P and radius as shown in
Figure 1.3. Integrating on of equation (1.2.16) we obtain
u (r )dA (r )dA
2 *
(1.2.21)
(r )dA 1 (1.2.22)
u (r )dA 1
2 *
Hence (1.2.23)
Applying the divergence theorem on the left hand side of (1.2.21) we get
u *
u (r )dA u . n ds n ds
2 * *
(1.2.24)
u * C
(1.2.25)
n r
u * 2 C
n ds 0 r rd 2C (1.2.26)
1
C (1.2.27)
2
7
1
u * (r ) ln r (1.2.28)
2
1
u * (r ) (1.2.29)
4r
More details of both fundamental solutions can be seen in Gilbert and Howard
(1990)or Paris and Cañas (1997).
Historically, there have been several ways to obtain the integral representation of the
potential at an internal point. We follow the approach of Davies and Crann (1996).
From Figure 1.2.3, applying the second form of Green’s theorem (1.2.14) with
v u * in the region we have
u * u
(u u u u )dA u * )ds
2 * * 2
(u (1.2.30)
n n
Since, from (1.2.7) and (1.2.17), both u and u* are harmonic in the integral
on the left hand side is zero, therefore
u * u u * u
(u
n
u * )ds (u
n
n
u * )ds
n
(1.2.31)
u
Substituting the fundamental solution (1.2.28) in (1.2.31) and the fact that q
n
and then taking the limit as 0 we have
1 (ln r ) 1 1
2 (u
n
q ln r )ds lim u
0
n
(
2
ln r )ds lim (
0
2
q ln r )ds
u u P and q q P
Hence
1 (ln r ) 1 1
2
(u
n
q ln r )ds u P lim (
0
2
)ds q P lim (
0
2
ln )ds (1.2.32)
Computing the integral on the right hand side and rearranging the equation we obtain
8
1 (ln r )
uP
2
(u
n
q ln r )ds (1.2.33)
1 1
or uP
2
( ln r )q u ( 2 r . nds
r
(1.2.34)
u*
i.e. u P (u q uq )ds , where q
* * *
(1.2.35)
n
Equation (1.2.35) is known as the basic formula for the internal solution. It is obtained
whenever the value of u and q are known everywhere on the boundary. However, for
properly-posed problems we can specify only one of u and q at each boundary point
(Weinberger, 1965) and consequently equation (1.2.35) is of little help as it stands. In
section 1.2.8 we show how to develop a boundary integral equation from which we
are able to obtain u and q everywhere on the boundary and hence, via (1.2.35),
everywhere inside the domain.
Suppose that P is a point on the boundary at which there is a kink with angle as
shown in Figure 1.2.4 ( a ). If the boundary is smooth at P then = .
P P
2 1
n
*
(a) (b)
From Figure 1.2.4 ( b ), applying the second form of Green's theorem (1.2.14) with
v u * to the region , where * is the difference between and , then
u * u u * u
(u u u u )dA u* )ds (u u*
2 * * 2
(u )ds (1.2.36)
n n
* n n
9
u * u u * u
(u
n
u * )ds (u
n n
u * )ds
n
(1.2.37)
*
u * * u u * u
(u n u n )ds lim
0 (u
n
)ds lim (u * )ds
0 n
(1.2.38)
*
*
Therefore
1 (ln r ) 2
1 2
1
2
(u
n
q ln r )ds u P lim
0
1 2
d q P lim ( ln )d
0
1 2
Computing the integral on the right hand side and using the fact that 2 1 ,
we obtain
(ln r )
u P (u q ln r )ds (1.2.39)
n
1
or u P ( ln r )q u ( 2 r.n)ds (1.2.40)
r
i.e. cu P (u * q uq * )ds , where c (1.2.41)
2
Equation (1.2.41) is known as the basic formula for the boundary element method. It
is clear that if P is an interior point then 2 and hence c 1 . If P is a
1
boundary point and the curve is smooth, then and c .
2
cu uq * u * qds (1.3.1)
N N
Target element
j+1 [j] j
…
N 10 …
… N
We choose a suitable set of basis function w j (s) : j 1,2,..., N where s is the
distance around the boundary, N , and consider the boundary element approximation
N
u~( s) w j ( s)U j (1.3.2)
j 1
N
~(s)
q w j (s)Q j
j 1
(1.3.3)
N N
ci u i w j (s)U j q ds w
*
j ( s)Q j u * ds (1.3.4)
N j 1 N j 1
N N
ci u i Hˆ ijU j Gij Q j (1.3.5)
j 1 j 1
where
Hˆ ij w (s)q ds
j
*
(1.3.6)
N
Gij w
N
j ( s)u * ds (1.3.7)
We shall consider linear elements in which w j (s) is the usual ‘hat’ function based at
node j, as shown in Figure 13.2.
11
w j (s)
=1
Node j+1
= -1
Node j
l
l
Global co-ordinate s s j
2
Figure 1.3.3 Relation between the global co-ordinate and the
2
dimensionless local co-ordinate, ( s s j )
l
With the basis function w j (s) being the hat function shown in Figure 1.3.3, we see
that the values of u and q at any point of an element can be defined in terms of their
nodal values and the linear interpolation functions 1 and 2 , often called shape
function such that
U
u ( ) 1U j 2U j 1 1 2 j (1.3.8)
U j 1
Q
q ( ) 1Q j 2 Q j 1 1 2 j (1.3.9)
Q j 1
12
where is the dimensionless co-ordinate and 1 and 2 are the usual Lagrange
interpolation polynomials
1
1 (1 ) (1.3.10)
2
1
2 (1 ) (1.3.11)
2
The integral along the element [ j] in equation (1.3.4) becomes, for the left hand side,
U 1 N U
N
jw q *
ds
1 2 q *
ds U h j1 h j2 1 (1.3.12)
N j 1 [ j ] 2 j 1 U 2
h j1 1q * ds (1.3.13)
[ j]
h j 2 2 q * ds (1.3.14)
[ j]
Qj N Qj
N
jw u *
ds
1 2 u *
ds Q g j1 g j2 (1.3.15)
N j 1 [ j ] j 1 j 1 Q j 1
g j1 1u * ds (1.3.16)
[ j]
g j 2 2 u * ds (1.3.17)
[ j]
Hˆ ij h j1 h( j 1) 2 (1.3.18)
13
N N
ciU i Hˆ ijU j Gij Q j (1.3.20)
j 1 j 1
In equation (1.3.20) j denotes a node at which the basis function is defined. Such a
node is called a functional node. Also in equation (1.3.20) i denotes a node at which
the equations are collocated. Such a node is called a collocation node. Also nodes
define the shape of the boundary and in this context they are called geometric nodes.
In this case the functional and collocation nodes are the same. We shall see later that
there are circumstances in which it is helpful for them to be different.
Let us define
H ij Hˆ ij where i j (1.3.21)
H ij Hˆ ij ci where i j (1.3.22)
N N
H ijU j Gij Q j
j 1 j 1
(1.3.23)
HU GQ (1.3.24)
Applying the boundary conditions to identify the Dirichlet and Neumann boundary
regions we can partition the system (1.3.24) in the form
H 1 U Q
H 2 1 G 1 G 2 1 (1.3.25)
U 2 Q 2
Ax y (1.3.26)
Q 1
Where A G1 H 2 , x , and y [G 2 Q 2 H1 U1 ]
U 2
Once the values of U2 and Q1 on the whole boundary are known we can calculate the
value of u at any internal point Pk by using equation (1.2.35) and the boundary
discretisation to give u k in the form
14
N N
u k Gkj Q j H kjU j (1.3.27)
j 1 j 1
We note that this formulation uses linear elements and is traditionally called the
standard collocation method. Other elements such as constant elements and parabolic
elements are explained in París and Cañas (1997). We shall introduce a program
implementing the standard collocation method in section 1.4 and then discuss some
computational results.
The LINBEM code (Mushtaq, 1995) is implemented using the standard collocation
method as described in the section 1.3. It is based on the code in Brebbia (1978). It
works well with smooth domain boundary problems; but, in problems with domains
containing corners, the computational results are not good enough. We illustrate this
in the rest of this section and give an alternative method to overcome this problem in
the next chapter.
2u 0 (1.4.1)
y
(0,6) q=0 (6,6)
10 7
u = 300 u=0
1 4
(0,0) q=0 (6,0) x
We partition the boundary into 12 elements with 5 internal nodes. The internal
solution and the percentage error are shown in Table 1.4.1.
15
Table 1.4.1 The internal solution and percentage error
We see from Table 1.4.1 that the internal solution from LINBEM is quite good.
However the normal derivative at the corners as shown in Figure 1.4.2 needs to be
improved.
60
Normal derivative
40
20
-20
-40
-60
1 2 3 4 5 6 7 8 9 10 11 12 1
Exact LINBEM
We observe from equation (1.3.24) that the method uses one-to-one mapping between
boundary nodes and normal derivative variables. This gives rise to an ambiguity in the
u
normal derivative at the corners e.g. at (0,6) we have two normal derivatives, on
y
u
y 1 and on x 0 . In chapter 2 we present an approach to overcome the
x
problems caused by corners and discontinuous boundary conditions.
16
2. The multiple node method
At the beginning of this chapter we introduce a problem caused by corners and /or
discontinuous boundary conditions. The next section presents a multiple node method
to overcome the problem. After that we implement a program using the method.
Various problems are introduced to test the program. We also present another
approach of the method to resolve such problems.
2.1 Introduction
In the traditional BEM, as explained in the previous chapter, the boundary of the
domain is discretised into boundary elements. The main process is to construct a
system of algebraic equations to solve for values of function and normal derivative at
the boundary points and then use these values to compute function values at internal
points in the domain.
The boundary of the domain often contains corners and points with discontinuous
boundary conditions, which leads to certain difficulties. The difficulty for problems
whose domain contains corners is caused by the ambiguity in the normal derivative at
a corner. The difficulty for problems with discontinuous boundary conditions is
caused by the fact that the shape functions maintain the continuity of the boundary
approximations. This has been the subject of considerable research over the past
decade and several methods have been developed to overcome these difficulties.
These are divided in two categories namely, non-conforming element methods
(Manolis and Benerjee, 1986) and multiple node methods (Mitra and Ingber, 1993).
The non-conforming methods include the use of discontinuous elements and semi-
discontinuous elements. In these methods, nodes are moved away from corners so that
fluxes need not be evaluated at corners. The multiple node method introduces
additional unknowns into the formulation, and hence additional equations are required
to close the system of equations.
The relative merit of the two methods is discussed by Brebbia and Niku (1988). A
comparison between the two methods has been made (Subia, Ingber and Mitra, 1995).
However, in this work we shall focus on the multiple node method only, because of
its accuracy and also the ease of programming. There are also several kinds of
multiple node method. This work considers the multiple node method with auxiliary
boundary collocation approaches (Subia and Ingber, 1993) and the gradient approach
(Alarcón, et al, 1979, Paris and Cañas, 1997)
This chapter gives the mathematical background of the multiple node method. It is
divided into four main sections. In the first section we discuss the algorithm of the
multiple node method. In the second section we explain how to implement the
MULBEM code using the multiple node method. In the third section we discuss
computational results from the MULBEM program compared with those from the
LINBEM program and known exact solutions. Furthermore, we discuss the efficiency
of the program on several classic test problems compared with the BETIS program
(Paris and Cañas, 1997) which uses specially constructed singular elements to take
account of discontinuities in geometry or in the unknowns. In particular the well-
17
known Motz problem (Motz, 1946) is introduced as a test for the MULBEM program.
Finally, in the last section of the chapter we present the gradient approach (Paris,
Martin and Alarcon, 1981) which is a variation of the multiple node method.
In the previous chapter we summarised the mathematical background needed for the
standard collocation form of the BEM using linear continuous elements. The system
of equations is expressed in matrix form
HU GQ (2.2.1)
In this equation it is assumed that all elements of Q are single-valued. However, this
condition is violated at corners and points with discontinuous boundary conditions as
shown in Figure 2.2.1.
X Z X [1] Y [2] Z
q q1 (c) q q2
(a)
u
At corners, since, q where n is the outward normal vector to the element, q
n
is not single-valued. The same is true at points with discontinuous boundary
conditions. One method to resolve this ambiguity is to put multiple nodes at these
points. In the case of M multiple nodes, the dimension of the vectors U and Q
becomes ( N M ) 1. Consequently, we need to add extra collocation points to obtain
a square system of linear equations. Before discussing how to allocate such points we
detail the main structure of the algorithm.
Traditionally, Laplace problems are classified into three categories depending on the
boundary conditions. A Dirichlet condition is that we know the function value, a
18
Neumann condition is that we know the normal derivative, a Robin condition is that
we know a linear combination of the two.
All the features of the BEM are illustrated using only the Dirichlet and Neumann
conditions. The Robin condition is easily incorporated and since it adds nothing to the
understanding of the BEM we shall not consider such problems. For details see Paris
and Cañas (1997). We shall consider the mixed problem as shown in Figure 1 in
which we have a Dirichlet condition on a part 1, of the boundary, together with a
Neumann condition on the remainder, 2.
First of all, we define the technical terms of two elements which have a multiple node
in common as shown in Figure 2.2.2. We note that we use a counter-clockwise
integration process. This integration direction leads to the terminology for the
elements associated with node Y: the previous element and the adjacent element, see
Figure 2.2.2.
[previous] Y [adjacent]
Y1 Y2
To implement the program we identify the multiple node by means of the boundary
conditions in the two elements it belongs to in the following way:
We are now ready to continue discussing the multiple node approach. Suppose that
we partition the boundary into N elements with the necessity of M multiple nodes.
The elegance of the method lies in the fact that the additional equations are obtained
from the framework of collocation in the boundary integral equation method and
without resorting to other laws, theorems, differentiation or finite differencing
(Manolis and Benerjee, 1986, Paris and Cañas, 1997). We emphasise that four
quantities, namely, two potentials and two fluxes, are associated with each multiple
node. Among these four quantities, two are given as boundary conditions and two are
unknown.
19
Consider the situations in Figure 2.2.1(a), in which a corner is shown, and in Figures
2.2.1(b), (c), in which discontinuous boundary conditions are shown. In all these
situations, an ambiguity in q may exist at the geometric node Y. In order to resolve
the ambiguity, two functional nodes, Y1 and Y2 , are placed at the geometric node Y.
After the addition of the multiple nodes we rearrange the geometric nodes into
functional nodes and collocation nodes. Collocation at N nodes will not provide a
sufficient number of equations for the solution. To derive a sufficient number of
equations we proceed as follows:
Code 1 Referring to Figure 2.2.1(a), we place the first multiple collocation node in
the previous element. We cannot place the second one at the same position as the
functional node because collocation equation (1.3.20) at this point will be linearly
dependent on the previous one so that the solution of the system of equations is not
unique. To resolve this problem, we collocate the equation at any boundary point on
the adjacent element and the value of u, as given by the Dirichlet condition, can be
inserted in the left hand side of equation (1.3.20).
Code 2 Referring to Figure 2.2.1(b), we place the first multiple node in the previous
element as usual. Furthermore, we can collocate the second one at the same functional
node because no linear dependence occurs. Consider the multiple functional nodes
that are numbered i and j, where the conditions are Dirichlet and Neumann
respectively. Then from equation (1.3.20), the diagonal element on the ith row of the
coefficient matrix is ci, and the diagonal element on the jth row of the coefficient
matrix is cj. The elements in the jth column of the ith row and on the ith column
of the jth row are zero. This different placement of the zeros makes the two
equations linearly independent. The rest of this scheme uses the continuity of u to
close the system of equations by the equation
Code 3 This scheme is similar to Code 2 , the difference being the closing of system
of equations of with the equation
Code 4 This scheme is similar to the previous two schemes, except that we close the
system with the equation
However, there is a major difference between this scheme and the others in the case of
the Neumann problem i.e. the problem which has only Neumann conditions. The
solution of the Neumann problem is unique only up to an arbitrary constant.
Consequently, the numerical solution of such problems is likely to exhibit this
attribute
20
2.3 The MULBEM program
The LINBEM code is modified in order to handle problems caused by corners and
discontinuous boundary conditions by applying the multiple node method. We call the
new code MULBEM. The program is implemented using the algorithm as described
in the previous section.The effect of the position of the collocation points is also
investigated in the example. The influence of the number of Gauss points is
examined. Further test problems are discussed in the next section 2.4.
In this problem we return to investigate Example 1.4.1 in chapter 1. We also use the
situation and the boundary conditions as shown in Figure 1.4.1.
Initially the boundary is modelled with 12 boundary elements and 4 multiple nodes at
the corners. The calculation of the potential at five internal points is required at the
points shown in Figure 2.3.1.
9 8 12 11 10 9
10 7 13 8
11 6 14 7
12 5 15 6
16 5
1 2 3 4 1 2 3 4
The normal derivatives at the corners obtained using MULBEM are compared with
those of the LINBEM code and the exact solution in Figure 2.3.3
21
u 300-50x (0,6) q0 (6,6)
(0,6) (6,6)
60
Normal derivative
40
20
-20
-40
-60
1 2 3 4 5 6 7 8 9 10 11 12 1
We see from Figure 2.3.3 that the errors in normal derivatives at the corners are
substantially reduced using a multiple node approach.
The internal solutions obtained using MULBEM program are compared with those
from LINBEM and the exact solution in Table 2.3.1.
Table 2.3.1 Internal values for the mixed boundary condition problem
Internal 1 2 3 4 5
point (2,2) (2,4) (3,3) (4,2) (4,4)
22
Not only is the error in the normal derivative substantially reduced, we can see from
Table 2.3.1 that the internal solutions using the MULBEM are also better than those
of using the LINBEM.
This example we consider the problem as in example 2.3.1 in the case of the Dirichlet
problem. The Dirichlet boundary condition is shown in Figure 2.3.4.
u 300 - 50x
u 300 u0
u 300 - 50x
Internal 1 2 3 4 5
point (2,2) (2,4) (3,3) (4,2) (4,4)
23
As in the previous case, the error in normal derivative is substantially reduced using
MULBEM as shown in Figure 2.3.5.
60
40
Normal derivative
20
-20
-40
-60
1 2 3 4 5 6 7 8 9 10 11 12 1
We have shown that, for all cases of the boundary conditions, the MULBEM
approach works better than does the LINBEM. We also use this Dirichlet problem to
investigate the effect of the position of the collocation point. Since we move the
collocation point off the functional node only at the multiple node with the Dirichlet
condition, we examine the effect in the Dirichlet problem. We concentrated on the
position close to the multiple node and found that the nearer to the multiple node the
more accuracy was obtained. The calculated values at internal points of the Dirichlet
problem using percent of collocation as 0.10, 0.50, 1.00 and 5.00 is demonstrated in
Table 2.3.3.
24
We note that if the percent of collocation less than 0.10 there is an error caused by
linearly dependent of the system of equations as mentioned in section 2.2.
We conclude this example by examining the effect of the number of Gauss points in
the numerical quadrature. We see as the number of Gauss points increases so does the
accuracy of the solutions as shown in Figure 2.3.6.
0.8
Error(%)
0.6
0.4
0.2
0
3 4 5 6 7 8 9 10
MULBEM
We observe from Figure 2.3.6 that the Gauss four-point seems to be taken as the
standard. We can't really see how well because errors in 3 point quadrature are
relatively poor. There is only a small change in the error as we increase the number of
Gauss points above 4. However, the computational effort increases significantly.
Consequently we shall use four-point quadrature in our BEM implementation.
In this section, we have shown that the multiple node method can reduce substantially
the errors caused by corners. In section 2.4 we consider a further set of test problems.
This section describes the solution of several test problems by the multiple node
method. All problems involve corners and/or discontinuous boundary conditions. The
first example is a Dirichlet problem in the unit square. In this problem we compare the
analytic solution using separation of variables with the numerical results from the
program. The second example concerns a circular region to examine the boundary
solutions on of the boundary. We compare the results with the standard BEM solution
using linear elements. The third example is the Motz problem (Motz, 1946) which
has become a standard problem to test a new BEM approach. We compare the
computational result with the BETIS program (Paris and Cañas, 1997) and other
researchers (Symm, 1973, Whiteman and Papamichael, 1972, Lefeber, 1989). The
final example is defined on an L-shaped region with a re-entrant corner.
25
Example 2.4.1 The Dirichlet problem
Consider the Dirichlet problem in the unit square with boundary conditions as shown
in Figure 2.4.1.
y
u = 1 (1,1)
(0,1)
u = 0 u = 0
x
(0,0) (1,0)
u = 0
2
(1 (1) n ) sin nx sinh ny
u ( x, y )
n1 n sinh n
(2.4.1)
To compare with this exact solution we discretize the boundary into 16, 32, 64, and
128 elements together with 4 multiple nodes at the corners. The resulting values at
the internal points are shown in Table 2.4.1.
26
We can see that the MULBEM solutions approach the exact solutions as the number
of elements increases as shown in Figure 2.4.2.
2.50
2.00
error(%)
1.50
1.00
0.50
0.00
(.25,.75) (.50,.75) (.75,.75) (.50,.60) (.50,.50) (.50,.40) (.50,.25)
16 32 64 128
For this problem, we concentrate on the internal points on the line x = 0.5 close to the
edge y = 1. This edge contains points (0,1) and (1,1) which are corners and points
with discontinuous boundary conditions. It is clear from the 16-elements solution that
the errors increase near the edge as shown in Figure 2.4.3.
0.6
0.5
errors (%)
0.4
0.3
0.2
0.1
0
0.05 0.15 0.25 0.35 0.5
16 elements
Table 2.4.2 Relation between percentage errors of each partition and distance from
the edge y = 1
We have a 'rule of thumb' which says that points which are less than approximately
one half an element length from the boundary yield relatively poor results. We can see
from Table 2.4.2 that the results are consistent with this rule of thumb. We would
expect that the error of the internal solution at points near the edge can be reduced by
increasing the number of elements.
This example is concerned with the potential flow past a cylinder between two
parallel flat plates. By virtue of the symmetry of the problem, we consider the region
shown in Figure 2.4.4 in which an artificial boundary is set at four radii from the
centre of the cylinder. We discretize the boundary into 60 elements and use five
multiple nodes at the corners. The partition is shown in Figure 2.4.4.
54 34 (4,4)
(0,4)
33
55
28
27
65 17
(0,0) (4,0)
16
1
2u 0 (2.4.2)
subject to
u = 0 on y=0
u = 0 on ( x 4) 2 y 2 1, y 0
q = 0 on x = 4, 1< y < 2
u = 2 on y=2
u = y on x=0
Po in t
Point MULBEM LINBEM
In traditional BEM, we recall that the greater the number of elements we use the more
accuracy we obtain. For this problem, although we do not have the exact solution, we
see from Table 2.4.4 that the LINBEM solutions approach the MULBEM solutions as
the number of elements increases.This implies that the MULBEM (60 elements) result
is better than that of the LINBEMwith 60 elements.
29
Table 2.4.4 The normal derivatives of the boundary solution on the cylindrical
boundary
Point
Point Boudary solution
Boundary solution
MULBEM LINBEM
This problem is often used as a test for a new method since the solution has a
singularity. The geometry and boundary conditions of the problem are shown in
Figure 2.4.5. We seek the solution to Laplace's equation in the rectangular region
subject to the given boundary conditions.
q=0
7 q=0 r u = 1000
O
A B
u = 500 q=0
14
7 7
30
The analytic solution (Motz, 1946) in the neighbourhood of the point (7,0) in polar
form is
1 3 3 5 5
u ( r , ) 0 1 r 2
cos 2 r 2 cos 3 r 2 cos ... (2.4.3)
2 2 2
where r is the distance from (7,0) and is measured anti-clockwise from the line
y = 0, x > 7. Whiteman and Papamichael (1972) showed, using a conformal
transformation method, that the first three i are given by
To compare with the BETIS program (Paris and Cañas, 1997), we discretize the
boundary into 56 elements. We see that the domain of this problem contains 4 corners
and 1 discontinuous boundary point, so that we include 5 multiple nodes at such
points as shown in Figure 2.4.6.
53 39
54 38
O
61 31
A 1 5 15,16 26 30 B
We also compare the result with a variety of reference solutions (Lefeber, 1989,
Symm, 1973, Whiteman and Papamichael, 1972) as shown in Table 2.4.5.
We can see from Table 2.4.5 that the MULBEM result agrees well with the BETIS
result. Both results are better than those from the LINBEM. For the points which the
references have solutions, MULBEM solutions agree well with those solutions.
31
Table 2.4.5 The solution of the Motz problem
The symbol " ****** " in the table means there is no solution from that reference
We also investigate the solution in the neighbourhood of the point (7,0) and we see
that the algorithm works well; the solution is indistinguishable from that computed
using the BETIS code as shown in Figure 2.4.7.
0.80
Potential
0.60
0.40
0.20
0.00
1 2 3 4 5 6 7
Furthermore, we test the accuracy of the program using the solutions near the point
(7,0) to approximate the coefficients 0 and 1 in equation (2.4.3) and to compare
with the reference values in equation (2.4.4). We refine the mesh near the point (7,0)
32
and the whole boundary contains 87 elements. We use the solution near the origin in
Table 2.10 to approximate such values.
Table 2.4.6 Solutions of the points near the singular point O where r is distance from
O along OB (see Figure 2.4.5)
u r
500.0000 0.00
513.9760 0.01
520.5151 0.02
529.6550 0.04
536.6196 0.06
542.4766 0.08
547.2065 0.10
1
For small r we should have, along OB, u 0 1r 2
. Applying the least squares
1
method to fit the function u 0 1r 2
to the data in Table 2.4.6, we obtain the
coefficients
0 499.5
1 151.0
which are very close to the coefficients of analytic solution, 0 500 and
1 151.625 in equation (2.4.4).
Consider the potential problem defined in the L-shaped region as shown in Figure
2.4.8. We note that the domain of this problem contains 6 corners and one of those is
re-entrant. We are interested particularly in the accuracy of solutions near this point.
where r is the distance from O along OB and , is the internal angle at the
2
re-entrant corner O.
33
20 u = -4 1 A
21 93
q=0
12
q=0
80 u = -16 56
73 B
68 55
O 72
q=0 8
25 47
26 q=0 46
95 5
3 1
In this problem we have , hence and we obtain
2 3
2 2
q 1r 3
2 3r 3
... (2.4.6)
The dominant term close to the singular point is the first term and hence
1
q 1 1
in the neighbourhood of O.
r
Letting k 1 be the order of the singularity, we see that the relation between
log q and log r is a straight line whose slope is k.
We now approximate the value of 1 and the order of the singularity k using the
solution near the singular point O in Table 2.4.7. Applying the least squares method to
fit the function log q log 1 k log r to this data we obtain the order of singularity
k and the coefficient 1 and these are compared with those of the BETIS results as
shown in Table 2.4.8.
34
Table 2.4.7 The normal derivative at the points near the re-entrant O, r being the
distance from O along OB
q r
1.4489820 1.00
1.2576840 1.25
1.1207070 1.50
1.0177490 1.75
0.93566 2.00
Table 2.4.8 Approximation of the order of the singularity k and the coefficient 1
The symbol " ****** " in the table means there is no solution in the item
We see from Table 2.4.8 that the MULBEM program works as well as the BETIS
code.
In the next section, we introduce a variant of the multiple node method, the gradient
approach.
35
2.5 The gradient approach
The multiple node method is one of the efficient methods to resolve corner problems
in the BEM. There are several other methods similar to this approach (Subia and
Ingbar, 1995). However, they are not suitable for use with the dual reciprocity
method, a technique for solving non-homogeneous problems, which we consider in
chapter 3. The multiple node method yields a singular system of equations when used
with the dual reciprocity method.The gradient approach (Alarcón, et al, 1979, Paris
and Canãs, 1997) is a further variation of the multiple node method which we are
going to discuss in this section. We give the mathematical algorithm and show
computational results at the end of the section. Although it is more complicated than
the previous multiple node method from a programming point of view it is suitable to
apply with the dual reciprocity method.
The gradient approach is based on relating two unknown fluxes associated with a
corner, extending to the domain close to the corner the hypothesis of linear
interpolation along the elements adjacent to the corner. This way there will be three
variables at each node: one potential and two fluxes. The continuity of potential and
boundary conditions are suitable to reduce the variables to have one unknown for all
case of boundary conditions except the Dirichlet condition. We discuss only the case
when the Dirichlet conditions apply at consecutive elements.
The approach requires the knowledge of the potential at every node. Suppose that we
have potential u k 1 , u k and u k 1 at node k-1, k and k+1 respectively, as shown in
Figure 2.5.1. Then we can obtain a linear approximation to u, from the nodal value,
over the triangle defined by the three nodes. We can then find grad u over the triangle
and knowing the unit normal vectors to each of the elements we can obtain the fluxes
at node k in each of such elements.
u k 1
uk
k+1
( x3 , y 3 )
( x2 , y 2 ) k
u k 1
( x1 , y1 )
k-1
36
Due to the fact that the potential is known at nodes k 1 , k, and k 1 , the equation of
the plane represented in Figure 2.5.1 can be easily generated as
x x2 y y2 u uk
x1 x 2 y3 y 2 u k 1 u k 0 (2.5.1)
x3 x 2 y3 y 2 u k 1 u k
( y 1 y 2 )(u 3 u 2 ) ( y 3 y 2 )(u1 u 2 )
u u2 ( x x2 )
( y 1 y 2 )( x3 x2 ) ( x1 x2 )( y3 y 2 )
( x 3 x2 )(u1 u 2 ) ( x1 x2 )(u 3 u 2 )
( y y2 ) (2.5.2)
( y 1 y 2 )( x3 x2 ) ( x1 x2 )( y3 y 2 )
Using this expression, the components of the gradient at point 2, i.e. node k, can be
generated.
Let us define as
( y 1 y 2 )( x3 x2 ) ( x1 x2 )( y3 y 2 ) (2.5.3)
u y1 y 2 y y2
(u 3 u 2 ) 3 (u1 u 2 ) (2.5.4)
x
u x3 x 2 x x2
(u1 u 2 ) 1 (u 3 u 2 ) (2.5.5)
y
In general, if the gradient is known at the node k, then any flux along a certain
direction can be expressed projecting the gradient onto this direction.
Let v k be the unit vector along the gradient direction as shown in Figure 2.5.2. The
vector n R,k denotes the unit normal vector to the right element. The vector n L,k
denotes the unit normal vector to the left element.
37
(u) k
q L ,k q R ,k
vk
n L,k n R,k
k
k-1
k+1
Figure 2.5.2 The gradient approach at the node k
The unit normal vector is
u u
x y
vk = , (2.5.6)
u u
2 2
u u
2 2
x y x y k
Due to the fact that if the right-hand sides of equations (2.5.7) and (2.5.8) are known,
then they can be used to calculate the two fluxes at the node k. In this way these
equations express both fluxes as functions of a single unknown (the modulus of the
gradient). The advantage of this approach is that we need not to introduce new
variables to the system of equations.
The GRABEM program is implemented using the idea as described above. The
gradient approach is added in the process of applying the boundary condition to the
final system of equations to be solved. It must be pointed out that the number of
variables associated with a node k is initially 3, as shown in Figure 2.5.3.
q R ,k
q L ,k
q1 (k )
uk
k
[k-1] k-1
[k]
k+1
Figure 2.5.3 Three variables at node k
38
In Figure 2.5.3, u k is a potential variable and q L ,k , q R ,k are two flux variables at the
node k. There are five situations which can occur at each node.
1. Neumann-Neumann condition
2. Neumann-Dirichlet condition
3. Dirichlet-Neumann condition
4. Dirichlet-Dirichlet condition at corner
5. Dirichlet-Dirichlet condition with smooth boundary
We recall from chapter 1 that we must apply the boundary conditions to the system of
equations (1.3.24)
HU GQ (2.5.9)
Ax y (2.5.10)
We use the number before each condition to be the code of the boundary condition in
the program. The scheme of applying these conditions is shown in Table 2.5.1.
Code uk q L,k q R ,k
1 Unknown Known Known
2 Known Unknown Known
3 Known Known Unknown
4 Known Unknown Unknown
5 Known Unknown Unknown
We see from the scheme that each condition in the first three codes has only one
unknown. Hence we can apply to equation (2.5.9) directly. Code 4 is the most
difficult case to apply the boundary condition. The last case is consequent of case 4
using the fact that the two fluxes are equal. We give details for all cases in the rest of
this section.
This implies that the outward normal derivative of the potential is known at both sides
of node k and consequently, we obtain equation k as follows.
Aik H ik
39
xk u k
This condition implies that the potential is known along the element [k] and the
normal derivative is known along the element [k-1]. The application of the boundary
condition is then immediately applied as follows.
Aik G2k ,k
xk q R ,k
This condition implies that the potential is known along the element k-1 and the
normal derivative is known along the element k. The application of the boundary
condition is then immediately applied as follows.
xk q L,k
This condition implies that the potential is known along the element [k-1] and element
[k]. Before applying the boundary condition, let us define new variables in equation
(2.5.7) and (2.5.8) as
Dk (u ) k (2.5.11)
40
q R,k Dk C R,k (2.5.15)
We can see from equations (2.5.14) and (2.5.15) that the two fluxes are functions of
one variable, namely, D k . Therefore, we can now apply the boundary condition as
follows.
x k Dk
y k ... H i ,k u k ...
This condition implies that the potential is known along the element [k-1] and element
[k]. According to the smoothness of the boundary, the angle between the normal
vector and the gradient vector is zero, hence
C L,k C R,k 1
We can now use the algorithm for Code 4 and we obtain the final system of equation
as follows.
x k Dk
y k ... H i ,k u k ...
After the system of equations has been solved, the solutions of most cases are directly
obtained from the variables of their codes except those of the code 4. The solution for
case 4 is the modulus of the gradients as in equation (2.5.11). To obtain the two
normal derivatives at the node, we substitute that value in equation (2.5.14) and
(2.5.15). We discuss a few GRABEM results in the following examples. Further
applications will be focused in the next chapter.
We use the problem in Example 2.3.1 with mixed boundary conditions, using the
same discretization. The internal solutions are compared with those of using LINBEM
and MULBEM and shown in Table 2.5.2.
41
Table 2.5.2 The potential at the internal points
We see from Table 2.5.2 that the GRABEM result is the best among three methods.
The normal derivatives on the boundary nodes are shown in Table 2.5.3.
We also see from Table 2.5.3 that the normal derivative using GRABEM is the best
result. To examine more, we now test GRABEM with the Dirichlet problem.
We use the problem in Example 2.3.2 with the Dirichlet conditions, using the same
discretization. The internal solutions are compared with those of using LINBEM and
MULBEM and shown in Table 2.5.4.
42
As in the previous example, we see from Table 2.5.4 that the internal solution using
GRABEM is better compared with that from LINBEM or MULBEM. However, the
GRABEM result is only slightly different compared with that from MULBEM. The
normal derivatives on the boundary nodes are shown in Table 2.5.5.
As we expected, we see from Table 2.5.5 that the normal derivative using the gradient
approach is the most accurate.
There are two main methods for overcoming problems caused by geometric corners
and/or discontinuous boundary conditions in the boundary element method. In the
first method, usually known as the non-conforming method, functional and
collocation nodes are moved off geometric corners or points of discontinuous
boundary conditions. In the second method, the multiple node method, multiple nodes
are placed at corners and points with discontinuous boundary conditions. Various
authors claim one method better than the other (Brebbia and Niku, 1988, Manolis and
Benerjee, 1988). However, Subia and Ingber (1995) suggest that both methods are
equally reliable.
From the programming point of view, we believe that the multiple node approach is
simpler. The elegance of this method is that the additional equations from the new
collocation nodes at the multiple nodes are obtained within the framework of the
boundary element method. We need not use other laws, theorems, differentiation or
finite differencing (Subia and Ingber, 1995). We have shown that the multiple node
method works well and is a significant improvement on the standard collocation
method. The slight increase in computational cost and complexity of input data is
more than compensated by the increased accuracy.
43
For the gradient approach, it is more complicated than the multiple node from the
programming point of view. The gradient approach needs more mathematical
background as explained in section 2.5. On the other hand the multiple node method
uses only the framework of BEM. However the gradient approach is suitable to apply
for the dual reciprocity method to solve Poisson-type equations for which the multiple
node is unsuitable. We will discuss this capability together with further problems in
chapter 3.
44
3. The dual reciprocity boundary element method
This chapter is the main part of this work. The first section introduces the difficulty
associated with domain integrals which may occur in the solution of Poisson's
equation. The second section surveys the features of Poisson's equation. The third
section gives the foundation of the dual reciprocity method to handle the domain
integral problem. The fourth section investigates the two frequently-used radial basis
functions in the method. The fifth section contains the modification of the existing
program to overcome such problem. The sixth section details the formulation of the
dual reciprocity boundary element method. The seventh section implements a
program applying both radial basis functions in the dual reciprocity boundary element
method. The eighth section discusses how to solve variation on Poisson's equation
and Poisson-type equation. The ninth section implements a program using the
gradient approach in the dual reciprocity method. The tenth section discusses how to
solve non-linear problems. The eleventh section introduces the Newton's iteration
method to solve the resulting non-linear equations. The last section gives concluding
remarks.
3.1 Introduction
There are three classical numerical methods for solving problems in engineering and
applied science. The first approach is the Finite Difference Method (FDM). This
technique approximates the derivatives in the governing differential equation which
using some type of truncated Taylor expansion. The second is the Finite Element
Method (FEM). This method involves the approximation of variables over small
parts of the domain, called element, in term of polynomial interpolation functions.
The disadvantages of FEM are that large quantities of data are required to discretise
the full domain. The third one is the Boundary Element Method (BEM). This
approach is developed as a response to the FEM discretisation problem. The method
requires discretisation of the boundary only, thus reducing the quantity of data
necessary to run a program.
However, there are some difficulties in extending the technique to applications such
as non-homogeneous, non-linear and time-dependent problems. The main drawback
in these cases is the need to discretize the domain into a series of internal cells to deal
with the terms not taken to the boundary by application of the fundamental solution.
This additional discretization destroys some of the attraction of the method in terms of
the data required to run the program and the complexity of the extra operations
involved.
45
The Dual Reciprocity Method (DRM) is essentially a generalised way of constructing
particular solutions that can be used to solve non-linear and time-dependent
problems as well as to represent any internal source distribution.
2u b (3.2.1)
1 2
n
1 2
qq
u u
u bu d q q u d u u q d
2 * * *
(3.2.2)
2 1
which can be integrated by parts twice to produce (Partridge and Brebbia, 1992)
u ud bu d qu * d qu * d uq * d u q * d
2 * *
(3.2.3)
2 1 2 1
After substituting the fundamental solution u * of the Laplace operator into (3.2.3) and
grouping all boundary terms together we obtain
Notice that although the function b is known and consequently the integral in does
not introduce any new unknowns, the problem has changed in character as we need
now to carry out a domain integral as well as the boundary integral. The constant ci
depends only on the boundary geometry at the point i under consideration.
46
The simplest way of computing the domain integral term in equation (3.2.4) is by
subdividing the region into a series of internal cells, over each of which a numerical
integration scheme such as Gauss quadrature can be applied. However, this technique
loses the attraction of its boundary only character. Several methods have been
proposed by different authors; the most important of them are:
5. The Dual Reciprocity Method This is the subject of this work and constitutes
the only general technique other than cell integration.
2u b (3.3.1)
where b bx, y .
The solution to equation (3.3.1) can be expressed as the sum of the solution of a
homogenous and a particular solution as
47
u u h uˆ (3.3.2)
2uˆ b (3.3.3)
If there are N boundary nodes and L internal nodes, as shown in Figure 3.3.1, there
will be N L interpolation functions, f j , and consequently N+L particular solution,
û j .
: Boundary node
: Internal node
N L
b ( x, y )
j 1
j f j ( x, y ) (3.3.4)
If bi and f ij are the values of b and f j at node i respectively, then we have a matrix
equation for the unknown coefficients j : j 1,2,..., N L as
b F (3.3.5)
The particular solutions û j , and the interpolation functions, f j are linked through the
relation
48
2uˆ j f j (3.3.6)
uˆ
N L
b j
2
j (3.3.7)
j 1
Equation (3.3.7) can be substituted into the original equation (3.3.1) to give the
following expression
uˆ
N L
2u j
2
j (3.3.8)
j 1
Multiplying by the fundamental solution and integrating by parts over the domain, we
obtain
uˆ u d
N L
( 2 u )u * d 2 *
j j (3.3.9)
j 1
( u)u d bu d
2 * *
(3.3.10)
Integrating by parts the Laplacian terms in (3.3.9) produce the following integral
equation for each source node i (Partridge and Brebbia, 1992):
c uˆ
N L
ci u i q *ud u * qd
j 1
j i ij q *uˆ j d u * qˆ j d
(3.3.11)
uˆ j
The term q̂ j in equation (3.3.11) is defined as qˆ j , where n is the unit outward
n
normal to , and can be written as
uˆ j x uˆ j y
qˆ j
(3.3.12)
x n y n
Note that equation (3.3.11) involves no domain integrals. The next step is to write
equation (3.3.11) in discretised form, with summations over the boundary elements
replacing the integrals. This gives, for a source node i, the expression
N N N L
N N
ciui q*ud u *qd c uˆ
j i ij q*uˆ j d u *qˆ j d (3.3.13)
k 1
k
k 1
k
j 1 k 1
k
k 1
k
After introducing the interpolation function and integrating over each boundary
element, the above equation can be written in terms of nodal values as
49
N N N L
N N
ci u i H ik u k Gik q k j ci uˆ ij H ik uˆ kj Gik qˆ kj (3.3.14)
k 1 k 1 j 1 k 1 k 1
where the definition of the terms H ik and Gik is defined as in chapter 1. The index k is
used for the boundary nodes which are the field points. After application to all
boundary nodes, using a collocation technique, equation (3.3.14) can be expressed in
matrix form as
N L
Hu Gq j 1
j (Huˆ j Gqˆ j ) (3.3.15)
ˆ
ˆ GQ
Hu Gq HU (3.3.16)
Note that equation (3.3.14) contains no domain integrals. The source term b in (3.3.1)
has been substituted by equivalent boundary integrals. This was done by first
approximating b using equation (3.3.7), and then expressing both the right and left
hand sides of the resulting expression as boundary integrals using the second form of
Green's theorem or a reciprocity principle. It is this operation which gives the name to
the method: reciprocity has been applied to both side of (3.3.9) to take all the terms to
the boundary, hence Dual Reciprocity Boundary Element Method.
The Dual Reciprocity Method (DRM) is a technique by which the domain integral is
transferred to an equivalent boundary integral by using suitable interpolation
functions. The interpolation functions can be constructed in several ways. In the
engineering community, the basis function f 1 r , is widely used for this purpose,
where r is the usual Euclidean distance function. Such functions are called the Radial
Basis Functions (RBF) and are used in the context of interpolation. In the past, the
choice of approximation functions in DRM has been largely empirical and the lack of
a sound theoretical basis has been a major criticism. Golberg and Chen (1994) use
the theory of RBF to provide a mathematical foundation for DRM. Among several
available choices of RBF, the frequently used function f 1 r , and the augmented
thin plate spline f r 2 log( r ) are studied in this work. We examine the Root Mean
Square (RMS) error of both approximation terms for a variety of functions. We will
develop a program using both radial basis functions in the section 3.7. In this section
we give the mathematical concept of radial basis functions and discuss approximation
results. Although for a variety of problems the convergence of the DRM is still under
investigation (Chen and Rashed, 1998), we examine the convergence of the solution
using both radial basis functions with a few problems in section 3.7.
50
3.4.1 Mathematical concepts
The main process of the DRM is to avoid the domain integral and to use an equivalent
boundary integral instead. This is achieved by interpolating the function b in equation
(3.3.4) as
N L
b
j 1
j fj (3.4.1)
in terms of radial basis functions at some chosen number, N, of boundary and internal
nodes, L, in the domain. These nodes are called interpolation nodes. Nodes used for
evaluation in the domain will be called evaluation nodes.
In this study, two radial basis functions f 1 r and f r 2 log( r ) are considered
for use in (3.4.1) where r is the distance from some fixed point. Computational results
will be discussed in sub-section 3.4.2.
Following the notation of Partridge and Brebbia (1990), we define the position
vector, r, relative to a fixed point, P, by
r rx i ry j (3.4.2)
with rx x x P and ry y y P ,
so that
r 2 rx ry
2 2
(3.4.3)
where rx and ry are the components of r in the direction of the x and y axes.
f 1 r (3.4.4)
N L
b
j 1
j (1 r j ) (3.4.5)
where r j2 ( x x j ) 2 ( y y j ) 2 .
N L
b( x i , y i )
j 1
j (1 rij ) (3.4.6)
51
where rij2 ( xi x j ) 2 ( y i y j ) 2 which leads to the equation
b F (3.4.7)
r2 r3
uˆ (3.4.8)
4 9
x y 1 r
qˆ rx ry (3.4.9)
n n 2 3
f r 2 log r (3.4.10)
N L
b ( x, y )
j 1
r log r j a bx cy
j j
2
(3.4.11)
N L N L N L
where
j 1
j
j 1
j xj
j 1
j yj 0
f k rk log rk
2
for k = 1, 2, 3, …, N+L (3.4.12)
f N L 1 1 (3.4.13)
f N L2 x (3.4.14)
f N L 3 y (3.4.15)
52
Using relation (3.3.6) we obtain
4 4
rk log rk rk
uˆ k for k = 1, 2, 3, …, N+ L (3.4.16)
16 32
x2 y2
uˆ N L 1 (3.4.17)
4
x3
uˆ N L 2 (3.4.18)
6
y3
uˆ N L 3 (3.4.19)
6
r 2 log rk rk 2 x y
ˆqk k rx ry for k = 1, 2, 3, …, N+L (3.4.20)
4
16 n n k
x x
qˆ N L 1 (3.4.21)
2 n
x 2 x
qˆ N L 2 (3.4.22)
2 n
y 2 y
qˆ N L 3 (3.4.23)
2 n
There are various comments about these two radial basis functions. Golberg (1995)
suggested that the augmented thin plate spline might be a better and more
mathematically defensible choice because of its convergence to the exact solution and
the accuracy of solutions. However, several authors indicated that in practice TPS
showed little or no improvement over the traditional one (Agnantiaris, Polyzos and
Beskos,1996). More comments can found in (Golberg, Chen, Bowman and Power,
1998). To make our comments about this subject we examine the two radial basis
functions in the following part.
From the point of view of the boundary element method, the convergence of
approximations to b is not the primary issue, the convergence of the corresponding
algorithms for solving a given boundary value problem are (Golberg, 1995). In this
part we investigate the approximation of the two radial basis functions. The
convergence is examined by mean of Root Mean Square (RMS) error.
53
We begin with the definition of the RMS error and show the results later. An overall
estimator of the interpolation error is the root mean square (RMS) error. The RMS
error is defined as,
b
M
1 ~ 2
RMS error = i bi (3.4.24)
M i 1
~
where bi and bi are the exact and approximate values, respectively, at an evaluation
node i.
In this work, we use the 24 24 square with uniform meshes to perform the
interpolation. The varieties of interpolation nodes to investigate the convergence are
25, 49, 81 and 169. The domain to interpolate 144 evaluation nodes with 25
interpolation nodes is shown in Figure 3.4.1. The evaluation nodes are located at the
centre of each grid.
(0,24) 13 (24,24)
12 11 10 9
14 8
15 7
16 6
(0,0) 1 2 3 4 5 (24,0)
1. b( x, y ) 2
2. b( x, y ) x
3. b( x, y ) x 2
4. b( x, y ) e 10x
5. b( x, y ) sin( x y )
54
Case 1 b( x, y ) 2
The RMS errors for the source term b( x, y ) 2 with various number of
interpolation nodes are shown in Figure 3.4.1.
We see from Table 3.4.1 that the RMS errors using the augmented thin plate spline
are exactly zero. The excellent result is expected by virtue of the linear terms in the
augmented TPS function. As we would expect the TPS coefficient are found to be
i 0 for i 1, 2, 3, ..., N L, b c 0, and a 2 . On the other hand, for the
function f 1 r the coefficients are non-zero. However, the RMS errors from this
function is reduced when the number of interpolation nodes increased.
Case 2 b( x, y ) x
Similarly, for the source term b( x, y ) x with the same condition, the RMS errors
behave as the previous case as shown in Table 3.4.2.
We make similar comments here since we expect the augmented TPS to yield the
exact values since we are approximating a linear term.
55
Case 3 b( x, y ) x 2
The approximation for this case is different from the first two cases. The RMS errors
are shown in Table 3.4.3.
For the quadratic function, we can see from Table 3.4.3 that the RMS errors from both
radial basis functions decrease as the number of interpolation nodes increased. This
shows that the approximation is convergence. The convergence from TPS is more
rapid than from f 1 r.
Case 4 b( x, y ) e 10x
We see again that, for the exponential function, the TPS result is better than that of
f 1 r as shown in Table 3.4.4.
Table 3.4.4 The RMS errors for the source term b( x, y ) e 10x
Case 5 b( x, y ) sin( x y )
For the case of trigonometric function, the RMS errors are shown in Table 3.4.5.
Although the RMS errors from TPS are greater that those from f 1 r at the
beginning, they reduce more rapidly.
56
Table 3.4.5 The RMS errors for the source term b( x, y ) sin( x y )
We have shown that the interpolation using the augmented thin plate spline is better
than that using f 1 r for all cases. Further discussion can be found in Karur and
Ramachandran (1994). Although the approximation is not the only one factor for the
convergence of solutions, we shall use both radial basis functions to investigate the
accuracy of the DRM solution in section 3.7.
In this section we develop the method in a manner, such that we can easily adapt the
existing MULBEM code to solve the Poisson equation. We name the modified
program MULDRM. This program firstly uses the radial basis function f 1 r .
Consider the Poisson equation (3.3.1) with the boundary conditions as shown in
Figure 3.3.1. From equation (3.3.4) we have, for each node
N L
b
j 1
j fj (3.5.1)
Following the work of Partridge and Brebbia (1989) we shall expand b in a series of
radial basis functions as described in section 3.4. If P is any point in the domain then
a radial basis function may be written in the form (r ) where r is the distance of the
point ( x, y ) from P i.e. r is given by r 2 ( x x P ) 2 ( y y P ) 2 .
f j 1 rj (3.5.2)
57
r j2 r j3
uˆ j (3.5.3)
4 9
Set
N L
U u j uˆ j (3.5.4)
j 1
and so
N L
2U 2 u j 2 (uˆ j ) (3.5.5)
j 1
N L
2U b j f j (3.5.6)
j 1
2U 0 (3.5.7)
After the Laplace equation (3.5.27) has been solved, the values of U and Q are
known and hence we also obtain the solution of the Poisson equation (3.2.5) via
N L
u U j uˆ j (3.5.10)
j 1
N L
and q Q j qˆ j (3.5.11)
j 1
58
3.5.2 The MULDRM program
The MULBEM code is developed using the mathematical idea as described above.
We name the modified program MULDRM. This program firstly uses the radial basis
function f 1 r as discussed in section 3.4. We now test MULDRM on two
examples.
2 u 1 (3.5.12)
The problem is to be solved over an isotropic square plate occupying the region
6 x 6 and 6 y 6 . The symmetry of the region means we need consider
only the second quadrant as shown Figure 3.5.1, so that q 0 on x 0 and y 0 ,
u 0 on x 6 and y 6 .
(-6,6) 12 11 10 9 (0,6)
13 8
u 1
2
14 7
15 6
16 5
(-6,0) (0,0)
1 2 3 4
59
A particular solution is
x2
up 18 (3.5.13)
2
( 2n 1) ( 2n 1) x
2
u ( x, y ) a n cos x cosh
y 18 (3.5.14)
n 1 12 12 2
where
(2n 1)
3 sin
1 12 2
a n (3.5.15)
3 (2n 1) ( 2 n 1)
cosh
2
Initially the boundary is modelled with 12 boundary elements and 4 multiple nodes at
the corners. The calculation of the potential is required at the five internal points
shown in Figure 3.5.1.
The internal solutions obtained using MULDRM are compared with those using the
Monte Carlo method (Gipson, 1985) and the exact solution in Table 3.5.1. The Monte
Carlo method used the number of integrating points as 500, 1000 and 3000.
We see from Table 3.5.1 that the Monte Carlo result for 3000 integration points is
better compared with the other integration points. From MULDRM the multiple node
result is better than the standard collocation. To clarify more, the percentage errors are
calculated and shown in Figure 3.5.2.
60
Percentage errors in potential
3.5
3
2.5
Error (%)
2
1.5
1
0.5
0
(-2,2) (-4,2) (-3,3) (-2,4) (-4,4)
As Gipson (1985) comments, the Monte Carlo result for 3000 integration points is
within four percent tolerance hence, in an engineering application, the accuracy is
excellent. We can see from Figure 3.5.2 that percentage errors in MULDRM using
the multiple node are less than 1.5. Hence the MULDRM result is satisfactory.
Furthermore, it is the best among the other methods.
Unfortunately, we do not have the normal derivative from Monte Carlo method to
compare with. However, the normal derivative from MULDRM is compared with the
exact solution and shown in Figure 3.5.3.
0
-0.5
-1
-1.5
-2
-2.5
-3
-3.5
-4
-4.5
(0,6) (-2,6) (-4,6) (-6,6) (-6,4) (-6,2) (-6,0)
61
We can see from Figure 3.5.3 that the normal derivative from MULDRM using
multiple nodes agrees well with the exact solution. On the other hand, the normal
derivative using the standard collocation gets worse at the corners such as (0,6) and
(6,0) .
We refer to the problem given by Partridge and Brebbia (1989). Consider the Poisson
equation
2 u b ( x, y ) (3.5.16)
on an elliptical domain.
The discretisation is shown in Figure 3.5.4. The boundary conditions will be given in
each of the three cases which follows.
(0,1)
14 13 12
15 11
16 10
(-2,0) 1 9 (2,0)
2 8
3 7
4 6
5
(0,-1)
x2
y2 1 (3.5.17)
4
62
(a) The case b( x, y ) 2 (torsion problem)
This case and others in this section will be modelled using the element geometry
shown in Figure 3.5. The governing equation is
2 u 2 (3.5.18)
x2
u 0.8 y 2 1 (3.5.19)
4
q 0.2( x 2 8 y 2 ) (3.5.20)
We use the number of multiple nodes as 4, 8 and 16. The four multiple nodes are
shown in Figure 3.5.1. The internal solutions are compared with the cell integration
method (Partridge and Brebbia, 1990) and the exact solution as shown in Table 3.5.2.
We observe from Table 3.5.2 that the results using MULDRM are better than those
using cell integration. Percentage errors in potential at the internal points are shown in
Figure 3.5.5.
63
Percentage errors in potential
6
Error (%) 5
0
(1.5,0.0) (1.2,-0.35) (0.6,-0.45) (0.0,-0.45) (0.9,0.0) (0.3,0.0) (0.0,0.0)
For the constant case, we see that from Figure3.5.5 that MULDRM result is excellent
since percentage errors are less than 2 and the solution using the method is better than
that using the cell integration method. However, the multiple node approach does not
seem to help very much. In this problem the boundary is smooth and does not contain
' real ' corners.
The normal derivative on the boundary from MULDRM is shown in Figure 3.5.6.
0.00
Normal derivative
-0.50
-1.00
-1.50
-2.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Node
64
Although MULDRM gives the excellent results for the internal solution, we observe
from Figure 3.5.6 that it gives poor normal derivatives. In this case the discretised
boundary N is different from the boundary . We should see in section 3.7 that the
direct use of the dual reciprocity method does not suffer in the same way.
2u x (3.5.21)
2x x 2
u y 2 1 (3.5.22)
7 4
which satisfies the boundary condition u 0 on and produces
x 3x 2
q 2 y 2 2 (3.5.23)
14 2
Results for both MULDRM (with different numbers of multiple nodes) and Cell
Integration are given in Table 3.5.3.
For the linear case, we can see from Table 3.5.3 that the multiple nodes do not help
much as in the previous case. We observe from Figure 3.5.7 that percentage errors of
MULDRM results are less than 5. Therefore the results are satisfactory.
65
Percentage errors of potential
7
6
5
Error (%)
4
3
2
1
0
(1.5,0.0) (1.2,-0.35) (0.6,-0.45) (0.9,0.0) (0.3,0.0)
As mentioned in the case (a), we can see from Figure 3.5.8 that the normal derivative
is not good enough. We also make similar comments to the previous case.
1.5
1
Normal derivative
0.5
-0.5
-1
-1.5
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Node
66
(c) The case 2 u x 2
2u x 2 (3.5.24)
x2
u
1
50 x 2 8 y 2 33.6 y 2 1 (3.5.25)
246 4
q
1
246
x
50 x 3 96 xy 2 83.2 x
1
2 246
96 x 2 y 32 y 3 83.2 y y (3.5.26)
Results for the MULDRM program with varieties of multiple nodes and the exact
solution are given in Table 3.5.4.
For the quadratic case, as the result in the first two cases, we can see from Table 3.5.4
that the multiple node approach does not seem to help much. However, the result from
the method are better compared with those of the other methods. For the multiple
nodes, it is sufficient to take 4 multiple nodes to get the considerable solution.
Percentage errors of potential at the internal points are shown in Figure 3.5.9.
67
Percentage errors in potential
12
10
Error (%)
0
(1.5,0.0) (1.2,-0.35) (0.6,-0.45) (0.0,-0.45) (0.9,0.0) (0.3,0.0) (0.0,0.0)
We see from Figure 3.5.9 that errors at some points are greater than 5. The errors are
also greater than those for the constant and linear case. These are caused by the
approximation of the position function.
We also make the same comments as in the first two cases for the normal derivative
from MULDRM as shown in Figure 3.5.10.
0.00
-0.20
Normal derivative
-0.40
-0.60
-0.80
-1.00
-1.20
-1.40
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Node
68
We have shown that the Poisson equation can be solved using the transformation to
the Laplace equation. The MULDRM solution is satisfactory in the case of constant
and linear source terms. However, for higher order source terms, the solution is not
good enough. This is due to the approximation of the right-hand side function. The
normal derivative for the problem posed on the domain which the boundary and the
discretised boundary are different is needed to be improved. Furthermore, the
transformation to Laplace equation cannot apply to the problem whose the source
terms contain the problem variable or its derivative. The dual reciprocity method
together with radial basis functions will be introduced to handle such problems in
section 3.6 and those which follow.
69
3.6 Formulation of the dual reciprocity boundary element method
In the previous section we developed the DRM from Poisson's equation in the case
b b( x, y ) for which we can transform the partial differential equation to Laplace's
equation. In this section we continue discussing the formulation of the dual
reciprocity boundary element method from section 3.3 for cases which cannot be
transformed to Laplace's equation. We will implement a program using this
formulation in section 3.7. Furthermore, it is suitable to be modified for solving non-
linear problems later on.
The vector in equation (3.3.16) will now be considered. It was seen in equation
(3.3.4) that
N L
b
j 1
j fj (3.6.1)
b F (3.6.2)
where each column of F consists of a vector f j containing the values of the function
f j at the ( N L) DRM collocation points. In the case of the problems considered in
this section, the function b in (3.3.1) and (3.3.3) is a known function of position. Thus
equation (3.6.2) may be inverted to obtain , i.e.
F 1b (3.6.3)
The right-hand side of equation (3.3.16) is thus a known vector. It can be written as
Hu Gq d (3.6.4)
ˆ
ˆ GQ
d HU (3.6.5)
Ax y (3.6.6)
70
case of internal nodes, as was explained in section 3.3, ci 1 and equation (3.3.14)
becomes
N N N L
N N
u i H ik u k Gik q k j ˆ
u ij H ˆ
u
ik kj Gik qˆ kj (3.6.7)
k 1 k 1 j 1 k 1 k 1
The r function was adopted first by Nardini and Brebbia (1983) and then by most
researchers as the simplest and most accurate alternative. In this case the function f of
r is the radial basis function as discussed in section 3.4.
r is given by r rx i ry j
so that
r 2 rx2 i ry2 (3.6.8)
For the radial basis function f r , it can easily be shown that the corresponding û
r2
function is , in the two-dimensional case.
9
The function q̂ is by
qˆ
r
3
rx cos(n, x) ry cos(n, y) (3.6.9)
In the above, the direction cosine refers to the outward normal at the boundary with
respect to the x and y axes. Formula (3.6.9) may be easily obtained using (3.3.12)
r rx r ry
and remembering that and .
x r y r
Furthermore, Partridge and Brebbia (1992) suggest that f r can in fact be regarded
as just one component of the series
f 1 r r 2 ... r m (3.6.10)
71
The û and q̂ functions corresponding to (3.6.10) are :
r2 r3 r m 2
uˆ ... (3.6.11)
4 9 (m 2) 2
x y 1 r rm
qˆ rx ry ... (3.6.12)
n n 2 3 (m 2)
However, this work focuses on the two radial basis functions: the linear function
f 1 r and the augmented thin plate spline f r 2 log r as discussed in section
3.4.
We will implement a program using both radial basis functions with the DRM
formulation as described above in the section 3.7.
DRBEM1 program is based on the idea of the previous section. We note that in the
previous section the implementation of the boundary and internal solutions were
separate processes. In this section we develop a procedure in which the boundary and
the internal solutions are computed simultaneously. This section is divided into two
parts. The first part details the scheme of matrix equations of the program, using both
the radial basis functions as described in section 3.4. The second part discusses
computational results from DRBEM1. They are compared with those obtained using
MULDRM.
The basic system of equations for the DRM is derived from equation (3.3.16) as
Hu Gq (HU ˆ )
ˆ GQ (3.7.1)
N
NxN N - NxN N = NxN Nx(N+L) - NxN Nx(N+L) +
L
H u G q H Û G Q̂
Notice that the matrices on the right hand side are known. H and G are from the
main frame of the standard collocation method. The matrix Û is derived from
72
equation (3.4.5) for f 1 r and from equations (3.4.13) to (3.4.16) for the
augmented thin plate spline. The matrix Q̂ is derived from equation (3.4.6) for Linear
and from equations (3.4.17) to (3.4.20) for TPS. The vector is obtained by solving
equation (3.6.2). The result on the right hand side is stored in the vector d as in
equation (3.6.4). After applying the boundary conditions, the system (3.6.6) will be
solved. Having obtained the values of unknowns on the boundary, the next step is to
calculate values at the internal points. For this, equation (3.6.7)
N N N L
N N
u i H ik u k Gik q k j uˆ ij H ik uˆ kj Gik qˆ kj (3.7.2)
k 1 k 1 j 1 k 1 k 1
The scheme will be described term by term. On the left hand side the identity matrix I
of size L L multiplies the L values of u at internal nodes. This is due to the ci
terms, which are all unity at internal nodes as explained in (1.2.41). Obviously,
values of u and q are obtained by solving equation (3.7.1). The same matrix appears
in the last term on the right hand side. The H and G matrices partitions of size
L N which appear on both sides of equation (3.7.2) are produced by integrating
over the boundary from each internal node. The matrices Û and Q̂ together with the
vector are the same as used in equation (3.7.1).
We emphasis that DRBEM1 uses both schemes for its implementation. However, the
process to execute the internal solution is separated from the boundary solution. This
can be done only in the case that the function b is known. For the case when the
function b contains the unknown, u, the two schemes will be integrated in one global
scheme. We will discuss this case in section 3.8. Next we will investigate the
computational results of the program.
This sub-section we test DRBEM1 on two problems. Both problems have been tested
by MULDRM in section 3.5. From this section and later on, Linear stands for the
radial basis function f 1 r and TPS is for thin plate spline f r 2 log r augmented
with the linear term.
73
Example 3.7.1 Square plate with constant source term
This example uses the same problem as in Example 3.5.1. The internal solutions
obtained using DRBEM1 are compared with those of MULDRM, the Monte Carlo
method (Gipson, 1987) and the exact solution in Table 3.7.1. We can use multiple
nodes here because these problems can be transformed to the Laplace equation as in
section 3.5.
Table 3.7.1 Internal solutions of the square plate problem in Example 3.7.1
Percentage errors of the potentials at the internal points are shown in Figure 3.7.3.
3.5
3
2.5
Error (%)
2
1.5
1
0.5
0
(-2,2) (-4,2) (-3,3) (-2,4) (-4,4)
We see from Figure 3.7.3 that all results are satisfactory since their percentage errors
are usually less than 3.5. The result from the multiple node method are better among
the other methods. DRBEM1 using Linear works as well as the standard method. The
TPS result is better than the Linear one. The normal derivative is shown in Figure
3.7.4.
74
Normal derivative from DRBEM1
0
-0.5
Normal derivative
-1
-1.5
-2
-2.5
-3
-3.5
-4
-4.5
(0,6) (-2,6) (-4,6) (-6,6) (-6,4) (-6,2) (-6,0)
We see from Figure 3.7.4 that the normal derivative is poor, particularly, at corners.
DRBEM1 needs to be improved and we shall discuss this subject in section 3.9.
This example uses the same problem as in Example 3.5.2. We will also discuss three
cases: constant, linear and quadratic.
The internal solutions obtained using DRBEM1 are compared with those from
MULDRM, and the exact solution as shown in Table 3.7.2.
75
We can see from Table 3.7.2 that it is difficult to distinguish which is the better result.
Percentage errors of internal solution are calculated to clarify as shown in Figure
3.7.5.
2
Eorror (%)
1.5
0.5
0
1 2 3 4 5 6 7
For the constant function, we see that from Figure 3.7.5 that each solution is
satisfactory since its percentage error is less than 2. The MULDRM result using the
standard collocation is poorer among the other methods. The normal derivative is
shown in Figure 3.7.6.
0.0
-0.2
Normal derivative
-0.4
-0.6
-0.8
-1.0
-1.2
-1.4
-1.6
-1.8
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
76
As mentioned in Example 3.5.2 for the case (a) in Figure 3.5.5, we see from Figure
3.7.6 that the normal derivative on the boundary from DRBEM1 for both radial basis
functions agrees well with the exact solution.
The internal solutions obtained using DRBEM1 are compared with those from
MULDRM and the exact solution as shown in Table 3.7.3.
The percentage errors in potential at the internal points are shown in Figure 3.7.7.
12
10
8
6
4
2
0
(-1.5,0) (-0.9,0) (-0.3,0) (0.3,0) (0.9,0) (1.5,0)
For the linear function, we can see from Figure 3.7.7 that the DRBEM1 result is still
satisfactory because each percentage error is less than 4. The MULDRM result using
77
the multiple node method is also satisfactory and better than that using the standard
collocation. The normal derivative from DRBEM1 is shown in Figure 3.7.8.
0.8
0.6
Normal derivative
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
As we expected, we see from Figure 3.7.8 that the normal derivative from DRBEM1
agrees well with the analytic value except at the corners. Furthermore the normal
derivative is better than that from MULDRM as seen in Figure 3.5.7
The internal solutions obtained using DRBEM1 are compared with those from
MULDRM and the exact solution as shown in Table 3.7.4.
For the quadratic function, we see from Table 3.7.4 that the MULDRM result using
the multiple node method is still better than that using the standard collocation
78
method. The DRBEM1 result is nearly the same as the multiple node result. For
DRBEM1, the solution using Linear is slightly better than that using TPS. The
percentage errors of the solutions are shown in Figure 3.7.9.
18
16
14
12
10
8
6
4
2
0
(-1.5,0) (-0.9,0) (-0.3,0) (0,0) (0.3,0) (0.9,0) (1.5,0)
We can see from Figure 3.7.9 that the solution at some points for each method is not
satisfactory since their percentage errors are greater than 5. This problem needs to
increase the interpolation nodes. We shall discuss such problem in terms of the
convergence of solutions in sub-section 3.7.3. The normal derivative from DRBEM1
is shown in Figure 3.7.10.
0.0
Normal derivative
-0.2
-0.4
-0.6
-0.8
-1.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
79
Although the internal solution from DRBEM1 is worse at some points, we can see
from Figure 3.7.10 that the normal derivative from DRBEM1 agrees well with the
analytic value except at the corners. The normal derivative is better than that from
MULDRM as seen in Figure 3.5.9.
We have shown that on the problem in which the boundary N and contains
corners, MULDRM works better than DRBEM1 (see Example 3.5.1 and 3.7.1). The
DRBEM1 will be improved to resolve this problem in section 3.9. On the other hand,
on the problem in which the boundary N , DRBEM1 gives more accuracy in
normal derivative than MULDRM (see Example 3.5.2 and 3.7.2). Furthermore,
DRBEM1 can be modified to solve problems whose the source term contains the
problem variable but MULDRM cannot. This is because such problems are not able to
transform to Laplace's equation. We discuss this subject in section 3.8.
This subsection we examine the convergence of the solution obtained by both radial
basis functions as introduced in section 3.4. We investigate two problems. The first
problem is from Example 3.7.2. We note that in such problems the boundary is
different from N . The second problem is a mixed problem. Boundary conditions
comprise both Dirichlet and Neumann. Furthermore, some parts of its boundary N
are the same as .
This example uses the same problem as in Example 3.7.2. To examine the
convergence, we partition the boundary into 8, 16, 32, 64 and 128 elements with 17
internal points. We shall discuss all three cases as in Example 3.7.2:
The internal solution for each partition is considered for both radial basis functions.
The percentage errors are shown in Table 3.7.5.
We can see from Table 3.7.5 that all solutions are satisfactory when the number of
elements is greater than or equal to 16 since percentage errors are less than 3. The
results from TPS are better than those from Linear. The convergence of the internal
solution at the point (1.5, 0) is shown in Figure 3.7.11.
80
Table 3.7.5 Percentage errors of the internal solution for each partition
Number of elements
Point 8 16 32 64 128
Linear TPS Linear TPS Linear TPS Linear TPS Linear TPS
(1.5,0.0) 5.259 4.079 0.749 0.885 0.123 0.212 0.052 0.052 0.377 0.013
(1.2,-0.35) 7.039 6.282 1.809 1.700 0.193 0.407 0.100 0.100 0.279 0.025
(0.6,-0.45) 8.895 8.362 2.279 2.053 0.359 0.501 0.123 0.123 0.217 0.030
(0.0,-0.45) 9.048 7.615 2.265 2.034 0.372 0.500 0.123 0.123 0.202 0.030
(0.0,0.0) 7.495 6.207 1.68 1.534 0.25 0.374 0.092 0.092 0.178 0.023
(0.3,0.0) 7.597 6.300 1.663 1.525 0.242 0.371 0.091 0.091 0.183 0.023
(0.9,0.0) 7.575 6.245 1.501 1.415 0.175 0.343 0.084 0.084 0.219 0.021
6
5
Error (%)
4
3
2
1
0
8 16 32 64 128
Number of elements
Linear TPS
We see from Figure 3.7.11 that the errors using TPS are less than those using Linear
at almost every internal point. However, the solution using Linear or TPS converges
to the exact solution as the number of elements increases.
The internal solution for this case is similar to the previous case. The convergence of
the internal solution at the point (1.5,0) is shown in Figure 3.7.12.
We observe from Figure 3.7.12 that the error using TPS is greater than that of using
the Linear for small numbers of nodes. However, the solution eventually converges to
the exact solution as the number of elements increases more rapidly than does Linear.
81
Convergence of internal solution at (1.5,0)
10
9
8
7
Error (%)
6
5
4
3
2
1
0
8 16 32 64 128
Linear TPS
In this case the function is a quadratic function. The effect of the function
approximation is more apparent than inthe first two cases. The convergence of the
internal solution at the point (1.5, 0) is shown in Figure 3.7.13.
25
20
Error (%)
15
10
0
8 16 32 64 128
Linear TPS
82
It is clear from Figure 3.7.13 that, although the percentage errors from TPS are much
greater than those from Linear for small number of nodes the solution converges to
the exact solution as the number of elements increases more rapidly for TPS.
Moreover TPS behaviour is steady improving while the Linear case shown more
irregular error variation.
We note that for a problem when the position function is constant, linear or quadratic,
the convergence of solutions using TPS is better than that using Linear. We examine
the case of the function of position being an exponential function in the following
example.
Example 3.7.4 The mixed problem with exponential function right-hand side
2 u 5e 2 x y (3.7.3)
(0,1)
9 7
11 5
u e 2 x y
2 x y
q 2e 13
3
15
(0,0) (1,0)
17 19 21 23 1
q e 2 x y
Boundary node Internal node
We can see from Table 3.7.6 that all solutions are satisfactory when the number of
elements is more than or equal to 24 because the percentage errors are less than 3. The
percentage errors from TPS are less than those from Linear except some points near
83
corners such as (0.25, 0.25). The convergence of the internal solution at the point
(0.75, 0.25) is shown in Figure 3.7.15.
Table 3.7.6 Percentage errors in the internal solution for each partition
Number of elements
Point 12 24 48 96 144
Linear TPS Linear TPS Linear TPS Linear TPS Linear TPS
(0.75,0.25) 1.837 1.065 0.559 0.227 0.361 0.141 0.298 0.104 0.281 0.074
(0.50,0.25) 2.354 2.012 1.028 0.9 0.581 0.505 0.385 0.318 0.325 0.205
(0.25,0.25) 4.482 5.187 1.999 2.619 0.901 1.496 0.374 0.95 0.205 0.623
(0.25,0.50) 1.95 2.412 0.996 1.218 0.577 0.73 0.351 0.492 0.279 0.349
(0.50,0.50) 1.793 1.648 0.822 0.614 0.533 0.305 0.422 0.19 0.39 0.127
(0.75,0.50) 1.598 3.612 0.521 0.366 0.269 0.128 0.199 0.066 0.185 0.046
(0.50,0.75) 0.996 3.606 0.297 0.316 0.195 0.121 0.161 0.067 0.152 0.047
1.5
Error (%)
0.5
0
12 24 48 96 144
Number of elements
Linea r TPS
Figure 3.7.15 Convergence of the internal solution at the point (0.75, 0.25)
We can see from Figure 3.7.15 that the error using TPS is less than that of using
Linear. The solution converges to the exact solution as the number of elements
increases.
For all cases of the position function, constant, linear, quadratic and exponential, we
have shown that solution from DRBEM1 using TPS is better than that using Linear
provided we have enough elements. However, DRBEM1 needs to improve in two
ways. The first one is the treatment of corners. This feature will be discussed in
section 3.9. The second one is to solve Poisson-type problem when the right-hand
side function b contains the problem variable. We discuss this feature in the following
section.
84
3.8 The DRBEM2 Program
In this section we discuss how to modify the DRBEM1 code in order to solve various
of Poisson-type problems. We note that the process of finding the internal solution
using DRBEM1 is separate from finding the boundary solution. For Poisson-type
problems in which the right-hand side contains the problem variable or its derivative,
including non-linear problems, we are not able to simply expand this term. We set up
a solution process in which the determination of the boundary solution and the
internal solution are obtained simultaneously. There is a cost, however, in that this
approach requires the inverse of an ( N L) ( N L) matrix. The program DRBEM2
is modification of the DRBEM1 code in such a manner. The new program finds all
variables at the same time. Furthermore, it includes both radial basis functions. We
explain how to implement it in the following subsection.
2u b (3.8.1)
As described in the formulation of the dual reciprocity method, the matrix form of the
system of equations is obtained from Equation (3.6.3) to (3.6.5) as
Hu Gq (HU ˆ )F 1b
ˆ GQ (3.8.2)
For a discretisation with N boundary nodes and L internal points, this may be
represent in schematic form as shown in Figure 3.8.1.
Figure 3.8.1 Scheme of the whole system of equations, after Partridge and
Brebbia (1992)
85
We note that the scheme in Figure 3.8.1 is derived from the two schemes in Figure
3.7.1 and Figure 3.7.2.
In general, the right hand side of equation (3.8.1) is an unknown function so that the
right hand side of equation (3.8.2) is unknown.
Define
S (HU ˆ )F 1
ˆ GQ (3.8.3)
Hu Gq Sb (3.8.4)
From the formulation (3.8.4), in this case the vector b is known. Setting
R Sb (3.8.5)
Hu Gq R (3.8.6)
Ax y (3.8.7)
To test the program, we examine a few examples for this kind of Poisson equations as
follows:
We test the program with the problem in Example 2.3.1. The results are compared
with those using LINBEM and MULBEM programs and are shown in Table 3.8.1.
We see from Table 3.8.1 that DRBEM2 works well but the accuracy is less than that
of MULBEM. This is because the domain contains corners. The normal derivative on
the boundary from DRBEM2 is shown in Figure 3.8.2
86
Table 3.8.1The potential at the internal points
60
40
Normal derivative
20
-20
-40
-60
1 2 3 4 5 6 7 8 9 10 11 12 1
Although the internal solution is satisfactory, we can see from Figure 3.8.2 that the
normal derivative needs to be improved. We shall improve DRBEM2 to resolve such
problems in section 3.9. We note that the Linear and TPS results are the same because
the program does not require interpolation of the right-hand side term in Laplace
problems.
We also test the program with the problem in Example 2.3.2 with the Dirichlet
problem. The results are compared with those using LINBEM and MULBEM
programs and are shown in Table 3.8.2.
We also see from Table 3.8.2 that DRBEM2 works well but the accuracy is less than
that of MULBEM. We make similar comments as in the previous example. The
normal derivative on the boundary from DRBEM2 is shown in Figure 3.8.3
87
Table 3.8.2 The potential at the internal points
60
40
Normal derivative
20
-20
-40
-60
1 2 3 4 5 6 7 8 9 10 11 12 1
We also see from Figure 3.8.3 that the normal derivative from DRBEM2 for this case
is quite poor. We also make the same comments as in Example 3.8.1.
We use the problem as in the case (c) of Example 3.7.2. The potential values at the
internal nodes are shown in Table 3.8.3 and the normal derivatives on the boundary
are shown in Table 3.8.4.
We observe from Table 3.8.3 and Table 3.8.4 that the results from both programs are
nearly the same. This shows that the separation of solving the boundary and the
internal solution does not affect to the solution. Consequently,The normal derivative
from both programs is also the same as in Figure 3.7.10.
88
Table 3.8.3 The potential at the internal nodes of the problem
We test the program with a problem which the exponential function as in Example
3.7.4.
.
Consider the Poisson equation
2 u 5e 2 x y (3.8.8)
where the boundary conditions are as in the mixed problem of Example 3.7.4.
The potential solutions at the internal nodes are shown in Table 3.8.5. We observe
from Table 3.8.5 that the solutions from both programs are almost identical. This fact
is true for all problems in section 3.7. We note that the simultaneous solution for
internal and boundary values yields the same solution as the separated approach. So
that it is sufficient to investigate only DRBEM2 for the rest of this section.
89
Table 3.8.5 The potential at the internal nodes of Example 3.8.4
25
20
Normal derivative
15
10
5
0
-5
-10
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
The normal derivative on the boundary as shown in Figure 3.8.4 agrees well with the
exact value except at the corners. We shall discuss how to handle such problems in
section 3.9.
b c( x, y )u (3.8.9)
b Cu (3.8.10)
90
where C is the diagonal matrix
c( x1 , y1 ) 0 0 ... 0
0 c( x 2 , y 2 ) 0 ... 0
C 0 0 c( x3 , y 3 ) ... 0 (3.8.11)
... ... ... ... ...
0 0 0 ... c( x N L , y N L )
Setting
R SC (3.8.12)
(H R)u Gq (3.8.13)
Ax y (3.8.14)
2 u u (3.8.15)
on the domain with the boundary as shown in Figure 3.5.3 of Example 3.5.2. We
consider the Dirichlet problem for which the exact solution is
u sin x (3.8.16)
The solution of the problem is also compared with the DRM solution given by
Partridge and Brebbia (1990) and the exact solution in Table 3.8.6.
We see from Table 3.8.6 that DRBEM2 works slightly better than the reference
program. In this example, it is clear that the solution using the thin plate spline is
better than that using the other radial basis functions. The normal derivative on the
boundary is shown in Figure 3.8.5.
91
Table 3.8.6 The potential at the internal nodes of the problem
0.50
0.40
Normal derivative
0.30
0.20
0.10
0.00
-0.10
-0.20
-0.30
-0.40
-0.50
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
We see from Figure 3.8.5 that the normal derivative on the boundary agrees with the
exact value..
u u
3.8.4 The Poisson-type equation 2 u b( x, y, , )
x y
u
b c( x, y ) (3.8.17)
x
92
where c( x, y ) is a function of position. If we collocate at the nodes we can rewrite
(3.8.17) in matrix form as
u
bC (3.8.18)
x
u u u u
T
( x1 , y1 ) ( x2 , y 2 ) ... ( x N L , y N L ) (3.8.19)
x x x x
b F (3.8.20)
A similar equation may be written for the nodal value of u in matrix form as
u F (3.8.21)
u F
(3.8.22)
x x
where
F f j f j f j
T
( x1 , y1 ) ( x2 , y 2 ) ... ( x N L , y N L ) (3.8.23)
x x x x
u F 1
F u (3.8.24)
x x
F 1
Hu Gq SC F u (3.8.25)
x
Setting
F 1
R SC F (3.8.26)
x
(H R)u Gq (3.8.27)
93
Applying the boundary condition as described in equation (1.3.25) we obtain the
system of equations in matrix form as
Ax y (3.8.28)
u
We can use a similar process to handle . We end this section with two examples.
y
u ex (3.8.30)
The solution of the problem is also compared with the DRM solution given by
Partridge and Brebbia (1990) and the exact solutions in Table 3.8.7.
We see from Table 3.8.7 that DRBEM2 using TPS works better than the reference
program using f 1 r and f 1 r r 2 r 3 . The normal derivative from
DRBEM2 is shown in Figure 3.8.6.
We see from Figure 3.8.6 that the normal derivative from DRBEM2 agrees well with
the exact value.
94
Normal derivative from DRBEM2
8.0
Normal derivative 7.0
6.0
5.0
4.0
3.0
2.0
1.0
0.0
-1.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 1
u u
Example 3.8.7 The case 2 u
x y
Consider the Poisson-type equation
u u
2u (3.8.31)
x y
on domain with the boundary as in Figure 3.5.3 of Example 3.5.2. We consider the
Dirichlet problem for which
u ex e y (3.8.32)
The solution of the problem is also compared with the DRM solution given by
Partridge and Brebbia (1990) and the exact solutions in Table 3.8.8.
As we expected, we can see from Table 3.8.8 that the solution using TPS is better
than that using the other radial basis functions at almost every internal point.
95
The normal derivative is shown in Figure 3.8.7.
8.0
7.0
6.0
Normal derivative
5.0
4.0
3.0
2.0
1.0
0.0
-1.0
-2.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 1
We see from Figure 3.8.7 that the normal derivative agrees well with the exact
solution. However, we have seen that, for the problem in which the boundary N ,
the normal derivative is better than for that in which the boundary N .
From Example 3.8.5 to 3.8.7, we have shown that DRBEM2 works slightly better
than Partridge and Brebbia method (1990) for the problem considered so far in which
the boundary is smooth. However, we need to resolve corner problems as we
mentioned previously. We discuss a modification of DRBEM2 to deal with such
problems in the following section.
As we mentioned in Example 3.8.1, the solution using DRBEM2 is not quite good
enough for the problem when the boundary contains corners. Unfortunately, the
multiple node method is not able to be used with DRM. Collocation at the same point
forces the matrix F in equation (3.6.2) to be singular. Consequently, the vector
cannot be obtained. The gradient approach as discussed in section 2.5 is introduced to
resolve the problem. It is suitable for the DRM formulation. DRBEM2 is modified
using this gradient approach. We call the new code GRADRM. To test on several
problems, GRADRM is implemented to solve the Poisson-type equation in the form
96
u u
2 u p1 ( x, y) p 2 ( x, y)u p3 ( x, y) p 4 ( x, y) (3.9.1)
x y
We apply both the radial basis function f 1 r and the augmented thin plate
spline f r log r a bx cy . The program is tested on seven cases:
2
Case 1: p1 p 2 p3 p 4 0
Case 2: p 2 p3 p 4 0 , p1 ( x, y) 1
Case 3: p 2 p3 p 4 0 , p1 ( x, y) x 2
Case 4: p 2 p3 p 4 0 , p1 ( x, y ) 5e 2 x y
Case 5: p1 p3 p 4 0 , p2 ( x, y) 1
Case 6: p1 p2 p4 0 , p3 ( x, y ) 1
Case 7: p1 p2 0 , p3 ( x, y ) 1, p 4 ( x, y ) 1
Case 1 p1 p 2 p3 p 4 0
This case is the Laplace equation. We test the program and compare the result with
the previous programs. We discuss two example for this case.
2u 0 (3.9.2)
with boundary conditions as used in the mixed boundary problem in Example 2.3.1.
The internal solutions are compared with those of using LINBEM, MULBEM and
DRBEM2 and shown in Table 3.9.1
We see from Table 3.9.1 that the solution obtained using GRADRM is better than the
other three methods. We note that there is no difference between two radial basis
functions because the Laplace equation does not use the interpolation process in the
program. The normal derivatives on the boundary nodes are shown in Figure 3.9.1.
97
Normal derivative from DRBEM2 and GRADRM
60
40
Normal derivative
20
-20
-40
-60
1 2 3 4 5 6 7 8 9 10 11 12 1
Figure 3.9.1 Normal derivative on the boundary from DRBEM2 and GRADRM
We see from Figure 3.9.1 that the normal derivative obtained using GRADRM agree
well with the exact value. It is much better compared with that using DRBEM2.
2u 0 (3.9.3)
with the Dirichlet conditions used in Example 2.3.2. The internal solutions are
compared with those of using LINBEM, MULBEM and DRBEM2 and shown in
Table 3.9.2.
We see from Table 3.9.2 that the solution obtained using GRADRM is still better
than the other three methods. The normal derivatives on the boundary are shown in
Figure 3.9.2.
98
Normal derivative from DRBEM2 and GRADRM
60
40
Normal derivative
20
-20
-40
-60
1 2 3 4 5 6 7 8 9 10 11 12 1
Figure 3.9.2 Normal derivative on the boundary from DRBEM2 and GRADRM
We see from Figure 3.9.2 that the normal derivative obtained using GRADRM is
much better than that using DRBEM2 as we expected. There is no difference between
two radial basis functions because the Laplace equation does not use the interpolation
process in the program as in the previous problem. We note that GRADRM works
better than the other methods for Laplace's equation, particularly, the problem posed
on the boundary in which N .
Case 2 p 2 p3 p 4 0 , p1 ( x, y) 1
In this case we return to the Poisson problem whose domain has corners as in
Example 3.5.1.
2 u 1 (3.9.4)
with the boundary condition as used in Example 3.5.1. The result from GRADRM is
compared with the previous programs as shown in Table 3.9.3.
We can see from Table 3.9.3 that GRADRM using Linear works almost as well as
MULDRM. The result is better compared with that of the other programs. The
GRADRM solution using TPS is worse compared with that using Linear. As we
mentioned in section 3.4.4, it should be better as the number of elements increases. To
examine this proposition we partition the boundary into 24 elements. The result is
shown in Table 3.9.4.
99
Table 3.9.3 The potential at the internal points for 12 elements
We can see from Table 3.9.4 that GRDRM using Linear works as well as MULBEM
using multiple nodes. We compute the RMS error of the GRADRM solution and find
that it is 0.016 for TPS and 0.043 for Linear. Therefore, the GRADRM result using
TPS is the best compared with that of the other programs as expected. We note that
TPS result improves more than Linear as the number of elements increases. The
normal derivative on the boundary is shown in Figure 3.9.3
0
-0.5
Normal derivative
-1
-1.5
-2
-2.5
-3
-3.5
-4
-4.5
(0,6) (-2,6) (-4,6) (-6,6) (-6,4) (-6,2) (-6,0)
100
We can see from Figure 3.9.3 that the normal derivative from DRBEM2 is worse at
the corners such as (0,6) and (-6,0). On the other hand, the derivative from GRADRM
agrees well with the exact value.
Case 3 p 2 p3 p 4 0 , p1 ( x, y) x 2
This case we again examine the torsion problem in case (c) of Example 3.5.2.
2u x 2 (3.9.5)
with boundary solution as in case (c) of Example 3.5.2. The internal solution at the
some internal nodes on the line y 0 is shown in Table 3.9.5.
Since the boundary is smooth, we can see from Table 3.9.5 that the solution from
DRBEM2 and GRADRM are the same. We also see that the Linear result is slightly
better than the TPS result because of the small numbers of interpolation elements.
However the convergence of the TPS solution is more rapid than that of Linear as
seen in Figure 3.7.13 of Example 3.7.3.
We note that this problem the boundary N which the normal derivative from
MULDRM is poor (see Figure 3.5.9). However, the normal derivatives from
DRBEM2 and GRADRM agree well with the exact value except at the end of the
major axis of the elliptical boundary since the big change of the normal vectors at
these points as shown in Figure 3.9.4
101
Normal derivative from DRBEM2 and GRADRM
0.0
Normal derivative -0.2
-0.4
-0.6
-0.8
-1.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Case 4 p 2 p3 p 4 0 , p1 ( x, y ) 5e 2 x y
2 u 5e 2 x y (3.9.6)
with boundary as in Example 3.7.3. The potential solutions at the internal nodes are
shown in Table 3.9.6.
We see from Table 3.9.6 that the internal solution using GRADRM is better than that
using DRBEM2. The RMS error of the GRADRM solution using TPS is 0.011 and is
0.016 using Linear. Hence, the GRADRM result using TPS is the better of the two
methods.
102
The normal derivative on the boundary is shown in Figure 3.9.5.
25
Normal derivative
20
15
10
5
0
-5
-10
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
We see from Figure 3.9.5 that the normal derivative agrees well with the exact value
except at the corners such as the node 1 and 9. Furthermore, the normal derivative
using GRADRM at such nodes is better than those using DRBEM2.
Case 5 p1 p3 p 4 0 , p2 ( x, y) 1
2 u u (3.9.7)
with boundary condition as used in Example 3.8.5.The solution of the problem is also
compared with that of the method of Partridge and Brebbia (1992) and the exact
solution in Table 3.9.7.
We see from Table 3.9.7 that the solutions obtained GRADRM and DRBEM2 give
the same result. This is because the domain of the problem has a smooth boundary.
However, it is slightly better than the results in the reference at almost every point.
The TPS solution is better than the Linear solution as the absolute errors are shown in
Figure 3.9.6.
103
Table 3.9.7 The internal solutions of the problem in Example
0.004
0.0035
Absolute error
0.003
0.0025
0.002
0.0015
0.001
0.0005
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
GRADRM(Linear) GRADRM(TPS)
We can see from Figure 3.9.7 that the normal derivative from GRADRM is the same
as that form DRBEM2 (see Figure 3.8.4). This is because the boundary is smooth and
N .
104
Normal derivative from GRADRM
0.50
0.40
Normal derivative
0.30
0.20
0.10
0.00
-0.10
-0.20
-0.30
-0.40
-0.50
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Case 6 p1 p2 p4 0 , p3 ( x, y ) 1
105
As for the previous example, we see from Table 3.9.8 that the solutions obtained
using GRADRM and the DRBEM2 give the same result. Once again, the TPS result
is still better than that of using Linear. The normal derivative on the boundary is
shown in Figure 3.9.8.
8.0
Normal derivative
6.0
4.0
2.0
0.0
-2.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 1
We see from Figure 3.9.8 that the normal derivative from both radial basis functions
agrees well with the exact value.
Case 7 p1 p2 0 , p3 ( x, y ) 1, p 4 ( x, y ) 1
u u
Example 3.9.8 The case 2 u
x y
u u
2u (3.9.9)
x y
with the boundary conditions as in Example 3.8.7. The solution of the problem is also
compared with that of the method in Partridge and Brebbia (1992) and the exact
solutions in Table 3.9.9.
As before we have seen that GRADRM and DRBEM2 give the same result for a
problem with a smooth boundary. We also can see from Table 3.9.9 that this
statement is true for this problem. Furthermore, the GRADRM result using TPS is
more accurate compared with that using Linear.
106
Table 3.9.9 The internal solutions of the problem in Example 3.9.9
8.0
7.0
Normal derivative
6.0
5.0
4.0
3.0
2.0
1.0
0.0
-1.0
-2.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 1
We can see from Figure 3.9.9 that the normal derivative from GRADRM agrees well
with the exact value. The TPS result is more accurate than the Linear one.
In the Laplace case, we have shown that GRADRM works better than MULBEM on
problems whose domain contains corners. However, from a programming point of
view, the MULBEM code is simpler. For radial basis functions the augmented thin
plate spline is more accurate than the linear function f 1 r . In general, GRADRM
works better than DRBEM2. However, in the problems containing a smooth domain,
DRBEM2 works as well as GRADRM. So far, the work has dealt with linear
problems. We are going to develop GRADRM to solve non-linear problems in section
3.10.
107
3.10 Non-linear problems
2 u c1 ( x, y )u n (3.10.1)
b c1 ( x, y )u n (3.10.2)
If we collocate at the nodes then equation (3.10.2) can be written in matrix form
b (C1U)u (3.10.3)
(u1 ) n 1 0 0 ... 0
0 (u 2 ) n 1 0 ... 0
U 0 0 (u 3 ) n 1 ... 0 (3.10.4)
... ... ... ... ...
0 ... (u N L ) n 1
0 0
ˆ GQ
Hu Gq HU
ˆ (3.10.5)
ˆ GQ
Hu Gq HU
ˆ F 1b (3.10.6)
108
ˆ GQ
Hu Gq HU
ˆ F 1 (C U)u
1 (3.10.7)
ˆ F 1 (C U)
ˆ GQ
R HU 1 (3.10.8)
(H R)u Gq (3.10.9)
The solution procedure is now iterative since the coefficient of matrix R depends on
u. We use equation (3.10.9) to develop an iterative scheme in the form
with ˆ GQ
R (k ) HU
ˆ F 1 (C U ( k ) )
1
After applying the boundarys condition to equation (3.10.10), the system of equations
is of the form
A ( k ) x ( k 1) y ( k ) (3.10.11)
and we stop when Max xi( k 1) xi( k ) , for some tolerance, .
i
GRADRM is modified using the idea of this direct iteration process as described
above. We test the program on some problems whose domain contains corners in the
following examples.
2 u 2u 3 (3.10.12)
in a square plate 1 x 5 , 1 y 5
1
with solution u
x
We partition the boundary into 16 elements with 9 internal points. The boundary
conditions for the function and normal derivative are shown in Figure 3.10.1.
109
q0
13 12 11 10 9
(1,5) (5,5)
14 8
1 23 24 25 1
u 15 7 u
x x
20 21 22
16
6
17 18 19
(1,1) (5,1)
1 2 3 4 5
q0
With a tolerance 0.01, the convergence of the solution using Linear was obtained in 8
iterations. On the other hand, the solution using TPS was obtained in 9 iterations. The
internal solution is compared with the exact solution in Table 3.10.1.
Although the number of iterations for TPS is greater than for Linear, we can see from
Table 3.10.1 that the TPS solution is better than the Linear solution at almost every
point. The normal derivative at the boundary nodes is shown in Table 3.10.2.
110
Table 3.10.2 The normal derivative at the boundary nodes
We observed from Table 3.10.2 that the normal derivative for TPS is less than for
Linear at some corners such as (5,1) and (5,5). However, the solution using TPS
should be better as the number of elements increases. For the reason we have
discussed in section 3.7.3.
2 u 2u 3 (3.10.13)
on the domain as used in Example 3.10.1 with the Dirichlet boundary condition
1
u . (3.10.14)
x
With the tolerance 0.01, the convergence of the solution for both radial basis
functions was obtained in 5 iterations. The internal solution is compared with the
exact solution as shown in Table 3.10.3.
111
As for the previous problem, we see from Table 3.10.3 that the TPS solution is better
than the Linear solution. The normal derivatives on x 1 and x 5 are shown in
Table 3.10.4.
For the Dirichlet problem, we can see from Table 3.10.4 that the normal derivative
from TPS is still more accurate compared with that from Linear. The normal
derivative on the boundary is shown is Figure 3.10.2.
0.2
0
Normal derivative
-0.2
-0.4
-0.6
-0.8
-1
-1.2
-1.4
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
112
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
We can see from Figure 3.10.2 that the normal derivative from GRADRM agrees well
with the exact solution. The normal derivative using TPS is slightly better than that
using Linear.
u
16. 3.10.2 The case 2 u c2 ( x, y)u
x
113
(H R)u Gq (3.10.18)
where
ˆ GQ
R HU
x
ˆ F 1U F F 1 (3.10.19)
GRADRM is modified to solve the non-linear problem using a direct iteration method
as described above. The program is tested with the following example.
(1,0) 1 9 (5,0)
2 8
3 7
4 6
5
(3,-1)
2.50
Normal derivative
2.00
1.50
1.00
0.50
0.00
-0.50 114
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 1
The internal solution is compared with the method in Partridge and Brebbia (1992)
and the exact solution in Table 3.10.5.
We see from Table 3.10.5 that the GRADRM solution using Linear is the same as the
Partridge solution using Linear because the boundary is smooth. The GRADRM result
using TPS is slightly worse than that of using Linear. This situation occurs as in
Example 3.7.2 case (b) and (c). The solution using the TPS should be better when the
number of elements increases as the RMS errors shown in Table 3.10.6.
Table 3.10.6 RMS error of internal solutions from GRADRM using 17 internal points
associated with 16, 32 and 64 elements.
16 0.00203 0.00718
32 0.00312 0.00347
64 0.00355 0.00198
115
Although this problem does not use uniform meshes, as discussed in section 3.4, the
solution from TPS converges to the exact solution as the number of elements
increases. On the other hand, the Linear solution seem to be worse for this situation.
The problem considered so far are of only moderate size but we require of the order
of 10 iterations for convergence i.e. we must solve a set of linear solution ten times.
We now investigate Newton's method with a view to reducing the number of
iterations.
Solving non-linear problem needs iteration methods. The previous method uses only
the value of the right hand side function on each iteration. Newton's method offers
potentially faster convergence because it also uses the first derivative of the function.
Hence this method is applicable only to problems when the function is differentiable.
We discuss briefly the algorithm of the method in this section. For further
applications, see the references in Bartholomew-Biggs (1998). A recent application to
DRM can be found in Bulgakov, Sarler and Kuhn (1998). We discuss four examples
in this section. The first three were examined on the problem in the previous section
and the number of iterations will be investigated again. The last, new one, is
concerned with the non-linear Poisson in which the inhomogeneous term contains the
square root of a partial derivative. This example is considered in Xu and Kamiya
(1999).
g ( x) 0 (3.11.1)
116
If we assume that x is close to x * , and hence that is small, then we can neglect the
higher order terms in (3.11.2) and get
which leads to
J 1 (x)g(x) (3.11.5)
From (3.11.5) we obtain the basic Newton iteration method to solve (3.11.1) as
x ( k 1) x ( k ) ( k ) (3.11.6)
We are ready to discuss how to apply the algorithm to the DRM formulation.
Equations (3.6.3) to (3.6.5) give
ˆ GQ
Hu Gq HU
ˆ F 1b (3.11.7)
Letting
ˆ F 1
ˆ GQ
S HU (3.11.8)
which leads to
Hu Gq Sb (3.11.9)
g(u, q) Hu Gq Sb 0 (3.11.10)
g i
Gij (3.11.11)
q j
117
g i b
H ij S i (3.11.12)
u j u j
We develop GRADRM using this definition of the Jacobian matrix to reform the
Newton's iteration method to solve (3.11.10) and discuss computational results in the
following part.
Consider the problem as discussed in Example 3.10.1. For the tolerance value 0.01,
The internal solutions are shown in Table 3.11.1.
We see from Table 3.11.1 that the number of iterations is substantially reduced for
both radial basis functions. Furthermore, for the TPS result, GRADRM(NT) is more
accurate than GRADRM. The normal derivative at the boundary is shown in Table
3.11.2.
For the tolerance value 0.00001, the internal solution is shown in Table 3.11.3.
We see from Table 3.11.3 that the number of iteration is substantially reduced. The
internal solution using GRADRM(NT) and that of the old one are the same, as would
be expected. We observe from Table 3.11.1 and 3.11.3 that GRADRM(NT) with
tolerance 0.01 gives nearly the same result as the old one with tolerance 0.00001.
This implies that the Newton method helps both accuracy and iteration.
The normal derivative from GRADRM(NT) is the same as that from GRADRM as
shown in Table 3.11.4.
Table 3.11.4 The normal derivative from GRADRM(NT) with tolerance 0.00001
119
Example 3.11.2 Dirichlet Problem
Consider the problem as discussed in Example 3.10.2. For the tolerance value 0.01,
The internal solution is shown in Table 3.11.5.
We see from Table 3.11.5 that the number of iterations is reduced. Furthermore, the
internal solution using GRADRM(NT) is better than that of the old one.
We see from Table 3.11.6 that the normal derivative using GRADRM(NT) is also
usually better than that of the old one.
We see from Table 3.11.7 that the number of iterations is substantially reduced. At
each point, the internal solution using GRADRM(NT) and that of the old one
converge to the same value. Furthermore, as for the previous example, we observe
from Table 3.11.5 and 3.11.7 that GRADRM(NT) with tolerance 0.01 gives the same
result as the old one with tolerance 0.00001. This indicates that the Newton's iteration
method is an efficient tool for solving non-linear equations. The normal derivative
from GRADRM(NT) is the same as that from GRADRM and shown in Table 3.11.8.
120
For the tolerance value 0.00001, the internal solution is shown in Table 3.11.7.
Table 3.11.8 The normal derivative from GRADRM(NT) with tolerance 0.00001
For this Dirichlet problem, the normal derivative on the boundary agrees well with the
exact solution as shown in Figure 3.11.1. We can see that the result from TPS is
slightly better than that from Linear as expected.
0.2
0
Normal derivative
-0.2
-0.4
-0.6
-0.8
-1
-1.2
-1.4
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Linear
121 TPS Exact
Consider the problem as discussed in Example 3.10.3. For the tolerance value 0.01,
The internal solution is shown in Table 3.11.9.
We see from Table 3.11.9 that the number of iterations is reduced. The internal
solution using the GRADRM(NT) and the old one are the same. Therefore
GRADRM(NT)is more efficient because its iteration is less but give the same result.
The normal derivative on the boundary also agrees well with the exact solution as
shown in Figure 3.11.2.
2.50
Normal derivative
2.00
1.50
1.00
0.50
0.00
-0.50
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 1
This example is referred to the example in Xu and Kamiya (1999). Consider the
Poisson-type equation
2 u 6 y u x (3.11.13)
(1,2) 1 31 (2,2)
(1,1) 11 21 (2,1)
The boundary conditions, which come from the exact solution can be expressed as
u 3x 3 y (3.11.14)
The boundaries perpendicular to the x axis are set to Dirichlet conditions and those
perpendicular to the y axis to Neumann conditions. The problem is tested using the
GRADRM(NT). The process converged within four iterations for the convergence
tolerance 0.001. The solution along the boundary agrees well with the exact solution
as shown in Figure 3.11.4.
60
50
40
Solution
30
20
10 123
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
The normal derivative on the boundary also agrees well with the exact value as shown
in Figure 3.11.5.
80
Normal derivative
60
40
20
-20
-40
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
Exact GRADRM
124
We see from Table 3.11.10 that the solutions using Linear are slightly better than
those using TPS at only some points. However the RMS error of the linear solution
is 0.0149 but the RMS error of the TPS solution is 0.0098. Hence it is fair to say that
the augmented thin plate spline is better than the linear function f 1 r .
So far, the Newton iteration method works well with Poisson-type problems in which
the non-homogeneous term depends on the problem variable. It is our purpose to
study another case in which an unknown parameter appears in the equation.
GRADRM will be developed for solving the non-linear problem which both of these
situations appeared. It is intended to study in the future work.
We have developed a variety of programs to solve various problems both linear and
non-linear. The development is shown in Figure 1 of the preface. The list of the
problems relative to the programs is also presented in the preface.
In linear problems, we begin with Laplace's equation. The existing program LINBEM
using the standard collocation method works well with smooth boundary problems.
However, for corner and discontinuous boundary condition problems, LIBEM gives
poor solution at the corners and the points with discontinuous boundary condition.
MULBEM, using the multiple node method with auxiliary collocation, can resolve
such problems. GRABEM is an alternative to resolve such problems but it is more
complicated from the programming point of view. However it is suitable to apply in
the dual reciprocity method.
For Poisson's equation, with the right-hand side function of position, there are two
ways to handle this problem. The first one uses the idea of transformation from
Poisson's equation to Laplace's equation. In this case, MULBEM is modified to
MULDRM. The second one uses the direct dual reciprocity method. In this case the
DRBEM1 is implemented. For problems in which the boundary N , MULDRM
works well with corner problems. On the other hand DRBEM1 needs to be improved.
However, if N the normal derivative from DRBEM1 is better than that from
MULDRM.
125
In non-linear problems, GRADRM is modified using direct iteration method. The
number of iteration process is not good enough. The Newton iteration method is
introduced to reduce such problem. In this way, GRADRM is developed to
GRADRM(NT) using the Newton iteration method. GRADRM(NT) gives
encouraging result, for non-linear problems.
For radial basis functions we have shown that the augmented thin plate spline
f r 2 log r a bx cy works better than the linear function f 1 r .
126
4. Future Work
Non-linear problems in which the non-homogeneous term contains both the problem
variable and a parameter needs a double iteration process, one for the variable and one
for the parameter. The present program needs to be developed to solve such problems.
2 u f ( x, y, u ) ( x, y ) (4.1)
u
a bu 0 ( x, y) (4.2)
x
The solution of such problem depends on the value and its relation to the
eigenvalues of the operator 2 over . In fact the linear operator has a discrete
spectrum of non-negative eigenvalues. However, the non-linear problem has a
continuous spectrum, a solution existing only for values of such that 0 c ,
where c is the principal eigenvalue of 2 , N.B. the eigenvalues depend on the
region and the values of a and b. The DRM can be used to find the critical value,
c , together with the solution u.
2 u e u ( x, y) (4.3)
u0 ( x, y ) (4.4)
The dual reciprocity method find the boundary solution and the internal solution at
the same time. The system of equations obtained by the method is quite large. The
multi-partition method for boundary element problems is a suitable domain
decomposition strategy to map the boundary element method onto an MIMD parallel
architecture such as the nCUBE. This approach has been successful in a variety of
situations (Davies, et al, 1996, Davies and Mushtaq, 1996, Crann, et al, 1998, Davies,
et al, 2000). Our future work is to develop the program using the dual reciprocity
method in a multi-partition context.
127
Non-linear material problems will be considered in a parallel environment. The
typical equation to be considered is
. ( ) 0 (4.6)
128
References
Alarcón, E. Martin, A. and Paris (1979) Boundary elements in potential and elasticity
theory. Computers & Structures 10, 351-362.
Anderson, C.A. and Zienkiewicz, O.C. (1974) Spontaneous ignition: Finite element
solutions for steady and transient conditions. Transactions of the ASME 8,
398-404.
Brebbia, C.A. and Dominguez, J. (1977) Boundary element methods for potential
problems. Applied Mathematical Modelling 1, 7
Brebbia, C.A. (1978) The boundary element method for engineers, Pentech Press.
Brebbia, C.A. and Niku, S.M. (1988) Conforming versus non-conforming boundary
elements in three-dimensional elastostatics: Letter to the editor. International
Journal for Numerical Methods in Engineering 26, 283-287.
Brown, J.W. and Churchill, R.V. (1993) Fourier series and boundary value
problems, 5 Ed. Highstown: McGraw-Hill.
Chan, C.L. and Chandra, A. (1991) An algorithm for handling corners in the
boundary element method: Application to conduction-convection equations.
Applied Mathematical Modelling 15, 244-255.
Chen, C.S. and Rashed, Y.F. (1998) Evaluation of thin plate spline based particular
solutions for Helmholtz-type operators for the DRM. Mechanic Research
Communication 25, 195-201.
Ciskawski, R.D. and Brebbia, C.A. (1991) Boundary element method in acoustics,
Addison-Wesley.
Crann, D., Davies, A.J. and Mushtaq (1998) Parallel Laplace transform boundary
element methods for diffusion problems. Computational Mechanics
Publications , 259-268.
129
Cruse, T.A. (1969) Numerical solution in three dimensional elastostatics. Int. J. Solids
and Structures 5, 1259-1374.
Davies, A.J. (1980) The finite element method: A first approach, Oxford University
Press.
Davies, A.J. and Crann, D. (1996) Using a spreadsheet for the boundary element
method: Interior and exterior problems. Technical Report 13, Mathematics
Department: University of Hertfordshire.
Davies, A.J. and Mushtaq, J (1996) The domain decomposition boundary element
method on a network of transputers. Computational Mechanics Publications,
397-406.
Davies, A.J., Crann, D. and Mushtaq, J (2000) A parallel Laplace transform method
for diffusion problems with discontinuous boundary conditions. Application of
High-performance Computers in Engineering VI, 3-10
Gerald, C.F. and Wheatley, P.O. (1994) Applied numerical analysis, Addison-
Wesley.
Gilbert, R.P. and Howard, H.C. (1990) Ordinary and partial differential equations
with applications, West Sussex: Ellis Horwood Limited.
Gipson, G.S. (1987) Boundary element fundamentals: Basic concept and recent
developments in the Poisson'equation, Southampton: Computational
Mechanics Publishers.
Golberg, M.A. (1994) The theory of radial basis functions applied to the BEM for
inhomogeneous partial differential equations. Boundary Elements
Communications 5, 57-61.
Golberg, M.A. (1995) The numerical evaluation of particular solutions in the BEM:
A review. Boundary Elements Communications 6, 99-106.
Golberg, M.A. (1996) The method of fundamental solutions for Poisson's equation.
Engineering Analysis with Boundary Elements 16, 205-213.
Golberg, M.A., Chen, C.S., Bowman, H. and Power, H. (1998) Some comments on
the use of radial basis functions in the dual reciprocity method.
Computational Mechanics 22, 61-69.
Gray, L.J. and Lutz, E. (1990) On the treatment of corners in the boundary element
method. Journal of Computational and Applied Mathematics 32, 369-386.
Hall, W.S. (1994) The boundary element method, Dordrecht: Kluwer Academic
Publishers.
130
Hess, J. and Smith, A.M.O. (1964) Calculation of non-lifting potential flow about
arbitrary three-dimension bodies. Journal of Ship Research 8, 22-44.
Humi, M. and Miller, W.B. (1992) Boundary value problems and partial differential
equation, PWS-KENT.
Ingham, D.B. and Yuan, Y. (1994) The boundary element method for solving
improperly posed problems, Southampton: Computational Mechanics
Publications.
Jaswon, M.A. (1963) Integral equation methods in potential theory I. Proc. Roy. Soc.
275, 23-32.
Karur, S.R. and Ramachandran, P.A. (1994) Radial basis function approximation in
the dual reciprocity method. Mathematical and Computer Modelling 20, 59-
70.
Mitra, A.K. and Ingber, M.S. (1993) A multiple-node method to resolve the
difficulties in the boundary integral equation method caused by coners and
discontinuous boundary conditions. International Journal for Numerical
Methods in Engineering 36, 1735-1746.
Niku, S.M. and Brebbia, C.A. (1988) Dual reciprocity boundary element formulation
for potential problems with arbitrarily distributed sources. Engineering
Analysis 5, 46-48.
Paris, F., Martin, A. and Alarcon, E. (1981) Potential theory. Progress in boundary
element method 1, 45-83.
Partridge, P.W. and Brebbia, C.A. (1989) Computer implementation of the BEM dual
reciprocity method for the solution of Poisson type equations. Software for
Engineering Workstations 5, 199-207.
131
Partridge, P.W. and Brebbia, C.A. (1990) Computer implementation of the BEM dual
reciprocity method for the solution of general field equations.
Communications in Applied Numerical Methods 6, 83-92.
Partridge, P.W. and Wrobel, L.C. (1990) The dual reciprocity boundary element
method for spontaneous ignition. International Journal for Numerical
Methods in Engineering 30, 953-963.
Partridge, P.W., Brebbia, C.A. and Wrobel, L.C. (1992) The dual reciprocity
boundary element method, Southampton: Computational Mechanics
Publications.
Partridge, P.W. (1994) Dual reciprocity BEM: Local versus global approximation
functions for diffusion, convection and other problems. Engineering Analysis
with Boundary Elements 14, 349-356.
Partridge, P.W. and Sensale, B. (1997) Hybrid approximation functions in the dual
reciprocity boundary element method. Communications in Numerical
Methods in Engineering 13, 83-94.
Powell, M.J.D. (1994) The uniform convergence of thin plate spline interpolation in
two dimensions. Numerical Mathematics 68, 107-128.
Power, H. and Casares Long, J.J. (1997) Applications of high performance computing
in engineering V, Southampton: Computational Mechanics Publications.
Rizzo, F.J. and Shippy, D.J. (1970) A method of solution for certain problems of
transient heat conduction. AIAA Journal 8, 2004-2009.
Shaw, R.P. (1970) An integral equation approach to acoustic radiation and scattering.
Topics in Ocean Engineering II, 143-163, Gulf Publishing Co., Houston.
Subia, S.R. and Ingber, M.S. (1995) A comparison of the semi-discontinuous element
and multiple node with auxiliary collocation approaches for the boundary
element method. Engineering Analysis with Boundary Elements 15, 19-27.
Symm, G.T. (1963) Integral equation methods in potential theory II. Proc. Roy. Soc.
275, 33-46.
Walker, S.P. and Fenner, R.T. (1989) Treatment of corners in BIE analysis of
potential problems. International Journal for Numerical Methods in
Engineering 28, 2569-2581.
132
Weinberger, H.F. (1965) A first course in partial differential equations (Lexington:
Xerox)
Wrobel, L.C. and Brebbia, C.A. (1987) The dual reciprocity boundary element
formulation for nonlinear diffusion problems. Computer Methods in Applied
Mechanics and Engineering 65, 147-164.
Xu, S.Q. and Kamiya, N. (1999) A formulation and solution for boundary element
analysis of inhomogeneous-nonlinear problem: The case involving derivatives
of unknown function. Engineering Analysis with Boundary Elements 23,
391-397.
======================================
133