Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Engineering Analysis with Boundary Elements 50 (2015) 97–108

Contents lists available at ScienceDirect

Engineering Analysis with Boundary Elements


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

Direct use of radial basis interpolation functions for modelling source


terms with the boundary element method
Carlos F. Loeffler a,n, Átila L. Cruz a, André Bulcão b
a
PPGEM, Departamento de Engenharia Mecânica, Universidade Federal do Espírito Santo, Av. Fernando Ferrari, 514, Goiabeiras, 29075-910 Vitória, ES, Brasil
b
PETROBRAS, CENPES—Centro de Pesquisa e Desenvolvimento Leopoldo A. Miguez de Mello, Av Horácio de Macedo—Cidade Universitária, 950, Ilha do
Fundão, 21941-915 Rio de Janeiro, RJ, Brasil

art ic l e i nf o a b s t r a c t

Article history: In this paper a new technique is presented for transforming the domain integral related to the source
Received 6 December 2013 term that characterizes the Poisson Equation, within the scope of the boundary element method, for
Received in revised form two-dimensional problems. Similarly to the Dual Reciprocity Technique, the proposed scheme avoids
6 June 2014
domain discretization using primitive radial basis functions; however, it transforms the domain integral
Accepted 26 July 2014
into a single boundary integral directly. The proposed procedure is simpler, more versatile and some
useful and modern techniques related to radial basis function theory can be applied. Numerical tests
Keywords: show the accuracy of the proposed technique for a simple class of complete radial interpolation
Radial basis approximation functions, pointing out the importance of internal poles and the potential of applying fitting interpola-
Poisson’s problems
tion schemes to minimize the computational storage, particularly considering more complex future
Boundary element method
approaches, in which a mass matrix may be generated. For the analysis of the accuracy and convergence
of the proposed method, results are compared with those obtained using Dual Reciprocity, using known
analytical solutions for reference.
& 2014 Elsevier Ltd. All rights reserved.

1. Introduction Anticipating by some years the use of meshless approaches [5],


the DRBEM also uses radial basis functions to interpolate the
The boundary element method (BEM) has been shown to be an variable that comprises the kernel of the domain integrals [6].
effective numerical technique for modeling many engineering However, the DRBEM was developed using radial functions with
problems. Considering problems governed by self-adjoint differ- complete support, which are effective for interpolation [7,8] or
ential operators, numerical results of the BEM are of high accuracy, fitting approximation. These classic functions provide satisfactory
for which the properties of the fundamental solution play an accuracy for application to moderately sized data sets, but diffi-
important role. culties can arise for larger data sets [9].
In many cases, intrinsic BEM limitations have been overcome Considering the advances with the Compact Radial Basis
gradually through new strategies, such as the resolution of internal Functions (CRBFs) [10,11], many criticisms have been made of
domain terms, for example, that which appears in the Poisson the complete radial basis functions traditionally used in the
Equation. Despite the existence of many techniques that model DRBEM. CRBFs present advantages especially in cases in which
this kind of problem, the Dual Reciprocity Technique (DRBEM) is the interpolation matrix is comprised of a large number of basis
still currently the most general option [1,2]. It was created for points and it can also be ill-conditioned.
solving eigenvalue and dynamic problems, but its methodology In fact, computational tests have shown that many classic
has been successfully generalized and applied to other cases, complete radial basis functions become inadequate with the
especially to model the domain integrals that appear in time DRBEM in certain applications. Some studies indicate lack of
dependent cases, diffusive-advective problems and other heat and convergence when traditional complete radial functions are used
mass transfer problems [3,4]. in DRBEM in conjunction with iterative procedures [12]. It should
be emphasized that the use of radial functions in this formulation
differs from the simple interpolation procedure and also the
techniques of solving differential equations, since it generates
 Corresponding author. Tel.: þ55 27 33352669. two primitive functions from the original interpolation function,
E-mail addresses: carlosloeffler@bol.com.br (C.F. Loeffler), forming auxiliary matrices that may produce additional harmful
atila.lc@hotmail.com (Á.L. Cruz), bulcao@petrobras.com.br (A. Bulcão). numerical effects.

http://dx.doi.org/10.1016/j.enganabound.2014.07.007
0955-7997/& 2014 Elsevier Ltd. All rights reserved.
98 C.F. Loeffler et al. / Engineering Analysis with Boundary Elements 50 (2015) 97–108

Therefore, the objective of this paper is to present an alter- The coefficient δ in these previous equations means the greatest
native formulation, hereafter named DIBEM, to approximate the distance between two nodal points.
complete kernel of the domain integral related to the body force Since the fundamental solution in the DIBEM is the kernel, the
term, such as is done in simple interpolation. Similarly to the source point ξ must have different positions than the information
DRBEM, the standard approach of the BEM to the Laplace operator points to avoid singularities. In fact, the body force is a known
is maintained since it is assumed that it has satisfactory accuracy. value, thus it can be taken at another position. Considering that
The proposed technique is also simpler and more versatile, since linear boundary elements are used, the values of the body force
any type of radial basis function may be used according to a simple are taken at the center.
scheme that avoids domain integrations, as it will be shown Similarly to the DRBEM, the proposed method also uses a
herein. In order to present preliminary results, only classic radial primitive function Ψj, such as
basis functions with complete support are tested. Z Z Z Z
zðξ; XÞdΩ ¼ ðξ α j ψ j;ii ðXÞÞdΩ ¼ ðξ α j ψ j;i ðXÞni ðXÞÞdΓ ¼ ξ α j ηj ðXÞdΓ
Ω Ω Γ Γ
ð6Þ
2. Boundary integral equation
The numerical evaluation of the previous boundary integral is very
Consider a two dimensional domain Ω(X) constituted by a simple. Using Eq. (6) and the well-known BEM procedure for the
homogeneous and isotropic medium, in which a scalar potential u discretization, Eq. (1) may be rewritten as:
(X) and a body force p(X) act. The domain is limited by a boundary
H 11 u1 þ …H 1n un  G11 q1  …G1n qn ¼ 1 α1 N 1 þ 1 α2 N 2 þ ::1 αn N n
Γ(X), where potential conditions are prescribed on Γu(X) and
normal derivatives of potential q(X) on a complementary bound- H 21 u1 þ::::H 2n un  G21 q1  …G2n qn ¼ 2 α1 N1 þ 2 α2 N 2 þ::2 αn N n
ary Γq(X). Taking an auxiliary function un(ξ;X) and its normal ………………
derivative qn(ξ;X), where ξ is an arbitrary source point, it is H n1 u1 þ…H nn un  Gn1 q1  …Gnn qn ¼ n α1 N1 þ n α2 N 2 þ::n αn N n ð7Þ
possible to establish an equivalent inverse integral form related
In a matrix form, for convenience:
to Poisson's Equation, given by [13]:
Z Z Z ½Hfug ½Gfqg ¼ ½AfNg ¼ fPg ð8Þ
cðξÞuðξÞ þ uðXÞqn ðξ; XÞdΓ  qðXÞun ðξ; XÞdΓ ¼  pðXÞun ðξ; XÞdΩ ξ
Γ Γ Ω In Eq. (8), the lines of matrix A are comprised by vectors α, which
ð1Þ may be obtained from following the basic interpolation equation

In Eq. (1), the value of the coefficient c(ξ) depends not only on the ½F½ξ α  ¼ ½ξ Λ u ð9Þ
positioning of the source point ξ with respect to the physical ξ
Where the matrix Λ is composed of values of the fundamental
domain Ω(X) but also on the boundary smoothness at the same solution. For each source point ξ, the right hand side of Eq. (9) may
point [14]. be rewritten as

½ξ Λ u ¼ ½ξ Λ ½Fα ð10Þ
3. Interpolation procedure
Thus, the last two equations can be equaled, resulting in:

The aim is to resolve the integral term in the right hand side of ½ξ α  ¼ ½F  1 ½ξ Λ ½Fα ¼ ½F  1 ½ξ Λ ½p ð11Þ
Eq. (1) by interpolation, using radial basis functions in a similar
For Poisson type problems, it is possible to optimize the computa-
approach as that used in the DRBEM procedure. Thus, the ξ
tional operations, since the matrix Λ is diagonal. In both
complete kernel of the domain integral is interpolated directly,
boundary formulations, DRBEM and DIBEM, an inverse must be
according to the following expression:
computed once. However, in the first the construction of two
pðXÞun ðξ; XÞ ¼ zðξ; XÞ ¼ ξ α j F j ðX j ; XÞ ð2Þ interpolation matrices and additional products of them by H and
G, both being all full matrices [2], are required. Thus, computa-
For each source point ξ, the interpolation given by Eq. (2) is done tional time is decreased for the DIBEM, which requires only the
by considering each of the base points Xj in relation to the domain ξ
product of matrix Λ versus F  1 for each source point.
ξ
points X, weighted by the coefficients αj. The quantity of basis
points X must be equal to the known values of z(ξ;X).
j

The interpolation functions Fj used belong to a class of radial 4. Internal basis points
functions, that is, the argument is composed by the Euclidian
distance r(Xj;X) between base points Xj and domain points X, As happens in the DRBEM, the distribution of the body force p(X)
hereafter called information points. inside the domain is not well approximated if the interpolation
Non radial functions could also be used with success [15,16], basis points are taken exclusively on the boundary. A first improve-
but it is important to point out that in the DIBEM the kernel of the ment to the accuracy of the results is simply to introduce basis
integral to be approximated is composed of the product of the points inside the domain, also named poles. Since the DIBEM
fundamental solution and a function that describes the distribu- interpolates directly all functions that compose the kernel of
tion of the body force on the domain. Therefore, for body forces the domain integral, including the fundamental solution, a larger
with more general shapes, the radial basis functions become number of internal poles is required for better performance.
advantageous. However, unlike with the DRBEM, increasing the number of internal
Regarding the interpolation functions, the following complete basis points does not alter the results, since the mathematical aim
radial basis functions are tested: of the DIBEM is closely similar to an interpolation technique.
F j ðX j ; XÞ ¼ ðr=δÞðsimple radialÞ ð3Þ
5. Least square curve-fitting
F j ðX j ; XÞ ¼ ðr=δÞ3 ðcubic radialÞ ð4Þ
The similarity of the DIBEM to a direct interpolation procedure is
F j ðX j ; XÞ ¼ ðr=δÞ2 lnðr=δÞðthin plate radialÞ ð5Þ advantageous, due to its simplicity. Thus, some auxiliary numerical
C.F. Loeffler et al. / Engineering Analysis with Boundary Elements 50 (2015) 97–108 99

resources can be used. One of these consists of considering that this 6. Numerical simulations
process of approximation may be done through an average curve on
the domain. In this case, the number of information points Xi must As mentioned, examples that possess analytical solutions are
be greater than the number of basis points Xj. This fitting is a used to evaluate the accuracy of the proposed formulation. At the
mathematical process based on the Least Squares Theory [17]. same time, the same problems are solved by the DRBEM to allow
According to this method, the determination of the αj coefficients for a more efficient comparison between the two formulations.
is done as follows: Linear elements were used for both formulations. Meshes from 32
to 320 boundary elements are used. The number of poles ranged
FðX i ; X m ÞT ðFðX i ; X j Þαj ¼ ½FðX m ; X i Þf ðX i Þ; i ¼ 1; …; K; j; m ¼ 1; …L; K 4 L from zero to 576. The distribution inside the domain was as
ð12Þ uniform as possible.
Performance analysis is done using the concept of average
In Eq. (12), L is the number of basis points, while K refers to the percentage error for each mesh. For each node, nodal error is
number of information points. calculated by the difference between the numerical and the
The use of the fitting scheme allows the size of the final matrix analytical values, divided by the largest analytical value presented
system, which is built initially with a large number of data points, in the problem, excepting the penultimate example, in which the
to be reduced. concept of the absolute error was adopted.
The main idea here is to compute the necessary information Concerning the behavior of radial functions with DRBEM, the
about the source term inside the domain, reducing the final cost of radial cubic function presented good results only for constant
processing by the elimination of a large number of internal domain forces, not showing convergence for linear and more
information points. It is expected that lowering the degree of the elaborate distributions. The thin plate produces reasonable results
radial approximation should not affect strongly the quality of overall, but commonly the behavior of the curve of average
results and that the basis points located on the boundary have percentage errors shows high sensitivity to the arrangement of
more importance in the approximation procedure. the internal points employed. Therefore, for sake of space, only the
For the sake of simplicity, fitting tests are implemented using two results obtained with the simple radial basis function are pre-
different strategies for numerical evaluation: firstly, excluding only sented for the DRBEM.
internal information points; secondly, excluding not only internal Interpolation and fitting results are separated for convenience.
information points, but also two additional information points
inserted between the extreme nodes of each boundary element. 6.1. First example: rod subjected to a constant vertical body force
The fitting technique may be used to avoid these extreme
conditions, which are used here deliberately for clearer numerical Fig. 1 shows the physical and geometric features of the first
evaluation, despite being particularly interesting to test the effect example.
of the complete exclusion of internal data points. Internal points The length is L, the specific mass is ρo, the Young Modulus is E
are very important for the proper interpolation, but become and the body force is g. For simplicity, all these variables are taken
expensive in large applications, particularly for the cases of time equal to unity. The governing equation for this problem is given by
marching problems.
d u ρ0 g
2
¼ ð13Þ
dx21 E

6.1.1. Interpolation procedure


Initially, the numerical results are presented for displacements
considering global radial basis functions. Thus, Fig. 2 presents
results for the DRBEM and the DIBEM using in both cases the
simple radial function; Fig. 3 presents exclusively results for the
DIBEM, using cubic radial and thin plate radial basis functions.
As expected, due to its mathematical features, the accuracy of
Fig. 1. Homogeneous vertical bar subjected to gravity. the solutions with the DIBEM without internal points presented

Fig. 2. Error comparison between the DRBEM and the DIBEM for displacements for the first example.
100 C.F. Loeffler et al. / Engineering Analysis with Boundary Elements 50 (2015) 97–108

Fig. 3. Error comparison between cubic radial and thin plate radial functions for displacements with the DIBEM for the first example.

Fig. 4. Error comparison between the DRBEM and the DIBEM for tractions for the first example.

Fig. 5. Error comparison between cubic radial and thin plate radial functions for tractions with the DIBEM for the first example.

Fig. 6. Comparison between two different fitting approximations for displacements with the DIBEM using the simple radial function for the first example
C.F. Loeffler et al. / Engineering Analysis with Boundary Elements 50 (2015) 97–108 101

higher errors than those obtained by the DRBEM. However, functions; Figs. 5 and 6 present results for the DIBEM using cubic
when enough internal points are introduced, results for the DIBEM radial and thin plate radial functions.
become better than the DRBEM, in which the introduction of many The DRBEM was more effective for the calculation of tractions
internal poles is not effective. It may also be noted that finer than for displacements, even though the DIBEM has a better
boundary meshes with few internal points frequently have inferior performance. Despite the better final results, in this example the
performance than poorer boundary meshes with many internal thin plate radial function did not repeat the previous performance,
points for the DIBEM. Still, as a consequence of the necessity of since the simple radial function presented a superior rate of
internal information about the body force, the error curve for convergence. The cubic radial function was the worst radial function
refined boundary meshes with DIBEM tends to present initially a and its percentage average error in the absence of internal poles
slower rate of convergence as additional internal points are was the highest.
introduced if compared to coarser boundary meshes, although
this behavior also depends on the complexity of the source term to
6.1.2. Fitting approximation
be approximated.
Fitting tests are implemented exclusively for the DIBEM, using
Comparing the classic radial basis functions used, they had
the two different strategies previously mentioned. The results for
similar performance for the DIBEM, but the thin plate was slightly
both strategies are shown next side by side in graphic form. Only
superior, excluding the extreme situation without internal poles.
the coarsest and finest boundary meshes were simulated. Thus,
Usually, a certain loss of monotonicity in the answer was observed
results for displacements using simple radial function are depicted
for the thin plate radial error curve for an increase in number of
initially in Fig. 6.
internal poles, particularly considering poor boundary meshes.
These graphs show that the performance of both fitting strategies
This behavior is only due to greater sensitivity of this radial
were less effective than that of the interpolation technique. Indeed,
function to the position the internal interpolation points.
more accurate results than the interpolation procedure were not
The good performance of the thin plate function in this case
expected for these two kinds of fitting strategies proposed. These
may be justified due to the interpolated kernel consisting of the
tests were performed due to the possibility of reaching an acceptable
fundamental solution, which for two dimensional problems is
accuracy with computational economy in future applications.
given by a logarithm.
The comparison between the numerical behaviors of these two
The following figures present results for tractions: Fig. 4 shows
schemes suggests clearly that excessive insertion of information
results for the DRBEM and the DIBEM using simple radial
points on the boundary is not effective, unlike what occurs with
the inclusion of more information points inside. Thus, considering
a sufficient amount of points inside to simulate the domain force,
finer boundary meshes are unproductive.
The scheme that uses additional information points located on
the boundary had the worst performance whilst results of reason-
able quality were achieved for the scheme that excludes only
internal information points. A stronger increase in the rate of
convergence error curves was observed by including a larger
number of internal points. Therefore, a larger number of internal
points was only included in the fitting simulations to show how
superior accuracy can be achieved using finer boundary meshes
under these conditions.
Fig. 7 presents results of fitting that exclude only internal basis
points, now using the thin plate radial function. As observed, this
function really presents faster convergence and better quality in
the final results than the simple radial function.

6.2. Second example: rod subjected to variable body force

Fig. 7. Fitting approximation for displacements using the thin plate radial basis Considering the same rod tested in the first example, the
function for the first example. density is now assumed to be linear, starting from a unitary value

Fig. 8. Error comparison between the DRBEM and the DIBEM for displacements using simple radial basis function for the second example.
102 C.F. Loeffler et al. / Engineering Analysis with Boundary Elements 50 (2015) 97–108

at the origin becoming zero at the free end. The governing finer boundary meshes for problems with simple distributions of body
equation is given by force. Thus, Fig. 12 presents graphically the percentage average error
curve for two meshes, with 36 and 164 boundary nodes, as a function
d u ρ0 ðL x1 Þg
2
¼ ð14Þ of the number of internal points for the simple radial function and the
dx21 EL thin plate radial function. Hereafter, a single scheme of fitting is used
for simulations, which excludes all internal points. These points only
are used to compute information about the body force inside.
6.2.1. Interpolation procedure Similarly to what occurs in the previous example, for the
Initially, Fig. 8 presents displacement results using the classic simple radial function the rate of convergence decreased when
simple radial basis function for the DRBEM and the DIBEM. The considering excessive numbers of boundary nodes, requiring
performance of the proposed formulation is significantly superior. numerous internal points to improve accuracy. For the thin plate
The thin plate radial function presents slightly better accuracy radial function, numerical behavior in this case was mainly
than simple radial basis function, but just for the finer boundary improved. Faster convergence can also be observed for the finer
mesh, as shown in Fig. 9. Therefore, both radial functions in the boundary mesh.
DIBEM give higher quality results than the DRBEM.
Concerning the tractions, results for the DIBEM and the DRBEM
both using simple radial function are presented in Fig. 10. Results
6.3. Third example: membrane subjected to variable domain force
for the proposed formulation are superior than the DRBEM and
the accuracy of the tractions were similar to that achieved for
The third problem presented consists of a square membrane
displacements. Rate of convergence for the DRBEM is superior for
with unitary edges L, subject to a domain force p(x1, x2) so that the
less internal points, but the quality of its results do not improve
governing equation is
from a given level of refinement. It must be pointed out that
results for DIBEM using the thin plate radial function were still x1 π x2
∇2 uðx1 ; x2 Þ ¼  cos ð15Þ
better, as shown in Fig. 11, despite slight oscillations in the error L L
curve for the coarse mesh.
The membrane is clamped on two parallel sides and is free in the
other two sides. The analytical solution for displacements u(x1, x2)
6.2.2. Fitting procedure is given by
The next test emphasizes the feature of the fitting procedure, in  
which poorer boundary meshes offer more satisfactory results than L π x1 x1 π x2
uðx1 ; x2 Þ ¼ 2 sinh  2 cos ð16Þ
π sinhðπ Þ L π L

Fig. 9. Error for the DIBEM displacements using thin plate radial basis function for Fig. 11. Error for tractions for the DIBEM using thin plate radial function for the
the second example. second example.

Fig. 10. Error comparison between the DRBEM and the DIBEM for tractions using simple radial basis function for the second example.
C.F. Loeffler et al. / Engineering Analysis with Boundary Elements 50 (2015) 97–108 103

Fig. 12. Comparison between simple radial and thin plate radial functions for the fitting procedure with the DIBEM in displacements for the second example.

Fig. 13. Membrane submitted to variable domain loading.

Fig. 14. Error comparison between the DRBEM and the DIBEM for displacements using simple radial basis function for the third example.

Fig. 13 shows the geometric and loading features of this third


problem.

6.3.1. Interpolation procedure


Displacement results for the DRBEM and the DIBEM are shown
in Fig. 14, both using the simple radial function for interpolation.
Despite a high value of error for meshes with few internal basis
points (not presented in graphics which are for presenting the
final results) the DIBEM performance for displacements was
clearly superior to DRBEM, with the error curves presenting a
convergence behavior with the increase of internal points for all
boundary meshes used, despite the oscillations in the traction
error curve for the coarser mesh. The behavior of thin plate radial
function with the DIBEM in the final results was still better, as Fig. 15. Error for the DIBEM displacements using thin plate radial basis function for
presented in Fig. 15. It is worth noting that the thin plate function the third example.
104 C.F. Loeffler et al. / Engineering Analysis with Boundary Elements 50 (2015) 97–108

Fig. 16. Comparison between error curves for tractions with the DRBEM and the DIBEM using simple radial basis function for the third example.

results in approximation with radial functions does not depend


solely on the affinity between the kernel of the integral to
be interpolated and the type of radial function chosen. Thus, when
a smoothing procedure such as fitting is performed, certain
characteristics stand out to achieve better final results, such as
the simplicity of the simple radial function.
Although the accuracy obtained previously with the interpola-
tion procedure is superior, the fitting procedure presents reason-
able final results. It should be noted that this accuracy was
achieved with the final matrix system composed exclusively of
boundary basis points, which may lead to considerable computa-
tional savings in applications related to dynamics.

6.4. Fourth example: clamped membrane with non uniform loading

Fig. 17. Error for the DIBEM tractions using thin plate radial basis function for the This example aims at evaluating the performance of both
third example. formulations in a situation in which exclusively Dirichlet condi-
tions are prescribed and the imposed domain action presents
does not have a good performance with the DRBEM, presenting higher gradients. It consists of a square membrane with unitary
strong absence of monotonicity. sides, governed by the following differential equation:
The graphs in Fig. 16 show tractions calculated using the simple
radial interpolation function with the DRBEM and the DIBEM. As ∇2 uðx1 ; x2 Þ ¼ ð12x21  12x1 þ 2Þx22 ðx22  2x2 þ 1Þ
it may be observed, the introduction of internal poles without þ ð12x22 12x2 þ2Þx21 ðx21  2x1 þ 1Þ ð17Þ
insertion of boundary nodes is not suitable with the DRBEM,
particularly for traction results. However, the rate of convergence The displacement field is given as follows:
of the DIBEM was very slow in this case, requiring many internal uðx1 ; x2 Þ ¼ ðx41  2x31 þ x21 Þðx42  2x32 þ x22 Þ ð18Þ
points to reach a value error similar to the DRBEM.
As mentioned, certain slight oscillations are common in inter- Due to the particular distribution of the domain forces the values
polation procedures using radial functions with internal poles, due of the normal derivatives of the potential are also null on the
to the different positions where these poles may be placed. The boundary and the criterion for performance verification now is
thin plate radial function is more sensitive to this effect as shown given in terms of average absolute error given at the nodal points.
in Fig. 17, particularly if coarse meshes are used, which suggests
that it is more appropriate to evaluate the convergence curve 6.4.1. Interpolation procedure
considering the general tendency of results. Initially results for tractions are presented in Fig. 19 obtained
for the DRBEM and DIBEM using the simple radial basis function.
Now, the non uniform distribution of the domain forces caused the
6.3.2. Fitting procedure DRBEM performance not to be as efficient for boundary meshes
The next tests present the average percentage error for dis- with a reduced number of internal points as previously, due to the
placements given by the fitting approximation using DIBEM. In fact that the more complex distribution of the domain action
Fig. 18, the graph on the right shows the thin plate radial function needs a larger number of internal points for good approximation.
while the graph on the left is related to the simple radial function. The minimum error value found for the DRBEM using finer meshes
All internal points used to give information about the domain was also well below that of the DIBEM, confirming the expectation
force are excluded. that the latter is more effective for representing the domain forces,
The thin plate function presents a better solution for fewer since the requirement of a large amount of poles adversely affects
boundary nodes than the simple radial function. However, the the DRBEM results.
finer boundary mesh with simple radial function achieves the The rate of convergence of the thin plate radial function with
results obtained by the coarser mesh faster, once enough informa- the DIBEM was inferior to that obtained with the simple radial
tion points are introduced. This is justified because the quality of function, although for meshes with higher numbers of internal
C.F. Loeffler et al. / Engineering Analysis with Boundary Elements 50 (2015) 97–108 105

Fig. 18. Performance of simple radial and thin plate radial functions for the fitting procedure with the DIBEM in displacements results for the third example.

Fig. 19. Error curves for tractions with the DRBEM and the DIBEM using the simple radial function for the fourth example.

located in different parts of the domain, the fitting procedure


cannot do without some internal basis points in the final discrete
equation system to be solved (Fig. 21).

6.5. Fifth example: square rod submitted to uniform torsion

This example consists of a prismatic rod of square cross section


under uniform torsion, as shown schematically in Fig. 22:
The partial differential equation governing this problem is a
Poisson Equation, written strategically in terms of a potential
torsion function u(x1,x2), such that
∇2 uðx1 ; x2 Þ ¼  2Gθ ð19Þ
In Eq. (19), θ represents the angular displacement per unit length
and G is the shear modulus of the constitutive material of the rod.
Fig. 20. Error curves for tractions with the DIBEM using the thin plate radial
The distribution of shear stresses in the cross section is given by
function for the fourth example. ∂uðx1 ; x2 Þ ∂uðx1 ; x2 Þ
τx1 ¼ ; τx2 ¼  ð20Þ
∂x2 ∂x1
basis points the final results for the thin plate function tend to be
better, as shown in Fig. 20. The analytical expression for stresses in each direction is given by
2    
32Gθ 1 1 sin mπax1 cos nπax2
τxi ¼ ∑ ∑ 4δi1  
6.4.2. Fitting procedure aπ m ¼ 1;3;::: n ¼ 1;3;:::
3
m m aþ2 n
2 2

For this example, the behavior of the fitting procedure was only    3
satisfactory for the simple radial function, since the error curve for cos mπax1 sin nπax2 5
 δi2   ð21Þ
the thin plate radial function is not monotonic and its rate of n m aþ2 n
2 2

convergence is also very slow, as shown in Fig. 21.


It is noteworthy that the highest concentration of the domain The distribution of shear stresses over the cross section is given by
forces in the central region of the membrane was also favorable for the spatial derivatives of the torsion potential. The values of the
obtaining good results. In the cases in which these forces are shear modulus of the constitutive material of the rod, the angular
106 C.F. Loeffler et al. / Engineering Analysis with Boundary Elements 50 (2015) 97–108

Fig. 21. Comparison between the performance of simple radial and thin plate radial functions for the fitting procedure with the DIBEM in traction results for the fourth
example.

Fig. 22. Uniform torsion in a square rod.

Fig. 23. Error curves for shear stress using the simple radial basis function with the DRBEM and the DIBEM for the fifth example.

displacement per unit length and edges A are also considered


unitary in the simulations. The prescribed condition on all bound-
aries is the torsion potential equal to zero.

6.5.1. Interpolation procedure


Fig. 23 presents the results for the average percentage error in
shear stresses for the DRBEM and the DIBEM using the simple
radial function. As usual, an excessive number of internal points
compared to the number of boundary nodes decreases the
performance of the DRBEM. On the other hand, for the DIBEM,
the strong dependence on internal poles to achieve good numer-
ical results is now increased, especially for meshes with few
internal points. However, in this example, the performance of
the DIBEM was not notably superior to the DRBEM, since the
values of average percentage error only were reduced with the Fig. 24. Error curves for shear stress using the thin plate radial function with the
DRBEM for the fifth example.
increase number of internal poles and boundary nodes. A finer
mesh with 320 boundary elements is included to emphasize this
behavior. Nevertheless, the average error curves for finer boundary Fig. 24 presents the behavior of the thin plate radial function
meshes with the DIBEM are preponderantly convergent, whilst the with the DIBEM, which was not very close to that of the simple
DRBEM are not. radial function, particularly considering the results for finer
C.F. Loeffler et al. / Engineering Analysis with Boundary Elements 50 (2015) 97–108 107

Fig. 25. Performance of the simple radial and the thin plate radial functions for the fitting procedure with the DIBEM in stress results for the fifth example.

meshes. This behavior is due to the slight local disturbances main one and is recommended to reduce the size of the final
observed in the numerical results for nodes on the boundary, matrix system, which implies lower computer run times. Success-
which are composed exclusively of normal derivatives in this two fully, results show that the fitting requires a reduced number of
dimensional case. This hypothesis will be confirmed hereafter boundary basis points to reach reasonable results, although the
with fitting tests. internal distribution of the domain forces is the most important
factor to define the amount of internal points necessary for an
6.5.2. Fitting procedure efficient approximation.
Fig. 25 shows the performance graphs considering the fitting Global functions in a Goldberg [16] sense can also be imple-
procedure. The final values for the simple radial and thin plate mented together with internal poles, improving the quality of
radial functions are very similar and present the same level of numerical results in certain cases. The use of the modern CRBFs is
error reached in the third example. This confirms the hypothesis also immediate, without any of the difficulties observed with the
that the oscillations in the nodal results are distributed around the DRBEM. This class of functions can be applied not only for
mean values and can be eliminated by the fitting procedure, in reducing the computation effort, providing arrays with many null
which a lowest-order curve approximates the behavior of the elements, but also to reduce the risk of poor matrix conditioning,
domain function in an average sense. since a diagonal dominance is generated.
According to these features, the DIBEM must be tested for more
sophisticated applications, such as dynamic situations, given by the
7. Conclusions Helmholtz and Wave Equations. However, these advanced approaches
necessarily involve further research in addition to the successful
Nowadays, one important BEM challenge is to suitably model solutions of the problems which were examined in this article.
certain mathematical components that can appear in partial Not only potential problems can be solved with the aid of the DIBEM,
differential equations, such as domain source terms, for example. but also problems governed by Navier's Equation in stationary
In this context, the proposed DIBEM is an alternative for finding a conditions with domain actions and elastodynamic cases as well,
better approximation for solving the Poisson type equations. provided that accessible adaptations are implemented.
The DIBEM is mathematically more simple, robust and general
than the other available BEM alternatives that avoid domain
integration. Unlike the DRBEM, the proposed formulation does References
not require the construction of two auxiliary arrays multiplying
the H and G matrices; its method approximates directly the [1] Nardini D, Brebbia CA. A new approach for free vibration analysis using
boundary elements. In: Brebbia CA, (editor.), Boundary element methods in
complete kernel, similarly to what is done in the interpolation engineering; 1982, p. 312–26.
procedure, making use of only a primitive function. Just the [2] Partridge PW, Brebbia CA, Wrobel LC. The dual reciprocity boundary element
transformation of the domain integral into a boundary integral method. 1st ed.. . London, UK: Computational Mechanics Pub. and Elsevier
Applied Science; 1992.
makes the DIBEM different to a basic interpolation procedure. [3] Loeffler CF, Mansur WJ. Analysis of time integration schemes for boundary
Thus, a wider range of different radial functions can be used element applications to transient wave propagation problems. In: Brebbia CA,
without instability problems. editor. Boundary element techniques: applications in stress analysis and heat
transfer. UK: Computational Mechanics Publishing; 1987. p. 105–24.
On the other hand, because of this interpolation feature, the [4] Ramachandran PA. Boundary element methods in transport phenomena. 1st
importance of internal poles is increased in the DIBEM. Refined ed.. . London, UK: Computational Mechanics Publication and Elsevier Applied
boundary meshes may present worse results if few internal points Science; 1994.
[5] Shaback R, Wendland H. Using compactly supported radial basis functions to
are considered. However, considering enough internal points to solve partial differential equations. In: Chen CS, Brebbia CA, Pepper DW,
simulate the domain forces, the performance of the DIBEM is editors. Boundary element technology XIII. Southampton, Boston: WitPress;
superior, obtaining better results than the DRBEM in all examples 1999. p. 311–24.
[6] Karur SR, Ramachandran PA. Radial basis function approximation in the dual
simulated. In the DRBEM, when the number of internal basis
reciprocity method. Math Comput Model 1994;20(7):59–70.
points reaches a certain number in relation to the number of [7] Franke R. Scattered data interpolation: test of some methods. Math Comput
points located at the boundary, the results commonly degenerate, 1982;38(157):181–200.
increasing the percentage error. [8] Buhmann MD. Radial basis functions: theory and implementations. 1st ed..
New York: Cambridge University Press; 2003.
Many interesting resources may also be easily used in associa- [9] Schaback R. Error estimates and condition numbers for radial basis function
tion with the DIBEM without difficulties. The fitting scheme is the interpolation. Adv Comp Math 1995;3:251–64.
108 C.F. Loeffler et al. / Engineering Analysis with Boundary Elements 50 (2015) 97–108

[10] Wendland H. Piecewise polynomial, positive definite and compactly sup- [14] Brebbia CA, Dominguez J. Boundary elements—an introductory course. 1st ed..
ported radial functions of minimal degree. Adv Comput Math 1995;4:389–96. London, UK: Computational Mechanics Publication; 1992.
[11] Floater M, Iske A. Multistep scattered data interpolation using compactly [15] Wang JG, Liu GR. A point interpolation meshless method based on radial basis
supported radial basis functions. J Comput Appl Math 1996;73:65–78. functions. Int. J. Numer Methods Eng 2002;54:1623–48.
[12] Cheng AHD, Young DL, Tsai CC. Solution of Poisson's equation by iterative [16] Golberg MA, Chen CS. The theory of radial basis functions applied to the BEM
DRBEM using compactly supported, positive definite radial basis function. Eng for inhomogeneous partial differential equations. BE Commun 1994;5:57–61.
Anal Bound Elem 2000;24:549–57. [17] Stark PA. Introduction to numerical methods. 1st ed.. New York and London:
[13] Brebbia CA, Telles JCF, Wrobel LC. Boundary element techniques. first ed.. Macmillan Publishers and Collier Macmillan; 1970.
Berlin, Heidelberg: Springer Verlag; 1984.

You might also like