Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

This article appeared in a journal published by Elsevier.

The attached
copy is furnished to the author for internal non-commercial research
and education use, including for instruction at the authors institution
and sharing with colleagues.
Other uses, including reproduction and distribution, or selling or
licensing copies, or posting to personal, institutional or third party
websites are prohibited.
In most cases authors are permitted to post their version of the
article (e.g. in Word or Tex form) to their personal website or
institutional repository. Authors requiring further information
regarding Elseviers archiving and manuscript policies are
encouraged to visit:
http://www.elsevier.com/copyright

Author's personal copy


ARTICLE IN PRESS
Engineering Analysis with Boundary Elements 33 (2009) 10621073

Contents lists available at ScienceDirect

Engineering Analysis with Boundary Elements


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

On the enriched RBF method for singular potential problems


Francisco Bernal , Manuel Kindelan
n Institute of Modeling, Simulation and Industrial Mathematics, Universidad Carlos III, 28911 Leganes, Madrid, Spain
G. Milla

a r t i c l e in f o

a b s t r a c t

Article history:
Received 14 October 2008
Accepted 5 March 2009
Available online 17 April 2009

The performance of Kansas method in the solution of elliptic partial differential equations (PDEs) with
singular boundary conditions is addressed. Like in all global numerical schemes, low-order singularities
bring about Gibbs oscillations that deteriorate the accuracy and the convergence rate of Kansas method
to a great extent. Moreover, they may render it uncapable of handling common problems of
incompressible ow. Focussing on a problem of Laplacian ow which is linked to the benchmark Motz
problem, it is shown how all these difculties can be overcome by enriching the radial basis function
(RBF) interpolant with the proper singularity-capturing terms. This simple modication even enables
Kansas method to outperform the nite element method (FEM) in the conservation of Laplacian ow
through an irregular channel.
& 2009 Elsevier Ltd. All rights reserved.

Keywords:
Radial basis function
Kansas method
Meshless
Motz problem
Singular elliptic PDE

1. Introduction
Currently, the method of choice for solving partial differential
equations (PDEs) over non-regular shaped domains is the method
of the nite element (FEM). As an alternative to it, collocation
meshless methods are becoming an increasingly used approach.
In some of them, the approximate solution takes the form of
an expansion of radial basis functions (RBFs) over a scattered
discretization of the domain under consideration, and therefore
no partitioning of the latter into a mesh is requiredwhich often
is a non-trivial and time-consuming preprocessing step. Moreover,
certain problemssuch as those involving free boundaries
or crack propagationare inherently hard to reconcile with the
notion of a structured numerical support, and a mesh or a grid
becomes even more of a burden. In these cases, a meshless
description in terms of free nodes which may be easily generated
or moved is understandably appealing.
The rst and most popular collocation scheme with RBFs for
solving PDEs was introduced by Kansa in 1990 [16,17], and is
usually known as Kansas, or the RBF, method. Kansas idea is to
modify the RBF interpolation scheme over scattered nodes [13] in
order to solve PDEs. By derivating the RBFs in the RBF interpolant,
the individual numerical derivatives which appear in the PDE may
be computed, and the interpolant coefcients are solved for with
collocation. Since the PDE is solved in strong form, no numerical
integrations are needed, and Kansas method is truly meshless.
There are further advantages to Kansas method than geometric exibility and ease of implementation. Not only is the

 Corresponding author.

E-mail address: fco_bernal@hotmail.com (F. Bernal).


0955-7997/$ - see front matter & 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.enganabound.2009.03.002

approximate solution (which we will call the RBF solution from


now on) continuous throughout the domain, but also its
derivatives, overcoming the ambiguity which occurs along interelement sides in the FEM. Because of their radial dependence,
the RBFs are blind to the domain dimension, and free of the
dimensionality curse which befalls traditional schemes [8]. Most
importantly, many popular RBFssuch as the multiquadric (MQ)
or the Gaussianmay exhibit spectral convergence in the solution
of elliptic PDEs [10]. Therefore, with a numerical discretization
whose size is much smaller than those required by nite
differences (FD) or FEM, Kansas method may achieve under
regular assumptions the same accuracy, since the former methods
enjoy just algebraic convergence.
In fact, Kansas method has been reported to outperform FEM
in the solution of PDEs [23]at least, as long as the solution of the
PDE is smooth. Research of both theoretical and implementation
aspects of the RBF method carries on at an increasing pace, and,
among many other applications, it has been used to simulate
uid-mechanical free-boundary problems [3,5]. However, there
are fewer papers devoted to this kind of problems than might be
expected, given that it is one of the obvious targets for meshless
methods. According to our own experience in the simulation of
incompressible ows with Kansas method, this is partly due to
the difculty of any numerical method in strong form to enforce
mass conservation. While a Galerkin-type formulation such as
FEM implicitly enforces mass conservation by integrating the
divergence of the velocity eld across each element, there is no
such a safeguard in the RBF scheme.
Moreover, uid-mechanical problems are typically modeled by
PDEs whose boundary conditions (BCs) may vary non-smoothly
along the boundary. For instance, in the benchmark problem of
the Stokes ow in a lid-driven box, the velocity eld exhibits a

Author's personal copy


ARTICLE IN PRESS
F. Bernal, M. Kindelan / Engineering Analysis with Boundary Elements 33 (2009) 10621073

jump discontinuity at both ends of the lid [9]. In other problems


(such as those addressed in this paper), there may occur
singularities of a higher order. Typically, in points where the
nature of the BC changes abruptly from Neumann to Dirichlet, the
solution may be continuous and yet exhibit a sharp feature.
Another common source of singularity in problems of engineering
interest is the non-smoothness of the domain boundary itself, i.e.
the presence of re-entrant corners.
Unfortunately, boundary singularities pose challenging difculties to all kinds of numerical methods, regardless of whether
their formulation is weak or strong. Discontinuities, for instance,
trigger the well-known Gibbs oscillatory phenomenon [21], and
deteriorate the performance of the used method, very especially if
it has global support. Since the BCs are an inherent part of the PDE
model and cannot be disposed of, the chosen numerical scheme
must be ableor modied in such a wayto cope with them.
Regarding Kansas method, the accuracy of the RBF interpolant
is dramatically reduced compared with the results achieved in
elliptic PDEs whose solution is smooth. Moreover, one of the key
features of Kansas method, namely its spectral convergence, is
entirely lost.
This paper is concerned with the application of Kansas method
to singular incompressible ow problems. The remainder of it is
organized as follows. Section 2 surveys precedents in the context
of collocation meshless methods, which are mostly based in the
enrichment of the RBF interpolant with singularity-capturing
terms of one sort or another, as we will also do. In Section 3, the
solution of the benchmark Motz problem by Kansas method is
discussed and the analytical term required is constructed. As we
shall see, not only does the addition of this term greatly improve
the accuracy of the RBF solution, but, more importantly, spectral
convergence is restored, which was lost because of the presence of
the singularity. Section 4 reports on how the same idea recovers
the harmonicity of the RBF solution of a Laplacian ow problem in
a highly irregular domain. Finally, Section 5 concludes the paper.

2. RBF interpolation of non-smooth functions


The reason why the RBF method is inaccurate when dealing
with this type of problems is the inability of the RBF expansion to
capture the discontinuities of the solution or of its low-order
derivatives. In fact, the RBF method fails to yield optimalor even
acceptableresults whenever the exact solution exhibits a sharp
feature that does not belong to the interpolation space of the
smooth RBF used.
In the context of standard (meshed) numerical methods, the
usual approach to overcome these difculties is grid renement in
the vicinity of the singularity. However, adaptive grid renement
schemes cause a signicant increase in computational cost and
their efciency is not always satisfactory. A better alternative is to
incorporate analytical information into the numerical scheme
about the singularities of the problem. This notion has already a
number of precedents in the literature on meshless collocation.
When solving a crack propagation problem, Alves et al. [1]
improved the performance of the Method of the Fundamental
Solution (MFS) by enriching the RBF basis with the lowest terms
of the series expansion of the locally exact solution close to the
crack tip. The same ideas had been used before in the related
Element-Free Galerkin Method (EFGM) (see [4]). Notice that in
this kind of problems, the singularity takes place inside the
domain rather that on its boundary. A similar situation is the loss
of accuracy undergone by the RBF approximation in the vicinity
of a re-entrant corner in the eigenvalue problem analyzed by
Platte and Driscoll [27]. There, the exact solution has also singular
behavior, due in this case to the non-smoothness of the boundary.

1063

They also did not observe improvement either through boundary


collocation of the PDE nor through node clustering around the
singularity. Instead, they restored convergence by enlarging the
interpolation space with the lower powers of the series expansion
of the corresponding eigenfunction. Hu et al. [15] used similar
ideas for the radial basis collocation method (RBCM) with inverse
multiquadrics (IMQRB) and Gaussian (GRB) radial basis functions.
In their work, collocation is treated as a RitzGalerkin method
involving integration approximation.
Other authors have tackled singular BVPs with meshfree
methods without resorting to analytic information. Jung [20]
considers the multiquadric interpolation of 1D discontinuous
functions, and achieves a remarkable improvement by closely
placing a center at either side of the discontinuity and making
both shape parameters zero. Notice that the interpolant is actually
enriched with two non-smooth terms, albeit not problemdependent. While this strategy can be applied to differential
equations as well, its extension to more dimensions is ambiguous.
Finally, Chantasiriwan [9] showed that the global support of
Kansas method amplies the effect of the oscillations brought
about by boundary singularities. Instead, he suggests to use
locally-supported RBF collocation for these kind of problems.
Concretely, he discusses the use of the local RBF-based differential
quadrature method (LRBFDQ), introduced by Shu et al. [32]. The
main advantage of this approach is that the resulting system of
equations is sparse, and does not suffer from ill-conditioning
while enjoying just algebraic error convergence rates. Thus, it
may be used to solve large 2D or 3D problems. As shown by
Chantasiriwan, the LRBFDQ is also preferable to Kansas method
for singular BVPs in terms of accuracy and stability. However, if
the analytic form of the singularity is available, the enriched
version of Kansas method (which will be explained next)
outperforms the LRBFDQ method, as reported in [6]. The situation
when there are singular lines or planes rather than isolated
singularities has been less explored in meshfree methods. In the
EFGM, Rabczuk and Belytschko introduced a discontinuous
description of the discontinuity (cracking particles) for handling
a problem of crack dynamics [29]. In RBF collocation, Kansa et al.
[18] have solved a turbulent combustion problem, in which the
ame is a discontinuous (shock) free boundary. They have
modeled it as the product of a Heaviside function in the normal
direction multiplied by a smooth curve in the tangencial direction,
and enriched the RBF interpolation space with it. Their approach
benets from the accuracy and efciency of Kansas method while
avoiding a smeared ame line because of modeling or numerical
artifacts.

2.1. Enriched Kansas method


In this section, we discuss a method to improve the accuracy of
Kansas method in dealing with this kind of problems, which is
based on the addition of analytical information into the numerical
scheme in order to capture the discontinuity. For the sake of the
argument, consider a general linear, 2D, elliptic PDE as

Lu~
x 0;

~
x2O

(1)

Bu~
x 0;

~
x 2 @O

(2)

Kansas method discretizes O and @O into NI and NB centers,


x fjj~
x ~
xj jj; j 1; . . . ; N I N B
respectively, places an RBF fj ~
on each of them, and looks for an approximation to u~
x as
u

NX
I N B
j1

aj fj ~
x

(3)

Author's personal copy


ARTICLE IN PRESS
1064

F. Bernal, M. Kindelan / Engineering Analysis with Boundary Elements 33 (2009) 10621073

Enforcement of the PDE and the BCs on NI NB collocation nodes


(which are usually the same set of RBF centers), leads to the
following system of linear equations:
NX
I N B

aj Lfj ~
xi 0 if ~
xi 2 O

(4)

aj Bfj ~
xi 0 if ~
xi 2 @O

(5)

j1

NX
I N B
j1

Inversion of the above system yields the set of coefcients aj ,


which allow to construct the RBF solution anywhere in O [ @O.
Notice, however, that this latter step may be tricky since the
matrix is fully populated (due to the global support of the bestperforming RBFs) and ill-conditioned. Moreover, Schabacks
uncertainty principle postulates a trade-off between accuracy
and stability [31]. An alternative to strict collocationwhich leads
to square systemsis to solve (4) and (5) in a least squares sense.
In that case, the PDE and/or the BCs should be sampled on more
than NI NB points. While overdetermined systems may be
preferable (for instance, if the data is known to be contaminated
by noise), in this paper we have restricted ourselves to square
systems, which is also the original and most usual approach.
Another well-known drawback of RBF collocation is the
relative inaccuracy of the RBF interpolant (3) near the boundaries
(specially those with derivative BCs). This can be partly alleviated
through the so-called PDEBC strategy [12], which consists in
enforcing both the PDE and the corresponding BC on each of the
boundary collocation nodes. Since there are N B new collocation
equations, the interpolant (3) must be supplemented with
a matching number of RBF centers, in order to keep the system
xNI NB 1 ; . . . ; ~
xNI 2NB g are
square. Such NB extra RBF centers f~
usually arranged at a xed distance off the boundary and outside
the domain.
The precedent scheme yields poor approximations to solutions
u which contain boundary singularities, unless the interpolation
space be equipped to reproduce them. For the sake of simplicity,
assume that there is a single singularity at the origin, r; y are the
polar coordinates centered at it, and let fmk r; y; k 1; . . . ; ngbe a
set of n special functions designed to capture the required
behavior at the origin. The enriched RBF interpolant (with extra
centers for PDEBC) is
u~
x

NIX
2NB

aj fj ~
x

j1

n
X

aNI 2NB k mk r; y

(6)

NIX
2NB

ai mk ri ; yi 0 k 1; . . . ; n

(11)

i1

where r i ; yi are the polar coordinates of the ith RBF center. It is


straightforward to extend this approach to the case where several
singularities lie along the boundary (as in Fig. 9).
Notice that the mk r; y may be used in different domains as
long as the PDE to be solved and the exact form of the singularity
are the same. However, instead of using a series expansion of the
locally exact solution like in most enriched-meshed methods, we
just add singular terms into the RBF interpolant which absorb the
local singular behavior of the solution, and let the RBFs interpolate
the smooth remainder of it. Far away from the origin, the special
functions mk r; y may depart from the BCs, but the exibility of
the RBFs makes up for it. Therefore, the basic idea is to remove the
non-smooth features in u~
x by means of the mk r; y terms,
implicitly recasting the problem into a smooth one which is
suitable for the RBFs. Moreover, we do not restrict the action
of the special functions to a subdomain around the singularity,
but allow them to act in the whole domain O, thus eliminating the
need for domain decomposition and solution matching which is
characteristic of the enriched traditional methods.
Nonetheless, since in general the inclusion of the singular
terms in the RBF interpolant will only be benecial in the portion
of the domain surrounding the associated singularity, it might be
convenient to modify the scheme in such a way that the
augmenting terms are conned to it. At least two approaches
can be envisaged:
1. Domain Decomposition Methods (DDMs). One of them consists in partitioning the geometry into overlapping subdomains
and employing the classical Schwarzs alternating algorithm
(explained in Section 4) for solving the PDE. In this approach,
the singular terms associated to each singularity only exist in
the subdomain containing it (see [1]).
2. Matching the mk r; y with a spline that vanishes at a nite
distance. The problem with this approach is that such a spline
need not satisfy L and therefore may produce residual. For
this reason we have preferred the DDM approach in our
experiments.
3. The Motz problem
3.1. Introduction

k1

The enriching terms must comply with the differential operators


L (globally) and B (in a vicinity d of it)

Lmk r; y 0;

r40

(7)

Bmk r; y 0;

d4r40

(8)

The Motz problem [26] is a benchmark for testing the


performance of numerical methods in the solution of elliptic
problems with singular boundary conditions. The problem
(as modied by Wait and Mitchell [34]) consists in solving
Laplaces equation in the domain depicted in Fig. 1 (left). In this
problem, the partial derivatives are singular at the origin, where
@u=@x undergoes an innite jump.

In order for the system of collocation equations to be square of


dimension N I 2NB n, n additional equations must be supplemented. We follow [27] in enforcing the same orthogonality
conditions that are required for enriching polynomial terms
whenever they are added into the interpolant. The full system of
collocation equations is then
NIX
2N B

aj Lfj ~
xi 0 if ~
xi 2 O [ @O

(9)

j1

NIX
2N B
j1

aj Bfj ~
xi

n
X
k1

aNI 2NB k Bmk ri ; yi 0 if ~


xi 2 @O

(10)

Fig. 1. The Motz problem (left). Polar coordinates and local BCs close to the
singularity. u0 0 (PDEs in Section 2), 1 and 10 (PDEs in Section 3).

Author's personal copy


ARTICLE IN PRESS
F. Bernal, M. Kindelan / Engineering Analysis with Boundary Elements 33 (2009) 10621073

Absolute Error

105

1010
0

20
N

40

Absolute Error

A1
A5
A10
A15

100

1065

A1
A5
A10
A15

6
20

40
Np

60

Fig. 2. Trefftz method for the Motz problem. Left: absolute error as a function of N (N p 120). Right: absolute error as a function of N p (N 20).

Albeit non-smooth, the solution of the Motz problem is


continuous at the singular point. Moreover, it may be expressed
as a series

 
1
X
2k  1
(12)
ur; y
Ak r 2k1=2 cos
y
2
k1
where r and y are polar coordinates (see Fig. 1, right). The above
solution is valid in the entire solution domain. In [30], the rst
20 coefcients Ak were computed exactly by a conformal mapping
technique, and it was shown that the radius of convergence is at
least 2. Previous work on Motzs problem with meshless methods
can be found in Refs. [15,19,28].
3.2. Boundary method
In Ref. [2], a simple boundary method is proposed to compute
the coefcients, Ai , associated with the asymptotic solution of
elliptic equations having a large convergence radius in the vicinity
of singular points. It follows previous work of Li et al. [22], and
reduces the problem to that of minimizing a discrete functional in
order to approximately satisfy the boundary conditions. Denoting
by GL, GU , GR the left, upper and right boundaries of the domain,
let us dene

2
2
X @u
X @u
X
500  uxi ; yi 2
xi ; yi
xi ; yi
@x
@y
~
~
~
x 2G
x 2G
x 2G
i

(13)
where ux; y; @u=@x, and @u=@y are given by (12) and its derivatives
xi , are
truncated to N terms. In (13) a total of N p 4N nodes, ~
distributed equispaced along GL [ GU [ GR . The coefcients Ai
which minimize this functional for a given N are obtained by
solving the linear system of equations resulting from
@J
0;
@Ai

i 1; 2; . . . ; N

(14)

i.e. by imposing the boundary conditions away from the


singularity in a least square sense.
The left part of Fig. 2 shows the absolute error of 4 selected
coefcients, as a function of N, for a xed number of boundary
nodes (N p 120). Notice that the method exhibits exponential
convergence rate for each of the coefcients. The right part of Fig. 2
shows the absolute error of the same coefcients as a function of N p ,
for a xed number of terms in the asymptotic expansion (N 20).
In this case there is little benet from increasing the number of
nodes at the boundary. Therefore, using a collocation method
(N p N) instead of residual minimization at the boundary, leads to
comparable results.

3.3. Enriched multiquadric interpolant


According to the general approach described in Section 2, we
seek to enrich the RBF interpolant with a suitable term m1 r; y
which captures the derivative discontinuity at the origin. Given
the simplicity of the problem, formulas (7) and (8) can be readily
solved analytically
@2 m1 1 @m1 1 @2 m1
2

0
r @r
@r 2
r @y2

(15)

@m1
r; y 0 0
@y

(16)

m1 r; y p 0

(17)

Because (15)(17) is not well posed, its solution is not unique.


Not surprisingly, the Motz functions

 
2k  1
f k r; y r 2k1=2 cos
(18)
y ; kX1
2
which appear in the exact solution (12) comply with the PDE (15)
and the local BCs at the origin (16) and (17), and are therefore
eligible as the singular enriching functions mk r; y in (6). Notice
that each f k has a singularity of order k at the origin. Therefore, in
order to reproduce the derivative jump, we just pick a single
enriching function, namely the rst one
p
m1 r; y f 1 r; y r cosy=2
(19)
In the remainder of the paper we drop the general mk r; y
notation and will refer to the enriching functions for the concrete
Laplacian singularity (15)(17) as f k r; y; k 1; 2; . . .. Among the
many available RBFs (see for instance [11] for a list) we have
chosen the multiquadric
q
(20)
fk ~
x k~
x ~
xk k2 c2k
rst proposed by Hardy [13] and which has been used extensively
both for interpolation and for the solution of PDEs. To specify
the MQ basis (20), one must specify both the centers f~
xk g and
the shape parameters fck g. Regarding the latter, although they
critically affect the accuracy, there is no xed procedure to pick an
optimal set of values. Usually, a single value ck c is used for all
the MQs. As c ! 1, the approximation error steadily drops until
numerical blow-up occursfor the MQ becomes atter and
atter, causing numerical instability.
The MQ centers are arranged into a square grid of constant h
over the rectangular domain in Fig. 1, plus the extra layer of
centers at a distance h off the boundary as required for PDEBC.
To characterize the accuracy of the RBF approximation we use a
scattered set of j 1; . . . ; N ev 5305 evaluation nodes and dene

Author's personal copy


ARTICLE IN PRESS
1066

F. Bernal, M. Kindelan / Engineering Analysis with Boundary Elements 33 (2009) 10621073

450
400
350
300
250
200
150
100
50
0

500
400
300
200
100
0
1
0.5
y

0 1

0.5

0
x

0.5

500
400
300
200
100
0
1
0.5
y

0.5 0

0.5

450
400
350
300
250
200
150
100
50

Fig. 3. Motz problem. Left: MQ-only solution of the Motz problem. Right: enriched solution (MQs first Motz function f 1

150
100
50
0
50
100
1
0.5
y

0
0 1 0.5 x

0.5

80
60
40
20
0
20
40
60
80
100

500
400
300
200
100
0
1
0.5
y

0
0 1 0.5 x

0.5

400
350
300
250
200
150
100
50

Fig. 4. Motz problem. Left: contribution of the MQs to the solution. Right: contribution of the rst Motz function f 1

the root mean square (RMS) of the error  as


v
u
N ev
u 1 X
RMS t
u~
xj  uex ~
xj 2
N ev j1

(21)

where u~
xj is the RBF solution (with and without enriching term)
xj is the reference solution taking the rst 20 terms in
and uex ~
[30].
As expected, the MQ-only interpolant cannot provide a good t
for the non-smooth solution of the PDE and gives rise to Gibbs
oscillations around the origin, as shown on the left side of Fig. 3.
On the contrary, the singularity at the origin is neatly reproduced
if the interpolant is enriched with the term f 1 (right side). Fig. 4
illustrates the role of the two parts of the RBF solution. The Motzs
function f 1 (right) is responsible for tting the singular shape of
the solution surface around the origin, thus removing any nonsmooth features from the remainder MQ expansion. The left side
of Fig. 4 shows that the MQs are now effectively interpolating a
surface which is smooth everywhere. As we shall see later, the fact
that the MQs are in practice dealing with differentiable surfaces
(once the singular term absorbs the non-differentiable points),
leads to the restoration of the exponential convergence law which
is usual for Kansas method. Notice also that the contribution of f 1
(in Fig. 4, right) departs largely from the solution far away from
the origin. Because of the exibility of the multiquadric, the MQ
contribution makes up for it, so that both parts of the interpolant
merge seamlessly into a global approximation which is valid
everywhere.
On the left side of Fig. 5, four different gridlike point sets made
1
1
1
, 15
, 20
,
up of N 96; 291; 586, and 981 nodes (h 15, 10
respectively) are compared. It is clear that the gain in accuracy
for all of them resulting from the enrichment with f 1 amounts at
least half an order of magnitude. Moreover, the MQs f 1 curves
(solid) exhibit convergence as c ! 1, while the MQ-only ones
(dashed) are insensitive to the shape parameter. The addition of f 1
allows to take full advantage of tuning the MQ shape parameter,
so that the respective minima of the solid and dashed curve for
each point set are at least one and a half order of magnitude apart.

p
r cosy=2).

p
r cosy=2.

These results are competitive with FEM, which yields just


RMSFEM 1:71 with a triangulation of N 1405 vertices, and
RMSFEM 0:013 with N 11 578 if the mesh is adaptively
rened.
Further gains in accuracy are obtained if more Motz functions
are included into the RBF interpolant than just f 1 (shown in Fig. 5,
right). Since such functions make up a basis for the exact solution,
the improvement should not be surprising. However, we can see
that, regarding the minimum of each of the curves, there is little
improvement in adding more than three enriching functions. We
remark that, contrary to Trefftz methods, is not only some ff n r; yg
in series (12) that are being collocated, but there are also
multiquadrics that do not belong to the exact solution basis.
A different interpretation is also possible: given that there are
innite singularities in the exact solution of the Motz problemnotice that the kth term in (12) has a discontinuity in the
derivative of the same order, the added f 2 ; f 3 ; . . . ; f M to the
RBF interpolant are in fact removing them, until the effect of the
M 1th singularity is so small that it is no longer perceived by
the MQ basis.
Finally, we point out that the present meshless solution with
the enriched, strict collocation approach yields results comparable
to the methods (adding method and subtracting method) in
Ref. [15], for the Gaussian RBF and the inverse MQ. Notice,
however, that the methods in [15] can only be applied in problems
for which a series solution like (12) exists. As shown in the next
problem, this is not the general case.

3.4. Motz variant with a hole


In order to further explore these ideas, we consider a variant of
the Motz problem where a circular hole of radius 0:25 has been
added to the original domain at 0; 0:5, along which a Dirichlet BC
u 200 applies. A FEM solution of the modied PDE is shown to
1
, right).
the left of Fig. 6, as well as a sample point set (for h  10
Notice that PDEBC is to be enforced also on the interior boundary.
For this variant of the Motz problem, the boundary method in [2]

Author's personal copy


ARTICLE IN PRESS
F. Bernal, M. Kindelan / Engineering Analysis with Boundary Elements 33 (2009) 10621073

1
log10 RMS ()

log10 RMS ()

N = 96
N = 291
N = 586
N = 981

1067

1
0
1

n=0
n=1
n=3
n=5
n = 10

0
1
2
3
4
0.1

2
0

0.2

0.4

0.6

0.8

0.2

0.3

0.4
c

0.5

0.6

0.7

Fig. 5. Left: effect of the rst Motzs function on the accuracy. Right: effect of several Motzs functions (N 981).

500

400

0.8
300

0.6

200

0.4

200

0.2

0
1
y

0.5
0 1

100

0
x

0.5

0.5

Fig. 6. Left: FEM solution of the Motz variant with a circular hole. Right: RBF point set. Notice that the y-axis is stretched.

with a modied functional


X
J0 J
u~
xi  2002

Table 1
Motz variant with a circular hole.

(22)

~
xi 2GH

(where GH is the circular boundary), fails. Notice that along those


rays radiating from the origin which intersect the hole, more
than one BC must be complied with. Seemingly, more basis
functions should be included in the series (12) than in the Motz
problem. For up to 30 terms , the errors (to the reference FEM
solution) are as large as 50, regardless of the number of boundary
collocation points N p . However, the series cannot be truncated
with more than  30 terms, for the large exponents involved lead
to numerical blow-up. On the other hand, the MQ interpolant
enriched through f 1 turns out to deliver a good accuracy
regardless of the presence of the hole. Since we have chosen to
use evenly distributed RBF point sets, the only condition for high
accuracy is that the distance between nodes, h, must be small
enough for the RBF point set to adequately sample the circular
boundary. This can be seen in Fig. 8 (right), where the coarse point
set with h  10 fails to provide adequate resolution. Since in this
problem the Motz functions do not provide a basis for the exact
solution, there is no benet in adding more enriching functions
than f 1 (n 1) to the RBF interpolant, as shown in Table 1.
Therefore, as long as the singularity remains the same, the
performance of the enriched RBF method is rather unaffected by
the shape of the remainder of the domain.

3.5. Restoration of spectral convergence


The removal of non-smooth features from the MQ contribution
in the enriched interpolant should convey the restoration of the
well-known exponential h2c convergence established for RBF
collocation. Madych [25] gave an error estimate for multiquadrics

n0

n1

n2

n3

n5

0.0504
0.0713
0.1008
0.1234
0.1425

1.2774
1.1993
1.0954
1.0351
1.3103

0.2732
0.1391
0.0646
0.0376
0.0322

0.2334
0.1200
0.0608
0.0393
0.0330

0.2235
0.1199
0.0616
0.0398
0.0309

0.2102
0.1193
0.0604
0.0383
0.0305

Effect of the number of enriching functions f k r; y; k 0; . . . ; n on RMS for N


1969 and increasing c.

c=h

as Oeac l , where h is the mesh constant, c the shape parameter,


and a and 0olo1 are constants. This result has a great impact,
because it implies that the accuracy can be increased through two
independent ways: either by rening the mesh (thus increasing
the computational cost), or by simply increasing the shape
parameter cuntil the onset of ill-condition dictates in practice
a minimum of the error at some nite value of the shape
parameter. In the presence of singular points, however, this
pattern is no longer valid. Indeed, the standard Kansas method
for the Motz problem loses any c-convergence as it is shown by
the dashed curves in Fig. 5 (left). Once f 1 is included (solid line),
the error curves become smooth again, and decrease as c ! 1,
featuring a neat minimum in the case N 291.
Although Madychs results [25] were derived for the interpolation problem, several authors have numerically investigated
the convergence properties of the RBF method in the solution
of elliptic PDEs, and have also found exponential convergence for
that type of problems. Cheng et al. [10] carry out a numerical
investigation to show that using inverse multiquadrics
(IMQ),
p
c=h
. This law is
their results converge at the exponential rate Ol
usually accepted also in the case of MQs. However, these results

Author's personal copy


ARTICLE IN PRESS
1068

F. Bernal, M. Kindelan / Engineering Analysis with Boundary Elements 33 (2009) 10621073

101
N = 96
N = 291
N = 586
N = 981

102

RMS ()

RMS ()

103

101

100

100

N = 96
N = 291
N = 586
N = 981

101

102
0

10
c1/2/ h

15

20

10

15

c1/2/ h

Fig. 7. Motz problem. Left: loss of spectral convergence for MQ-only interpolant. Right: restoration of spectral convergence through the addition of a singularity-capturing
function (f 1 ).

101

101

100

100

h 1/10

RMS ()

RMS ()

h 1/20

101

h = 1/5
h = 1/10
h = 1/15
h = 1/20

4
6
8
(c1/20.1504)/h

h 1/40

101

102

102
0

h 1/30

10

12

6
8
10
(c1/2 0.075)/h

12

p
Fig. 8. Restoration of h2c convergence: best t for n 1 (MQs r cos y=2). Only enriched-RBF collocation matrices with condition number in the range 106 pkp1014
(lled markers) are considered for the t. Outliers to the left of the x-axis correspond to ko106 , while the right tail is made up of ill-conditioned cases (k41014 . Left side:
1
) have
Motz problem. Right side: Motz variant with a circular hole. Notice the different resolutions used in each problem. On the right plot, the coarsest point sets (h  10
been excluded from the t regardless of their condition number.

apply only to smooth solutions, as is apparent in Fig. 7 for our


error data. Without enrichment, a convergence pattern can hardly
be made out from the data (left side). In a previous study on
elliptic PDEs, Hu et al. [15] considered the Motz problem and
solved it with a similar enriched-RBF technique, and followed
c=h
Madych
in considering that the error goes like Ol rather than
p
c=h
Ol
. Our own results, on the other hand, do approximately
agree with the latter t, as shown in Fig. 7 (right). However, the
experiments for the four resolutions fail to collapse intopa single
c=h
, but
curve. Therefore, the error convergence is not simply Ol
the t is lacking an additionalpdependence on h. In fact, if we plot
c=hb=h
, the error of all the different
our results as a function of l
numerical experiments (in the range of condition number
106 oko1014 ) do coalesce into a single curve, both for the Motz
problem (Fig. 8, left)
p
 6:54  0:5 c0:1478=h
(23)
and for the Motz variant with a hole (Fig. 8, right)
p
 2:66  0:6 c0:0753=h

respect to c) has dropped for each of them as n grows. It might be


argued that, as more and more terms of the exact solution are
included, the contribution of the MQs in the interpolant becomes
smaller; on the other hand, the c-convergence is linked to the
multiquadrics alone.

4. Application to singular Laplacian ows


In this section we apply the enriched Kansas method to a
singular potential ow problem which arises in the simulation of
plastic injection molding (PIM) [14]. Because it features a freeboundary problem conned inside a 2D, irregular shape (the mold
oor plan), PIM makes a good case for a meshless formulation.
At a given time t, the PDE domain Ot is dened by the ooded
region of the mold oor plan. Its boundary
@Ot @OI t [ @OF t [ @OW t

(24)

1
In the latter case, the data relative to h  10
have been removed
from the t because of poor resolution. As it can be seen, the
circular hole slows down the convergence rate, but does not
change the convergence law itself.
Regarding the convergence pattern if more enriching terms
than just f 1 are present, the following comments are in place.
From Fig. 5 (right side), we can see that the convergence curves for
n 3, 5, and 10 are not as smooth as the curve for n 1.
Moreover, since the three curves for n 3, 5, and 10 collapse
together at about c 0:5, the mean rate of convergence (with

(25)

is made up of the injection gate(s) @OI t (through which the


polymer is fed into the mold by an injection machine), the
impenetrable walls @OW t, and the front @OF t. The goal is to
solve the pressure eld which drives the motion of the uid
at time t, compute the velocity eld v, and accordingly advance
the front to its new position at t Dt, thus updating Ot Dt and
@Ot Dt, until the mold is completely lled. Under certain
assumptions [5] the velocity eld vx; y can be regarded as
potential, v / rp, and the instantaneous pressure px; y obeys
the following Laplace PDE:

r2 p 0 if ~
x2O

(26)

Author's personal copy


ARTICLE IN PRESS
F. Bernal, M. Kindelan / Engineering Analysis with Boundary Elements 33 (2009) 10621073

1069

Fig. 9. Conservation of mass in the RBF interpolant: only MQs (top row) and enriched (bottom row). In both cases c 0:1 and k O1011 :

p pI

if ~
x 2 @OI ;

p0

if ~
x 2 @OF ;

@p
0
@n

if ~
x 2 @OW
(27)

where we have dropped the argument t. The pressure exerted by


the injection machine, pI , is assumed constant along the injection
segment. Notice that a similar singularity to that of the Motzs
problem arises, with the BCs changing abruptly from @p=@n 0
to p pI at both ends of the injection segment. Therefore, a
straightforward application of Kansas method produces overshoots at the gate edges of the mold, not only in the RBF solution
but in the higher derivatives as well. This is illustrated in the
example shown in Fig. 9 (top row). The semicircular domain
(which may be regarded as the initial condition of a PIM
simulation) is dened by
q
(28)
@OF : fx; yj x2 y2 0:62
@OW : fx; yjx 0; 0:2pjyjp0:6

(29)

@OI : fx; yjjyjp0:2

(30)

and pI 1. It can be seen that the amplitude of the oscillations


increase by some three orders of magnitude between the RBF
2
solution and the residual Rx; y r p. Notice that the latter is
equivalent to non-physical mass sources or sinks (depending on
the sign). The crucial point is that such oscillations will not, in
general, cancel out, thus preventing the simulation of the ow
evolution from conserving mass at each iteration. We remark that
this kind of PDE and the associated difculties also appear, for
instance, in the context of seepage ow [23], or as the zeroth stage
of a linearization procedure to solve the full non-linear PIM
equations [7].
On the other hand, the enrichment of the interpolant provides
a mass-conserving interpolant (Fig. 9, bottom row). Let X 1 ; Y 1
0; 0:8 and X 2 ; Y 2 0; 1:2 be the gate edges. According to
(15)(17), the f k r i ; yi ; kX1 (18), are local solutions of the
Laplacian close to the ith gate edge of coordinates
X i ; Y i ; i 1; 2, where
q
r i x  X i 2 y  Y i 2
(31)


yi arctan


y  Yi
;
x  Xi

poyi pp

(32)

Fig. 10. Bugle domain. The dashed- and thick-lines depict the injection gate and
the front, respectively. The colored sector is the overlap between DDM
subdomains, limited by the articial boundaries GA and GB .

We pick k 1 in order to capture the derivative discontinuity. The


enriched RBF interpolant with N 1177 MQs and n 1 special
functions per singularity is therefore
p

N
X
i1

 p

 p

ai fk~
x ~
xi k aN1 f 1 r 1 ; y1 aN2 f 1 r 2 ;  y2
2

(33)
Not only does the residual drop by orders of magnitude but
there also is an improvement in the error (with respect to a ne
FEM solution).
4.1. The classical Schwarz alternating algorithm
A more challenging test is the non-convex domain O depicted
in Fig. 10, which will be referred to as bugle. It is obtained by
removing a wedge of amplitude 2:1294 rad from the annular
region between two circles of radii 0:5 and 1:1 whose centers are
offset a distance 0:5 along the y-axis. The wedge vertex is the
center of the inner circle, which is taken as the origin.
The MQ centers are evenly scattered throughout the domain at
an average distance to the closest neighbor given by hhi 0:067,
plus the required extra centers for PDEBC, which are set off at a
distance hhi of every boundary node in the direction of the
outward normal. The shape parameter is the same for all of the
RBF centers. In addition to them, one more extra layer of interior

Author's personal copy


ARTICLE IN PRESS
1070

F. Bernal, M. Kindelan / Engineering Analysis with Boundary Elements 33 (2009) 10621073

nodes has been arranged at hhi=2 off the boundary (but inside the
domain), with the goal of increasing accuracy close to it [27]. Also,
the nodes have been removed from several straight corners. All in
all, the effect of such tweakings on the accuracy is small.
However, because of the irregular, non-convex shape involved
and the presence of square corners, it turns out to be necessary to
decompose the bugle domain into simpler subdomains which are
solved recursively using the classical Schwarz alternating algorithm [33], which will be outlined next for this particular
problem. More details about DDM in the context of Kansas
method may be found in [24]. The domain is partitioned into two
overlapping injection and front subdomains labeled as OA and OB ,
respectively, so that OA [ OB O and OA \ OB aGB . The overlap is
dened by articial straight boundaries GA and gB radiating at
angles 0:08 and 3p=4 radians from the origin (seeFig. 10), along
which extra nodes are added to enforce a Dirichlet boundary
condition.
The jth iterate of p in the subdomain A, pAj , is the solution of

r2 pAj 0; ~
x 2 OA
pAj 10;

~
x 2 @OI ;

(34)
@pAj
@n

~
x 2 @OAW ;

0;

pAj pBj ;

~
x 2 GA

@OAW

pBj

(35)

@OBW

where
is the iterate j over subdomain B, and
and
are
the portions of the wall belonging to the injection and front
subdomains, respectively. The nodal values of pAj are used to
update the Dirichlet BC along GB , in order to yield pBj1

r2 pBj1 0; ~
x 2 OB
x 2 @OF ;
pBj1 0; ~

(36)
@pBj
@n

0; ~
x 2 @OBW ;

pBj1 pAj ; ~
x 2 GB

(37)

Let pAj ji be the nodal pressure, at iteration j, on each the nodes i in


point set A which lie along GA . Dene analogously pBj jk and
X
X
(38)
Xj
pAj i
pBj k
i

Convergence is declared when jXj  Xj1 jo106 , which typically


takes place in a few iterations. Moreover, the convergence rate is
exponential, with the exponent a growing function of the overlap
area.
The resulting point sets are shown in Fig. 11. Notice that new
points have been placed off the articial boundaries GA and GB in
order to perform PDEBC also along it. The RBF interpolant for the
injection subdomain consists of 341 MQ centers plus the enriching

functions, while that for the front subdomain is made up of just


675 MQ centers. The nodes in the overlapping region belong to
both subdomains. The action of the enriching functions is
therefore restricted to the injection subdomain.

4.2. Results
Instead of performing an exact inversion of the collocation
algebraic system, we have rather employed Penroses pseudoinverse (implemented by the MATLAB command pinv). At condition
numbers below the ill-condition threshold (which is about k
1014 in our double-precision MATLAB environment), the pseudoinverse is indistinguishable from the inverse, and beyond it, it is
equivalent to inverting the system with singular value decomposition (SVD). With MATLABs default parameters, the pseudoinverse provides better stability for ill-conditioned systems up to
k  1018 ) and allows for picking the shape parameter from a
wider range without the concern of triggering a sudden numerical
blow-up.
For lack of an analytical solution, a FEM approximation has
served as a reference solution, which has been obtained with the
PDETool of MATLAB 7.0. In order to optimize accuracy, the solver
runs in the adaptive mode, which means that the mesh is
adaptively rened throughout several iterations. The nal mesh of
NFEM 424 732 vertices is therefore denser in regions containing
corners, narrow passages or high gradients. However, since we
would like to sample the RBF solutions over an evenly distributed
set of monitoring points, we take the preprocessing step of
linearly interpolating the values of the FEM reference solution
yielded by the adaptive mesh over a different, balanced set of
Nev 5N FEM scattered points, which are also generated by the
PDETool. The magnitude of Nev 7329 means a trade-off between
an adequate sampling of the boundary region (along which
the largest errors take place) and a reasonable CPU time
(for interpolation of the RBF approximation).
In each of the experiments, the quality of the RBF solution is
assessed through the RMS values of the error  and of the residual
R. The latter is simply the RMS of
2

R r p

(39)

sampled over the set of evaluation nodes. A related test of the


harmonicity of the numerical solution is obtained by computing
the total ow across the domain.

1.5
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0.2

0.5

0.5

0
1.5

0.5

0.5

0.5

1.5

Fig. 11. Bugle: DDM subdomains. The green, red, purple, navy, and light-blue points are the PDE, front Dirichlet, articial Dirichlet, Neumann collocation nodes and the
outlying MQ centers for PDEBC. The green circles indicate PDEBC. (For interpretation of the references to color in this gure legend, the reader is referred to the web version
of this article.)

Author's personal copy


ARTICLE IN PRESS
F. Bernal, M. Kindelan / Engineering Analysis with Boundary Elements 33 (2009) 10621073

Table 2
Bugle domain, DDM MQs.

1071

Table 3
Bugle domain, DDM MQs f 1 r1 ; p=2 y1 f 1 r 2 ; p=2  y2 .

c=hhi

RMS()

RMS(R)

Inow

Outow

DF

c=hhi

RMS()

RMS(R)

Inow

Outow

1.00

0.046

1.07

0.675

0.524

0.151

1:2  109

1.00

0.0267

1.00

0.673

0.699

0.026

1:2  109

1.41

0.028

0.96

0.782

0.562

0.220

8:6  109

1.41

0.0134

0.77

0.780

0.711

0.069

8:6  109

1.73

0.025

1.01

0.800

0.577

0.223

9:5  1010

1.73

0.0126

0.66

0.798

0.713

0.084

9:5  1010

2.24

0.023

1.16

0.791

0.588

0.202

4:8  1012

2.24

0.0098

0.54

0.788

0.714

0.074

4:8  1012

3.32

0.024

1.41

0.752

0.587

0.165

5:9  1016

3.32

0.0037

0.36

0.750

0.713

0.036

5:9  1016

3.87

0.026

1.48

0.741

0.579

0.163

2:5  1017

3.87

0.0021

0.30

0.739

0.713

0.025

2:5  1017

4.47

0.030

1.52

0.735

0.565

0.170

2:5  1018

4.47

0.0013

0.26

0.732

0.713

0.019

2:5  1018

5.00

0.034

1.55

0.732

0.551

0.182

1:2  1019

5.00

0.0008

0.23

0.729

0.713

0.016

1:2  1019

5.48

0.039

1.55

0.731

0.534

0.197

1:9  1019

5.48

0.0014

0.22

0.727

0.712

0.015

1:9  1019

5.92

0.044

1.55

0.730

0.518

0.212

3:6  1018

5.92

0.0011

0.20

0.726

0.712

0.014

3:6  1018

6.32

0.048

1.55

0.730

0.502

0.227

1:6  1019

6.32

0.0009

0.19

0.725

0.712

0.013

1:6  1019

6.71

0.053

1.54

0.730

0.486

0.243

20

2:2  10

6.71

0.0009

0.18

0.725

0.712

0.013

2:2  1020

7.07

0.058

1.52

0.730

0.471

0.259

2:4  1019

7.07

0.0010

0.17

0.724

0.712

0.013

2:4  1019

7.42

0.010

0.44

0.725

0.656

0.069

19

2:9  10

7.42

0.0020

0.19

0.724

0.711

0.014

2:9  1019

7.75

0.018

0.26

0.725

0.606

0.119

2:2  1019

7.75

0.0008

0.16

0.724

0.715

0.009

2:2  1019

8.06

0.031

0.57

0.707

0.549

0.159

19

8.06

0.0025

0.25

0.705

0.718

0.014

2:6  1019

2:6  10

Let us dene the in- and out-ows as the line integrals


Z
@p
FI
dl
GI @n

FF

Z
GF

@p
dl
@n

(40)

(41)

where @p=@n stands for the derivative of p along the outward


normal. Because of the impenetrability of the walls, the ow
balance

DF FI FF

(42)

vanishes in the case of the exact solution. FI is straightforward to


compute analytically as long as GI is a straight segment, while we
have preferred Simpsons quadrature over a partition of GF into
5000 subintervals for FF .
The effect of including the rst Motzs function into the RBF
interpolant on the estimators RMS(), RMS(R), and DF is to let
them drop by roughly one order of magnitude each, compared to
the MQ-only solution. In fact, the MQ-only solution in Table 2
even with domain decompositioncan hardly be regarded as
harmonic, as it yields unacceptable ow balances, with violations
of mass conservation of the order of 20% or even larger. Again, this
can be xed by the inclusion of a singularity-capturing function at
both ends of the injection gate, as shown in Table 3. Notice that
the inow grows (in absolute value) with c, while the outow
decreases. For this reason, the estimator DF must be interpreted
with care. However, the convergence is clear: all the three
estimators RMS(), RMS(R), and DF drop steadily with c, until
reaching a common minimum, neatly located at c 7:75hhi. This
pattern allows us to regard DF as a trustworthy indicator of
convergence. For comparison purposes, the ows of the reference
FEM solution (with N FEM vertices) were also computed (they are
easy to compute exactly, for the gradient is constant in every
element, and therefore the ow prole along a segment is
staggered). The FEM inow and outow were 0:714 and 0:720,
respectively, thus yielding 0:006. On the other hand, if the
collocation nodes in the RBF point set (which are generated as
well by MATLABs PDETool) were the vertices of a mesh, the
resulting value of the ow balance for that FEM approximation
would be 0.237. It must also be noted that, notwithstanding the
fact that most of the entries in Tables 2 and 3 take place beyond

DF

our ill-condition threshold, results are numerically stable. While


the use of the pseudoinverse does not necessarily improve the
results beyond the ill-conditioning, it preventsor at least puts
offa numerical blow-up. Again, there is no advantage in adding
2n; n41 Motz functions into the RBF interpolant, as is apparent in
Fig. 12 (left). The curves for n f1; 2; 3; 5; 10g converge together in
a smooth manner until the onset of ill-condition at about
c 3:5hhi. From the graphs it is also clear that the convergence
of both RMS() is exponentialat least in the nal stretch of the
shape parameter before ill-condition (approximately between c
1:5hhi and 3:5hhi, where the condition number is k  1016 ).
Finally, the right side of Fig. 12 shows the t to the convergence
law
p
c0:0464=hhi

RMS 0:10  0:56

(43)

This pattern is still acceptable in spite of the irregular domain, the


use of DDM, and the fact that the two point sets involved in
each piece of data are unevenly scattered. Specially the front
subdomain has clusters of nodes along the narrow channel and
close to the front corners, which cause that its condition number
kB be about 105 times larger than that of the injection
subdomain, kA . Data actually considered for the t
(106 pkA p1014 ) are represented by lled markers.

4.3. Comparison to FEM


Fig. 13 compares mass conservation for the bugle domain as
computed by FEM (running on the adaptive-mesh mode) and
Kansas method with multiquadrics plus the rst Motzs function
f 1 . The y-axis represents the absolute value of the ow balance
relative to the mean ow, while the x-axis is the number of nodes
(mesh vertices in the FEM case). The enriched RBF approximation
clearly outperforms the FEM solution both in convergence rate
and in efciency, before breaking down due to round-off error at
about N 2500. Before it, there is a range where the convergence
is approximately potential, with an exponent ve times higher
than that of FEM. The shape parameter is c 55h. In the case of
FEM, the fact that the mesh is adaptive and hence there is a large
distribution of distances to closest neighbors may account for the
rather unusual exponent.

Author's personal copy


ARTICLE IN PRESS
F. Bernal, M. Kindelan / Engineering Analysis with Boundary Elements 33 (2009) 10621073

102

101

102
n=0
n=1
n=2
n=3
n=5
n = 10

103

104

RMS ()

RMS |error| to FEM solution

1072

103

104
0

4 5
c/<h>

h 0.067
h 0.050
h 0.040
h 0.030

10
12
(c1/2 0.05)/ h

14

Fig. 12. Bugle: error (hhi 0:0067) and convergence t (n 1).

100

2 (out+in) / (out in)

Adaptive FEM
Enriched RBF
101
slope = 0.69
102

103
slope= 3.52

104

103

104

105

106

N
Fig. 13. Flow conservation of RBF vs. FEM in the bugle domain.

5. Conclusions
The work presented in this paper investigates the performance
the RBF collocation method in elliptic PDES with singular
boundary conditions. As we have seen in the case of Newtonian
ow in PIM, the effect of sharp features in the solution may be
devastating for the validity of straightforward RBF approximations. Moreover, the presence of singularities causes the loss of
one of the most interesting properties of RBF collocation, namely
its spectral convergence.
Contrary to Trefftz methods, it is the combination of RBFs
and analytical terms that solves the PDE. The removal of the
singularity recovers the h  c exponential convergence rate of
Kansas method discovered by Cheng et al. for smooth elliptic
PDEs. The enriching term can be constructed analytically and
incorporated into the interpolant regardless of the shape of the
domain far away from the singularity, as we have seen that in the
Motz, Motz variant, and related Newtonian PIM ow problems
discussed in this paper. In all these cases, the singular behavior
can be adequately (or entirely) accounted for by a single enriching
term. Elsewhere, we have applied the same idea to somewhat
more intricated elliptic singular problems, where several singularity-capturing terms are needed for absorbing the non-regular
features of the solution at singular points [6].
It has been suggested [23] that Kansas method may be
superior to the FEM in solving elliptic boundary-value problems,

due to its simpler coding, meshless quality and spectral


convergence. We have tested this comparison focusing on mass
conservation, which is critical in the numerical simulation of
incompressible ows. The FEM works on the weak, or variational,
formulation of the PDE. By construction, harmonicity of the
numerical approximation is built-in, since the divergence theorem
has been implicitly applied on each element when deriving the
system of equations for the nodal values. Moreover, if interpolation is linear within each element, the Laplacian of the FEM
solution is zero everywhere. On the other hand, Kansas method
collocates the strong formulation of the PDE at certain points in
the domain. There is no safeguard in Kansas method to enforce
ow conservation or harmonicity of the RBF interpolant (since the
RBFs are not divergence-free in the rst place). As we have seen,
even natural-looking problems of potential ow can be intractable
by the standard Kansas method, unless a specialized treatment be
undertaken. The enrichment with singular functions (as long as
they are available for the concrete problem) seems a simple and
unexpensive modication of Kansas method, which enables it to
be competitive also in this kind of non-regular elliptic problems.
As reported in Section 5, the gain over FEM in terms of the size
of the numerical support may be between two and three orders of
magnitude, even in highly irregular domains.

Acknowledgments
This work has been supported by the Spanish MECD Grants
MAT2005-05730, FIS2004-03767 and FIS2007-62673 and by
Madrid Autonomous Region Grant S-0505. The constructive
remarks by an anonymous referee have been greatly appreciated.
References
[1] Alves CJS, Leitao VMA. Crack analysis using an enriched MFS domain
decomposition technique. Eng Anal Bound Elem 2006;30(3):1606.
[2] Arad M, Yosibash Z, Ben-Dor G, Yakhot A. Computing ux intensity factors by
a boundary method for elliptic equations with singularities. Commun Numer
Methods Eng 1998;14:65770.
[3] Behrens J, Iske A, Pohn S. Effective node adaption for grid-free semiLagrangian advection. In: Sonar T, Thomas I, editors. Discrete modelling and
discrete algorithms in continuum mechanics. Berlin: Logos; 2001. p. 1109.
[4] Belytschko T, Fleming M. Smoothing, enrichment and contact in the elementfree Galerkin method. Comput Struct 1999;71:17395.
[5] Bernal F, Kindelan M. RBF meshless modeling of non-Newtonian HeleShaw
ow. Eng Anal Bound Elem 2007;31:86374.
[6] Bernal F, Gutierrez G, Kindelan M. Use of singularity capturing functions in
the solution of problems with discontinuous boundary conditions. Eng Anal
Bound Elem 2009;33(2):2008.
[7] Bernal F, Kindelan M. A meshless solution to the p-Laplace equation. In:
Ferreira AJM, Kansa EJ, Fasshauer GE, Leitao V, editors. Progress on meshless
methods. Berlin: Springer, to appear.

Author's personal copy


ARTICLE IN PRESS
F. Bernal, M. Kindelan / Engineering Analysis with Boundary Elements 33 (2009) 10621073

[8] Buhmann MD. Multivariate cardinal interpolation with radial basis functions.
Constr Approx 1990;6:22555.
[9] Chantasiriwan S. Solutions to harmonic and biharmonic problems with
discontinuous boundary conditions by collocation methods using multiquadrics as basis functions. Inter Commun Heat Mass Transfer 2007;34:
31320.
[10] Cheng AHD, Golberg MA, Kansa EJ, Zammito T. Exponential convergence and
hc multiquadric collocation method for partial differential equations. Numer
Methods Part Differ Equat 2003;19:57194.
[11] Fasshauer GE. Meshfree approximation methods with Matlab. In: Interdisciplinary mathematical sciences, vol. 6. Singapore: World Scientic Publishers;
2007.
[12] Fedoseyev AI, Friedman MJ, Kansa EJ. Continuation for nonlinear elliptic
partial differential equations discretized by the multiquadric method. Int J
Bifur Chaos 2000;10.
[13] Hardy RL. Multiquadric equations of topography and other irregular surfaces.
J Geophys Res 1971;176:190515.
[14] Hieber CA, Shen SF. A nite-element/nite-difference simulation of the
injection-molding lling process. J Non-Newtonian Fluid Mech 1980;7:132.
[15] Hu HY, Li ZC, Cheng AHD. Radial basis collocation methods for elliptic
boundary value problems. Comput Math Appl 2005;50:289320.
[16] Kansa EJ. Multiquadricsa scattered data approximation scheme with
applications to computational uid-dynamics. I. Surface approximations
and partial derivative estimates. Comput Math Appl 1990;19:12745.
[17] Kansa EJ. Multiquadricsa scattered data approximation scheme with
applications to computational uid-dynamics. II. Solutions to parabolic,
hyperbolic and elliptic partial differential equations. Comput Math Appl
1990;19:14761.
[18] Kansa EJ, Aldredge RC, Ling L. Numerical simulation of two-dimensional
combustion using mesh-free methods. Eng Anal Bound Elem, in press,
doi:10.1016/j.enganabound.2009.02.008.
[19] Karageorghis A. Modied methods of fundamental solutions for harmonic
and biharmonic problems with boundary singularities. Num Meth Part Differ
Equat 1992;8:119.

1073

[20] Jung JH. A note on the Gibbs phenomenon with multiquadric radial basis
functions. Appl Numer Math 2007;57:21329.
[21] Lanczos C. Discourse on Fourier series. New York: Hafner; 1996.
[22] Li Z-C, Mathon R, Sermer P. Boundary methods for solving elliptic problems
with singularities and interfaces. SIAM J Numer Anal 1987;24:48798.
[23] Li J, Cheng AHD, Chen CS. A comparison of efciency and error convergence of
multiquadric collocation method and nite element method. Eng Anal Bound
Elem 2003;27:2517.
[24] Ling L, Kansa EJ. Preconditioning for radial basis functions with domain
decomposition methods. Math Comput Model 2004;40:141327.
[25] Madych WR. Miscellaneous error bounds for multiquadric and related
interpolators. Comput Math Appl 1992;24:12138.
[26] Motz H. The treatment of singularities of partial differential equations by
relaxation methods. Q Appl Math 1947;4:3737.
[27] Platte R, Driscoll TA. Computing eigenmodes of elliptic operators using radial
basis functions. Comput Math Appl 2004;48:56176.
[28] Poullikkas A, Karageorghis A, Georgiou G. Methods of fundamental solutions
for harmonic and biharmonic boundary value problems. Comput Mech
1998;21:41623.
[29] Rabczuk T, Belytschko T. Cracking particles: a simplied meshfree method for
arbitrary evolving cracks. Int J Numer Methods Eng 2004;61:231643.
[30] Rosser JB, Papamichael N. A power series solution of a harmonic mixed
boundary value problem. MRC Technical Summary Report 1405, University of
Wisconsin; 1975.
[31] Schaback R. Error estimates and condition numbers for radial basis function
interpolation. Adv Comput Math 1995;3:25164.
[32] Shu C, Ding H, Yeo KS. Local radial basis function-based differential
quadrature method and its application to solve two-dimensional incompressible NavierStokes equations. Comput Methods Appl Mech Eng 2003;192(3):
94154.
[33] Schwarz HA. Gesammelte mathematische abhandlungen. Bronx, NY: Chelsea
Publishing Co.; 1972 Band 1, p. 11.
[34] Wait R, Mitchell AR. Corner singularities in elliptic problems by nite element
methods. J Comput Phys 1971;8:45.

You might also like