(K. Abe (Auth.), S. Kobayashi, N. Nishimura (Eds.) PDF

Download as pdf or txt
Download as pdf or txt
You are on page 1of 422
At a glance
Powered by AI
The document discusses strategies for boundary element analysis of slender bodies and thin plates, including sub-element division schemes and sub-region division methods to improve integral accuracy, as well as use of the direct regularization method to model fixed end conditions.

The strategies proposed include improving accuracy of numerical integrals between upper/lower planes and edges using sub-element division, using sub-element division for slender elements in 3D analysis of thin plates/shells, and using sub-region division for problems with cracks/notches.

Sub-element division schemes are used to improve accuracy of numerical integrals in quasi-singular integrals between upper/lower planes and edges of slender bodies.

S. Kobayashi N.

Nishimura Editors

Boundary
Element Methods
Fundamentals and Applications

s. Kobayashi .

N. Nishimura (Eds.)

Boundary Element
Methods
Fundamentals and Applications
Proceedings of the IABEM Symposium
Kyoto, Japan, October 14-17, 1991

Springer-Verlag
Berlin Heidelberg GmbH

S. Kobayashi
N. Nishimura
Division of Global Environment Engineering
Kyoto University
Kyoto, 606-01
Japan

ISBN 978-3-662-06155-8
Library of Congress-in-Publication Data
IABEM (Organisation). Symposium (1991: Kyoto. Japan)
Boundary clement methods: fundamentals and applications: proceedings of the IABEM
Symposiom, Kyoto, Japan, October 14-17.19911 S. Kobayashi and N. Nishimura, cds.
ISBN 978-3-662-06155-8
ISBN 978-3-662-06153-4 (eBook)
DOI 10.1007/978-3-662-06153-4
1. Boundary element methods -- Congresses
I. Kobayashi, S. (Shoichi). II. Nishimura, N. m. TItle
TA347.B69121991 620'.00I'SI53S-dc20 92-32708
This work is subject to copyright. All rights are reserved. whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations. recitation. broadcasting.
reproduction on microfilm or in other ways, and storage in data banks. Duplication of this publication or
parts thereof is pennltted only under the provisions of the Gennan Copyright Law of September 9, 1965.
in its current version, and pennission Cor use must always be obtained from Springer-Verlag Berlin Heidelberg
GmbH. Violations are liable for prosecution act under German Copyright Law.

e Springer-Verlag Berlin Heidelberg 1992


Originally published by Springer-Verlag Berlin Heidelberg New York in 1992
The USC of general descriptive names, registered names. trademarks. etc. in this publication does not imply,
even in the absence of a specific statement, that such names are exempt from the relevant protective laws
and regulations and therefore free for general usc.
Typesetting: Camera ready by author
61/3020-543 2 1 0 - Printed on acid -free paper

IABEM-91: International Symposium


on Boundary Element Methods
Organized by
International Association for
Boundary Element Methods (IABEM)
Sponsored by
Japan Society for Computational Methods
in Engineering (JASCOME)

Organizing Committee
Prof. S. Kobayashi (ChaIrman)
Kyoto University, JAPAN
Prof. L. Morino
University of Rome "La Sapienza", ITALY
Prof. W.L. Wendland
University of Stuttgart, GERMANY
Prof. F.J. Rizzo
Iowa State University, USA
Prof. T .A. Cruse
Vanderbilt University, USA
Prof. A. H-D. Cheng
University of Delaware, USA

Local Organizing Committee


S. Kobayashi, Kyoto University
T. Fukui, Fukui University
T. Honma, Hokkaido University
Y. Kagawa, Okayama University
N. Kamiya, Nagoya University
K. Kishimoto, Tokyo Institute of Technology
M. Kitahara, Tokai University
N. Nishimura, Kyoto University
K. Ohnishi, Tokyo Science University
M. Tanaka, Shinshu University
N. Tosaka, Nihon University
R Yuuki, University of Tokyo

Scientific Advisory Committee


Prof. S.N. Atluri
Prof. D.E. Beskos
Prof. T. Burczynski
Prof. J. Dominguez
Dr. F. Farassat
Dr. J.L. Hess
Prof. S. Kobayashi
Prof. L. Morino
Pre!. J.C. Nedelec
Prof. C. Polizzotto
Prof. P.D. Sclavounos
Prof. E.P. Stephan
Dr. J.C.F. Telles
Prof. J.O. WatSG!l
Dr. RB. Wilson
Dr. E.C. Yates, Jr.
Prof. F. Ziegler

Prof. P.K. Banerjee


Dr. H.D. Bui
Prof. T .A. Cruse
Prof. Qinghua Du
Prof. T.L. Geers
Dr. H.W.M. Hoeijmakers
Prof. G. Maier
Prof. S. Mukherjee
Prof. R Piva
Prof. F.J. Rizzo
Prof. RP. Shaw
Prof. M. Tanaka
Prof. N. Tosaka
Prof. W. Wendland
Prof. J.C. Wu
Prof. P.J. Zandbergen

Preface
The Boundary Element Methods (BEM) has become one of the most efficient tools for
solving various kinds of problems in engineering science. The International Association for
Boundary Element Methods (IABEM) was established in order to promote and facilitate
the exchange of scientific ideas related to the theory and applications of boundary element
methods.
The aim of this symposium is to provide a forum for researchers in boundary element
methods and boundary-integral formulations in general to present contemporary concepts
and techniques leading to the advancement of capabilities and understanding of this computational methodology.
The topics covered in this symposium include mathematical and computational aspects,
applications to solid mechanics, fluid mechanics, acoustics, electromagnetics, heat transfer,
optimization, control, inverse problems and other interdisciplinary problems. Papers dealing with the coupling of the boundary element method with other computational methods
are also included.
The editors hope that this volume presents some innovative techniques and useful knowledge for the development of the boundary element methods.
February, 1992
S. Kobayashi
N. Nishimura

Contents
Abe, K.
An h- adaptive boundary element method and evaluation of error
Ali, A. and D. Ostergaad
Implementation of FE-BE hybrid techniques into finite element programs

1
11

Amaya, K., S. Aoki and K. Kishimoto


Optimum location of electrode in cathodic protection system
Antes, H. and W. Cheng
Vibrations of Reissner plates by BEM and DjBEM: A comparison
Bassanini, P., M.R. Lancia, R. Piva and C.M. Casciola
Numerical approximation of boundary integral equations in three dimensional aerodynamics

30

Burczynski, T.
Shape sensitivity analysis of uncertain static and vibrating systems using stochastic boundary elements

49

Chahrour, A.H. and M. Ohtsu


BEM analysis of crack propagation in concrete based on fracture
mechanics

59

Cheng, A.H-D. and H. Antes


On free space Green's function for high order Helmholtz equations
Dominguez, J. and O. Maeso
Eoundary element model for the seismic analysis of arch dams
Fukui, T. and K. Tani
Stability of time domain boundary element method in wave propagation
problems
Futagami, T. and H. Ogata
Computational aspect of ground-water management
Hirose, S.
Dynamic crack contact analysis by boundary integral equation method
Hunt, B. and A.P. Adamson
Current status of the GENESIS methodology for knowledge-based treatment of transonic flows, with emphasis on shock fitting and nondissipative suppression of expansion shocks
Igarashi, H., M. Kuroda and T. Honma
A boundary element analysis of space charge fields
Ishiguro, S. and M. Tanaka
A time-stepping boundary element method applied to transient thermoelasticity

21

40

67
72
82

92
102
111

121
129

IX

Kamemoto, K and T. Mine


Calculation of aerodynamic characteristics of a heaving and pitching
airfoil by a panel-vortex method
Kanoh, M. and T. Kuroki
BEM- WFDM coupled analysis for groundwater resources

139

149

Kawamura, R
Approximation analysis on 3-dimensional coupled electric potential and
hole or electron by boundary element method

159

Kim, S.
The completed double layer BIEM: A boundary integral method for complex microstructures in a viscous fluid

167

Kimura, M.
Boundary element methods for crack problems

176

Kita, E. and N. Kamiya


An easy adaptive boundary mesh for 2D elastic problem

184

Kitahara, M. and K Nakagawa


Boundary element - node condensed finite element method for three
dimensional elastic wave propagation problems
Knopke, B.
The analytical integration of boundary integrals for plate bending

194

202

Kojima, K
Identification of corrosion profiles in computed thermal tomography by
a Volterra integral equation of the second kind

215

Leung, KL., H.H. Chyou and P.B. Zavareh


A mixed-hybrid variational formulation for coupling BEM and FEM in
elastostatics

222

Matsumoto, T., M. Tanaka and H. Hirata


Sensitivity analysis of steady-state heat conduction problems using
boundary element method

232

Morino, L., M. Gennaretti and P. Petrocchi


A general theory of potential aerodynamics with applications to helicopter rotor-fuselage interaction

242

Nishimura, N., A. Furukawa and S. Kobayashi


Regularised boundary integral equations for an inverse problem of crack
determination in time domain

252

Ohtsuka, K
Boundary integral equation defined on a crack approximated with voids
Olukoko, O.A., A.A. Becker, and RT. Fenner
Boundary element analysis of frictional slip behaviour in contact
problems

262
269

x
Onishi, K. and R. Kawamura
Three-dimensional long time prediction of nuclides migration by boundary element method
Piltner, R.
Trefftz-type boundary elements for plate problems
Poterasu, V.F. and N. Mihalache
Brittle fracture materials by BEM and R-functions
Sasaki, N., T. Kuwahara and T. Takeda
The sweeping primitive used for 3-D automatic boundary element mesh
generation

279

288
298
308

Sugino, R. and N. Tosaka


Direct differentiation approach to boundary element method for nonlinear water wave problems

314

Tabata, T., T. Kuwahara and T. Takeda


A formula for the boundary integral in the potential problem of twodimensional anisotropic materials and its evaluation

324

Takeda, T. and T. Kuwahara


BEM analysis applied to the electric field near the edge of compound
dielectric materials

333

Tanaka, M., M. Nakamura and H. Ishikawa


Analysis of certain inverse problems in transient heat conduction by the
boundary element method

343

Taniguchi, T. and Y. Fukuoka


Grid generation of arbitrary 3D surface
Tosaka, N., T. Honma and T. Manabe
Integral equation analysis of viscous flow between two rotating coaxial
disks
Tosaka, N., K. Kakuda, H. Yoshikawa and A. Anjyu
The coupling approach of boundary and finite element methods to incompressible viscous flow problems
Yamashita, S. and N. Tosaka
Boundary element method for couple stress theory of elasticity
Yu, X., M. Isobe, A. Watanabe and K. Sakai
Analysis of wave motion over submerged plate by boundary element
method
Yuuki, R., G.Q. Cao and K. Maekawa
Accurate BEM elastostatic analysis for very slender body and thin plate

353
363

373

383
393

403

An h-Adaptive Boundary Element Method and Evaluation


of Error
K.Abe
Department of Construction, Niigata University,
Igarashi-2-Nocho,Niigata,950-21,Japan

SUMMARY
In this paper an h-adaptive refinement process is constituted for boundary
element method.

A new res idue is introduced to define error indicator and

estimator. In addition, nodal error is evaluated by using the fact that the
relation between these error and the residue is expressed as a matrix equation.
The efficiency of the proposed strategy is proved through numerical example.

INTRODUCTION
Recently adaptive mesh refinements in boundary element method had been proposed
by a number of researchers, based on the ~o called h- [1J, p- [2J. hp- [3J and
r-[4] versions. Although the hp-version converges faster than the h- and
p-versions. implementation to the existing code is easier in the h-version.
Thus the development of the h-adaptive boundary element method has a meaning.
Constitution of an effective error indicator and estimator is significant for
the development
been defined by
the discretized
from a boundary

of adaptive mesh refinement method. So far, these indices had


many researchers based on the residue "r" which is defined from
boundary integral equation [3J. The residue r can be evaluated
element solution. However r is to be zero at all collocation

points, if the collocation technique is employed in boundary element method.


Thus it is necessary to evaluate the residue at several points on each element,
in order to estimate the distribution of r properly. This process requires a
computational effort.

Then modified versions were proposed to improve the

efficiency of the evaluation and computational economy [5-7J.

Evaluation of error was studied from a point of view of development of error


indicator and estimator so far [8,9J. Although evaluation of nodal error has
advantages in practice, it is impossible to know the nodal error and to correct
tho numerical solution directly by those proposed evaluation technique.
The author introduce a new residue "R" and define an error indicator and
estimator based on this residue in h-version boundary element method. Since the
residue R is not null at each collocation point, it is convenient for
evaluation of this distribution. Furthermore, the relation between the error
and R at each collocation point is expressed as a matrix equation. Then an
error evaluation procedure is proposed by using this relation. In the last
section, a numerical example is shown to demonstrate the efficiency of this
method.

DEFINITION OF RESIDUE R
Let us consider the following two-dimensional potential problem.

AU = 0

u =

q =

in Q
(1)

u on fu
q onr q

where u and q is the potential and flux respectively, Q is the domain, and
is the boundary. u and q is defined as iT on r u and as if on r q respectively.
The boundary integral equation for this problem is expressed as
c(x)u(x)

+ Jfq*(X,Y)U(Y)dfy

- Jru*<x,y)q<Y)dry
(X

=0

(2)

r)

where u* and q* are the fundamental solution for the Laplace's equation, and
C(x) is the free-term.
The boundary value functions are approximated by
U(X) =
""

'" q,.(X)U.""
N

4..

(3)

where u and 'Q are the approximated boundary functions. N is degree of freedom,
and 1> i is shape function. In this paper, the function 1> i is assumed to be
constant on boundary element r i , then the boundary values are approximated by
piecewise constant functions.

Substituting eq. (3) in eq. (2) we obtain the residue r:

By enforcing the res idue to be zero at collocation points, we obtain a


boundary el ement equation
(5)

in which x. is the collocation point, u(x)= Uj on the boundary element


Ci=C( x;).
The boundary element equation is reduced to a matrix equation

AX :: b

: Exact solution

t : Interpolation o f

f
solution

r.

I
e

i-I

j ,

and

(6)

where A is a coefficient matrix, X is constituted from the unknowns of


and b is a known vector calculated from the prescribed part of u and q.
Subtracting eq.(5) from eq.(2), the following equation is obtained

? : Interpolation of
f : Boundary elellent

+1

Fig.l. Definition of each function.

u and q,

c. (u.
I

~ y -u.I ) + J q *(Xi ,y)(u-u)dr

Jr *

~ y = 0
U (Xi ,y)(q-q)dr

(7)

(i=i,,N)

where Ui=U(Xi).
Introducing an interpolation of u and q by using the shape function

as

depicted in Fig.l :
N

= L$(X)q.

G<X) = L$.(X)U.
I

(8)

then the equation (7) can be rewritten as follows


(9)

where EU=IHl', Eq=q-q, Eui=Ui-lli, and the new residue R is defined by

Eq. (0) implies that the residue R is describing the appropriateness of


discretization of a solution in the boundary integral equation.
The matrix form of eq. (9) [10] is

AE"

(11)

where i is a nodal error vector constituted from Eui and Eqi.


From equation (9) or (11) we obtain a relation between the nodal error and
the new residue. These equations imply that error indicator and estimator can
be defined by the residue R.
Since R is constituted from the exact solution and its interpolatidn as in
eq. (10), we can not evaluate R exactly. Thus the functions u-u and q-q- are
approximated by using the boundary element solution as

on

rI

(12)

r.

where f is meaning u or q,

collocation points, and

defined by quadratic shape functions I/J ikon the

middle element

i,

1-f is

is an interpolation of

as depicted in Fig.!.

over the three

The approximation of Ri is obtained as


"'-'

Rj =

where
)1,
I\jj ,=

Jr. q

L. (-kI.J +A")
IJ

(13)

ill

(Xj

~.'V

,y)(u-u)df

)2,
I\jj
,=

J '"
fJ

U (Xi

It"'
,y)(q-q)dr

(14)

ERROR INDICATOR AND ESTIMATOR


In this paper, error indicator A i j is defined from eq, (14) as

A"
'=
IJ'

1
2
-~ IJ
.. +t\"IJ

(15)

l i j is a measure of contribution of element r j to the residue Ri. Since


the residue Ri at a nodal point xi is not always reflecting the degree of

approximation on

i,

elements refined in the next step are decided as follows

Find the refined element


for nodal point xi :

r j:

for which
such that

,1

IRi

j=Max{ l

i k,

k=l, " " N}

I ;?; tlMax{IRkl,k=l,

,N}

(0< tl ;:::;;1)
Here the parameter tl influences on the number of elements refined at one
step and on the computational economy. Through numerical experiments the
optimum value of 0,5 was found for {3,
Error estimator is defined as
(16)

where hi is element length of r i ,


Since the residue R is not null at each point x"
approximation of a normalized norm

eq. (16) wi 11 be an

From eq. (11) we can deduce the following inequality

where I'l, and 1'1 2 are constants independent of the discretization, and

It i

is

defined by

L I EUi I hj

II Eu II :=
'V

Jr !U-U!2 df

(18)

Thus it is possible to utilize 17 as an estimator. The relation in eq. (17)


had been confirmed through numerical experiments.

EVALUATION OF NODAL ERROR


By using 'R for R in eq. (11), we can evaluate a nodal error i approximately.
However in some cases it will be insufficient, since 'j( is obtained by a
boundary element solution, then polluted by the nodal error f. To improve the
accuracy of evaluation of the error, R is decomposed into two parts in eq. (13)
as:

where ~i is defined as

In eq. (19) and (20), the summation in


adjacent three elements. Eq.(20) is obtained
q by quadratic interpolations of them in
solutions are piecewise quadratic functions,

the index "k" is carried out for


by approximating the solution u or
eq. (10). Thus, when the boundary
the residue R is obtained exactly

byt

Using ~ for R in eq. (10), we obtain the following equation


jO\

a ..
E = R.I
IJ J
=

If. - L<L J
I

k fk

q cpok df
J

EUj
A

+ L<Ljr; u cp.k df ) Eqj


j
k k J
where ai j is an element of the matrix A.

(21)

Arranging the above equation, the following matrix equation is obtained

AE =

(22)

where

aIJ.
(23)

or
in which a;j is an element of the matrix A, and when u is unknown on r
defined by the upper expression; otherwise by the lower one.
The nodal error can then be evaluated properly by using eq.(22).

,j

aij is

NUMERICAL EXAMPLE
To verify the efficiency of the new residue in adaptive refinement strategy, a
numerical test is demonstrated for the following potential problem.
The geometrical and boundary conditions are depicted in Fig.2. The boundary
is discretized into four constant elements at the first step.
Tolerance of O.5XIO- 3 for 'YJ was specified in the adaptive refinement. The
boundary was discretized into 99 elements at the final step as shown in Fig.3.
Improvement of accuracy in the adaptive refinement process is compared with a
series of uniform refinement in Fig.4. As shown in the figure reduction of 'YJ
in the proposed method is faster than in the uniform refinement. In Fig.4 the

n =i-x

99 ELEMENTS
o
!I

\:1

II
o

10'

u=o
Fig.2. Configuration and boundary
condition of the example.

Fig.3. Adaptive mesh refinement in the


final step for tolerance of O.5xIO- 3

Uniform refinement
Present (Exact)
Present (Eq. (12))

10'

Fig. 4. Convergence of the estimator

7].

estimator evaluated in the program based on the numerical solution is shown and
compared with the one obtained from the exact solution. Through this comparison
it is found that the evaluation of distribution of the error by eq. (12) is
sufficiently enough for calculation of the error estimator.
Distribution of nodal error along each boundary is shown in Fig.5. In these
figures nodal error evaluated by eq. (22) and exact error were plotted. These
figures show that the evaluation of nodal error by eq. (22) is excellent. The
error is concentrating near the corner (x=l.y=l). Then the mesh was divided
very fine in the neighborhood of the corner as shown in Fig.3. It implies that
the error indicator ~ i j is available in the adaptive mesh refinement strategy.
CONCLUSIONS
An h-adaptive boundary element method and evaluation of nodal error has been
constituted based on a new residue R.
The residue R was defined as an influence of discretization of a solution on
the boundary integral equation. Error indicator and estimator were constituted
from the residue R at each collocation point. R was evaluated approximately
from a boundary element solution.

XlO -

-0- Exact
---G-- Eq. (22)

- 0 - Exact

--D-- Eq. (22)

0'0

-1

-1

-2

-2
(a)

( b)

xlO-<
xlO- 7

~'f'"

0'

0 IiarJIJ.._It8t!I4-......."I!F'~t=iHj::::e::.:::1

-1

-2

-3

--0-- Exact
- --0--- Eq. (22)

--0- Exact
--D- - Eq. (22)

~::l

::l

-4
-5
- I)

-3
(c)

(d)

Fig.5. Comparison of nodal error evaluated by eq. (22) with exact one along each
boundary. (a)error of flux along y=O (b)error of flux along x=l (c)error
of flux along y=l (d)error of potential along x=O.
Relation between a nodal error and the residue was expressed in a matrix
equation. Then an evaluation of nodal error was proposed based on that
relation. To improve accuracy of the evaluation of the error, influence of the
error on the residue was removed.
The efficiency of the proposed error indicator and estimator and reliability
of the nodal error evaluation has been proved by a numerical experiment.

REFERENCES
1. Rencis,J.J. and Mullen,R.L. :A self-adaptive mesh refinement technique for

boundary element solution of the Laplace equation. Compo Mech. 3(1988)309-319.

10

2. Alarcon,E.,Reverter,A. and Molina,J. :Hierarchical boundary elements. Compo ~


Struct. 20(1985)151-156.
3. Rank,E.:Adaptive h-,p- and hp-versions for boundary integral element
methods. Int.J.Num.Meth.Eng. 28(1989)1335-1349.
4. Ingber,M.S. and Mitra, A. K. :Grid optimization for the boundary element
method. Int.J.Num.Meth.Eng. 23(1986)2121-2136.
5. Parreira,P. : Further developments on error indicators and estimators for
adaptive hierarchical boundary elements. Advances in Boundary Elements (C.A.
Brebbia and J.J.Connor Ed). Vol. 1(1989)105-122. Springer-Verlag.
6. Cerrolaza,M. :Further applications of p-adaptive boundary elements. Boundary
ElementsX (C.A.Brebbia Ed). Vol.l(1988)123-135.Springer-Verlag.
7. Parreira,P. : Residue interpolation for error estimation in p-adaptive
hierarchical collocation boundary elements. Boundary Elements in Mechanical
and Electrical Engineering (C. A. Brebbia and A. Chaudouet-Miranda Ed). (1990)
483-498. Springer-Verlag.
8. Alarcon,E. and Reverter,A. :p-adaptive boundary elements. Int.J.Num.Meth.Eng.
23(1986)801-829.
9. Rencis,J.J. and Jong,K.Y.:Error estimation for boundary element analysis. J.
of E.M.(ASCE). 115(1989)1993-2010.
10. Iso, Y., Takahashi, S. and Onishi, K. :Numerical convergence of boundary
solutions in transient heat conduction problems. Chap.l in Topics in Boundary
Element Research (C.A.Brebbia Ed). Vol. 3(1987)1-24. Springer-Verlag.

hnplementation of FE-BE Hybrid Teclmiques into Finite


Element Programs
Ashraf Ali and Dale Ostergaard
Swanson Analysis Systems. Inc. P.O.Box 65, Johnson Road
Houston. PA 15342-0065

Infinite boundary elements (]BE) in both 2-D and 3-D have been implemented in a general purpose finite element
program for use in conjunction with finite elements (FE) to solve unbounded magnetic field problems. The
advantage of the !BE's, in addition to reduced model size. is that they do not destroy the bandedness of the finite
element matrices as do boundary elements (BE). In contrast to the infinite finite element (]FE) usage, these !BE's
do not require definition of nodes in excess of the ones already introduced for the interior FE domain, thus
simplifying modelling efforts further. The !BE formulations presented here produce unsymmetric matrices. An
attempt is made to symmetrize the !BE matrices. The example problems presented demonstrate that the loss of
accuracy in solutions due to the symmetrization is insignificant

Introduction
Electromagnetic field problems are governed by the Maxwell field equations. These equations are solved by the
finite element method using potential functions [1]. For unbounded exterior regions, the magnetic potential and the
flux are theoretically zero at infinity. The finite element method attempts to solve this class of electromagnetic
field problems by truncating the exterior domain by placing the model boundary at a finite distance from the region
of interest and setting the potential there to zero. Often, a large distance is required to attain high accuracies in the
region of interest. Nevertheless. the finite element method proves to be the most versatile technique, especially for
solving problems containing inhomogeneities and/or nonlinearities.
Boundary elements, on the contrary. naturally model the behavior at infinity and are especially suited for linear and
homogeneous domains. Consequently, a combined FE-BE technique appears to be an optimal procedure for
modelling exterior domain problems. In this technique, the domain containing material nonlinearities,
inhomogeneities and current sources can be modelled by finite elements. The surrounding air, extending to infinity
and governed by Laplace's equalion, is modelled using boundary elements. A considerable amount of research bas
been carried out in this area [2~.
The above mentioned procedure suffers from a serious drawback, especially so in a general purpose finite element
environment, in that the boundary element matrices destroy the bandedness of the finite element matrices.
Recently, Kagawa, et aI. [7] presented an infinite boundary element in 2-D which preserves the bandedness of the
finite element matrices. The formulation bas been extended to 3-D to solve 3-D potential problems in an
unbounded domain [8]. These infinite boundary elements in 2-D and 3-D have been implemented in the
ANSYS program to be used with finite elements for solving field problems.
It is to be noted that other techniques exist in the literature to model the far-field effects, viz., ballooning and
infinite finite elements. Ballooning was developed by Silvester, et al. [9] in which the outer boundary of the finite
element model is moved outward from a center location by means of a recursive technique. Infinite finite elements
were introduced by Ungless [10] as well as by Bettess [I1J. These elements use either a lIr or an exponential type
of decay function to model the far-field effects. Owing to the intrinsic conceptual accuracy in modelling the

12
behavior at infinity, an infinite boundary element that preserves the bandedness of the finite element matrices is
certainly a better choice and is singled out for teview of its implementation into finite element packages.
~

The theory of the infinite boundary elements both for 2-D and 3-D cases is briefly summarized here. For a

detailed description of the theory of mE, the reader is tefemd to Kagawa, et al. [7] for the 2-D case and
Kaljevic' , et al. [8] for the 3-D case. The objective bere is to develop a 2-noded line mE in 2-D and a 3-noded
triangular or 4-noded quadrilater.l1 surface mE in 3-D that can be used with 2-D and 3-D solid finite elements
such that the bandedness of finite element assembled matrices is preserved.
For unbounded field applications, the problem domain is set up to consist of an internal FE domain. OF. and a
series of exterior BE subdomains, OB, as shown in Figures (1) and (2) for 2-D and 3-D respectively.

Figure 1: Definition or Z-D IBE Subdomaio and the Characteristics of the Z-D mE

s
..... . ; I - - - - y

x
(a)

(b)

Flgure 2: A ~D mE Snbdomaio and the Corresponding Subeiements

A.

B9!JM a'! In!eeraI Equations for Subdomajn

A set of boundary element equations can be written for each BE subdomain. 0B. In 2-D the BE segments in OB
become 2-noded lines, e.g.. lines 1-2.2-3.4-1 and 3-4; line segment 3-4 is assumed to be located at infinity

13
(Figure 1). In:>-D the BE segments in 0B are surfaces (Figure 2). All equations are developed with reference to a
basic triangular interface 1-2-3. A quadrilateral surface is easily treated by subdividing it into four overlapping
triangles. The surface 4-5-6 in Figure 2 is assumed to be located at infinity.
The BE subdomain OB is governed by Laplace's equation. The fundamental solutions are given by:

2-D:

." ... (r)

=~

3-D:

." ... (r)

=-

where: v

1.(~)

(1)

47rv

(2)

'= material constant and r =distance between field and source points

For the 2-D case, the potential, and flux q =

o.p / an are interpolated linearly on the interface between the points 1

and 2 (Figure 1). On the segments 2-3 and 1-4, the potential and flux are taken to be proportinal to 1/r and 1/r2
respectively, where r = variable radial distance from 2 towards 3 or 1 towards 4. In:>-D the potential and flux are
also interpolated linearly on the interface triangle 1-2-3 (Figure 2), whereas the decay functions for potential and
flux are assumed to be proportional to 1/r2 and I/rl respectively on any subelement, e.g., 1-2-5-4, from the edge

1-2 towards the edge 4-5. The potential and flux vary linearly on the edges 1-2,2-3 and 3-1.
The boundary element equation is now written for the subdomain 0B:

C,

where: Ci

t/Jt(~)

fr.

.,,(x) q'" (x,~) df(x)

fr.

q(x)

t/J .. (x,~)

(3)

df(x)

=boundary integral jump term, x = field point, ~ =source point and r B =boundary of 0B.

B Flux Conservation Equation

Since there is no source or sink in the subdomain, 0B. the total flux over the subdomain must be zero:

f.

q(x) df(x)

=0

(4)

After discretization, the equations (3) and (4) can be expressed in terms of nodal unknown potentials and fluxes.
These discretized equations are then added together to yield:
[H]!.p\ = [GJ{q,,1 + [G,J!q..1

(5)

where:

{.pI = L "'1

""J T
{q.J = L q., q..'J T
(q..) =

Lq..1 q... J T

(.pI =
2-D

L.p,

.p. t/J. JT

{q.1 = L q., q..

q..J T

(3) (1) (1) (2) (2) (3)


!q,1 = Lq" q,. q... q... q...

q..,J T

3-D (Figure 2b)

In 2-D all the integrations of equations (3) and (4) are performed in closed form. For the 3-D case the integrations
on the triangle 1-2-3 and those in the radial infinite direction on the subelements 1-2-5-4,2-3-6-5 and 3-1-4-6

14
are also perfonned in closed fonn. The diagonal tenns of [H] in equation (3) vanish except for the jump tenn,
which is evaluated in closed fonn. The integrations in the finite direction for the off-diagonal tenns of [H] and the
non-singular tenns of [0.,;] in equation (3) are perfonned using Gaussian quadratwe, wbereas those for the tenns
of [0.,;] containing logarithmic singularities are perfonned using a combination of standard Gaussian quadrature
and logarithmic quadrature rules [8].

Since the finite element equations are posed in tenns of nodal potential. alone, the fluxes 'In and q't must be
eliminated from equation (5).
C. Eqpllibrium at the Interface

In order to eliminate the fluxes 'In and q't, the equilibrium condition between the FE and BE domains must be

considered. Note that the compatibility of potentials is easily satisfied at the interface, as both FE and BE
fonnulations use linear shape functions to interpolate potentials on the interface. The equilibrium conditions are
implicitly enforced by computing an equivalent nodal load vector (PI due to the fluxes 'In and q't. To achieve this,
the energy flow quantity, W, from OB is written as:
W

= r. q(x) tP(x) dI"(x)

(6)

W should also equal the energy flow due to the equivalent oodalload, IP I,

W=itP}1p}

(7)

where: I tP I = nodal potentials, and

ip} = /p.} + ip.}

(8)

The expression for this load vector can be developed in a number of ways. Here the expressions are developed
only for the 2-D case. For the details of the alternative assumptions and the derivation of the equations for {Pnl
and IP't I for the 3-D case, please refer to Kagawa, et aL [6] and Kaljevic' ,et aL [8] respectively. Inserting the

shape and decay functions for

tP

and q for the BE segments 1-2,2-3 and 1-4 (Figure 1) into equation (6) we

obtain, after performing integration,


(9)

Equating this to equation (7) yields:

/pl

= [A"j{q"l

+ [A.j{Cir}

(10)

where:

[A,,] = ~

U ~] [A.] -2 [rio]
o
1

f2

(11)

L =length of BE segment 1-2


D " - q. - q. Geometric ielariOW!

In order to complete the coupling process of the BE equations with the FE equations it still remains to express the

nux vectors {'In I and {q't I in terms of the nodal potentials I tP

I. Note that the fluxes Iq't I are introduced into the

15
problem by the artificial boundaries 1-4 and 2-3 (Figure 1) for 2-D or 1-2-5-4, 2-3-6-5 and 3-1-4-6 (Figure 2a)
for the 3-D case. The normal gradient qn at any nodal point can be uniquely computed in terms of flux
components along the radial r direction and those along the normals to the artificial boundaries meeting at that
point. Consider the 2-D case for simplicity (refer to Kaljevic , et al. [8J for 3-D derivations). At node i,
(12)

q'lcosa,-~sina,

But,

= -1', -

[,

Therefore, equation (12) reduces to:

q..

= q~

sin a,
cos a, + 1'1 - [,

which, written for node 1 and 2 (Figure 1) becomes,


where:

{q,,1 = [q {q..l + [D]{tf>l

(13)

[ sin

aO,/r,

sin

0]

(14)

a,IT,

E. Eq,uiva1ent Coefficient Matrix !KJ


All the ingradients are now available to form an equivalent coefficient matrix for 0B. It can be done by simply
eliminating {~l and {q'tl vectors from equations (5), (10) and (13) in order to arrive at:

[pI = [KH1'1

(15)

in which the coefficient matrix [K] is given by:


~

[Kl

= [A.] ([C}[G.:r-'[G.l + [l])-,([C}[G.r'[H] + [D])


+ [A.] ([G.][q + [G,Jt([H] - [G.J[Dl)

(16)

H!;
[K] = [A.][q + (lA.][D][A.] + [1]) ( ([Gol[C] +

(IHJ - [GoJ[C])
where: [IJ

=identity matrix.

[o.J) [A.1

t
(17)

16
'Ibis coefficient matrix can conveniently be assembled with the finite element matrix to obtain banded global
coefficient matrix. Note that the BE equiValent matrices are generally unsymmetric. If desired. they can. however.
be symmetrized by computing the off-itiagonal tenns as:
(18)

The IBE fonnulation has been implemented in the ANSYS finite element program [12] for general use in 2-D and

3-D linear/noolinear magnetic/thennal field analysis. The 2-D IBE element, known as 1NFIN9. is represented as a
2-node "line" element that can be defined on the surface of the companion first order quadrilateral/triangular

magnetic field element PLANB13 (or PLANE55 for thermal). The 3-D IBE element. designated as lNFIN47. is
represented as a 3-noded triangular or a 4-noded quadrilateral element that can be used with the first order 3-D
magnetic solid element SOLID96 (or SOLID70 for thennal).
Example 1: Overlapping Cylinder Problem
The first example problem consists of calculating the 2-D magnetic field fonned by intersecting two infinitely long

cylinders with opposite current density +1. and -1. The problem is commonly referred to as the overlapping
cylinder problem. A detailed study on the problem that considers a finite element-only solution to the unbounded
domain configuration can be found in reference [13].
A unifonn magnetic dipole field is formed in the region of intersection oCthe cylinders as shown in Figure 3. The

exact solution of the field in the overlap region is given as:

B,

=p.

(1.) (s)/2

(19)

where Po =4.1r X 10-7 Him and s is the cylinder offset as shown in the figure.
A study was performed on the solution accuracy in the overlap region by varying the radial location of the outer
boundary of the .finite element model (and thus the location of the infinite boundary element). Results for By were

compared at the central field location (x,y) = (0.0). Also recorded was the maximlUll variation of calculated field
By over a small bore region of radius 0 < r < O.Olm.

s
R

=
=

9.6mm
25mm
40mm < Ib < 120mm

Figure 3: Overlapping Cylinder Collfiguratioo

17
The finite element mesh was constructed automatically using the solid meshing capability in the ANSYS program
which fills the region with quadrilateral and triangular elements. The outer boundary at fb is meshed with the mE.
For each test case, the mesh density specification was left unaltered. Figure 4 illustrates the mesh for an outer
boundary radius. fb = O.04m. The solution flux lines are displayed in Figure 5.

Figure 4: Mesh for rb = 40 mm test case

Figure 5: Flux Lines for rb = 40 mm test case

Table I compares the solution results for several test cases of varying fb. An outer boundary placement of less than
2-1f2 diameters is shown to give a central field accuracy of one part per million and a field variation in the central
bore (0 < r < Olm)
.
of 15 parts per million. This is in contrast to a central field accuracy of one part in 4.000 for an
outer boundary placement of 20 diameters in a finite elemenHmly analysis using a Dirichlet boundary condition
[13].
Table 1
Overlapping Cylinder Test Case Results

Central Field Accuracy for Varying Outer Radius (rb) (Theoretical Solution: By 2.533380T)
By(T)
Part per Million
Part per Million
% Error
No. of Elements
Variation 0<r<10 mm
Accuracy
1582
None
2.533581
.00793
2869
79
162
2.533485
00415
40
4006
41
12
2.533397
.0067
9894
7
80

fb (mm)

120

17419

2.533383

.00012

15

lnfinite Boundary element placed on exterior surface of cylinder


Example 2' MalWfltic flux Density Due to Current in a Solenoid
A solenoid of length L = 0.254 m. radius R = 0.0127 m, and comprised of N = 231 closely spaced turns of the coil
is considered next. The coil carries a current of I

=0.5 amps.

The permeability of the air is taken to be II.

=41r x

10-7 HIm. The solenoid was modelled as an equivalent permanent magnet. The magnetic formulation makes use
of a reduced scalar potential [12] . Due to the axial and the longitudinal symmetry of the problem. only a 5 degree
wedge of half of the length of the solenoid is modeled in 3-D using a 5 x 25 mesh of eight-noded linear finite
elements. The analysis is done for three different discretizations of the unbounded domain: (a) the solenoid is
modeled using a 5 x 25 mesh of finite elements as shown in Figure 6 and infinite boundary elements are placed
directly at the outer surface of the solenoid; (b) the finite element mesh is extended to cover a region of dimensions
0.0899 m x 0.0899 m as shown in Figure 6 and the infinite boundary elements are placed at the outer surface of this

18
region; and (c) the finite element mesh is further extended to cover a region of dimensions 0.127 m x 0.127 m. as
shown in Figure 6. No infinite elements are used in the case (c) and the same density of the mesh in the inner
0.0899 m x 0.0899 m domain is used as in case (b). The results for the tlux density along the axis of the cylinder

are shown in Table 2 for the three analysis cases. The closed fonn analytical solution [14] for this position is also
shown in Table 2 for comparison. It is noted that an excellent agreement of the present results with the analytical
solution is obtained for the case when a portion of the unbounded domain is modeled using finite elements before
employing the infinite boundary elements.
y

5 mE's on
the top face

~[
~

~I I.

10mE'son

the top face

5xl0 5xlO

35 mE's on

the side face

5x25 5x25 ,/
mesh mesh

5x25
mesh

I-.J

0.0889

0.0127

.1

~[ ::

mesh

I--~---'---x

I.

(b) Mesh 2

(a) Mesh 1

8x25
mesh

0.127

.1

(c) Mesh 3

Figure 6: The Flnite Element Disc:retizatiODS Employed in the Solenoid Problem

Table 2
Magnetic Flux Density due to Current in a Solenoid
y [10-2m]
0.000

0.508
1.016
1.524
2.032
2.540
3.048
3.556
4.064

4.572
5.080
5.588
6.096
6.350

Meshl
1.1159
1.1156
1.1146
1.1130
1.1109
1.1083
1.1047
1.0992
1.0890
1.0680
1.0231
0.9324
0.7830
0.6966

Magnetic dux density [lo--:7fJ


Mesh 2
Mesh 3
Exact Solution[ 14J
0.8570
1.1207
1.1120
1.1196
0.8570
1.1203
1.1184
0.8569
1.1190
1.1162
0.8567
1.1167
1.1127
0.8563
1.1132
1.1108
0.8554
1.1078
1.0998
0.8535
1.0996
1.0879
0.8493
1.0870
1.0691
0.8402
1.0667
1.0373
0.8202
1.0326
0.9792
0.7765
0.9720
0.8686
0.6858
0.8621
0.5279
0.6829
0.6805
0.3855
0.5711
0.5689

Example 3' Go and Retwn Conductors - Test ofSvmrnetrir.atlon

It was mentioned earlier that the mE coefficient matrices are, in general, unsymmetric. A "go" and "return"
conductor problem is considered in order to test bow the results are intluenced when the unsymmetric coefficient
matrices are symmetrized. The conductors are assumed to be infinitely loog and carry uniform currents in opposite
directions (Figure 7). One-quarter of the domain is modelled. Homogeneous Dirichlet ('" =0) and Neumann

19
( &p / an = 0) boundary conditions ue applied at the edges X=O and Y=0 respectively. Should the outer boundary
be taken as a circular arc centered at the origin and meshed with the line mE's, then the coefficient matrices of
these mE's would all be symmetric. Instead a square boundary is chosen in order to produce unsymmetric mE
coefficient matrices (Figure 8).

5mB's
on
this side

~--4---~---r---r~~---- x

a-2mm

b .. 8 mm
Current I OK 1.0 Amp
Permeability of air 110 41t x 10-7 H/mm

Conductor Bouooary

Figure 8: Mesh for Go and Return Conductors

Figure 7: Go and Return Conductors

Problem
Table 3
Potentials at the boundary of the right conductor (Figure 7) of the go and return conductor problem.

cjI* 10-6
8degrees

Unsymmetric mE Symmetrized mE
Matrix
Matrix

QosedForm
[15)

% Error

0.3177

0.3177

0.3219

1.30

11.25

0.3171

0.3170

0.3213

1.31

22.5

0 .3156

0.3156

0.3194

1.19

33.75

0.3125

0.3125

0.3163

1.20

45

0.3085

0.3084

0.3120

1.12

56.25

0.3033

0.3033

0.3065

1.04

67.5

0.2969

0.2969

0.2999

1.00

78.75

0.2893

0.2894

0.2921

0.96

90

0.2813

0.2813

0.2833

0.71

101.25

0.2718

0.2718

0.2737

0.69

112.5

0.2615

0.2615

0.2635

0.76

123.75

0.2508

0.2508

0.2530

0.87

135

0.2408

0.2408

0.2429

0.86

146.25

0.2321

Q.232 I

0.2337

0.68
0.80

157.5

0.2245

0.2245

0.2263

168.75

0.2200

0.2200

0.2214

0.63

180

0.2184

0.2184

0.2197

0.59

As expected, the unsymmetry was seen to be mild. Two sets of results ue produced with this model: (a) solution
with unsymmetric mE matrix, and (b) solution with symmetrized matrix. Results ue reported in Table 3 along

20
with the closed fonn analytical solutions [15]. It is seen that the solutions remain unaltered despite symmelrization
of the mE matrices. It is to be noted that if the mE's are not used for the mesh of figure (8), it produces
completely wrong results.

Conclusions
An infinite boundary element formulation is presented in 2-D and 3-D in order to solve unbounded field problems

in an efficient manner. 1be fonnulation is based essentially on the boundary integral method. 1be objectives of
the present formulation are: (1) to develop a local equivalent coefficient matrix involving only the potential

degrees of freedom associated with the nodes of the exterior edge of a single finite element, and (2) to allow a
straightforward coupling of the boundary element equations with the finite element equations. Since these
boundary integral equations are developed locally, in contrast to the usual global approach, they preserve the
bandedness of the assembled finite element matrix. 1be use of the infinite elements result in considerable
reductions to the model size for unbounded problems. A further advantage of the infinite boundary elements, in
contrast to the infinite finite element usage, is that they do not require definition of nodes in excess of the ooes
already introduced for the interior finite element domain. A few example problems are discussed to demonstrate
the effectiveness and accuracy of the infinite boundary elements. It is also demonstrated that the infinite boundary

element coefficient matrix can be symmetrized without significant loss of accuracy in solutions.
References
1.

P. P. Silvester and R. L. Ferrari: "Finite Elements for Electrical Engineers", Cambridge Univ. Press, 1983.

2.

O. C. Zienkiewicz. D. W. Kelley and P. Bettess: "The Coupling of the Finite Element and Boundary Solution
Procedures",Intemationai Journal for Numerical Methods in Engineering, VoL 11, pp. 35.5-375, 1977.

3.

C. A. Brebbia and S. Walker: "Boundary Element Tedmiques in Engineering", Newne&-Butterworth. 1980.

4.

S. J. Salon: "1be Hybrid Finite Element-Boundary Element Method in Electromagnetics", IEEE


Transactions on Magnetics, Vol. Mag-21, No.5, pp. 1829-1834, September 1985.

5.

S. J. Salon and J. M. Scbneider: "A Hybrid Finite Element Boundary Integral Fonnulation of the Eddy
Current Problem",lEEE 'Iiansactions on Magnetics, Vol. Mag-18, No.2, pp. 461-466, March 1982.

6.

Y. Kagawa, T. Murai and S. Kitagami: "On the Compatibility of Finite Element-Boundary Element
Coupling in Field Problems", COMPEL, VoL I, No.4, pp. 197-217, 1982.

7.

Y. Kagawa. T. Yamabuchi and S. Kitagami: "Infinite Boundary Element and its Application to a Combined
Finite-Boundary Element Technique for Unbounded Field Problems", Boundary Elements VII, ed. C. A.
Brebbia, Springer-Verlag, New YotX, 1986.

8.

I. Kaljevic' ,S. Saigal and A. Ali: "An Infinite Boundary Element Formulation for Three-Dimensional
Potential Problems", to appear in International Journal for Numerical Methods in Engineering.

9.

P. P. Silvester, D. A. Lowther, C. J. Carpenter and B. A. Wyatt: "Exterior Fmite Elements for 2-Dimensional
Field Problems with Open Boundaries", Proc. IEE, 24, December 11]77.

10.

R. L. Ungless: "An Infinite Finite Element", MA Sc. 1besis, University of British Columbia, 1973.

11.

P. Bettess: "Infinite Elements",International Journal for Numerical Methods in Engineering, Vol. II, pp.
53-64,1977.

12.

G. 1. DeSalvo and R. W. Gorman: "ANSYS User's Manual Rev. 5.0", 1991.

13.

M. S. Chapman: "Calculation of the Maguetic Field of Two Intersecting Cylinders Using the ANSYS Finite
Element Program", Report No. MDTA-142. Lawrence Livennore Labs, August 1989.

14.

W. B. Boast: "Principles ofEleclric and Maguetic Fields", Harper and Brothers Publishers, New Yolk, 1948.

15.

J. M. Schneider: '"The Fmite Element-Boundary Integral Hybrid Method and Its Application to
Two-Dimensional Electromagnetic Field Problems", Doctoral Thesis, Rensselaer Polytechnic Institute; 1982.

Optimum Location of Electrode in Cathodic Protection System

K. Amaya*, S. Aoki* and K. Kishimoto**

*Department of Mechanical Engineering Science


Tokyo Institute of Technology,
O-okayama, Meguro-ku, Tokyo, Japan
**Department of Mechanical Engineering for Production
Tokyo Institute of Technology,
O-okayama, Meguro-ku, Tokyo, Japan

ABSTRACf

This paper presents a boundary element application to determine the optimum


impressed current densities and optimum location of electrode in a cathodic
protection system. The potential within the electrolyte is described by Poisson's
equation with nonlinear boundary conditions which are enforced based on
experimentally determined electrochemical polarization curves. The optimal
impressed current densities are determined in order to minimize the power supply
for protection under the protecting conditions that the electric potential of
every part of the structure to be protected should be less than some critical
value. The solution is obtained by using the conjugate gradient method in which
the protecting conditions are taken into account by the penalty function method.
The boundary element method is employed to discretize the governing equations.
Several numerical examples are presented to demonstrate the practical
applicability of the proposed method.
Key Words: Boundary Element Method, Optimization, Penalty Function Method,
Cathodic Protection, Electrochemical Polarization Curve,
Conjugate Gradient Method
INTRODUCfION

The economic loss due to galvanic corrosion in structures, such as underground


pipelines, off-shore structures and chemical plants, is estimated to be
tremendously great. In order to reduce the loss, it would be necessary to design
such structures and protection systems based on a precise analysis of the
corrosion behaviors. This paper deals with the application of the boundary
element method to the cathodic protection systems for corrosion prevention.

22

The electric current versus electrical potential difference between a metal (M)
and a reference electrode (e.g., SCE=Saturated Calomel Electrode) curve is called
a polarization curve of the metal, which is schematically shown in Fig.l. The
relationships between the current density q and the potential E for anodic and
cathodic reactions (solid curves) are not obtained individually but nominal
relationship E =f (q) for the two reactions (dashed curve) is measured. In the
natural state, the reactions become in equilibrium at point C, and the current
densities QC shown in the figure flow from the anode to cathode. Corrosion rate
is proportional to this current density.
There are two ways to cathodically protect a structure: (1) by a sacrificial
anode or (2) by the an extermal power supply (see Fontana, 1986). In the first
method, the metal to be protected is short-circuited to a less nobel metal to be
corroded at the expense of protecting the more valuable one, while in the second
method currents are impressed by an external power supply to reduce the potential
of the metal to the critical value Ep(see Fig 2). The cathodic protection
systems are usually designed on the basis of past experience together with a
limited number of empirical formulas applicable for a narrow range of conditions.
In this paper we will focus on the extermal power supply system and develop a
method to determine the optimum impressed current densities and optimum location
of electrodes for preventing the corrosion.

o
Fig.1 Polarization curve

current

tank
ele trode

Fig.2 Cathodic protection system

GOVERNING E'JJATIONS

Let us assume that the surface of the electrolyte domain Q is surrounded by


r (= r d+ r n+ r m) as shown in Fig. 3 where the potential values and current

23

densities are prescribed on r dand r n' respectively while r m is the metal


surface . In the dOOlain Q there are m number of electrodes.

Fig.3 Boundary conditions for cathodic protection


If we assume that the electrolyte is homogeneous and that there is no
accumulation or loss of ions except the points of electrodes, the
electropotential within the electrolyte, 1> (::- E), obeys the Poisson's equation:
(1)

where x denotes the conductivity of the electrolyte , (} ( . )' the delta function,
Q e the impressed current densitys and xe the location of electrodes. The
boundary conditions associated with (1) are described as:

1> = o

q[=xn] qo
-1>=f(q)

on rd

(2)

on rn

(3)

on rm

(4)

where q denotes the density of current across the boundary and a/ a n the outward
normal derivative. It is noted that the electrical potential is defined with
referring to the metal and has the inverse sign of that employed in the corrosion
science in which the potential is defined to a reference electrode such as a
saturated calomel electrode (SeE). o and qo are the prescribed values of the
potential and the current density, respectively. The function f(q) is the
experimentally determined polarization curve and, in general, is a nonlinear
function of q.
The condition of cathodic protection in which we wish to lower the potential
values of metal below a critical value E p is given by:

24

-;,

~E

(5)

In the present paper, the optimal impressed current densities are to be


determined in such a way that the necessary power to achieve the complete
protection becomes minimum, i.e, we employ the following objective function:

00

P=J{;'+fJq)}qdr
re

where r e is the circle whose center is ~e and radius is E (E is very small


compared with the size of 0) and f e ( q ) is the polarization curve of
electrode. Therefore, the problem considered here is reduced to finding the
couple of Qe and ~e that minimize (6) subjected to the conditions (1)-(5).
NUMERICAL METIIOD

Since the knowledge of the quantities, such as potentials and current densities
on the surface of materials is important for the corrosion problems, the boundary
element method is employed to discretize the governing equation. Also this method
is useful for changing the locations of electrodes in the optimization procedure.
By Green's theorem we obtain

"s O(;'*~;'-;'~;'*) dO
=T (;, *q - ; , q*) dr

(7)

where r = r n + r d + r m' ;, * is the funamental solution to Laplace's equation,


and q *= " 8 ;, */8 n. Substitution of Eqn. (3) into Eqn(7) yields

(8)

where c denotes a known constant resulting from the cauchy principal value of
the surface integral. The boundary r is discretized into boundary elements.
" [H]

t-

[G]

~= Qe~*(~)

(9)

where , the i-th component of the vector it* is given by

"

gi(~)=";'ie = ~ln

O/r ie )

(lO)

and r ie is the distance between the points i and e. Substituting (2) through
( 4) to (9), we obtain the following simul taneous equations
" [H] {

.~:
-f(qm)

} - [G]

I;:}
qm

~j[*(~)

(11)

25

where the quantities with suffix d, nand m represent those related to the
boundary r d' r nand r m' respectively.
Considering that (5) is the constraint condition for minimizing (6), we define
the following objective function:
P *=P (Qe' ~) +k 1/ Q (Qe' )Se) II 2

(12)

where k is the penal ty coefficient, and II 1/ represents the norm of the


vector. The k-th component of the vector II is given by
(13)
where u( . ) denotes the unit step function. The function II is 0 when (5)is
satisfied ,and if (5) is not satisfied, II has some value. Qe and ~e minimizing
p* for k-400, satisfy (5), and also make the necessary power Pminimal. In the
computation we will find Qe' ~e which minimize the objective function p* for a
sufficient large value of k. We can obtain the optimal solution by following
procedure.
1. Assume Qe and ~e'
2. Solve Eqn.(1~4) by B.E.H.
3. Calculate the value of P by (6).
4. Calculate the value of II by (13).
5. Obtain the value of P *.
6. Change Qe and ~ethen go back to 1
NUMERICAL EXAMPLES

In order to examine the availability of the method, example problems have been
solved.
(I) The first example is shown in Fig.4 in which the rectangular storage tank of
the size a x O. 5a is protected by one electrode. 30 Constant elements(see Fig.4)
were used for boundary element discretization. The polarization curve of the
metal were assumed as shown in Fig.5 with the critical value Ep= -1. To minimize
the objective function p*, the conjugate gradient method (see Beveridge and
Schechter, 1970) were employed.
Fig.6 shows the location of the electrode. It is noted that at 15 iterations the
optimum was achieved for the penalty factor k=10 5 . The electrode is finally
located at the center of the tank. Fig.7 shows the results of initial and optimal
potential distributions along the surface of the tank. It is seen that the
potentials at C and D are higher than the critical value -1 in the initial

26

distribution. After optimization, the whole surface of the metal is cathodically


protected . Fig.8 shows the current density distribution. The necessary total
power P/(

El>

is calculated as 19.18.

ELECTRODE

Fig.4 Model of cathodic protection case 1

E IEp

Fig.5 Polarization curve


0.50r-----r-----Tr---~----~

UJ

<: coo 0.25


t--

en

=O~----~I------L-I----~----~

0.0

0.5

DISTANCE x/a
Fig.6 Initial and optimum locations of electrode

t.o

27

~t

w
'-

-0 . 5

-=
.....

0-

- - I NIT I Al
-OPT IMAL

-1. 0

, - - .... ,
'eo \
....

".

:=
0-

e>

0...

-1.5

0 . 5 1.0

1.5

2.0

2.5

DISTANCE s/a

3.0

Fig.7 Potential distribution along surface


O,,~,,-.-.-.-.-.-r-r-r,

-1

=0--

-2
\

-4.

\
\

\
\

,\
,,

'

I)

-Ii

0.5

- -INITIAL
-OPTIMAL

I \

~ co -5
cc
::::::. c-

Co.:>

,\

\
\

\I

1.0

1.5

DI STANCE

2.0

SI

2.5

3. 0

Fig.8 Current density distribution


(2)The example is shown in Fig.9 in which the asymmetric storage tank of the size
a x o. 5a is protected cathodically by two electrodes . 60 constant elements were
used for boundary element discretization. other conditions are the same as
example(!) . Fig.!O shows the location of electrodes, and the solid circles in the
figure indicate the optimized locations. It is noted that at 45 iterations the
optimum was achieved. Fig.!! shows the initial and optimized potential
distributions. It is seen that the potentials at A,B and F are higher than the
critical value -1 in the initial distribution. After optimization, the whole
surface of the metal is cathodically protected. If the impressed current is
smaller than this solution, the condition of cathodic protection is not satisfied
at A,B,D,E and F. Therefore this solution is considered to be optimal. Fig.12
shows the current density distributions. The necessary total power
P/( K Ep2) is obtained as 6.88.

28

y 11\

. -.

60 Im4111 , llIIlls

..

tw

ELECTRODES

ELECTROLYTE

"

METAL
_.

It!

..

.....

-L

IIiSULATOR

Fig.9 Model of cathodic protection case 2

O. 50 1-----r----.1r---

-...1 --...

(J

::z
<as

1--0.25f-

en >-

OL..--- - ' - --

0. 0

--'-_ _ _'--_ _-L.....

0. 5

1. 0

DISTANCE x/a

Fig.l0 Initial and optimum locations of electrodes

- -IN I TI At
-OPTIIIAL

w
....... -0.5

-1.5

~~~~~~~~-L~~~~

0.5

1.0

1.5

2.0

2.5

DIS TANCE s / a

Fig. 11 Potential distribution along surface

3.0

29

00-

en

-=
......
<::>

=-=
.....
00-

cc:
ex:

==~

- - I NI T 1AL

:>-

.......,

-1

:I

-3

0.

=-

-2

-4-

-s

O.S

1. 0

I.S

2. 0

2.S

DIS TANCE s / a

3. 0

Fig.12 Current density distribution


CONCLUSlOOS
This paper presents a boundary element application to determine the optimum
impressed current densities and optimum location of electrode in a cathodic
protection system. The electrodes are asumed as sources of currents, and the
electropotential within the electrolyte is modeled by the Poisson's equation. In
this optimun problem location of source point and power of source are designed.
The optimum solution is obtained by using the conjugate gradient method by taking
into account of the protection condition with the penalty function method has
been proposed . The boundary element method has been employed to discretize the
governing equations. The numerical examples demonstrate that the present method
can effectively be used to design the cathodic protection systems of actual
structures.

REFERENCES

Aoki, S. and K.Kishimoto {1989}. Prediction of galvanic corrosion rate by the


boundary element method. ComPut. Math. Appl.
Beveridge, G.S.G. and R.S.Schechter (1970). qptimization: Theor,y and Practice,
McGraw-Hill, New York.
Brebbia , C.A. (1978). The Boundary Element Method for Engineers. Pentech Press,
London.
Fontana, M.G. (1986). Corrosion Engineering. 3rd Edition. McGraw-Hill, New York.

Vibrations of Reissner Plates by BEM and D/BEM:


A Comparison
H. Antes and W. Cheng
Institut fur Angewandte Mechanik, Technische Universitat Carolo Wilhelmina,
D-3300 Braunschweig, Germany
Summary
During the last decade, the Boundary Element Method (BEM) has emerged as an
accurate and efficient numerical method for plate dynamic analysis (for a
survey, see [1]). But, although it is known that in certain cases, e.. for
thick plates, the effects of shear deformation must be included, dynanuc BE
formulations for the respective theory, i.e. for Reissner I Mindlin plates
are rare.
As already successfully used in the case of thin Kirchhoff plates [2]
and plates with variable thicknesses [3], the static fundamental solution
can be employed also for dynamic analyses. These formulations require a
discretization of both the perimeter and the interior surface plate, but the
simpler form of the static fundamental solution reduces the computational
difficulties, and, the domain discretization is used only to evaluate
integrals and not to discretize the structure. Consequently, the method
retains most of the advantages of a BEM solution over a pure domain
discretization method.
In the present paper, based on a direct BE formulation for Reissner
plates [4], this approach, but, since the dynamic fundamental solution has
been derived recently [5,6], also the really dynamic BEM is used for the
dynamic analysis in the frequency domain. A number of numerical examples is
presented in order to illustrate the correctness of the new fundamental
solution, the applicability of the different formulations, and, moreover,
the higher order theory effects.
Reissner's Equations of Motion
Taking Reissner's assumptions concerning the stresses (f ap' (1a3' and (133'
i.e. his definitions of the corresponding internal stress-resultants, the
bending stres-couples MaP and the shear-stress resultants Qa [4], as "displacements" the rotations q,a and the deflection w, and considering loadings
q and body forces Fi varying sinusoidally with time, the threedimensional
theory equations for the dynamic equilibrium yield straight forward [6]

v
= - Fa - r-v

, -2

= - F3 - q - ph

h3

q,a - P IT OJ ~ a

(1)

c/ w ,

(2)

where p is the density of the plate material, h is the plate thickness, AZ =


101hZ is a characteristic quantity of Reissner's theory, and CI) is the cir-

31

cular frequency. If one substitutes via the constitutive equations and the
strain-displacement relations [6] the rotations tP a and the deflection w in
(112), the equations of motion will be obtained as

*D u.(x) = -p.(x) ,
.I..
1J
J
1

(3)

where the total load is defmed as

+ r-v..t
+q.

Pa = Fa

P3

There,

F3

-2

q,a

(4)

(5)

L1:Y denotes the components of the dynamic, frequency domain operator:


*D

.laP = N

*D

L1a3

1- v
-z-

{[.I2
1 + v 0aP }
-..t + c(w) ] 0aP + r:v

*D
I-v 2
= - L13a
= - N -z-..t

L1*D
33 -

N -ZI-v ..t2 {.I

+ "56

whith the bending stiffnes

N,

(6)

0a

(7)

c(w) }

(8)

v, and the frequency

the Poisson's ratio

dependent factor c(w) : = 2(1 + v) p w 2/E.


The Domain/Boundary Element Method
The static-like form of these equations (1/2) as well as of (6/8) for c=O
suggests using the integral identities for Reissner'S plate statics [4] and
the static fundamental solution [7] in order to achieve an integral representation of the problem. Then, one starts from Betti's reciprocal theorem
for two Reissner states: One state consists of the known and unknown state
variables of the given problem, and the second state represents the
fundamental solution tP (k)(x,~, w(k)(x,~ due to unit static loadings at the
a
point ~ of an inftnite plate. This leads to (k = 1,2,3):

~aW tPa(~ + ~WwW

f{lPhw2[ Z tPix)l/>~k)(x,~ + w(x)w(k)(x,~ ]:


-

-2- -

--.

D~-----------------+ q(x)[ w(k)(x,~ - 1~ v ~ tP~~1(X,~ ] }lDx

f {Mnn(X)tP~k)(X,~ + Mnt(X)tP~k)(X,~ + Qn(X)w(k)(X'~} dl'x

.Ii {
' \ k )(x,~
-)'
tPn(x)Mnn

1\ (k)
tPt(x)M
nt (x,~

t"T'

w(x)Qn(k) (x,~.r x'

(9)

32

O.5"f

"f

where <id({) is
and
for points c! on smooth boundary curves and in
the interior, respectively. By using these equations (9), k=1,2,3, the rotations 4>olf;) and the deflection w(f;) can be determined at any arbitrary point
~ in the interior of the plate a, when all unknown boundary reactions have
been previously determined. The relevant static fundamental solutions are
[6,7]:

4>~P)

rlNn l~vB(Z) - In(i)+ 1 ] "aP- [ l~vA(Z) +2 ] r,ar,p}

w(fJ) = - 4>~3)
P

w(3)

where
A(z)

=~
01tl'!

= ~ j.z { [i
= Ko(z) + ~Z

r r
-

,a

(In(zz) - 1)

l~v]

!]
Z

(11)

1n(C) - 2i }

B(z) = Ko(z) + !Z [K 1(z)


contain the modified Bessel functions Ko(z) and KI(z) with the
z=lr. A(z) is continuous while B(z) has the singularity In(z) [7].
gular couples ~(k) ~(k) and shear forces Q(k) can be obtained by
nn' nt
n
basic displacement - stress relations [6,7].
[K 1(z) -

(10)

and

(12)

!]
Z

argument
The sinusing the

For the numerical solution of these integral equations (9), the boundary r
and the interior domain a are both discretized into boundary and interior
elementl", respectively. Then, via the usual approximation procedures
introduction of appropriate shape functions, point collocation (here, at
both boundary and interior nodes) and numerical integration - the discretized versions are obtained:

(13)

where u tb and tlb are the vectors of the unknown geometrical and statical
nodal boundary variables, respectively, while the vectors u lb and t lb contain the known, i.e. prescribed, nodal boundary values. Finally, u.1 is the
vector of the unknown interior nodal values. The index at the vectors q, and
the first of the lower two indices at the boundary element integral matrices
U and T, as well as at the interior element inertia integral matrix M,
indicate the location of the collocation point ~.

33

For the special case in which around the whole boundary r all geometric
boundary values are prescribed to be zero, i.e. for a clamped plate, the
above system can be simplified [8]. As a first step, the unknown boundary
reactions tb can be expressed by
tb = - [ Ubb

rl { ol [ Mbi ]. u

qb}

(14)

Then, by substituting t b , the remaining equations are reduced to


{ I -

ol[ Mii - Uil> [ Ubb r l Mbl ]} u; = ~ - Uib [Ubbrl q"

Moreover, for the free vibration problem, i.e.


reduces to the eigenvalue problem

q"

(15)

0, equation (15)
(16)

The matrix H is real and, in general, non-sparse, non-symmetric, and


positive definite. Solutions of (16) may be found iteratively using
algorithm developed by Smith et al. [8]. For any harmonic excitation
circular frequency OJ, the responses, i.e. the boundary reactions and
interior rotations and deflection, are obtained directly from Equ. (13).

nonthe
with
the

The Dynamic Boundary Element Method


Obviously, in the above DIBEM formulation of the dynamic problem the inertia
term domain integrals affect the basic boundary character (similar to any
problem where body forces are taken into account) . The only way to avoid
these inertia term domain integrals, is to employ the fundamental solutions
of the complete system of dynamic equations (3), i.e. the solution for a
dynamic unit point load at an arbitrary point {:

*D
D(k)
_
k
LI .. (==) u (x,{) - - J(x-{) J.
1 J uX
J
1

(17)

34

B~3)

z-k a2~a3 J.-2{ I:V A2 - alAO }

with

(20)

Ao = Ko<-lil) +
Yo(V"fii;lr), Aa = aa Ko(ra:;r) -a3 Yo<"v'f3;Tr)
= ra:; KI<ra:;r) + VTi;T Y1(VTi;Tr) ,
C 1 = a 1 Ko(~r) - a2 Ko(v'i\r) , C2 = ~ Kl(~r) - v'i\ K1(v'i\r).
where the frequency dependent factors ai' a2 and a3 are given by

Ba

a1

a2,3

= J.2 - c(.r)
= i [- c (liV

~) ,fc2(liV - ;)2+ ~(1-V)CJ.21

(21)

(22)

and Ko and Yo is the modified Bessel function of second kind and the Neumann
function, respectively. The corresponding singular "stresses", i.e. the
singular stress couples ~~~) and stress resultants 6.~k) are even more
lengthy; their explicit expressions may be found in [5].
Now, by employing this dynamic fundamental solution, the inertia domain integrals, i.e. the marked terms in (9) are avoided. Hence, the approximation procedure for gaining the discretized version of this system needs
point collocations only at the boundary nodes of the shape functions. Thus,
in the notation as used before in Equ. (13), the final system of algebraic
equations is given as

In this system, only boundary reactions. i.e.

u1l>

and t lb , are unknown.

Some Numerical Checks for D/BEM and Dynamic BEM


First, free vibrations of a thin (h=0.05a), simply supported square plate
with side length a=0.2 m (v= 0.34; E = 1.02-1011 N/m2 ; p = 8350 kg/m) are
determined by DIBEM. In the numerical treatment of (9), a discretization is
employed with 4 as well as with 8 elements of equal length LI a for each of
the plate edges, i.e. for the boundary integrals, and, in addition, 16 and
64, respectively, quadrilateral interior elements (of size Lla-Lla) for the
domain integrals. Quadratic and biquadratic shape functions are employed for
approximating all the states along the boundary and the "displacements" in
the interior, respectively. Thus, point collocation creates the system (13)
with 195 and 675 equations, respectively.

35

Table: Eigenfrequencies of a thin simply supported square plate (h=0.05a)


Method
C0 3
C0 6
CO 2
C0 5
COl
CO"
Leissa
BEMS'!'"
BEMST"
BEMDY
Bezine
SAPIV
D/BEM 4-4
D/BEM 8-8
E-theory
Mindlin
BPC [14]

19.739
19.756
19.866
19.803
19.866
19.764
19.593
19.557
19.440
19.640
19.560

49.348
49.445
49.671
49.420
50.145
51.801
48.622
48.248

98.696
99.058
100.987
98.604
101.864
97.987
96.602
94.528

78.957
79.496
81.051
79.210
80.971
84.780
78.012
76.271

128.31
129.91
133.88
128.60
133.68
133.88
127.70
121.65

167.78
168.94
169.14
167.78
173.71
157.70
156.56

Deflecllon(dlmenslonless)

.
...
.
..
:1 };

~r---~--~~---~----~----.~.----~----~----~--~----,

30

JLV:

20

10

plate thickness

" I
I.... ~'
_ _ _ _ _-<!l-_ _......a::.::--"I
_0. _ ..- ;
-

"

" \\

,,~ :0" -.-:.-:.:::::;.,/1'~o.;:-L-'-~-'""-""-~~4>::--=l~r---=--=--=-9-

I#':

J:
I,

-20

-10

h. 0.0! ..
h. 0.02 ..
h. O.O~ ..

:: f

It'

I ,

" "

.,

If"
I,

-30

"
"
"

-40

_soL-__

__

~~

10

__

____

15

20

____

~~

25

__

30

__

____

35

40

____

45

__

50

Frequency(dlmenslonless)
Fig.l. Central deflection of a square plate with 2 opposite simply supported
and two fixed edges under harmonic uniform load versus frequency:
Influence of the plate thickness (DIBEM solution)
In the table, DIBEM results are compared to those found by other authors,
other methods, and different theories. In Kirchhoff's theory, Leissa [11]
gives a series expansion solution, Beskos and Bezine use BE formulations
with the static fundamental solution and quadratic (BEMS'!"') and constant
r(BEMST", Bezine) shape functions [1,2] as well as with the dynamic

36

fundamental solution (BEMDY, [2]), while SAPIV is a four-node quadrilateral


FE approximation [2]. Moreover, the first eigenfrequency is also known from
an 'exact' elasticity solution [12] and from two other papers using
Mindlin's theory [13,14). The DIBEM delivers values which are, on the whole,
lower than those from Kirchhoff's theory, but the agreement of the first
eigenfrequency with the 'exact' one is excellent. Thus, these lower values
should be an effect of the higher order theory.

A second example with other boundary conditions, a square plate with two opposite simply supported and two fixed edges, confU'DlS this higher order
theory effect (see Fig. 1); the thicker the plate, the lower the eigenfrequency.
Next, the correctness of the dynamic BEM employing the new dynamic fundamental solution (D-F) is demonstrated. Figures 2 and 3 depict in dimension= w phol/qo versus the excitation
less values the central delection
frequency W = (J) al,[jifiTN in the case of a rather thin plate (h/a = 0.05) and
a moderately thick plate (hla = 0.1).

Detlectlon(dlmenslonless)

150

h/a-O.05

100

-S-F

-D-F

50

o
-50

I(

'\

l-

-100

-150

20

40

60
80
100
120
140
Frequency(dimenslonless)

160

Fig.2. Central deflection of a simply supported square plate under


harmonic uniform load versus frequency - h = 0.05 a:
Comparison of dynamic BEM (D-F) and D/BEM (S-F) solution

180

200

37

Obviously, although using a rather coarse grid with only 44 boundary


elements and for the D/BEM additionally 16 interior cells, in the whole, the
agreement is excellent. Differences are found only for frequencies 0> higher
than 120 with a certain increase in the case of the thicker plate. But,
further numerical studies are necessary to decide whether it is caused by
the rather coarse grid or it is really due to the use of the different
fundamental solutions.
Deflectlon(dlmensionless)

200

hfaO.l

150

-S-F

-D-F

100
50

'\

\..

If

-50

-100

-ISO
-200

20

40

60
80
100
120
140
Frequency(dlmenslonless)

160

180

200

Fig.3. Central deflection of a simply supported square plate under


harmonic uniform load versus frequency - h = 0.1 a:
Comparison of dynamic BEM (D-F) and D/BEM (S-F) solution
A second study analyses two problems: the importance of using the sixth
order Reissner' theory and the effect of the plate thickness. In Figure 4,
the central deflection versus frequency is shown for Kirchhoff's theory (h =
0.05 a) and for Reissner's theory (using the dynamic BEM) in the case of two
different thicknesses, h = 0.05 a and h = 0.10 a. It becomes apparent that
the behavior in the range of low frequencies (0) < SO) is almost identical,
but especially the higher natural frequencies are shifted, the more, the
thicker the plate and the higher the frequency.

38

Detlectlon(dlmenslonless)

200r-----------------------------------------------~

150
100

Kirchhoff

--- h/a-O.05

...... h/a-OJ

!
'.

50

::

it,~

.~,

..

...+.:~~--~--~.'b_~.~.~j~
...~/!~
O ~~Jt::===_~==~;l\.~
; -_ ...
(

. U~
~"'\:

-50
-100

~\....

: ,: .,"

I( il

__~
180
200

-150~--~----L---~----L----L--~L---~--~L----L

20

40

60
80
100
120
140
Frequency(dlmenslonless)

I
U

1::,1

160

Fig.4. Central deflection of a simply supported square plate under


harmonic uniform load versus frequency - h = 0.05 a:
Comparison of dynamic Kirchhoff' BEM and Reissner' BEM solution
Conclusions
On the basis of the preceeding studies the following conclusions can be
drawn:
1. Two different boundary element procedures have been presented for the
dynamic analysis of plates following Reissner's theory. The first works
with the static fundamental solution, while the second uses the dynamic
fundamental solution.
2. A comparison of the two methods revealed that both can efficiently
provide results of high accuracy. Depending on the particular problem to
be solved, one method can be more efficient than the other.
3. Both methods have been successfully used for the study of the thickness
effect of plates, especially On the eigenfrequencies.
4. More extensive parametric studies are required for all possible boundary
conditions and Ioadmgs.
Acknowledgement:
The author likes to acknowledge the financial support of the Deutsche
Akademische Austauschdienst grant DAAD 324/306/00519 which allows W. Cheng
to work on Reissner's plate dynamics at the Institut fur Angewandte
Mechanik, Technische Universitat Braunschweig.

39

References
1.

2.
3.
4.
5.
6.
7.
8.
9.

to.
11.
12.
13.
14.

Providakis, C.P. and Beskos, D.E., Dynamic analysis of elastic plates,


BEM X, Vol. 4, Heidelberg-New York: Springer 1988.
Providakis, C.P. and Beskos, D.E., Free and forced vibrations of plates
by boundary and interior elements, Int. J. for Numerical Methods in
Engineering 28 (1989) 1977-1994.
Katsikadelis, J.T. and Sapountza]ds, E.l., A BEM solution to dynamic
analysis of plates with variable thickness, 2nd Nat. Congo on Mechanics,
Athens 1989.
Antes, H., On a regular boundary integral equation and a modified
Trefftz method in Reissner's plate theory, Engineering Analysis 1 (1984)
149-153.
Cheng, W.; Antes, H.: The dynamic fundamental solution of the
Reissner-Mindlin plate theory. (in preparation).
Antes, H., Static and dynamic analysis of Reissner-Mindlin plates, chap.
8 in: Boundary element analysis of plates and shells (Ed.: D.E.,
Beskos), Heidelberg-New York: Springer 1991.
Vander Weeen, F., Application of the boundary integral equation method
to Reissner's plate model, Int. J. for Numer. Methods in Engng. 18
(1982) 1-10.
Beskos, D.E., Dynamic analysis of plates and shallow shells by the
D/BEM, in: Advances in the Theory of Plates and Shells (Eds.: Voyiadjis,
O.Z. and Karamanlidis, D.) Amsterdam: Elsevier 1989.
Smith, B.T., Boyle, I.M., Dongarra, 1.1., Garbow, B.S., Ibeke, Y.,
Klema, V.C. and Moler, C.B., Matrix eigen-system routines - EISPACK
guide, Berlin: Springer 1976.
Hdrmander, H.: Linear partial differential operators. Springer Verlag,
Berlin 1963.
Leissa, A.W., Vibration of plates. NASA SP-160, Washington D.C.: NASA
1969.
Levinson, M., Free vibration of a simply supported rectangular plate:
Exact elasticity solution. 1. Sound Vibration 98 (1985) 289-298.
Mindlin, R.D. and Deresiewicz, H., Flexural vibrations of rectangular
plates. 1. Applied Mechanics 23 (1956) 430-436.
Akkarl, M.M., Hutchinson, J.R., Boundary point collocation method for
the vibration of a Mindlin thick plate, pp. 465-476 in: Boundary
Elements X, Vol. 3, Stress Analysis (Ed.: Brebbia, C.A.), Berlin-New
York: Springer 1988.

Numerical Approximation of Boundary Integral Equations


in Three Dimensional Aerodynamics

P. Bassanini, M.R. Lancia, R. Piva


Universita. di Roma La Sapienza, Roma Italy
C.M. Casciola
I.N.S.E.A.N., Roma Italy

SUMMARY
A general integral formulation based on the Poincare representation for the velocity
field is investigated in the 3D case and a simple numerical application is presented. An
approximating procedure for the boundary integral equation in the unknown tangent
trace of the velocity field is presented with a particular emphasis on the approximation
of the tangent fields by means of finite elements.

INTRODUCTION

In the present paper we analyze a boundary integral formulation for 3D Aerodynamics


and the related problem of its numerical approximation.
From the physical point of view if we consider an attached incompressible flow about
a sharp edged body, the bulk of the fluid may be accurately modeled by the incompressible Euler equations. In order to retain the basic effects of a real flow, a Kutta type
condition is required to describe the vortex shedding. By an analogous approximation
the rotational wake may be represented by a sheet of concentrated vorticity, whose non
linear evolution may be described through a suitable form of the Helmholtz equations
to track the wake motion and to account for the vortex stretching. The concentrated
vorticity is given by n /\ [u], where the covariant components of the jump [u] of the
velocity across the wake are constant along the trajectories of wake points.
The complete problem has been analyzed in details under the hypotesys of 2D flow
in a previous paper [2], where several numerical results are also discussed.
Here, rather than discussing the aspects related to vorticity dynamics, we focus our
attention on the kinematics of the flow. Namely we study the solution of the boundary
value problem for the velocity field, once the vorticity distribution is known and the
boundary conditions on the normal velocity components at the body boundary are

41

provided. The general kinematical problem, both for the 2D and the 3D problem, has
been studied in [1], where a weak extension of the Poincare formula is developed. With
the same approach, here the velocity field is still represented by the Poincare formula,
which accounts for the vorticity field by a Biot-Savart like integral. The effect of the
solid boundary is described through two boundary integrals involving the normal and
the tangential traces (n . u , T u respectively).
Several differences appear when comparing the 3D problem with the 2D one. In
2D aerodynamics, it is well known that the circulation, directly related to the forces
exerted on the body, cannot be determined by purely kinematical analysis, because
a suitable irrotational flow circulating around the body may always be added to any
solution without affecting the vorticity distribution. This point is reflected into the
equations following from the 2D boundary integral formula following from the Poincare
formula.
In 3D we have to consider the vector circulation and the picture is completely different.
Actually a theorem [s] shows that the vector circulation is completely determined by
kinematical considerations (we deal here, for the sake of simplicity, with zero genus
boundaries, i.e. homeomorfic to a sphere). This shows that no circulating eigen solution
exist for the boundary value problem (that is purely circulating flows with zero vorticity
and zero normal velocity at the boundary). Consistently uniqueness theorems may be
proved for the relevant boundary integral equations.
Concerning the numerical approximation, we consider here only one of the two equations that may be obtained from the Poincare formula. Namely the one corresponding
to the projection of the identity on the tangent plane, which turns out to be a second
kind Fredholm equation.
Although the kernel of this equation is weakly singular the numerical integration is
rather cumbersome to perform. Moreover the right hand side of our equation requires
the evaluation of a Cauchy type integral.
Instead of a direct evaluation of these integrals we recasted our basic identity into a
new form by means of the generalized vector Stokes theorem. This allows to deal with
simpler kernels (also avoiding the presence of the Cauchy integral), but as a counterpart
a differential operator acting upon our unknown, the unknown tangent trace of the
velocity field, appears in the formulation. In .section 2, after a brief account of the
theory, we show the procedure to obtain this alternative form.
A crucial point in the approximation procedure, is related to the nature of the
unknown, which is a tangent vector field. We introduce a suitable discrete form for
this fields, obtained by means of finite elements. The details are given in section 3
while numerical results for the flow about simple shaped bodies are discussed in the last
section in order to demonstrate the feasibility of this approach in view of more general
applications to flows with vorticity.

THE POINCARE IDENTITY AND THE BOUNDARY INTEGRAL


EQUATIONS

Consider a bounded domain 0 C ]R3 with smooth COO connected boundary r, and a
3D vector field u(x) defined in the exterior domain 0' = ]R3 - 11. Then u(x) can be

42

represented by the weak form of the Poincare identity [1]:

= grad(x) + rotA(x)

u(x)

A(x) =

~(x)

L y)~(y)dy +
L
G(x,

,x E 0'

(1)

V(n /\ u)(x)

g(x,y)Q(y)dy - V(n u)(x)

where n is the inner normal to 0' ,~ = rotu is the vorticity, Q = divu is the expansion,
G(x, y) = 1/(41Tlx - yD, and V is the simple layer operator [7]

V(J)(x) =

1r G(x,y)f(y)ds

lI

EO'

or its weak extension when x E r [7], and /\ denotes the exterior (cross) product in R3.
This identity holds for u(x) square summable in 0', with Q and ~ square summable and
with compact support D in 0', so that u(x) is harmonic in R3 - D and
u(x)

= O(1/lxI2)

for x

(2)

-+ 00

Thus the normal and the tangential traces u . n, n /\ u belong to H- 1/ 2 (r), and n . ~
0- 1/ 2

belongs to H
(r). As proved in [8], the vector circulation Ir n/\u dS and the vorticity
n . ~ are connected on r by the relations:

(3)
Hence if the field vorticity is zero, the vector circulation is also zero. Then from (1) it
follows that the asymptotic expansion

holds for x -+ 00.


Thus the leading term depends on the flux across r, Ir n u ds and the average
expansion,ID Qdy, which can be assigned independently. IT they satisfy the relation

1r nu ds + Iv Qdy =

(5)

then u = O(1/lxI 3 ), and viceversa [1], [10]. Now if u(x) represents the flow field past
a body r, we must regard n . u, Q, ~ as given and n /\ u as unknown. In the solutions
we have in mind, Q = 0, n . u = n U oo (Uoo the uniform translational velocity of the
body) and ~ can be computed from the dynamical equation.
Then by taking the limit of n u and n /\ u from (1) as x approaches the boundary
we have found [1] the two simultaneous boundary integral equations
n(x)'rot"Vn/\u = Fdnu,Q,~l xEr

(6)

43

2n 1\ u(x) - n

1\

rot" V n

1\

= F [n . u, Q,~]

xEr

(7)

where

= n u + :n [Iv G(x,y) Qdy -

Fdn. u,Q,~]

F [n u,Q,~]

1\

rot

10 G(x,y) ~ dy + n

V(n u)] - n rot

1\

'V

10 G(x,y) ~dy

(10 G(x,y) Qdy - V(n .u)

(8)

(9)

Equations (6), (7) are equivalent and either one can be solved for nl\u. Equation (6) is of
the first kind and has been studied in [1]. Equation (7) is a Fredholm integral equation of
the second kind with weakly singular kernel and the existence and uniqueness of smooth
(Holder continuous) solutions was investigated by Werner [3], who proved uniqueness
for surfaces of genus zero (homeomorfic to a sphere). When r has genus p, there exist p
indipendent eigensolutions of class C""(r). Let us consider eq. (7), although the kernel
is weakly singular, its numerical integration presents some difficulties. The right hand
side (9) of the equation leads to the calculation of Cauchy type integral which requires
a particular care. In order to avoid such difficulties the equations are recasted in a more
convenient form. We take as unknowns the components of the tangential velocity, u,
which, as mentioned in the introduction are more convenient when dealing with vortex
layers. The key steps to derive the new formulation are reported here for completeness.
We compute the term rot V(n 1\ u) (x EO') from the Poincare formula (1) as follows:
ret V (n

1\

-Ir 'V vG(x, y)

u)

1\

(n(y)

1\

u(y)) dSv

-Ir [U(y) . 'V vG(x,y) n(y) - u(y)


- ( u(y)

lr

1\

(n(y)

1\

'VvG(x,y)) dSv

a~~: y)]
+

(u(y)

lr

(10)

a~(x,y)
nv

dSv

-Iru(y).n(y)'V"G(x,y)dSv
Thus from Poincare identity we obtain, for Q

~u(x) -

= 0 and x E r:

Ir u(y) a~~~;f) ds v + Ir u(y)

1\

(n(y)

1\

'Vl/G(x, v)) ds v

(11)

= rot 10 G(x,y)~dy
Projecting this equation on the covariant basis ga (Section 3) yields two scalar BIE's
with weakly singular kernels in the unknowns U a = U, . gao An alternative approach
which appears more convenient for numerical purposes may be based on the integration
by parts in the second integral on the left end side of eq. (11).

44

By applying the vector Stokes theorem we find:

-Ir

u(y}

1\

(n(y)

1\

V v G(x, y)) dS~

Ir G(x, y)

(n(y)

1\

V v)

1\

u(y)dS"

Hence
rot V (n

1\

u) = (u(y)

lr

a~X'Y)
ny

+ V(n 1\ V)

dS y

1\

u)

V V (n u)

(12)

u(y)dSII

(13)

Thus the Poincare identity can be written in the form


u

(u(y)

lr
V

so that as x E
1
-u
2

r :

fD

a~(x,y)
nil

G Q dy

+ rot

ru(y) a~x,y)
dS
nil

lr
+V

dSv

II

G Qdy + rot

(G(x,y) (n(y)

lr

fo G ~ dy
-

x E 11'

rG(x,y) (n(y)

Jr

fo G ~ dy

By decomposing u = n Un + u. on
equation with weakly singular kernel

1\ VII) 1\

II VII) II

u(y)dSII

(14)

xEr

we finally obtain the boundary integral

(15)

-Ir G(x,y)

(n(y)

II VII) II

u.(y)dSII

= f[u",~, QJ

where f is the sum of three terms. The first has the same form of the left hand side
with u"n replacing Un and the other two give the contribution of the vorticity and of
the expansion respectively (see eq. (14)).

Numerical approximation

In the previous section we derived a vector boundary integral equation in the unknown
tangential trace u.(x) (eq. 15). We present here a discretization of this equation in
terms of finite elements. We consider CO elements to discretize the geometry of the
surface, and we take the vertices as collocation points for the discrete version of eq.(15).
It is worth noticing that by this choice it is not possible to have a continuous tangent
plane across adjacent elements, so that some care is needed in the approximation of the
tangent vector fields which have to be continuous at the collocation points.
Let us assume that the surface is decomposed in several patches ( r = U~'::l r i ) such
that each patch r i , described in terms of curvilinear coordinates by

(16)

45

has in common with the others a portion of its boundary at most. At each point of r
a local basis is introduced through:

(17)

With the above assumptions the restriction ulri of the velocity field is expressed
(with the usual summation convention on repeated indexes) by

(18)
By substituting this expression into eq.( 15) and projecting on the local basis, we write
a system of two scalar equation in the unknown contravariant velocity components u"':

(1 - c)u"'(x) (g",(x) . gp(x)) - E, Jr, u"'(y) (g",(y) . gp(x))


+Ei

hi

gp(X) . [(n 1\ V)

-(1 - c)u" (n(x) . gp(x))


-E,

1\

aG
an dS(y)

u"'(y)g", (y)] G dS(y)

+ Ei Jrr ; un (n(y) . gp(x)) aG


an dS(y)

hi gp(X) .

[(n 1\ V)

1\

u"n(y)] G dS(y)

(19)

In order to construct the discrete version of this system, we need to introduce an


appro:dmation for the surface geometry as well as an approximation for the vector
field. We thus introduce a quadrangularization Qi = U qi on each domain D i . An
isoparametric finite element is then associated with each element qi E Q,. By this
procedure we construct an approximation for the i-th surface patch fhi' given by:

(20)
where Xk are the coordinates of the nodes (belonging to the exact surface f) related to
the i-th patch, Ngi is the number of nodes in the patch and N" are the basis functions
of the finite element space (shape functions).
In the actual calculation we use an element with eight nodes, four of which are vertices
while the others define the edge midpoints. As usual we require that the quadrangularizations of the different patches must give rise to a compatible quadrangularization for
the whole surface.
A similar type of procedure is provided to describe the vector fields by using four nodes
elements. In the following we name "V-node" a node used to discretize vectors and
"G-node" one from the quadrangularization of the geometry. The two set of nodes are
arranged such that any V-node coincides with a vertex of the geometric quadrangularization.

46

At each V-node we may thus associate the true basis vectors defined on the exact
surface (eq. (17)). By this procedure we discretize the tangent and the normal vector
fields in the form:

(21)
(22)
where gar,k, nk are basis vectors at the k-th node (belonging to the exact surface r)
related to the i-th patch, Nut is the corresponding number of nodes and MI< are the
shape functions.
Using this approach we construct a linear algebraic system for the unknown contravariant velocity components u~ (the normal field is given as a boundary condition).
Some additional care is required to treat those nodes which belong to two different
patches of the surface. In this case we actually have two different maps describing the
same region of the surface so that two different representation are induced for the tangent
vectors (that is we have two different bases at the same point). We also have different
contravariant components for the same vector depending on the base. This means that
at these special points we have four unknowns (two contravariant components for each
basis). We must thus supply two more algebraic equations which follow from the change
of the coordinate system.
The resulting system of equations, which is non singular, as expected from the theoretical analysis, may be solved by standard techniques. In the following section we present
briefly some numerical results for a very simple test case.

Numerical Results and Conclusions

As a preliminary validation of this model we selected a simple test case for which an
analytical solution is available, namely the irrotational flow about a sphere moving in
a uniform stream. Fig. (a) shows the computational grid. The discretization employed
a three patches description for the surface. The subsequent figure presents a plot of
the tangent velocity field at the surface. We may observe the typical flowfield given
by the potential theory, and in particular the stagnation point in the aft-end of the
body. This solution is compared with the analytical one. The relative error, defined
by IIUr - Vrll/IlVrll, where Vr is the exact solution, is everywhere less than .1 percent
for this case. The maximum error is attained at the four nodes dose to the stagnation
points, where the norm of velocity is very small, whereas in other nodes the relative error
is smaller by several order of magnitude. The above simple result shows the accuracy
of the proposed description for the tangent velocity field, which is a crucial point in the
extension of this approach to the general case of flows with vortex wakes.

Acknowledgement

This research was partially supported by C.N.R under contract n. 91.01319.ct01.

47

References
[IJ P. Ba.ssanini, C.M. Ca.sciola, M.R. Lancia, R. Piva: "A Boundary Integral Formulation for the Kinetic Field in Aerodynamics. I. Mathematical Analisys.", Eur. J.
Mech., B/Fluids, 10, n. 4, 1991.
[2J P. Ba.ssanini, C.M. Ca.sciola, M.R. Lancia, R. Piva: "A Boundary Integral Formulation for the Kinetic Field in Aerodynamics. II. Applications to Unsteady 2D
Flows.", Eur. J. Mech., B/Fluids, 1, n. , 1992.
[3] P. Werner: "On an Integral Equation in Electromagnetic Diffraction Theory", Duke
Math. J. 7,411-444, 1966
[4] Batchelor G.K., 1967, "An Introduction to Fluid-Dynamics", Cambridge, Cambridge University Press

[51 Ca.sciola C.M., Lancia M.R., Piva R., 1989, "A general approach to unsteady flows
in Aerodynamics: Cla.ssical results and perspectives". In: ISBEM 89 (Symp. of the
Intern. Assoc. for Boundary Element Methods), Ea.st Hartford, Conn. USA
[6] Costabel M., Wendland W.L., 1986, " Strong ellipticity of boundary integral operators", J.Reine Angew. Math. 372, 34-63
[71 Hsiao G.C., 1988, "On boundary integral equations of the first kind". In: ChinaUS Seminar on Boundary Integral Equations and Boundary Element Methods in
Physics and Engineering, Dec.27,1987- Jan. 1988, Xi'an Jiaotong Univ., Xi'an,
China (PRC)
[8J Milne-Thomson M.L., 1968, "Theoretical Hydrodynamics", MacMillan, London
[9J Nedelec J.C., 1977, "Approximation des equations integrales en mecanique et en
physique", Lecture Notes, Centre de Mathematiques Appliquees, Ecole Poly technique, Palaiseaux, France
[lOJ P. Ba.ssanini, C.M. Ca.sciola, M.R. Lancia, R. Piva: "A General Integral Formulation for rotational flows in Aerodynamics" Proc. IABEM90

48

a)

b)

lrrotational flow about a sphere in uniform translation . a) Computational grid, b) Tangent


velocity field

Shape Sensitivity Analysis of Uncertain Static and Vibrating


Systems Using Stochastic Boundary Elements
Tadeusz BURCZYNSKI
Department of En~eering Mechanics, The Silesian Technical University of
Gliwice, 44-100 Gliwice, Konarskiego 18A, Poland

Summary
The
application
of
stochastic
boundary
elements
to
examining
displacements, strains, stresses, natural frequencies and in the general
case - an arbitrary functional with respect to stochastic shape of the
boundary is presented. In order to solve this problem the idea of
stochastic shape sensitivity analysis is used. It is assumed that the
boundary undergoes small random variations which create the stochastic
tolerance range. Analytical boundary integral expressions for random
variations of displacements, stresses and natural frequencies are derived
and boundary elements are used in computational procedures.
Introduction
Models and solutions widely used in classical boundary value
problems are too idealized and can not reflect the complexity of real
physical phenomena. Therefore one can obtain a more adequate description
by replacing deterministic function, which describes boundary conditions,
material properties or shape of the boundary, by stochastic functions.
Such description makes better use pf knowledge about known physical
phenomena.
Many
different
interpretations
are
possible
for
terminology
Stochastic Boundory Element Methods (SBEM). This term can be used to
refer to boundary element methods which account for uncertainties in:
boundary conditions or material properties of a structure, as well as the
shape of a boundary. Such uncertainties are usually distributed on a
boundary (l) or within a domain (.0) of the structure and should be
modelled as spatial stochastic fields r(x,y), where x denotes spatial
position in IJ or r, y is an elementary event (YE W). Such stochastic
fields are defined on the probability space ('fJ,I,f1') where 'fJ is the space
of elementary events, I is a a-algebra of subsets of 'fJ and l1' is a
probability.

50

Applications of SBEM appear to have been initiated in the early


1980's. The earliest application used the boundary integral equation
method to solving stochastic boundary value problems of elastostatics
(Burczyliski [1]). Later SBEM was used to stochastic potential problems
(Burczyliski
[2]),
steady-state
heat
conduction
(Drewniak
[15]),
stochastic groundwater flow (Cheng and Lafe [14]) and dynamical problems
(Burczyliski [3,6] and Burczynski and John [11,12]). SBEM was also
extended to problems with random media (Burczynski [4] and Ettouney at
01.[16]) and to problems with uncertain boundaries (Nakagiri et 01.
[17,18] and Burczynski [5,7] and Burczynski and FedeliIiski [10]). A
description of applications of SBEM to stochastic boundary value
problems, wave propagation in stochastic media and eigenvalue problems of
elastic bodies with random boundaries is given by Burczyliski in [8]. The
overview of the basic concepts underlying the Stochastic Boundary Element
Methods and developments SBEM for solving variety static and dynamic
problems is presented by BurczyDski in [9].
In this paper SBEM is used to problems with stochastic shape of the
boundary.
Stochastic Shape of the Boundary
One knows that boundaries bounding real bodies are very complicated
as far as their geometrical shape is concerned. Usually they are uneven
and the irregularities do not easily lead to a unique deterministic
description. Therefore boundaries of such bodies can be defined
stochastically. SBEM is a very useful and natural technique for modelling
such problems. In order to' solve these problems it is possible to use the
perturbation technique (cf. [18,19]) or the method based on the idea of
stochastic shape sensitivity analysis (cf. [5,7,9,10]). The application
of the later approach to examining stresses, strains, displacements,
natural frequencies and in the general case an arbitrary functional with
respect to stochastic shape of the boundary is presented.

One assumes that stochastic shape of the boundary r may be defined


by prescribing a stochastic vector field 6g(x,y)=(6gk (x,y, so that:

x (y)

=x +

6g(x,y),

E6g=0,

(1)

where the deterministic variable l[ is related to the baseline of the


boundary r.
The
stochastic transformation field g(x,a(y modifies the external
boundary r, where a(y) =(8r(y, r=I,2, ..R, is a set of random shape

51

parameters, which specifies the actual stochastic shape of the structure.


The variation of the transformation field og is expressed as
(2)

where Vk=ogk/oa, can be considered as a velocity transformation field


which is associated with a shape design parameter a,(y). One assumes that
random shape parameters can be expressed as follows:
(3)
a(y) = a o + oa(y) , Eoa(y) =0,
where oa(y) represents fluctuation (variation) of random parameters and
ao=Ea(y) is the mean value of shape parameters a(y) which represent
deterministic base shape parameters.
If a(y) has the Gaussian distribution then it is completely described by
the mean value a o and the covariance matrix [K] is given as follows:
T

(4)

[K] = E[oaoa ].

Because
the
boundary
undergoes
small
random
variations
then
of
stresses
O'(x,y),
strains
e(x,y),
resulting
stochastic
fields
displacements u(x,y) and natural frequency w(y) can be expressed as
follows:
(5)
xe.o,
O'(x,y) = O'o(x) + oO'(x,y) , EoO'(x,y) =0,
e(x,y) = eo(x)
U(x,y) = uo(x)

+ OU(x,y),

w(y) = Wo

Eoe(x,y) =0,

oe(x,y) ,

Eou(x,y) =0,
ow(y) ,

xe.o,

xe.o or xer,

Eow(y) =0,

(6)

(7)
(8)

where qo=Eq(x,y), q=O',e,u and wo=Ew(y) are identified with the mean value
of state fields calculated for the untransformed boundary r with the
deterministic base shape parameters ao=(ao,)' r=1,2, .. R.
In the general case it is possible to consider an arbitrary functional
J= J'P(O',e,U)d.o

Jq>(U,p)dr,

(9)

.o(a)
r(a)
where 'P(O',e,u) is an arbitrary function of stresses 0', strains e and
.
*
displacements U within the domain .0 =.o(a), and q>(u,p) is an arbitrary
function of displacements U and tractions p on the boundary * =r(a).
The functional J can express global mechanical characteristics (e.g.
potential or complementary energy) as well as local stresses, strains or
displacements at point Xc.

52

Applying the Dirac


expressed as follows:

function

O'q.(x,,)

6(x-x,,),

the

6(x - x,,)6qIt6110'k1dtJ

stress

tJ
Comparing (9) with (10) one has
!P(O')

= 6(x

component

O'q.

is

!P(O')dtJ.

(10)

= o.

(11)

tJ

- x,,)6qIt6 110'k1 and ,(u,p)

In the similiar way the strain component tq. can be obtained:


tq.(x,,)

6(x - x,,)6qIt6I1tkldtJ

tJ

where

!P(8)dtJ,

(12)

tJ

In the case of the displacement component uq(x.,), xoetJ, one obtains

uq(xo)

6(x - x,,)6qkultdtJ

tJ

where

!P(u)dtJ,

(14)

tJ
(IS)

If one wishes to find the displacement component uq at a boundary point


x"er then one should introduce functions:
qJ(u)

= 6(x

- x,,)6qItUk and

!P(0',8,U)

= 0,

(16)

and the displacement component Uq can be expressed as follows:


uq(x,,)

,(u)dr

6(x - x,,)6qltultdr.

(17)

Due to stochastic shape variation the functional J can be expressed as


follows:
(18)
J(ao + 6a(y = J(aO> + 6J(y) ,
where the first variation of the functional 6J()I) can be expressed
analytically utilizing the adjoint approach (cf. [17]). Independently of
the primary system (PS), a concept of an adjoint system (AS) with the
adjoint solution u, t and 0'. is introduced.
The adjoint system is an elastic body with identical configuration and
physical properties as the primary system but with other boundary
conditions and body forces. On the boundary
of (AS) there are
prescribed boundary conditions in the form:

53
o
U __

ocb

Po = ocb

(u,p)

op

and within the domain D initial strain


.
body forces b are specIfied:

.."ai_- a'l'oa(a,e,u)

ai

(u,p)

(19)

OU

ai

e, and stress a

ai

fields and

= a'l'ae(O',e,u)

The constitutive law for the adjoint system has the form:

0'

= C(e

Ai

- e) -

ai

(21)

0' ,

where
Ce=c;jklekl' Cijkl =AOijOk I + II (OikOjl + UOjk), A and II are the
Lame constants.
In view of (11), (13), (15), (16) and (19) and (20) it is seen that in
the case of local characteristics one should introduce for adjoint
structure:
- a unit initial strain field
.i

ekl

0'1'
= -,,-UO'kl

(22)

= o(x - Xo)OqkOsl'

in the case when one seeks the stress component O'q. at the point xoED,
- a unit initial stress field
(23)

in the case when one seeks the strains component eq. at the point Xg E D,
- a unit body force
(24)

in the case when one seeks the


xoED,
- a unit traction
.i
a'I'
Pk = au;-

displ~cement

component uq at the point

(25)

o(x - Xg)Oqk'

in the case when one seeks uq at the boundary point Xg E

r.

The boundary element discretization


The boundary integral equations for primary (PS) and adjoint (AS) systems
have the form:
c(x)u w (x)

u *(x,y)pw(y) - p *(x,y)u',,(y)] dr(y)

Z ....(x),

(26)

w - (PS) ,(AS)

where Z W depends on body forces in the case the primary system (PS) and

54
ai

ai

on the initial strains t , stresses (1


and body forces b
of the adjoint system (AS).
The first variation of the functional J can be expressed as:
T
cSJ(y) = {S} {cSa(y)},
where

in the case

(27)
(28)

is the matrix of variation of random shape parameters, and

{S} =

[Sl,~""SW,SR]

(29)

is the sensitivity matrix whose elements Sr' r=1,2, ..R, are expressed by
total material derivative of the functional J with respect shape
parameters, Le.
_
Sr -

DJ
Dar'

(30)

The total material derivatives of J has the form (cf.[17]):


Sf =

J['P

(1

t&

+ b u + (q,

+ p u)'n-2(q,

J(~: -pa) (g:: -U:kV~)dl'l + J(~:

r1

J[q,

rz

+ P

+ u)

+ p

u\~nkV~dl'

(g( -P:kV~)dl'Z (31)

u1V~dL,

inte~and

where
[q,+p u1=(q,+p uJ.+- (q,+p urrepresents the discontinuity
of (q,+p u) along the curve L, which separates two parts of the boundary
r 1 and r z, n=(nk) is the unit normal vector, :f( is the mean curvature of
the boundary.
For the specified boundary conditions u(x)=u(x), xer1 and p(x)=p(x),
xerz, r1urZ=r, their total material derivatives DuIDa, on r 1 and DpIDa,
on r z are known and can be expressed in terms of the transformation
velocity field and gradients of UO and pO, respectively.
It is seen from (31) that sensitivities of J depend only on boundary
state variables of the primary system (PS) and the adjoint system (AS).
This fact
gives significant advantages in numerical calculations by
means of boundary element methods.
The discrete versions of the integral equations (26) are obtained
approximating the boundary r by series of boundary elements rO
(e=I,2, ..E). The Cartesian coordinates x are approximated by means of

55
en

shape functions L and the nodal coordinates x :


e
eo
X = Lx
Boundary displacements and tractions are approximated
e
boundary element r as follows:
o
eu

q = Nq
q=u,p,u ,p .

over

(32)
each
(33)

The domain a is divided into internal cells a (c=l,2, .. ,C) and body
forces and initial strains and stresses are approximated in terms of
"nodal values and shape functions N
r

"" cn

r=b,b

Nr ,

,8

ai

Ai

,a .

(34)

Finally, the discrete version of integral equations (26) has the form:
VI

HU = GP

VI

VI

Z,

(35)
w=(PS),(AS)
where U W and pVI are column matrices of nodal displacements and
tractions, respectively, H and G are square matrices which depend on
VI
boundary integrals of fundamental solutions and shape functions. Z is a
column matrix dependent on body forces b, in the c~se of the primary
system, or on body forces b, and initial strams 8 and stresses a,
in the case of adjoint system.
It is worth noting that H and G are the same for the primary system
(PS) as well as the adjoint system (AS) and therefore are calculated only
once. Taking into account boundary conditions equation (35) can be
reordered in such a way that all unknown variables are written in a
VI
VI
vector X and known variables in Y . The final form can be written:

AX'"

= BY +
VI

'"

Z .

(36)
w=(pS),(AS)

Solving equation (36) one obtains state fields of the primary and
adjoint
system,
which are needed to calculate
sensitivities
of
the functional J.
The variation of natural frequency is given by:

c5w(y) = {SW}T{c5a(y)},

(37)

where elements of sensitivity matrix {Sw} are evaluated by total material


derivative of natural frequencies with respect to shape parameters
(cf.[5,9,1O]):

sO; =

DC:

1= -2wo

[a(u) 8(U) - Wo P u u]nltvltdr.

where p is the mass density and u(x) is the displacement amplitude.

(38)

56

It is interesting to notice that ~co(y) is expressed by the boundary


integral and depends on the natural frequency COO' the mode u(x) and the
stochastic fluctuation of shape parameters ~a(y). THis is very important
in numerical calculations using boundary elements.
The mean value COo can be evaluated using the dual reciprocity
approach for the free vibration problem. After discretization of the
boundary r by means of boundary elements the following equation is
obtained (cf.[lO]):
:z
(39)
H U = COo M U,
where U is the column matrix of nodal values of displacement amplitude, M
is the mass matrix.
The covariance of stochastic stresses, strains and displacements for the
two fixed points Xl and X:z can be calculated from:
Kq(x1,x:z} = E{[q(X1,y) - qo(Xl)][q(X:z,y) - qo(x:z}]} =

= E[~q(xl,y)~q(X2'Y)] =

{S(x l)} [K] {S(x:z}{

(41)

The variance of the random natural frequency can be calculated using the
equation:
Var(co) = E[co(y) - col = E(~co(y)]:Z = {Sco} [K] {Scot
(42)
In order to calculate the sensitivity matrices {S} or {Sco} one should
r
determine the transformation velocity field vk which is associated with
the shape parameter a.(y). The selection of shape design parameters is
the key element in the shape sensitivity analysis and optimal design.
The most natural way for shape presentation is the use of boundary
co
nodes. Let the set of all boundary elements r, e= 1,2, ..E, be described
by means of the union (cf.[8]):
P

{ro} = U {r;},

(43)

p-I

where {r;} is a set of all these boundary elements, which are joined at
the p-th boundary vertex (p= 1,2, ..P).
Sensitivities can be computed using the following expressions (cf.[8]):

t J Wn.:-h~Wctr;W,
B

Smp-h

We

0-1

rp

p-l,2, .. P

(44)

h-:Z,l,O for d-3


h-l,O for d-2

where Ep denotes a number of boundary elements which join themselves in

57

the boundary node p, functions W are determined on each boundary element


by state fields of primary (PS) and adjoint (AS) systems and ~({) is
an interpolation function of the transformation field g(x;a) on the
o
boundary rp.

r;

Numerical examples
The problem of stochastic shape sensitivity analysis is considered for a
steel bracket (Fig. I) and a steel arch (Fig.2). It is assumed that the

stochastic tolerance range

-.-.----L.

Fig. 1

Fig.2

boundary undergoes small random variation measured along the normal to


boundary with elements of the cqvariance matrix k..=~rsc, where
-6 2
c=9.10 [m). These stochastic variations create the random tolerance
range. The mean values of the first natural frequencies
are
w o=2122[rad/s) for the bracket and wo=3314[rad/s] for the arch. The
-2
variances
equal
Var(w(y=34.96[s )
for
the
bracket
and
-2
Var(w(y=42.40[s) for the arch, respectively. Assuming that w(y) has
the . Gaussian distribution the probability 9' that the natural frequency is
included in the interval Iw(y)-wol <3/ Var(w)' is 911=0.9973.
References
1. Burczynski T.: Stochastic boundary value problems of elastostatics in
terms of the boundary integral equation method. Papers of Inst Civil
.
Engn. , Wroclaw Technical University, Vol. 1, No.28, pp.61-68, Wroclaw
1981 (in Polish).
2. Burczynski T.: The boundary element method for stochastic potential
problems. Applied Mathematical Modelling , Vo1.9, No.3, pp. 189-194,
1985.

58

3. Burczyl'lski T.: Stochastic formulation of wheelset dynamics by means of


the boundary element method. Proc. Sth Intemotional Wheelset
Congress, VoU, pp.n.3/1-3115, Madrid 1985.
4. Burczynski T.: Stochastic boundary element method for wave propagation
through continuous random media. In: BETECH'S6-MIT (Eds. J.J.Connor
and C.A.Brebbia), CMP 1986.
5. Burczynski T.: The boundary element procedure for dependence of
eigenvalues with respect to stochastic shape of elastic systems. Proc.
25th Symp. on Modelling in Mechanics, PTMTS, Vo1.2 , pp.235-238,
Gliwice-Kudowa 1986.
6. Burczynski T.: The boundary element method in stochastic problems of
elasticity. Mechanics and Computer, Vo1.7, pp.125-147, PWN Warsaw 1988
(in Polish).
7. Burczynski .T.: Boundary element method for deterministic and
stochastic shape design sensitivity analysis. In: Advanced Boundary
Element Methods (Ed. C.T.Cruse), pp.73-80, Springer-Verlag 1988.
8. Burczynski T.: The Boundary Element Method for Selected Analysis and
Optimization Problems of Deformable Bodies. Silesian Tech. Univ.
Publications, Gliwice 1989 (in Polish).
9. Burczyl'lski T.: Stochastic boundary element methods. In: Advances in
Boundary Element Techniques (Eds.J.H.Kane, G.Maier, N.Tosaka and
S.N.Atluri), Springer-Verlag (in press).
10. Burczyl'lski T. and Fedelinski P.: Shape sensitivity analysis of natural
frequencies
of
structures
using
boundary
elements.
Structural
Optimization, 2, pp.47-54, 1990.
11. Burczyl'lski T. and John A.: Stochastic dynamic analysis of elastic and
viscoelastic systems by means of the boundary element method. In:
Boundary Elements VII (Eds. C.A.Brebbia and G.Maier), VoU, pp.6/536/61, Springer-Verlag 1985.
12. Burczynski T. and John A.: The stochastic boundary integral equation
method for random vibrations of continuous media. Opole Technical
University Publications, No.I09, pp.61-74, Opole 1985.
13. Burczynski T. and John A.: Application of the stochastic boundary
element method to analysis of elastic and viscoelastic structures.
Proc. X Polish Cont. Computer Methods in Mechanics, pp.47-54,
Swinoujscie 1991.
14. Cheng A.H-D. and Lafe O.E.: Boundary element solution for stochastic
groundwater flow: random boundary condition and recharge. Water
Resources, Vo1.27, No.2, pp.231-242, 1991.
15. Drewniak J.: Boundary elements for random heat conduction problems.
Engineering Analysis, Vol.2, No.3, 1985
16. Ettouney M., Benaroya H. and Wright J.: Probabilistic boundary element
methods. Chapter 5 in: Computational Mechanics of Probabilistic and
Reliability Analysis (Eds. W.K.Liu and T.Belytscbko), Elmepress
International 1989.
17. Mr6z Z.: Variational approach to shape sensitivity analysis and
optimal design. in: The Optimum Shape (Eds. J.A.Bennet and
M.E.Botkin), Plenum Press, New York 1986, pp.79-110.
18. Nakagiri S., Suzuki K. and Hisada T.: Stochastic boundary element
method applied to stress analysis. In: Boundary Elements V (Eds.
C.A.Brebbia, T.Futugami and M.Tanaka), pp.439-448, Springer-Verlag
1983.
.
19. Nakagiri S. and Suzuki K.: Indeterminate shape modification for stress
reduction based on boundary element sensitiVity analysis. In: Computer
Aided Optimum Design of Structures: Recent Advances (Eds. C.A.Brebbia
and S.Hernandez), pp.85-94, CMP, Springer-Verlag 1989.

BEM Analysis of Crack Propagation in Concrete Based on


Fracture Mechanics
A.H.Chahrour* and M.Ohtsu**

*.*

Graduate School of Kumamoto University.


Professor, Department of Civil & Environmental
Engineering, Kumamoto University, Japan.

Summary
The boundary element method (BEM) is applied to fracture mechanics. The procedure is
utilized to clarify crack mechanisms and to predict crack propagation in concrete beams.
Mode-I fracture of center-notched concrete beams is analyzed, on the basis of a critical stress
intensity factor (K1r) criterion of the linear elastic fracture mechanics LEFM and a tensile
softening model of the non-linear fracture mechanics NLFM. Results show that the tensile
softening model reasonably simulates global behaviors of notched beams, although the ~c
criterion could recover an essential feature of mechanical behavior. 1\vo-domain BEM analysis is introduced to analyze crack propagation in arbitrary directions. A mixed-mode fracture
of a center-notched beam subjected to anti-symmetric loading is analyzed, based on the criterion of maximum circumferential stress. Crack trajectory observed in an experiment is in
good agreement with that of BEM analysis.
Introduction
The applicability of fracture mechanics to the failure analysis of concrete structures is actively
investigated. In this respect, the finite element method FEM is widely employed as a numerical procedure (1), (2).
In a discrete model of fracture mechanics, the whole domain is referred to as homogeneous
and linearly elastic. An analytical aspect is limited to the treatment of crack boundaries.
Both a stress-free crack of LEFM and a fracture process zone of NLFM correspond to nucleating new boundaries in the numerical procedure. Consequently, FEM is not a smart choice
to analyze the discrete crack model because it requires internal nodes and elements to model
the domain. In contrast, BEM deals with the discretization of only the boundary and holds a
privileged applicability to the discrete crack model.

In the present paper, 2-D BEM analysis is carried out to investigate mode-I fracture of
simply supported center-notched beams under four-point bending. Analytical results are
compared with results obtained from the experiments, Here, the objective is to predict loadCMOD (crack mouth opening displacement) curves. 1\vo criteria based on LEFM (linear
elastic fracture mechanics) andNLFM (non-linear fracture mechanics) are employed.
Mixed-mode fracture of center notched concrete beams subjected to anti-symmetric loading
is investigated by using a two domain BEM analysis (4). Crack trajectory based on the criterion of maximum circumferential tensile stress is investigated and compared with experimental observation.

60

Experiment
Mechanical properties of concrete were as follows: compressive strength was 36.3 Mpa, tensile strength was 3.2 Mpa, Young's modulus E was 29.4 Gpa, and Poisson's ratio v was 0.21.
These values are averaged results of three cylindrical specimens of 10 ern diameter and 20
ern height, tested after 28 days of moisture curing. Center-notched beams with 2 ern and 4cm
notch depths were cast. All specimens were of dimension 40 ern in length, 10 ern in depth,
and 10 ern in width.
I.

52

;of

Load Cell

%1
IE

51

Z"I

Fig. 1 Experimental set-up for a four-point qending test


Mode-I fracture was investigated by four-point bending tests, where a servo-valve controlled loading machine was employed and complete load-CMOD curves were obtained. An
experimental set-up is shown in Fig. 1. CMOD was measured by using a clip gauge inserted
at the notch bottom. Load was applied through a load cell at the rate of O.OOOSmm/s and a
load-CMOD curve was drawn in a X-Y plotter. Acoustic emission (AE) events were monitored by an AE sensor, a pre-amplifier, a discriminator, and an event counter. The critical
stress intensity factor ~c of concrete was obtained as 0.526 Mpam.s, which was determined
from the load level responsible for the nucleation of the fracture process zone as detected by
AE. The fracture energy GF was approximated by a formula (1- V2)KldE in the case of
plain strain (3).

P/2

10

Notched-beam specimen
(eM I

10

....

=J:

15

Fig. 2 Experimental set-up for mixed-mode fracture


An experimental set-up for studying mixed-mode crack propagation in a concrete beam
subjected to anti-symmetric loading is shown in Fig. 2. Center notched beams were loaded
until a crack surface was generated all the way from the notch tip to the top of the beam in a
tortuous manner.

61

BEM Formulation and analytical models


The boundary integral equation is given, as follows:

fJ Gik(X,y)tk(y) - Tik(x,y)u.(y) ]dS

Cuj(x) =

(1)

where C is the configuration coefficient determined from eliminating rigid motion in Eq. 1.
Gik.(x,y) is the fundamental singular solution for the 2-D (two dimenSional) linear elastic field
at the plane strain state, and Tik(x,y) is the associated traction solution.
Mode-I fracture is analyzed by considering the right-half portion of the beam shown in Fig. 1
due to symmetry. A model is referred to as a half beam model shown in Fig. 3. The boundary of the specimen is discretized into elements L G=l, ... ,N), and thus Eq.1 becomes,

Nf

N(

Cuj(x) + . E T ik(x,y)u.(y)dS = . E J?ik(x,y)t.(y)dS


J=1 ~
J=l Lj

(2)

In the case of a linear element indicated in Fig. 4, displacement {u} and traction {t} on the
element is represented by,

{u(a)}

=(1 -

a)u/2 + (1 + a)u k+1/2

(3)

{t(a)} = (1 - a)t/2 + (1 + a)tk+l/2

(4)

Substituting Eqs. 3 and 4 into Eq. 2, the resulting equation can be solved by adopting the 4point Gaussian integration for the displacements and tractions at nodes. In Fig. 3, a fine
mesh of 2.5mm length is assigned on the boundary of the notch and the ligament, while a
coarse mesh is taken on the other boundaries.
y l --

--"'::..:..:....--1

L.ot-------------~----------------~~--------~
)(
scm
-"
ISCm

Fig. 3 Half beam model (mode-I fracture)

To analyze crack propagation in a whole domain, a multi-domain approach was employed


(4). Here, a constant element is adopted for traction and a linear element is assigned for
displacement on the element,

62
YYk

~+1
/

il
a

11

11

10

Fig. 4 Linear boundary element


{t(a)}

=tj(YYk) =Constant

(5)

{u(a)} = (1 - a)uj (yJf2 + (1 + a)uj (Yk+l)12

where a is the variable of the local coordinates as shown in Fig. 4. The integration of each
element is carried out by the Gaussian 4-point scheme. Taking into account diagonal terms
in Eq.1, we obtain the matrix form,

(6)
To make the numerical scheme comparable to the conventional FEM code, Eq. 6 is converted
into the equation, not of traction on the element but of nodal forces. Eventually the following
equation is obtained,

(7)
where N.Jgn is the shape function for obtaining nodal forces from tractions. Eq.7 is of the
same stittness matrix as in FEM, and thus is readily coupled with the other domain.
The two domain BEM was applied to analyzing mode-I fracture based on the LEFM criterion.
Fig. 5 shows an analytical model which is referred to as a whole beam model, where the
same mesh division as the half beam model is taken.
y

1(!Jam

P/2

P/2

\V

]I

Q)I
.A.
15cm

150m

..
.,.

5cm

..

Fig. 5 Whole beam model (mode-I fracture)


A model for the analysis of mixed-mode fracture is given in Fig. 6 where double nodes are
assigned along the notch, and a broken line indicates the stitching boundary. When a crack
propagates, a node at the tip is separated into two nodes. It implies adding a new node on

63
Scm

P/2

"tI'-

.,

_"tI'

IDem

IDem

..

./

II

15001

5em

Fig. 6 Two-domain BEM model (mixed-mode fracture)


the stress-free crack surface, resulting in the increase of the total number of nodes.
The stress intensity factors at the crack tip were determined by Smith's one point
formulae (5). IS and K". are determined by using displacement u. in the X-direction and
1
displacement Vj in the Y-direction, shown in Fig. 7,
~

=(21T /L)1I2E[ vj(B) - vlA) ]/4(1 -

Ku =(21T 1L)1I2E[ uj(B) -

v~

(8)

uj(A) ]/4(1 - v~

(9)

Here L is the distance from the notch tip.

V,Y

Fig. 7 Crack tip element

Analytical procedure
(1) Mode-I Fracture
In the analysis of mode-I fracture based on the tensile softening model, a linear model
shown in Fig. 8 was selected to model the fracture process zone. A Clack propagates when

64

stress at the crack tip is equal to the tensile strength of concrete. Cohesive forces obtained
from the relation In Fig. 8 are applied at the nodes of the fictitious crack boundary. The
analytical procedure is the same as developed for FEM (1).

Fracture
energy GF
w.

Crack opening displacement


Fig. 8 Tensile softening curve
In the analysis of mode-I fracture based on LEFM, both the half beam model and the whole

beam model are considered. In the computation of the stress intensity factors, only ~ is
considered. The analytical procedure is iterated, as follows: at a stage, ~ at the crack tip is
computed due to unit load and actual load is determined as equal to ~cfKt. Then, the location of the crack tip is advanced up to the next node, removing the constraint at the previous
notch tip.
(2) Analysis of Mixed-Mode Crack Propagation
Mixed-mode crack propagation in a center-notched concrete beam subjected to anti-symmetric IOdding was analyzed. Due to the lack of symmetry, the whole domain of the concrete
beam is taken into account By employing the two domain approach, the region of the left
hand side and that of the right hand side to the crack tip arc stitched, holding the continuity of
displacement and traction along the ligament. The direction of crack growth is determined
based on the criterion of maximum circumferential stress (6).
Jin e + Ku(3cose -1) =0,

(10)

where the stress intensity factors ~ and Ku are computed from Eqs. 8 and 9, by substituting
relative displacements at the nodes on the crack tip element. From Eq. 10. the direction of
crack propagation is given. Here, e is taken to be the angle corresponding to the bigger
maximum tensile strength as shown in Fig. 9.

O~----------------

Fig. 9 Crack growth direction

65
In each step, the crack propagates by 5mm length at a time and a new coplanar stress-free
boundary is added in front of the crack tip. Then, the stitching boundary is created by
connecting the crack tip to one node on the top surface of the beam, which is denoted by Q in
Fig. 6.
Analytical results and discussion
In the numerical analysis, it is assumed that a crack starts at the crack tip and propagates
upward through the ligament until failure of the beam occurs. In the case of 2cm notched
beams undergoing mode-I fracture, the load-CMOD curves are shown in Fig. 10. An
experimental result is labeled as "Expt.". An analytical result by the tensile softening model
is indicated as "Coh.". Results of the half beam model and the whole beam model are shown,
as "LEFM(half)" and "LEFM(whole)", respectively. Although the agreement between the
experimental result and the analytical results is not excellent, all analytical results reproduce
essential feature of the experimentalload-CMOD curve.
1400

1200

,--.
"-

1000

100

600

'-'

//

...-.'

,,.

.:!

;;:-,

.,

~~.'

~~

.. ,

::: ! . . . >.

.. '

: ....

0.1

0.05

0.15

CMOD(mm)

0.2

0.25

Fig. 10 Load versus CMOD curves in a 2 cm notched beam


IDem

8em

_,

BEM

----Crack trace

IDem

IDem

15cm

5em

14

Fig. 11 Crack trajectories in the BEM analysis and experiment

66
It is noted that the result of IEFM(whole) evaluates less load-bearing capacity than that of
IEFM(half). It may result from the fact that the boundary conditions and the global stiffness
of the models are different. It is obvious that the LEFM models could not simulate the
nonlinear behavior proceeding the peak load.

An analytical result obtained for the case of a 2 em notched-beam undergoing mixed-mode


fracture is shown in Fig. 11. Crack trajectories are compared between an analytical result and
an experimental result. Since the crack elements of 5 mm length were employed, the determined profile of crack trajectory is not smooth and tortuous. It is observed that the agreement is reasonable up to the stage where the simulation analysis becomes unstable.

Concluding remarks
The applicability of BEM analysis to the discrete crack model of mode-I and mixed-mode
fracture is investigated on the center-notched concrete beams. The results obtained for
notched beams undergoing mode-I fracture are quite acceptable, while the analysis of
mixed-mode fracture need further research to improve the procedure. The applicability of
BEM analysis to fracture mechanics is so promising that BEM can be considered as a powerful tool, in comparison with the complicated aspects in FEM application to fracture mechanics.
References
1. A.Hillerborg, M.Modeer and P.E.Petersson, "Analysis of Crack Formation and Crack
Growth in Concrete by Means of Fracture Mechanics and Finite Elements", Cement & Concrete Research No.6, 1976, pp. 773-782.
2. A.R.Ingraffea & V.Saouma, "Numerical Modeling of Discrete Crack Propagation in Reinforced and Plain Concrete", Fracture Mechanics of Concrete: Structural Application and
Numerical Calculation, Martinus Nijhoff Publishers, 1985, pp. 171-225.
3. A.H.Cilahrour and M.Ohtsu, Proceedings of Japan Concrete Institute JCI, Vol. 12, No.2,
1990, pp. 865-870.
4. T.A.Cruse and E.Z.Polch, Proceedings of 3rd Japan Symposium on Boundary Element
Methods, JASCOME, 1986, pp.111-133.
5. R.N.L.Smith and J.C.Mason, itA Boundary Element Method for Curved Crack Problems in
Two Dimensions", Boundary Element Methods in Engineering, edited by C.A.Brebbia,
Springer, Berlin, 1982, pp. 472-484.
6. M. Ohtsu, Theoretical and Applied Fracture Mechanics, Vol. 9, No.1, 1988, pp. 55-60.

On Free Space Green's Function for High Order

Helmholtz Equations
A. H-D. Cheng
DEPARTMENT OF CIVIL ENGINEERING, UNIVERSITY OF DELAWARE
NEWARK, DELAWARE, U.S.A.

H. Antes
INSTITUT FUR ANGEWANDTE MECHANIK, TECHNISCHE UNIVERSITAT BRAUNSCHWEIG
D-3300 BRAUNSCHWEIG, GERMANY

Abstract
Many coupled elliptic PDE system can be decoupled by using the linear partial
differential operator theory. Sometimes the decoupled system results in a high order
Helmholtz equation. Closed form Green's function in two and three dimensions for
such a system is derived in this work.

Introduction
Many non-transient problems in applied mechanics are governed by a system of coupled,
linear, constant coefficient elliptic partial differential equations. For transient problems,
similar equations result if the Laplace or Fourier transform is applied to eliminate the
time parameter and derivatives. The system of equations can be decoupled by applying
the linear partial differential operator theory} in which the operators can be manipulated as algebraic quantities. The system is therefore formally 'solved' (decoupled) by
applying the matrix rules for linear system of equations. Recently, in two drastically
different engineering applications, one concerns the groundwater flow in a multi-layer
aquifer system,2 and the other involves the vibration of Reissner plates,3 the same fundamental decoupled equation is derived which is in the form of a high order Helmholtz
equation.
The generalized form of this equation, written in the form of a free-space Green's
function solution, is the following:
(1)
In the above the constants Ai, i = 1, ... , n, are eigenvalues of the system, and 6(:.:, :.:1) is
the Dirac delta function. In the Reissner plate problem, n = 3 and Ai are real numbers
(positive and negative); while in the groundwater problem, n is the number of aquifer

layers to be treated and Ai are all negative.

68

In the present work, complex eigenvalues will be considered for generality. We notice
that in both application the governing equations exist only in two spatial dimensions.
Again to generalize, we shall consider herein both two and three-dimensional Green's
functions.

Two-Dimensional Green's Function


We first notice that (1) can be written in the recursive form:
(\7 2+ An)

4>,. = 4>,.-1
(\7 2+ A,._I) 4>,.-1 = 4>,.-2
(\7 2 + A2)

4>2
(\7 2 + AI) 4>1

(2a)
(2b)

4>1

(2c)

= t5(:e - :e')

(2d)

The solution of the last equation (2d) is well known4 5

4>1

-'4i Ho(1) (V" AIr)

(3)

whe:-e i = A, r = l:e - :ell, and H~I) is the Hankel function of the first kind of order
zero.6 We now solve the next equation (2c)
(4)

By inspection, we find the solution of (4) as


(5)

We can easily prove the above by the substitution of (5) into (4)
(\7 2 + A2)4>2 =

(\7 2 + A2)

[_i4 H~I)(Al -Ar)]


A2

+ [\7 2 + Al + ('\2 -

Ad] [

!H~:(~r)]

ht5(:e,:e ' ) + ht5(:e,:e' ) - -4i


"I -

"2

= -~ H~I)( Ftr)

"2 -

"I

H~I)(Ftr)
(6)

(7)

69
This equation is similar to (4) and each term on the right hand side will generate two
similar terms as (5), but with different coefficients. After some rearrangement, we obtain

1>

i [

-4

Hb

l )(

~r)

(A2 - AdP'3 - Ad

Hb

+ (AI -

JX;"r)
A2)(Aa -

Hb

l )(

).2)

+ ().I -

l )( y%r)
]
A3)(A2 - A3)

(8)

The solution can be repeated and the general solution takes the form:

1>" =

E "H(I)(

."

-.:

rr )

yAir

4 i=1 I1j=I.#i(Aj - A.)

(9)

The above solution can also be specialized for real and positive eigenvalues:

1>.. =

!.

"Yo( ../Air)

4 i=1 I1j=l.j~;().j - Ai)

(10)

where Yo is the Bessel function of the second kind of order zero. For real and negative
eigenvalues, defining k. = -Ai, we obtain

1>..

= _2.. t

. Ko( v'fir)

2rr i=1 I1i=l,i~i( ki - k j )

(11)

where Ko is the modified Bessel function of the second kind of order zero.

Three-Dimensional Green's Function


For equations in three-dimension, the solution of (2d) is again well known
A.

'/'1

= -4rrr
-1eivAir
1

(12)

For higher order equations, a close examination of the procedure reveals that the rules
used in the preceding section do not change for the three-dimensional case. We hence
have

and the general solution


1

eivrir

1>,. = - - E = ..::---~---::-:41!'r

For real and positive Ai, (15) becomes

1>" = __1_
41!'r

i=1 I1j=I.#i().j - Ai)

t .

cos( Ar)
.=1 I1j=l.i#i(Aj - Ai)

(15)

(16)

and for negative Ai = -ki'

(17)

70

One-Dimensional Green's Function


Although not of much use to boundary integral equation method, we also present the
one-dimensional results. The governing equation is
(18)

The solutions are


(19)
(20)

<Pn =

(21)

For positive and negative Ai, we have respectively


(22)
(23)

Conclusion
We have derived in this paper the free space Green's function for the high order
Helmholtz equation with complex or real eigenvalues, in one through three spatial dimensions. The success of the solution hinges on the decomposition of the equation into a
series of inhomogeneous Helmholtz equations, each defined from the previous one. The
fundamental solutions are built according the observation of the recurrence behavior.
These fundamental solutions however are just the raw forms for constructing the
actual Green's functions of interest. The additional operation often involves carrying out
the differentiation according to a 'cofactor matrix'.2.3 In carrying out the differentiation,
the recurrence relations (2) are again helpful in the mathematical manipulation.
The real purpose for deriving these solutions are for the BEM solution. Those results
are reported elsewhere7 8 hence not repeated here.

Acknowledgment
The a.uthors wish to a.cknowledge the financial support of the Deutsche Forschungsgemeinschaft through the project "Randelementmethoden" which allows the collaboration
to take place during the first author's four-week visit to Technische Universitat Braunschweig.

71

References
[1] Hormander, H., Linear Partial Differential Operator", Springer-Verlag, 1963.
[2] Cheng, A.H-D. and Morohunfola, O.K., "Multiple leaky aquifer system: 1. Pumping well solution," submitted to Water Resour. Res., 1991.
[3J Cheng, W. and Antes, H., "The dynamic fundamental solution of the ReissnerMindlin plate theory," in preparation.
[4] Morse, P.M. and Feshbach, H., Methods of Theoretical Physics, Pt. I & II, McGrawHill,1953.
[5] Greenberg, M.D., Application of Green's Functions in Science and Engineering,
Prentice-Hall, 1971.
[6] Abramowitz, M. and Stegun, LA., Handbook of Mathematical Functions, Dover,
1972.
[7] Cheng, A.H-D. and Morohunfola, O.K., "Boundary element formulation for multiaquifer systems," Computational Engineering with Boundary Elements, 1: Fluid
and Potential Problems, BETECH90, Univ. Delaware, eds. S. Grilli, C.A. Brebbia
and A.H-D. Cheng, Compo Mech. Publ., 145-156, 1990.

[8) Antes, H. and Cheng, W., "Vibrations of Reissner plates by BEM and D/BEM: A
comparison," Proc. Symp. Int. Assoc. Boundary Element Methods (this Proceedings), Kyoto, Japan, Oct. 14-17, 1991.

Boundary Element Model for the Seismic Analysis


of Arch Dams
J.Dominguez and O.Maeso.
Escuela Superior de Ingenieros Industriales, Universidad de
Sevilla, Av.Reina Mercedes, sin, 41012-Sevilla, Spain.

Summary
A three-dimensional boundary element model for the seismic
analysis of arch dams is presented. The soil and the dam are
assume to be viscoelastic domains the former being boundless.
The water is assume to be compressible subject to small
amplitude motions. The soil-water, soil-dam and water-dam
dynamic interactions are taken into account rigurously. The
analysis is done in the frequency domain and factors such as an
infinite soil, the effect of the local topography, the actual
non-uniform geometry of the reservoir and the spacial
distribution of the ground motion about the canyon walls are
taken into account in a direct way as apposite to the existing
FEM models which are not able to consider those factors.

Introduction
The analysis of the seismic response of dam-reservoir systems is
an important problem within the field of earthquake engineering.
The seismic response can not be studied considering the
structure as an isolated body under the influence of the base
motion. There are important effects due to dam-water,
dam-foundation and water-foundation interaction that make
necessary the use of models including the three media and the
interaction between them. Among the factors affecting the
response one can mention the soil characteristics at the site
(topography, mechanical properties, layering), the water
compressibility and the geometry of the botton. A great effort
has been dedicated to this subject in the last two decades. The
st udies of Chopra and his co-workers

[1,2]

who analysed the

significance of a number of factors contributing to the response


shold be metioned. Those studies have been done using the Finite
Element Method.

However,

due to the great complexity of the

73

problem

the

models

include

important

simplifications.

promising alternative to the FEM is the Boundary Element Method


(BEM).

Medina et

al. [3,4]

showed that the BEM,

due to

its

particular characteristics, is very well suited for this kind of


problems and allows to overcome some of the most important
shortcomings of the FE models. The abovementioned B.E. studies
refer to gravety dams
two-dimensional models.

which

can

be

represented

by

The analysis of arch dams requires of a three-dimensional model


and is in this context where

the shortcomings of existing F.E.

models become more obvious. These models represent the soil as a


massless elastic finite domain; the reservoir, as a small
irregular region connected to an infinite channel with uniform
cross section and the water-soil interaction by means of an
absortion coefficient which takes into account in a simplified
manner the transmition of waves from the water to the soil.
Because of these simplifications the F.E. models are unable to
represent

the

propagation

of

the

waves

in

the

soil

and

consequently the different values of the excitation at different


points of the dam foundation and the diffracted waves going into
the soil. This effect is particularly important when the
wavelength of the incident field is similar to the dimensions of
the dam. The F.E. model is also unable to represent properly,
the soil topography, large irregular geometries of the
reservoir,
the water-soil interac.tion and the combine
water-soil-structure interaction.' On the contrary, the B. E.
model presented in this paper is able to represent in an easy
and direct form the abovementioned factors which have important
effects on the seismic-response of arch dams [5] .
Model for the solid domains. Dam-soil interaction.
The system to be analysed, shown in Figure 1, consists of two
solid regions (the soil and the dam) and a fluid region (the
water). A dynamic interaction exist among the three regions and
the analysis of the response of the dam to incident seismic
waves requires of the study of the whole dam-soil-water system.

74
Time harmonic waves propagating vertically from the soil which
produce a unit free field vertical or horizontal motion either
along the transversal or the upstream direction are assumed.

Fig.i. Arch dam-water-foundation system.

Fig.2. Boundary elements model of the arch dam substructure.


In the present paper the Morrow-Point dam-soil-reservoir system
previously

studied

using

F.E.

by

Fok

and

Chopra

[2J

is

considered. The dam is modeled as a viscoelastic domain using

75

nine node quadrilateral and six node triangular elements with


quadratic shape functions in two directions both for the field
variables and the geometry. Figure 2 shows the B.E. model for
the darn. The properties of the concrete are: density
Kg/m 3 ;

Poisson's ratio

0.2;

shear modulus

= 2481.55

11500 MPa and

= 0.05 . Because of the available space only the

damping ratio

response of the darn to a horizontal upstream excitation is shown


in this paper. Figure 3 shows the amplification of the upstream
motion at

the

dam crest versus

frequency when the

soil

is

assumed to be rigid and the reservoir empty. The results show a


very

good agreement with the F.E. ones obtained using shell

elements (2). The frequency is normalyzed by the first natural


frequency of the darn. The darn crest displacement is normalyzed
by the horizontal motion produced by the incident waves at the
soil free surface far from the dam.

so
BEM

Fok &' Chopra

30

10

w/w1
Fig.3. Radial amplification. Rigid foundation, empty reservoir.
The effect of the soil-dam interaction is analysed by doing
coupled models including the dam and the soil. The soil is
modeled as

properties:

density

shear modulus

linear viscoelastic

=2641.65

solid with the

Kg/m3;

following

Poisson's ratio

11500 MFa and damping ratio

0.2;

= 0.05. The soil is

discretized using the same kind of elements as for the darn. The

76
discretization is left open at a certain distance from the dam.
Figure

4. a

shows

the

dam-soil

model

for

the

Morrow-Point

reservoir in which the water basin is consider to extend


uniformly to infinity for the sake of comparison with the F.E.
model. In Figure 4.b the model for the more realistic case of a
finite reservoir is shown.
symmetric models.

The Figure shows one half of the

The analysis of the dam response for both

geometries when the reservoir is empty produces almost identical


results. This shows that the soil topography far from the dam
does not have influence on the dam response for empty reservoir
conditions.

Fig.4. Geometry and boundary elements discretization of the


coupled system. (a) finite reservoir; (b) infinite reservoir.
Figure 5 shows the upstream amplification at the dam crest for
the dam-soil system. The results obtained show important
differences with those obtained by Fok and Chopra

[2]

using

finite elements. It can be easily shown that the differences are


due to the fact that the F.E. model, which includes a massless
soil

assumption,

can

not

take

into

account

the

spatial

distribution of the excitation. If the B.E. code is run assuming


density =0,

the obtained results are very close to the F.E.

results as shown in Figure 5.

n
-BEM
.... Fok &: Q.opra
. - BEM zero dcDsity
30

20

10

oo

Fig.5.

Radial

amplification.

Compliant

foundation,

empty

reservoir.
Model for the fluid domain

pam-soil-water interaction.

The water is modeled as an inviscid compressible fluid under


small amplitude harmonic motion. It's behaviour is governed by
the Helmholz aquation. The B.E. for the fluid are quadratic nine
or six node elements as in the solid case.

The

interaction

conditions between the water and the solid regions

(dam and

soil) are: equal normal pressure and displacement on the solid


and fluid boundaries in contact, and-zero shear traction on the
solid elements of the contact surfaces. Two geometries for the
reservoir are considered as shown in Figure 4. One includes an
infinite uniform rectangular cross section channel to represent
the reservoir extending towards a very long distance; the other
is finite consisting of an uniform region closed by a spherical
surface. For simplicity, the geometry of the reservoir has been
consider to be uniform; nevertheless, non uniform geometries of
the

reservoir

can

be

easily

modeled

with

the

quadratic

isoparametric elements used. The infinite channel in one of the

78
geometries allows for the propagation of standing waves which
send energy out of the system. The boundary conditions for this
infinite channel can be easily established as a relation between
water pressure and it's normal derivative at the cross section
[6]

BEM
Fok & Cbopra

.a

i
~

30

2D

III

w/w1
Radial amplification.
reservoir full of water.

Rigid

Fig.6.

SD

~I

40

It
~

30

..

2D

]~-~!

infinite

HEM infinite c:bumd

!NJ~/ ~'~\\;

III

foundation,

w/w'1
Fig.7.

Radial

amplification.

Rigid

infinite reservoir full of water.

foundation,

finite

and

79
Figure 6 shows the radial amplification at the dam crest for
full

reservoir,

rigid

foundation

and

infinite

reservoir

geometry. The results show a good agreement with the F.E. ones
[2]. The important effect of the reservoir geometry for rigid
foundation can be seen in Figure 7 where the difference between
the dam response for the infinite and the finite reservoir cases
is

shown.

reservoir

The

importance

geometry,

of

good

representation

provided easily by the

B. E.

of

models

the
as

opposite to F.E., is obvious.


The full
interaction problem including the dam,
the
compliant foundation and the water is analysed next. Figure 8
shows the radial amplification at the dam crest for the model
including the infinite channel. The results are compared with
those obtained with the existing F.E. model [2]. Substantial
differences between the B.E. and the F.E. results exist. This
differences are due to the simplifications of the F.E. model as
can be shown by running the B.E. code with zero density for the
soil

and

an

absorption

coefficient

for

the

water-soil

interaction as the F.E. model does. In such a case, results very


similar to those corresponding to F.E. are obtained as shown in
Figure 8.
A comparison between the dam crest radial
amplification for the infinite and the finite reservoir models

is shown in Figure 9. The important effect of the reservoir


geometry on the seismic response of the dam is clearly shown.
Again, the importance of a good representation of the reservoir
geometry, as the B.E. can easily accomplish, becomes clear.
Conclusions
A three-dimensional boundary element model for the seismic
analysis of dam-soil-reservoir systems has been presented. The
model takes into account rigurously the spatial distribution of
the excitation, the wave propagation in a viscoelastic boundless
soil,

the

soil-water

interaction and

the

actual

reservoir

geometry. All this factors have an important effect on the


seismic response of arch dams ad can not be taken into account
in a proper way by the existing F.E.

models.

Previous B.E.

80
models took only into account some of the interaction effects
and include important simplifications.

J
i
t

30

211

10

w/v1
Fig.8. Radial amplification.
reservoir full of water.

"
~

iii

foundation,

infinite

HEM iDliDite clJamJel


.... HEM finite cIwmel

.=8

Compliant

30

..

211

to
0

w/v1
Fig. 9.

Radial amplification . Compliant foundation,

infinite reservoir full of

water.

finite and

81

References
1. Hall,

J.F.;

including

Chopra,

A.K.:

hydrodynamic

(1983) 149-163.
2.Fok, K.L.; Chopra,

Dynamic

effects.

A.K.:

analysis

of

arch

J.Eng.Mech.Div.

dams

ASCE.109

Earthquake analysis of arch dams

including dam-water interaction, reservoir boundary absorption


and foundation flexibility. Earthq.Eng.Struct.Dyn.

14

(1986)

159-184.
3.Medina, F.; Dominguez, J.: Boundary elements for the analysis
of dams including dam-water-foundation interaction effects. I.
Eng.Analysis with Boundary Elements. 6 (1989) 151-157.
4.Medina, F.; Dominguez, J.; Tassoulas, J.L.: Response of dams to
earthquakes

including

the

effects

of

sediments.

J.Struct.Eng.ASCE. 116, No. 11 (1990) 3108-3121.


5.Clough, R.W.; Zhang Guangdon: Findings and Comments. Earhquake
behavior of arch dams. Proceedings of the China-U.S. Workshop.
Pergamon Press (1987).
6. Dominguez, J; Meise,

T.:

On the

use of the

BEM

for

wave

propagation in infinite domains. Eng .Analysis with Boundary


Elements. 8, No.3 (1991) 132-138.

Stability of Time Domain Boundary Element Method


in Wave Propagation Problems
T. Fukui* and K. Tani**

* Department of Architecture and Civil Engineering, Fukui University, Fukui 910, Japan
** The Japan Research Institute, Nishi-ku, Osaka 550, Japan

Sununary
Stability of the time domain boundary element method in two-dimensional wave propagation
problems was studied. A numerical testing method based on finding the characteristic roots
of the time marching boundary element equation was used to decide the stability of typical
problems. The results given by the tests are: the boundary element method for exterior
Neumann problems is stable; the boundary element method for Dirichlet problems is
unstable.

1. Introduction

This paper is concerned with the stability property of the time domain boundary element method in two-dimensional wave propagation problems. Stability consideration
is an important step in developing time domain boundary element methods. In the
previotLs works,1,2,3 we studied the property of error propagation of numerical boundary element systems. The numerical experiments led to some phenomenological
results: the error propagates but decays with time; the error decaying is slow if the
time step interval is small; the error may grow unless the coefficient matrices of
boundary element system is calculated accurately. Even though these statements are
useful as the guidelines to design boundary element models, they still cannot be precise criteria to decide the stability of the models.
In this paper, the stability of the boundary element method is examined by a numerical method in which the stability criterion is precisely written by the characteristic
roots of the numerical boundary element system. In two-dimensional wave propagation problems, the boundary element equation is an indefinite discrete-time series
equation (difference equation) in vector form, for which it is difficult to obtain the
characteristic equation. To avoid this difficulty, the original equation is truncated
and is decomposed into a set of scalar equations of each modes. Therefore, the
characteristic roots are obtained for each modal equations. The stability of typical
problems with both Neumann and Dirichlet boundary conditions is tested by this
method, and the dependency of stability condition on discretization parameters and
other influential factor of the numerical boundary element system is clarified.

83

2. Time Domain Boundary Element Method in Wave Propagation Problems


The formulation of the time marching boundary element method in two-dimensional
wave propagation problems will be reviewed here.
The initial-boundary value problem in a region B with its boundary aB is defined by
\7 2 u

= ~ azu
c2

in B

at2

au
at

m B

-(z,O) = vo(z)
u( z,t)

(1)

8u
-(z,t) = s(z,t)

~(x,t)

an

where c is the wave velocity, \7 2 is the Laplace operator in two-dimensional space,


and 01
implies the normal derivation on the boundary. Uo and Vo are given initial
values, and
and; are given boundary values. The region B is exterior or interior,
and the boundary aB is assumed to be smooth here.

an

A boundary integral equation corresponding to the problem (1) is given by4

ftl

au (lI,t-r)ds ll dr = u(z,t)
(2)
o 8B27r'V~-r2/c2 an
Heaviside's function, r=lz-yl, u=au/ot, and the subscript y implies
- In

H(r-r/c)

,J

where H(t) is
that the operation is carried out on the point y. ;;. is a known function that is the
Cauchy solution in the infinite region by the given initial condition or an incident
wave.
Introducing approximations of u(z,t) and au(z,t)lan

u(x,t) =

EEu K'J</>K(t)'l/Jj{z),
Kj

au
an

-(z,t) =

EEs K'J</>'K{t)'I/J/z)

(3)

Kj

substituting them into the integral equation (2), and using the collocation method to
satisfy the equation, a series of discretized equations are obtained. The N-th step
equation becomes
BOuN + B1uN- 1 + B 2u N- 2

+ ... + BN-1u1
(4)

where the components of the coefficient matrices and vectors are defined by

84

[u-KIi

= u(zi,Kdt)

where r;= Iz;-yl, and 0KO is Kronecker's delta.


In Dirichlet problems u(z,t)=u(z,t) is given on aE in (1), that is, uK is given at each
time steps. Thus, the general expression (4) becomes
_ AOs N - Al s N- 1

A2sN-2 _ ... _ AN-lsI

= f'l

(5)

where f'l is a known vector. At N-th time step, since the previous solutions s1,
sN-l have already been determined, (5) is a linear equation about sN. On the other
hand, in Neumann problems sK is given, and (4) becomes
BOuN + B l u N- 1

+ B 2 u N- 2 + ... + BN-lul = gN

(6)

with a known vector gN. The boundary element equations (5) and (6) have a form of
linear filter, which can be considered as a difference equation.

3. Nmnerical Method of Stability Examination


Tl.e stability of a difference equation can be decided by examining its characteristic
roots. To derive the characteristic equation of the boundary element equation (5) or
(6) numerically, it is convenient to transform the original equation into a set of
reduced scalar equations by two kinds of operations: truncation and modal decomposition. This transformation reduces the degree of the obtained characteristic equation
so that it can be solved accurately by the usual polynomial equation solver. The
Neumann problem (6) will be considered hereafter as the example.
If N in (6) is an immense number. the corresponding characteristic equation may be
of so high degree that the accurate roots cannot obtained numerically. However, the
truncated equation of (6)

BOuN + B l u N- l

+ B2 u N- 2 + .. , + BMuN- M = gN

(7)

where M < N is the truncation length, has good approximate characteristic roots for
the original equation (6) if M is sufficiently large. S Therefore, the truncated equation
(7) is considered useful to examine the stability of (6). The appropriate value of M
for the usual equation solver is less than a few hundreds.
Moreover, the modal decomposition reduces the degree of freedom of (7), and it leads
to scalar form of characteristic equation that is convenient to solve numerically. Let
Xi (i=l, 2, ... ) be a set of linearly independent vectors, by which the unknown
vector u is expressed as

(8)

85
If Xi satisfies the orthogonality condition

(9)

i=loj

(7) can be decomposed into a set of scalar equations

tf + bltf-1 + bltf-2 + ... + Wtf- M

g[l

(i=1,2, ... )

(10)

by the modes. Xi may be obtained as eigen-vectors of the boundary unknown vector


u using the corresponding frequency domain boundary element method. 6
The characteristic equation of (10) becomes
(11)

The solutions of (11) are the characteristic roots of j-th mode.


The criteria of unstability of the equation (7) is: the solution of (10), i.e., the solution of (7), can be unstable, if there exists at least one root whose absolute value is
larger than 1, or if there exists a double root whose value is 1. 7
4. Stability of Exterior Ne1llll8DD. Problem
The stability of the boundary element method in exterior Neumann problems is
examined here. Since the stability does not depend on the shape of the boundary, an
exterior region of a circular boundary with radius a is chosen as the test problem
because the eigen-vector Xi that satisfies (9) can be obtained easily for a circular
boundary. In this example, 4>K(t) and tPi(z) of (3) are the bases of piecewise linear
function and B-spline of the fourth order, respectively. The discretization parameters
are the time step interval t:.t and the number of knots m on the boundary. (6) is
truncated into M+1 terms as in (7). The time for the wave running through the circle, T o=2a/ c, is chosen as the unit measure of time interval.
4.1. Characteristic Roots
Fig. 1 shows the distribution of the characteristic roots of some lower modes. They
are plotted on the complex plane, where the unit circle is also drawn. In each figures,
there are two kinds of roots: external and internal ones. The external roots are close
to the unit circle, but none of them is out of it. The internal roots are mostly
arranged on an internal circle. The higher is the mode, the smaller is the radius of
the internal circle.
Fig. 2 shows the root distribution dependency on the truncation length M in the
lowest mode, and Table 1 shows the absolute values of the external four roots. All
values in Table 1 are less than 1. The external roots depend little on M. The absolute values of them are almost equal when M? 3 To/ t:. t. On the other hand, the
number of the internal roots increases, and the radius of the internal circle becomes

86

..

(al Oth mode

0. 0

'0

(b) 1st mode

2nd mode

....
.
o

(d) 3rd mode

(e)

00

(el 4th lode

Fig. 1. Distribution of the characteristic roots of some lower modes in


exterior Neumann problem. m=16, tlot= T o/16, and M=3To/t::.t.

(al H=1.25T.!At

(b) H=3T.! At

(el H=8T./At

Fig. 2. Distribution of the characteristic roots of the Oth mode for


different truncation parameters M in exterior Neumann problem.
m=S and t::.t= To/S,
large as M becomes a large number.
Fig. 3 shows the root distribution dependency on the time step interval t::.t in the
lowest mode. The numbers of both external and internal roots increase as t::.t
decreases. The number of the external roots equals To/ t::.t. The radius of the internal circle also becomes larger as t::. t decreases.

87

Table 1. Absolute values of the external roots of the Oth mode in exterior
Neumann problem. m=8 and ~t= To/8.
Root

1.25To/~t

3To/~t

8To/~t

1
2
3
4

0.988565
0.981508
0.917766
0.787098

0.996466
0.973984
0.921312
0.826513

0.995966
0.973638
0.920926
0.870402

I
(a) li t=T./8

(b) li

t=Te/16

(c) li

t=T./24

Fig. 3. Distribution of the characteristic roots of the Oth mode for


different time step intervals ~t in exterior Neumann problem.
m=16 and M=3To/~t.

1. OOf----t-l-..:::----f------l

..

'N

O. 99f----+---~-f-_"<_----l

.ode 2

O. 98 ~O---::-':;-;:----;;-';-;;--"--;:-'0.15

A tIT.

Fig. 4. Maximum absolute values of the characteristic roots of some lower


modes v.s. ~t in exterior Neumann problem. m=16 and
M=3To/~t.

Fig. 4 shows the relation between the maximum absolute values of lower mode
characteristic roots and the size of the time step interval ~t. As ~t decreases, the
maximum absolute values increase up to I, but they never exceed 1. The value of

88

the Oth mode is closest to 1. The higher is the mode, the smaller is the maximum
absolute value when ~t is small.
The above results show that the system's characteristic is expressed by the external
roots. Only the time step interval effects to change the positions and the number of
them, but the truncation length effects little. All roots are within the unit circle;
therefore, the boundary element equation (7), i.e., Neumann problem (6), must be
stable. The lowest mode is most sensitive to the stability of the equation.
4.2. Effect of Integration Error on Stability

As the truncation length M becomes large or the time step interval ~t becomes
small, the internal roots are getting large and they somehow push the external roots
toward the unit circle. This fact suggests that for large M or small ~t the absolute
values of external roots almost equal to 1, and that the discretized system may be
unstable with additional numerical error in the system.
Table 2 shows the effect of error in the system of (7). Integration error that is controlled by the adaptive integration methodS is introduced artificially in calculating
the coefficient matrices of (7). The table shows that the absolute value of the biggest
root exceeds 1 when ;:::10- 3 That is, if the coefficient matrices of (7) have error,
the system can be unstable. This fact agrees with the result obtained by numerical
experiments on error propagation of boundary element systems. 1,2,3
Table 2. Influence of numerical error on absolute values of the external
roots of the Oth mode in exterior Neumann problem. m=4,
~t= To/16, and M =3 To/ ~t.
Root

=10- 7

<:=10- 4

=10- 3

=10- 2

1
2
3
4
5
6
7
8

0.999802
0.997285
0.992423
0.984334
0.971989
0.953140
0.923687
0.883648

0.999839
0.997277
0.992398
0.984341
0.972001
0.953200
0.923719
0.883742

1.000046
0.997611
0.992727
0.984786
0.972569
0.953739
0.924238
0.883664

1.001829
0.997591
0.994116
0.982156
0.976624
0.952462
0.928521
0.881444

5. Stability of Dirichlet Problem

The stability of the boundary element method in Dirichlet problems, i.e., the equation (5), is examined here. The boundary element equations of the exterior and interior problems are same in (5) except the definition of sK. A circular boundary is
chosen in the test problem. The discretization is same with the problem in the
previous section except that t/>'K(t) of (3) is the base of piecewise constant function.

89
Fig. 5 shows the distribution of the characteristic roots of some lower modes. The
distribution patterns are similar with that of the Neumann problem (Fig. 1). However, the radius of the internal circle is larger than that of Fig. 1, and a few roots are
out of the unit circle in the lowest mode.

(a) Oth lode

.
. ...

(b) 1st lode

(el 2nd lode

0.

0.

(d) 3rd mode

(e) 4th lIode

Fig. 5. Distribution of the characteristic roots of some lower modes in


Dirichlet problem. m=16, At=To/16, and M=4To/At.
Fig. 6 shows the root distribution dependency on the truncation length M in the
lowest mode, and Table 3 shows the absolute values of the external three roots.
Some of the values in Table 3 are larger than 1. Still the root distribution dependency on the truncation length M has the similar property with that of Fig. 2, but
the internal circle is close to the unit circle.
Table 3. Absolute values of the external roots of the Oth mode in
Dirichlet problem. m=8 and At= To/S.
Root

2T.o/At

4To/At

6To/At

8 To/ At

0.995294

1.013411

1.002323

1.005842

2
3

1.006256

0.998547
0.978884

0.994711
0.980692

0.995517
0.981506

0.995522

90

(b) 1I=4T,1 At

(a) 1I=2TaI At

(e) 1I=8T.1 At

Fig. 6. Distribution of the characteristic roots of the Oth mode for


different truncation parameters M in Dirichlet problem. m=8 and
Llt= T o/8.

(a) A t=T8/4

(c) A t=T./16

(b) A t=T./8

Fig. 7. Distribution of the characteristic roots of the Oth mode for


different time step intervals Llt in Dirichlet problem. m=8 and
M=4To/Llt.
1.02

1.0

to

oS

""

---

~ .ode 1
~

aode

.ode 0

----

,Ode 2

1~.Ode
1\

0.1

\\

0.2

0.3

A tIT.

Fig. 8. Maximum a.bsolute values of the chara.cteristic roots of some lower


modes v.s. At in Dirichlet problem. m=lB and M=4To/At.

91

Fig. 7 shows the root distribution dependency on the time step interval Llt in the
lowest mode, and Fig. 8 shows the relation between the maximum absolute values of
the lower mode characteristic roots and the size of Llt. The value of the Oth mode is
larger than 1, but it is approaching to 1 as Llt becomes small.
In Figs. 5-8 a few characteristic roots are always out of the unit circle in Dirichlet
problem; therefore, the boundary element equation (5) can be unstable.
6. Conclusions

A numerical stability test based on finding characteristic roots of difference equation


is proposed and is applied to the time domain boundary element method in twodimensional wave propagation problems. The boundary element equations of typical
problems are examined by this method. The followings are concluded from the tests:
1. The boundary element method for exterior Neumann problems is stable.

2. The boundary element method for Dirichlet problems is unstable.


3. Even in exterior Neumann problems, unless the coefficient matrices of the boundary element equation are accurate enough, the system may be unstable.

The present method of stability testing is general in time domain boundary element
method and is applicable to other type of boundary conditions, three-dimensional
problems, and other physical problems such as heat conduction problems and elastic
wave propagation problems.
Refere:lces
1.

Fukui, T., A numerical study on errors of time marching boundary integral equation
method in wave problems, in Theory and Applications of Boundary Element Methods, ed.
Q-H. Du and M. Tanaka, Tsinghua Univ. Press (1988), 221-228.

2.

Fukui, T. and K. Tani, On numerical error of time marching methods in wave problems,
Proc. 5th Japan Nat. Symp. on Boundary Element Methods, JASCOM (1988),269-274.
Fukui, T. and K. Tani, Error and stability of numerical solution of time marching boundary element method in wave problems, in Advances in Boundary Elements, Vol. 2, ed.
C.A. Brebbia and J.J. Conner, Springer (1989), 221-230.
Fukui, T., Time marching analysis of boundary integral equations in two dimensional
elastodynamics, in Innovative Numerical MethotU in Engineering, ed. R.P. Shaw, et al.,
Springer (1986), 405-410.
Fukui, T. and K. Tani, On truncation scheme for time marching BEM in twodimensional wave problems, in Theory and Applications of Boundary Element Methods,
ed. Q-H. Du and M. Tanaka, Pergamon Press (1991).
Kitahara, M., Boundary Integral Equation Methods in Eigenvalue Problems of Elastodynamics and Thin Plates, Elsevier (1985).
Koopmans, L.H., The Spectral Analysis of Time Series, Academic Press (1974).
Fukui, T., Some numerical integration schemes of the boundary singular integrals in 2D
BIEM formulations, Proc. 4th Japan Nat. Symp. on Boundary Element Methods,
JASCOM (1987),329-334.

3.

4.

5.

6.
7.
8.

Computational Aspect of Ground-Water Management

Tanehiro FUTAGAMI and Hiroyuki OGATA

Department of Civil Engineering, Hiroshima Institute of Technology


Miyake, Itsukaichi, Saeki, Hiroshima, 731-51, Japan

Abstracts
In recent years, a great deal of research work of distributed parameter control systems
has been carried out by scientists and enginers.
By combining the boundary element method with linear programming, the boundary
element and linear programming method (the BE&LP method) has been developed and
systematized in order to solve distributed parameter control systems.
In this paper an efficient computational algorithm of the BE&LP method for optimal
control problems of ground - water management in steady - state confined aquifer is
studied by noting its special structures. The BE&LP method utilizes the advantages of
established numerical techniques of both the boundary element method and linear
programming. The tractability in both the boundary conditions and the complicated
equality or inequality constraints makes sure that the method becomes a powerful
technique for various problems of optimal control and design in distributed parameter
systems.

1. Introduction
Recently, a great deal of research work for optimization of distributed parameter
systems characterized by partial differential equation systems have been studied (11).
By combining the boundary element method with linear programming, the boundary
element and linear programming method (the BE&LP method) has been developed in
order to optimize partial differential equation systems with both equality or inequality
constraints and a linear objective function (8). Such systems are frequently encountered
in various problems of control and optimal design and are of special interest in field
problems (ground - water flow, diffusion, heat conduction, electric/magnetic potential,
etc.). The BE&LP method utilizes the advantages of established numerical techniques of
both the boundary element method and linear programming.
In this paper an efficient computational algorithm of the BE&LP method in ground-

water management is studied in order to save computer time and memory due to its

93
special structures. The finite element and linear programming method (the FE&LP
method) has been also developed by the combined use of the finite element method
with linear programming (6, 7). The combined use of the finite difference method with
linear programming has been also studied in the field of ground-water management (I).
In order to optimize transient partial differential systems the finite element and dynamic
programming method (the FE&DP method) (7, 10), and the boundary element and
dynamic programming method (the BE&DP method) (9) have been studied. The BE&LP
method and these related methods may become useful techniques for optimization of
distributed parameter control systems because of their tractability in the boundary
conditions, the initial and final conditions and the equality or inequality constraints.

2. Formulation of the BE&LP Method in Ground Water Management


2.1 Systems of Basic Equations
The basic equation systems describing steady - state confined aquifer management with
constraints and an objective function are as follows (see Fig. 1):
Objective Function

Z = Opt. f( {</l}' (q), ( e))

(e)

(throughout the whole domain Q)

(\)

in Q

(2)

subject to:
Equilibrium Equations
Governing Differential Equation
2

TV </l=e+w
Boundary Conditions

</l=</l
d</l

q =T;:m =q

on

r1

(3)

on

r2

(4)

in

rl

Constraints
~::;</l::;<P

.e::;e ::;e

at x*

and

(5)
(6)

in which <I> = the state variable (ground - water head above datum); q ::: the groundwater flux (the in - flux (+; e ::: the decision variable (control variable, controllable sink

94
(+)/so~rce

term); 4>

(-)term); W = known sink (+)/source (-) term (uncontrollable sink/source

= the

prescribed ground - water head;

= transmissivi~;
requirement); $

q = the

prescribed ground - water flux; T

~ = the lower limit of the state (lower ground - water head


= upper limit of the state (upper ground - water head requirement); e_=
of the decision variable (lower limit of the controllable sink / source); e

the lower limit


upper limit of the decision variable (upper limit of the controllable sink / source);

= the

a = the

whole domain;

rl

r2

= the boundaries; Or = subdomain associated

with the state-constraints; x = point fitted for the controllable load; and
boundary associated with the constraints.

@ : Point

G = sub

Filled for ConlJOliable Sink I Source (Decision Variable)

0 : Point Filled for Uncontrollable Sink / Source


6.: Regulated Point in Ground Water Head

Fig. 1 Confined Aquifer Divided into Boundary Elements


The objective function may be composed of the state variables (ground - water head
distribution) {4>}, the ground - water fl ux distribution (q) and the decision variables
(controllable sinks / sources) (e) in general as shown in Eq. 1.

2.2. Discretization by the Boundary Element Method


In order to discretize the above mentioned differential systems, the boundary element
method (2, 3, 4) is used. By using the weighted residual method, the eqUilibrium
equations (Eqs. 2 - 4) are replaced by the following integral formulation.

95

...

{TV' <\>-6-WI<l> dfl +

,....

{<I>-<\T a~* dr+

_ *
J {q-ql<l>
dr=O

rl

The weighting function

<I>

(7)

rz

has continuous first derivatives. Later this function will also

be required to satisfy the governing differential equation (Eq. 2). Integrating the first
term of Eq. 7 by parts twice, we obtain:

TV'2<1>*<l>d fl

-1

(6+W)q,*dfl+

q<l>*dr+L q<l>*dr-L 4xI*dr-L cpq*dr=O

(8)

where

(9)
The weighting function is now taken as the fundamental solution to Laplace's equation.
The fundamental solution to two-dimentional Laplace's equation is given by
</1* =-LlrJ
21t r

(10)

where r is the distance from the point of application of the unit source to the point
under consideration. Although several types of boundary elements have been presented,
the constant boundary element method is used in this study (see Fig. 1). Therefore,
substitution of Eq. 10 into Eq. 8 yields the following Equation.

1tT4t+f i-'r2fi'<l>dr+f

T r. n<l>dr=f qlrJdr+f qlrJdr.


6im+Wslnlr2
r
r .1:\
r'n s:\
r~
Ii
n
Ii

(11)

6j means the controllable point sink I source at internal or boundary point T, and Ws
means the uncontrollable point sink I source at internal or boundary point's'. From now
on in this study we consider only such point sinks I sources. Then, the Eq. 11 for a
given node 'n' becomes in the following discretized form:
1tT<I>n +

~\

~\

~I

L hni$j = L gnjqj + L dni6j + Pn

Therefore, the whole set of the equilibrium equations for the N

(12)

= Nj + N2

nodes on the

whole boundary is expressed in the following matrix - vector form:


[H] (<I>n) = [G] { qn} + [D] (6i) + (Pn)

(13)

where

hnj=hnj

(n= j)

(14)

(n;ej)

(15)

. Eq. 13 is rewritten in such a way that all the unknowns are on the left hand side as

96

follows:
[A]{Xnl - [D]{a;} = {bn}
NxN
NxI

(16)

where (X n ) is the vector of unknown cP and q. Then, the matrix - vector forms of the
BE&LP method in the ground - water management are obtained as shown below:
Objective Function

(throughout the whole domain 0)


N

L.

Z = Opt. f({CPn},{ qn},{ ad) = opt. f({X n},{ aiD = Opt. <'L, anXn +
~iai)
{ad
{ail
{a;} n=1
i=1

(17)

Subject to:
Equilibrium Equations (N-Eqs.)
[A]{X n} - [D]{ ad = {bn }

(18)

Constraints ( (2L + 2K + 21) - Eqs.)

or,
(for interior points in Or)

(19)

(for interior points in Or)

(20)

or,

h<l>m} 2: {~m},

or,

[R]{ Xn) 2: {~m}

(for mid - nodes on

G)

(21)

h<Pm}

or,

[RJ{ Xn}

(for mid - nodes on

r';)

(22)

{ad 2: LBd

(for the points fitted for the controllable loads)

(23)

{G;} $; {e;}

(for the points fitted for the controllable loads)

(24)

$;

h<l>m},

$;

h<l>m}

Nonnegative Conditions
{ Xn} 2: 0, { e;} 2! 0

(25)

Therefore, the BE&LP method in the ground - water management is one that optimizes
the objective function (Eq. 17) under the conditions of the equilibrium equations (Eq.
18), the constraints (Eq. 19 - 24) and the nonegative conditions (Eq. 25),

97

3. Efficient Computational Algorithm. Finding of Initial Basic Feasible


Solution without Introduction of Artificial Variables
The discretized equation systems (Eqs. 17 - 25) include inequality constraints (Eqs. 1924). In developing of the solution procedures of the BE&LP method, we shall find it
much easier to work with equality constraints.
By converting the inequalities to the equalities through the use of slack variables {2.0J},

{AI}, {lite}, {Ilk}, (Yi}, and {Vi}, we obtain the following standard systems.

Objective Function
N

<L

Z = Opt.
unXn +
~i9i)
lei) n=l
i=l

(26)

Subject to:
Equilibrium Equations (N-Eqs.)
[A]{Xnl - [O]{9tl = {bnl

(27)

Constraints 2L + 2K + 21) - Eqs.)

[Ao]{Xnl - [OoHEld - {1zl:::: LEd

(28)

[Ao]{Xnl - [OoHEld + {Ad:::: (Pil

(29)

[R]{Xnl - (Ilk! = {~ml

(30)

[R]{Xnl + {!!kl = {k<l>m}

(31)

{eil- {Yd = L~d

(32)

(9d + {vil:::: (ed

(33)

Nonnegative Conditions

The obtained standard systems are solved by using the simplex method (5). Before the
use of the simplex method, however, it is necessary to find a solution which satisfies
Eqs. 27 - 34. Such a solution which satisfies the given conditions (Eqs. 27 - 34) is
called a basic feasible solution. The simplex method is always initiated with a program
whose equations are in canonical systems which apparently show a basic feasible
solution. In the case where a canonical form, or an initial basic feasible solution can
not be found, we have to use the two phases (Phase I and Phase II) of the simplex
method. Phase I is to find an initial basic feasible solution and Phase II is to find an
optimal feasible solution, if one exists. And Phase I of the simplex method augments

98
the systems to include a basic set of artificial variables. Therefore, Phase I of the
simplex method requires considerable computer time and memory because of the
introduction of the artificial variables.
In this research, however, by use of the special structures of the BE&LP method, an
efficient computational algorithm to find the initial basic feasible solution is developed
without the introduction of artificial variables. Therefore, the troublesome Phase I of the
simplex method becomes unnecessary. The efficient computational algorithm starts with
the standard systems (Eqs. 26 - 34). At first the equilibrium equations (Eq. 27) in the
standard systems are transformed by using Gaussian elimination as follows:
(35)
where
(36)

Substitution of Eq. 35 into the standard systems (Eqs. 26 - 34) yields the following
equations.
Objective Function

Z = Opt

{9d

n=1

i=1

(L <XoXo + L l3i 9i)

(37)

subject to:
Equilibrium Equations (N-Eqs.)
{Xn} - [D](9;} = {Bn}

(38)

Constraints 2L + 2K + 21) - Eqs.)

- L~tl = {]~; I

(39)

[D]{9d + {Ad = (13;)

(40)

[D]{9d - {~k} = Uh:l

(41)

[DJ{ ad + {ilk! = ( 13k)

(42)

{ad - {Yd = {.ed

(43)

{ad + (vd = led

(44)

[DJ{ 9il

Nonnegative Conditions

{X n } ~ 0, {9d ~ o. {Ad ~ o. {A.Ll 2! 0, {~k} 2! 0,


where

{ilk! 2! 0, {Yil 2! o. {vd 2! 0

(45)

99

_0

{B,J==fPd-[A]{Bn}, !.Hkl=={~mJ-[RJ{Bn), (Bkl==(k<I>m}-[R]{Bnl

(46)

Now let us consider a typical problem of optimal control in the ground - water
management with the following objective function and constraints.
Objecteve Function
I

J==Max.

fe J

L 9i

(47)

i=1

Constraints
(48)
(49)

Such a typical problem or its arrangements may be frequently encountered in various


practical distributed parameter control systems. In such a case by chosing {9 j ) to be
equal to {~d, we can easily find an initial basic feasible solution as follows:

(ed == (Bil, {Xn} == [D'HBil + {Bn}, {All == [6](~il - U~;}, (id == - [6]{~d + {B;},
(~k}==[D]{~d-(Jh), (!lkl==-[D](~d-{Bkl, {Yil=O, {vil={ed-{~il

(50)

Since {yd are equal to zero, {Yd become non-basic variables, and other non-zero
variables {Xn}, lei}, (AI), (AI), (~k), (~k) and {Vi} become basic variables in the
initial simplex tableau. Thus, the initial simplex tableau of the problem is obtained
without the introduction of artificial variables. Therefore, computer time and memory are
saved considerably.

4. Numerical Example
A numerical example is conducted as shown in Fig. 2. Computed results are shown in
Fig. 3.

5.

Conclusions

The boundary element and linear progrmming method (the BE&LP method) was studied
to solve control problem of steady - state confined aquifer management. In order to save
computer time and memory an efficient computational algorithm of the BE&LP method
was developed by noting the fact that the method has special structures. The proposed

100

=103 m2/day
W =0 (Uncontrollable Load)
T

q,m ~

9:J

91

o
(4,8)

o
(12,8)

~o
16m

(iFo r2 (Impermeable Wall)


)01

Max. Z = 91 + fu + 93

Fig.2 Input Data and Unknowns on Simple Model


of Confined Aquifer Management

q,1O=24.9 m q,9'=24.7 m q,s=24.7 m $7=24.9 m

q,1=24.9 m q,z=24.7 m $3=24.7 m $4=24.9 m


3
Optimal Objective Z

=91 + fu + 9)
=1500", 500 + 1500 =3500

Fig. 3 Results

= 24.0 (m=2,3,8,9)

rl q,m ~ ~ =24.9 (m=I,4,7,1O)


o=.eS 9 s e= 1500 m3/day

m3/day

101

efficient computational algorithm makes it possible to obtain the initial basic feasible
solution of the simplex method without the introduction of artificial variables and to
omit the troublesome Phase I of the simplex method.
Therefore, the proposed efficient algorithm allows us to solve large scale control
problems in ground - water management.
The tractability in both the boundary conditions and the equality or inequality constraints
makes sure that the BE&LP method becomes a powerful technique for many problems
of optimal control and design in various industries.

References
1. Aguado, E. and Remson, I. (1974), "Ground-Water Hydraulics in Aquifer Manage-

ment", Journal of the Hydraulics Division, ASCE, Vol. 100, No. HYI, pp. 103-118.
2. Bolteus, L. and Tullberg, O. (1980), "Boundary Element Method Applied to TwoDimentional Heat Conduction in Non-Homogeneous Media", Advances in
Engineering Software, Vol. 2, No.3, pp. 131-137.
3. Brebbia, C. A. (1978), "The Boundary Element Method for Engineers", Pentech
Press.
4. Brebbia, C. A. and Walker, S. (1980), "Boundary Element Techniques in
Engineering", Newnes-Butterworths.
.
5. Dantzing, B. G. (1963), "Linear Programming and Extensions", Princeton University
Press.
6. Futagami, T., Tamai, N. and Yatsuzuka, M. (1976), "FEM Coupled with LP for
Water Pollution Control", Journal of Hydraulics Division, ASCE, Vol. 102, HY 7,
pp. 881-897.
7. Futagami, T. (1976), "Several Mathematical Methods in Water Pollution Control THE FINITE ELEMENT & LINEAR PROGRAMMING METHOD and the
Related Methods", HIT-C-EH-l, Department of Civil Engineering, Hiroshima
Institute of Technology.
8. Futagami, T. (1981), "Boundary Element & Linear Programming Method in
Optimization of Partial Differential Systems", Boundary Element Methods,
3rd Internationl Seminar on Boundary Element Methods, Brebbia, C. A.,
ed,. Springer-Verlag, pp,457-471.
9. Futagami, T. (1982), "Boundary Element & Dynamic Programming Method in
Optimization of Transient Partial Differential Systems", Boundary Element Methods
in Engineering, Fourth International Seminar on Boundary Element Methods,
Brebbia, C. A. ed., pp. 58-71.
to. Futagami, T. Fukuhara, T. and Tomita, M. (1977), "Transient Finite Element
& Linear Programming Method in Environmental Systems Control-The Efficient
Computational Algorithm" Proceedings, IFAC Environmental Symposium, Akashi.
H., ed., Pergamon Press, pp. 143-150.
II. Futagami, T., Tzafestas, S. and Sunahara, Y. (Editors), "Distributed Parameter
Systems: Modelling and simulation", North Holland, 1989.

Dynamic Crack Contact Analysis by Boundary Integral


Equation Method
S. Hirose
Department of Civil Engineering, Okayama University,
Tsushima-Naka, Okayama 700, Japan

Summary
A crack analysis is usually carried out on the assumption that crack faces are free of
traction. A number of papers on the crack analysis in both frequency and time domains
have been published [lJ. In an actual situation, however, crack faces may not always
satisfy the assumption. Let us consider a fatigue crack as an example. Crack-opening
displacements across a fatigue crack are very small in some case. There is a possibility
of dynamic contact phenomena on the crack faces when the crack is subjected to the
incident wave with a large amplitude.
The present paper is concerned with the dynamic contact problem of a penny-shaped
crack subjected to a longitudinal incident wave. Since this problem is a boundary-type
nonlinear one, superposition is not applicable, A time-domain boundary integral equation
based on the traction formulation is used and solved by a step-by-step time marching
algorithm. Nonlinear contact conditions are satisfied at each time step. Scattered farfields are investigated as well as near-field solutions. A similar approach was applied to
the scattering problem of elastic waves by a nonlinear interface [2].
Problem
Let S be a penny-shaped crack in a three-dimensional infinite body D as shown in Fig. 1.
The domain D is a homogeneous, isotropic, linearly elastic solid. The crack S is assumed
to be in a static linear displacement field of the form

1r = 6i3?T(1- 2v)/8(1- v)-2i1.ox3/a

at the initial stage. Here, v is the Poisson's ratio, a is the radius of the crack, and 10
denotes the amplitude of the static displacement. The linear displacement field yields the
static crack-opening displacement given by

(1)
For time t

0, the crack is subjected to an incident wave as shown in Fig. 1. The incident

wave is a longitudinal plane wave defined by

(2)
where

CL

is the velocity of the longitudinal wave, T is a dimensionless period, Uo is

the amplitude of the incident wave, and

HO is the

Heaviside step function. The total

103

Figure 1: A penny-shaped crack subjected to an incident L-wave.


displacement field u can be expressed as

(3)
where 'Ii is the static displacement before the L-wave incidence,

u is the incident displace-

ment field and u' C is the scattered wave generated by the interaction of the incident wave
with the penny-shaped crack. Since there is no displacement discontinuity in the incident
field, i.e., [u]

= 0, the total crack-opening displacement may be written as


(4)

In a usual crack analysis, the crack faces are assumed to be free of traction. The boundary
conditions are then given by

(5)
where

aij

are stress components of the total wave field. In this paper, however, the

following contact conditions are considered.

i- 0,

[U3]

1.

0,

0'33

0'33

on St

0'13

0'23

on

(6)

where St and S .. represent a non-contact area and a contact area of the crack face, respectively.
As seen in eq.(7), we assume that there exists no friction in the tangential directions
on the crack. Also, the problem is concerned with the plane crack subjected to a normal

104

incidence of the longitudinal wave. Then variables in the normal direction are independent
of variables in the tangential directions. Therefore we can solve the problem by using the
boundary integral equations which relate
and

0"33

are abbreviated as [u] and

0",

[U3]

with 0"33 only. For simplicity, hereafter,

[U3]

respectively.

Boundary Integral Equations


Though the boundary conditions are nonlinear as seen in eq.(7), the domain D is linearly
elastic at all times. The following boundary integral equations based on the traction
formula can be used for solving the crack contact problem as in a linear problem [3].

-a(z,t) = p.f. f W(z-y,t)*[u(y,t)]dSy zESt> tE[O,oo)

O"(z, t) - a(z, t)

ls,

f W(;;r: ls,

y,t)

* [u(y,t)]dSy

z E Su, t E [0,00)

(7)
(8)

where a is the stress component related to the incident wave, the asterisk indicates the
convolution integral with respect to time, and p.f. denotes the finite part of the integral.
In eqs.(7) and (8), W is the hypersingular integral kernel defined as

W(;;r:,t)

()2

= -C33kIC33PQ-{)
() Gkp(;;r:,t).
XI Xq

(9)

Here Cijkl are elastic moduli and Gkp are fundamental solutions given by

Gkp(;;r:,t) =

41r~1;;r:1 {e~l~p -

OkP)

1~1 xo(t -

xl;;r:l)dx

[~O(t-l!l)-~o(t-l!l)] + OkPO(t_l::t)}
Izl2 ci
CL
c}
CT
c}
CT

+Xk Xp

(10)

where 0(') is the Dirac delta function, p is the mass density, and CT and CL are the velocities
of longitudinal and transverse waves, respectively.
Discretization of eqs.(7) and (8) yields a system of equations with [ul and
tenns. It depends on the contact conditions given by (7) whether [u] or

(J

0"

as unknown
becomes the

unknown value on a crack element. In this sense, the system of equations is nonlinear.
The numerical implementation will be discussed later.
After determining the crack-opening displacement by solving the boundary integral equations (7) and (8), the scattered wave field at an arbitrary point can be calculated by
substituting the solution [u] into the following equation.

un;;r:, t)

= ls,f C33PQ{)OYq Gkp(z -

y, t) * [u(y, t)]dSy

zED

(11)

105

1a!1 > Iyl can be expressed in a simple manner


by introducing the approximation of Ia! - yl : : : 1a!1- 0' y (0 = a!fla!D in eq.(ll). For
Particularly, the scattered wave at far-field

example, the longitudinal component

as follows.
ui,C(a!,t)

uLc

of the far-field scattered wave may be written

- IOL (z,t -1a!lfeL)


= _1
411" 1a!

(12)

where OL is the far-field amplitude defined by

(13)

Numerical Implementation
In the numerical analysis of the boundary integral equations (7) and (8), the crackopening displacement [u] is represented by a piecewise linear function in time and by the
constant elements for the spatial distribution except for the crack-tip elements. For the
crack-tip elements, the crack-opening displacement for the corresponding static problem
is introduced to express the behavior of ..;;. at the crack edge, where c is a distance from
the crack tip. The time axis is discretized into a sequence of small increment At, and the
crack face is divided into Q elements, i.e., S = L:~l Sq. Then the unknown crack-opening
displacement [ul may be written as

[u(y, t)

= L <pq(y) L 1]m(t)[u];;'
q=l

(14)

where [u];;' is the coefficient for the crack-opening displacement for the q-th element at
the time t = mAt, <pq denotes the spatial shape function for the q-th element, and

1]m

is

the interpolation function with respect to time given by

(15)
where i7m(t)

= (t -

mAt)fAtH(t - mAt).

Substituting eq.(14) into eqs.(7) and (8) and using the collocation technique yield the
following matrix form.

0] {[U]M } _ 1=1

WO
( WO
-1

CT M

M-l [

M-m

w M- m

where 1 is the unit matrix, the superscripts m and Min


steps. Also, W

M- m

{[u] } -

{(j-M}
(tM

fur

CT M

and

is the matrix having the components defined by

(16)

indicate the time

106

Compute the matrix components

W~

(m=O,l'."tMmd,~-l,
p, q, = 1,2, ... , Q)

Assemble the mahices

W~

and

Compute the R.H.S. and L.R.S. in eq.(I6)

Figure 2: Numerical algorithm.


where

2p

(p

1, ... , Q) are collocation points. Since the equation (16) is nonlinear,

several iterations are necessary to obtain the convergence of solutions in each time step.
The numerical algorithm is shown in Fig. 2.
Examples
In numerical calculations, the time increment cTAt/a was taken as 0.05 and the pennyshaped crack was divided into 452 quadrilateral elements. Taking account of the axisymmetric property of the problem considered here, only twelve field points were taken
as independent collocation points. This procedure drastically reduces the dimension of
the problem. For all calculations shown in the sequel, the Poisson's ratio is 0.25, the
nondimensional period T is 4, and the ratio of u.o/ft.o is 1.
Figure 3 shows the crack-opening displacements [U3] versus the distance r from the crack
center at several early time steps. The corresponding normal stresses are shown in Fig. 4.

107
1~--~-----'----------1

0.5

0.15
0.2,0.25

ria

0.5

Figure 3: Crack-opening displacements vs. r /a at several time steps.


It can be seen from these figures that the crack contact phenomena stem from the crack

= 0.1 and move to the center of the crack. After that, the entire
crack faces come into contact with each other at the time cT~t/a = 0.2.

edge at the time

cT~t/a

Figure 5 shows the time variation of the crack-opening displacement [u 3C] and the stress
component

0"33

at the nearly crack center of

ria = 0.05.

For comparison, the results

for a stress-free crack are shown in Fig. 6. Figure 5 shows that the stress component
has impulsive responses every time the contact phenomena are generated at [ujc]/uo

-0.9987. As seen in Fig. 6, on the other hand, the linear analysis for the stress-free crack

shows the smooth response in the crack-opening displacement.


Figures 7 and 8 show the far-field amplitudes

flL ,

defined by eq.(13), of the back-scattered

longitudinal wave as a function of time, corresponding to Figs. 5 and 6, respectively. Fig. 7


shows the stepwise behaviors, which may occur due to nonlinear contact phenomena on
the crack and include high frequency components.
Conclusion
The time-domain boundary integral equation was applied to the dynamic crack contact
problem. The problem is a boundary-type nonlinear one, which is suitable for the boundaryelement analysis. Numerical solutions in near- and far-fields show nonlinear responses,
which may include high frequency components called higher harmonics.

108

Figure 4: Normal stresses vs.

rIa at several time steps.

15

I
I

I
I
I

,I

,,
I

i,' aO"33

" fioJ..l.

,,
I

I ,

: I

:'

I'

-15L---------~--------~--------~

cLt/a

10

15

Figure 5: Time variation of the crack-opening displacement [U3J and the normal stress
at rIa = 0.05; crack with a contact condition.

0"33

109

15

-15LO--------~5--------~1~O--------~15

cLtja

Figure 6: Time variation of the crack-opening displacement


at ria = 0.05; crack with a stress-free condition.

[U3]

and the normal stress

0"33

20r---------~----------~---------.

-20~--------~----------~--------~

(CLt

-I

zl)ja

10

15

Figure 7: Far-field amplitude 11L of the back-scattered wave as a function of time; crack
with a contact condition.

110
20r---------~--------~--------,

-20~O--------~--------~1~O--------~15
5 (eLt -I:vl)/a
Figure 8: Far-field amplitude
with a stress-free condition.

~h

of the back-scattered wave as a function of time; crack

References
1. e.g., P. A. Martin and G. R. Wickham, Diffraction of elastic waves by a pennyshaped crack: analytical and numerical results, Proc. R. Soc. Lond. A 390 (1983)
91-129.

2. S. Hirose and M. Kitahara, BIE analysis of elastic wave scattering by a nonlinear interface, submitted to Review of Progress in Quantitative Nondestructive Evaluation
11, eds. D. O. Thompson and D. E. Chimenti, Plenum, New York (1992).
3. S. Hirose and J. D. Achenbach, Time-domain boundary element analysis of elastic
wave interaction with a crack, Int. J. Numer. Meth. Eng. 28 (1989) 629-644.

Current Status of the GENESIS Methodology for KnowledgeBased Treatment of Transonic Flows, with Emphasis on Shock
Fitting and Non-Dissipative Supression of Expansion Shocks
Barry Hunt and Arthur P. Adamson (retired)
GE Aircraft Engines, Cincinnati, Ohio 45215

Summary
The Boundary Integral Method ("Panel Method") is generalized [1,2,3.4] to nonlinear problems by
iteratively updating the normal-velocity boundary condition V: in an equivalent pseudo-Lap/ace (PL)
problem. This allows existing linear 3-D panel codes to be adapted to the solution of the compressible
Euler equations over realistic configurations, without the need for a computational mesh in the external
field. This paper gives an overview of the current status of this developing methodology. known as
GENESIS: GEneralized Nonlinear Extension of Surface Integral Schemes.
The direct (integral) link between cause and effect, intrinsic to integral methods in general, allows
GENESIS to couple knowledge-based or phenomenological techniques with the numerical computation.
The application of such a coupling to the problems of shockJitting and of expansion shock suppression is
described. A class of three "slightly rotational" fields is introduced, offering progressively closer
approximations to the compressible Euler equations. The first is a variant of the linear Prandtl--Glauert
field; the other two are nonlinear, and offer new insights into shock phenomena. GENESIS uses a
semi-analytic approach to solve these fields. While sufficiently accurate in their own right for many
purposes, these "Slightly rotational" solutions can be further corrected by using the technique introduced
in [1]. to yield accurate solutions of the Euler equations, but using only a fraction of the computational and
human resources required by the current generation of Computational Fluid Dynamics (CFD) codes.
1 The GENESIS flowfield representation
1.1 Reduction to boundary form
The GENESIS methodology [1,2,3,4] is aimed at compressible, rotational flows over three~imensional
aeronautical configurations in steady-state translational or rotatory motion. We confine our attention here
to the tw~imensional translational case. Nonetheless. we retain the usual three~imensional notation:

t. J. k. and the components of the velocity


J oriented into the paper. We denote the unit normal and

in a cartesian system x. Y. z, the unit vectors are respectively

V are still

U, V, W, but with V = 0 and with

tangential vectors by
For a velocity field

n and S ,with

V in a domain

s= Jx n. The corresponding velocity components are VII and V.

Q, with divergence (field source) VV and rotationality (vorticity)

VXV, GENESIS constructs a semi-analytic vector field F(x,y,z) as a function of V(x,y, z), such that:

VF

=l:(x,y,z) -

VV

and

VXF

=r(x,y,z) -

VXV

(1)

with F=0 at infinity in the case of an unbounded flow. The perturbation velocity field Vp induced by the
field distributions l: and r may then be replaced exactly by that induced by boundary distributions of
source

UF

and vorticity

iJF =(J)FJ on the body boundary S (augmented, for points lying on S or in the

interior of Q , by the local value of

F).

These equivalent boundary distributions are defined by:


(2)

112

Suppose the nonlinear flow has a required boundary condition VII = VnBC (zero for a solid, stationary
boundary). Then GENESIS reduces the problem to a sequence of pseudo-Laplace (PL) linear problems to
be solved for the tangential velocity V;, but now enforcing a boundary condition V,: dermed by:
V:

=V"",,+OF

At convergence, the required tangential velocity component V. of the nonlinear field


~=~-~

(3)

V is given by:
00

At each iterative cycle, the current estimate V. given by (4) is used to update the field vector F , which then
updates OF and WF using (2); the new 0, is then used in (3) to update V,: for the next cycle.
t.2 Extraneous constraints
In addition to the suppression of expansion shocks [4], two further constraints need to be injected into the
model to enable it to give a realistic flow simulation. Both of these involve the use of eigen-distributions
of source or vorticity, the mathematical relevance of which is discussed in Sec. 1.4.
1 .2.1 Vorticity constraint This is necessary for bodies with a sharp trailing edge - it is the familiar
KUlta condition. The smooth behavior observed in a real flow over an airfoil is the result of physical
features absent from the idealized model: the non-dissipative Euler equations do not support the creation
or diffusion of vorticity. In GENESIS, as with most kinds of boundary integral method, the
physically-observed behavior is enforced extraneously by requiring the solution to contain the correct
multiple of a circulatory (vorticity) eigen-distribution We' This multiple is the one which yields the
required smooth flow in the nonlinear field, i.e. after application of the tangential correction (4).
Fo' rounded bodies. such as an ellipse, different multiples of We yield different flow solutions, each
satisfying all the other constraints of the problem. None of these should be considered "incorrect";
solutions obtained with different multiples of We correspond to flows in which the "solid" surface of the
body has different "slipping" motions. A simple example is a circular cylinder with uniform rate of spin 0
about its axis. In a uniform cross-flow. this would generate an asymmetric flow with finite overall vorticity
(the circulation) and a lift force perpendicular to the cross-flow, both the circulation and lift being
proportional to 0 . The flow is symmetric only when 0 =0, corresponding to a zero multiple of we.
Without a specific constraint, the multiple of (J}e selected for a rounded body depends on the solution
algorithm employed. Forexample. a panel method using only boundary sources would necessarily always
have a zero multiple of We. Other algorithms may preferentially select a multiple dependent on
initialization values, or on the grid employed. Others yet may never converge. but simply drift around
without converging. or diverge. This offers a complete explanation of the bizarre results presented in [5].
1 .2.2 Gross source constraint The second constraint is particular to GENESIS. and ensures that the
gross amount G of "equivalent boundary source" OF in the model matches the gross amount of "field
source" in the domain g. In a physical system. the field source is V V. When g is the unbounded
domain exterior to a solid body. Gauss' theorem applied to V V tells us that the gross amount of source in
the physical system is zero. We therefore assume for GENESIS the gross-source constraint:

G=f1~~=o

(If we apply a correction to account for the difference between the Euler field source and its GENESIS
approximation. then this constraint (5) applies to the overall field.) The need for this gross-source
constraint is related to the existence of a source eigen-distribution 0", for a closed contour S . Different
multiples of Ob present in a converged GENESIS solution will lead to different flow solutions satisfying
all the other constraints of the problem; only the (unique) solution also satisfying (5) is physically

113

admissible. We may speculate. however. that modifications to this constraint will yield solution variations
analogous to those obtained with the many arbitrarily "conservative" finite-difference schemes.
1.3 Tbe Laplace-solver: SAVER
1 .3.1 Linear case: the A-field The Laplace-solver in GENESIS has the acronym SAVER (Source
And Vortex Evaluation by Relaxation) [3]. SAVER fixes a boundary source distribution u' using:
(6)
u' =-n Vo+ V;
is given by (3) with a prescribed UF. and Vo is the freestream velOCity. For an incompressible.

where V;
irrotational flow we choose F= O. so that. from (2). both UF and O)F are zero; for reference. we designate
the resulting ("Laplacian") field the A-field. and the corresponding tangential velocity solution VsA .

Relaxation Starting from some initial guess for an unknown boundary vorticity 0)'. the tangential
component ~.. of the perturbation velocity induced by a' and 0)' is computed by the usual
influence-matrix multiplication. The boundary vorticity 0)' is then relaxed towards V~. (The simplest
GENESIS scheme uses 0)' = V~. but generally a relaxation factor of about 1.3 seems to work
better.) The cycle is repeated some number of times. retaining the q' defined by (6). but updating 0)'. For
the A-field. the cycle is run to convergence (typically about 10 iterations for a lifting airfoil). SAVER has
the option to increase accuracy. if needed. by periodically applying a Jacobi update to the source q' to
eliminate numerical discretization errors in the normal velocity [4].
Kutta condition SAVER maintains the Kutta condition by periodically adding to 0)' the appropriate
multiple of a "unit circulatory" distribution O)e This vorticity distribution 0) e is itself pre-computed by
SAVER. by relaxing O)e to be equal to the tangential component of the perturbation velocity induced by
O)e itself(i.e with V. =0 and q' =0). In this preliminary computation. O)e is re-scaled at each iteration to
maintain a unit sum of the values of O)e at adjacent trailing points. The resulting "unit circulatory
distribution" O)e is of small magnitude except near the trailing-edge. Any multiple of O)e induces a
purely ci!Culatory external field (tangential at the body boundary). with

V= 0 in the interior of the body.

Discretization The SAVER discretization uses piecewise circular-arc panels carrying piecewisequadratic 0' and 0)'. A centered quadratic fit between midpoint values of 0' and 0)' is used. unless there
is a local discontinuity in these distributions. or in the geometry; in such cases the fit is biased forward or
backward. Exact analytic expressions for the vector influence coefficients are employed. This yields the
accuracy required to enable good solutions to be obtained with few panels (say. 40). for shocked airfoils.
1.3.2 Nonlinearadaptation Fornonlinearflows. F [and q' from (2). (3) and (6)] must be adjusted as
SAVER need not be run to convergence each time V; in (6) is updated.
[Typically. the overall number of iterations required changes little from that for the A-field.]

F follows the changing value v..

At convergence. the required component V. of the nonlinear field V is now given by (4). Note from (3)
that the boundary condition VnBC of the nonlinear flow is also recovered by subtracting UF from the
converged V; obtained by the SAVER algorithm in the PL flow. Note also that if the vector F issuchthat
the trailing-edge values of O)F obtained from (2) sum to a non-zero value e. then on the basis of (4) the
Kutta correction applied in SAVER must force the trailing-edge values of
to also sum to e .

v.:

The (pseudo-Laplace) field for a fIXed V,: is linear: if we scale the entire distribution UF [used in (3) to
derme V,.'] by a constant factor f. the incremental tangential velocity (V; - V.,A ) will also scale by f.
For a particular shape of the V; distribution. we can therefore write for every point on the boundary:

V; -

Vs~

= constant x aF

(7)

114
Analogously to the (Kutta) vorticity constraint, SAVER needs periodically to apply the gross-source
constraint (8) to Up. This control is necessary because of the existence ofa source eigen-distribution, any
multiple of which could otherwise pollute the solution.
1.4 Eigen-distributions of boundary singularities
When SAVER is converged, the fictitious velocity field throughout the interior of the body is equal to the
freestream velocity Vo;

more succinctly, the GENESIS I SAVER algorithms drive the interior

perturbation velocityfield to zero. However, an infinite number of different fields 1 exist, with 1 nonzero in the domain Q exterior to a boWldary S ,but with F =0 throughout the domain interior to S, and

for which vl 0 and V xl =0 throughout Q, with 1 = 0 at infinity (in the case of an WlboWlded
domain Q). Before examining two special cases of interest here. we first discuss the general case.
1.4.1 General case Consider the case where

is the (Laplacian) perturbation velocity field VLzp

induced, in any domain H with closed boWldary S, by any combination of sources, vortices, etc., all

external to H. Any such field has in the domain H the properties:

vl =0, V xl =0, 1 =0 at infinity.

It follows [1, 2,3,4] that the surface S carrying boWldary source uft,p
(J}ft,p = - j. Ii X VLzJ" where the unit nonnal

=- Ii Vl"P

Ii points outwards from

and boWldary vorticity

H, will induce throughout H a

velocity field identical to V Lzp Furthennore, it will induce a velocity field of precisely zero everywhere
outside H. Conversely, it follows that, in the presence of the originally stated set of external singularities,
boundary distributions u4. Ii.VLzp and (J}f:.p
Ii X VLzp will annihilate the original perturbation

=J.

field VLzp in the interior of H, but will themselves cause no perturbation at points exterior to H -

for

example in the vicinity of the original setof singularities. The velocity jump crossing the interface S outof
the domain H is, as may be expected, (J}f:.pjx Ii +ui:pli = VLzp. Since any multiple of this combination
uf:.p (J}i:p will induce zero perturbation outside H, we may reasonably tenn it an eigen-distribution.

1 .4.2 Vorticityeigen-distribution One special example of an eigen-distribution is the unit circulatory


distribution used to maintain the Kutta condition in SAVER. In this case, H is the domain exterior to a
boWldary S ,(i.e. the domain elsewhere designated Q), and the vorticity (J}Lzp = (J}e is assumed to lie on
the inner face of S. The source content ULzp

=Ii V lap

of this eigen-distribution is zero, since its

self-induced velocity field V Lzp is by definition tangential to S. The vector velocity induced by (J}e is
zero throughout the domain interior to S .
1 .4.3 Source eigen-distribution A second special example of an "eigeIHiistribution" is the unit
blowing distribution. This is a source distribution ULzp = Ub on the boWldary S, constructed, analogously
to the unit circulatory distribution, to be locally equal in magnitude to the normal component of the
velocity induced at S by that source distribution Ub. This can be obtained directly from (J}e: a source
distribution on the boWldary S. of precisely the same fonn as the unit circulatory vorticity distribution (J}e
described above, must also induce a velocity of zero throughout the domain interior to the boWldary S.
The unit blowing distribution is obtained by dividing this distribution by the total amoWltof source on S;
(8)
The vorticity content of this eigeIHiistribution is zero, since its self-induced velocity is nonnal to S .

115

2 Approximation of the Euler field source and vorticity


The field functions 1:(x.y. z) and r(x.y. z) used by GENESIS are approximations of the divergence V V
and vorticity V xV of the physical velocity vector V. In the case of the non-dissipative Euler equations,
the field vorticity V x V is precisely zero for any fluid particle which has not passed through a shock;
the exact vorticity field downstream of an Rankine-Hugoniot shock is described in [1]. The field source
distribution for the non-dissipative Euler equations is exactly VV =M2 ;)V.las, where av.las is the
flow acceleration along the local streamline, and M is the local Mach number, given in terms of the local
flowspeed V. and the local speed of sound a by:
M=V.
a

where

(9)

where y is the ratio of specific heats and Vm is the thermodynamically maximum-possible flowspeed,
given for a particular value of the freestream Mach number Mo by:
2 _
-

Vm

2
1 + -;::;--M'
v-I)

(10)

The original GENESIS approximations introduced in [I,2,3J make l:(x.y. z) and r(x.y.z) match the Euler
V V and V x V values exactly at the body boundary. but approximate their spatial variation W.f.t. z or n
(normal to S) as a weighted sum of some set of pre--defmed profile shapes; the computed V. will be in
error if the resulting composite profiles do not match the (unknown) physical profiles adequately.
We introduce here a variant of this "profile" approach. Instead of constructing an F field giving an
approximate solution of the exact equations. we defme an approximation of the Euler equations. then
construct an exact solution of the approximate problem. We defme a class of three slightly rotational
"analytic" fields which form progressive extensions to the (Laplacian) A-field of Sec 1 .3 .1 : the B-field
is a variant of the (linear) Prandtl-Glauert equation; the C- and D-tields are nonlinear, and closely
approximate the Euler field in most parts of g , even when transonic. GENESIS computes these fields
exactly, without introducing approximate profiles. Inspection of the solution offers a retrospective test of
the Validity of using the chosen analytic approximant; if required, the "profile" approach [1] can then be
applied to the error field, i.e. to the difference between the 1: and r required in an Euler field and those
implied by the "analytic" approximant.
2.1 "Analytic" approximation of the Euler compressible flow field
The "analytic" B-. C- and D-fields employ different representations of the Euler field source
1:. = M2 av./as. They all replace av./as by au/ax, Le. the horizontal derivative of the horizontal
velocity component, U; this is generally a good approximation except near a vertical surface, e.g. a
leading edge. The three fields differ in their approximation m of the local Mach number M [seeEq(9)],as

detailed below. Each field carries (physically spurious) vorticity defined by

r =m2 au/ az, hypothesized

to be small relative to 1: ; this hypothesiS can be tested retrospectively. For each field, it is simple to derive
an analytical expression for a vector field F from which the equivalent boundary functions Up and {J)p
are readily obtained. A shock (in the C- and D-fields) is represented explicitly as a discontinuity in Up and
{J)F: there is no need for an external "source sheet" to represent the shock.

116

2 .1.1 The B-tield "analytic" representation


The B-field approximates the local M by the (constant) freestream value Mo ; thus 1:8 =Mol auI ax and

r8 =MJ aulaz. We choose a unit horizontal free stream:


F8=F8i

satisfies V.F8

Vo =Ii . Thus a vector F8 , where:

withFB=MJ(U-l)

(11)

=1:8 and Vx FB =Jr8 ,with F8 =0 at infinity. Ifthe body is solid (Le. V.sc =0 ),then:

and observing that J. 11 x

U=~~

O~

i =n. , we obtain from (2) with F =FB :


(IF=-n,,F

and

(JIF=-nzF

(13)

Iteration cycle Two methods have been developed for the GENESIS iteration cycle for (I, and (JI,;
these are described below for the B-field. and their variants for the C- and D-fields outlined later.
Method Bl The simplest GENESIS iteration cycle consists of first initializing the estimate of the
"physical"

(e.g. to the Laplace solution

~A

), USing (12) to obtain U,

inserting this in (11) for FB,

then using (13) for (IF. This is integrated to form the gross source G, and the gross-source constraint (5) is
enforced by subtracting this multiple of the unit blowing distribution (8) from (IF:
(14)

Thecorrected (IF is put into (3) for V': , to be used in (6) for (I'. SAVER is cycled a number of times using
this (I' to update V;. then (4) is applied for the new estimate of ~. This cycle is iterated to convergence.
Method B2 An accelerated method, developed primarily for the C- and D-fields in transonic flow, also
works well for the B-field. Here, given the current V;, we write U, by elimination between the four
(linear) equations (12), (4), (13) and (11), in the following form (strictly valid only at convergence):
2M2
,
U_nz V.-nz 0
I-n?M6

(15)

This "accelerated" U is used in (11), then the resulting F8 used in (13) for (IF. The gross-source
constraint is applied, as for Method Bl, to give the corrected (IF. Eq (3) gives the accelerated update of
V,: to be used as the boundary condition for the next cycle of SAVER, giving an updated value of V; .
The cycle is then repeated to convergence.
We observe that at any point where n?M6= 1. the value of U defined by (15) becomes infinite, unless the
numerator of the expression is also zero at that point We may identify such points with "superinclined"
panels within the context of supersonic panel methods for the solution of the linearized supersonic
Prandtl-Glauert equation [6].
2 .1.2 The C-tield "analytic" representation
The C-field approximates M by m =Ulao =MoU with ao= liMo the freestream sonic speed; thus

::Ec =Mo2 u 2aulax and re =MJu2aulaz. We again use a free stream Vo =11. A vector Fe, where:

(16)

then satisfies the requirements

V'Fe =1:e

and

VXFe

=Jre, with Fe =0

at infinity.

117
Iteration cycle Method Cl is identical to Method BI, but with Fs replaced by Fe defined by (16).

v; .

Method C2 Analogously to Method B2. given the current


we here combine the four (nonlinear)
equations (12). (4). (13) and (16). to obtain the following cubic form (suictly valid only at convergence):

V'

=V -

n?Fe

=V

+ t3(1_V3)

where V'

=nz V:

and

t3

= tM6 n;

(17)

Note that the symbol V' introduced here is not equal to the x - component of the PL velocity vector
since the nonnal component

V;

V'.

is not zero. Note also that t3 is a different constant for each boundary

point. We thus require to solve the cubic equation (17) for the accelerated update U to be inserted in (16).
The remainder of the C-field algorithm is identical to Method B2, with Fs replaced by Fe .
Three roots of cubic Fig 1 shows a generic graphical representation of (17) for a particular value of
Mo nz ,Le. at one particular boundary point P where the vertical component of the unit nonnal is nz, ,for a
particular freestream Mach number Mo. Suppose that at some iteration of GENESIS the value of the PL
velocity at P is

v.p . from which (17) then gives a value Vj..

Fig I shows that if this were a converged

solution. there would be three candidate values of Vp, marked X,Yand Z in the figure. The point Z
represents a feasible reverse-flow solution, and will not be considered further here. The point X
represents a subsonic solution UsubP, while the point Y represents a supersonic solution UsUfUrp .

,
Jump location

V4:;iz

Fig. 2

Fig. 1. U' = V -

nlF

for one value of Mo nz Fig. 3

Location of Pr and PJ

Transonic velocity disuibution

We observe that for this Mo nz there is at point P a positive limiting value V' = ULP shown by Lj. in
Fig. I; this implies a feasible limit V;LP to the local
computed by GENESIS. The corresponding
positive turning-point value of Vp at Lj. is Up =Vrp+. Some simple algebra gives these values as:

v.p

VTP+ =Mo In,1 ;

v.

_ Vfp+

sLP -

Inzl

(18)

118

Let us define for each point P at the current iteration an exceedance ep defined by:
ep

= IVsp I - v,&

(19)

Considering a single point P :


If ep < 0, the local candidate solutions UswbP and UsuperP at P are both feasible.

If ep = 0 precisely, the local candidate solutions U.ubP and UsuperP at P coincide (point Lj.).
If ep> 0, there is nopositivesubsonicorsupersonicsolution Up of the cubicdefmed at P by (17).
Considering the entire set 0/points on the boundary:
If ep < 0 at every point on the boundary at the current iteration, then the flow is/ully subsonic.
If ep =0 anywhere, but with ep > 0 nowhere. then we have a/easible transonic solution.
If ep > 0 at any point on the boundary at the current iteration. the flow is an infeasible transonic flow.

Subsonic candidate solution GENESIS obtains the local subsonic candidate solution UsubP by using a
second-()roer Newton-Raphson (N-R) iteration, initialized either with a previous subsonic solution
UsubP. or with the value Up; if this results in an unsatisfactory N-R iteration. a smaller fraction of Up is
used for initialization. This N-R cycle does not usually need to be taken to convergence.
Supersonic candidate solution GENESIS obtains the local supersonic candidate solution UsuperP by
first computing UsubP then taking the positive root of the quadratic resulting from removal of the factor
(U - UsubP) from the cubic (17). With t3 defmed by (17) and Up by (17), we have:
where

(20)

Scaling of infeasible solution If any point P has an exceedance ep> 0 in (18). this implies that the
source distribution UF used in (3) to define the SAVER V; is "too strong". We therefore apply a uniform
scaling to UF, independently on the upper and lower surfaces of the body. Suppose the maximum value of
the exceedance on one of these surfaces occurs at a point PT' Then (7) indicates that the exceedance at PT
will become zero if we scale the distribution UF by a factor /;. , with initially k = 1 :

VLi>

/A: = IV.'

-IV..,v.1
v.' I

,p -

sAP

evaluated at PT ;

UF -

/A:UF

(21)

and then run SAVER again with V; re-<iefined by (3) with this scaled UF. The location PT of the
maximum exceedance will now have shifted. so this scaling cycle needs to be repeated until PT stabilizes.
The scaling factor / relative to the original UF distribution is thus the product of the successive /A: values:

/=/11213 ' ..

(22)
for SAVER. its output V: will be precisely that
required to force the subsonic candidate solution UsubP defined above to exactly match its turning-point
value Up = UTP+ at the (stabilized) point PT. The candidate solutions U.ubP and UsuperP will then be

If we now use the value /UF instead of UF to define

V;

identical at PT. which is now a sonic transition point All other points will have "feasible" values with two
different candidate solutions U.ubP and UsuperP. We thus have the basis for a/easible transonic solution.
subject to the reservation that we have achieved this at the expense of scaling up by the factor /.
Subsonic and supersonic regions The choice between the candidate solutions UswbP and UsuperP is
made initially by arbitrarily defining on each surface ajwnp point PJ, downstream of the point PT (see
Fig2). Points lying between PT and P, take their local candidate solution UsuperP, while the remaining

119

points take U.IlbP The resulting transonic U distn1rution inserted in (16) gives the new value of Fe for
the next cycle of GENESIS. Atconvergence, U and ~ transition smoothly from subsonic to supersonic
flow across PT, but jump discontinuously from supersonic to subsonic at P, (Fig 3).
Discontinuities in SAVER It is instructive to examine the behavior of the SAVER algorithm in the
presence of a V.: which is discontinuous at a point P,. By virtue of (6), this converts in SAVER to a
discontinuity in the source density (I' at P,. This induces, in both the exterior and interior domains, a
velocity with a locally singular tangential component, locally proportional to the logarithm of the distance
from P,. This velocity is uni-directional, with the fonn of a logarithmic "spike" at P, : it is "continuous
but locally infinite". The only way SAVER can induce a globally zero velocity field in the interior domain
in the PL problem is for the vorticity (1)' to adopt the same logarithmic-spike fonn at P,. In an infinitely
accurate implementation, the computed V; would display this same behavior. Now, SAVER computes
the velocity at panel midpoints; the immediate vicinity of a discontinuity at a panel comer is never
examined. The computed V; therefore inevitably becomes "discontinuous but finite". Consequently,on
the basis of (4), the computed shock jump in the physical velocity

is the sum of this discontinuity in

V; andthediscontinuityin(1)F implied by (13) and (16). Although a full discussion lies outside the scope
of the present paper, it seems plausible that a mathematical explanation of the characteristic shape of the
velocity distributions observed in the near vicinity of a shock lies in the interplay between singular edge
discontinuities, which are made explicit in this type of integral representation.
Jump strength The shockjump forthe C-field is close, but not identical, to the Rankine-Hugoniotjump
[1 J. GENESIS automatically computes the correct C-field jump. An analytic expression forthisjumpcan
readily be established in the special case where the shock is located at a point PJ where the surface has
n,,=O and liz = I andiscontinuous. (This condition is closely met over most ofa typical airfoil.) Inthis
case, we perceive from (13) that here (IF =0 and (1)F = F, so that the SAVER boundary condition V,: is
continuous across PJ. The computed V; will therefore be continuous, and the entire jump XY in Fig. I is
attributed to the jump in F. Some algebraic manipulation of (20) in this case gives the shock ratio:

r_

U2 = ![/t 2 - 3m
1]
with ml =MoUI (pseudo-Mach number)
(23)
UI
2
ml
where UI = VII is the shock-entry velocity, and U2 = V.2 the exit velocity (Fig. 2). We observe in this
special case that the value ml =13 yields Va =0 ,indicatingthat any higher value of ml would demand
taking the third root of the cubic (17) marlted by Z in Fig. 1; we can speculate that this represents a
shock-induced separation.
Note also on the basis of (16) that we can write for thejieid source: l:c = Mo2 u2aulax = aFlax. Ifwe
consider a shock as the limit of the continuous distribution :E, we can equate the change in U between the

entry and exit points to the shock source strength: U2- U I = l:cdx= Fcz-Fcl =

tMJ(ul-ul).

Some simple algebra then leads to the same relation (23) for the shock ratio.
Shock location iteration The Scaling of infeasible solutions and the automatic jump modeling discussed
above ensure the absence of an expansion shock at the transition point PT, and the presence of a
compression shock jump at P, (Le. surface coordinate s =s,), with a ratio (23) consistent with the
C-field. This is achieved at the expense of a scaling factor f s: 1 defined by (21),(22). This factor will
only take thevalue f = I, on each surface. when the arbitrary locations P, guessed forthe shocks are both
correct. The shock relaxation process proposed for GENESIS consists of using successive locations

s,

120

and their corresponding values I to establish a sensitivity derivative al/asJ for each I, SJ pair, and using a
Newtofi-Raphson iteration on SJ to drive both scaling factors to 1= 1 . The details of this iteration are
currently under development. The current GENESIS implementation (for a single-shock lifting airfoil, or
fora nOfi-lifting symmetric body) successfully employs an inverse scheme: the location PJ is prescribed,
and an iteration determines the freestream Mach number Mo which will give that location.
2 .1.3 The D-field "analytic" representation

The D-field approximates the local Mach number M by m = U/ aD , where [by analogy with (9)1

aZ ="; 1 (V~ -lP) ,with V~ defined by (to). Thus


2
lP av
1:D =-- - - - -

,,-1 V,;-lP ax

and

FD,

where:

It can be shown that the vector function


with

Fo(U)

lP

=_2_ [V~ In(V~-1 V.. + U) _


1'-1

au

r D =-----,,-1 V,;-lP az

V.. + 1 V,.-U

(24)

(U _ I)]

(25)

satisfies the requirements V'Fo =1:0 and Vx Fo =jro with FD =0 atinfinity. The D-field algorithm
is analogous to that for the C-field, with some minor differences. A nonlinear equation again has to be
solved for U, given U' . Instead of (17). we now have to solve the transcendental equation:
V' = V -

n;FD

with, again, V' = n, V;

(26)

The generic graphical representation of (26) over the relevant range is very close to that of (16) for the
C-fteld(Fig.l},exceptthatthe U transition value UTP+ now implies a slightly smaller limit value of U' :

.= j 1 +V~:'1 n,' ;

VTP

Vu = Vi".
In)

(27)

The subsonic candidate solution U,ubP is obtained exactly as for the C-field, but now the supersonic root
UsvperP also has to be obtained by aN-R iteration, initialized ala value greater than UTP+ A satisfactory
process is to first obtain U.uperP for the C-field, then to use this as initial value in the N-R iteration for
U.vperP for the D-field. The shock jump ratio can also be obtained by a N-R iteration on a related
transcendental equation, but once again this is unnecessary; GENESIS computes it automatically.
ACKNOWLEDGMENT The authors gratefully acknowledge thecOilstructive comments, relating to some of the
speculative COilCepts introduced in this paper, made by Jim Keith, Frank Parsons and Ron Plybon at GE Aircraft
Engines, and also by Carson Yates (NASA LaRC), and David Nixon (Nielsen Engineering & Research Inc.).

References
[1] HUlIJ, B., Plybon, R.C.: Generalization of the bowtdary-integral method to nonlinear problems of compressible fluid flow:
The no-mesh alternative. ISBEM..<J9, East Hartford, Conn . Oct. 1989.
[2] HUlIJ, B.: Generalized nonlinear boundary integral methods for compressible, viscous flows over arbitrary bodies:
Knowledge-based CFD. Japan/USA Boundary Elements Symposium. Palo Alto, CA. June 1990.
[3] Hunt, B.: GENESIS: A mes&-free, knowledge-based, nonlinear Bl. methodology for compressible, viscous flows over
arbitrary bodies: Theoretical framework and basic physical principles. IABEM-90. University of Rome, Italy, Oct. 1990.
[4] HUlIJ, B.: Knowledge-based, nonlinear bowtdary-integral models of COMp'essibie, viscous flows over arbitrary bodies:
taking C.F.D. back to basics. SecD7ld Pan-American Congress ofApplied Mechanics(pACAM II), Valparaiso, Chile, January
2 - 4, 1991. Also in expanded form in Applied Mechanics Reviews Supplement Mechanics Ptm-Amuica, October. 1991.
[5] Pulliam T.H.: Computational challenge: Euler solution for ellipses. AIM Journall8(10) (1990).
[6] Hunt B. Hewitt B.L.: The indirect boundary integral formulation for elliptic, hyperbolic, and nonlinear fluid flows.
DevelopntUlls in Boundary E1eme:nt Methods - 4, PK Banerjee. JO Watson (eels), Elsevier Applied Science Publishers.
Barking. UK, 1986, 227-339.

A Boundary Element Analysis of Space Charge Fields


H. Igarashi, M. Kuroda and T. Horuna
Department of Electrical Engineering, Faculty of Engineering,
Hokkaido University
Kita-13, Nishi-B, Kita-ku, Sapporo, 060 Japan
ABSTRACT
This paper presents a boundary element analysis of space charge fields which
are governed by the Poisson equation and the equation of current continuity. The
equation of current continuity is solved by the method of characteristics ( MOC ),
in which the charge density is calculated from the values of the electric field on
characteristic lines. On the other hand, BEM is used to solve the Poisson equation
because, in general, space charge problems include open boundaries. In MOC, in
order to reduce computing times, the electric fields on characteristic lines are
linearly interpolated in cells in solution region. The above mentioned solution
procedure is carried out by turns until a self-consistent solution is obtained.
INTRODUCTION
Space charge fields, which are governed by the Poisson equation and the
equation of current continuity, appear in high voltage (HV) DC transmission [1],
electric precipitation [2] and corona discharge in xerographic processes [3] and so
on. In space charge fields, since electric fields and space charges strongly affect
each other, the distribution of them must be determined simultaneously.
Therefore, the space charge field has been analyzed by numerically solving the
above equations. Lean et al. have analyzed the space charge field in a charge
transport channel on the basis of combination of BEM with method of
characteristics (MOC) [4]. Authors have solved the space charge field around a
HYDC transmission line by iteratively using BEM and MOC [5].
On the other hand, recently, a number of experimental and theoretical studies
on corona devices for xerographic processes have been reported [e.g. 6-B]. A typical
structure of a corona device is illustrated in Fig. 1. As shown in this figure, a high
voltage is applied between the thin discharge wire and the shielding plate. Corona
discharge is caused by the nonuniform electric field in the vicinity of the wire and
then ions are emitted from the wire, and consequently, they electrify a
photoconductor (PC) drum or paper moving in the transverse direction under the
corona device. The corona device is used for charging of PC, toner transfer from

122

t;" /

shield

.:.::...........:--

space
charge
field

PC or paper

..... ions

Fig. 1 A space charge field in a corona device


PC to a paper and cleaning of PC. The corona device, therefore, plays an important
role in xerographic processes. However, few studies on computational methods for
the analysis of the corona device have been reported. The main purpose of this
paper is to show reliability and performance of the combination ofBEM with MOC
for the analysis of corona devices.
The reminder of this paper is organized as follows. In the next section, BEM and
MOC for space charge fields are formulated. Then, the solution procedure is
described in some detail and two ways of electric field calculation for the BEMMOC combination are presented. In addition, the conventional method based on
Deutsch's assumption is briefly explained. Next, the accuracy and computing time
of the proposed BEM-MOC combination and Deutsch's method are investigated.
Conclusions are given in the last section.
FORMULATION
The governing equations of unipolar ion flow fields are given by
P
v2<1>= --,

(1)

ap
- +vJ=O.

(2)

J=ppE,

(3)

&0

at

where ~ is the electric potential, E is the intensity of electric fields, p is the charge
density, EO is the permittivity of free space, J is the current density, and 11 is the
mobility ofions. The substitution of eqs.(l) and (3) into eq.(2) yields

123
ap
p2
-+VVP+Il-=O.

at

eo

(4)

where v is the velocity of ions defined by pE. On characteristic lines ( lines of


electric force ), a solution to eq.(4) can be expressed in the form

(5)
where Po is the charge density on the discharge wire and t is the time required to
move along a characteristic line from a field point to the surface of the wire
electrode. The method in which a physical quantity is determined from an
equation on the characteristic line like (5) is called MOC.
The electric potential 4> depends on the charge density P as explicitly shown in
eq. (1) and P depends on 4> as implicitly shown in eq. (5). Hence, the space charge
field must be self-consistently determined using eqs. (1) and (5). In this paper, eq.
(1) is solved by BEM and the electric field E in the domain, which is used in MOC,
is calculated as
E

=-

Ij Jrj q VG cis + Ij Jrj ell v( aG)


cis - !. I J V (pG) dS
an
eo k Ok

(6)

where q is the flux on the boundary r of the solution region Q and G is the
fundamental solution to the two-dimensional Laplace equation. On the other
hand, in MOC, the times t for all nodal points is computed as follows:
t=

~ l.I(IlIE.I}.
L..

(7)

where Ii is the length of the segments of the characteristic lines which is adjusted
to accurately calculate t.
SOLUTION PROCEDURE
At the beginning of the solution procedure, the electric field Eo is obtained by
solving Laplace equation using BEM. Then, the charge densities at all nodal
points in the solution region are calculated by MOC, that is, virtual ions are
moved along lines of electric force at the velocity -pEo from nodal points to the
surface of the wire and the charge densities PI are calculated by inserting the
moving times (7) in eq. (5). After this MOC process, the Poisson equation (1) with
the charge density PI is solved again by BEM. Those solution procedures, which
are shown in Fig. 2, are repeated until a self-consistent solution is obtained.

124

Initial set up

Tracing along the characteristic lines backward in


time from nodes to the wire (MOC)

r"0W = rold + IlEold Ati

Calculation of Pk at nodes

p. = {llpo+ (llIco)!:At)-'

NO

Fig2. Flow diagram


CALCULATION OF ELECTRIC FIELDS

As mentioned in the previous section, the electric field calculations must be


carried out on all characteristic lines which start from nodal points in the solution
region. Since they are performed in all iterative steps, they may be the most timeconsuming work in the present method. In this paper, the following two types of
method for the electric field calculation are considered.
(a)Direct method
The electric fields are computed on the characteristic lines directly from eq.
(6). Although this method may be accurate, the calculation of the electric
field, which contains computing of G and aGlan with respect to all cells and
boundary elements, must be repeated a large number of times, i.e., the
number of calculation points on a characteristic line X the number of the
characteristic lines (=nodal points in the solution region) X the number of
iterative steps.
(b)Interpolation Method
In this method, the functions G and aGlan are calculated only for nodal points
in the solution region before the iterative procedure, and then, on the

125

characteristic lines, the electric fields are interpolated in cells using those
nodal values. Since this method does not need the calculation of G and aGlan
at arbitrary points, the electric field can be obtained just only by multiplying
the field quantities <p, q and p by the stored values of G and aGlan. The results
of this method may strongly depend on the discretization of the solution
method while this method may enable us to efficiently calculate the electric
field.
DEUTSCH'S ASSUMPTION [9]
In theoretical analyses of space charge fields, Deutsch's assumption, in which
the strength of electric fields is affected by space charges while the direction of
them is not, has been often adopted. In accordance with Deutsch's assumption, the
electric field E with space charges can be written in terms of the vacuum field Ev
as follows :

--J -

Ec)2 2Ec Po ,I ds }112


+
E
E E l : olE I
u'

E= { ( -

uo

VOO

(8)

IJ

where the integration is performed on the characteristic line of vacuum fields, l is


the length of the line, Ec and Euo are the electric fields on the wire with and
without space charges, respectively. The field calculation based on this
assu~ption can be carried out fast compared with the numerical method proposed
in this paper. The accuracy and computing time of those two methods will be
compared in the next section.
NUMERICAL RESULTS
A numerical model for a corona device is shown in Fig.3. As shown in this
figure, the boundary conditions are chosen as

tP=tPo'

tP=o.

p::: Po'

(on the wire surface) ,


(on the shield and plate).

(9)

In the model, the symmetry of right and left is assumed and the plane of
symmetry is shown by a broken line in Fig. 3. In addition, this model is
anti symmetric with respect to top and bottom since the plate is grounded. Taking
account of these types of symmetry, the following fundamental solution G is used
inBEM.

126

plane of symmetry
I ~
I

:
:

$=$0
Y

W
wire (x o' yo)
radius a

V
!

$=0

I
I

j ----

shield

cells

d ~t
h

..

plate

O~:~:__~~~~~t~/~~$=_o~
Fig. 3 A model for a corona device
1
[(x-xi + (y+yi][(x+xl + (y+yll
G(x,y;x.,y.) = -log
,
,
,
'.

I
4n
[(x-xi + (y-y.)2J[(x+xi + (y-yiJ
'"
,

(10)

Figure 4 shows the resultant characteristic lines for the space charge field (solid
lines) and vacuum field (broken lines). In this calculation, the geometric
parameters are chosen as W=H=11.176mm, h=16.764mm, d=1.0mm, (x o'
Yo)=(O, 16.764mm), a=88.9pm, and the applied voltage $0 is taken to be 5.0kV.
The solution procedure converges after several iterations under the condition that
the relative changes of the charge densities at all nodal points are smaller than 1
%. In Fig. 4, we can see that trajectories of ions in the space charge field are quite
different from those in vacuum. Therefore, in this case, the simplification based on
Deutsch's assumption seems to be inappropriate. Figure 5 shows the profiles of the
y-component of the electric field on the plate, which are calculated by the present
method and the method based on Deutsch's assumption. (The vacuum solution is
also shown in this figure for reference.) There are clear discrepancies between the
two numerical solutions.
Table 1 shows the computing times and the ratio of the current flowing into the
plate to the total current, for the methods (a) and (b), the Deutsch's method and
the experiment [6]. From this result, we conclude that, from practical point of
view, the BEM - MOe combination with the electric field calculation based on the
interpolation is most useful for the analysis of space charge fields in the corona
devices.

127

Fig. 4 Lines of electric force in the space charge field ( solid


line) and vacuum field (broken lines).

Deutoch'. _hod

o BEM +MOC

r;

2.

."

-.;

""Ii...
II

i>i

1.0

10

20

30

hnm)

Fig.5 Profiles of y-component of the electric field on the plate

128

Table 1 The ratio of the current flowing into the shield to total current and
computing times.
Direct
method

Interpolation
method

Deutsch's
assumption

Experiment[6J

Shield
current(%)

80.4

83.7

72.7

80

Computing
time (sec)

1457

60

16

--

CONCLUSIONS
A numerical method using BEM and MOC for analyzing space charge fields has
been proposed. The accuracy and computing time of the present method have been
compared with Deutsch's method. It is concluded that, when calculating the
electric field by the interpolation in cells, we can efficiently analyze the space
charge field in sufficiently good accuracy by the proposed method.
REFERENCES
[1] J. L Davis and J. F. Hoburg, "HVDC transmission line computations using
finite element and characteristics methods", J. Electrostat, 18, pp.I-22, 1986.
[2) P. L. Levin and J. F. Hoburg, "Donor cell-finite element descriptions of wireduct precipitator fields, charges, and efficiencies", IEEE Trans. Industry
applications, 26, pp. 662-670, 1990.
[3] D. M. Burland and L. B. Schein, "Physics of electrophotography", Physics
Today, pp.46-53, May, 1986.
[4] M. H. Lean and G. A. Domoto, "Charge transport in viscous vortex flows", J.
Appl. Phys., 61, pp.3931-3933, 1987.
[5] H. Igarashi and T. Honma, "Analysis of ion flow fields by combination ofBEM
with MOC", Proceedings of the third Japan-China symposium on boundary
element method, Pergamon Press, Oxford, pp. 351-360, 1990.
[6] F. W. Hudson andJ. E. Cranch, ''The effect of shield design on positive corona
charging currents", IEEE Trans. Industry applications, 13, pp. 366-369,1977.
[7] K. Tateishi and Y. Hoshino, "Corona current control by mesh grid", IEEE
Trans. Industry applications, 17, pp. 606-609, 1981.
[8) M. Omodani, T. Tanaka and Y. Hoshino, "Ion flow control characteristics for
high quality continuous-tone printing", J. Phys., D, 20, pp.1224-1229, 1987.
[9] W. Deutsch, "Uber die Dichteverteilung Unipolarer Ionestrome", Annalen
der Physik, 5, pp.588-612, 1933.

A Time-Stepping Boundary Element Method Applied to


Transient Thermoelasticity
Shuji ISHIGURO" and Masataka TANAKA""
Nagano Prefectural Research Institute for Inforlation Technology
Matsuloto 399, Japan
Dept. of Mech. Systels Eng Faculty of Engineering. Shinshu University
Nagano 380. Japan

ABSTRACT
This paper is concerned with a new boundary elelent approach based on the
tile-stepping schele for two-dilensional therloelasticity in unsteady heat
conduction states. The tile derivative in the resulting differential
equations is approxilated by the tile-stepping schele. The reduced
differential equations are transforled into a set of boundary integral
equations by using the exact fundalental solutions which can be derived by
the Horlander lethod. The lathelatical forlulation is presented in detail.
The proposed solution procedure is applied to sOle exalPle problels. whereby
its potential usefulness is delonstrated.
1. INTRODUCT ION
The boundary elelent techniques applicable to the tile-dependent problems
including transient therloelasticity can be classified as follows:
(1) Transfor. lethod using the Laplace or Fourier transforl.
(2) Direct lethod forlulated in terls of the tile-dependent fundamental
solution. and
(3) TI.e-Steppln. let hod based on a tile-stepping approxilation of the time
derivatives.
The Transfor. lethod solves the problel in the transforled dOlain by the
standard boundary elelent lethod. and then obtains the tile dependent
solution by the inverse Laplace or Fourier transforl. Difficulties lay arise
in this inverse transforl. Recently. however. FFT techniques have been
successfully applied to the inverse transforl. Therefore. it can be seen
that the usefulness of the Transforl Method is now re-evaluated. An
application of this lethod to the transient therloelasticity is presented in
Suh and Tosaka l . In the Direct lethod. the governing differential equations
are transforled into a set of boundary integral equations in space and tile.

130

which are solved b introducing tile elelents as well as boundar elelents.


The Direct Method, therefore, can provide directl the tile-dependent
solution. This let hod can be applied successfull to the problels in which
the exact fundalental solutions are known or can be easil treated in
nUierlcal ilPlelentation. However, in genera\' it is ver difficult to
obtain the exact fundalental solutions for lan tile-dependent problels
including full coupled problels. Until now, this lethod has been applied
onl to uncoupled problels 2 - 4
The Tile-Stepping lethod can also treat directl the tile-dependent
quantities as well as the Direct Method. Hence, if this advantage can be
full used, the Tile-Stepping Method la becole an effective lethod for the
tile-dependent problels, the fundalental solutions of which are difficult to
obtain. Recentl, this lethod has been applied to transient heat conduction
problels 5 and to elastodnalic proble.s 6 However, there is onl a
theoretical investigation7 for coupled dnalic therloelasticit.
This paper is concerned with a new boundar V elelent approach based on the
Tile-Stepping Method for two-dilensional uncoupled therloelasticit in
transient heat conduction states. The lathelatical forlulation is presented
in detail. NUlerical results are also shown to delonstrate the effectiveness
of the proposed lethod.

2. BOUNDARY INTEGRAL FORMULATION AND ITS IMPLEMENTATION


2.1 Fundalental solution
We shall consider two-dilensional uncoupled therloelasticit in unsteadv
heat conduction states. Under the assulPtion of no bod force and no heat
source, the governing differential equations of the displacelent and
telperature fields are written as follows:
(I,J=1,2.3)

(1)

where UJ(x,t) denotes the phsical variables at a spatial point of x at


at tile t, i. e. Ul and U2 are the displac8lent cOIPonents in Xl and
X2, respectivel. and Ua is the t8lperature. Throughout this paper, the
sUllation convention on repeated indices is used. Denoting bv the sVlbol 0 I
differentiation with respect to a spatial co-ordinate XI and bv Ot the
tile derivative. we can express the linear differential operator of Eq.(I)
as follows:

131

(2)

where
()I=()/()XI

~=()12+()22

,()t=()/()t

(3)

(4)

7=G/(l-21i)

P=27(1+Ii)a

In Eq.(2) and (4) G is the shear lodulus,

II

Poisson's Ratio. a the

coefficient of linear therlal expansion and ka the therlal diffusivit


defined as k0=il/ (caP)
density of

Now, we shall denote b

CII

denotes the specific heat, P the

the infinitesi.al tile interval, and approxilate

~t

the tile derivative at tile


() t U 3

where

and il the thenal conductivit.

laSS

t+~t

t + Ll t )

(x,

=.

as follows:
U3

(x. t + Ll

t) - U 3

Ll t

(x,

t)

(5)

Then, Eq.(l) can be reduced to the following for.:


L JJ
where

U J

[L IJ]

( x. t+Llt) +BI=O

[L J J] =

[B J] =
Using the

[B J]

and

IT

(6)

are as follows:

.,'+Gt.

7 ()1()2

7()1()2

7 ()22+ G~

G'(X.t)/LlJ

H~rlander

(J,J:::1,2.3)

lethod 8

9,

(7)

(8)

we can derive the fundalental solution. The

procedure of this lethod is sUI.arized as follows:


1) Express the adjoint operator J..JJ of the differential operator LJJ,

i. e.
7 ()~ +G~

[ J..I J] =

[ 7 ()

P ()

I ()

(9)

'2) Calculate the deteninant J.. 01 the adjoint operator J..IJ; in this case

132

.t=G (r+G) 1::.1::. (klll::.-1/Llt)

(10)

3) Obtain the two-point function .p(x.!.!) which satisfies the differential


equation:
.t.p' (x, Y)

+ 0 (x,

y)

=0

(1 I)

where 0 (x,!.!) is the Dirac delta function. It is interesting to note that


the deterlinant .t given b Eq.(lO) is the product of the biharlonic
operator ~ and the lodified Hellholtz operator (ka~-l/~t) and constant
G(r+ G). Hence, we introduce a new function O'(x,!.!) defined b the
following equation:
0"

(x,

y)

= G (r + G) I::. I::. .p' (x, y)

(12)

We can express Eq.(ll) in the fori:


( kill::. - 1 / Ll t ) 0"

(x, y)

+0

(x,

y)

=0

(]3)

The solution of Eq.(13) is given blll:


(14)
with constants CI and C2. In Eq.(14), r is the distance between two
points x and y. IIl(r) the zero-order lodified Bessel function of first
kind, Ke(r) the zero-order lodified Bessel function of second kind.
Finall. we can obtain
iJ'(x,

y)

kllLlt 2
21CG(r+G)

(Ke(./kIlLlt)+
(C-l) (2./ k:21t ) 2+

[1 + (2./ kllLlt ) 2] In 2./ kllLlt +C)

where C is the Euler constant.


4) Calculate the transposed cofactoral differential operator
adjoint operator .tlJ. which can be defined b

(15)

~lJ

of the
(16)

(I.J.K=1.2.3)

where OIK is the Kronecker delta. The operator


follows:

~lJ

can be given as

133

rke02

[MIJ]

~+Gke~~-r02 /At-G~/At

-rkBOI02~+rOI02/At

-fJGOI~

rkB012~+Gkll~~-r O//At-G~/At

o
o

-fJG02~

G (r+G)

- r k

0 2~ + r 0 I 02/ At

(17)

5) Calculate the funda.ental solution


using the following relation:
u j J (x, y)

= M I J 1>'

uiJ(x.~)

(x. y)

which can be derived by

O.J=1.2.3)

(18)

The detailed expressions of U jJ(x,~) are presented in Appendix 1.


2.2 Integral Equation For.ulation
We can start the integral equation for.ulation using the following identity
which can be obtained by integrating Eq.(6) over the whole do.ain .ultiplied
with the fundnental solution
$s (LIJUJ (x,

UjK(X,~),

that is.

t+At) +BI] UiK (x, y) dS (x) =0


(I.J,K=I.2.3)

(19)

Applying the divergence theore. to Eq.(19). we can finally obtain the


boundary integral equation expressed as follows:
CUJ(Y.

=-IrUI

Ir P I

t+At)
t+At) piJ (x .y) dr (x) +

(x.

(x,

$SA\U3 (x,

+ At)

j J (x. y) d

t) U3J (x,

y)

(x)

dS (x)

+
O.J=I.2.3)

(20)

where P I (x. t+,dt) 0=1. 2) is the traction co.ponent. p a (x. t+,dt) the heat
flux.

piJ(x.~)

the funda.ental solution concerning tractions and heat flux

which can be derived frol the funda.ental solution

ujJ(x,~).

The detailed

expressions of pjJ(x.~) are presented in Apppendix 2. In Eq.(20) C denotes


the shape coefficient. which equals 1.0 when the source point y is located
inside the donin S. and equals 0.5 when the source point y on a slooth
boundary.

134

2.3 Boundar~ Elelent Anal~sis


The standard boundar~ elelent let hod can be applied to nUlerical cOIPutation
of Eq.(20) to obtain the results at tile t+dt. Tile integration can be lade
in the following wu. We first calculate UI(X,t+dt) and PI(X,t+dt)
0=1.2,3) at tile t+.1t using the initial condition for U3(X,t) at tile t.
Then, we proceed to calculation at t+2dt using the obtained U3(X.t+dt) as
the pseudo initial condition. In this wa~, cOIPutation is advanced until the
final stage in tile is real ized. It should be noted here that in this lethod
the internal dOlain lust be discretized into cells. Although this lay result
in increase in cOIPutation tile, cOIPutation can be carried out ver~
effectivel~ if the sale At is used for all tile steps.

3. NUMERICAL RESULTS AND DISCUSSION


Therlal stresses occurring in a square plate and a solid c~linder are
cOIPuted b~ the proposed tile-stepping boundary elelent lethod. The present
boundar~ solutions are cOIPared with the exact ones available in the
literature.
Figure 1 shows the boundar~ elelent lesh and the internal cell subdivision
used in the anal~sis of the square region, where the whole boundar~ is
divived into 40 equal linear elelents and the dOlain into 100 equal linear
rectangu lar cells. I t is assuled that the telPerature at the initial state
is IOO'C and that the upper and lower sides of the plate are always kept at

Exact

(]

[>

120

~ 100

Pl =0

--.&

d>

Lb=OC

..

I...

100

Terrperatu-e
Stress o"xx

00- 0 -0 0

0
"0

\0

100mm

Fig.l Elelents and cells


of the square plate

Y mm

Fig.2 Te.perature and stress of


the square plate (t=5s, x=50'1)

135

Exact
120

lrl 10

!
~

0
0

60
40
20

0"

-100

:\

zoo

/
0"'-0

Temperature

Stress

100

O'xx

,,;:

120

"'-0-"

/0

Exact

100

Temperature
Stress 0'..

III

Of)

t
-200

<1\

300

~__~__~~~~__~-~O

20

40

60

BO

100

V mm

V mm

Fig.3 TelPerature and stress of


the square plate (t=15s, X=5011)

Fig.4 TelPerature and stress of


the square plate (t=30s, x=50.1I)

OC. The telPerature and stress O'xx at each tile step are cOlputed. It is
also assuled that G=80.8GPa ( Young's lodulus is 210GPa ), II =0.3,
a=1.IXIO-6/C, and ka=ISI1 2 /s. In nUlerical cOIPutation we use the tille
step ~t=5s and assule the plane strain state. In Figs.2 to 4 are shown the
results obtained at tiles 5s, 15s and 30s, respectively. There is no
appreciable difference between the nUlerical results obtained for various
values of the co-ordinate x. COIParison is lade between the present
boundary elelent solutions obtained at x=5011 and the exact ones I2 13
Fairly good agreelent can be recognized between both the solutions.
Figure 5 shows the boundary elelent lesh and the internal cell subdivision,
used in the analysis of a quarter portion of the cylinder by cosidering its
SYlletry. The whole boundary is divided into 40 linear elelents, and the
dOlain into 200 linear triangle cells. as shown in Fig.5. When the initial
telPerature is assuled to be 100'C and the telPerature on the surface is
assuled to be OC, we calculate the telperature, radial and hoop stresses at
each tile step. It is assuled that the aaterial constants are the same as
those of the square plate. In nUlerical cOIPutation we also use the tile
step ~t=5s and assule the plane strain state. In Figs.S to 8 are shown the
results obtained at tiles 5s. 15s and 30s, respectively. The present
boundary elelent solutions obtained along the radius inclined at 45' frol
the co-ordinate x are cOIPared with the exact ones l4 Fairly good agreelent
can be recognized between both the solutions. It is interesting to note that

136

there is no appreciable difference between the results along various radii


of the cylinder,

Exact

120

:-> 100

Te~rature

Radial Stress

Hoop StrE'ss

300

~o-o-o-o-o-o'"

QI

\g

:; 80

~ 60
i.0

20

..

100mm

20

Temperature

Radial SlrE'55

Hoop Stress

~ ZOO

\'/\

:; 80
60

40

20 t--B _a_a_e-=oe::::o.-e--"
/
0

20

40

60

Radius

120

300

~O-O-O-o-o'o'b

QI

-100
100

80

Exact

iii

,!l!

100 V)

mm

Rdd lus

Exact

~ 100

VI
VI

Fig,6 Te.perature and stresses


of the cylinder(t=Ss. 45')

Fir,S Ele.ents and cells


of the cylinder

120

GO

40

If

200 ~
~

-,-,-.-,_o_o_~~

80

100

s:

!1J

..g:

Temperature

o
6

Radiat Streoss

:-> 100 "--o-o<:r--..


iii

('(I

c:J'.....

QI

80

0"

0\

AI

E 60

40
0

-100
100

mm

Fig,7 Te.perature and stresses


of the cylinder (t=lSs. 4S')

300

Hoop Stress

20
0

ZOO ~

11\

100 Ui

/ 6 0\
/6/ ~/~

1-21_a~-O0

20

40

60

Radius

80

-100
100

mm

Fig,S Te.perature and stresses


of the cylinder (t=30s. 4S')

137

4. CONCLUSION
In this paper. the boundary elelent forlulation based on the tile-stepping
schele has been presented for two-dilensional uncoupled thermoelasticity.
Through nUlerical cOIPutation of a few exalPles. it is revealed that the
present lethod can provide accurate solutions in an efficient lanner.
As the future research work. we lay recollend application of the proposed
tile-stepping approach to three dilensional therloelasticity. and loreover
to fully coupled therloelasticity.

ACKNOWLEDGEMENT

Parts of this work were supported by the Grant-In-Aid for General Scientific
Research (e) of the Japanese Ministry of Education. One of the authors
(M.T.) wishes to express his cordial thanks for this financial support.

REFERENCES

1. Suh. I. and Tosaka. N. Boundary elelent analysis of linear coupled

2.

3.
4.

5.
6.
7.
8.

9.
10.
11.
12.
13.

therloelasticity problels. Proc. 4th JASCOME Symposiul on BEM. (1987).


pp.99-104.
Tanaka. M. Togoh. H. and Kikuta, M. Fracture lechanics application in
therloelastic states. Topics in Boundary Elelent Research. Vol.I. ed. by
C. A. Brebbia. Springer-Verlag. (1984). pp.59-77.
Chaudouet, A. Three-dilensional transient therloelastic analysis by the
BIE lethod. Int. J. NUl. Meth. Eng . 24(1987). pp.25-45.
Dargush. G. F. and Banerjee. P. K. Boundan elelent lethods for
poroelastic and therloelastic analyses. DeveloPlents in Boundary Element
Methods-5. ed. by P. K. Banerjee and R. B. Wilson. Elsevier Appl. Sci.
(1989). pp.l19-156.
Roures. V. and Alarcon. E. Transient heat conduction problems using
B.I.E.M COIPuters &Structures. 18-6(1983). pp.717-730.
Tanaka. M. and Matsuloto. T. Transient elastodynalic boundary element
forlulations based on the tile-stepping schele. Int. J. Pres. & Piping.
42(1990), pp.317-332.
Tanaka. M. and Asahi. A New boundary elelent lethods for therloelastic
problels. Materials Systel, 8(1984). pp.157-162. (in Japanese).
H~rlander. L. Linear Partitial Differential Operators. Springer-Verlag,
(1969). p.3.
Tosaka. N. Boundary Integral equation forlulations for linear coupled
therloelasticity. Proc. 3rd JASCOME SYIPosiul on BEM. (1986),pp.207-212.
MiuIOtO, T., Handbook of Matheut ics. ed. by Yano. K. MorokitaShuppan. (1985). p.500, (in Japanese).
Brebbia. C. A. and Georgiou. P., COlbination of boundary and finite
elelents for elastostatics. Appl. Math. Model . 3(1979). pp.212-220.
Carslaw. H. S. and Jaeger. J. C Conduction of Heat in Solids. 2nd ed
Oxford at Clarendon Press. (1987). pp.92-105.
Boley. B. A. and Weiner. J. H., Theory of Therlal Stresses. John Wiley &
Sons, (1960). pp.258-260.

138
14. Takeuti, Y. and Noda, N., Analues of Therlal Stresses, NisshinShuppan, (1989), p.119, (in Japanese).
APpendix 1 Fundaaental solution
Uil(X,\,I)

uiJ(x,\,I)

1
r
1
81lG(l-1I) ( [-(S-h)(C+ln2Jkedt)--2-l 81l+ r,l r,l)
(i.j=I.2)

(AI)
(A2)

(AS)
(A4)

where KI(r) is the first-order .odifled Bessel function of second kind and
the notation ( ).1 denotes the differentiation of the variable ( ) with
respect to a spatial co-ordinate x I.
APpend I x 2

Funda.enta I so I ut I on

p iK(X.\,I)=

p i J (x. \,I)

r (1-2&1) n l u jK. 1(x.\,I)+2


GnJUIK,J(X,\,I)

r 11

n 1 u iKol (X. \,I) +


(K=1,2.S;i,j=I.2)

(A5)

(K=1.2.3;i=1.2)

(A6)

piJ(x.\,I)=-41l(l:1I)r ([(1-2&l)8IJ+2r.1 roll


(1-211) (n

r l - n Jr.

~~

I) )

(j,j=1.2)

(A7)
(A8)

(AlO)

where
n 2.

is the un i t nona I vector on the boundar)' with co.ponents n 1 and

Calculation of Aerodynamic Characteristics of a Heaving


and Pitching Airfoil by a Panel-Vortex Method
K.Kamemoto and T.Mine
Department of Mechanical Engineering, Yokohama
National University,
Yokohama 240, Japan

Summary
This paper discusses an application of a two-dimensional vortex model to
investigation of boundary layer characteristics and unsteady flow around a
heaving and pitching airfoil with or without boundary layer separation in a
uniform flow. In this method, the turbulent boundary layer around an airfoil
is represented by a distribution of vortex sheet elements at the distance of
local displacement thickness from the surface, and the free shear layers
shed from the airfoil are modelled with discrete vortices which are
introduced step by step at the trailing edge or a separation point. As a
result of the calculations of flow patterns, unsteady fluid forces and
powers done by the flow for a harmonically heaving and pitching NACA-0012
airfoil, it was concluded that the present method based on BEM and DVM
provides reasonable prediction of the aerodynamic characteristics.
Introduction
In this study, in relation to an alternative to the windmill, the unsteady
flow around a two-dimensional airfoil, which is heaving and pitching in a
uniform flow, is analyzed numerically, using the Panel-Vortex Method
presented by Kamemoto & Suzuki[1]. It is known, if the motion in heave and
pitch is simple harmonic of a frequency,

so low that the forces acting on

the airfoil can be taken as quasi steady, the work done by the flow per
cycle becomes the maximum when the phase angle by which heave lags behind
pitch is 90 and it is independent from the frequency. As the frequency is
increased, however, the quasi steady condition becomes inadequate due to the
unsteady characteristics of fluid forces, and in the actual flow, the
boundary layer separation will occur either at a higher frequency or at a
larger pitch amplitude. Therefore, it seems interest to examine the work
various frequencies and pitch amplitude.
done per cycle of the motion for
The unsteady flow around an airfoil is solved at successive intervals of
time starting with an initial flow at t=O, by the Panel-Vortex Method which
is a direct viscid-inviscid interaction scheme based on BEM and DVM. In the
present paper, calculation results of flow patterns and coefficients of lift

140
and drag forces acted on an airfoil, which is heaving and pitching in a
uniform flow, are presented. Furthermore, the work done by the flow is
estimated for various frequencies.
Mathematical model of flow around a heaving and pitching airfoil
In this study, a NACA 0012 airfoil was treated as a heaving and pitching
airfoil, and the motion in heave and pitch is assumed to be simple harmonic,
so that the heave oscillation and the absolute angle of pitching are
respectively expressed as
y=y 0 sin

GI t

(1 )

a=ao cos Glt

(2)

where the heave displacement y

~~-.-.-.-/-.-

~L
Cd

=-

is positive upwards and the


pi tch displacement
a
is
Figure
Airfoil heaving and
pitching in the uniform flow.
positive nose up and yo and ao
are their maximum
values.
Equa tions( 1) and (2) show that the phase angle by which heave lags behind
pi tch is 90,. Figure 1 shows the positions of the airfoil at four
consecutive quarter-cycle intervals for the case given by Eqs (1) and (2).
Let the heave velocity of the center of action of the airfoil be written
GlYo cos Glt
and let the lift on the airfoil per unit span L be positive
upwards, the rate P at which work is done by the airfoil is

The work done per cycle is the integral of this rate and is

(4)
It is obvious that the feature of change of the lift with incidence has
considerable effect on the work. As described by Niblett[2J, if the
frequency GI of heave and pitch is so low that the aerodynamic
characteristics can be considered as quasi steady, and if the rate of change
of lift coefficient with incidence dCli da can be given a constant a, the
work done per cycle obtained from Eq.(4) becomes W' :J(pU:C a ao Yo/2 .' here
is the density of the fluid, U.is its velocity, C is the chord of the

airfoil. It is known, this work is larger than those for any other phase

141

angles between heave and pitch than 90. On the other hand, when the
frequency is not so low, even if the aerodynamic features are still
considered as quasi steady, the heave velocity V= QlYo cos Qlt can not be
ignored, and so the rate of change of lift coefficient should be considered
with relative incidence fJ instead of absolute incidence a. Furthermore,
when the frequency becomes much larger, the quasi steady condition is not
adequate because of its appearance of unsteady characteristics of fluid
forces, and even the boundary layer separation will occur at a considerably
high frequency or at severely high incidence.
In modelling of the flow around a heaving and pitching airfoil by the PanelVortex Method, the boundary layer over the surface is represented by a
distribution of vortex sheet panels. The strength of vortex sheet is the
circulation r corresponding to the total vorticity existing in the boundary
layer per unit length of x along the airfoil surface which is expressed as
r = W,where W denotes the inviscid relative flow velocity outside the
boundary layer. Distributing the vortex sheet panels at the height of local
displacement thickness from the surface of the airfoil configuration, the
boundary layer displacement effect is accounted for by replacement of
instantaneous profile formed with displacement thicknesses by vortex panels
as shown in Fig.2.

r. ",r. ."""'''''''''''''.",..;!J
ilii,,=~

x control point
(a) Separation point
(b) Trailing edge
Figure 2 Modelling with vortex panels and discrete vortices.
The unsteady boundary layer correction to the inviscid calculation is
introduced by displacing each panel corner point step by step in the
direction normal to the airfoil surface by the movement during a small time
interval dt, due to the displacement velocity, which can be defined from the
tWO-dimensional unsteady, incompressible boundary layer equation and the
continuity equation as

a8*= _ 6* aw _ (1 +_2_) 6* aw +W

at

w at

ax

(-!;
aH
H ax

_...L a8 +...L Cf)


H ax 2

142

The shape factor H and friction coefficient Cf are defined as 6*18 and
Tw/(p W~

12) respectively, where 8 is momentum thickness, Twdenotes the

surface stress, and Wo is the relatively approaching flow velocity.


In the present calculation, the boundary layer is assumed to be turbulent
from the stagnation point without transition, and the friction coefficient
Cf is given qy the empirical formula of Ludwieg-Tillman[3J, and variation of
the shape factor H is calculated from the entrainment equation presented by
Head[4], where the condition of turbulent boundary layer separation is given
as

H~

2. O.

As illustrated in Fig.2, the continuous leaving of ei ther upper or lower


boundary layer is individually represented by corresponding discrete
vortices for either case of fully-attached or separated flow. An extra panel
of length aI and constant circulation strength is attached to the downstream
corner of the vortex panel representing displacement thickness at the
trailing edge or a separation point, in the direction of extending the
displacement thickness panel. In subsequent time, the extra panel is
replaced by a discrete vortex which convects downstream according to the
local velocity. The strength of the discrete vortex is given as
dr=r 2 d V2

When boundary layer separation occurs, the displacement


thickness in the separated flow region is assumed to be the initial value,
because it is thought to have little effect on the inviscid flow
calculation.
In order to consider the effect of viscocity on vorticity diffusion, a
viscous vortex core of its radius is introduced so that the vortex has the
characteristics of a forced vortex within the core radius instead of that
of a free vorte~ The initial core radius for each vortex is given to be the
displacement thickness at its leaving point and the change of core radius
from to + de due to viscous diffusion during a time interval dt is
estimated from a solution of two dimensional vorticity diffusion equation as
d=c 2

)1

d tl (2 ) where c=2.24,

and)l is the kinematic viscosity.

Procedure of calculation
The flow of computation with the present method is as follows. (1) The
unsteady flow around a heaving and pitching airfoil is solved at successive
intervals of time, starting with an initial flow and motion of heave and
pi tch at t=O. (2) In order to determine the bound vortex sheet strength on

143

each panel, the boundary condition of normal velocity component


corresponding to the displacement veloci ty given by Eq.(5), is applied at
the control point on the panel resulting in the following system of
equations.
(6)

The second, third and forth terms are the normal velocities induced at the
i-th control point by the bound vortex panels and the two extra panels at
time t respectively, these terms contain the unknown factors of vortex
strength, where as the first, fifth and sixth terms are the normal velocity
components of the free stream, the velocity induced by all discrete vortices
and the movement due to heave and pitch respectively,

which can be

completely evaluated. The coefficient k in the right hand term is

relaxation factor for numerical stability. (3) The calculation of vortex


strength on panels gives both a new instantaneous velocity distribution
around the airfoil and convective velocities of discrete vortices. (4) The
unsteady boundary layer calculation for the new relative velocity
distribution yields new information on displacement corrections o6i 0 t and H
to the subsequent inviscid flow calculation. (5) Convective movement of
discrete vortices and displacement of vortex panels during the time interval
dt are solved. (6) New discrete vortices are introduced in the proximity of
either the trailing edge or a separation point.
Then, the unsteady flow around the heaving and pitching airfoil is solved
step by step by repetition of this procedure until given time to. It should
be noted here that in the present calculations, non-dimensional values based
on the approaching absolute velocity Uo and the chord length C of the
airfoil were used.
Results and discussion
Using the present numerical method, the unsteady and incompressible flow
around a NACA-0012 airfoil harmonically heaving and pitching in a uniform
stream was investigated, where this airfoil has the center of action at 35%
cord from its leading edge. Several combinations of non-dimensional period
and maximum angle of absolute incidence were chosen among values of
= 5 to 20 and

T=TUo/C

ao=10 to 60, keeping the maximum heave Ya/C in Eq.(1)

unity. The Reynolds number Re = UoC/v, the initial displacement thicknesses

144

Figure 3

Flow pattern at t = 40, absolute and relative incidence angles,


lift and drag coefficients calculated for the period T = 20 and
the pitching amplitude a 0 = 20
l)

ci

ci

~:~~~p~~,
I

Figure 4 Flow pattern at t = 40, absolute and relative inciden~e angles,


lift and drag coefficients calculated for the period T = 20 and
the pitching amplitude ao= 10.

145

...

/~<;:.:-~: .

... ...
:

Q)
Q)

'"

~.:

...

':,':

"

Relative an91~

"

I." :

I ...

bQ
Q)

<l:l.. 't:l .,;


0

......

.....

0
0

CI

ug
~O
u&

't:l

15.0

&
N

Figure 5 Flow pattern at t = 15, absolute and relative incidence angles,


lift and drag coefficients calculated for the period T = 7.5 and
the pitching amplitude Qo= 50.

and time step size were respectively taken as Re=5x 106


dt=0.05. Figure 3

ii"'/C =0.001

and

shows a flow pattern at a non-dimensional time t=40 after

impulsi vely starting the heave and pitch motion of period T=20 and pi tch
ampli tude Q. = 20" , and the figure also shows variation of lift and drag
coefficients with time, where Cl and Cd are defined as Cl= L/(PU:'C/2) and
Cd= D/(PU:C/2) respectively. It can be seen in this figure that the phase of
variation of lift leads that of relative incidence to some extent, although
the oscillation period T=20 is rather large, ie. the frequency is rather
low. This result suggests that the assumption of quasi steady state becomes
much more inadequate as the period becomes smaller.
Figure 4 shows a flow pattern at t=40 and variation of lift and drag
coefficients calculated for T=20 and ao = 10 , at which the relative
incidence has opposite sign with respect to the heave displacement at any
time. In this case, i t is shown in this figure, the phase relation of the
lift with the relative incidence becomes phase-lag instead of phase-lead.

146

~ ~

;..

o~ ."~
'ii

'+--==~:::---'r--::::~=---I::-:--=:::'k::.:---1:-::-:-:::;:;=-'I~,,-..
.0
t= 15.0

ci

'j

15.0

Figure 6 Flow pattern at t = 15, absolute and relative incidence angles,


lift and drag coefficients calculated for the period T = 7.5 and
the pitching amplitude a 0 30.
=0

Calculated results for a smaller period are demonstrated in Figs.5 and 6.


Figure 5 shows a flow pattern at t=o15 and variation of lift and drag
coefficients for T=7.5 andao= 50, at which the relative incidence has the
same sign as the heave displacement at any time. In this figure, it is seen
that the phase-lead of the lift variation is much more considerable and
negative drag appears periodically.
Figure 6 shows also a flow pattern at t=15 and variation of lift and drag
coefficients for T=7.5 but ao= 30, and in this case the relative incidence
has opposite sign wi th respect to the heave displacement at any time. In
this figure, it can be detected that the phase-lag of the lift variation
concerning the relative incidence becomes considerable and the appearance of
negative drag is remarkable in a period.
Numerically integrating the product of the lift coefficient Cf and heave
veloci ty ;;Yo cos iltfor a period

T according

to Eq.(4),

the non-dimensional

147

~
IL-________________________________

Figure 7 Work done by the flow per cycle calculated for various periods and
pitching amplitudes.

~----~~----~----~------~----~----~----,

C>

(j

10'

30'
40'
ANGLE a.

50'

60'

Figure 8 Time mean power done by the flow estimated from the work per
cycle.

work done by the flow per cycle can be estimated. Figure 7 shows the
calculated relations between the work per cycle and absolute incidence
amplitude oro for various oscillation period, where the hatched regions show

148

the magnitude of fluctuation of the work especially due to the unsteady


boundary layer separation.
In Figure 8, values of non-dimensional time-mean-power

P done

by the flow,

which is obtained by dividing the non-dimensional work per cycle by period

T,

are shown together with curves of constant relative pitch amplitude Po

It is clear in this figure that, as far as the relative pitch amplitude Po


is a value between 10 and -10 at which no significant separation of flow
seems to appear, larger power is obtained at a smaller period ie. a higher
frequency.
Conclusions
Basic aerodynamic characteristics of a NACA-0012 airfoil heaving and
pitching in a uniform stream were numerically investigated by applying the
Panel-Vortex Method. The following conclusions are obtained.
(1) Although the non-dimensional period of simple harmonic heave and pitch
is not so small as T=20, the quasi steady condition becomes inadequate.

(2) According to the combination of period ""


T and absolute pitch amplitude

a.,

the variation of lift coefficient with time has a phase-lead feature

with respect to the relative incidence when the signs of relative incidence
and heave velocity are the same, and, on the other hand, it has a phase-lag
featuze when their signs are opposite to each other.
(3) As far as the relative pitch amplitude Po is a value between 10 and

-10,larger absolute value of power is obtained at a smaller period of


heave and pitch ie. a higher frequency.
References
(1) Kamemoto, K and Suzuki, T.:A vortex method for predicting
unsteady flow and boundary layer development around an airfoil.
Boundary Element Methods in Engineering, Springer-Verlag

(Proc. of

ISBEM 89', East Hartford) (1990) 51-57.


(2) Niblett, L.T.: A guide to classical flutter. Aeron. J.(Nov. 1988)
354.
(3) Ludwieg,

H.

and Tillmann,

W.:

339-

Untersuchungen liber die

Wandschubspannung in turbulent Reibunsschichten.Ing-Arch. 17(1949) 288299.


(4) Head, M.R.: Entrainment in the turbulent boundary layer.
(1958).

ARC RM 3152

BEM-WFDM Coupled Analysis for Groundwater Resources


M. Kanoh
T. Kuroki

Civil Engineering Department, Kyushu Sangyo University, Fukuoka 813, Japan


Civil Engineering Department, Fukuoka University, Fukuoka 81401, Japan

ABSTRACT
The purpose of our paper is to present an application of the boundary element method
combined with the weighted finite difference method to the analysis of groundwater seepage
flow and saltwater wedge diffusion by the coast. The domain is roughly subdivided into linear
and nonlinear flow zones. Boundary element method is used in the linear flow zone, coupled
with nonlinear flow zone. In the weighted finite difference method, the rectangular local
coordinates ; and 11 are introduced in the nonlinear flow zone in such a way that 11 is
tangential to the interface between the two zones in order to evaluate the flux across the
interface. The new weighted fmite difference scheme, in which the continuity of flux across the
interface can be strictly treated, is proposed in the paper. The combined method is tested to the
salt water wedge diffusion in an experimental vessel as an example of groundwater resources
by the coast. The numerical solutions of the combined method for the problem are in good
agreement with the experimental results. As compared to the BEM-FEM coupled
analysis(Kanoh et. aI., 1990), the method gives rise to an accurate solution for the problem
with much save of the computational time.

KEYWORDS
Groundwater resource; Coupling; Weighted finite difference method; Local coordinates;
Continuity of velocity flux

INTRODUCTION
Groundwater resources by the coast or by the ancient ocean shut in the underground, which
can be observed in the Scandinavia Peninsula, sometimes encounters the saltwater pollution
problem. In the problem saltwater wedge diffusion occurs, where the very delicate balance of
density is kept and the velocity of the seepage flow delicately changes the direction and the
magnitude. Up until now by BEM, FEM and BEM-FEM coupled method we can not succeed
to analyze the subtle phenomena. On the other hand by weighted finite difference
method(referred herein as WFDM) we could obtain the comparable solutions with the
experimental results(Kanoh et. aI., 1989). However the WFDM by itself requires a large
amount of computational memory and CPU time. Then in this paper BEM-WFDM coupled
method is discussed.

GOVERNING EQUATIONS
The phenomena of fluid flow and saltwater diffusion in a porous medium by the coast can be
described by a set of coupled partial differential equations. These equations can be derived as
follows.
(a) Flow equations
We assume that the seepage is two-dimensional and the flow is in Darcy flow range. With the
rectangular coordinates Xi (Xl and X2 : let Xl be a horizontal axis) and the rectangular local
coordinates;,11 (let; be tangential to the interface between the two zones), the governing
equations for the seepage flow are given by
(1)

150

a 2H

aH
at

{e ( H ) + as} -

a 2H

pap

- k - - - k - - = k ~ ...:.......cos9) + k ~ _. sin 9) , (2)


a~2
a1]2
a~ PI
a1] PI

where qH) is the specific water capacity, a is a constant(a =1 in a saturated zone,a =0


in an unsaturated zone), S is the specific coefficient of storage. H is the water head, k is the
permeability. and 9 is the angle formed where 1] and Xl cross each other.
Here P is the density of the mixing water given by

o ~e

'5.] ,

(3)

with the density of fresh water Pi and the density of salt water Ps and the salt water
concentration e. The velocity components \/j \/ ~ and \/" of Darcy flow are defmed by
\/j

oH

(lXj

PI

\/J<=_k(oH +.E..cos9)

..

=-k(-~-+Oj2-)

o~

PI

'

(4)

\/'1=_k(oH +.E..sin9)

d1]

PI

'

(5)

with Kronecker's delta OJ 2


(b) Convective diffusion equations
The governing equations for the convective diffusion of the salt concentration are given by

ae
ae
-ae = -\/~--\/,,-+
at

ae
at

-- =

a~

ae
aXl

-\/1---\/Z

a1]

ae
axz

a~2

a 2e
a1]2

(6)

a 2e
ax/

a 2e
axz 2

(7)

a 2e

d~--+d,,--

--+ d1 - - + dz - -

where dl , d2 ,de and d" are diffusivities, \/1 , \/2 ,ve and \/" Darcy flow velocities
respectively. Here dl , d2 ,de, d" ,\/1 , \/2 '\/e and v" are assumed to be constant

WFDM FORMULATION
Referring to the Fig.t, we shall divide the whole region into three subregions ill ,il2 and
are fresh and salt water zones respectively, in which the
concentrations are constant, therefore seepage flow may only be considered. The region il2 is
a mixing zone governed by the nonlinear coupled equations for water head and concentration.
We shall use WFDM in il2 with the rectangular local coordinates ~ .1/ ,coupled with BEM in
ill and il3
il 3' The regions ill and il 3

(a) WFDM for flow equations including the interfaces


In order to evaluate the flux across the interface. the alternative WFDM scheme, in which the
continuity of the flux across the interface can be rigorously treated, is considered. Here the
velocities(\/n) in the external normal direction on the interfaces rl2 and r23 are known. On
the other hand the velocities V'I equal \//1 on the interfaces Tl2

in Fig.2. Modifying Eq.(2) as the following description,

and T23 as illustrated

151

Fig. 1. Subdomain method with WFDM meshes

aH

at

a 2H

aH

p.

- K - - = K ~ ~ose) + ~ K ( - - + -sine)}
a~2
a~ PI
aT]
aT]
PI

(8)

then substituting v1'/ of the Eq.(5) to Eq.(8) the seepage flow equation yields

aH

_K

at

a~2

= K ~ ..!!.-.cos e)
a~ PI

_~ vl1 )
aT] A

(9)

wltere K is kl{C(H)+aS} and A is (C(H) + as}.


When Eq.(9) is considered as the equation of H, the right hand side denoted simply by G can
be regarded as an inhomogeneous term. Then WFDM for the inhomogeneous seepage flow
equation can be obtained similarly by the procedure(Kanoh et. al., 1989):
l.WFDM corresponding homogeneous seepage flow equation(referred herein to homogeneous
WFDM) is considered.
2.The homogeneous WFDM is obtained by the similar method to the advective diffusion
equation(Kanoh et. aI., 1986). Namely the neighboring points in the scheme illustrated in
Fig.3 are considered, the homogeneous WFDM has the form
H (i ,g j )=DI {H (i +I,g j )+H (i -I,g j )+H (i ,g +1 j )+H (i ,g -I j )}+D2 -H (i ,g j -I)
+DJ { H ( i +I,g j -I) + H ( i -I,g j -I) +H ( i ,g +1 j -I) +H ( i ,g -I j-I))
. (10)
The weights DI,D2 ,DJ are determined by the linear equations

-Kn

1-4Kn

6K~

1-12Kn
+24K~

[::1 [:]
=

(11)

The linear equations can be obtained, if we substitute the three H values discretised as Eq.(12)
into Eq.(lO)

(12)

152

Fig.2 Coupling on interface r12


t. j

mi

J---------- ~.i
g I

i1

iIi+1

II

&+1

1\ . II

~}

unknown

:} known

Fig. 3. WFDM scheme example


where PI ,P2, q = 0, 1, 2, .. , h and.1t are space and time increments, and Kn is
K.1tlhl . Here Eq.(l2) can describe the discretised polynomials which satisfy the homogeneous
equation.
3.WFDM for the inhomogeneous equation (referred herein to inhomogeneous WFDM) is
obtained based on the homogeneous WFDM. With the neighboring points in Fig.3, the
inhomogeneous WFDM is
H (i ,g j) =DI (H( i +l.g j )+H( i -l.g j )+H( i ,g +lj )+H( i ,g -lj)}+ DzH(i.g j -1)
+DJ {H (i +l.g j -1)+ H (i -l,g j -l)+H (i.g +lj -l)+H (i.g -lj -l)}+EJ G (i ,g j -1)
+E2 (G(i+1.gj-1)+G(i-1,gj-J)+G(i,g+lj-J)+G(i.g-1j-1)} ,
(13)

153

where C(i , g ,j) is the right hand side of Eq.(9), which includes the derivative of v71
point. The weights E} and El are detennined by the linear equations

4 ][E1] _ Ii. [2 -Kn

2-4Kn

El

113 K~/2

- K

4Kn

21I3-2Kn+

2K~

[D1]

'

at the

(14)

~:

The equations are obtained, if we substitute the two sets of C and H values obtained from
Eq.(15) into Eq.(13).

= :Jf-r( Pl)l + qKn} ,


K
H(4) = =-!?t 2(p})4 + (p})2 qKn

C(2)= 1 , H(2)
C(4) = hl{ (P1)1 + qKn)

21

2'

41

21

(qKn)l )
4
'

(15)

Eq.(15) describes the sets of discretised polynomials which satisfy Eq.(9).


(b) WFDM for flow equations excluding the interfaces
The neighboring points in the scheme illustrated in Fig.3 are considered, WFDM scheme for
seepage flow equation(Eq.(2 in D2 excluding the interfaces 12 and 23 can be
obtained by the procedure (Kanoh et al.1989) as the following equation:

H (i ,g J )=AJ {H (i +i,g J )+H (i -I,g J )+H (i ,g +iJ )+H (i ,g -IJ )}+A2H (i ,g J -i)
+AJ {H (i +i,g J -1)+H (i -i,g J -1)+H (i ,g +iJ -i)+H (i ,g -IJ -1)}+B]'P (i ,g j -1)
+Bz{ P (i +i,g J -1) + P (i -I,g J -1) +P (i ,g +IJ -1) +P (i ,g -iJ -i)}
(16)

where H(i ,g ,j) is the unknown water head at the point (i ii~, g iil1,jiit) ,A 1 ,A2 ,AJ ,B 1
and B2 are the weights. P(i ,g J) is the right hand side of Eq.(2). The weights are determined
by the linear equations (17) and (18) respectively
4

-Kn

1-4Kn

12K~

1-24Kn
+4BK;,

1
-Kn

4
1-4Kn

l::] [:1

][B1]_~[

B2 - K

2 -Kn
116 K2n

]
2-4Kn
1I6-2Kn+ 2K;,

(17)

[~l]
A~

(18)

(c) WFDM for convective diffusion in porous media


On the interfaces the concentrations(C) are known. namely C =0 on 12 and C =1 on 23
,and the derivatives of the concentration are unknown. Therefore it is not necessary to
investigate the continuity of the concentration flux across the interfaces. Then we can use the
same WFDM as the one for convective diffusion in porous media presented by Kanoh et
al.,(1989). With the some neighboring points selected from the points in Fig.3, the WFDM
scheme is

C (i ,g J) =Wd C (i -I,g J) + C ( i ,g -i j )}+ W2{C ( i + I,g j) +H ( i ,g +i J )}


+ WJ {C (i ,g J -1) }
(19)

154

The weights Wj ,Wz ,WJ are detennined by the linear equations


2

o
o

2( i-C,J
2( i-C,J2+4DII

where

CII

(20)

and DII are Courant and diffusion numbers respectively.

BEM FORMULATION
In the regions DI and ~ the densities are constant and the seepage flow can be treated as
steady. therefore only the Laplace equation for the seepage flow need to be considered.
The expression is
k.1H =0
(21)
Using the fundamental solution to the Laplacian
(22)

with r = PQ for two points P and Q. the unknown H


equation
c(P)H(P) +

where

and aHI an

VII

VII

Vi nj

an

ant

ToH v*dr

(23)

are written as

aH

= - k (- +

aXj

and hence

aH

ov*
HTdr=

Jnt

satisfies the boundary integral

VII

PI

p
PI

0 j 2' -

) nj

oH
P
=-k - k '"-fly

an

(24)

PI

= - - --ny

Then, in Dj Eq.(23) becomes

.i

c(P)H(P) +

elll!

H--dr=-ov*
1
0n
k

(25)

vAv*dr-

Jll!

n,v*dr

(26)

Jll!

By the boundary element collocation method using linear shape functions, the integral equation

is approximated in the form


Nt

'"

-(1)
~ Hij
Hj

j=l

Nt

-(1)
(1)
= - lk '"
~ G
ij (V,.)j - a j

i = 1,2, ... , Nl

(27)

j=l

with nodal unknowns Hj and (V,.h on the boundary aDt. Similarly in D3 ,we can obtain

c(P)H(P) +

.103

av*
1
HTdr=-n
k

v.v*dr+(l+/3)

155

(28)

nyv*dr

.1m

.103

i = 1,2, ... ,N3


where /3 (={Ps- pplp!)

(29)

is a constant.

COUPLING
As illustrated in Fig.2,

r 1Z

and

r Z3 are internal boundaries between regions 01,

Oz and

between regions Oz. 03 respectively. In the regions OJ and 03 the concentrations are
known, so that the seepage flow may only be considered. Along the interfaces JZ and Z3 '
Eqs.(13), (27) and (28) can be coupled by means of Vn and v l1 As a matter of fact,
referring to Fig.2, G included in Eq.(13) for the mesh adjacent to r lZ is written as

Gi1 t = G(i,gj-I) = IK

~Lcos8)~VA1J)
a~ Pi
a11

I. .

(30)

l,gJ-I

Using a finite difference scheme. Eq.(30) yields

Gi1' = G(i ,g j-I) = -fi;f8(pjf - Pi11 ) +}f(Vnh - (v l1hiJl

=G(i ,g +lj-l) = -k[8(Pi1t - Pi1t) +}f(Vnh - (VT/h3Jl


Gi1t =G(i ,g -IJ-l) = 2~ [8(Pi1 t - pj~/) + fr(Vnh - (vT/h2J]
Gl = G(i +1,g j-l) = f[ {-r(Vn)l - (VT/)ll)]
Gi1 t =G(i -1,g j-l) = /h [8(Pi1 t - Pi1') + }f(VT/)Jl - (VT/hlJl
Gj1'

Similarly referring to Eg.4. G included in Eq.(13) for the mesh adjacent to

as

(31)

r Z3

is written

Gj~t =G(i,gj-l) = 2~ [8(pji'-pj~') +{-r<VT/J26-(Vn )61]


Gji t

=G(i ,g +1j-l) = -k [8(Pi~' - Pi~')

+ }f(VT/h8 - (vJiJ]

= -k[8(Pi~t-Pj~')

+}f(VT/h7-(Vn)])]

Gj~t =G(i,g-IJ-l)

f! {-r(VllJJ6 - (vJ61]
Gi~t = G(i +1,g j-l) = k !8(Pii t - Pi1t) + -}r(VT/h6 - (vT/)]6)]

Gi t = GO -1,g j-l) =

(32)

where 8 is Kocos8lPt. Gii' and Piit denote the G value and the density of the point
28 at the time (t -.1t) respectively. Eq.(31) and (32) are generally written as
(33)

156

Fig.4 Coupling on interface r23


As Eq.(13) is a fmite difference equation concerning only the point(i ,g j) under consideration
and the several neighboring points, it is necessary that we develop the thinking point to the
each point of the region. Then the developed system can be written in matrix form as

[p(2)] [H(2)] = [R(2)] [a(2)] + [S(2)] [H-<1t(2)]


Substituting Eq.(33) to Eq.(34), consequently the assembly of Eq.(34) in

nz

(34)
results in
(35)

According to the boundaries Trz ,rZ3 and rzz=afh \ (rlZ UrZ3), we shall
express the first term on the right hand side in terms of submatrlces as follows.

On the other hand, Eqs.(27) and (29) can be written respectively in the form

UfO l] (H Ol ) = - tWO(!l] (V~!l)_ (d

ll )

[il(ll] (H(3)

-1 [GPl] (v: l)_ (!ill)

(37), (38)

Since[a(!l] and [a P)] are non-singular, we have

We notice that the exterior normals on the interfaces have opposite directions. Taking the
components corresponding to {v~12)1 and from Eqs. (39), and (40) respectively, we can write
.

(41), (42)

157

Substituting these expressions into Eq.(36), we can obtain

[Q(2l] {V~'l}

=k

[Q(l2 l] {lOll} (H(ll) + k [Q(23l] [pm] (H(3)}

+ [Q(21l] lV~nl} + [Q(i 2l](a (Ill) + [Q(m] [a(23)]

={M(il] (H(I l}+(M(3l] (H(3l}+(h<'l}+{b(ll}+{b(3)} .

(43)

Consequently, from Eq.(35) we have the combined system of equations

[M(I)] {H(ll} _[P(ll]{H(ll} + [M(Jl] (H(3)}

= - V'l(p 'V'1)} - (h(ll) - {b(ll} - {b(ll}

(44)

NUMERICAL EXAMPLE
We consider salt water wedge diffusion in the experimental vessel as shown in Fig.5. In the
vessel small beads of 2.0-2.6mm in diameters are densely packed and a porous medium is
composed. The cross section is 80cm wide and 50cm high. The head of fresh water on the right
hand side is 57.9cm, while the head of salt water on the left hand is 56.0em. The densities of
fresh and salt water are 1.0 and 1.025 respectively. We consider that the diffusivities of the
porous media are proportional to the seepage flow velocities at the corresponding points and
the proportional constant depends on the porosity at the point. The domain D under
consideration is divided into three subdomains Dt , Dz and D3 as shown in Fig.5. In the
first and third domains the water head H is calculated by BEM, while in the second domain the
salt concentration C and H are calculated by WFDM. Along the interfaces rn and rZ3
BEM and WFDM are coupled by means of VII and v'1. Fig.6 shows profiles of calculated
c0ncentration distribution by the coupled method compared with the distribution of the
experimental result. The results by the BEM-WFDM coupled method are in good agreement
with the experimental results.

Fig.5. BEM and WFDM meshes the coupling method

158

Experimental result
BEM-WFDM coupled
method result

Fig.6. Experimental and BEM-WFDM coupled


method salt concentration distribution

CONCLUSION
In the regions, where the concentrations can be considered as constant, we can save CPU time
and memory by BEM. In the region, where salt and fresh waters are mixed, we use WFDM in
order to analyse skillfully and accurately the nonlinearity. To investigate the continuity of flux
across the interfaces, the alternative WFDM, where the local rectangular coordinates ;. 71 are
applyed, is proposed in this paper. Then WFDM is coupled with BEM on the internal
boundaries by means of VII and vlI' The coupled method is efficient in handling the highly
nonlinear effect. As compared to the WFDM or BEM applied to the whole region, the method
gives rise to a small system of nonlinear equations. As compared to the results by the BEMFEM coupled method, the calculated results by the BEM-WFDM coupled method are in good
agreement with the experimental results using rather coarse meshes.

REFERENCES
Kanoh, M. and Ueda, T. (1986) ; Application of Weighted FInite Difference Method to
One-or Two-Dimensional Convective Diffusion Problem, J. Hydroscience & Hydraulic
Engng., Japan Society of Civil Engng. Vo1.4, pp.11-29.
Kanoh, M., Kuga, Y., and Ueda, T.(1989); Weighted Finite Difference Method for Diffusive
Transport, Proc. Int Sym. on Sediment Transport Modeling, pp.209-214.
Kanoh, M., Kuroki, T. and Onishi, K. (1990) ; BEM-FEM Coupled Analysis for Salt
Water Wedge Diffusion in Groundwater, Proc. 3rd Japan-China Sym. on Boundary
Element Methods, pp.53-62.

Approximation Analysis on 3-Dimensional Coupled Electric


Potential and Hole or Electron by Boundary Element Method
Ryuji KAWAMURA

Information and Mathematical Science Laboratory, Inc. 2-43-1


Ikebukuro, Toshimaku, Tokyo 171
KEYWORDS Boundary Element Method, Electric Potential. Hole and Electron
Concentration, Numerical Analysis, Transient Problem, Advective
Diffusion, Drift Term
ABSTRACT
BEM is applied to coupled problems of electric potential and hole or electron of
semiconductor with semi-classical treatment for further investigation. In the design of
semiconductor device simulation, this simulation analysis is combined with the analysis of
electric potential flow and hole or electron. The former electric problem consists of non-linear
potential problem, however, the latter hole or electron problem is slightly complex because of a
drift term such as its advective term. Coupled electric potential flow and advective diffusion of
hole or electron are formulated in this paper by the direct boundary element method using

respective fundamental solution in each phenomenon. Physical solution is obtained when Peclet
number is large or coarse meshes are used in the modeL A few numerical examples are
presented. The method aims at analyzing not only simple but also conventional treatment
1 INTRODUCTION
Semi-classical treatment has been formulated for electric potential and hole or electron for
transister device simulationl '. Coupled analyses of electric potential and hole or electron device
simulation by FEM and FDM have been presented by several authors2l )'. The previous paper"
describes the direct integration method by means of exact time-integration of the fundamental
solution with the constant velocity in advective diffusion equation. Coupled treatments have
difficulty in advective diffusion equation because an explicit form of the fundamental solution
can be derived only for the case of constant electric field. BEM approach can be safely applied
when coarse meshs are used in the model at large Peclet number".
2

FORMULATION
Three-dimensional electric potential is governed by the following equation:
(1)

160

in which the right hand tenn is non-linear, where electric field

E is given by :

E = -V<ll

(2)

where Q is also expressed by using Boltzman equation") :


Q=e(pJ!xp[ -~<ll]-naexp[+~<ll])

(3)

where j3=eIkBT, and kB : Boltzmann constant, T: absolute temperature.


Three-dimensional current equation of hole and electron are given in the following forms,
(4)
(5)

(6)

J. =+eDnVn-ef.lnnV<ll

(7)

Eqs.(4),(5),(6) and (7) give,


(8)

(9)
Electric field E is devided into two parts of average Eo and of difference E'.
Eqs. (8) and (9) are replaced by,

d,n -f.ln(V, E)n -J.ln(EO V)n =DnV2n +f.ln(E'V)n +G n- Un

(9)'

where 1C : directric constant(F/m), <ll : electric potential(V), e: electric charge(C), E:


electric field(V/m), Q: source rate of potential(C/m), ND: donnerconcentration(l/m1, NA :
accepter concentration(l!m), p: hole concentration(lIm), n: electron concentration(l/m),
and Gp : rates of generation of hole( lIm3sec), up: rates of recombination of hole( l/m3sec), Gn :
generation rates of electron(1/m3sec), Un: recombination rates of electrOn(1/m3sec), J.lp : hole
mobility(m2/Vsec), J.lo: electron mobility(m2/Vsec), operator V (0 I
I
I OZ), and
Dp: diffusion coefficient of hole(m2/sec), Dn: diffusion coefficient of electron(m%ec). In
Eqs.(8) and (9), the third terms of left hand corespond to advective terms and the second terms

= ax,o oy,o

161

corespond to damping tenn in advective diffusion equation. We consider initial condition of


electric potential <ll(x,y,z,O)= <1>0 for iteration and that of the concentration of hole and electron
p(x,y,z,O)=Po and n(x,y,z,O)=no We also consider boundary conditions of the type
<ll(xO'yO'zO't)= <1>1. p(xO'yO'zo,t)=PI. n(xO'yo,zo,t)=n 1 1C OIPI ov=-q and DpOPI 0 v=-q p and DnOnl
ov=-qn. where v is perpendicular unit vector on surface. The fundamental solutions of electric
potential and concentration hole and electron are given by,
<1>' (r;rj ) = 1 / (41tlCf'),
p~(r,t;lj,'t)

(10)

=exp[-~/Eo' r')/2D p -(~pEO)2t' 14Dp -

~/V. E)t'-r,2 /4D/] / (41tD/)3/2

(11)
p~(r, t;lj,'t) = exp[ +~n (Eo' r') I 2Dn - (~nEO)2t' 14D. + ~n(V . E)t' _r'2 14D.t'] / (41tD nt,)3/2

(12)

respectively, where r' = r -

Ii and t' =1:-t.

Boundary integral equations are derived in the fonn,

and integral equations for hole p and electron n are obtained by describing by (p,n):
E>i(p,n)(ri , 't) -

ff q'
o

p.n

(r, t;ri, 't)(p,n)(r, t)drdt

+J.'o fpp,n (r,t;ri, t)~p .(E'V)(p,n)dnctt

(14)

where E>i is determined at the point considered. We put


p' (r;lj) =-lCV<1>'(r;Ij)'v

(15)

q;.n(r,t;rj , t) = -{DV(p,n)(r, t;lj, 't) ~p.nEo(r)(p,n)(r, t;'I, t)}v

(16)

As the variations in p, C!p and n, q" with small t are neglected, time integrations of pOp, q 'p and
n'., q'n are carried out4 ) respectively.

162

[{erfc(r'N4Dt +~IJ.(V. E)H(IJ.Eo)2't / 4D) 'exp[2r'/--/4D '~IJ.(V E)+(!!E)2/4D]


+{erfc(r'N4Dt -~IJ.(V. E)'t + (!!EO)2't / 4D) exp[-2r' /..J4D '~IJ.(V. E) + (lJ. pE o)2 / 4D]]
(17)

p' p.n't(r,rJ = S:q;.n(r,t;rj,'t)dt =expF+WEo' r')/ 2D]/ (41tD)3/2[+IJ.(E o n){(41tDi /2


/4r"(erfc(r'N4Dt +~IJ.(V. E)'t+ (!!Eo)2't/4D)' exp[2r'/..J4D '~IJ.(V. E)+(IJ.E o)z/4D]
+erfc(r'N4Dt -~IJ.(V. E)H(IJ.EO)2't/ 4D)exp[-2r'/..J4D '~IJ.(V. E) + (IJ.E o? /4DJ)}
+(r'n)(4D p exp[-r,2/4D't + IJ.(V. E)'t- (!!EO)2't /4D] /2r'z.[t

+.fiiD / r 2--/4D /2r'-~IJ.(V. E) + (IJ.E o)2 /4D)


(erfc(r'N4D't +~IJ.(V. E)'t + (IJ.Eo)2't/ 4D)exp[2r'/--/4D '~IJ.{V. E)+(IJ.E o)2/4D]
+(--/4D /2r'+~IJ.(V. E) + (IJ.Eoi /4D)
(enc(r'N4D't -~IJ.(V. E)H(IJ.Eo)2't 14D)exp[-2r'/..J4i5 ~IJ.(V. E) + (IJ.Eoi /4D]))].
(18)

Then, discretizations of Eqs.(14) and (15) after time integratations yield matrix equations as
follows,

JP;..'t(r,r)(p,nXr, 't)dr
= - J'l'p.n't(r,r)qp. (r, 't)dr + Jp;. (r,O;rj, 't)po(r)dn
+ J'l'p.n't(r,r XGp. (r, 't) - U (r, 'tdn + I 'l'p.n't(r,rJlJ.p. (E' V)(p,n)dO.
Elj(p,nXrj, 't) -

(19)

The resulting coupled integral equations (13) and (19) can be solved numerically.

In the case of steady concentration, 't -

00 at Eqs.(17) and (18) give,

'l'p.n(r,rj,oo) = exp[+IJ.(E o r')/2Dl1 (41tDr) exp[-2r'/--/4D '~IJ.{V. E) + (IJ.E O)2 /4D]


(20)

163

and
p' p.n(r,lj,oo) =exp[+Il(E o . r') /2D] / (41tD/ 12 [+Il(E o . n) + (r'n){(2D / r2

+..,f4DIl(Y' E) + (IlEO)2 I r) exp[-2r'/..J4D ..,fIl(Y' E) + (IlE O)2 /4D])]

(21)

3 EXAMPLE and RESULTS


a) Example 1
Three-dimensional unifoon cubic model at horizontal type with electric field is tested. Its
size are x(O, 5Ilm),y(0, 41lm) and z(O, 2.5llm). Steady drift flow is assumed by constant electric
field, ct>(5Ilm,y,z)=0.3V, ct>(O,O,O)=OV, and with non-linear teon by Eq.(3).

In Fig.1 and

Fig.2, coupled transient diffusion problems with drift terms are given by contours in hole and
electron concentration at 6 nsec by time step l.Onsec. Fig.3 shows electric potential contour
with voltage O.3V.
b) Example 2
In three-dimensional static case in the case of potential Eq.(l), hole and electron by
asymmetric concentration examples at vertical type show in Fig.4 and Fig.S.

Fig.6 shows

electric potential contour.


4 CONCLUSION
Three-dimensional coupled electric potential and hole or electron by the boundary element
method is foonulated. This code will be tested by examples compared with benchmark test for
transister device simulation.
REFERENCES
W.Shockley, Electrons and holes in Semiconductors, D.V.Nostrand Co., p.301, (1950)
2 E.Kameda and Y.Kagawa, Finite Element Simulation Using Semi-infinite Element Joint ly
on Semiconductors Neighboring to the Contact-Surface, 11th Electric and Electronics
Symposium p.211, (1990)
3 C.M.Snowden, Introduction to Semiconductor Device Modeling, World Sci. Pub. Ltd.
p.32, (1986)
4 R.Kawamura and M.Fukuma, Analysis of Three-Dimensional Transient Advectio
Diffusion by BEM with Time Dependent Fundamental Solution.The 6th Japan Nat. Symp.
on BEM, pp.73-80, (1989)
5 M.Kuroda, Y.Tanaka, M.Igarashi and T.Honma, A Boundary Element Analysis of
Convective Diffusion Equation, 10th Symposium of Computational Electrical Engineering,
pp.l97-200, (1989)
6 N.Mikoshiba, Physics ofTransister, Baifukan. p.197. (1989)

164

THREE OIMENSID-IAL B.E . M EXN1PLE

cootourl1'.l .

ELCTRIC nELO Q.d ELECTRO'I a.d HO..E .1

6 .00(g(jE-1\l9

I
2
3
4

~
8

d.
'61\X\l(\)
'8
: I.
: I
:

I.

'00
'00
'00
'~I
1
'I

1 .6lIlOC'~ 1

1800E'~1

x vector pote nt\ ol

n\n:: 484E"OO

lI'Iox= 200E .. ~l

at

at

77

'I'I od,

ncei.

Fig. l Contour of Concentration of Hole Symmetric Case


THREE DIMENSIO'lAL B.E . M EXN1PLE

g.@IIXIlG-1Il9

ELCTRIC FIELD a.d ELECTRO'I Q,d HOLE .1

r:ll~r:~
3
: I.

: I.
: I.

Fig. 2 Contour of Concentration of Electron Symmetric Case

I.

I.

'~I

t~ 1

'~I

165
THREE DIMNSICl-IAL aEN EXAl"PLE

~ . !li00/i' - 00

ELCTR[C FIELD

and

l vector pote nt\al


_In: ISlOOEtOO Qt noc1e
3 '&4.E-& 1 ot node

WlO)l:

Fig.3 Contour of Stationary Electric Potential


co.

I:

'!:

Symmetric Case

THREE DIMENSIONAL a E N EXAMPLE. ELCTRIC FIELD

and

contOllr
' I;jI..
, I

lV8l8'r:

<I

<I

ELECTRON

5
6

Fig.4 Contour of Concentration of Hole Asymmetric Case

I. 'II

'I.
III

HOLE

and

1,

I
I

'I
+

166

coaa

T~EE

DIMENSIONAL aE.N EXAMPLE . ELCTRIC FIELD

and

ELECTRON

and

HOLE

I. ''mII

con ~O~~IV~kJ,

34
S

I
I
I

I
+ I
.. I

123
I

Fig.5 Contour of Concentration of Electron Asymmetric Case


co .. e

THREE DIMENSIONAL a.E.N EXAMPLE. ELCTRIC FIELD

a nd

C~~~~~~I'!~~I10LE

'.

4
S

I.

,~.
'
.

Fig.S Contour of Stationary Electric Potential Asymmetric Case

-~~
-II
- I
- I

The Completed Double Layer BIEM: A Boundary futegral


Method for Complex Microstructures in a Viscous Fluid
Sangtae Kim
Department of Chemical Engineering, University of Wisconsin
Madison, WI 53706 USA

Summary
A boundary integral equation for the Stokes (linearized Navier-Stokes) equations has
been developed. In contrast to the classical integral representation of Lorentz (which
follows from a Green's identity) the integral representation employed here consists of an
aphysical dipole distribution, and a combination of point forces and torques. The existence
of this class of representations fonows from the Fredholm-Riesz-Schauder theory. In
the problems of greatest interest in particle hydrodynamics, the classical representation
leads to a Fredholm integral equation of the first kind, whereas the new method leads
to that of the second kind. Upon performing Wielandt deflations, the compact operator
that appears in the governing equation may be cast as a contraction mapping, leading
to rapidly converging iterative solutions. Preliminary exper.iments on high-performance
parallel computers suggest that the algorithm can be adapted to scalable parallel computer
architectures.
Introduction
The motion of small particles in a vacuum is governed by Newton's laws of motion.
Given the inertia of each particle and the net external force on each particle, one can in
principle determine the trajectory for each particle. The analysis of the motion of small
particles in a viscous fluid is far more difficult. One now has to solve for the motion of the
intervening fluid. The end result is a BVP which couples the geometries of particle shape
and relative configurations in a nontrivial fashion [1,2]. These difficulties not withstanding,
the analysis of the motion of N rigid particles of arbitrary shape in a three-dimensional
viscous flow field constitutes the fundamental problem in computer-aided simulation of
solid/liquid dispersions.
Boundary integral methods furnish an attractive approach to these problems by transforming the governing three-dimensional PDE (the Stokes equations):
(1)

to a two-dimensional boundary integral equation.


The classical velocity representationa of Lorentz [3] lead to a Fredholm integral equations of the first kind for the unknown surface tractions - and the usual mathematical
and computational difficulties associated with first kind equations. Recently, a new class
of integral representation developed by Power and Miranda [4] has received considerable
attention. These lead to a second-kind boundary integral equation for an aphysical surface
dipole density, thanks to the jump properties of the double layer potential [5].

168

_ ----------;:;NE:....CSX-2

X-MP

Cray 2S

WT3364

MFLOPs

80387

+ 80287
1980

82

84

86

88

1990

Figure 1: Evolution of floating point performance (supercomputer and micro) during the
80's: 1000xlOOO LINPACK.

We have developed a version of the double layer method, that we call the Completed
Double Layer Boundary Integral Equation Method (CDL-BIEM), that features an iterative solution of the integral equation [2,6,7]. Boundary element implementation of these
ideas yields a computational algorithm with scalable demands on inter-processor communications, well-suited for the next generation of parallel supercomputers. The road is thus
open for direct simulation of microstructure evolution in multi phase dispersions.
High-Performance Parallel Supercomputers
With the rapid development of supercomputing at the chip level (see Figure 1), one can
envision computer architectures that are essentially a network of many tightly integrated
"supercomputers on a chip", or what is commonly called a high-performance parallel
computer. We may expect to see in the near future, a computing environment consisting
of over 16,000 processors with an aggregate performance exceeding the TeraFLOPS (one
trillion floating point operations per second) leveL But this level of performance can only
be achieved by algorithms designed to exploit concurrent computing at each processor,
with minimal communication requirements between processors - so called highly parallel
algorithms. Our development of CDL-BIEM for particulate Stokes flow problems suggest
that it maps quite naturally onto a many-processor environment, provided that particular
attention is paid to overcoming inter-processor communication bottlenecks and memory
limitations.
The Completed Double Layer Boundary Integral Equation
The steps in this method are:

169

1. Construction of the relevant Green's function 9 (for the free-space or bounded


domain).

2. Construction of the kernel. First derive the expression for the stress field 0'( z; 9)
using the Newtonian constitutive equation. The kernel of the integral operator
follows as: K(z, y) = -2n O'(z - y; 9).
3. Construction of the integral representation for the velocity field in the fluid domain,(i.e. z E Vj ). This introduces the completed double layer representation:

1I(z)

= 1I RG (z) + is K(z,y) ",(y)dS(y) .

The so called range completer, 1I RG (z), is known and specified a priori and handles
the forces and torques applied on the particle.
4. On the boundary of the fluid domain, set the no-slip velocity boundary condition.
This leads to, on z E S = avj, the boundary integral equation:
11

_11 00

+ [u" + w" X (z -

zoe)]

= 1I RC + '" + .q",)

5. To allow simple iterative solution, perform Wielandt deflations to remove the outermost eigenvalues (.\ = -1) of the integral operator, i.e. map K -+ 11.

= -L<"',"">""

[u"+W,, x (z-z",)]

6N

L < """,i > + K(",)

11.(",)

'" + 11.(",)
The eigenfunctions

",i

",i

1=1

_11 00 -

for the eigenspace at ..\

1I RC

=:

= -1 are known analytically.

6. A second set of Wielandt deflations at ..\ = +1 maps 11. -+ il where the latter is a
contraction map. The resulting boundary integral equation,

is amenable to iterative solution for

<p.

7. Physical results such as the particle velocities, are obtained by "post processing" of
the solution <p. Actually, since the relations

[u"

Wa

(z - :1:,,)]

= - L < """,I > ",i


i

where computed to perform the eigenspace deflations, the physical result for the
particle velocities were computed as part of the solution procedure.
We now describe these basic steps in more detail. To derive the standard BEM for
Stokes flow, we start with the governing Stokes equations for an incompressible Newtonian
fluid of viscosity 1',

(2)

170

and recast them as a two dimensional (boundary) integral equation,

v(z)

= 8~J.l

is

t(e) . g(z - e) dS(e)

+~

is

K(z - e) . v.(e) dS(e) .

(3)

The unknowns correspond to the surface tractions t = (T n and surface velocities v .


Here, n is the surface normal pointing out of the particle and into the fluid. The symbol
9 denotes the Oseen tensor, the fundamental solution or Green's function of the Stokes
equations, which is given explicitly by

(4)
The integral representation given above has a long history dating back at least to Lorentz
[3], and numerical solution of the Stokes equations via discretization of this equation is
also well known [8]. The two integral terms on the RHS of the integral representation are
known as the single layer and double layer potentials [5], in analogy with the corresponding
terminology in electrostatics. The kernel of the double layer term is given by

K(z - e)

= -2n(e) . 1;(z -

e) ,

(5)

where 1;(z) is the stress field of 9(z)/87rp., or more explicitly,


1;(z)

3 zzz

= ---I
47r z 15

(6)

When the boundary velocities v.(e) are provided, the integral equation becomes a Fredholm integral equation of the first kind with the tractions t(e) as the unknown density.
(This is the most common type of problem encountered in particle hydrodynamics).
Application of this boundary integral formulation is not feasible for two important
class of hydrodynamic problems: geometries with sharp surface curvatures [9]; and manybody problems. Indeed, this observation is a well-known result from the theory of linear
operators and integral equations. The single layer integral operator is compact, and in
infinite-dimensional Hilbert spaces, the inverse of a compact operator is unbounded. Fredholm equations of the first kind are ill-posed, since very different inputs are mapped by the
compact operator to very similar outputs. This ill-posedness is not manifest in problems
where coarse meshes yield acceptable results, such as flow past a single sphere [8].
The solution of very large, dense systems of the kind obtained above by LU factorization is an expensive if not impossible proposition. On the other hand, our formulation of
the boundary integral equation (described in the next section) is amenable to fast iterative
solutions. Indeed, large systems on the order of 10,000 by 10,000 have been solved without
difficulty on minicomputers, and larger systems on the order of 1,000,000 by 1,000,000 are
easily solved on conventional supercomputers. Since the method allows tradeoffs between
processor workloads and memory usage, the future potential of this method is limited
mainly by economic constraints on the design of new architectures.
We tum our attention now to the derivation of the new BEM that yields a secondkind equation. From Odqvist's work [5] it is known that the double layer potential is
discontinuous at the surface. Specifically, the jump is exactly twice the local value of the
the double layer density so that at the point z on the surface,

171

lim

<-0+

{IIs K(z + en - {) . <p({) dS(e)}

::: <p(z) + t K(z - {) . <p({) dS({)


lim

<-0+

{IIs K(z ::: -<p(z)

n -

(7)

e) . <pee) dS(e)}

+ tK(z -

e) <p(e)dS(e) .

(8)

An integral representation based on just the double layer potential alone,

v(z) ::: t K(z - e) . <p({) dS(e) ,

(9)

after imposition of the no-slip boundary condition, v(z} ::: v.(z), z E S, yields the
boundary integral equation,

zES,

(10)

which is a Fredholm integral equation of the second kind. However, most flows of interest
do not fall into this category, because the double layer potential alone cannot represent
flows that exert a net hydrodynamic force or torque on the submerged body.
The key idea in the work of Power and Miranda [4] is that the single layer potential
may be replaced with the velocity fields of a point force and point torque, leading to an
integral representation of the form,

v(z)

_FH.

98(z) + -21 (TH x V). 98(:,:) + 1 K(z - {). <p(e) dS(e) .


7rJl7rJlIs

(11)

Here, FH and TH are the hydrodynamic force and torque on the particle, and <pee) represents the unknown double layer density which is to be found. The operator notation,

will be used for the last term in the boundary integral equation.
The three components of the force and torque introduce 6 new unknowns. However,
from the Fredholm alternative, the null space of 1 + IC is nontrivial; in fact, N(l + IC) :::
R(l + IC*).!. and since there are 6 'independent things' missing from the range (dimR(l +
K*).!. ::: dimR(l+K).!. ::: 6) there should be 6 linearly independent nontrivial null solutions
of the equation r.p + 1C(<p) ::: O. The null solutions are readily determined to be the six
independent rigid body motions of the particle surface. The removal of this degeneracy
provides a criterion for a unique solution. The method employed by Power and Miranda,

Fr ::: - f

rr : : -f (ei

ei .

<pee) dS(e) ::: -(e;, <p({) ,

e) . <pee) dS(e) = -(ei

e, <pee) ,

(12)
(13)

is but one possibility.


The interpretation using the dimensions of the various subspaces associated with 1 + IC
was first given by Karrila and Kim [6], along with a number of different completions

172

schemes for multi particle problems in unbounded and bounded flows. It seems natural
to call this new method, as it involves the idea of completing the deficient range of the
double layer operator, the Completed Double Layer Boundary Integral Equation Method.
Let us now consider mobility problems where the force and torque on each particle is
specified, and the unknown particle motions are to be determined. We write the boundary
integral equation as
U

+w X

g(z)

-F . -

87r jj

1 H
+ -(T
2

g(z)
x V). 87r jj

+ <p(z) + K(rp) ,

z E

S.

(14)

Here, U + w X z is the rigid-body motion to be determined. The equation is written for


just one particle but the approach is readily generalized to multiparticle settings.
The completion procedure now involves adding 6 extra conditions,

f ei . rp(e) dS(e)
f dS(e)

(15)

(w x z) = _ t(ei x z) f(ej X e) rp(e)dS(e)


j=1
f(ej X e) . (ej X e) dS(e)

(16)

The expressions appearing in the denominators are normalization factors. Insertion of


these 6 equations into the boundary integral equation yields
6

rp(i)(rp, rp(i)}

+ <p(Z) + K( rp)

FH.

j=1

<p(z)

or:

+ 1i(rp)

g(z)
87r jj

= FH. g(z)
87rp

+ ~(TH
2

V). g(z)
87r jj

+ ~(TH X V).
2

g(z) .
87rp

(17)

(18)

The operator 1i defined above is obtained from Wielandt deflations of the double layer
operator, K, since ej and ej X z, j = 1,2,3 are the 6 null eigenfunctions of the operator
1 +K.1
The term deflation refers to algorithms for moving eigenvalues to the origin. Wielandt's
deflation uses only the eigenvectors of the operator, and not those of the adjoint. In our
problem, at ..\ = -1, the eigenfunctions of K are known analytically (they are simply the
rigid body motion null solutions), while the eigenfunctions of K' are not).
We start with a matrix of rank 2 in a 3-dimensional space.

l+[(=A=

1 0)
( 1001
1 1 1

The null solution is [1, -1, Ojt, while the orthogonal complement of the range is spanned
by [1, -1, 1]. The eigenvalues of A are 0, 1 and 2. In this example also, the null solution
contains a nonzero projection in [R(A)jl. so we may complete the 'deficient operator' as
follows:

1
( 110)
001
1 1

(1)
-1

(1 -1 0) =

(l+a
I-a
1+ 0)
0

a 1

a 1

01

lWe reserve the notation rpU), j = 1, ... ,6 for the orthonormalized null eigenfunctions.

(rpU>, rp(k) =

fljle

i.e.,

173

I
xn+}=Mx n+b

Figure 2: Implementation of the Jacobi iteration on a parallel computer.


Now the new (matrix) operator 1 + K has eigenvalues at 2a, 1 and 2. If we choose
a = 1/2, the eigenvalue of K which was at -1 has been Wielandt deflated to O. The 'new'
K operator has eigenvalues at 0, 0 and 1. An iterative procedure is possible, after the
eigenvalue at +1 is also deflated [2] . Then in both the example and CDL-BIEM system
of equd.tions, there is now a unique solution, accessible by an iterative procedure.
At the beginning of this section, we pointed out that the BEM density function <p is
an aphysical quantity and alluded to the need for some sort of post-processing to extract
the physical results, such as the particle mobilities. However, we see that the Wielandt
deflations provide explicit formulae for converting <p to particle motions. Thus the postprocessing step is strictly speaking, not really a post-processing operation at all, but
rather forms an integral part of the solution procedure.
Asynchronous Iterations
Iterative algorithms are ideally suited for parallel machines. As shown in Figure 2, each
element of the product vector may be computed in parallel. However, additional factors
become important when we consider large (in number of unknowns) scale problems the system matrix can become so large that updating and collation of information from
the network of processors to reconstruct the unknown vector for subsequent redistribution
to all becomes a nontrivial task. In other words, communication of information between
processors becomes the bottleneck.
Interactions between boundary elements on neighboring particles are much stronger
than interactions between elements on distant particles. We can reduce the amount
of communication by replacing the synchronous iterative scheme described earlier (the
point-Jacobi iterations) with asynchronous iterations. The part of the solution vector
corresponding to a distant boundary element is most likely stored on a distant memory

174

2.2

r..

"::>
Ct:

::t
I:
co

2 .0

1.8

1.6

1.4

~- ----------------------L.....I'--L--'---'-....L......L......L-..L...-L.....J..........--'---'-....L......L......L-..L...-L.....JL.....I

10

12

Iteration

14

16

18 20

Figure 1: Convergence of iterations for tetrahedron, cube, octahedron and icosahedron.


bank. We receive this information on a less frequent basis, since access is likely to be
more difficult. But while this 'distant' information is on its way, we can perform many
local iterations. Preliminary work on this idea has shown that asynchronous iterations
significantly reduce the number of global iterations and the communication requirements.
Asynchronous iterations can introduce instabilities (nonconvergence) in a system stable to synchronous iterations. A matrix with spectral radius less than one, may have
a sub-block with a spectral radius greater than one. The local iterations on that subblock will diverge. In our case, the matrix is divided so that all boundary elements on
the same particle reside in the same block. Every submatrix corresponds to a particlecluster problem and must conform to the CDL-BIEM spectral theory. The submatrices
all have spectral radius less than one, so local iterations converge to a local fixed point.
Information arriving from distant regions will nudge these results and the overall system
eventually converges to the true global fixed point solution.
Illustrative Results
CDL-BIEM has been tested against established analytic, numerical and experimental
results for particle mobility and drag coefficients [2J. For example, in Figure 3, we show
the comparison of the method with experimental results for the sedimentation velocity
of various platonic solids. Despite the presence of sharp corners, the results compare
quite favorably with those found in the laboratory (horizontal dashed line) [101 . The
convergence of the iterative scheme as shown in this Figure is quite typical.
Conclusions
For a quite general range of particle and container geometries, we map the solution of
three-dimensional Stokes flow problems onto a boundary integral equation involving an

175

unknown surface dipole density. The governing integral equations may be recast as a
contraction mapping, by deflating analytically, the outermost eigenvalues of the double
layer operator. The resulting algorithm is well-suited for parallel computers since the
computations are readily divided into parallel tasks of equal size. For very large problems, algorithm and architecture must be considered together to produce optimal solution
strategies.
Acknowledgements
This material is based on work supported by the National Science Foundation Grant CTS9015377 and the American Chemical Society Petroleum Research Fund. Computational
facilities were provided by the Center for Parallel Computational Engineering - ParCE.
References
1. Happel, J.; and Brenner, H.: Low Reynolds Number Hydrodynamics, Martinus

Nijhoff, The Hague, 1983.


2. Kim, S.; and Karrila, S. J.: Microhydrodynamics: Principles and Selected Applications, Butterworth-Heinemann, Boston, 1990.
3. Lorentz, H. A.: A General Theorem Concerning the Motion of a Viscous Fluid and
a Few Consequences Derived from It, Vers!' Konig!. Akad. Wetensch. Arnst., Vo1.5,
pp. 168, 1896 (see also Collected Papers, Vol.4, pp. 7-14, Martinus Nijhoff, The
Hague, 1937).
4. Power, H.j and Miranda, G.: Second Kind Integral Equation Formulation of Stokes'
Flows Past a Particle of Arbitrary Shape, SIAM J. Appl. Math., Vo1.47, pp. 689698, 1987.
5. Odqvist,F.K.G.: Uber die Randwertaufgaben der Hydrodynamik zaher Fliissigkeiten (On the Boundary Value Problems in Hydrodynamics of Viscous Fluids), Math.
Z., Vo1.32, pp. 329-375, 1930.
6. Karrila, S. J.j and Kim, S.: Integral Equations of the Second Kind for Stokes Flow:
Direct Solution for Physical Variables Removal of Inherent Accuracy Limitations,
Chern. Engng. Commun., Vol.82, pp. 123-161, 1989.
7. Karrila, S. J.; Fuentes, Y. 0.; and Kim, S.: Parallel Computational Strategies for
Hydrodynamic Interactions between Rigid Particles of Arbitrary Shape in a Viscous
Fluid, J. Rheology, Vo1.33, pp. 913-947, 1989.
8. Youngren, G. K.; and Acrivos, A.: Stokes Flow Past a Particle of Arbitrary Shape:
A Numerical Method of Solution, J. Fluid Mech., Vol.69, pp. 377-403, 1975.
9. Tran-Cong, T.; and Phan-Thien, N.: Stokes Problems in Multiparticle Systems: A
Numerical Method for Arbitrary Flows, Physics of Fluids, Vol.A1, pp. 453-461,
1989.
10. Pettyjohn, E. S.j and Christiansen, E. B.: Effect of Particle Shape on Free-Settling
Rates of Isometric Particles, Chern. Engr. Prog., Vo1.44, pp. 157-172,1948.

Boundary Element Methods for Crack Problems


Masato Kimura
Research Institute for Mathematical Sciences
Kyoto University
Kitashirakawa, Sakyo Kyoto 606-01, JAPAN
Fax: +81-(0)75-753-7272

Introd uction

Recently, various types of numerical computation for crack problems are carried out by
using the finite element methods (e.g. [1]) and the boundary element methods (e.g.
[2]). According to linear elastic fracture mechanics, the displacement and the stress up
to a crack tip have singularities of O(rt) and O(r-t) respectively, where r denotes the
distance from the crack tip. It is well known that the coefficient of this singularity is
called a stress intensity factor, which plays an important role in engineering. When we
use the finite element methods or the boundary element methods for crack problems, the
approximation 01 these singularities is essential in accurate computation. Therefore, we
have to use singularity elements. One of the most popular singularity elements is the
quater-point element, which is constructed by using a quadratic isoparametric element.
(See [1], [2], for example.) The merit of the quarter-point element is that the singularities
are represented only by a few modifications of an isoparametric element, but there are
some limitations of the quarter-point element.
The aim of this research is to clarify, from a mathematical point of view, usefulness
and limitation of the singularity elements, espesially of the quater-point element. For this
purpose, it is enough to consider Laplace's equation in R2, because singularities of the
solution have the same property as that of linear elastic fracture mechanics.
There are some ways in applying the boundary element methods to crack problems; for
example, the method dividing the domain by the imaginary boundary including a crack
(the sub-region method), and the method to obtain the crack opening displacement using
the finite parts of divergent integrals in the sense of Hadamard, etc. In this research, we
restrict ourselves to the symmetric caces, and we reduce the crack problems to the mixed
boundary value problems of Laplace's equation, in order to focus on the discretization
errors from the use of the quarter-point element.

Mathematical Formulation of a Crack Problem

We suppose a domain n to satisfy as follows:


Let
be a simply connected bounded domain in R2, and let its boundary :=
be
piecewisely of class C 2 We assume that there is no cusp point on
Let 0 be a point in
and let P be on
We denote a open line segment OP by S, which represents a crack.

n,

t.

r.

r an

177

We assume that 5 crosses

at P with a positive angle. Then we define nand

by

0",6)

0'--'-----==5"--_ _

o ----""-----/ P

X-,

Figure 1: The Domain n with a Crack 5


Under the above hypothesis, we consider the following crack problem;
Problem 1 Find u E V such that

~~::
{-= -- =
8u+
8n

8uan

::~',
on 5,

where 9 E C(r) is a given function and

V .-

.-

{v Ivv has
E C (n) n CO(n u r) n BO(ny
a Ci-class extension up to 5
2

on each side of S.

U {P}

'

The next proposition holds for Problem 1.


Proposition 1 There exists a unique solution u of Problem 1, and the solution u is
expanded in a neighborhood of 0 as follows:
u(r , (J)

V'u(r,{J)

=(

au
-8
Xl

au

8X2

co

. {J

= L..
~ J{- . r t . cos L2'
j=O

j
1.
=I:-.Kj.r 2 - 1

cos(L. - 1)0
2

00

j=12

.(

- sm( - - 1){J
2

),

178

where K j (j = 0,1, .) are constants dependent on u, and where (r,8) denotes the polar
coordinate.
Especially in the case f! is a unit disk and 0 is its center, then

Kj

~ r" g(t)dt
21r 10
{
=
1

21f

-.; 10

jt

cos"2' g(t)dt

(j

= 0 ),

(j

= 1,

2, ... ).

Proposition 1 shows that the solution of Problem 1 has a singularity at O.


Definition 1 (Stress Intensity Factor) We call Kl defined in Proposition 1 a "stress
intensity factor" of u.

Our purpose is accurate computation of stress intensity factors by the boundary element
methods. But we can not obtain a boundary integral equation in usual ways because
of a crack. In the present paper, we assume that n is symmetrical with respect to S
and that 9 is an odd function whose variable is the length from Q along r. Under these
assumptions, we can reduce Problem 1 to the following mixed boundary value problem of
Laplace's equation. (See Figure 2.)
Problem 2 Find u such that

=0
ulr+ = f

in

n+

on

r+

ul L =0

on L

oul

on S,

6.u

an s

=0

where a function f E CO(r+) is given and satisfies f( Q)

= o.

rt
----I

P=lQi----:-----o---:--...J P

Figure 2: Reduction to the Mixed Bounda.ry Value Problem

179

Quarter-point Element

The boundary integral equation for Problem 2 is as follows:

au (y)du
faan+ E(x - Y)-a
II

where

an+. = r+ u

a E(x = c(x)u(x) + faan+ -a


lIy

y)u(y)du y

(x E

an+),

L uS and E(x) is the fundamental solution of Laplace's equation;

E(x)

= - 211" log Ixl

The unknown functions in the integral equation are uls and ~~ Ir+ U L' It is well known
that uls and ~~IL have singularities of O(r!) and O(r-!) respectively at 0, and we must
use singularity elements in order to carry out accurate computations.
One of the most popular singularity elements is a quarter-point element. (For details,
see [2j.) The quarter-point element with singularity of O(r!) is constructed by placing
the midpoint node of the quadratic isoparametric element at the quarter-point. This
element is used for approximating u\s. In order to approximate ~~ IL' we adopt the
singularity element which is generated by multiplication of 2~ to the above quarterpoint element. This has singularity of O(r-!). The merit of the quarter-point element
is that the singularities are represented only by a few modifications of the isoparametric
elements.

Q ---------------

4h

4i?: qU"'e<p~~n"'"""

'"

__ " ______P

quarter-point element

represents the collocation points on the quarter-point elements.

Figure 3: The Quarter-point Elements


The base functions of these quarter-point elements have the next expressions:
p(r)

q(r)

= aIr! + a2r

= blr-t + b2 + b3 r!

on S,
on L.

p( r) and q( r) represents quarter-point elements on Sand L respectively, where


unknown coefficients to be determined. Since we have

aj,

bj are

180

we can define two approximate values of the stress intensity factor by


siJp:=

siJq:= 2b 1 .

aI,

Numerical Examples

The auther gives some numerical examples under the following conditions:
n+ is the upper half of a unit disk. The elements used in discretization are linear elements
along r+ and quadratic elements on Su L. The two elements including 0 on Su L are the
quarter-point elements. A parameter h denotes the length of the quarter-point element.
The computations are carried out for the following cases:
1
h=- -

1111
-

20' 40' 60' 80' 160' 240' 320' 480'

J(t)

= cos 2'

3t

5t

7t

1
and-

640'

cos 2"' cos 2"' cos 2"' and smt.

Our numerical results are shown in Table 1 and Figure 4. Our computations were done
on the Apollo-workstation (DSP 10010) of Reserch Institute of Mathematical Sciences,
Kyoto Univercity.
Remark 1 Let n+ be the upper halJ oj a unit disk with 0 its center, and suppose
J(t)

jt

= cos 2"

(j = 1,3,5,),

then we have

on S,
on L.

On Accuracy of "sifp"

According to our numerical results of Table 1, we notice that the accuracy of siJp is worse
than that of siJq except for the case of J(t) = cos because u(r,O) is exactly represented
by the quarter-point element in this case.
In order to clarify the reason for it, we analyze it using the value K(f), which is
defined as follows. Suppose u( r, 0) is the solution of Problem 2 on S, and determine Cl
and C2 by solving the following equations

t,

p(h) = u(h,O)
p(4h) = u(4h,0),

where the function p(r) is the base function of the quarter-point element on Sand p(r)
clrt + C2r. Then we obtain
1
= 2h-'i.
{4u(h,0) 1

Cl

Here we define I< (f) by

K(f) :=

Cll

u(4h,0)}.

181

and the value of K(f) is a truncated value of Kl by using the quarter-point element.
Since we know
00

u(r,O)=

2:=

Kj.r t ,

j=lJ:odd

K (f) has the following expression;


K(f)

=
i=l,j:odd

According to the definition of K(f), sifp is expected not to be more accurate than
K(f), and we notice that K(f) is very close to sifp from Table 1. Figure 4 suggests that
when f(t) = cos 2mrt (m = 1, 2, 3 ), sifp converges to Kl in O(hm). We can explain
the reason of this behavior by the following remark.

Remark 2 Suppose
f(t)

= cos -2m+l
2-t

(m

= 0,

1, 2, ... ),

then

f(t) -

h
1/20
1/40
1/60
1/80
1/160
1/240
1/320
1/480
1/640

0.999742
0.999746
0.999746
0.999746
0.999746
0.999746
0.999747
0.999748
0.999747

f(t) -

h
1/20
1/40
1/60
1/80
1/160
1/240
1/320
1/480
1/640

cos(t/2) ,
sifq

sifp

0.999987
0.999990
0.999990
0.999990
0.999991
0.999991
0.999991
0.999992
0.999991

cos(3t/2) ,
sifq

-2.08654e-4
-4.1l452e-5
-1.98252e-5
-1.28779e-5
-5.47358e-6
-3.60658e-6
-2.56869e-6
-5.44658e-7
-5.49149e-7

exact sif

10
K(f)
1.000000
1.000000
1.000000
1.000000
1.000000
1.000000
1.000000
1.000000
1.000000

exact sif - 00
sifp
K(f)

-0.10012
-5.00183e-2
-3.33393e-2
-2.50027e-2
-1.25004e-2
-8.33361e-3
-6.25008e-3
-4.16556e-3
-3.12431e-3

-0.10000
-5.00000e-2
-3.33333e-2
-2.50000e-2
-1.25000e-2
-8.33333e-3
-6.25000e-3
-4.16666e-3
-3.12500e-3

II

II

f(t) = cos(5t/2) ,

h
1/20
1/40
1/60
1/80
1/160
1/240
1/320
1/480
1/640
f(t)

h
1/20
1/40
1/60
1/80
1/160
1/240
1/320
1/480
1/640

exact sif - 00

sifq

sifp

1.46831e-2
3.61221e-3
1.59728e-3
8.96383e-4
2.23290e4
9.9048ge-5
5.57884e-5
2.59216e-5
1.45830e-5

-3.34747e-2
-8.46083e-3
-3.77013e-3
-2.12309e-3
-5.31628e-4
-2.36471e-4
-1.32944e-4
-5.79763e-5
-3.26077e-5

= sin(t) ,

exact sif -

sifq

0.846068
0.847911
0.848283
0.848420
0.848559
0.848587
0.848597
0.848605
0.848607

Table 1:. Numerical Results for Problem 2

sifp

0.903815
0.875127
0.866138
0.861732
0.855223
0.853078
0.852009
0.850943
0.850411

K(f)
-3.50000e-2
-8.75000e-3
-3.88888e-3
-2.18750e-3
-5.46875e-4
-2.43055e-4
-1.36718e-4
-6.07638e-5
-3.41796e-5

084882636
K(f)
0.904499
0.875410
0.866291
0.861831
0.855259
0.853100
0.852026
0.850955
0.850422

182

.(5t/2),.lfq
co.(7t/l),.ifp
8(7t/2).eUq

x - 109 h. Y - log

Ierror I

1e-3 < x < 0.1. le1 < y < 1

The horizontal axis represents log Isifp - K 11 and log Isifq - K11, and the vertical axis represents log h,
where 4h is the length of quarter-point element as shown in Figure 3.

Figure 4: The Graph of sifp and sifq

Conclusions

According to our numerical results, sifq seems to be more accurate than sifp except for
the specipal cases. For accurate computation of the stress intensity factor, sifp is of no
practical use, becauseK(f) suggests that we can not expect sifp to be more accurate
than our numerical results as far as we use the quarter-point element.
In the future, it will become important to propose much accurate singularity elements
with mathematical estimates. For this purpose, we will have to use the proper singularity
elements baesd on the mathematical analysis.

183

References
[1] Barsoum, R. S. : On the Use of Isoparametric Finite Elements in Linear Fracture
Mechanics, Int. J. Num. Meth. Eng., vol. 10, p. 25-37, (1976).

[2] Blandford, C. E. et ~l.: Two-dimensional Stress Intensity Factor Compldations


Using the Boundary Element Method, Int. J. Num. Meth. Eng., vol. 17, p. 387-404,
(1981 ).

An Easy Adaptive Boundary Mesh for 2D Elastic Problem

E. KITA and N. KAMIYA


Department of Mechanical Engineering,
Nagoya University, Nagoya, 464-01 Japan

Summary
This paper concerns a new adaptive boundary mesh generation, specifically focusing on
computational efficiency. For this purpose, the following facilities are employed: relatively
fine initial mesh, simple interpolation error estimator and advanced h-adaptive mesh
refinement process which can remove the nodes with small contribution to error decrease.
In this paper, it is applied to the optimal mesh design of the two-dimensional elastic
problem for confirming its validity.

Introd uction
Boundary element method (BEM) has become very popular method for numerical analysis
and its application to CAD/CAE is studied eagerly. CAD/CAE system should be userfriendly, which can be operated easily even by inexperienced user. One of the important
problems for the such system is pre-processor, especially, mesh generation facility which
must be simple but reliable. Many studies have been devoted to construct adaptive
meshes including error estimation [1-9]. They, however, have some difficulties from the
practical viewpoint; low computational efficiency, inaccurate error estimation for rough
mesh and so on. Therefore, the authors presented a new adaptive mesh generation for
the two-dimensional potential problem [10]. This method has simple algorithm defined
as an optimization problem and can make the accurate mesh by only fewer iterations. In
this paper, the method is extended to the two-dimensional elastic problem, for which the
vector-valued quantaties are defined on each node.
The chapters of this paper are as follows: first, the basic concept of the present method
is explained; next, the optimal mesh generation problem is defined as the optimization

185

problem and related mesh design process is explained; finally, the method is applied to
the two-dimensional elastic problems.

Basic concept
The existing methods for construction of the adaptive boundary element meshes have the
following difficulties:
1. low computational efficiency,
2. inaccurate error estimation for rough mesh, and
3. total number of degrees of freedom (DOF) of the refined mesh may exceed hardware

constraint.
In order to overcome these difficulties, the present method will employ the following
facilities:
1. relatively fine initial mesh within the limit of DOF specified by operator,

2. simple interpolation error estimation, and


3. advanced h-adaptive mesh refinement process which can remove the nodes with
small contribution to decrease the error.
Fine initial mesh satisfies a requirement for rigorous error estimation, which often failed
in some previous works by simple estimation tool, and simultaneously saves number of
iterations.
When, however, fine initial mesh is adopted, the ordinary h-adaptive method yields too
fine final mesh whose total number of DOF may exceed hardware constraint. Therefore,
the present method will adopt a new h-type adaptive algorithm; computational error
is improved by dividing the elements with large error (like ordinary h-adaptive) and

simultaneously, the nodes with small contribution to decrease the error are removed.
This algorithm can make it possible to keep the total number of DOF below the specified
limit during mesh refinement process.

186

Optimal mesh generation problem


In this chapter, we define the object function, the design variables and the constraint
condition for the optimal mesh generation problem.

Object function

The object function is defined by an error estimator named the "error

area" and its definition is explained first.


Variation of nodal variable to be computed by linear boundary element is a piecewise linear as depicted P(s) with respect to the tangential coordinate s. More accurate
solution Q(s) is predicted by spline interpolation of the BEM solution (Fig.l). The error
area A is defined as follows:

11

P(s) - Q(s) I ds

(1)

where c is the global boundary of the object under consideration. By the way, on the
boundary where the variable is specified by the boundary condition, P(s )-Q(s) is thought
to disappear and therefore, the above equation is rewritten as follows:

A
where

La

1 P(s)

(2)

- Q(s) I ds

c is the boundary where the variable is specified.

For the two-dimensional elastic problem, one must consider the four variables: xdisplacement

'Ill>

y-displacement

'U2,

x-traction tl and y-traction

t2

and related four error

areas are defined: AUll A u2 , All and A t2 The object function, which is minimized, for
the two-dimensional elastic problem is defined as follows:
W(A u l>A..2 ,A/l,At2 )

= (~~:r + (~~:r + (~~:r + (~~:r

(3)

where the superscript 0 indicates for the initial mesh.

Design variables

Degrees of freedom (DOF) are taken as the design variables in this

scheme. The object function is minimized by adding DOF to the element with relatively
large error and by removing DOF from the element with relatively small error. We may
follow two approaches to vary DOF:
1. the approach based on the ordinary h-version, which varies number of elements or

nodes without variation of the order of interpolation function.

187

cu
~ L,..,-1DJ:l:l>~

>

BEM solution

Z
u
c

Spl ine interpolation

lL.

Tangential coordinate (s)

Figure 1: Error Area


2. the approach based on the ordinary p-version, which varies the order of interpolation
function under constant number of elements or nodes.
In this paper, the first will be employed because of its easy implement in BEM computation.
Constraint condition Since the present method employs relatively fine initial mesh,
too fine final mesh may be generated if DOF is simply added. Therefore, the maximum
number of DOF is specified in advance and is taken as the constraint condition:
(4)

Mesh design process


This chapter concerns mesh design process, which requires the variation of the object
function due to change of the design variables; i.e., mesh design sensitivities. We describe
how to calculate the sensitivities first and explain the related mesh design process.
Mesh design sensitivity In the mesh design process we will adopt here, the two design
sensitivities are required: the decrease of the object function by adding nodes (positive
sensitivity, 5Wn and the increase of the object function by removing nodes (negative
sensitivity, 5W;+).
Positive sensitivity

In Fig.2( a), an element AB is divided to minimize the object function

and a node C is added. Before adding the node C, the error area of the element AB is
the area surrounded by the line AB and the curve BCA. After adding the node C, the

188

..2
>
'"c:
o

Error area decrease

..2

Error area increase

'"
>

c:

zu

zu

c:

c:

::l

;:J

U.

U.

Tangential coordinate (5)


(a) Error area decrease

Tangential coordinate (5)


(b)Error area increase

Figure 2: Error area decrease and increase

error areas of the elements AC and CB are the area by the line AC and the curve CA,
and the area by the line CB and the curve BC, respectively. In this case, the error
area is decreased by the triangle ABC. One can calculate the error area decrease for the
four variables; 6A;l, 6A;2, 6A;J and 6A;;, respectively. The object function decreased by
dividing element i, positive sensitivity of the element i, is calculated as follows:

Negative sensitivity

In Fig.2(b), a node B is removed and a new element AC is generated

from the elements AB and BC. Before removing THE node B, the error areas of the
elements AB and BC are the area surrounded by the line AB and the curve BA, and the
area by the line BC and the curve CB, respectively. After removing the node B, the error
area of the element AC is the area by the line AC and the curve CBA. In this case, the
error area is increased by the triangle ACB. One calculates the error area increases for
the four variables; 6A~l' 6A~2' 6A~ and oAii, respectively. The object function increased
by removing node i, negative sensitivity of the node i, is calculated as follows:

Mesh refinement process In the mesh refinement process, firstly, one determines the
order to add a new node (Node Addition Order, NAO) and the. order to remove the
existing node (Node Removal Order, NRO). NAO is specified at each element by the

189

descending order of the positive sensitivity and NRO is specified at each node by the
ascending order of the positive sensitivity. Mesh refinement is carried out as follows:
1. BEM analysis is carried out and the object function is calculated.

2. If the magnitude of the object function for the present iteration step is greater than
one for the former iteration step, the process is terminated. If this is not the case,
the mesh refinement is carried out as follows:
(a) NAO and NRO are specified.
(b) One selects the element with the highest NAO (the element k) and the node
with the highest NRO (the node I).
(c) The object function of the next iteration step Wi+l is predicted by the one for
the present iteration step

Wi, oW;

and

owt

as follows:

(7)
(d) If Wi+! is smaller than Wi, the selected element is divided and the selected
node is removed, and the process goes to step ( a).
(e) If not so, mesh refinement is terminated and the process goes to step 3.
3. In this step, the nodes with small contribution to object function decrease are removed as follows:
(a) NRO is re-specified for the refined mesh.
(b) Nodes with higher NRO are removed while the error area increase by removing
nodes is smaller than the specified reference value, which is taken as 1% of the
object function in this paper.
4. Return to step 1.

Example and discussion


As a numerical example, the plane stress analysis of a square plate under uni-axial tension
is considered here(Fig.3). Poisson ratio is taken as

/I

= 0.3. We take two initial meshes:

one has 52 equal length elements and the other 84 equal length elements.

190

Distributed load P

Figure 3: Object under consideration

Initial mesh
52 D.O.F.

Final mesh
52 D.O.F.
2 iterations

B
(b)Final mesh

(a)Initial mesh
Figure 4: 52 elements mesh

D
Final mesh
84 D.O.F.
3 iterations

Initial mesh
84 D.O.F.
B

A
(b)Final mesh

(a)Initial mesh
Figure 5: 84 elements mesh

191

(x 10- 5 )
1

(xlO- 5 )
5

-.....

~O

{;

\
\
\

-1~==~==~======~
2

'
, :

'----t~7p

,-----

0
1

-1
2
_____ 1

t2/p

,?O
-2

(a)Initial function distribution

S""
-4

(b )Final function distribution

Final mesh
68 D.O.F.
1 iteration

Figure 7: Final mesh with 68 elements

0
1
0

,,

tdp

Figure 6: 84 elements mesh

o~

__________

~~c

8
Figure 8: Result of Kamiya et. al.

192

'al mes hes


I erent ophm
T a ble 1: Companson 0 f dft
Constant DO F
Less DOF
68
84
Degrees of freedom (DOF)
1
Iteration number
3
77
50
Object function decrease (%)

Firstly, optimal meshes are generated subject to constant number of DOF. For the 52
elements mesh, the initial and optimal meshes are shown in Fig.4. For the 84 elements
mesh, they are shown in Fig.5 and their stress distributions are shown in Fig.6.

In

the optimal meshes obtained, the boundary near stress concentration points A and B
is divided minutely to improve computational accuracy and on the other hand, some
nodes are removed from the remaining part of the boundary. Therefore, the function
distribution near the stress concentration points is improved excellently although total
number of elements has kept invariant (Fig.6). Iteration numbers for these examples are
2 and 3, respectively; convergence rate for the former is faster.
Next, the 84 elements mesh is considered as an initial one and is refined subject to
invariant or less number of elements. Final mesh is shown in Fig.7. In Table 1, DOF,
iteration number and the object function decrease in Fig.5{b) and Fig.7 are compared.
In the optimal mesh shown in Fig.7, the object function decrease is 50%, although DOF
is decreased by 16 (19%) but the decrease is smaller than that of the optimal mesh in
Fig.5(b).
Besides, the same example was considered by Kamiya et. al.[7] using different adaptive scheme shown. Their optimal mesh in Fig.8 is divided minutely near the stress
concentration points, which is quite similar to the present result.

Conclusion
In this paper, a new adaptive boundary mesh generation scheme was presented and applied
to the two-dimensional elastic problem. The following conclusions can be drawn:
1. Optimal mesh could be generated by only few iterations.

2. Optimal mesh could be generated subject to equal and less number of elements.

193

3. The numerical examples show that the present method is very effective for the twodimensional elastic problem, as well as the two-dimensional potential problem shown
in previous work [10].

References
[1] E. Alarcon and A. Reverter. p-Adaptive boundary elements. Int. J. Num. Meth.
Eng., 23:801-829, 1986.
[2] J. J. Rencis and R. L. Mullen. Soltion of elasticity problems by a self-adaptive mesh
refinement technique for boundary element computation. Int. J. Num. Meth. Eng.,
23:1509-1527, 1986.
[3] G. F. Carey and S. Kennon. Adaptive mesh redistribution for a boundary element
(panel) method. Int. J. Num. Meth. Eng., 24:2315-2325, 1987.
[4] E. Rank. Adaptive h-,p- and hp- versions for boundary integral element methods.
Int. J. Num. Meth. Eng., 28:1335-1349, 1989.
[5] M. Guiggiani. Error indicators for adaptive mesh refinement in the boundary element
method - A new approach. Int. J. Num. Meth. Eng., 29:1247-1269, 1990.
[6] P. Parreira. Further developments on error indicators and estimators for adaptive
hierarchical boundary elements. In C. A. Brebbia and J. J. Connor, editors, Froc.
Boundary Elements Xl, pages 95-103, Compo Mech. Pub. / Spring. Ver., 1989.
[7] N. Karniya and K. Kawaguchi. An attempt of adaptive boundary elements. In
M. Tanaka, C. A. Brebbia, and T. Honma, editors, Froc. Boundary Elements XII,
pages 527-538, Compo Mech. Pub. / Spring. Ver., 1990.
[8] W. Sun and N. G. Zamani. Adaptive mesh redistribution for the boundary element
in elastostatics. Compo Struct., 36(6):1081-1088, 1990.
[9] R. Yuuki, G.-Q. Cao, and S. Ueda. Error estimator for boundary element analysis
by using the direct regular method. In M. Tanaka., C. A. Brebbia, and T. Honma,
editors, Froc. Boundary Elements XII, pages 539-550, Compo Mech. Pub. / Spring.
Ver., 1990.
[10] E. Kita. and N. Ka.miya. A new ada.ptive boundary element refinement based on
simple algorithm. Meeh. Res. Comm., 18(4):177-186, 1991.

Boundary Element-Node Condensed Finite Element Method


for 'Three Dimensional Elastic Wave Propagation Problems
M. Kitahara
Faculty of Marine Science and Technology, Toka; University
3-20-1 Orido, Shimizu, Shizuoka 424, Japan
K. Nakagawa
Total System Institute
1-18-6 Asagaya-minami, Suginami-ku, Tokyo 166. Japan

INTRODUCTION
In the analysis of wave proragation problems for an elastic body with local inhomogeneity, it is
convenient to introduce a combined method of boundary elements and finite elements [1], [2].
In this case, the outer homogeneous region is treated by the boundary element (BE) method and
the inner compact inhomogeneous region is analyzed by the finite element (FE) method. One
problem of this method in the case of the three dimensional elastic wave propagation problem
is the dimension of the final system matrix which comes from the boundary element and finite
element formulations. In order to reduce the dimension of the BE-FE system, a method to
condense the interior nodes for the finite element region is introduced in this paper. The method
is based on the constraint modes and normal modes [3]. After showing how the boundary
elements and the finite elements are combined for the condensed model. we show one numerical
result in the three dimensional elastodynamics for the spherical inclusion.
BOUNDARY INTEGRAL EQUATIONS
The three dimensional elastic scattering problem is considered. as shown in Fig.1. In this figure.
u(I) is the incident wave and U(5) is the scattered wave. The boundary integral equation is
formulated for the homogeneous exterior domain D and the finite element formulation is adopted
for the inhomogeneous interior domain DC .
In the exterior domain D , the boundary integral equation for the total displacement field u (=
+ u(5 is reduced to the following well known form:

u(I)

on the boundary surface S . where Uij (x, y) is the fundamental solution for the time-harmonic
field and 'Iij (x, y) is the traction kermel

195

D
U(I)

Fig.1 Exterior domain D , interior domain De , and boundary S .


U(x y ) =1'J'
41rJ.l

[e-5+---1 /) {) {e- - eikTT

ikTT

'J

k}

/)Zi /)Zj

ikLT

}]

'

(2)
(3)

In what follows the time-harmonic factor exp( -iwt) is suppressed in all physical quantities. For
the discretization of Eq.(l), the linear shape function with four nodes is introduced

Nt = (1 - e)(1 - 11)/4 , N2 = (1 + e)(1- 11)/4 ,


N3 = (1 + e)(1 + 11)/4, N4 = (1 - e)(1 + 1})/4 ,

(4)

where -1 5 5 1 , -1 51) 5 1 . By using the rectangular element in Eq.(4), the boundary


integral equation (1) is discretized as

(5)
where

uf and tf are values of


Ht/ =
G't/ =

Uj

and tj at the f3 th node, and Ht/ and G':/ have the form

1
L1
L

qE,

s.

qE~

S.

1';j(x",y)NfJ dS1J+Cij (xJ5gfl,

(6a)

Uij(x",y)NpdSII

(6b)

In equation (6), the under bar in the index g. means no summation. The Ep is the set of elements
which have the f3 th node as a common node, and the summation is over the elements q which
have the f3 th node as a common node. More precise numerical procedures can be found in
Ref.[4]. For simplicity. we write Eq.(5) as
Hu = Gt + u(l) .

Operating

AG- 1

(7)

on both hand sides of Eq.(7), we get the boundary element equation


Z(B)U

r(B)

+g ,

(8)

196

where

(9a, b, c)
and A is the matrix that transforms the surface traction t to the so called equivalent nodal force
r(B) :

(10)

FINITE elEMENT EQUATIONS


The finite element formulations in the inhomogeneous domain DC lead to the following equation

(11)
where the time-harmonic factor of exp( -iwt) has been assumed. The mass matrix M . damping
matrix C . stiffness matrix K and equivalent nodal force r(F) have the forms
Mit =

E lD.
{ NaNpSijdV,

ct/ = alMijP + a2 K ,(/ '

K~P = E

{ (BTDB)ijP dV ,
f,(F) = E { Nat,dS.
ls,
lD.

(12a, b)
(12c, d)

In this expression. N a is the linear shape function with eight nodes. B is the strain-nodal displacement matrix. D is the stress-strain matrix. and the sign E means to take element matrix
and to make final system matrix. Rayleigh damping has been assumed in Eq.(12c). where al
and a2 are constants. For convenience. we rewrite Eq.(ll) as
z(F)u

(13a)

r(F) ,

where
z(F)

= -w 2 M -

In more detail. Eq.(13a} can be written


[

Z(F)
ss

Z~~)

Z(F)]
Sd

Z~~)

iwC + K .

[u]
S

Ud

[r(Fl]
S
ryl '

(13b)

(13c)

where the subscript S designates the boundary quantities and d indicates the domain quantities.
NODE CONDENSATION
In the huge FE domain of DC the degree of freedom of Ud in Eq.(13c). which is the displacement
field in DC becomes very large. This is critical in the actual computation in the case of the
three dimensional elastodynamic problem. An example of FE domain DC and the boundary S
of BE and FE domains is shown in Fig.2. A method to reduce the degree of freedom of Ud is
discussed in this section.

197

anb

"

(a)

(b)

Fig.2 An example of FE domain DC , and boundary S of BE and FE.


Constraint modes
In equation (13), we set M = 0 , C = 0 , and

reF) = 0 , and get

~~: ~~:] [:~] = [~]

(14)

The second equation of Eq.(14) can be written


(15)
From equation (15), the interior nodal displacement Ud is expressed by the boundary displacement
Us in the form
(C)

ud

_.;0;.
= - K-1K
dd
dSUS = ""cUS

(16a)

when:
(16b)

The meaning of matrix +C is the interior displacement fields in DC induced by the unit displacements on the boundary S . We call u~c} as constraint modes.
In the actual problem. the dimension of Kdd is very large and it is not convenient to take
Instead it is convenient to use Eq.(15) to make the constraint matrix +c In the actual
calculation, we set u~) = (0 0,1 (ith component) , 0 ... 0 ) and solve the following
equation

Kil.

K ddUd(i) = - K dSUS(i)

)
,= 1,2
, ... ,ns

(.

(17 )

for U~i) where ns is the degree of freedom of Us . Irons' Frontal Solution method [5] can be
advantageously used as a solution strategy. It is to be remarked that the matrix Kdd is same
for all cases of u~) In this sense it is convenient to take forward elimination once and store it.
From the obtained U~i) we can make the constraint matrix

""c .;0;.

[U(l)
d

U(2)
.. u(nS}j
d'
d

(18)

198

Normal modes
In equation (13), we put C
_w 2

= 0 and r<F) = 0, then get

[:~; :~:] [:~] + [~~; ~~:] [:~] = [~]

(19)

After the eigenvalue analysis for the condition

(20)

Us =0

onS.weget

(21)
where w? and A, are eigenvalUes and eigenvectors. respectively. These eigenvalues and eigenvectors have the following properties
AfM,14Aj = 0 (i ::/: i) , AfK,14Aj = 0
AfKddA, = w?(AfMddA.)
(i = j) .

(i ::/: i) ,

(22a,b)
(22c)

It is usual to normarize to be AfMddAi = 1. In this case. Eq.(22c) reduces to be


AfMddA, = w? .
Now we define the normal modes
n

u~n)

= LA,S; == +NS
,=1

as a linear combination of eigen modes. where

+N = [A1

+N

(23)

is the normal mode matrix given by

, A2 , ... , AnI,

(24)

and E is the unknown vector of n dimension. The number n in Eqs.(23) and (24) can be taken
up to the degree of freedom of Ud in DC In the actual problem. however. the n can be a
very small number compared with the degree of freedom of Ud The most effective case is the
slendar FE domain of DC as shown in Fig.2 (b) . In this case the motion can be approximated
by the first several eigenmodes in the engineering sense.
It is convenient to adopt the subspace method for the present case of eigenvalue analysis. In the
algorithm of the subspace method. it should be noted that the results of forward elimination of
the matrix Kdd have been already stored in the process to make the constraint mode +C
Condensations

For the displacement field

Ud

in the FE domain

DC

we now set

(25)

199

as a sum of the constraint and the normal mode. Equation (25) together with the trivial relation
n. = u. leads to the following transformation
(26a)
where T is the node condensation matrix and has the form

T:=

[;c

(26b)

+ON]

With the help of the matrix T we get the condensed stiffness matrix

K:= TT [Kss
KdS

KSd]
Kda

T= [Kss
_0],
0
Kdd

(27a)

where
(27b, c)
From equation (22c) Kdd is a diagonal matrix. The condensed mass matrix M takes the form

M = TT [MoSS

0] T= [~SS
MdS

Mdd

~Sd]
Mdd

'

(28a)

where we put MSd = MdS = 0 . and


Mss = Mss + +~Mdd+c, MSd = +rMdd+N ,
MdS = MId, Mdd = +~Mdd+N .

(28b, c)
(28d,e)

Matrices Mdd and Kdd have a simple relation from Eq.(22c)


(29)

wr

where 0 2 is a diagonal matrix with eigenvalues


as the diagonal terms. The condensed
damping matrix and the equivalent nodal force vector can be expressed as
(30)
(31a)
(31b, c)

Condensed FE equations
Equations (27).(28).(30). and (31) lead to the condensed finite element equations

-w 2 M [ ~ ] - iwe [ ~ ]

+ K [~ ]

= f(F)

(32)

200

This is the condensed version of the original FE equation of (13). It is convenient to write Eq.(32)
as
Z(F) [

~]

= rtF) ,

(33a)

where
(33b)

BE-CONDENSED FE COMBINED SYSTEM


Boundary element equations in Eq.(8) for the exterior homogeneous domain D take the form

(34)
on the boundary surface S . Node condensed finite element equations in Eq.(33) for the interior
inhomogeneous domain DC can be writtn

[ Z<[}
Z~~)

zV}] [us] _ [f~Fl + c)Ifyl]


E

Z~~l

C)~f~F)

(35)

by refering Eq.(31) for rtF) . In equation (35), Us is the nodal displacements on the boundary
surface Sand E is the node condensed unknown vector in DC .
The continuity conditions for nodal forces and displacements can be written as
f(B)

+ f~F) = 0,

Us

(in Eq.(34

= Us

(in Eq.(35

(36a, b)

on the boundary surface S . Equations (34) and (35) with continuity conditions in Eq.(36) lead
to the following BE-FE combined system

+ Z(B) zV2] [us] _ [g]


[ z~7Z(F)
Z(F)
E - 0
dS

dd

(37)

where f~F) = 0 has been assumed for the body force in FE domain.
NUMERICAL EXAMPLE
A spherical inclusion model. as shown in Fig.3, was chosen for the numerical test of Eq.(37). The
incident wave is the plane longitudinal wave travelling along the :1:3 axis:

(38)
Material prorerties are E/ E = 2 PIp = 1 and ii = /I = 1/4 where the bar denotes the
quantities of the inclusion. Fig.4 shows the boundary displacements on the spherical inclusion for
akL = 1.0 where a is the radius of the sphere. left hand side of Fig.4 is the radial displacement
component and right hand side is the circumferential displacement component. White circles are
BE-FE results and solid lines are analytical results which were obtained from the spherical wave
function expansion. In the present BE-FE combined method, boundary elements=56 ( (J direction

201

: 7. t/J direction: 8). nodes=64. total degree of freedom=192 for BEM ; finite elements=224 (
(J : 7. t/J : 8. r direction: 4 ). nodes=320. total degree offreedom=960 for fEM. were used.
2.5

I U r I / I u(I) I
fig.3 Spherical inclusion model.

2.5

I Us I / I u(I) I

Fig.4 Boundary displacements.

REfERENCES

1. Zienkiewicz.O.C.. Kelly,D.W. and Bettess.P.; Marriage a la mode-the best of both worlds


(finite elements and boundary integrals). Int. Symp. on Innovative Num. Analy. in Appl.
Eng. Sci. CETIM. Senlis. france. pp.19-26. 1971.
2. Kobayashi.S . Nishimura.N. and Mori.K.; Applications of boundary element-finite element
combined method to three-dimensional viscoelastodynamic problems. Boundary Elements.
Ed. Du Qinghua. Pergamon. pp.67-74. 1986.
3. Craig Jr . R.R., and Bampton. M.c'c'; Coupling of substructures for dynamic analyses.
AIAA J . VoU. pp.1313-1319. 1968.
4. Kitahara.M . Nakagawa.K. and Achenbach. J.D.; Boundary-integral equation method for
elastodynamic scattering by a compact inhomogeneity. Computational Mechanics. VoI,S.
pp.129-144. 1989.
5. Irons.B.M.: A frontal solution program. Int. J. Num. Meth. Eng . Vol.2, pp.5-32. 1970.

The Analytical Integration of Boundary Integrals


for Plate Bending
B. Knapke
Gh Kassel, FB 14, Lehrstuhl fUr Baustatik
Manchebergstr. 1, D3500 Kassel, Germany

Summary
Serious problems arise when Gaussian quadrature is employed for the integrations
required in the direct boundary element method (DBEM). You will run in a lot of
trouble if you are going to compute domain points near the boundary. These problems can be overcome if you carry out the integration analytically. The technique is
developped in this paper.

Introduction
One of the major problems in the application of the direct boundary element
method (DBEM) to elliptic partial differential equations of fourth order are the many
hypersingular kernels within. This distinguishes fourth order equations from second
order equations. In second order theory you can avoid these kernels altogether if you
use an appropriate formulation.
In the case of the plate equation the integral equation for the slope ~~ already
contains a hypersingular kernel (r- 2 ). For the calculation of the internal actions (bending moment and shear force) you need to know the second and the third derivatives of
the deflection w. In the representation formulas for these terms appear even stronger
singular kernels (r- 3 and r- 4 ) respectively.
The program BE-PLATTE, a professional program for the analysis and design of
concrete and steel plates, is based on the direct boundary element method and solves
the integral equations of the deflection wand the slope ~: by a collocation method.
The program is restricted to straight elements.
The design of the reinforcement within a plate structure depends on the distribution of the bending moments in the interior of the plate. For that purpose the program
lays after a choice of a mesh width by the user a mesh of stress points over the plate
and calculates in this points the bending moments. Intuitively you would expect that
a regular mesh is chosen. Unfortunately the generated mesh which depends on the
shape of the plate is in some cases very irregular. This leads to illegible reinforcement
layouts, s. Fig. 1
The reasons are to be found in the restrictions which the program has to obey in
generating the mesh. These restrictions are due to the fact that the program renders

203

...

+
+

I
I

,+
,

,
,+
,

+ +

+
+

++

,
\

.. ~

'.

+
,

++t

+:,

....

+I
'"
I
I'"

'"

~-+-+++

'*'

III

...

~
~

""-- . :PI
~~~:~
,

"

+ +

+ +

'*'

..

+
+
o

+ +

'*'
.t

+ ++

.#

I
I

-".

, ..

+ + -1-++--- --+--+

+ ++++

il

t')~

' . , ., _

+ +++
C")
,...

"'..(D
C")~

. . . ...

+- ;:: ~

+-'""t-

.',
I

+
,

,
I

C
41

41

SC

...

....o

+
I

_
D

..
....o

V
tID

+:

41

+,

,I

--

'*'

'"

-;t +

(')

++

+ +,+++- -+-;.t

++

+ +

'*'

+ + +++

'"

'*'

+ +

+:

'*'

~<'rN

'*'

+ +

++

""+

+ + +

++
+

+ +

'+
, ......

++t

..,L")

,+

\I-

+ + +

+ + +++++ 0 CZI .;F.o.. +

++

't
+

+ + +
+

++

+++

++

+~

+
+

+
+

.t

204
totally unreliable results for the internal actions in stress points near the boundary
because Gaussian quadrature is unsuitable for the integration of singular functions.
But for straight elements it is possible to evaluate the boundary integrals analytically.
The article is organized as follows: l is concerned with the integral equations of the
plate problem, in 2 we briefly mention the discretization and in 3 and 4 we dicuss
the analytical integration of the regular boundary integrals.

Integral equations for the plate problem


Of the four Cauchy-data of a plate two are prescribed and two are unknown so we
need two integral equations. For the deflection w we have the following representation
formula

c(x)w(x) = j[go(Y,x)V,,(w)(y) - :vgo(y,X)Mv(w)(y) - V,,(go(Y,x)w(y)


r

+ M,,(go(Y, x ~: (y)] ds y + j

go(Y, x)p(y) dDy

(J

+ L[go(YC, x)F(w )(yC) where

w(yC)F(go)(YC, XC)

(2)

1 2
go(Y,x) == 81rKr lnr

and for the slope ~~

Cl(X)W,l (x)

+ C2(X)W,2 (x) = j[gJ(Y'x)V,,(w)(y)


r

a
avg1(y,x)M,,(w)(y) -

V,,(gl(Y,X[w(y) - w(x)]

+ Mv(gJ(y, x ~:] ds y + j

gJ(Y,x)p(y) dDy

(J

+ L[gJ(yC,x)F(w)(yC) where

gl(Y,X)

[w(yC) - w(x)]F(gd(yC,x)]

(3)

= an x go(y,x) = 81rKr(1 + 2lnr)r n .

The sum 2: in (2) and (3) is to be taken over all comer points yC of the plate. If the
source point x coincides with one of the comer points, x = yC, then the contribution
of this point to the sum is neglected. The characteristic function in equation (2) is

xED,
I,
c(x)= { i1<p/21r, XET,
0,
x E DC

= R2 \

D.

205

And the characteristic functions Cl and Cz in equation (3 ) are defined as follows:

Their boundary values

. () Ll'P
v [1 . 2
. Z
J'P
Cl x =2;nl + 211" 2 sm 'Pnl +sm 'Pnz 'P~'

(4)

1 sin2'Pn2J~~
cz(x) =Ll'P nz + ~[sinZ'Pnl - -2
(5)
21r
21r
r
depend on the angles 'Pl und 'Pz which the two tangents at the boundary point form
with the xl-axis. At smooth points, where Ll'P = 'PI - 'P2 is 180=11" the brackets in
Equations (4) and (5) vanish,
[sin 2'PJ~; = sin 2'PI - sin 2'P2 = 0,
[sin 2 'PJ:; = sin 2 'PI - sin 2 'P2 = 0
so that at such points
CI(X)W,l (x)

+ C2(X)W,2 (x) = ~ :: (x).

If we place the point x on the boundary then Equations (2) and (3) constitute
two integral equations for the four boundary values

w,

ow

an'

Mn(w),

Vn(w)

and the corner terms

W(X C ),

F(xC )

of a Kirchhoff plate. As two of the four boundary functions and one of the two conjugated corner terms are determined by support and loading conditions the two integral
equations can be solved for the two unknown functions.

The discretization of the integral equations


For the approximate solution of the two integral equations (2) and (3) we use point
collocation. For that we choose K points Xk on the boundary, the collocation points.
We interpolate the boundary functions by polynomials which are defined piecewise
over subregions of the boundary called boundary elements.
To determine the unknown nodal values of the piecewise polynomial functions,
you don't demand any more that both integral equations are fulfilled for all x E r ,
but for the collocation points Xk, k = 1, .. ,K.
Now for the boundary integrals to exist, the piecewise polynomial functions must
be sufficiently smooth at the collocation points. The approximation of the deflection must be Cl at the node (the tangential derivative must be continuous) and the
approximation of the slope must be Co, i. e. continuous.
Therefore we approximate the deflection W along the boundary elements re by
Hermite-polynomials and the other boundary functions by piecewise linear polynomials.

206

Calculation of the boundary integrals

The values of the Cauchy principal values in the case of straight elements are given
in [1]
Now we turn to the regular boundary integrals. Over straight elements the product
of kernel functions times boundary layer can be integrated easily analytically if we
switch to a system of coordinates $, d That is centered at the source point x and whose
axis are parallel (s) and orthogonal (d) respectively to the element and its normal
vector, s. Fig. 2.

Fig.2 d-s-coordinate system

In that case the distance

varies only with s (which is up to a constant 'translational' term the arc-length s on


the element) and the normal and tangential derivative of the distance r become

r 11

d
= -,
r

rr

= -,
r
$

= fixed,

so that the kernel functions can be written as

1 2
go(y,x) =S7rKr lnr,

1
avgo(Y,x) =S7rKd(l +2lnr),
1
~
M II (go(y,x)) = - -S [(1 + 3v) + 2(1 + v)lnr + 2(1- v)"Z],
r

7r

VII (90(y, x)

=-

1
-S [2(1
7r

d
d3
+ v)"Z + 4(1- v)-:tl
r

207

and
1
gl(Y,X) = - n v 81rKd(1

8vg1(Y'x) = - n v 81rK[1

+ 2Inr) -

1
n r 81rKs(1

tP

Ids

+ 21nr + 2 r 2 ]- n r 41rK r2'


d

+ 2Inr),

ds 2

Mv(gl(Y, x =n v4- [(1 + v)2" + 2(1 - V)4]


11'
r
r
1
s
d2 s
+ n r-[(l + v)- - 2(1- v)-],
4
411'

r2

Vv(gl(Y,X =n v 411' [4(2v -

1\4 + 8(1- v)
d2

ds

+ n r 411'[2(1 + v) r 4 + 8(1 -

d4

(1

r6 -

+ v) r21

d3 s
v)-;:6]'

The polynomial basis functions defined on an element r; can be expressed in terms


of s as welL If we multiply these boundary layers with the kernel functions then we see
that all the boundary integrals can be expressed in terms of the following five integrals

lk

JIt:

=j

=j

It: _

Lit:

sit: In r ds

sk

2" ds
r

= _l_[sk+lIn r k+ 1
{[Ok'::: _tPJ k-2],
In r,

d1 arctan d'

ds'"
4

_
ds -

d[

r
d3Sk

--ds
r6

-1

2d - -;:r
p,

=j

'

k~2,

k = 1,
k = 0,

1)J k -

>2
, k,

2]

k = 1,

~[# + arctan iI,


k = 0,
k
2
{ ~[+ (k _1)K - ], k ~ 2,

d';.-'

4T"

H~

Mk

+ (k -

JH2]

+ 3KO],

d5 It:
{f[_d3~:_' +(k-l)Lk-2],
_s_
ds
=
_ d"
r8
ij?",

l[d"
6 r6

+ 5LO]

= 1,

k = 0,

k~2,

k = 1,
k = O.

All these integrals are indefinite integrals. To apply these formulas to an integral over
a given element ri=[Sa , Sb] you, naturally, have to substitute the integration limits as
in

s.
j s2In r ds = ~[s3In r - J41:!

etc.

'0
With these formulas an analytical integration of the regular boundary for any given
order n of polynomial shape functions can be carried out. This facilitates the implementation of an adaptive boundary element procedure for the plate equation.
These analytical integration formulas for the regular boundary integrals has been
implemented in the programm BE-PLATTE. A comparison with results obtained with
Gaussian quadrature shows that there is no measurable gain of accuracy in the boundary values of the plate bending solution. Different versions of the code that work with

208

analytical integration and numerical quadrature render nearly identical results. But
the gain in accuracy is remarkable when we turn to the calculations of the bending
moments at points close to the boundary. In Fig. 3 you see the distribution of the
principal moments for a slightly skew-angled plate. The stress-lines run parallel to the
x-axis and so at the upper and lower edge the stress-points come close to the boundary. Gaussian quadrature with ten points renders totally unreliable results. Fig. 4
shows the same results obtained with analytical integration. The gain in accuracy is
remarkable.
About the calculation of these results we want to talk next.

Internal actions
The bending moments
1'1'[11
Mn

= -K (W,ll +VWm),
= -K (wm +VW,ll ),

MI2

= -K (1 - V)W,j2'

are linear combinations of the second derivatives

so that, instead of formulating three separate representation formulas for the three
bending moments, it is more efficient to formulate only one representation formula for
a second order directional derivative

o 0
om on w.
Let the angles a, {3"

denote the orientation of the directions v, n, m


VI

ml

= cos a,
= cos {3,

V2 = sin a,

mz = sin{3,
n2

= sin,

We express the second order derivative I)!~n by the 'element intrinsic operators'
and
as follows

tr

02

omon = cos({3 -

02
a) cos(, - a) Ov 2

+ sin({3 + "I -

2a) O~~T

+ sin({3 -

a) sin(, - a)

tv

::2 .

The application of this operator to the kernel functions becomes particularly simple if
we note that
1
1 2
rrr = ;:rV'

209

Now to actually calculate the derivatives

we use the kernels

given in [I]. By making the substitutions

= -r"

rt = -r.,.

rm = -r"

r, = -r.,.

rn = -r"

rt = -r.,.

rm = -r.,.

r, = r"

rt = r"

rm = -r.,.

rp = r"

n=v m=v rn
n=v m=T
n=T

m=T rn

= -r"

(note that the two vectors T (= direction 1) and v (= direction 2) must form a righthanded system) we obtain the following results

~22 M,,(go(Y,x)) =4 1 2 ([(1 + v) + 2(1- v)r;](r; -

uV

~r

r;)

+4(1- v)r;r;)
t:o8:

uVuT

M,,(go(y,x =4 1 2 [2(1 + v) + 4(1- v)(r; - r;)]r"r.,.


~r

~22 M,,(go(y,x

uT

=4 1 2 ([1 + v) - 2(1- v)r;](r; - r;)


~r

+4(1 - v)r;r;),

!22

uV

v" (go(y, x =4 1 3 {2r,,([3 ~r

+ 4(1 -

v - 2(1- v)r;](r; - r;)

v)r;r;} - r.,.{12(1- v)r"r.,.(r; - r;)


- 4[3 - v - 2(1- v)r;]r"r.,.}},

210

02

li2 V... (gO (y,


uT

0:

uVuT

= - 43 {2rT{ -2[3 - v - 2(1 - v)r;]2rVrT


~r

- 4(1 - V)r~rT} + rv{ -8(1 - v)r;r;


- 2[3 - v - 2(1- v)r;](r; - r;)
+4(1 - v)r;(r; - 3r;)}}

VII(go(y,

x =4 1

~r

{2rT{[3 - V - 2(1 - v)r;](r; - r;)

+ 4(1- v)r;r;} + r...{4(1- v)rVrTCr; -

r;)

- 2[3 - V - 2(1 - v)r;J2rvrT

+ 4(1- v)(2r"r; - r;rT -

r~rT)}}'

Then we replace rv by ~ and rT by;. The integrals of the transformed kernels times
their boundary layers can also be expressed as linear combinations of the five integrals
Jk, Jk, Kk, Lk and Mk.
The calculation of the shear forces is done in a similiar way. The differential
operator

ol&mon
is expressed by the 'element intrinsic operators'

til

and

tT

as follows

o 0 0
03
&l om on = cos(.5 - a) cos(,B - a) cos(-y - a) ov 3

+ [cos(.5 - a) cos(,B - a) sin(-y - a)


+ cos(.5 - a) sin(,B - a) cosb - a)

+ sin(.5 - a) cos(,B - a) cos(-y + [cos(.5 - a) sin(,B - a) sine'Y -

a)] O~OT

a)
+ sin(.5 - a) cos(,B - a) sin(-y - a)

+ sin(.5 -

a) sin(,B - a) cosh - a)]

+ sin(8 -

a) sin(,B - a) sinh - a)

0:;T

::3'

where the angle 8 denotes the orientation of the direction 1. The following steps are
still to be done.

References

[lJ Abdel-Akher, A.; Hartley, G. A.:


Evaluation of Boundary Integrals for Plate Bending
Int. J. Numer. Methods Eng., vol. 28, 75 - 93 (1989).

211

[2J Abdel-Akher, A.; Hartley, G. A.:


Boundary Integration and Interpolation Procedures for Plate Bending
Int. J. Numer. Methods Eng., vol. 28, 1389 - 1408 (1989).

[3] Blum, H.:

Numerical Treatment of Corner and Crack Singularities


Saarbriicken, 1987.

(4] Hartmann, F.:


Introduction to Boundary Elements
Springer 1989.

[5] Hartmann, F.:


The Mathematical Foundation of Structural Mechanics
Springer 1985.

[6] Hartmann, F.:


ELASTOSTATICS
in: Brebbia, C.A.: Progress in BOUNDARY ELEMENT METHODS,
Pentech Press 1981.

[7] Hartmann, F.; Zotemantel, R:

The Direct Boundary Element Method in Plate Bending


Int. J. Numer. Methods Eng., vol. 23, 2049 - 2069 (1986).

[8] Melzer, H.; Rannacher, R:


Spannungskonzentrationen in Eckpunkten der Kirchhoffschen Platte
Bauingenieur 55 (1980) 181-184.

[9] Stern, M.:

A general boundary integral formulation


for the numerical solution of plate bending problems
Int. J. Solids Structures 15 (1979) pp. 769 - 782.

212

[101 Vable, M.:


An Algorithm based on the Boundary Element Method
for Problems in Engineering Mechanics
Int. J. Numer. Methods Eng., vol. 21, 1625 - 1640 (1985).

213

Q.I

.'". =
'-

<::

....

<II

'"

==

o:

-=~
I'C

'"r:

'"

=:

Q.,

0:

.~

.~

'"'"

<::

'"

.~

'-

Co

'-

<::

'"

>.

..:.
"0

........
Qj

.~

'" '"'"

<II

"0

.Q

.~

'-

....

'"

.~

214

...'"
c

C1/

E
0

0
.,...

.....

'"

Ol

.....C1/

.c

'" '"
Q..

.....

Q..

'+-

0
.,...

'"
'"

t:

.Q

"'0

....:::5 .....C1I

.0

.,...
~

.....

'"

:::5
U

..... '"W
In

-..

i
i:
~

iO

"'0

<I"

.....Ol

t..

Identification of Corrosion ProfIles in Computed Thennal


Tomography by a Volterra Integral Equation of the Second Kind
Fumio Kojima
Mechanical Engineering Department
Osaka Institute of Technology
5-16-1, Ohmiya, Asahi-ku, Osaka 535 JAPAN

Abstract
Motivated by the non-destructive evaluation in thermal testing of materials, a computational method for estimating corrosion shapes inside materials is considered. The problem
is formulated as a parameter estimation problem for a time-dependent boundary integral equation. The spline-based estimation technique is proposed using the trust region
method.

Introduction
Non-destructive evaluation method is very important technology for assessing the structural integrity of materials which are critically used in aerospace structures. Recently, due
to the recent advance in infrared imaging techniques, interest in using thermal properties
has grown for detecting structural flaws inside materials. Thermographic inspection of
aerospace materials has an advantage over conventional ultrasonic techniques, since the
heat can be injected with lasers, IR lamps or by hot air and the subsequent surface temperatures can be obtained from IR images without intimate contact with the structure.
It also permits one to scan a large area of the structure in a relatively short period of
time [1]. An approach investigated in this paper focuses on determination of the back
surface shape of a plate, where a portion of the back side of the plate has been removed
to simulate corrosion damage. Overall configulation of the infrared measurement system
is illustrated in Fig. 1. The time and spacial data obtained from an infrared measurement
system is compared to a solution for the temperature of the front surface of the remote
sample which is a function of the geometrical shape of the back surface. The shape of
the back surface is varied to minimize the least square error between the solution and the
measured data to give an estimate of the corrosion profiles in the back surface. Mathematically, this problem can be formulated as a geometrical heat inverse problem for a
thermal system. Previous efforts on this work has been studied by the author using the
finite element approach [2] [3] [4] [5]. In this paper, a new method is proposed for a shape
identification problem using the boundary integral equation method.

Integeral Equation Model for the System


Let G be a material shape which is a bounded domain in twcr dimensional Euclidean
space. The geometrical structure of corrosion shape is characterized by D. Then the
temperature distribution of the remote sample with damage is described by

ou

--.6.u=O in TxGID
at

(1)

with the initial and boundary conditions

u(to) = 0

on GID

(2)

216

MONITOR

Fig. 1 Infrared measurement system

au

au

=g on T x S
(3)
an = 0 on T x a(G/D)/S.
an
In the above equation, 9 denotes the heat source subjected to the surface of material S
and T denotes the time interval (0, t f) during which the process is observed. The surface
temperature recorded by the infrared camera is simply described by
-

y(t)

= u(t)ls

on T.

(4)

The non-destructive evaluation for this problem is to detect and characterize the corrosion shape D on the back surface of the material using the front surface data {y(t)}.
Throughout this paper, the boundary of G/ D is denoted by a(G/ D) which is assumed to
be simply connected CI-curve, i. e.,
a(G/D) = {~(o) = (6(O),M8))

10:::; 8:::; 1, ~ E C!([O, 1]) x C!([O, I])}

(5)

where C!([O, 1]) denotes the spa.ce of a continuously differentiable I-periodic functions on
[0,1]' We suppose that the front surface S is represented by
S = {~{O)IO :::;

: :; Bo}.

Using this para.metrization, any function defined on a(G/ D) is identified with a I-periodic
function on [0,1]. To simplify the notation, we define

>(t, 0) d;j u(t, ~(O,.)).


Then the thermal system (1-3) is reduced to the equivalent Volterra integral equation
defined on T x (0,1):

1
'2>(t,O)

rt

(lOr
10 dT 10 on(t,OiT,(J)>(T,(J)ds(q)

= ldT [r(t,OiT,(J)9(T,q)ds(q)
with

>(t,O)

= >(t, 1)

for 0 < t < tf

(6)

(7)

217

Corrosion Shope
Original Material

Fig. 2 Inspection boundary on the back surface

where

is the single-layer heat potential given by

r(t, 8; 7', a)
=

47r(t _ 7') exp

1(8) - (a W }
4{t _ 7')

= 0
and where

t> 7'

Otherwise

(8)

ar / an denotes the double-layer heat potential defined by


ar
an (t, 8; 1', a)

=
=

(n(a),(8) -(a ex { 1(8) - (qW}


87r(t - 7')2
P
4(t - 7')
Otherwise,
0

t>

l'

(9)

respectively. In the above equations, ds(a) and n(a) are a line element and an outer
normal derivative along the curve ( a) which are given by

s(q)

= ("
Jo

Id{da'a') Ida'

(d6(a) , _ dl(a) Ild(a)

n(a) =

dq

dO'

dO'

g(7',{8

{or0:58:580

(10)
(11)

The input {unction 9 is given by

g(T,8)

={

Otherwise

Spline-Based Parameter Estimation Technique


Let us set the inspection boundary I on the back surface. As shown in Fig. 2, the area I
is the part of the boundary

Ie 8(G/D).
We denote it by

(12)
It is a natural way to assume that

Ins = 0.

218

We shall parametrize the curve of inspection boundary I using the series of B-spline
functions,

T/(B,qM)

=L

iSM

qf!Bt,tAB)

where B~ denote B-spline functions of order d with the knot sequence AM and where

M) .
qiM = (M
qi,l' qi,2
The corrosion shape D is characterized by the parameter vector

M] .
qM = [M
ql , q2M, ... , qM
Those parameters qf1 must be chosen as

Thus the boundary o(O/D) is a function of the vector qM, i. e.,

As a result, the single and double layer heat potentials


explicitly depend on the
parameter vector qM. For convenience of discussions, we introduce the integral operators,

d2 10fl 0r
on (t,.,t:(ll[J,qM) jT,.,t:( O",q Mld~(O",qM)ld
dO"
0"

d;J. 10fl r (t, (M


B, q ); T, ~(M
CT, q) ) Id~(O",qM)1
dCT
dCT
respectively. Our numerical scheme for solving the integral equation (6) is based on the
spline collocation method [6]. Assuming some I-periodic B-spline mesh in s, the spline
being of order d, we approximate <fo{t, s) by

<fo(t,s) ~

<for(t)B~(s) ~

ipN(t).

B:(s)

(13)

'SN

where
ipN (t)

B:(s)

= [<fof (t), <fof (t) ... <fo~(t)r


[Bf,d(S), Bfd(S)'"

B~,d(S)r

Our scheme for the integral equation (6) is to find the time dependent coefficients ipN (t)
so as to satisfy the collocation equations

it

it

1
-(t,
Bi )
[K(t, Tj qM)( T)](f)i)dT
2 0 0
for

[L(t, Tj qM)g( T)](f).)dT

i= 1,2,,N.

(14)

219

The collocation equation (14) yields the dynamical system for the time dependent coefficient vector,

(15)
where eN, KN and GN denote the corresponding element matrices and vector given by

(fJ)
[eN] ij :::: BN
},d.
and

[GN(t,qM)L

= f[L(t, r; qM)g(r)J(O})dr,

respectively. We solve the above Volterra integral equation of the second kind numerically.
To this end, we present a numerical quadrature method. We set up some uniform mesh
with respect to time,
o :::: to < tl < t2 < ... < tn = t f
and we approximate the integral equation by a discrete sum,
i-I

Gf (qM) - L, WjKff (qM)<pf


j=O

for

i = 0, 1, ... ,n

(16)

where
and

Gf (qM) = GN(t;)(qM) = L, wjL N(t;, tji qM)g,


;=0

respectively. To obtain the numerical solution, we solve a linear equation for <Pi at each
time step in terms of the previously determined <Pj{j = 0,1,2"", i-I).
Thus the problem for estimating the corrosion shape is formulated as the parameter
estimation with respect to qM:
Given the data {Yd(ti)}, find the optimal solution q~ E Q which minimizes

(17)
subject to (16) where 7tN denotes an appropriate interpolation operator associated with measurement equipments and where QM is a compact subset of
RM.

Computational Algorithm by the Trust Region Method


Many optimization techniques for the problem (17) are readily applicable to our problem
since we can evaluate the gradient of the output least square function with respect to qM.
Inverse problems arising in non-destructive evaluations may exhibit a lack of continuous
dependence of the estimated parameters on the data. For that reason, there are many
cases where the estimated parameter sequence falls into one of local solutions of the

220
problem. In this paper, we propose the computational algorithm using the trust region
method. The iterative algorithm for finding cd! is stated as follows:
Algorithm: Trust Region Method
Let us choose WO and Ao be given.
Set JJ. E (0,1).
For k = 1,2,,
a) Compute IN(q(k l ) and construct the quadratic model

b) Solve the subproblem

and determine a feasible direction bq = Sk.


Compute
Pk = {IN(q<k l + Sk) - IN(q<k))}

/{</>(q(k l

+ Sk) _

IN(q(k l )}

d) If Pk > JJ. then


Oth,"!rwise
q(k+1l

= q(k).

e) Update the model if}k l , the scaling matrix Dk and A k .


The advantage of this method is its global convergence properties. Namely, the trust
region method makes it possible to obtain convergence to a optimal solution, even from
starting points that are far away from the solution. For detailed discussions and for the
implementation of that method, we refer to [7] and [8J, respectively.

References
[1] D. M. Heath, C. S. Welch and W. P. Winfree, "Quantitative thermal diffusivity
measurements of composites," in Review of Progress in Quantitative Nondestructive
Evaluation, Plenum Publ., Vol. 5B (1986), pp. 1125-1132.
[2] H. T. Banks and F. Kojima, "Approximation techniques for domain identification in
two-dimensional parabolic systems under boundary observations, " Proc. 20th IEEE
CDC Conference, Los Angeles, Dec. 9-11, (1987), pp.1411-1416.
[3] H. T. Banks and F. Kojima, "Boundary identification for 2-D parabolic systems
arising in thermal testing of materials," Proc. 21th IEEE CDC Conference, Austin,
TX, Dec. 7-9, (1988), pp.1678-1683.
[4] H. T. Banks and F. Kojima, "Boundary shape identification problems in twodimensional domains related to thermal test.ing of materials," Quart. Appl. Math.,
Vol. 47 (1989), pp. 273-293.

221

[5] H. T. Banks, F. Kojima and W. P. Winfree, "Boundary estimation problems arising


in thermal tomography," ICASE Report 89-44, NASA Langley Research Center,
November, 1989 j Inverse Problems, Vol. 6 (1990), pp. 897-921.

[6] D. N. Arnold and W. L. Wendland, "On the asymptotic convergence of collocation


methods," Math. Comp., Vol. 41 (1983), pp. 349-381.

[7] J.J. More, "Recent developments in algorithms and software for trust region methods," in Mathematical Programming: The State of the Art, Springer-Verlag, New
York, 1983.
[8] RG. Carter, "Safeguarding Hessian approximations in trust region algorithms,"
Technical Report TR 87-12, Department of Mathematical Sciences, Rice University,
1987.

A Mixed-Hybrid Variational Fonnulation for Coupling BEM


and FEM in Elastostatics
K. L. Leung. H. H. Chyou and P. B. Zavareh
United Technologies Corporation, USBf Co.,
Huntsville, Alabama 35807, U.SA.

Abstract
A mixed-hybrid variational formulation for boundary element (BE) and finite element (FE)
field coupling that provides accurate field values in the BE and FE subregions as well as
enforces displacement and traction continuity conditions at their interface is presented for
elastostatics problems. In this approach the mixed variational formulation for the FEM treats
the stresses and displacements in the FE subregion as independent variables while the hybrid
variational formulation for the BEM considers the displacements and tractions on the BE
boundary and displacements inside the BE subregion as independent variables. The
variational formulation ensures equilibrium and kinematic compatibility in the FE subregion
and provides accurate field values on the BE boundary as well as symmetric matrices. In this
paper, the mixed-hybrid variational formulation and numerical procedure of this BE/FE
coupling approach is described.

Introduction
It has been well recognized that the major strength of the FEM is its versatility in handling
various classes of engineering problems, including those involving nonlinearity and
inhomogeneity, while that of the BEM is its ease in data preparation, accuracy in predicting
stress concentrations, and its ability to handle problems involving the infinite domain.
Recently there have been numerous proposals of BEM/FEM coupling procedures with the

aim of exploiting the versatility of the FEM in handling nonlinear and inhomogeneous
problems while maintaining the accuracy of the BEM in areas of crack and stress
concentration. These coupling procedures provide an effective means for solving problems of
a localized nature directly within the global model. Two typical examples of such coupling
are as follows:
(a) Local stress concentration in a complex structure
In this case the BEM can be employed in the vicinity of a high stress gradient. whereas

the FEM can easily model the rest of the complex structure which may involve
inhomogeneity. anisotropy or nonlinearity.
(b) Local nonlinearity or inhomogeneity in the infinite domain

In this case the FEM can be used to handle the local nonlinearity or inhomogeneity,

whereas the BEM can be used to model the infinite domain which is assumed to be linear
and homogeneous.

223
One of the major difficulties of BE/FE coupling is the nonsymmetric and fully populated
nature of the stiffness matrix derived from a conventional BEM formulation. Assembling
such nonsymmetric BEM stiffness matrix with the symmetric and banded FEM stiffness
matrix will greatly degrade the computational efficiency in the solution of the global system
matrix equation. Another major difficulty of BEJFE coupling is the intrinsic difference
between the two solution techniques that causes some problem-dependent error of load
dispersion in the interface region, although global equilibrium is preserved [1].

One

important aspect of BE/FE coupling is the required simultaneous satisfaction of interface


traction equilibrium and compatibility. In addition, solving the localized high stress gradient
accurately in the BE subregion requires that accurate values of the field variables at the
interface be furnished by the FEM.

Cruse et al. [1], after attempting several different BEJFE coupling procedures that
involve non symmetric stiffness matrices, suggested that the development of a hybrid finite
element model is needed for a successful coupling. Zienkiewicz et. al. [2] and Brebbia et al.
[3] implemented a symmetrization technique for the BEM stiffness matrix that is based on
error minimization or energy consideration.

However, only displacement based FEM

formulation is employed in their coupling approaches. More recently, Dumont [4], Polizzotto
[5], Schnack et. al. [6], DeFigueiredo et al. [7] and other researchers have proposed various
variational formulations for the BEM.

Most variational formulations are based on

generalized principles of the type proposed by Hellinger-Reissner [8] and Wu-Washizu [9] in
which the field and boundary variables are taken to be independent of each other. One main
feature of variational formulations for the BEM is that the stiffness matrix obtained is
symmetric, which can be used for coupling with the FEM readily.

In the field of FEM, numerous researchers have developed various forms of variational
formulations to tackle different classes of problems. Sandhu and Pister [10] presented a
variational principle applicable to a wide class of linear coupled field problems. This
variational formulation approach has been particularly developed for the FEM. Sandhu and
Salaam [11] presented a variational formulation with nonhomogeneous boundary conditions
and jump discontinuities.

The Sandhu and Salaam [11] approach can generate dual-

complementary variational principles. Chyou [12] presented a mixed variational formulation


based on the potential and complementary energies of an elastic body to analyze laminated
composite structures.

In the present work, a mixed-hybrid variational formulation for coupling BEM and FEM
for elastostatics problems is presented. This approach, which is based on a variational
formulation characterized by a self-adjointness form with consistent boundary operators [1012], not only provides symmetric BEM stiffness matrix and exactly satisfies the displacement

and traction continuities at the interface but also furnishes accurate stress and displacement

224

field values by the FEM to interact with the BE subregion(s). The complete derivation and
mathematical foundation of the variational formulation applicable to general linear coupled
field static/dynamic problems in continuum mechanics is presented in [13]. The formulation
for elastostatics is summarized in this paper.

Basic Equations
Consider a boundary value problem expressed by

Au=f

on

=g

on

Cu
where

(1)

an

(2)

n is an open connected bounded region in an Euclidean space and an is the boundary

of n while A and C are linear bounded operators.

For linear elastostatics, an isotropic, linear elastic solid of volume

boundary r is considered. The basic field equations are given as follows:

enclosed by a

l. Equilibrium equations:

O'jjJ + fi = 0

(3)

2. Kinematic Relations:

1
ij = 1: ( ujJ + uj,i )

(4)

3. Constitutive equations:

O'ij = Eijld ~

(Sa)

Eij = Cijld O'ld

(Sb)

or

where the indicial notation and Cartesian reference frame are used. The quantities ui' ij' and
O'ij denote components of the displacement vector, the infmitesimal strain tensor, and the
symmetric Cauchy stress tensor respectively. E ijkl and <;jkl are the isothermal elasticity and
compliance tensors, respectively.

Variational Principle
For a positive linear bounded operator A defined on a dense set, Eqs. (1) and (2) have a
unique solution and if this solution exists, the functional
IT (u)

= (u, Au)o

- 2 ( u, f)o - (u, Cu-2g )00 - (u, (Cu)"-2g" )00

(6)

assumes its minimum value for that solution [11]. In the above equation, (a,b)o denotes

f a b dO for elastostatics and double primes denote jump discontinuities.


n

The self-adjointness form of the field equations take the form

225

a.

a. )

1
0
0
2 (~ik + ~jk
J

Eijld

-1

1
0
0
2(~~+~li~)

-1

{~} r}
::.

IJ

(7)

onO

where ~ik is the Kronecker's delta and (ota j ) denotes partial differentiation with respect to the
ith spatial coordinates. The consistent boundary operators for boundary conditions and jump
discontinuities are given as follows [l0-I2]:

1. Boundary conditions:

-nj uj
nj

aij

= -nj u"j

(8a)

= "tj

(8b)

with

2. Jump Discontinuity Conditions:

-(ni uj

)"

(ni a ij )"

= -g;;j

= ~~

(9a)

on

r!

(9b)

where ~j and tj are the specified boundary displacements and tractions respectively,

are surface elements embedded in

a across

r! and r!

which discontinuities occur and double primes

indicate jump discontinuities.


Thus, in view of Eqs. (1)-(9), the generalized governing functional for linear, coupled
field problems in linear elasticity can be written as [13]

II(Uj, Id' aij)

= - (ui' ajjJ >n

+ (Ejj' Ejjld fJd)n + (ajj' UjJ)n - 2 (aij' Ejj >n - 2 (Uj, fj)n

+ (Uj' ajj 11j - 2tj )r - (ajj' Uj 11j - 211j ~ >r + (llj' (ajj 11j)" - 2~j >r!
u

- (ajj' (Uj 11j)" - 2g1~j >ri

(10)

BEIFE Coupling
The domain

a is divided into two subregions of which one of them,

OB, is treated by the

BEM and the other, OF, by the FEM (Fig. 1). Similarly, r t is divided into rtBand rtF and ru
into ruB and ruF.

In the fmite element domain, OF,

aIJ.. and u.1 are chosen as the independent field variables

[13] while in the boundary element domain, OB, the independent field variables are taken to

be the displacement field variables, ~ , in the domain OB and the boundary variables, uj , and

226
the traction variables. Tj on the boundary ~ [7]. Along the interface between

o.D

and

o.F,

the displacement and traction continuities in a local sense are uj = ui and tj = r j respectively.

r=r'ur
a

I"'u . . . r=o
u

aU .

Fig. 1. BEM and FEM coupling.

Based on the chosen independent field variables. the assumptions [13] that
(i)

there are no jump discontinuities in stresses and displacements along ~ and

r:.

(ii) the constitutive relations Eij=C jjkl 0"1d are satisfied within o.F ,
(iii) the kinematic relations hold for o.B

(iv) glij = 0 and ~j =0

II.

If

(v)

Bi....B
=1u

0. ; and

ru

the differentiabilities of O"ij,j for

o.F

In

and Ujj for 0.8 are relaxed.

and the satisfaction of the displacement boundary conditions along

r: r:.
and

the governing

functional (10) can be written in the following form [13]:


fi l (ui' O"jj)

=2 (Ujj. Ejjld ukJ >oP + (O"jj. Ujnj

where ~ and

r:

(O"ij. qjld 0"1d

>r- - 2 (Uj. fj)OB

>oP - 2 (Uj. f;)oP - 2 (Uj. ~i >~ - (Uj. lj )~I


t

2 (iij. ~j)~ - (iij. r)~


t

- 2 (rj ur Uj )r- (11)

are the portions of the boundary along the interface of the BE/FE

subregions. Note that (Uj. O"jjj )OB vanishes when the singularities occurring at the source
points are excluded from ~ [7].
Variations of the functional fi l give the governing equations of the problem. As it is
pointed out in reference [13]. the above variational formulation is applicable to general linear
coupled field static/dynamic problems of BE/FE coupling.

227

The governing functional TIl for the BE/FE coupled domain can be written for
elastostatics problems as follows:
TIl

= J2 uij Ejjld uk,l dO - JIJij Cjjld IJId dO - J2 uj fj dO - J2 ui ti dr - Juj tj dr


d

JIJij u nj dO - J2 uJi dO - f 2 tj (u
j

0.8

j -

Ui) dr -

r'

0.&

f 2 ui tj dr - f ui ti dr
r:

r:

(12)

The governing functional, TIl' is used for the finite element and boundary element
discretization procedure. The discretization of the governing functional is given in the next
section.

Finite and Boundary Element Discretization


In the boundary element numerical method, the boundary of the BE subregion OR is

discretized into a series of mB segments called boundary elements. On the other hand, in the
finite element numerical method, the FE subregion OF is represented by a collection of mF
disjoint open subregions called finite elements.
approximated as
rn&

fil '"

The governing functional can be

rnP

L fi! + L fi~
e=l

(13)

e=l

Let the unknown field variables be approximated, in matrix form, as


for the FE subregion:
IJij =

'lie

IJe

in

OF

(14)

ind

(15)

in

(16)

Uj

= <Pc

Uj

= Ue

Se

for the BE subregion:

1j
_

ui
_

tj

= T e se
T

= clle

B
De

= TIe te

on
on

OB

rB
rB

on ~

(17)
(18)
(19)

where U and T are matrices composed of fundamental solutions of displacement and traction,
respectively, at a point in the Xi direction due to a unit load applied at a boundary point in the
Xj direction, and s denotes the forces acting at the boundary. Substituting Eqs. (14) through

228

(19) into Eq. (12) and neglecting the body forces, the discretized form of the governing
functional, iii' becomes [13]

where for the FE subregion,


F

KIlIl =

f <!leBEBT<!leT dOe

(21)

!t'e

f 'VeCV T dOe
e

(22)

!t'e

K~

f <!leie <ire

(23)

f <!leDVeT <ire

(24)

r teP
r icP

with B in Eq. (21) being the strain-displacement matrix, and for the BE subregion,

f UeTeT ~

K!

F~

=f

(25)

r te

cIIei e

(26)

KBts

~
KBu

f TleUeT ~

(27)

r ..

= f TleclleT ~
r ..
= f cIIe TleT ~

(28)

(29)

r icB

By partitioning the matrices in Eq. (20) into sub-matrices corresponding to interface and noninterface boundaries, denoted by subscripts I and D respectively, and taking variations on the
resulting equation leads to [13]

229

F
F
4Ku,.u,. 4Ku,.~
F

4~Iu,. 4KuuI I

~{

e=l

F
un

2FFn

F
-KII

F
Ur

2~

FT
-KII

-2~ncn -2K!nCI
F

-2K!l cn -2~c
I I

F
O"n

F
O"r

L=o

(30)

and

BT
BT
2KBss -2Kt"s -2K t.s
0
B
B
0
-2Kt"s 0
2Kt"u,.

L{
e=l
rn B

B
-2Kt.s

0
0
BT
BT
2Kt"u,. 2Kt,,~

2K~u,. 2K~~-KII

BT

BT

0
B

2K~u,.

BT

2Kt,,~ 2K~ul-K II

B
tn

tB
r

B
un

2FBn

B
ur

2FB
r

L=o

(31)

Assembling Eq. (31) for the BE subregion results in a matrix which can be partitioned as

(32)

where the matrices G and L contain sub-matrices corresponding to interface and noninterface boundaries. It should be noted that for the case when the whole domain is treated by
boundary elements only, that is, the FE subregion(s) and the interface boundaries are absent,
Eq. (32) reduces to the form presented in [7].
Defining R = G

T- l

L,

(33)

Eq. (32) can be written as

Kuuu

= 2FBt

where Kuu

=R

(34)

FR

or explicitly,

K~u,. K~~] { u:}


Ku1u,.

K~~

ur

= { 2F: }

2F I

By eliminating the variable, u:, Eq. (35) leads to

(35)

230
B

Hn UI = 2F I

B- 1

(36)

K;;,u.. Ku..u.. 2Fn

where
B

B- 1

Hn = Ku.u. - Ku.u.. K;;"u.. Ku..u.

(37)

After application of the displacement continuity and force equilibrium conditions at the
interface and assembly of Eq. (37) into Eq. (30), the system of equations can be stated as
F

4Ku..u..

4Ku..u.
P

4Ku.u.. 4Ku.u.+Hn
-Kn
P

-KuF

UI

B- 1

On

2~IGn - 2~{J,

- Ku.u.. Ku..u.. 2Fn

-2Kc;nGn -2K;,.nG,

pT

(38)

where u I = u I = u 1 and 0I =0I = or


After solving Eq. (38), u: can be recovered by using Eq. (35) as
B

un =

-1

Ku..u.. ( 2Fn - K;;"u. U I )

(39)

and t B can be calculated by using Eq. (32) as

= RuB

t B = G- 1 FR uB

(40)

(41)

The internal displacements and stresses in the BE subregion can be calculated by substituting
SB into Eq. (16) as [7]:
B

ui

0i

= UT RuB
= ST R uB

(42)
(43)

where S is a matrix composed of stresses obtained from the fundamental solutions of


displacement U.

Final Remarks
A variational formulation for linear coupled field elastostatics problems applicable to finite
and boundary elements coupling is presented. The derivation of the formulation is based on
the principle of the variational formulation with a self-adjointness form and consistent
boundary operators [10-13]. The formulation treats the hybrid field variables in the FE
subregion and the mixed variables within the domain and on the boundary of the BE
subregion as independent variables. The formulation can be easily extended to handle general
2D/3D dynamic and viscoelastic coupled field problems in continuum mechanics.

231

Acknowledgement
This research is supported by United Technologies Corporation, USBI Co. The authors are
grateful to Professor D.E. Beskos for his comments and suggestions. They are also indebted
to Dr. T.G.B. DeFigueiredo for her valuable discussion.

References
1.

Cruse T.A., Osias J.R and Wilson RB.: Boundary-Integral Equation Methodfor Elastic
Fracture Mechanics Analysis, Interim Report AFOSR-TR-76-0878, PWA-5406, Pratt &
Whitney Aircraft Group, United Technologies Corporation, East Hartford, Connecticut
(1976).

2.

Zienkiewicz O.C., Kelly D.W. and Bettess P.: The Coupling of the Finite Element
Method and Boundary Solution Procedures, Int. J. for Num. Meth. in Engng. 11 (1977)
355-375.

3.

Brebbia C.A., Telles J.C.F. and Wrobel L.C.: Boundary Element Techniques:- Theory
and Applications in Engineering, Springer-Verlag (1984).

4.

Dumont N.A.: The Hybrid Boundary Element Method in Elastostatics: Overview of the
Theory and Examples, Boundary Elements X, 1, ed. Brebbia C.A., Computational
Mechanics Publications, Southampton (1988) 43-57.

5.

Polizzotto C.: Variational Principles for Boundary Element Formulations in Structural


Mechanics, Boundary Elements X, I, ed. Brebbia C.A., Computational Mechanics
Publications, Southampton (1988) 19-32.

6.

Schnack E.: A HybridBEM Model,Int.J. Num. Meth. Engng. 24 (1987) 1015-1025.

7.

DeFigueiredo T.G.B. and Brebbia C.A.: A New Hybrid Displacement Variational


Formulation of BEM for Elastostatics, Advances in Boundary Elements, 1: Computations
and Fundamentals, eds. Brebbia C.A. and Connor 1.1., Computational Mechanics
Publications, Southampton (1989) 47-57.

8.

Reissner E.: On a Variational Theorem in Elasticity, J. Math. Phys. 29 (1950) 90-95.

9.

Washizu K.: Variational Methods in Elasticity and Plasticity, 3rd Edition, Pergamon
Press, Oxford, (1982).

10. Sandhu R.S. and Salaam U.: Variational Formulation of Linear Problems with
Nonhomogeneous Boundary Conditions and Internal Discontinuities, Compo Meth. Appl.
Mech. Engng. 7 (1975) 75-91.
11. Sandhu R S. and Pister K. S.: A Variational Principle for Linear Coupled Field
Problems in Continuum Mechanics,Int. J. Engng Sci. 8 (1970) 989-999.
12. Chyou H.H.: Variational Formulation and Finite Element Implementation of Pagano's
Theory of Laminated Plates, Ph.D. Dissertation, Ohio State University, Columbus, Ohio
(1989).
13. Chyou H.H., Leung K.L. and Zavareh P.B.: A Variational Formulation for Linear,
Coupled Field Problems in Continuum Mechanics for Finite and Boundary Elements
Interface Coupling, Internal !R&D Report, United Technologies Corporation, USBI Co.,
I:1untsville. Alabama (in progress).

Sensitivity Analysis of Steady-State Heat Conduction


Problems Using Boundary Element Method
T. Matsumoto, M. Tanaka and H. Hirata
Faculty of Engineering, Shinshu University,
500 Wakasato, Nagano 380, Japan

Introduction
In recent years, the boundary element method has been effectively applied to shape optimization
of continuum structures in various engineering problems. Shape optimization is one of the
fields in which the boundary element method is effectively applied. One of the reasons for this
is that the boundary element method can treat a problem by discretizing only the boundary in
most of the linear problems. In addition. the boundary element method can usually give us
more accurate boundary solutions than the methods based on the domain discretization.
Therefore, the boundary element method is suitable for such an optimization problem in which
the objective function and the constraints consist of the structural responses only on the
boundary or at small number of points in the domain.
In the iterative process of optimization, accurate and efficient computation of the sensitivities of
the system is important in order to obtain the optimal state successfully and economically. The
boundary integral design sensitivity formulation based on the implicit differentiation of the
boundary integral equation with respect to a design parameter has been successful in obtaining
accurately and efficiently the design sensitivities in elastostatic problems [1.2] and acoustic
problems [3]. In this method, the boundary integral equation or its discretized version is
differentiated implicitly with respect to a design parameter such as a shape design variable and
derived is a boundary integral equation or its discretized version relating the sensitivities to the
boundary responses.
In this paper, we present a boundary integral design sensitivity formulation for steady-state heat
conduction problems. There are many optimization and inverse problems on thermal
phenomena such as the optimal arrangement of heating or cooling lines in a die or a mold to
maintain a uniform temperature distribution on the cavity surface. The steady-state heat
conduction problem can be treated as a potential problem and the sensitivity formulation for it
based on the implicit differentiation of the boundary integral equation is the same as that for the
potential problem. Therefore, the previous work for the potential problem by the authors [4] is

233
valid also for the steady-state heat conduction problem. In this work, we present a treatment of
the convective boundary condition, which is inevitable in most of the heat conduction
problems, in addition to the previous work, and show some numerical examples. In the
formulation, every parameter which affects the system responses such as the boundary
condition itself, shape change, thermal conductivity, heat transfer coefficient and ambient
temperature can be adopted as a design variable. The boundary integral equation which can treat
the convective boundary condition is first regularized up to weakly singular order. Next, it is
differentiated with respect to a design variable and a boundary integral equation relating the
sensitivities of the boundary temperature and heat flux is derived. The sensitivity boundary
integral equation can also be solved like the original boundary integral equation by applying the
boundary conditions for the temperature and heat flux sensitivity and all the boundary
temperatures and heat fluxes obtained by the standard BEM analysis.

Formulation
The integral equation used in the BEM analysis of steady-state heat conduction problems for an
isotropic body is

8(y) +

f q*(x,y)6(x)dnx)

f O*(x,y)q(x)dnx),

YEQ

(I)

where Q is the domain of the body, r is its boundary, () is the temperature and q is the heat
flux. 0* is the fundamental solution of the Laplace equation and satisfies the following
diffe~ntial equation:

1
0 *.. = -8(x-y)
,II
It

(2)

where It is the thermal conductivity.

q* is the heat flux related to 0* and defined as


q* = -A.

ae*
an

(3)

For two-dimensional problems, 0* and q* are given as follows:

0* = ....=.LIn 1 ,
21CA. r

* _ _lor
q - 2trran

where r= I x-y I.
Equation (1) can be rearranged to the following form:

f
r

q*{O-6(y)}dr =

f
r

O*qdr

(4)

(5)

234

Equation (5) is a regularized form [5] of (1) and can be applied to a point y on the boundary. In
heat conduction problems, the treatment of the convective boundary condition is important. The
convective boundary condition is expressed in the form

(6)
where h is the heat transfer coefficient and 80 the ambient temperature. Denoting the rest part of
the boundary by r R and applying equation (6) to (5), we obtain

q*{8-8(y))dr =

8*qdr +

8*h(8-80 )dr

(7)

r
In equation (7), only the boundary temperature is the unknown function on

rc-

Discretization of equation (7) results in the following form of algebraic equation:

He = G q

(8)

Notice that in equation (8) the convective boundary condition has already been applied.
Let us assume that all the field functions in equation (7) depend on design variables, such as the
prescribed boundary temperature and heat-flux distribution, the heat transfer coefficient, the
ambient temperature, the thermal conductivity and the shape design parameters. Let us also
assume that the system responses are C I continuous with respect to the design variables.
Differentiating both sides of equation (7) when ye Q results in

-J

q*{8-8(y)dr -

q*{8-6(y)dr

O*qdr +

8*qdr

(9)

where a over-scribed dot ( . ) denotes the differentiation with respect to the design variable.
Equation (9) is applicable continuously from ye Q to ye r and all the boundary integrals
involved in equation (9) are at most weakly singular when y is located on r.

235
If we choose the design variable as a shape parameter, the derivatives with respect to the design
variable become as follows:
.*
(10)
= e~,I (i-.n
1
1

i/ = -A.8~. n.(i-Y)
t "
,1)

A.e~n.
.1 t

(11)

dr= (Xj),i - (Xj)J njnjldr

(12)

ni = (Xk),t nkntn j -

(13)

(Xk),j n k

Discretizing equation (9) results in the following algebraic equation:

He

Gq - ce + Dq

(14)

In order to calculate the design sensitivity coefficients of the boundary temperature and heat flux
by using equation (14), we must calculate the unknown boundary temperatures and heat fluxes
by applying the boundary conditions to equation (8). Notice here that the design sensitivities of
the temperatures or the heat fluxes are also known on the boundary where the values of the
temperatures or the heat fluxes are prescribed as the boundary conditions. In Figure 1 is shown
the computation procedure of design sensitivity coefficients using equations (8) and (14).

H.~~
He : G q~C. + Dq

Unknown

Boundary condition
fora and q

-."
""""'

Boundary condition

for a and q

e and q

Figure I Solution procedure for design sensitivity analysis.

Numerical Examples
The first example studied in this section is a thick cylinder with a uniform temperature
distribution along the inner wall and a uniform convective boundary condition along the outer

236
wall, as shown in Figure 2. A quarter region of the section of the cylinder is divided into 22
quadratic elements as shown in Figure 3. In this example, the design variables are chosen as the
radius of the inner wall a, thermal conductivity A, ambient temperature 00 and heat transfer
coefficient h. The theoretical solution of the temperature in the radial direction is given as
follows:
(J

= h (00-0) In(rla) + 0
h In(b/a)+Alb

(15)

where r is the distance of a point in the cylinder section from the center of the cylinder and (}a
the temperature on the inner wall. The theoretical solutions for the design sensitivity coefficients
of temperature can be obtained by differentiating equation (15) with respect to the design
variables. In Figures 4 - 7 are shown the results for the temperature sensitivities along AB in
Figure 3. All the present results show very good agreement with the theoretical solutions.

0=20C .
h=5W/(m2 K)

Figure 2 Section of cylinder with uniform temperature distribution on inner wall.

Figure 3 Boundary element discretization for quarter region of cylinder section.

237
xlO I

5. 0
I!I

4.0

Present
Theoretical

~ 3.0
<:I

~
"0

2.0

1.0

0.0

0.5

0.6

0.7

0.8
xI [m]

0.9
X

10 '

Figure 4 Temperature sensitivities along AB with respect to radius of inner wall.

10- 1
1.5
I!I

Present
Theoretical

1.0

..
~

"0

0.5

0.0

~~

o.s

__~~~~__~-L__~~__~~
0.6
0.7
0.8
0.9
X 10- 1
xI [m]

Figure 5 Temperature sensitivities along AB with respect to thermal conductivity.

238
)( 10.2

2.5
(!J

2.0

Present
Theoretical

1.5
~

~
"0

1.0

O.S

0.0

0.5

O.S

0.7

O.B
xI

0.9

em)

x 10"

Figure 6 Temperature sensitivities along AB with respect to ambient


temperature of outer wall.

10"

o.0

--r---,---.--..--..

(lj---.r--r--r---r-......

Present
Theorelical

(!J

-0.6

-1.2

:::G

1 -1.7
-2.3
-2.9

0.5

O.S

0.7

0.8
XI

0.9

em)

Figure 7 Temperature sensitivities along AB with respect to heat-transfer


coefficient of outer wall.

239
The next example is a doubly-walled thick cylinder as shown in Figure 8. Uniform temperature
distributions are given both on the inner and outer wall of the cylinder. The boundary of a
quarter region of the section of the cylinders divided into quadratic elements as shown in Figure

9. The theoretical solutions for the radial distribution of temperature are given as follows:

B,

= u,

(}2 =

Inr+

f3,

(16)

u 2 Inr+

f3 2

(17)

where subscripts 1 and 2 denote the inner and outer cylinder respectively. u" a 2 ,
are the constants given as follows:

u,u2 _
-

A2 ( ()a -())

and

f3 2

. .~-

(18)

...A'()a-BC>
-...~-..--.- - - -

(19)

A2 Ina+(A,-A 2 )lnh- A,lnc


-~-

f3,

Azlna+(A,-Az)lnb-A,lnc

f3 ,-- (A,-Az)lnh-A,lnc} ()a +Az())na

(20)

f3 ,-- (A,-Az)lnb+Azlna} Be-A, Baine

(21 )

Azlna+(A,-Az)lnb-A,lnc
Azlna+(A,-Az)lnb-A, Inc

Equations (16) - (21) contain all the parameters controlling the solution of this problem.
Therefore, the theoretical solutions of the sensitivity coefficients with respect to any design
variable can be derived by differentiating these equations with respect to one of such
parameters.

Figure 8 Section of doubly-walled thick cylinder with uniform


temperature distributions on inner and outer wall.

240
D
75 Nodes

Figure 9 Boundary element discretization for quarter region of


cylinder section.
3.0
Cl

2.0

~
c:

~
<Z:>
."

1.0

<0

0.0

IJJ

-1.0

- 2. 0
-3.0

C)

o 198.2%

Present(lmplicit Differentiation)
Forward Difference
Central Difference

6 10.1%

'f -8.7%

O.S 0 . 6 0 . 7 0 . 8 0.9 1 . 0 1.1 \.2 \.3 1. 4 \.S

x 10'\

Figure 10 Errors of temperature sensitivities along AC with respect


to radius of inner wall.

In Figure 10 are shown the computation errors of the results for temperature sensitivities along
AC with respect to the radius of the inner wall compared with those obtained by the finite
difference approach. The present approach gives accurate results in the entire range of edge AC.

241

Concluding Remarks

A boundary integral design sensitivity formulation has been presented for steady-state heat
conduction problems. First the convective boundary condition was incorporated into the
regularized boundary integral equation for steady-state heat conduction and then differentiated
with respect to the design variables such as boundary condition itself, shape parameter, thermal
conductivity, heat transfer coefficient and ambient temperature. The numerical examples for two
dimensional problems have shown that the present formulation can give us very accurate design
sensitivity coefficients of the boundary values.

References
1. J.H. Kane: Shape optimization utilizing a boundary element formulation. BETECH 86:
Proc. 2nd Boundary Element Technology Conf., Computational Mechanics Publications,
(1986),781-803.
2. M.R. Barone and R.J. Yang: Boundary integral equations for recovery of design
sensitivities in shape optimization. AIAA J., 2 (1988), 589-594.
3. T. Matsumoto, M. Tanaka and H. Egawa : Sensitivity analysis of frequency response in
acoustic problem using boundary element method. Proc. JSST 12th Symposium on
Calculations in Electrical and Electronics Engineering, (in Japanese), (1991), 433-437.
4. T. Matsumoto and M. Tanaka: Boundary integral formulations of design sensitivities for
potential problems. Engineering Analysis with Boundary Elements, 7(1990), 33-38.
5. H.D. Bui, B. Loret and M. Bonnet : Regularisation des equation integrales de
I 'elastostatique et de l'elastodinamique. Compte Rendu Acad. Sc. Paris, t.300, Serie 11(14),
(1985), 633-636.

A General Theory of Potential Aerodynamics with Applications


to Helicopter Rotor-Fuselage Interaction

L. Morino, M. Gennaretti and P. Petrocchi


Universita "La Sapienza", Rome, Italy

Introduction
This work presents a boundary-integral-equation methodology for the aerodynamic analysis of
unsteady, subsonic, compressible, potential flows around N bodies, each having arbitrary shape
and motion, with applications to rotor-fuselage configurations. Starting from the differential
problem, the unknown is expressed in terms of a boundary integral representation, which is
given by the sum of integrals over the surfaces of the N bodies. The novelty is that each
contribution is expressed in a frame of reference rigidly connected to the corresponding body.
This fact simplifies considerably both the derivation of the formulation and the evaluation of
the surface integrals (since in this case one deals with integrals over time-independent surfaces).
In this paper, first we present the boundary integral formulation for N surfaces including one
that surrounds the wake and is rigidly-connected to the undisturbed air.
In the following we present a brief outline of the development of integral equation methods (or
boundary-elements methods) in aerodynamics, for both airplanes and rotors, with emphasis on
the method introduced by Morino [1], which is the basis of this paper. For simplicity we deal
only with subsonic flows, with emphasis on helicopter applications (for supersonic and transonic
flows see [2]). In this case the non-linear terms in the differential equation may be neglected.
Boundary element methods are classified as 'indirect methods' and 'direct methods'. In indirect
methods, fundamental solutions (such as sources, doublets and vortices) are distributed on or
inside the surface of the aircraft, and the unknown intensity of the distributions is determined
by imposing the boundary condition on the surface of the body (typically that the air does
not penetrate the surface of the aircraft). In direct methods, the integral equation is obtained
from the use of the Green boundary integral representation and the unknown is directly that
of the differential equation (velocity or acceleration potential, or the pressure): hence the name
'direct'.
Consider first airplanes. The earliest examples of indirect formulations are probably due to
Prager [3] for two-dimensional flows and to Hess and Smith [4] for three-dimensional ones,
whereas the earliest example of direct formulations is probably that by Morino {II, who introduced a general integral equation for unsteady compressible potential flows. Numerical validation of the formulation are given in [5] and [6]. Next, consider potential compressible flows
around helicopter rotors. Probably the first work in this aerea is given in [7] (see also [2]). A
related but simpler approach for the analysis of unsteady, compressible potential flows was introduced by Gennaretti [8], a direct extension of [1] to the case of bodies in rotation and helicoidal

243

motion. The formulation has been extended to forward flight in [9J and to the case of N bodies
in [10]. Note that in the case of helicopter rotors in hover, the wake-transport problem is much
more important than for airplanes (for which the wake is typically assumed to be flat), because
the loads are strongly affected by the geometry of the wake used in the calculations. Therefore,
the use of an accurate wake geometry is essential to the solution of the problem: the wake may
be obtained, for instance, from theoretical (analytical or numerical) results or from experimental
data (this is referred to in the literature as "prescribed-wake" or "generalized-wake analysis")
or as an integral part of the problem ("free-wake analysis"). For a review of this subject see [2].
The present work is an extension of that presented in [10]; the novel aspect is the generality of
the formulation for the wake. First, we present the formulation. Then, we present the numerical
validation, both for the case of an isolated body (helicopter rotor, in hover and forward flight) and
for the case of rotor-fuselage aerodynamic interaction. The last item represents an application
where the methodology provides the greatest advantage with respect to the approaches most
commonly used. In fact, the problem of the rotor-fuselage aerodynamic interaction is typically
studied first by an independent aerodynamic analysis of the rotor and of the fuselage, and
then by determining the interactional effects by means of a first or higher order coupling (see
e.g., [11]-[15]). To the contrary, the present approach performs an aerodynamic analysis of the
complete rotor-fuselage configuration which automatically includes the interactional effects.
Boundary Integral Equation Formulation
Let (e, r) E (R 3 x R), denote an event in a space rigidly connected with the undisturbed air.
Consider N disjoint rigid bodies immersed in the air Vi C R3 (with nf:1V, = 0); the fluid volume
is then given by 'V
R3 \V, where V :::: uf:1V,. Let if> if>(e,r), with if> E C2 ('V x R), denote
the velocity potential. The linearized equation for the velocity potential is given by (for the
treatment of the non-linear terms, see [10])

- 0 if>

= 'V2~if> -

1 a2 if>
2"" -;-z
a"" vT

=0

(1)

where a oo is the undisturbed speed of sound.


In order to obtain an integral representation for this problem, it is convenient to recast it into an
equivalent infinite-space problem; to this end, we introduce a domain function E(e. r), defined
in (R3 x R) as E:::: 1 in 'V and E = 0 otherwise. Note that E =
E;, where E, :::: 1 in
R3 \V; and E,:::: 0 in Vi. Consider also the function ~(e,r) E(e,r)if>(e,r). Then, the original
problem (Eq. 1) may be replaced by the following infinite-space problem

nf:l

(2)
The solution to Eq. 2 is:

E(e.,r.)4>(e.,r.)::::

_+ lJJJ
a oo

fffi:

GVeE

V~if>d'Vdr +

fff[:

'V~. (if>'VeE)Gd'Vdr

{{{{OO GaEatPd'Vdr_+ {{{rJO ~(if>aE)Gd'VdT


-00
aT ar
aoo 1111- 00 aT
aT

(3)

where G is the solution of - 02G :::: See - e., T - T.), with G(e, (0) :::: G(e, (0) :::: 0 and G == 0 at
infinity (fundamental solution for the wave equation), or G(e-e., r-r.) = o(T-r.+fl/aoo)/41rfl,
where fl = lie - e.1I and S denotes the Dirac delta function.
Next, note that the integrals in Eq. 3 contain two Dirac delta functions, one stemming from the
derivatives of E, and the other from the expression for G. In order to perform the integrations,

244
note that E = nf:l Ei and nf:lVi = 0 imply VeE = Ef:l VeE; and oElor = Ef:l oE;jar.
Hence, it is possible to evaluate each i-th integral in a frame of reference fixed with the i-th
volume, Vi, since in such frame of reference Ei is not a function of time. Therefore, Eq. 3 may
be written as

E(e., T.)<,b(e., T.)

= 2: Ii

(4)

where Ii is the integral contribution of the i-th surface.


In order to evaluate Ii, let (x,t) denote an event in the space rigidly connected with Vi. Then,
consider the transformation relating such space to the space rigidly connected with the undisturbed air. Let this transformation be represented by

= e(x,t) = eo(t) + R(t)x

T=t

(5)

where eo denotes the air-space location of the body-space point x = 0, and R is an orthogonal isomorphism representing the rigid body rotation around eo. Since this is a rigid-body
transformation, we have J = 1, where J is the Jacobian of the transformation. Also, for any
differentiable function /, we have alar = alot - v Z ' V, where Vh = R-IVeh denotes the
body-space vector of the gradient of h, whereas v,,(x,t) = R-laVat is the body-space vector
of the velocity of a point x of the body space with respect to air space.
Next, recall that, for any hand g,

h(t) ]
/ 00 h(t)6[g(t)]dt=Lk , [-I'()I
9 t
-00

1=1.

=2:k /00

-00

h(t)
-I'()1 6(t-t k )dt
9 t

(6)

where tk are the roots of get) = 0, In our case, 9 = t - t. + lIe(x.,t.) - e(x,t)lIla oo and
iJ = 1+ (J,v/e aoo , where II = e(x.,t.)-e(x,t) and e = 111111. Here, for the sake of simplicity, we
limit ollrselves to the case in which the equation 9 = 0 has only one root, which we denote with
t. - 9. In other words, we assume that, at any time, all the points of the surface Si are inside
its sonic cylinder (subsonic flow). Combining the above expressions with Eq. 3 and performing
an integration by parts (with <,b = 0 at infinity) over the second and fourth integral on the right
hand side, one obtains I

Ii(X.,t.)

= ffl!: OVEiV<,bdVdt- Iff!:<,bVEj.VOdVdt

_+
aoo

ffff"" OdE; d<,b dVdt +

1111-

dt dt

00

+1111-00
a""

[[[[00

<,b dEi

dO dVdt
dt dt

where now dV denotes a volume element of the body space, whereas dl dt

I (J,v"IG=-.6(t-t.+9)=G o6(t-t.+9)
eaoo
4l1'e

G(x-x.,t-t.)= 1 + - with

= a I at -

-1

(7)
v z V, and

(8)

u= eliJl = ell + II' v leaool and Go = -1/4l1'U

Next, note that VG = VGo6(t- t. +9) + Go6(t - t. +9)V9 and oGlot


note that, for any a(x,t), if aE;jot 0 (i,e., if Vi is time independent),

Iff!: a V Ei 6 [t - to(x)]dVdt

fA

[a. n]t=to dS

= Go8(t- t. +9) Also,


(9)

lin performing the integration by parts over the last integral, we also assume that ;(~, -00) = ~(E. -00) = 0
and take advantage of the fact that V . v. = 0 (since v. is the velocity of a point of the body space with resped
to the undisturbed air. t.. a velocity corresponding to a rigid-body motion).

245

where Si

= aD;.

In particular, for any I(x,t),

ffff:

Id!iS[t - to(x)]d"Vdt =

-1fs; [Iv,. nlt=t. dS

(10)

Combining the above equation with Eq. 7 yields

Ii(e.,r.)

a4>
aGo +Goat/J (aD
v",' n)]
Ii [a-Go-t/J-aat a-+2-2n
n
n

++ 'J!s,
a oo

where

a oo

S,

j{

[t/J Go aa [v""
t

a/ail = a/an - v,,'n v,,' V.

n(l- v",' VD)]]

t=I.-8

t=t.-8

dS

dS

(11)

Equation 4, with Ii expressed by Eq. 11 is the boundary integral representation for the solution
to Eq. 1 with the initial conditions and infinity boundary conditions defined above. If
tends
to the boundary, Eq. 4 yields a compatibility condition that must be satisfied by the solution
of the problem. In our case at/J/an is known from the boundary conditions of impermeability
of the body surface, and therefore such compatibility condition is an integral equation for ,p on
the boundary. Once,p on the surface is known, ,p and hence v may be evaluated anywhere in
the field. Then, the pressure may be computed using the Bernoulli theorem.

e.

Helicopter Applications: Wake Contribution


In aeronautical applications, in particular in the case of helicopters, the problem is complicated
by the presence of the wake, which, for potential flows, manifests itself as a surface of discontinuity for ,p. The boundary conditions on the wake are v . n = Vw . n (where Vw is the velocity
of the wake) and Ap = O. These equations yield (see [2])

A at/J

an

=0

(12)

(13)
where Dw / Dt
velocity vw.

= a/at + Vw . V is the substantial derivative following a wake point Xw having

The contributions of all wake surfaces generated by the bodies in motion, may be determined
by the formulation derived in the preceding section. Nonetheless, they require a more accurate
analysis since, as it will be shown later, they give an additional line contribution.
For the sake of clarity, let us limit the discussion to the case of a single wake. In addition,
we assume, for simplicity, that the motion of the wake surface with respect to the air frame
of reference is negligible.' In this case, it is possible to treat the contribution of the wake by
including in the formulation of the preceding section an additional volume that surrounds the
wake. Under the above assumption, this volume is fixed in the air frame and hence v" = o.
Thus Eq. 7 yields

Iw(e.,r.)

00
-00

Iis:., (a4>'
-a
n

G-

<P

at;) dS dt
-a
n

(14)

2This is typically true for airplanes in maneuvering and helicopter rotors in forward flight. If this i. not true,
the formulation may be extended uaing the formulation for bodies tha.t move in arbitrary (not rigid-body) motion
(see [2J. Appendix 4). On the other hand, note that for airplanes in uniform translation and for helicopter rotors
in hover, the wake is fixed in the body frame. However, this formula.tion (used here for the hover results) is
relatively simple and is available in [2J; thus, for the sake of coincisenes., it is not discus.ed here.

246

where Slv is a closed surface that surrounds the open wake surface, Sw. In addition, in the air
frame of reference, Eq. 8 reduces to (note that now Go = -1/4lr!?)
G

= --1

4Jr!?

Ii ( t - t.

+ -!?

a""

= Go 5 ( t -

+ -!?

t.

a""

(15)

which is the fundamental solution of the wave equation. Performing the limit as the two sides
of Slv approach Sw and using the first wake boundary conditions, Eq. 12, one obtains

IW(e.,T.)=-/OO {[ AtP iJiJG dSdt


-00

JJ sw

(16)

Next, let us introduce a system of curvilinear coordinates, ,\ and 0<, over the surface of the wake.
Having assumed that the motion of the wake in the air frame of reference is negligible, we have
that the wake surface coincides with that swept by the trailing edge during its motion. Hence,
we may identify one of the two coordinates, with the arclength ,\ along the trailing edge. In
other words, the lines '\=constant identify the trailing-edge line at a certain time. Thus, the
other coordinate Ct identifies the trailing-edge location at a certain time. Hence, it is natural
to identify 0< with the time at which the trailing edge occupied that location. Next, consider
the covariant base vectors of the above coordinate system, a~ iJx/iJ'\ and a", = ax/iJCt. Note
that, because of the above choices made for the coordinates ,\ and 0<, a~ = iJx/iJ'\ is the unit
tangent to the trailing edge at time t = Ct, whereas a", = iJx/iJo< is the velocity, VTE, of the
trailing edge point, also at time t = Ct. In this coordinate system, we have dS
do< d'\ with
lIaa x a~11 II", where II" is the absolute value of the component of the velocity of the
trailing edge point in the direction v (which is normal to the trailing edge itself in the tangent
plane of the wake).

va =

= va

Next, note that the wake surface grows progressively in time because of the sweeping motion of
the trailing edge. Hence, because of the fact that, in compressible flows, a signal has a finite
speed of propagation, a point x. at time t. will be influenced by the not entire wake surface
already generated at time t., but only by the points that were present at the retarded time
t = t. - fJ. Aa shown later, this yields a line-integral contribution from the line separating the
influencing and the non-influencing portion of the wake (see Fig. 1). In order to demonstrate
this, it is important to note that, in Eq. 14, the complete wake surface generated up to time t.
is included (including the non-influencing portion of the wake). Therefore, the dpublet intensity
must be defined as one which is equal to A~(Ct,'\) on the portion of the wake already generated
at time t
t., and equal to zero otherwise (note that A~ is time independent because of Eq.
13). In other words, noting that, by definition of 0<, the coordinate O<TE(t) of the trailing-edge
at time t coincides with t, we have

AtP(Ct,'\,t)

= A~(Ct,~) H(t -

0<)

(17)

where H is the Heaviside function.


Combining Eqs. 15,16 and 17, we have, setting

Iw(e.,T.)

-1

00

-1-00 1>'f'"1 10fa. A~


00

f~' fa. A~ H(t -

-001~1

= _ f~2

fa. A~ H

-00

1~)

10

= t.,

H(t - 0<) aiJGo Ii


n

Ct) Go iJiJ!?
n

(t. - ...!.. - 0<)

1>.) 10
+/"" f~' fa. A~ Ii(t -

Ct.

aoo

5 (t - t. +...!..) ~
aoo

iJiJGo

0<) Go iJa!? Ii
n

(t - t. +...!..)
II" dec d'\ dt
aoo

tlv

a""

dCt

d~ dt

dCt d'\

(t - t. +...!..)
~ do< d'\ dt
aoo a""

(18)

247

where Al and A2 denote the extremes of the trailing edge.


In order to evaluate the last integral it is convenient to integrate first with respect to t. Then,
the integration with respect to 0< is performed by using Eq. 6 with g(o<) = t. - e(o<)/aoo - 0<.
This yields

Iw(e.,T.) = where

0<0

aGo dS + 1),' [A
1
ae llcp Go - 1
M -8
h~Sw. 1l"-8
n
),.
+ , n a""

tiv ]

<>=<>0

dA

(19)

is defined by g(o<o} = 0, and


1+M

= 8g = 1 + e' VTE
80<

aooe

>0

(20)

where 1 + M, is known as the Doppler factor. In addition, Sw is the portion of the wake
influencing the potential at the point x. at time t.. The last integral is evaluated for 0< = 0<0,
i.e., on the line
which separates the two regions of the wake.

.cw

For the validation of the formulation, both a prescribed-wake geometry and a free-wake geometry
analysis has been performed. In the first case the wake geometry is assumed to be either the
surface swept by the trailing-edge during its motion ("undeforming wake") or the one obtained
from experimental results, whereas, in the second case, the wake geometry is computed at every
time step.
Numerical Results and Comments
New numerical results for an isolated rotor in hover in steady state, and in forward flight, and for
a rotor-fuselage configuration are presented here. They have been obtained by using a zero-th
order boundary element numerical discretization. Hovering rotor results have been compared
with the experimental data given in [16], whereas the avancing rotor results have been compared
with th" numerical results given in [17]. For the rotor-fuselage configuration, an analysis of the
influence of the presence of the rotor on the pressure distribution over the fuselage surface, has
been performed.
Consider first the problem of an isolated hovering rotor. The case examined is that studied
experimentally in [16J: the rotor has two blades with a NACA 0012 section, radius of 45 in,
constant chord of7.5 in, no twist and a collective pitch of SO Figure 2 depicts the vertical section
with a plane tangent to the trailing edge, of the free-wake geometry computed by our method
for the incompressible case. Then, this wake geometry has been used in a prescribed-wake
analysis for the compressible case (tip Mach number equal to 0.727); the computed sectional
lift coefficients CI 3 are shown in Fig.3 where they are compared with the experimental data
and further numerical results obtained using both a helicoidal wake and a wake defined using
experimental data given in [16J. These results demonstrate that the use of a free-wake analysis
yields better results than the prescribed wake analysis.
Next, consider the problem of an isolated rotor in forward flight. For the one-bladed rotor case
suggested in [17J, both the free-wake and the undeformed-wake analysis has been performed both
for an incompressible case and for a compressible case (details on the rotor geometry are available
in [17]). Figure 4 (showing thrust coefficient CT tiS azimuth angle 0< for the incompressible case)
shows that, in this case, the free-wake analysis does not cause a very sensible change in the
results. Figures 5 and 6 depict the comparison between our undeformed-wake results and those
obtained in [17J using a lifting surface approach, for incompressible and compressible flows,
"The sectional lift coefficient is equal to the sectional lift divided by the product of the tip dynamic pressure
times the chord.

248

respectively. In addition, an articulated two-bladed untwisted rotor with a NACA 0012 section
in incompressible flow, has been studied. Our numerical results have been compared with the
data given by [18J. Figures 7,89, and 10 depicts, at four different blade positions, the comparison
between the numerical and experimental values of the differential pressure at the blade section
at r/ R = 0.75.
Finally, consider the problem of a rotor-fuselage configuration in forward flight. We analyzed an
axisymmetric fuselage with a two-bladed rotor at different distances from it. The fuselage shape
(see Fig. 11 where some streamlines are depicted for the isolated translating fuselage) may be
obtained [19] by suitable superposition of a source and two sinks aligned with the undisturbed
flow direction. Therefore, for the case of isolated translating fuselage, the exact solution of
the potential and hence of the pressure distribution is available. Figure 12 shows the pressure
distribution over the fuselage for the rotor at different distances from it. For the case of absence
of rotor (d = 00) the comparison between the exact pressure coefficient distribution and that
evaluated numerically by our formulation is also shown. In this case, the results are in good
agreement, even though the paneling used is not very dense. This analysis indicates that the
influence of the presence of the rotor on the aerodynamic loads on the fuselage surface is quite
strong and therefore further analysis is warranted.

Acknowledgements: This work was partially supported by ClRA grant no. 90.3014. VE.045,
to the University "La Sapienza" .
References
1. Morino, L.:

"A General Theory of Unsteady Compressible Potential Aerodynamics,"


NASA CR-2464, (1974).

2. Morino, L., and Tseng, K.: A General Theory of Unsteady Compressible Potential Flows
wi~h Applications to Aeroplanes and Rotors," Eds.: P. K. Banerjee and L. Morino, Developments in Boundary Element Methods, Volume 6: Nonlinear Problems of Fluid Dynamics,
Elsevier Applied Science Publishers, Barking, UK, (1990).
3. Prager, W.: "Die Druckverteilung an Koerpern in ebener Potentialstroemung", Physik.
Zeitscbr., Vol. XXIX, (1928), p. 865.
4. Hess, J., and Smith, A. M. 0.: "Calculation of Non-Lifting Potential Flow About Arbitrary
Three-Dimensional Bodies," Journal of Ship Research, Vol. 8, No.2, (1964), pp. 22-44.
5. Morino, L., and Kuo, C.-C: "Subsonic Potential Aerodynamics for Complex Configurations: A General Theory," AAlA Journal, Vol. 12, No.2, (1974), pp. 191-197.
6. Morino, L., Chen, L. T., and Suciu, E. 0.: "Steady and Oscillatory Subsonic and Supersonic Aerodynamics Around Complex Configurations," AlAA Journal, Vol. 13, No.3,
(1975), pp. 368-374.
7. Morino, L., Freedman, M. I., and Tseng, K.: "Unsteady Three-Dimensional Compressible
Potential Aerodynamics of Helicopter Rotors in Hover and Forward Flight - A Boundary Element Formulation," American Helicopter Society National Specialist Meeting on
Aerodynamics and Aeroacoustics, Arlington, TX, USA, (1987).
8. Gennaretti, M.: "Metodo delle Equazioni Integrali al Contorno per Analisi Aerodinamica e
Aeroacustica di Rotori in Flussi Potenziali, Compressibili, Non Stazionari," Tesi di Laurea,
School of Engineering, University of Rome "La Sapienza", Rome, Italy (in Italian), (1989).

249

9. Gennaretti, M., Macina, 0., and, Morino, L.: "A New Integral Equation for Potential
Compressible Aerodynamics of Rotors in Forward Flight," Proceedings of the International
Specialist's Meeting on Rotorcraft Basic Research, Atlanta, Georgia, (1991).
10. Morino, L., Gennaretti, M., and Macina, 0.: Toward the Unification of Unsteady Compressible Aerodynamics and Aeroacoustics," Proceedings of the VI Convegno Nazionale di
Meccanica Computazionale, Universita di Brescia, Brescia, Italy, (1991).
11. Clark, D. R., and Maskew, B.: "Study for Prediction of Rotor/Wake/Fuselage Interference," NASA CR-177340, (1985).
12. Clark, D. R., and Maskew, B.: "Calculation of Unsteady Rotor Blade Loads and
Blade/Fuselage Interference" , Proceedings of the Second International Conference on Rotorcraft Basic Research, College Park, MD, (1988).
13. Egolf, T. A., and Lorber, P. F.: "An Unsteady Rotor/Fuselage Interaction Method,"
Presented at the AHS Specialist's Meeting on Aerodynamics and Aeroacoustics, Arlington,
TX, (1987).
14. Bi, N., and Leishman, J. G.: "Experimental Study of Aerodynamic Interaction between a
Rotor and a Fuselage," AIAA Paper 89-2211, Presented at the AIAA 7th Applied Aerodynamics Conference, Seattle, WA, (1989).
15. Crouse, G. L., Leishman, J. G., and Bi, N.: Theoretical and Experimental Study of
Unsteady Rotor/Body Aerodynamic Interactions ," Presented at the 46th Annual Forum
of the American Helicopter Society, Washington D.C., (1990).
16. Caradonna, F. X., and Thng, C.: "Experimental and Analytical Studies of a Model Helicopter Rotor in Hover," USAAVRADCOM TR-81-A-23, (1981).
17. Tai, H., and Runyan, H. L.: "Application of a Lifting Surface Theory for a Helicopter
Rotor in Forward Flight," Vertica, Vol. 10, No 3/4, (1986), pp. 269-280.
18. Wagner, S., and Zerle, L.: private communication, (1991).
19. Caradonna, F. X.: private communication, (1991).

250

:-:/i_W
V
.

line of integration
-1

-2
-J

~
"

..

-4

-I,.. ....
- -~IM ...

-4

-3 -2

-1

Figure 1. Advanciog rotor wake: surface of iotegra- Figure 2. Vertical section of the free-wake geometry
tion $w and line of iotegration lw.
for a hoveriog rotor.
0 .35
.3

-_.

oo

clGsaic .ak.

;! 3.5

- ,.......s..

.25

2.5

.2

. 15

1.5

.1

.05

.5

50

100

I50 200 250 JOO J50

a lpha

Figure 3. C, VI r/ R for hoveriog rotor and compar- Figure 4. CT VI a for forward flight: undeformedwake and free-wake analysis for iocompresaible flow.
ison with experiments io 1161 (M..p = .727).
0
0
0

UNDEf'OR"'ING-WAKE

;! ,3 ,5
u

C TAl-RUNYAN

0
0
0

2.5

2.S

1.5

I.S

.5

.5

50

100

150 200 250 300

J50

UNOErOR""NG -WAKE

t; J .5

c TAl-RUNYAN

50

100

150 200 250 300 350

degree
degree
Figure 5. CT VI a for forward flight and comparison Figure 6. CT VI a for forward flight and comparison
with results io [171; iocompressible flow.
with results io [171; compressible llow.

251

au

~ 2.4
~

u
I

+
::I

EXPERIMENTAL
NUMERICAL

<>.

Q.

u
I

EXPERIMENTAL

NUMERICAL

1.6

1. 2

.8
.4

.2

.4

.6

.8

.2

1
>/e

Figure 7. Articulated rotor: -lJ.cp VI xlc at rlR =


0.75 for asimuthal position .p = 0.

CI.

"

EXPERIMENTAL
NUMERICAL

CI.

EXPERIMENTAL

u
I

NUMERICAL

./e

+
"

"I 1.6

.8

Figure 8. Articulated rotor: -lJ.cp VI xlc at rl R =


0.75 for azimuthal position .p = 90.

"::I
+

.6

.4

1.2

.8

.4

.2

.4

.6

.8

Figure 9. Articulated rotor: -lJ.cp VI


0.75 for azimuthal position .p = 180.

1.2

./e
xlc at rl R

.2

.6

.4

.8

x/e

Figure 10. Articulated rotor: -lJ.c p VI x I c at r I R =


0.75 for azimuthal position t/I = 2700

S INK

SOURCE

-I

-2

.....,-/

;.-

-3

-EXACT SOLUTION (0 . ..
44 _aD

." . RI:

...". RI_

-5

.4

Figure 11. Fuselage-like body geometry.

.8

1.2

1.6

2.4

Figure 12. cp VI x on the fuselage surface, for the


rotor at different distances.

Regularised Boundary futegral Equations for an fuverse Problem


of Crack Determination in Time Domain
N. Nishimura, A. Furukawa and S. Kobayashi
Division of Global Environment Engineering, Kyoto University,
Kyoto 606-01, Japan

Summary
An inverse problem of determining the shape of a crack from experiments using a physical quantity governed by the wave equation is considered. This inverse problem is converted into another of minimising a fit-to-data cost function. Powell's variable metric
method and a regularised boundary integral equation are used to solve this minimisation problem. Some 2D and 3D examples are presented to test the performance of the
proposed method.
Introduction
Assume that a domain D is known to contain a crack S having an unknown shape at
an unknown location. 'vVe are now interested in determining the geometry of this crack
experimentally using a physical quantity u governed by wave equation, as in ultrasonic
measurements. In the experiment one illuminates the unknown crack with known incident waves, and measures the time histories of the scattered waves at points far away
from S. Our inverse problem then attempts to determine S from the experimental data
thus obtained and the homogeneous Neumann condition, or the traction free condition
in physical terms, on the faces of the crack. In this paper we shall discuss methods of
solving this inverse problem with boundary integral equation methods, which are effective not only in sensitivity analysis [1] and crack problems [2J, but in crack determination
problems for elliptic differential equations [3] as well.
Crack determination problems have so far been solved numerically via a variety of
ways. For example Santosa & Vogelius [4] used FEM to solve a 2D crack determination
problem for Laplace's equation. Kubo et al. [5,6J considered a similar problem in 3D;
they used BIEM to solve many direct problems with given candidate cracks, and then
picked up the one which fits the experimental data the most. Nishimura & Kobayashi [3J
also solved a 3D crack determination problem with the help of regularised hypersingular
integral equations and a nonlinear programming technique. The method in [3] was later

253
extended to similar problems governed by Helmholtz' equation [7] or wave equation [8].
Tanaka et. al [9] also considered a related inverse problem in elastodynamics.
The objective of this paper is to extend the 2D method of near field data inversion
discussed in Furukawa et al. [8] to the far field inversion in 2D and 3D; we believe that
the latter is of more practical importance than the former considering the size of cracks
found in real structures. Specifically, we consider an infinite domain D containing one
single crack S, whose location and shape are unknown. The crack S is determined as
a minimiser of a cost function defined as the sum of the squared differences between
computed and measured scattered far fields produced by several known incident waves.
The minimisation is performed with the help of a nonlinear programming technique of
the quasi-Newton type (Powell's variable metric method), which needs the derivatives
of the cost function with respect to shape parameters of the crack. The computations of
these derivatives and the cost itself require solutions of hypersingular integral equations,
which we propose to solve with a variational approach. This approach uses Galerkin's
method for spatial variables and collocation for time. Finally we show several 2D and
3D numerical examples to test the efficiency and robustness of the proposed method.
2. Regularised Integral Equations for Direct and Inverse Problems
We consider 3D problems because the corresponding 2D results are obtained in a straight
forward manner. Let S be a smooth piece of curved surface in R 3 , bounded by a smooth
edge 8S. Also let the domain D be defined by D := R3 \S. In physical terms the surface

S represents a crack. The direct crack problem for wave equation is then formulated into
the following initial- boundary value problem: Find a solution u(z, t) which satisfies

Llu-ii.=O

(:~)+ =

inDx(t>O),
( : ) - =0

Uslt=o

= uslt=o = 0

onSx(t>O),

lim

in D,

x( ES)-xo( E8S)

cp(z, t)

=0

for t > 0,

and the radiation condition for Us, where us(:= U-U[) and Ul are scattered and incident
waves,

is the derivative with respect to time t, the superposed

+ (-)

indicates the

limit on S from the positive (negative) side of S, and cp := u+ - u- is the discontinuity


of u across S, or the crack opening displacement. The positive (negative) side of S is
the one into which the normal vector n to S (-n) points. Also, the wave speed has
been set equal to 1 with an appropriate scaling.

254
The solution to this problem is known to possess the following potential representation
in D x (t

> 0):
u(;c,t)

= U[(;c,t) +

G(;c-y,t-r)=

ll

aG(;c - y, t - r)
a
'P(y,r)drdSy ,

SOny

b( t - r - l;c - y I)
I - yI '
411";c

(1)

where G(;c - y, t - r) is the fundamental solution for the wave equation and b(.) is
Dirac's delta. The unknown crack opening displacement 'P is determined as the solution
to the following integral equation:
0=

aUI(;C,t)

On

ilt
S

~~G(;c - y, t

0 un x uny

where the integral with a superimposed

- r)'P(y, r)drdSy , (;c,t) E S x (t > 0) (2)

= denotes an integral in the sense of the finite

part.
An effective way of solving the hypersingular integral equation in (2) is to utilise variational formulations in which integration by parts reduces the singularity of kernel
functions to integrable one. An attempt of this type has been made by Becache in (10]
who used the Galerkin method for both spatial variables and time. She found, however,
that one has to take the time increment small in order to get a stable numerical result
[10]. In this paper we shall use the following variational equation in conjunction with
the Galerkin method applied to spatial variables and collocation to time:

(3)
where

81,2

is an arbitrary system of curvilinear coordinates on S,

eo{J

and

eijk

are 2 and

3 dimensional permutation symbols, and 'IjJ is a test function which vanishes on

as.

We

henceforth assume that Greek indices run from 1 to 2. Our experience shows that this
method stays stable even with a relatively large time increment.

255

The scattered field us given by


US('J!,t)

11

18G ('J!-y,t-r)

SOny

<.p(y,r)drd5y

approaches asymptotically to

41l'~'J!1 UF( i::, Tj 5, UI)


Izl

as
UF

~ 00

in the direction (viewed from the origin) given by i:: :=

'J!/I'J!I.

The factor

in this formula allows the following expression:

(4)
where T is a time parameter given by T := t

-I'J!I.

The quantity

UF

will be called 'far

field' in the rest of this paper.


We next consider the case where the location and shape of 5 are unknown. To determine
5 we illuminate the unknown crack with N known incident waves uy( z, t) (n
and measure the resulting far fields in }..;[ directions given by i:: m (m

= Tk

(k

= K;,m, ... ,K~,m)'

= 1, ... ,N)

= 1, ... , M)

at

where K;,m and K~,m are the first and the last time

steps at which the far field is measured. The inverse problem under consideration
attempts to reconstruct 5 from the data thus obtained. We here try to solve this
inverse problem by converting it into another of minimising a certajn cost function.
The cost function J(5) to be minimised is defined as follows: Given 5, solve

Is i
S

8 8
,,-,,-G(z - y,t - r)<.pn(y, T)dTd5 y
un" uny

&un

= - " I ('J!, t),

un
(z,t)E5x(t>O), n=l, ... ,N

for the unknown <.pn with (3). We then compute the far field

UF

by substituting the

solution <.pn into (4). The cost J(5) is then defined by

where

u'} stands for the far field obtained experimentally. With this cost the crack 5

is determined as a solution to following minimisation problem:


Minimise J(5).
S

(5)

256

In practice one may introduce simplifying assumptions on the shape of S and identify

S with a finite number of shape parameters. For example a penny shaped crack can
be described by 6 parameters. The cost is now a function of these parameters, and the
problem in (5) can be solved with nonlinear programming techniques. In this paper
we shall use Powell's variable metric method, which belongs to the family of the quasiNewton methods. Since this method of nonlinear programming requires derivatives of
J(S) with respect to shape parameters for S, one has to find an effective way to compute
these derivatives. To this end we view S to be an image of a fixed 'reference crack' So via
a mapping z

= z(X,a), z

E 5, X E So, where

ai

(i

= 1, ... ,L) are shape parameters

for S. The cost J(5) can now be differentiated directly with respect to

ai

in a manner

similar to taking material derivatives in the large deformation continuum mechanics.


Indeed, one has

where

stands for the differentiation with respect to one of the shape parameters.
v

Notice that the quantity <pn is known when one calculates J, because <pn is already
v

obtained in the evaluation of J, and the evaluation of J always precedes that of J in


Powell's variable metric method. Hence all the quantities in (6) are known except for

;Pn.

On the other hand one obtains a variational equation for

(3) with respect to

ai'

Indeed, one has

;pn

as one differentiates

257

11
1

1
11
S

1
l1
S

'"

o<pn(y,r)oxp(Y)
0
"drds I .ds2
s.\
us!,

G(z-y,t-r)e,xJ'

O<pn(y,r) o'f:p(Y)
0
"drds l .ds2
S,x
vs!,

o'f:j(Z) OXk(Z)
eijkt/J(z)-,,----,,--ds lz ds zz
VSl
VS2

l1

G(

Z -

)
.. n(
)oXp(y) oxq(y)d d d
y,t - r eipq<p y,r - " - - , , , - - r SI. S2.
VSI
VS2

.1,( )OXj(Z) o'f:k(Z)d d


eijk'f' Z - , , - - - - , , - - SI. S2 z
VSI
vS2

(
"n
oXp(y) oxq(y)
G z-y,t-r)eipq<p (y,r)-,,----,,--drds l .ds z
VSl
VSz
.1,( )OXj(Z) oXk(Z)d d
eijk'f' Z - " ' - - - - , , - - Si z S2.
VSl
vSz

l1
1
l1
S

'"

G,k(Z - y, t - r)(Xk(Z) - Xk(ye.\!'

ot/J(z) oXp(Z)d d
e",fj--",---,,-- Siz S2 z
vs",
VSfj

o<pn(y,r)oxp(Y)
0
-",--drds I .ds2
S.\
vs!,

ot/J(z) oXp(Z)d d
e",fj--",---,,-- Si z S2 z
vs",
vSfj

G(z - y,t - r)e.\!,

G ,I ( z-y,t-r )('"XI ()
v (
x -XI
y e;pq<p.. n( y,r )oXp(y)oxq(y)d
-,,-,--,,-- r dSI. d S2.
V~1
vS2
r
OXj(Z) OXk(X)
+J eijkt/J(X)-",--,,-ds lz ds z
5
liSI
US2
S

l1
1
l1
5

)
.. n(
)o'f:p(y)oxq(y)d
d
x-y,t-r e;pq<p y,r -"'----"'-- rds l S2.
VSl
vSz
.1,( )OXj(X) oXk(Z)d d
e"k'f' x - - - - - - SI S2
.)
OSI
OS2
z
z
5

1G(

1G(

)
.. n(
)OXp(y)o'f:q(y)d d d
X - y,t - r e;pq<p Y,r -,,----"'-- T SI. S2.
VSl
VS2

= { t/J(z)uj,i/(x,t)'f:,(x)n;(x)dS+ { t/Juj,;(x,t)eijkepqknp o'f:1(x)dS.

15

15

(7)

VXg
One can now solve (7) numerically for ~n using the same technique as in (3). See the

next section for the detail. In the special case of a penny shaped crack (3) and (7)
simplify considerably. The result of this simplification, however, is omitted because of
the space limitation.
3. Numerical Techniques
This section outlines the numerical techniques used to obtain the results to be shown
in the next section.

258
1. Di8cretisation

In 3D (3) is solved easily with the Galerkin method applied to space

variables and collocation to time. Namely, we substitute fiA for

t/J and

L:B fiBcpn,B(t)

for cpn in (3), where fiA is a shape function on 5, and cpn,A(t) is the time variation of cpn
at the Ath node on 5. The time function cpn,A(t) is further discretised with piecewise
linear interpolation functions, and the resulting equation is solved with collocation. As
regards fiA we follow Nedelec [11] to use ordinary piecewise linear triangular elements
neglecting the near tip singularity of cpot. In the discretised version of (3) thus obtained
we compute time integrals analytically, and the inner (outer) surface integrals analytically (numerically). Notice that the discretised equations obtained from (7) and (3) are
the same if one uses the same shape functions for

'(pn and cpR. Hence one constructs and

triangular-decomposes this matrix equation only when one computes J and reuses the
'7

results in the calculation of J.

2. Time Increment

The choice of the time increment Llt has to be made very carefully

in the present analysis because the ratio Llt/(element size) has an influence on the
numerical accuracy of the solution, and because the size of the crack changes during the
minimisation. One might possibly adjust Llt according to the current crack size, but this
would make the algorithm too complicated. We therefore chose to keep Llt constant in
the whole nonlinear programming process. In practice, one may set Llt to be sufficiently
small in comparison with the time scale of the variation of the experimental far field
data.
Time Steps

9.

Suppose that one calculates cp for K time steps with a fixed Llt.

Equation (4) then shows that one can determine the associated far field for K time
steps. In evaluating J(5) one has to compute far fields from a candidate crack at

= Tk = kLlt for k = K!,m, ... ,K~,m'

Now suppose K~,m

= K!,m + K

- 1, which

is a natural choice. If TK~.m is smaller than the computed arrival time TK. from
the candidate crack, there is no problem in the evaluation of the cost function. If

TK~.m

> TK., however, one may have to increase the number of time steps to compute

far fields for T > To, where To := (Ka +K -1)Llt. However, this 'variable K' approach
is inconvenient because it may possibly make the computational time uncontrollable.
We therefore decided to keep the number of time steps K constant in the calculation of
UF,

and set

UF(-, T;,)

equal to

UF(-, To;"')

for T;::: To.

259
4. Numerical Examples
This section shows a few 2D and 3D numerical examples. We here try inversions of
numerically simulated experimental data, rather than real ones. For the simulation we
used BIEM with the true crack geometry, and gave noise to the computed far field, thus
producing input data to the inverse problem solver. The number of time steps is 15 in
all the examples.

1. 2D Proble1Tl.8

The crack is assumed to be a straight line, which we can identify by

4 shape parameters, i.e.


given by u'l

= J(t -

Xl,2

coordinates of the tips. We consider two incident waves

tn - Xn), (n

= 1,2), where tn

the incident wave does not reach the crack at t

is a certain number chosen so that

= 0, Je-) is defined by
Os t S 21l'1

(8)

otherwise '

and I

= 2.

For the crack we used 9 piecewise linear boundary elements, and the time

increment is 1.5 times the average element size in the true crack. The directions of
far field measurements are i: m

= (cos am, sin am) with 01 ,2,3,4 =

45 0 ,135 0 , -45 0 , -135,

and 30% of random error is given to the data. Each tip is constrained to stay within
a circle having a radius of 3 centred at the origin, in order to avoid divergence of the
solution. Fig. 1 shows the true crack, initial guess and the intermediate crack locations
'V

at each evaluation of J in the minimisation process. As this figure shows the analysis
converged to a crack sufficiently close to the true one. The CPU time was about 2 sec.
on Fujitsu M1800 (a scalar processor).

2. 3D Problem3

The crack is assumed to be penny shaped, and is illuminated by three

incident waves given by u'l

= J(t -

tn - Xn) (n = 1,2,3) with I = 1, see (8). We use

6 shape parameters for the crack as in [3], and 21 piecewise linear shape functions for
unknowns on S. The time increment is set equal to the average element length in the
radial direction of the true crack. The far field measurement is made in 14 directions,
which consist of the coordinate directions (6 directions) and (1/,;3, 1/,;3, 1/,;3)
(8 directions). The data is noisy with 20% ofrandom error. Fig. 2 shows the true crack,
'V

initial guess and the intermediate crack locations at each evaluation of J. As this figure
shows the analysis converged to a crack sufficiently close to the true one. The CPU
time was about 45 sec. on Fujitsu VP2600 (a vector processor).

260

Fig. 1 Mode of convergence in 2D crack determination analysis in time domain.

initial
1

Fig. 2 Mode of convergence in 3D crack determination analysis in time domain.

261

5. Concluding Remark
The method discussed in this paper can easily be extended to elastodynamics. An
investigation along this line is now under way.
References
1. Barone, M.R.; Yang, R.-J.: Boundary integral equations for recovery of design
sensitivities in shape optimization. AIAA J. 26 (1988) 589-594.
2. Nishimura, N.j Kobayashi, S.: A regularized boundary integral equation method
for elastodynamic crack problems. Compo Mech. 4 (1989) 319-328.
3. Nishimura, N.j Kobayashi, S.: A boundary integral equation method for an inverse
problem related to crack detection. to appear in Int. J. Num. Meth. Eng. (1991).
4. Santosa, F.j Vogelius, M.: A computational algorithm to determine cracks from
electrostatic boundary measurements. Int. J. Eng. Sci. 29 (1991) 917-937.
5. Kubo, S.: Inverse problems related to the mechanics and fracture of solids and
structures. JSME Int. J. 31 (1988) 157-166.
6. Sakagami, T.j Kubo, S.j Ohji, Kj Yamamoto, Kj Nakatsuka, K: Identification of a
three-dimensional internal crack by the electric potential CT method. Trans. JSME
56 (1990) 27-32 (in Japanese).
7. Nishimura, N.: Regularised integral equations for crack shape determination problems. Inverse Problems in Engineering Sciences (Eds. M. Yamaguti et al.), Springer,
(1991) 59-65.
8. Furukawa, A.; Nishimura, N.; Kobayashi, S.: Crack determination by time-domain
BIEM. Proc. BTEC (1991) 67-70 (in Japanese).
9. Tanaka, M.; Nakamura, M.; Nakano, T.: Detection of cracks in structural components by the elastodynamic boundary element method. Proc. BEM12, Compo
Mech. Publ., 2 (1990), 413-424.
10. Becache, E.: Doctoral Thesis. Universite de Paris VI, (1990).
11. Nedelec, J .C.: Integral equations with non integrable kernels. Int.egral Eq. Operator
Th. 5 (1982) 562-572.

Boundary Integral Equation Defmed on a Crack Approximated


with Voids
KOHli OHTSUKA
HIROSHIMA-DENKI INSTITUTE OF TECHNOLOGY
6-20-1, NAI{ANO, AKI-KU,
HIROSHIMA 739-03, JAPAN

1.

Introduction
The equilibrium of a linear elastic body with a crack is described by a boundary value problem on a domain n c ~3. Here the undeformed shape n is expressed in the form of n ~3\C, where C is a surface with a smooth boundary.
Here we assume that the surface C is approximated with a family of domains
V. (0 ~ E ~ Eo) with smooth boundary, as ill Fig.1.

.~V(
..

.8V,

..

/
~....

~~~
..-!'~....n_r.:;;....::::s-_-~,
~_

...........-0:........

..

..,.

....

--..

.~.-:::.--=-~

/'

, - - - - C_ _

----111

Fig.i. Undeformed shapes of a crack C and voids V.


In domains n. = 1R 3 \V.. the elastic bodies subjected to forces
surface ave of voids V. is given by
in

it.(:!:)

-+

0'( it, )n,

O'ij(ti)

n.

q. on the

(zero body forces),

as:!: -+ 00

= (O'ii (ti. )ni) = q.

= CijHEkl(ti),

on

EH(ti)

av..

(1.1)

= ~(DkUI + DrUk)
Typeset by Aw-'!EX

263

where it. are displacement vectors, ii. = (n.,l, n.,2, n.,3) the outward unit normal
of BV. Precisely, it, are given by the solutions of the variational problems.
Problem (P) .: For a given
that

1
n,

0"( it. )e( ii)dx

q.

E (H- 1 / 2 (BV.)3, find it, E (Wl(U.)3 such

1 q..
av,

(1.2)

ii dS

Here Hl/ 2 (8V.) are Sobolev spaces of Qrder half in L 2 (8V.), whose norms
Iiqlb/2,av, are equivalent to
II q11 2l/2,av, = II q112L2(8V,)

+ InilV,

ilV,

Iq(r) - q(Y)1 2 d d
1X - Y14 S" S",

(1.3)

H-l/2(8V.) lihe dual spaces of the spaces H 1/ 2 (8V.) and the norms of Wl(U.)
are defined by
(1.4)
The problems (P) are uniquely solvable with Korn's inequality and La..'CMilgram Lemma (see e.g., [p.150, 2]).
The displa.cement vectors iE. satisfy the boundary integral equations
{

Jav,

0":&

(T(ny)(r - yn. "iE.(y)dSy

where T(n)(x) = (T;j(n))(x)

(-~ik;j(x)nk).

= q.(r)
Here we set

(1.5)
~i.t;j(X)

=(Jil:(Uj)(x) using the fundamental solution Uij, i,j = 1,2,3, defined by


We note that the integral kernel o"z: (T(n,,)(x - yii. J: is hypersingular, but

i -.

Jav,

(J" (T(ny)(x - y

n. z:i(y)dSy

is well-defined operator from Hl/2(8V,) onto H-l/2(8V.) (see e.g.[lJ).


In this paper, letting - . 0 under some assumptions, we prove the existence
it E (Wl(U)3 and the convergence it, -. it.

264

Crack C

Fig.2.

The mapping

~6(X)

2. An approximation of the crack with voids


We assume that the boundary av. of the voids intersect at the edge 8C of
the crack, and are divided into the following,

av{ = &+V{ u a-v{

(2.1)

U &C.

Moreover we suppose that there are C1-diffeomorphisms 4>~ from &:l:V{ onto
C\&C respectively. The approxima~ion means that
V'I C V.z

if e1

< e2,

av.: n &V' z =

and

n. v. = C,

if E1 :f:. El'
For a fixed 6 > 0, a ~ eo/4, we can construct a COO-function
defined on (-00,00) as follows (see Fig.2),
l(6; E)
For x

e V.o \C,

= 0 fore

8C

l(o; e)

~ 6,

we have uniquely a number

(2.2)
E

(2.3)
l(o; e)

= dore > Eo/2

E> 0 so

that

e &V.,

from the assumption (2.2), (2.3). Let I be the segment perpendicular to C. For
x e V'o \Vs, x E &V., we now denote by x(6; E) the point that intersects at
&Vl(s ). From the construction of l(o; -=),.

x( 6; e) = x if e > Eo/2,

and

x(o; 0)

e c.

We now define the mapping ~6(X) on 1i 3 \Vs by the following,


(

~6 x)

={

x(6; e)

x E V.o \Vi and x E &V"


3 -

x
xEIi \V w
Let d.(x, y) be the distance defined by the infimum of the length of the broken
lines connecting two points x,y E 1i 3 \V. and lying inside 1i 3 \V . Similarly we
define the distance d.(x, y) for x, y E 1i 3 \C.

265

Lemma 2.1.

The mappings

are Lipschitz continuous with respect to the distances d5 , d., i.e.,

with a constant G independent of .

For a domain Q in 1Ji3 and two points x, y of Q, we define the distance dq(x, y)
by the infimum of lengths of all broken lines connecting x and yin Q. We notice
that dq(x,y) is not equivalent to the Euclidean distance. If Q has the local
Lipschitz property, then dq is equivalent to the Euclidean distance.
We can easily show
Lemma 2.2.
Let Q and 0 be bounded domains in Q into O. Assume that
Q is covered by a family {Qj h=l .... ,m, of domains with local Lipschitz property
in 1Ji3 that satisfy the following:
There is a constant G such that 1iI>(x) - iI>(y)1 :5 Glx - yl for any
Qj, i = 1,'" ,m.
(2.5)
The image Qi of Qi under iI> is a domain with local Lipschitz property
for each i.

(2.4)

X,y E

Then there is a constant Go such that

(2.6)

do(q}(x),q}(y:5 Codq(x,y)

forallx,yEQ.

Proof.
Let B be an arbitrary broken line lying inside Q, which connects
y in Q. By considering a subdivision of B, if necessary, we may assume that
B consists of segments Zj_lZj, j = 1,," ,q, with Zo = x, Zq = y, where for each
j, Zj-l and Zj belong to QkU) for some leU), 1 :5 leU) :5 m. By (2.4), (2.5),
there is a constant Gl such that
X,

dO(q}(Zj_l),iI>(Zj)) ::; dq~(j)(q}(Zj-l,q}(Zi))

:5 G1 1q}(zj-l)-q}(zj)l:5 G1G!Zj_I-Zjl
for each j = 1, ... ,q. From now we have
q

do(x,y) :5 Ldo(iI>(Zj-l),q}(Zj)) :5 G1Clzj_1-zj\.


j=l

Therefore, we obtain (2.6) with Co = G1 C.

266

Proof of Lemma 2.1. The normal derivative of <p.(x), x E V.o \V. at the
point x E 8V. is almost d1(6;e)/d and
as 6 -+ O.

(2.7)

The directional derivatives in the parallel direction to C are zero. For x E !li3\C,
the gradient ~<1>.(x) is the identity matrix. From now, we complete the proof
0
of Lemma 2.1 from Lemma 2.2.

Remark 2.1.

We can prove that <p. : !li 3 \ Vs -+ !li3\C are 1-1 mappings.

Remark 2.2.
In the second derivative of <1>6, the curvature of surfaces 8V. will
appear. This indicates that an analogous result as in Lemma 2.1 does not hold
for the first derivative of <1>s.

From Lemma 2.1 and Remark 2.1, the maps vex) H <psv(x)
v(<1>s(x)) give
the homeomorphisms from WI(U) onto WI(U.).
Inverse (<p 6)-1 : WI(US) -+ WI(U) is given by 1P6w(x) = w(lPs(x, IPs =
1
<pS

3.

Approximation of solution

Letting IPs
<PiI, and using the change of variable in the left-hand side of
(1.2), we have for iiE (Wl(U)3,

We denote by -y% the trace operator from (WI(U)3 to (Hl/2(C)3, that is:
If ii is smooth up to C, -y+ii(x) (-y-ii(x, x E C, is the limit value of ii(y) as
y, over (under) C, moves closer and closer to x.
By the divergence theorem in weak form (see e.g.[p.9, 4]), we get

10 0'( 1P6iIs)e(ii)dx

{-y+ (O'(lPsiIs)ri) , -y+ii)c - (1'- (O'(lPsiIs)ri) , -y-ii)c (3.2)

- 10 divO'( 1P6iIs)iidx

= {-y+ (O'( IPs iIs )r!) , '1+ ii - -y- V)c


+ {(-y+ (0'( 1P6iIS)r!) - '1-(O'(lPsiIs)r!) , -y-ii)c

- In

divO'( IPs iIs )iidx

267

qS

Suppose, for the surface forces

E (Hl/2(8Vs)3,

~5(q61&+v.(x - ~s(ijsl&-v.(x

= (id&+v,)(l]is(x
as.5

and the existence

qE

({

-l-

(3.3)

- (q'5la-vJ(~s(x)

-l-

H~t2(C)}') 3 such that


(3.4)

where q61&:!:v are .the restriction of q6 on O=V respectively. Here the subspace

H~t2(C) of Hl/2(C) is defined by the norm

and

{H~t2(C)

denote the dual space of H~t2(C), and {-, )cdenote the duality

between ( { H52(C)},) 3 and ( {


For

H~t2(C)}') 3

v E (Wl(U)3, we have

(For the proof, refer to [Theorem 2.6. 5].)


Since
from (2.7)

divq(l]isu6)

-l-

0 in U in distribusion sence.

We then have from (3.3) and (3.4),

(1+ (q(l]isu6 )n) "1+ v- "1- V}C


I

+ (7+ (q(~Su6)n) -,- (q(l]isus)n) , -Y-V}c

- 10 divq(l]isus)vdx
-l-

(q, ,+v - ,-v)c.

268

Using the Lax-Milgram lemma, we can prove that W6it6 are Cauchy sequence in
(Wl(O)3. Therefore, there exists a function it E (Wl/2(O))3 as follows,

and it satisfy the integral equation

REFERENCES

[I} E.Becache and T.Ha Duong, Formulation variationnelle espace-temps associee


au potentiel de double couche des ondes elastiques, Rapport Interne nO 199,
(1989), CMAP, Ecole Poly technique.
[2] R. Dautray and J-L. Lions, Mathematical Analysis and Numerical Methods
for Science and Technology, Volume 4, "Integral Equations and Numerical
Methods", 1984-85.
[3] K. Ohtsuka, Generalized J-integral and three-dimensional fracture mechanics
n - Surface crack problem -, Hiroshima Math. J., 16 (1986), 327-352.
[4} R. Temam, Navier-Stokes Equations, North-Holland, 1979.

Boundary Element Analysis of Frictional Slip Behaviour


in Contact Problems
O.A. Olukoko*, A.A. Becker** and R.T. Fenner*
* Department of Mechanical Engineering, Imperial College, London SW7 2BX, United
Kingdom
** Department of Mechanical Engineering, University of Nottingham, Nottingham NG72RD,
United Kingdom

Abstract
A boundary element algorithm developed for frictional contact between two or more bodies has
been implemented in a multi-domain computer program for two-dimensional and axisymmetric
contact analysis. The solution method involves a rearrangement of the overall matrix equations
for the contacting bodies in order to incorporate the contact conditions at the common interface.
Several examples of two-dimensional and axisymmetric frictional contacts are analysed and the
results are in very good agreement with analytical and finite element solutions.
1 Introduction
In many structural mechanics problems, contact between two or more bodies often arise such as
in ball and roller bearings, meshing of gear teeth and the contact of a pneumatic tyre with the
ground. These problems are mostly nonlinear because there is usually no prior knowledge of the
extent of the contact area, which changes in size and shape as the load is applied, and the
presence of finite friction which introduces a stick-slip behaviour at the interface. Most analytical
approaches to the solution of contact problems are limited to idealised models thus making it
imperative for efficient numerical algorithms to be developed. The emergence of the finite
element (FE) method has resulted in the development of various numerical algorithms by many
authors to successfully tackle a wide variety of contact problems. for example. Francavilla and
Zienkiewicz (1] and Bathe and Chaudhary (2]. However. the inability of the displacement-based
FE equations to simultaneously satisfy the equilibrium and compatibility conditions at the contact
interface has limited the efficiency of the FE approaches.
The boundary element (BE) method is particularly well suited for contact problems since the
contact variables can be directly incorporated into the overall matrix formulation without
increasing the order of the equations. Modelling of only the boundary curves of the domains and
the ability to calculate the tractions and displacements to the same order of accuracy means that

270
relatively more accurate results can be obtained when compared to the FE method. The BE
method has been applied to two-dimensional elastic contact problems with friction by Andersson
and Allan-Persson [3], and to two-dimensional thermoelastic contacts by Karami and Fenner [4],
while Abdul-Mihsein et al. [5}, Becker and Plant [6] and Becker [7] applied it to axisymmetric
contact problems. This paper is a documentation of the further development of a BE algorithm

[7] for the analysis of two-dimensional and axisymmetric frictional elastic contact problems. The
appropriate contact conditions at the common interface are used in coupling the matrix equations
resulting from the BE formulation for the contacting bodies. The frictional conditions are
imposed using Coulomb's law of limiting friction and an efficient automatic iterative technique is
developed to handle the contact iterations. The results for the several examples analysed are
compared to analytical and FE solutions. The FE solutions are obtained using the commercial FE
package, ABAQUS [8J, which involves the introduction of special contact elements at the
common interface of the contacting bodies.
2 Contact Boundary Conditions
Consider two elastic bodies arbitrarily denoted as A and B which are in contact as shown in
Fig. I. On the common contact surface Se, the tractions and displacements are unknown but they
must satisfy the conditions of compatibility and equilibrium in accordance with the current
contact conditions. Contact problems may be classified as frictionless or frictional. For frictional
contacts, if the tangential tractions are lower than the limiting shear stress values for slip, then the
contacting surfaces are perfectly stuck to each other and relative tangential motion is impeded.
Hence, compatibility of displacements and equilibrium of tractions are maintained in the normal
and tangential directions. These can be expressed in global cartesian coordinates between the
nodes a and b on bodies A and B respectively (Fig. 1) which form a nodepair when in contact as
uxb+X b

uxa+xa

(1)

uyb + yb = u/ + ya
t,.,b = -lxa
lyb = _lya

(2)

(3)

(4)

where t and u are the tractions and displacements in the contact area respectively. Subscripts x
and y represent the global cartesian directions (r and z for axisymmetric problems) and x and y
are the cartesian coordinates. Hence, the only unknowns to be solved for sticking contact are Ilxa ,

ul, lxa and lya. If the tangential tractions, ~ exceed the limiting values, then slip must occur and

the local tangential tractions are set equal to the prescribed limiting values for slip according to
Coulomb's law of friction (I t,1
and the normal traction,

s I' 1tn I), that is, the product of the coefficient of friction, 1',

tn. The local tangential and normal directions are shown in Fig. 2. The

displacement constraints are removed in the tangential direction and the bodies are allowed to

271

deform tangentially. Coulomb's law of friction is applied globally by transforming the global
traction components tx and !.y to local coordinates. Hence, the compatibility and equilibrium
relationships for the slipping nodes are

fx3 = Italy8
lyb=_lya

(5)

(6)

fxb = -fxa = -llGlya


u yb

u ya + (ya _ yb) + (uxb -

(7)
UX8

)tanOC + (xb - xa)tanec

(8)

where

(9)

The correct sign of JA. in equation (9) is determined such that

is of the same sign as in the

previous iteration where a sticking status is assumed, and e c is the tangent angle as shown in
Fig. 2. Thus, for each slipping nodepair, the only unknowns to be solved for are tya, u ya , llxa
and uxb. Frictionless contact is a case of complete slipping at the interface with the tangential
tractions equal to zero. Hence, the slip relationships are applied in this case with the value of JA.
equal to zero. In most practical contact problems, the contact surface usually consists of regions
of slip and stick occurring simultaneously and the extent of each has to be determined iteratively.
3 Numerical Implementation
The detailed formulation of the BE method for two-dimensional and axisymmetric elastostatic
problems using isoparametric quadratic elements is well covered in the literature (for example,
Becker [9]) and will only be summarised here. For each of the bodies in contact, after
performing the necessary numerical integrations over the boundary elements, the reSUlting set of
linear algebraic equations can be written as

[A] {u} = [8] {t}

(10)

where [Al and [8] contain the traction and displacement kernels respectively. The prescribed
boundary conditions are applied, and by using the contact relationships in the previous section,
the matrices in equation (10) for each contacting body can be coupled together and rearranged to
form the final solution matrix
[C] {x} = {d}

(11)

272

where [C] is the coefficient matrix of the unknown vectors {x}, and {d} is a column matrix of
the known quantities. The equations are then solved by the Gaussian elimination technique.
4 Automatic Iterative Scheme
An efficient automatic iterative scheme is employed in the algorithm to handle the contact
iterations. At the end of each iteration or solution stage, the displacements of the nodes just
outside the initially assumed contact area are checked for overlap or interpenetration into the
opposite domain, the occurrence of which suggests that the contact area is too small for the
applied load. Hence, the overlapping nodes are included in the contact area for the next iteration.
For frictional contacts, all the contacting nodepairs are assigned a sticking status in the first
iteration and the ratio of the tangential to the normal traction is calculated. If the absolute value of
this ratio exceeds the coefficient of friction JA., then the affected nodepairs are allowed to slip in
the next iteration while a limit is imposed on the tangential traction according to Coulomb's law
of friction. Finally, nodepairs with normal tensile stresses are deleted from the contact area in the
next solution stage, which is an indication that the contact area is too large for the applied load.
The aforementioned checks are done automatically with no user-interaction and the analysis is
terminated when no changes are detected by these checks.
5 Examples
The BE program is applied to some practical two-dimensional and axisymmetric frictional contact
problems in which three-noded isoparametric elements are employed in the discretisation with
four Gaussian integration points. The results are compared with analytical and FE solutions
using ABAQUS. The BE meshes are plotted with the end-points of the elements shown as short
straight lines and the mid-points as small circles.
5.1 An Elastic Punch on a Foundation
The plane strain indentation of an elastic foundation by a flat elastic punch is considered. The
geometry and loading are shown in Fig. 3(a) with the following relative dimensions and material
properties; Hp/Wp = 2, WrlWp = 4, HrlWp = 4, ErfEp = I, v = 0.3. The BE mesh in Fig. 3(b)
consists of 76 elements for the symmetric half considered. The shear stress distributions for JA. =
0.2 and 0.3 are shown in Fig. 4(a) in comparison with the FE (ABAQUS) solutions. For JA. =
0.2, slip occurs in the range 0.45 s xlW p S 1.0 and decreases to 0.8 S xlWp S 1.0 for JA. = 0.3.
The dimensionless relative tangential slip at the contact interface corresponding to the JA. values is
shown in Fig. 5 together with the frictionless (JA. = 0) values. In the slip zones, the contacting
surfaces deform along their positive local tangents, that is in the positive x-direction for the lower
surface of the punch and in the negative x-direction for the upper face of the foundation in contact
with the punch. The effect of friction on the normal pressure distribution as shown in Fig. 4(b) is

273

a slight reduction from the frictionless values, with no significant change occurring for an
increase in J.1 from 0.2 to 0.3. Good agreement is obtained withthe FE solutions as evident from
Figs. 4 and 5.

5.2 A Cylindrical body on a Flat Plane


A cylindrical txxIy under normal and shear loads in contact with a flat elastic plane of the same
material under plane stress conditions is considered. The physical dimensions and loading for the
problem are shown in Fig. 6(a). The following data are employed in the analysis; E = 206 GPa,

0.155,
v = 0.3, R = 1.0 m, F = 2.352 x 1()6 N, T =3.523 x lOS N, ( l tan-1(T/F), Jl.
maximum contact pressure for Hertzian case, Po = 2.911 GPa. The BE mesh of 70 elements is
shown in Fig. 6(b). A comparison of the normal pressure distribution obtained for the model
with the Hertzian case (J.1

=0) having the same normal load gives a 3 percent reduction in the

maximum value as shown in Fig. 7(a). Also shown in this figure are the FE results obtained
using ABAQUS and those obtained by Sachdeva and Ramakrishnan [IO} using a flexibility
matrix (FE) method. It can be observed that the combined effects of the tangential load and
friction is a decrease in the normal pressures around the center of the contact zone and a slight
increase in the contact area in comparison with Hertz analytical results for the frictionless case.
Fig. 7(b) shows the shear stress distribution in which a state of complete slip along the interface
is prevented by the small adherence portion occurring close to the center of the contact zone. The
BE results are in good agreement with the FE results as evident from Fig. 7.

5.3 A Spherical Inclusion in an Infinite Solid


The non-Hertzian contact stresses occurring during partial separation between an infinite elastic
medium and an elastic spherical inclusion is analysed. The infinite txxIy is subjected to two
independent axisymmetric stress components at infinity and the dimensions of the infinite solid
relative to the inclusion are made sufficiently large so that die applied stresses can be reasonably
assumed to act at infinity. The geometry and loading are shown in Fig. 8(a) with the following

=5, HaIRA = 5, EaIEA = 1, v =0.3, Pr = -1,


=O. The contact angle e is measured from the z-axis. The axisymmetric BE mesh in Fig. 8(b)

relative dimensions and material properties; RBIRA


Pz

consists of 42 elements. The normal contact pressure distribution for different values of J.1 are
shown in Fig. 9(a) together with the analytical results (frictionless, J.1 = 0) of Wilson and Goree
[ll}. Excellent agreement is obtained with the analytical results for J.1 = 0 in which the initial
contact over the whole interface reduces to a spherical band in the interval 35 ~
However, for Jl. = 0.3, the contact zone is now obtained in the interval 30'

e ~ 90.

e ~ 90. Thus an

increase in Jl. leads to an increase in the contact area and a drop in the normal pressure values
which can be attributed to the load being distributed over a larger area. The shear stress
distributions for J.1 = 00 and 0.3 for the same loads are shown in Fig. 9(b) together with the FE

274
results from which a slip annulus is obtained within 30 s

es 77 for f.I. = 0.3. These results are

confinned by those obtained using the FE package ABAQUS as shown in Fig. 9.


6 Conclusions

A BE algorithm for elastic contact problems with friction has been implemented in a computer
program for two-dimensional and axisymmetric contact analysis. The algorithm employs
element-pairs along the common interface and couples the integral equations for the contacting
bodies by using the appropriate contact conditions. Frictional stick-slip is modelled according to
Coulomb's law of friction and an automatic iterative scheme to perfonn the contact itemtions with
no user-interaction is employed. The accuracy of the algorithm has been demonstmted with some
practical examples of two-dimensional and axisymmetric frictional contact problems involving
stick-slip behaviour at the interface, and the results are in very good agreement with FE and
analytical solutions.
References
1. Francavilla, A. and Zienkiewicz, O.c., A note on numerical computation of elastic contact
problems, Int. J. Num. Meth. Eng., 9, 913-924, 1975.
2. Bathe, K. J. and Chaudhary, A., A solution method for planar and axisymmetric contact
problems, lnt. J. Num. Meth. Eng., 21, 65-68, 1985.
3. Andersson, T. and Allan-Persson, B. G., The boundary element method applied to two
dimensional contact problems, in Progress in Boundary Element Methods, Vol. 2 (Edited by
Brebbia, C. A.), Pentech Press, London, 1983.
4. Karami, G. and Fenner, R. T., A two dimensional boundary element method for thermoelastic body forces contact problems, in Boundary Elements IX, Vol. 2 (Edited by Brebbia,
C. A.), CMP Publications, Springer-Verlag, 1987.
5. AbduI-Mihsein, M. J., Becker, A. A. and Parker, A. P., A boundary integral equation
method for axisymmetric elastic contact problems, Com put. Structures, 23, 7'ir1-m, 1986.
6. Becker, A. A. and Plant, R. C. A., Contact mechanics using the boundary element method, in
Proc. of the Int. Conf. on Tribology, Friction, Lubrication and Wear, 975-980, Institution of
Mechanical Engineers, London, 1987.
7. Becker, A. A., A boundary element computer progmm for practical contact problems, in
Modern Practice in Stress and Vibrational Analysis (Edited by Mottershead, J. E.), 313-321,
Pergamon Press, Oxford, 1989.
8. ABAQUS version 4.8 users' manual, HKS Inc., Rhode Island, U.S.A.
9. Becker, A. A., The Boundary Element Method in Engineering-A complete course, McGmwHill (UK), 1991.
10. Sachdeva, T. D. and Ramakrishnan, C. V., A finite element solution for the elastic contact
problems with friction, Int. J. Num. Meth. Eng., 17, 1257-1271, 1981.
11. Wilson, H. B. and Goree, J. G., Axisymmetric contact stresses about a smooth elastic
sphere in an infinite solid stressed unifonnly at infinity, Trans. ASME, J. Appl. Mech., 34,
960-966,1967.

275

Fig. 1 Two bodies in contact

Fig. 2 Local normal and tangential directions

Wp

:e---':Po

- -- - - - - -~"---""",f
Hp

~,
w,

contact lone
(a)

(b)

Fig. 3 An elastic punch on a foundation (a) Geometry and loading (b) BE mesh

276
0.6

I..

FE. ~. 0.2

1.6

0.)
IE.II-0.2
IE. ~-O. )
FE.~.

0.4

0::

.
~

0.2

I.'

IE. ~ 0.2. 0.3. 0.4


BE.,,-O

-FE.~.O

1.2
1.0
0.1

0.0 -i"":><;O"";"'"T"""--"T"""--"T"""--""""-0.0
0.2
0.1
0.4 >lVi, 0.6

(a)

0.6
0.0

0.2

0.4

>lVi,

0 .6

0.'

(b)

Fig. 4 Contact stresses for an elastic punch on a foundation. (a) Shear stress distribution
(b) Normal stress distribution

1.0

0.8
<:I

CI.

0.6

A.

r:

;;)

--FE.u-O
--FE,,, - 0.2
-FE.,,-O.l
o BE,Il- 0
D
BE." - 0.2
BE." -0.1

0.4

<l

0.2

0.0 :==-..-T"'""--.--,.-..0IIIIfI:.JZ:.'!:~~~r-1IIIIII=:-,1.0
0.8
0.6
0.2
0.0
0."
s/Wp

Fig. 5 Relative tangential displacement distribution for a punch on a foundation

277

--

r-..::===----::---c::::~__r---.

~------------------_________.~.

con~clzone

(b)

Fig. 6 A cylindrical body on a nat plane (a) Geometry and loading (b) BE mesh

1 .0r---:d~~=---

___

I .O~----~'------""

OJ

0.1
:.

0.6

0:

0.4

.:
-

0.2
0.0

-iWu. p.O.

....1.(1981)... -0..,'

l...J...~~::;::;::;:;:;::;:::::::;::::::;::;::::;:::..,.~~
-6

-5

-J

.1

_I
0
IObl1l

(a)

Sochdn.1l11. (1981).11- 0.15'

0.1

IE.II-O.m
PE.II-o.."

IE, II" 0.1"

+ PE.II" 0.15'

~,................."T'"'O""T""-.--T.......,,Ioio-1

0.0 +-.d...-.....,.......................
~

-5

-J

.::

-I
0
IOO.IIR

(b)

Fig. 7 Con~ct stresses for a cylindrical body on a plane (a) Normal stress distribution (b) Shear
stress distribution

278

P,

(b)

(a)

Fig. 8 A spherical inclusion in an infinite medium (a) Geometry and loading(b) BE mesh

0.6

0.4

_willoft .. ac..II961). .. O ~

-..,._0

...... --

005

BE.~.O.J

0.1

-.n.~.o.)

-FL~-

~o

~ z

~ ~

e(..-)

(a)

~ ~ ~

(b)

Fig. 9 Contact stresses for spherical inclusion example (a) Normal stress distribution (b) Shear
stress distribution

Three-Dimensional Long TIme Prediction of Nuclides


Migration by Boundary Element Method
K. Onishi and R. Kawamura

Department of Mathematics II, Science University of Tokyo,


and

26 Wakamiya-cho, Shinjuku-ku 162, Tokyo, Japan

Information and Mathematical Science Laboratory, Inc.,


2-43-1 Ikebukuro, Toshima-ku 171, Tokyo, Japan

Summary
Based on the stable characteristic property of the boundary element scheme,
this paper discusses a method aiming at the analysis of groundwater flow
and advective diffusion of radioactive nuclides in a large scale repository for
a long time interval. Transient flow and advective diffusion of radioactive
nuclides with a decay chain are formulated by the direct boundary element
method using respective fundamental solution to each phenomenon.

Introduction
Growing interest has been focused on quantitative long time prediction of
the migration of nuclides in the disposal of nuclear wastes in the ground after
reprocessing of nuclear fuel. The migration plays an important problem to
access the performance of the repository as well as thermal, radioactive and
structural problems.
The migration analysis is coupled to the analysis of groundwater flow.
Similar analyses of groundwater flow coupled to the saltwater intrusion by
the boundary element method have been presented by several authors [1],
[2]. Coupled treatment has difficulty in handling the advective diffusion
equation, because an explicit form of the fundamental solution can be derived only for the case of constant velocity. However, variable" velocity field
can be splitted into a sum of constant reference velocity plus its deviation.
Fundamental solutions corresponding to the constant reference velocity
are considered in this paper. As a basic study followed from Kawamura and
Onishi [3], transient problems of groundwater flow coupled to the migration
of nuclides in a long time interval of the order of lOS years are considered
for the ultimate goal of the optimum design of the repository. Mathematical discussions are same as above previous paper. The boundary integral

280

equation approach can be safely applied even if coarse meshes are used in
the model at large Peclet numbers.

Governing Equations
We consider a domain n in three dimensions with the rectangular coordinate
system (x, y, z). The domain is regarded as an aquifer of the ground, which
is saturated with groundwater. Let H denote the unknown water head. We
assume that the domain is hydraulically homogeneous and isotropic. Then,
transient groundwater flow in n is governed by the seepage equation:

sa:

-1et:.H = Q,

(1)

where S is the storage coefficient of the aquifer, Ie is the permeability coefficient, 6. is the Laplacean in three dimensions, and Q is the source rate of
groundwater. The velocity u of the groundwater flow is given by
Ie
u = -"8VH,
(2)
where () is the porosity of the aquifer, and V is the gradient operator.
We consider the migration of nuclides with a radioactive decay chain in
the aquifer. Let Cm denote the unknown concentration of mth nuclide, for
example, m = 1,2,3. Then, the migration is governed by the following
advective diffusion equations in n.

(}Rm a~m

+ (}Rm>'mCm + V . (}J) = (}Rm-l>'m-lCm-l + Fm ,

(3)

where the mass flux J is given by


J

= -DVCm+uCm.

(4)

Here, Rm is the retardation factor of the mth nuclide, Am is the decay


constant, Fm is the source term corresponding to the mth nuclide, and D
is the dispersion coefficient. We assume here that the dispersion coefficient
is independent on the species of nuclides.
The initial condition of waterhead and that of concentration in n are
given-by H(x, y, z, 0) HO and Cm(x, y, z, 0) C!, respectively. We also
consider the boundary conditions of the type:

H(x, y, z, t) = iI on r~),
Ie aHA
r(2)
)

(5)

on

H,

(6)

on

rg),
r (2)
0'

(7)

qH ( x,y,z,t :=

-s an = qH

(8)

where n denotes the unit exterior normal to the boundary r of the domain,
and r r~) + r~) = rg> + r~)j the partition of the boundary.

281

Boundary Integral Equations


Let rand ri are two positional vectors with the coordinates (x, y, z) and
(Xi, Yi, Zi), respectively. Put r' = r - ri and t' = r - t. Moreover, put K =
kjS and Km = Dj Rm. The flow velocity is splitted into a sum of constant
reference velocity plus its deviation. We denote Vm = uj Rm = Vo + v'.
The time-dependent fundamental solutions used for seepage equation and
advective diffusion equation are given by
*(

(S )3/2

H r,t,r"r = 41Tkt'

(Srt2)
exp - 4kt' ,

(9)

C*m (t.
1 )3/2
r , ,ri,r) -_ ( 41TK
mt'
')
Vo2t'
.\ I
r 12 )
( (Vo . r
2Km - 4Km - mt - 4Kmt' ,
exp

(10)

respectively, where r' = Ir'\' Vo = Ivol.


We put the corresponding boundary fluxes of these fundamental solutions
in the normal direction as
p~(r,t;ri,r) =

-K

oH*

an

(r,t;r"r),

(11)

p:(r,t;r;,r) = -Km o~;;'(r,tjr;,r)


-(vo' n)C;'(r, t;r"r).

(12)

Following the conventional mathematical procedure, we can derive from the


above governing equations the boundary integral equations respectively in
the form

e;H(r"r)
=-

10'"

-1o'r hP~(r,tjri,r)H(r,t)t:Irdt

fr H*(r,

tj rj, r)qH(r, t)t:Irdt

In H*(r,

OJ r" r)HO(r)dO

+ 10r 10r H*( r,tjr;,r )Q(r,t)d


-s- 0 dt,

e,Cm(r;, r)
= -

fr

-1o'r hP;"(r, t; ri, r)Cm(r, t)t:Irdt

Io-r C;'(r, tj r"

+ R;:'l Am_l!oT

- loT

(13)

In C;'(r,
rr

r)qm(r, t)t:Irdt +

In C;'(r,

tj ri, r)Cm -

In C;'(r,
1 (r,

0; rj, r)C!(r)dO

t)dOdt

t; r" r)(v' . V')Cm(r, t)dOdt

C*(
) Fm(r,t)di\d
+1010 mr,t;r"r BRm Ht,
where

e, is determined from geometric property at the point rio

(14)

282

We assume that the variations of H, qH and Cm, qm with respect to the


time in a small time increment T can be neglected. Let a = rrJ. /( 4Kt). Then,
the time integrations of H*, PH and C;;., q;;' can be carried out, respectively,
in the form

~lT(r,ri):= foT H*(r,tjri,T)dt


1

r'

= 411" Kr' {I - erf( v'4K Tn,


PHr(r, ri)

(15)

= foT PH(r, tj ri, T)dt


v'ierfc

(r' n)

= 211"3/2,-3 {v'a exp (-a) + T

( v'an,

(16)

and

r'
+erfc( ~
4KmT

V2T

Am T + -K)

2r'
v2
exp(- v'4Km Am + 4K)},

(17)

and also

PmT(r, ri)

:=

foT p;"(r, tj ri, T)dt


1

= (411"K
)3/2 exp( ,m

(vor')

4Km )

(r'
{- (vOr,){v'41rKm(
4r' erfc ~+
2r'

.1

eXP("'/4KmVAm

r'

+erfc(~ -

,----".--

\
v2T)
AmT + 4Km

V2

+ 41<:)
V 2T

Am T + 41<:)

2r' .1
V2
exp( - .../4Km Am + 4Km)}
4K
rrJ.
V2T
+(r' . n){ ,; exp( - - - - Am T - - - )
2r
4KmT
4Km

r=Km..;:rK;.

+v 11"-;=2 ""Tr' -

.1Am + 4Km
v2 )
V

283

(18)
Using these time integrations, we obtain the following integral equations at
the time r.

-ir PHT(r, ri)H(r, r)dr

0 iH(ri, r)
=

-ir WIT(r, ri)qH(r, r)dr + 10 H(r, 0; ri, r)HO(r)dn


f
Q(r, r)
+ in w1T(r,ri)-s-dn,

(19)

0 i Cm (ri, r) -irPmT(r, ri)Cm(r, r)dr

wzmT(r,ri)qm(r,r)df + 10 C;'(rJO;ri,r)C~(r)dn

=-

+R;~l ),m-11o wzmT (r,ri)Cm _ 1(r,r)dn

- 10 wzmT(r, ri)(v' . \i')Cm(r, r)dn


r) d
+ inf W2mT (r,ri) Fm(r,
eRn.

(20)

H.

Using constant elements in space, these equations are discretized further in


the form

0 i Hi( r) -

E Hi ( r) lrj PHT(r, ri)df


j=1
N

= - EqH;(r) f

irj

j=l

M Q,(r)
+E
s-

wIT(r,ri)r + EH? f H(r,Ojri,r)dn

1=1

0 i Cmi(r) = -

n,

1=1

in,

wIT(r,r;)dn,

~ Cmj(r)

ir/mT(r, r.)df

L qmj Jrj( W2mr(r, r.)r + 1=1


L C~, Jn,f C;'(r, 0; ri, r)dn
j=1

R;-l
Hm

),m-l

1=1

Cm - Ar)

f WZm,.(r, ri)dn

in,

(21)

284

(22)
The product form of 4 x 4 Gauss integration is used for the numerical
evaluation of the boundary integrals, and the product form of 2 x 2 x 2 Gauss
integration is used for the numerical evaluation of the volume integrals.
Resulting set of equations is a system of algebraic equations. After the
boundary conditions are inserted, the system can be solved numerically step
by step by using a time-marching scheme for each time step t" = t"-l + t::.t.

Numerical Examples
We shall consider the three-dimensional migration in a 100 x 100 x 100m3
cube with a decay chain U234 ~ Th 230 ~ Ra 226 Figure 1 shows boundary
element nodes and flow vectors. Figure 2 shows calculated concentrations
of U 23 \ T h 230 , and Ra226 at 105 years after the disposal.

Figure 1.

Boundary element mesh (a) and groundwater flow vectors (b).

285

Figure 3 shows the history of concentrations U23\ Th 230 , and Ra226 for the
time interval of 10" years.

~x

Figure 2.

Calculated concentrations of U234 (a), Th 230 (b),


and Ra 226 (c) at 105 years.

Conclusion
Boundary element method was applied to the prediction of transient threedimensional coupled groundwater flow and migration of nuclides for a long
time interval. An example of three-dimensional migration with a radioactive decay chain is presented. The method is useful to large scale long
time prediction of disposal of radioactive nuclides coupled with transient
groundwater flow.

286
UliHi
a?5GI
7.SG(:-SI

6zse:-SI

(a)

5 . . . $1
3.?$E-$t

2.o:a::-OI
1.25-01

(b)

".$OE-$2

(c)

1Qg(g2

_hh ....
..cuh...

Figure 3.

1.~Q.l

S.S9OlQ-Q;2

Concentration history of U234 (a), Th 230(b),


and Ra226 (c) for 105 years.

287

References
[1] M. Kano, Y. Hara,T. Kuroki, and G. Aramakij Boundary element
method for saltwater wedge diffusion. Proceedings of the 3rd Japan
National Symposium on BEM, JASCOME, pp.203-206 (1987).
[2] M. Morishita and N. Tosaka; Approximation of salt wedge diffusion
problems by the generarized boundary element method. Proceedings
of the 6th Japan National Symposium on BEM, JASCOME, pp.81-86
(1989).
[3] R. Kawamura and K. Onishi; Three-dimensional transient coupled
analysis of groundwater flow and nuclide migration by boundary element method. Advances in Boundary Element Techniques, Eds: J. H.
Kane, G. Maier, N. Tosaka, and S. N. Atluri, Springer-Verlag (in press),
1991.

Trefftz-lYPe Boundary Elements for Plate Problems


R. Piltner

Department of Civil Engineering, SEMM, University of California at Berkeley,


Berkeley, CA 94720, U.S.A.

Summary

For isotropic and anisotropic plates, in extension and under bending, solution representations in terms of arbitrary complex functions are used. With the aid ofthe solution representations, one can systematically construct infinite series oflinearly independent trial
functions for the displacements and stresses. For the complex functions Cauchy integrals
are chosen which are discretized along the boundary of the solution domain. The shape
functions for every boundary element must be chosen such that all limit values on the
boundary exist and all displacements and stresses remain finite everywhere in the solution domain and on the boundary. Since the trial functions satisfy the governing differential equations, they can be considered Trefftz-trial functions. The term "Trefftz-type
boundary elements" is used for the description of the presented algorithm because the
(Treffiz-) trial functions depend on a boundary element discretization.
Introd1.1 ction

In the Trefftz-method, trial functions are used which satisfy the governing differential
equations. The original Trefftz-method [1) can be generalized, and Trefftz-type finite elements [2-4) and boundary elements [4-8) can be derived.
It should be noted that the 3-dimensional displacement and stress representations
(listed in the next section) give us the variation of the physical quantities in the thickness direction. In the considered 2- and 3-dimensional representations, the complex functions are defined in the midplane of the plate. The midplane n is enclosed by the boundary r. Moreover, it should be noted that in the 3-dimensional cases the displacements
and stresses are defined in the volume domain V and on the surface S.
The 3-dimensional stress and displacement representations for plates are constructed
such that the boundary conditions on the upper and lower plate faces (i.e. on z = bI2)
are satisfied a priori. Using the 3-dimensional plate representations satisfying the governing differential equations and the boundary conditions on z =bI2, we are left with
the task of satisfying the boundary conditions on the lateral faces of the plate.

289

For the Trefftz-type boundary element algorithm considered in this paper, the following
steps are necessary in order to construct the approximation functions:
(i)

use a complex representation of the displacements and stresses,

(ii)

use Cauchy-integrals for the complex functions,

(iii) discretize the Cauchy-integrals along the boundary

r with the aid of boundary

elements,
(iv) assume complex basis functions containing nodal values for every boundary
element. (Remark: The basis functions have to be chosen such that all limit
values for the needed derivatives of the complex functions exist on the boundary.),
(v)

integrate the Cauchy-integral contributions for every boundary element analytically,

(vi) get displacements and stresses which (although they contain singular functions with poles on the boundary) are finite everywhere in the solution domain
V and on the boundary S. (For a 2-dimensional problem, the domain is denoted
by n and the boundary is f.)
The constructed approximation functions for the displacements and stresses depend on
the discrete values of the complex functions at the boundary element nodes ofthe chosen
bouncary element discretization. The boundary element nodal values are the coefficients
of the linearly independent trial functions constructed. We get the following form for the
approximation functions:

displacements: u

=U c + up

in V and on S

+ O"p in V and on S

(For 2 - D: in n and on r)
(For 2 - D: in n and on r)

stresses:

0'

tractions:

T=T'c+Tp onS (For2-D:onf)

(j. C

(1)

(2)
(3)

The vectors U', a', T* contain the linearly independent trial functions obtained from the
boundary element discretization. up, ap and Tp contain the particular solution for a
given load case. The vector c contains the real and imaginary parts of the nodal values of
the complex functions.
Since the approximation functions already satisfy the governing differential equations,
the task remains to calculate the nodal values c. For the calculation of the unknown
vector c, the following methods can be used:

290
(i)

point collocation (We require that the boundary conditions are satisfied at
selected points on the boundary and get a nonsymmetric coefficient matrix).

(ii)

variational formulation, containing only boundary integrals (gives a symmetric


coefficient matrix).

(iii) least

square method, for the minimization of the difference between prescribed


boundary values and approximated values on the boundary (gives a symmetric
coefficient matrix).

The simpliest method is to use of a point collocation procedure for the satisfaction of the
boundary conditions since no numerical integration along the boundary is needed for the
derived Trefftz-boundary element algorithm. In the point collocation method, the number of collocation points is chosen such that we get M linearly independent equations for
M unknowns. The possibility of using additional points and getting a solution from a
rectangular system of equations is discussed in reference [9].
Since the unknowns in this algorithm are the real and imaginary parts of the complex
functions at the boundary element nodes, we have an indirect boundary element algorithm. After we have calculated the complex node values, from the requirement that the
boundary conditions are satisfied in some defined optimal sense, we can immediately calculate the real displacements and stresses everywhere on the boundary and inside the
solution domain.

Complex representations for plate problems


In the following a list of homogeneous plate solutions is given in complex form:

Plane stress (and plane strain) in isotropic bodies

2jlU

=Re[K"<l>(z) -

z<1>'(z) - 'fez)]
(4)

2jlv = Im[K"<l>(z) - z<1>'(z) - 'fez)]

where
z =x +iy,
2jl = E/(l + v)

(5)

291

K={(3-V)/(1+V)
(3-4v)

for plane stress


for plane strain

Plane stress (and plane strain) in anisotropic bodies

(6)
where
Zl =
Z2

+ ,ulY =(x + ay) + iPy,

(7)

= X + .ll2Y = (x + yy) + ioy,

The complex coefficients Ph P2, ql, and ~ depend on the elastic compliance coefficients
aij, and,ub .ll2, fih il2 are the solutions of the characteristic equation

(8)
wbere
for plane stress
(i,j = 1,2,6)

for plane strain.

(9)

Bending of isotropic thin plates (Kirchhoff-plate theory)


w(x,y) = Re[zct>(z) + 'l'(z)]

(10)

Bending of anisotropic thin plates

(11)

where
Zl

=X + .lllY ,

Z2 =X + .ll2Y ,

and.llb .ll2, fil, fi2 are the solutions of the characteristic equation
(12)

292
where
(13)

The coefficients cij are the elastic constants and h is the thickness of the plate.

Three-dimensional in-plane solution involving powers of z

2JlU

=Re[

3- v
--; --;
V
h2
2
"
1 + v <11- (<II - Z ] - 2 1 + v [ 12 - z ] Re[ <II ],

3-v
V
h2
2
"
2JiV = Im[ 1 + v <11- (<II' - z' ] + 2 1 + v [ 12 - z lIm[<II],

(14)

v
,
2Jiw = -4z - - Re[<II],

l+v

where

(15)

( =x+iy.

Three-dimensional bending solution involving powers of z and trigonometric


functions of z

-12
z3
"
2Jiv =- z Im[ <II + ( <II' + z' ] + - - [ h z - 2(2 - v) -3 ] Im[ <II ],

I-v

(16)

2Jiw = Re[ ( <II + Z ] + -

2v

I-v

'

z Re[ <II ],

where gn(x,y) has to satisfy

(17)
and

aln

= ll1r1h (n=I,3,5, .... ).

Complex representation of gn(x, y):

293

(18)
or

JRe[(<1>~(t() Io(OJnrv'I'=t) dt,


1

gn(X, y) = Re[<1>n(O) Io(OJnr) +

(19)

where
10 =modified Bessel function of the first kind

r= ""x2 + y2.

(20)

The representation (18,19) is valid for a star-shaped domain. For engineering applications it is sufficient to use, from the series of functions gn(x,y) (n=I,3,5, ... ) in the representation (16), only the first function. A more detailed overview about complex representations is given in [4], where the corresponding formulas for the stresses are also given.

Construction of linealy independent trial functions


By using the complex representations (4)-(19) we ensure that with any choice of complex
functions the governing homogeneous differential equations are automatically satisfied.
Therefore the solution representations in terms of arbitrary complex functions are very
useful for a convenient construction of linearly independent trial functions.
For numerical applications, the complex functions are chosen with respect to the selected
method. For Trefftz-type finite element applications [2,4], we can use complex power
series for the functions <1>, Z, 'P and <1>n. In order to obtain a boundary element algorithm
the complex functions in the representations (4)-(19) are chosen in such a form that we
will have only unknowns on the boundary. This will be discussed in the next section.
In addition to the homogeneous solution contributions we need particular solutions. For
the example of a Kirchhoff-plate under a constant normal load p(x,y) =qo we have the
particular solution

w = 4~~ (x4 + y4).

(21)

Other particular solutions for plate problems are given in [4J. The particular solutions
satisfy the differential equations and take a load case under consideration into account.
There are no free parameters in the particular solution. All unknowns are in the homogeneous solution part.

294

Boundary element discretization


The complex functions <1>, X, 'I' and <1>n. in the representations (4)-(19) are chosen in the
form of Cauchy integrals which are evaluated along the boundary r of the plate midplane. Since for all complex functions the same kind of approximation is used, the discretization will be explained for one of the functions, say <l>(z). (Remark: Here the complex variable is z = x + iy. For functions of the other complex variables (, Zh Z2 the notation in the Cauchy integral representation will have to be changed appropriately.) After
dividing the boundary

into N boundary elements with the boundary portions

rj

G=I,2, ... ,N) the discretized function <l>(z) can be written as


1

<l>(z) = -.

LN

21rl j=l

,
-<V(O d( = -1. LN <I>l(z),
r (- z
2m j=l

(22)

where ( is a complex boundary coordinate. Using fifth-order shape functions Nj(s)


G=I,2, ... ,6) which are given in [4,5], the complex basis function for a boundary element
between nodes G-l) andj can be written as

(23)
where
( -_
Z'-l
s= _
J_,
Zj - Zj-l

(24)

and <1>j, <1>j, <1>; are discrete complex node values. The real parameters

are used to form the vector c of the representations (1)-(3). If we use straight line boundary elements the integration in (22) can be performed analytically. As result we obtain
an approximation function of the form

(25)
The functions F/z), Giz), Hj(z) are linearly independent and contain polynomials of z
and a logarithmic term which is multiplied by a shape function. For the example of fifth
order basis functions we get for Fiz), Gj(z) and Hj(z) the following expressions:

295
2
Zj-Z }
5
- - -K-3K
+ 11K3 -6K4 +2Ns(K)ln-G(Z)=(Z -Z'_l{ -11
J
]]
30 6
Zj-l - Z

(26)

Zj+l - Z }
+ (Zj+l - z { -4 - -13 L - 6L2 + 13L3 - 6L4 + 2N2(L) In ] 5 6
Zj-z

H-(z) = (Zj - Zj-l)


J

2{ -1

30

3 4

Zj-Z
+ -1 K + -1 K - -3 K + K + 2N6(K) In -}
12
3
2
Zj-l - Z

Zj+l - Z }
+ (Zj+l - Zj) 2{ -1 + -1 L - -11 L2 + -5 L3 - L4 + 2N3(L) I n
-20 4
6
2
Zj-z

where

K - z- Zj-l
Zj - Zj-l

L=

Z-Zj

(27)

Zj+1 - Zj

For more details see [4,5].

Numerical example
Tile numerical results for a variety of problems show that highly accurate values for the
displacements and stresses can be obtained with the Trefftz-type boundary elements [4].
Table 1: Results for a simply supported rectangular plate (a=6, b=4. h=0.2, E=l, v = O. 3,
uniform load p=l) for a discretization with 8 boundary elements.
h=0.2

w(x=O,y= O,z=O)

numerical
(8 elements)

exact (3-D)

error

[11]

2.7213 * lOS

2.7214 * 103

0.004

O'xx(x = 0, y = 0, Z = -hl2)

-1.198 * 102

-1.200 * 102

0.2

O'yy(x = 0, y = 0, Z = -hl2)

-1.950 * 102

-1.951 * 102

0.05

* 101

1.233 * 101

3.1

"ryz(x = 0, y = -bl2, Z = 0)

1.271

296

Some results for a simply supported rectangular plate (-a/2 S; x S; a/2, -b/2 S; y S; b/2)
with the thickness h=0.2 are given in Table 1. The numerical results obtained with the
Treffiz-type boundary element concept are compared to an exact three-dimensional solution. For every boundary element seventh-order basis functions have been used for the
Cauchy integrals.

References
[lJ E. Treffiz, "Ein Gegenstiick zum Ritzschen Verfahren", 2. Int. Kongr. f. Techn.
Mechanik, ZUrich, 131- 137, (1926).
[2] R Piltner, "Special finite elements with holes and internal cracks", Int. J. Numer.
Meth. Eng., 21,1471 - 1485, (1985).
[3] R Piltner, RL. Taylor, "The evaluation of stiffness matrices for elasticity problems
with the aid of boundary integrals", pp. 38 - 45 in: "NUMETA 90: Numerical Methods in Engineering: Theory and Applications", (Eds. G.N. Pande, J. Middleton),
Elsevier, LondonlNew York 1990
[4] R Piltner, "Finite and boundary elements for plate problems: complex function representations and generalized Treffiz-solutions", submitted.
[5] R Piltner, RL. Taylor, "A boundary element algorithm using compatible boundary
displacements and tractions", Int. J. Numer. Meth. Eng., 29, 1323 - 1341, (1990).
[6J R Piltner, RL. Taylor, "A boundary element algorithm for plate bending problems
based on Cauchy's integral formula", pp. 200 - 206 in: Proceedings of the International Symposium on Boundary Element Methods, United Technologies Research
Center, East Hartford, Connecticut, USA, October 1989: Boundary Element Methods in Engineering, (Editors: B.S. Annigeri/ K. Tseng), Springer, Berlin, Heidelberg,
New York, 1990
[7J R Piltner, RL. Taylor, "A boundary element procedure for plane elasticity based on
Cauchy's integral formula", pp. 15 - 25 in: Proceedings of the Eleventh International
Conference on Boundary Element Methods in Engineering, Cambridge, Massachusetts, USA, August 1989: Advances in Boundary Elements, Vol. 3: Stress
Analysis, (Editors: C.A. Brebbia/ J.J. Connor), Springer, Berlin, Heidelberg, New
York,1989
[8J R Piltner, R.L. Taylor, "The solution of plate bending problems with the aid of a
boundary element algorithm based on singular complex functions", pp. 437 - 445 in:
Proceedings of the 12th International Conference on Boundary Element Methods in
Engineering, Hokkaido University, Sapporo, Japan, September 1990: Advances in
Boundary Elements, Vol. 1: Application in Stress Analysis, Potential and Diffusion,

297

(Editors: M. Tanaka! C.A. Brebbia! T. Honma), Springer, Heidelberg, 1990


[9] E. Ramm, "Beitrag zur praktischen Berechnung dunner Kugelschalen bei nicht
rotationssymmetrischer Berandung nach der linearen Biegetheorie - Anwendung
der verbesserten Kollokation (diskretisiertes Fehlerquadratminimum) als Randverfahren", Doctoral Thesis, Universitiit Stuttgart, 1971.
[10} R. Piltner, "Three-dimensional stress and displacement representations for plate
problems", Mechanics Research Communications, 18 (1), 41- 49, (1991).
[11] R. Piltner, "The application of a complex 3-dimensional elasticity solution representation for the analysis of a thick rectangular plate", Acta Mechanica, 75, 77 - 91,
(1988).

Brittle Fracture Materials by BEM and R-Functions


V. F. Poterasu and N. Mihalache
Theoretical Mechanics and Structural Mechanics Departments
Polytechnic Institute of Iasi, Romania

Summary
The paper presents a combination between Boundary Element Methods and RFunctions Method for analysis of cracks propagations in elastic bodies based on the
new failure criterion developped by the authors. After theoretical presentation of displacement discontinuity method as the special variant of BEM in fracture mechanics
and the R-fundions Method is studied a numerical example concerning the behaviour
of a shear wall made by reinforced concrete.

Introduction
In the processes of structural design of the reinforced concrete or machine components
one of t!le most important step concerns the determination of the dimensions and the
material selection. This is usually done by applying a "failure criterion" [4] which in
general includes a comparison between the "critical load factor" representing the intensity of the applied load and the geometry of the part of material with a "characteristic
strength parameter". The authors propose an improved new relation for the MohrCoulomb failure criterion adapted for the reinforced concrete. This new relation has
the errors between 2-4 % in comparison with the classical Mohr-Coulomb failure criterion which presents errors between 10-30 %. BEM has been adapted for the fracture
mechanics analysis of two and three dimensional bodies. Inherent modelling accuracy
and efficiency have also lead to the use of BEM techniques for elastic fracture mechanics
analysis. Consider the geometry, illustrated in Fig.1 of a solid containing an open cavity
approximating a crack. Let S,

r+

and

r-

represent the regular, upper and lower crack

surface. The Somigliana identity of the displacements in an interior point p( al) and
boundary point m( a:) is given by

299

p+

p-

Lc: 2a

Fig.1 The cracked element

'U;(p)

=-

T;j(p,m)'Uj(m)ds

Js+r++r-

Uij(p,m)tj(m)ds

Js+r++r-

(1)

The kernels in eq.(l) are defined in Cruse [5J where it is seen that

(2)
except near the small crack tip region. Letting the two crack surfaces collapse to the
plane

r , eq.(1) is seen to give


'Ui(p)

-1
-L

T'j(p, m)'Uj(m)ds

T;j(p,m).d'Uj(m)ds

1
L

Uij(p, m)tj(m)ds

(3)

U'j(p,m) Lti(m)ds

where

Letting p( z) be a point on the plane


boundary integral we obtain
'U,(p} - .d'U,(p)/2

+
=

1
1

and taking the usual term for the jump in the

Tij(p,m}'Uj(m)ds
U;j(p,m)tj(m)ds

Tii(P, m).d'Ui (m)ds

(5)

300

CT;;

1
+L
=

D"ij(p,m)t,,(m)ds

-1

D"ij(p,m) Lt,,(m)ds -

S"ij(p,m)u,,(m)d,

(6)

S"i;(P, m)..1u" (m)d,

Displacement Discontinuity Method in Fracture Mecha.nics


Improvement and extensions in the fracture mechanics has given by Crouch and Starfield
[2J with the displacement discontinuity method. If the displacement discontinuity is
presented in the form

and the stresses, displacements are

+ Y/,:lIlIY) + 2GDlI (f,lIY + yf.YlIY)

CT., ..

= 2GD.,(2/,"1I

CTlIlI

= 2GD.,(-yf.:OYlI) + 2GDlI (f.1I11 - yl.lI'J11)

CT"y

= 2GD .. (f.lIY + yf.J'JY) + 2GDy ( -yl.lIYlI)

u .. =
Uy

D.,2(1 - 1/)/.11 - yl... .,

= D.. (l - 21/)/... - yl'''l1

+ DlI -

(7)

(1 - 21/)/... - y/..,..

+ Dl/2(1 - I/)/.l/ - yl.1I1I

(8)

It is easy to verify that the displacements are continuous everywhere in the infinite

body, except across the line segment


stresses along the line y =

lzl :::; a, y = 0, where they are discontinuous.

The

are given by

aG

CTl/l/(Z,O)

CT"'l1 z,

0)

= - 'Ir (1 -1/ ) D lI -z-a


2
2
=

aG

D __1__

(9)

(
).. z-a
2
2
'lr1-1/

In the problems of fracture mechanics is necessary an accurate solution near crack tip,
and a more sophisticated approa.ch is needed, developing the higher order displacement
discontinuity elements.

301

R - Fuctions Method
An attempt has been made to apply the R-functions method(RFM) to the linear elastic
fracture mechanics problems. Essential feature of this method consists in a conversion of
logical operations performed on sets into algebraic operations performed on elementary
functions. For a domain {} let us consider a boundary value problem given in the local
formulation

Au = f

in {}

c am,

u c X({})
m

l,ui = <Pi

on Si C S,

i = 1,2, ... , m

S=

USi
i=l

or in the global equivalent representation

r F(u)d{} + 15rG(u)da = min"

in

Here A. Ii. G are operators defined over a space of functions defined in the domain {}
and on S respectively. The general structure of the solution(GSS) of the problem is a
mapping B : m

X( 0) that satisfies the following conditions

->

liB(p)=<pi; i=1,2, ... ,mj VPEm


in whkh m is a function space defined over {}. Taking this into account we may present
the GSS in a more specifie form B = B( q; , 10 , <p) where

10 :

am

-+

R is a mapping that

defines the domain

0= {X E am, 1o(z) > O},


<P : R m -+

S = {X E Rm, 1o(z) = O}

R is a mapping defined over the domain 0 that strictly satisfies all prescribed

boundary conditions. The construction ofR-functions is presented in [7J, [8] . Consider


now a brittle cracked body, Fig.I. Let us assume that the stress and displacement field
in it may be obtained from a solution of the well known 2-D elasticity problem

given in terms of the Airy's stress function F , where

N(a) = t

15

p(.)ds

M(s)

tiS

(p.,y - p,lz)d.

302
are respectively the resultant of the shear forces and the moment alonhg the boundary
of the body due to an uniformly distributed load p(P"'PI/) . One of the possible ways
to get the required GSS is to find it as a sum of three components B

= Bl + B2 + Ba

that satisfy the following conditions on S

In order to determine the function Bl we may use the technique based on the formula
I

q, =

I: wi 1 q,i

:..:i=:.::~,- __

07

I: wil

q, = L

q,; IT

Wj (j

=f. i)

_...:..i=_l..,..,_ _ __

;=0 Wi

;:0

+ IT Wj{j = i)
j=l

The domain 0, Fig.I , may be defined by the theory of set operations F applied to
the known simple subdomains OJ, i
equation

Wi

= 1,2, ... , k

= O. The required function

logic F. In order to construct

each of which is determoned by an

is the R function associated with a function

it is sufficient to use the R-functions related to union

and intersection only. To obtain functions

q, is of numerical nature and it

begins with

discretization of the yet undetermined "free" function


M

~ = L~IINII(z)

k = I,2, .. ,M

11=1

We will obtain a function

q,

that meets the condition

~Is

= M but

~.mls

may still by

different from zero. It would be useful then to introduce the so-ca.lled normalized WN
function [8J. A function

is normalized to the k order if besides wNls = 0 it satisfies

the following conditions


WN ... S = 1

wN."il s

= 0,

i = 1,2, ... , k

in which n denotes the unit normal oriented toward the domain. The function WN
normalized to the first order may be obtained from the formula

303

If WN is used the following formula for differentiation of any function 9 taken in the

direction n normal to the boundary S holds

We may now write Bl

n , such that

= cP - WNcP

in which cP is any function defined in the domain

cP*IS = cP.m. Actually we have then

Finally we may write the GSS in the form

where the operator Dn is


D,.

= /)WN ~ + /)WN .!!....


tJz /)z

tJy tJy

where the last term involves a free function .p while the others satisfy the nonhomogeneous equation.
Reinfort:ed Concrete Knot for Pillar - Beam Connection
Using numerical methods studied above has been made analysis of a knot of reinforced
concrete of class 22.5, E=300,000 daN/cmsq.

,1/

= 0.2. The action of the knot is

shown in Fig.2a. The uniformly distributed force is considered constantly in Z direction


and the stress state analysis is made for a unity length, Fig.2b. The load necessary
for crack Pe was been determined experimentally by means of ultrasonic method. The
appearance and crack propagation have been made by boundary element method in the
nrst variant (constant interpolation functions and discontinuous displacement method
presented above) and the second variant (linear interpolation functions). The stress intensity factors are numerically determined by means of R-fundions. The discretization
in boundary elements, loads a~d bearings are presented in Fig.3a. In Fig.3b is shown
the discretization of the cracked model. The stress concentrations have been made by
means of a new improved Mohr-Coulomb criterion developed by the authors [4].

304
0"1

= O"s + (0"1

+ O"a) sin,8 =

c(2 cos,8 + tan,8)

(10)

b)
a)

P/lO

Fig.2

The Knot Model

p :70,7

Pc

l'IE'ment

Py

noncracked

19

16

..IE'ment

cracked

p~

39
40

b)

a)

Fig.3

The Discretization Model

305

In the classical Mohr-Coulomb model the relation is

(11)
By means of the experimental researchs are determined the cohesion c and the internal
friction angle [3]. Using the relation (11) the error is approximately 23.26% and with
the relation (10) is 1.2888%. For the reinforced concrete of the class C22.5 have
been experimental and statistical researches with the results 0"1=29.75 daN/cmsq. ;
0"]=-253 daN/cmsq. and the cohesion c=43.88 daN/cmsq. ,the internal friction angle

f3 = 5163'. Replacing these values in (11) the error is


(10) the error is

~=0.928%.

~=50.57%

and with the relation

The relation (10) has been introduced in a routine of

computer programbased on boundary elements for stress analysis in fracture situation.


The crack of the knot is made in the diagonal compressed and the stress diagram 0"11 is
presented in Fig.4 .

27

......

37

-0---

39

Fig.4

--1

~n
hf

--

~cracked (Gy ~ ",;

~'''''i~noncrnc~~d G).

\
\

- .....

....0" ", ;

i~
40

41

42

10

knots

The stress 0"1I diagram on the knot diagonal

The stress concentrations are presented in the points 41-42 and 37-46 from Fig.3b .
The culation of the stresses with the constant boundary elements(*) I linear{ 0) in the
non cracked situation( -) and cracked( - - -) shows that the results are better in the
linear variant, Fig.4, using the criterion (10) . The stresses and stress intensity factors
determined by means of the constant, linear boundary elements and R-functions are
presented in the Table 1 .

306

r(em)

0.01

0.1

0.3

(TO

63.05

62.63

61.92

59.10

56.42

54.79

53.59

KrO

421.16

132.3

75.3

39.48

26.65

21.13

18.02

KR
r

413.01

129.2

71.5

37.15

25.09

19.98

16.85

1.97

2.39

5.31

6.27

6.22

5.76

6.94

Table

Stress variation on the segment 40-41, Fig.4, determined by means of the boundary
linear elements in the cracked state is given by the parabola from Fig.5 .

37

em
6

5
4

3
2

,
41

,5

132.3
Fig.5

Stress Variation

O'll

and Kr

Conclusions
By means of numerical methods, the cracks appearance and propagations may be obtained more easily than those by experimental methods. It is possible to add new routines including new failure criteria. The R-functions give an actual and accurate value
of the variation of stress intensity factor. The error obtained for two variants analysed
may be comparaed with that by photoelasticity a.nd other experimental methods.

307

References
1. Atluri, N. S. ; Kobayashi, S. A. : Elasto-plastic Fracture (quasi-static) in Ed.Atluri,
N. S. :Computer Methods in the Mechanics of Fracture. ,sec.2(1985),55-84,NorthHolland.

2. Crouch, S. L. ; Starfield, A. M. : Boundary Element Methods in Solid Mechanics.(1983) George Allen & Unwin, London.
3. Mihalache, N. : Behaviour of the Shear Walls Made by Reinforced Concrete to the
Shear Forces Action.(1988) Dr.Thesis, Polyt. Inst. Jassy, Civil Eng. Fac. , Romania
4. Mihalache, N. , Poterasu, V. F. , Secu, AI. : Numerical and Experimental Studies
Concerning New Failure Criteria for Reinforced Concrete Elements, Proc. 10-th Int.
Conf. struct. Mech. in Reactor Technology (SMIRT), Q/02(1989), 53-58, Anaheim
USA.
5. Cruse, T, A. : Two and Three-dimensiona.l Problems of Fracture Mechanics. Ch.5 in
Eds. Banerjee, P. K. , Butterfield, R. : Developments in BEM.(1979)97-120, Applied
Science Pubi.
6. Peterasu, V. F. , Miha.lache, N. : Numerical Simulation for Crack Propagation
in a Flexible Dam by BEM Modelling. Simulation & Control, AMSE Press, B,
(1986),7,3,57-63.
7. Orkisz, J. , Rachowitz, W. , Wnuk, M. : Evaluation of stress Intensity Factors by
R-Functions Method. Int.J .Fracture ,24,( 1984) 107-126.
8. Rvachev, V. L. :Theory ofR-Functions and Their Applications (in russian). Naukova
Dumka(1982), Moscow.
9. Tanaka, M. , Hamada, M. ,Iwata, Y. : Computation of a Two-dimensional Stress
Intensity Factors by the BEM. Ingeniour-Arehiv, (1982)95-104.
10. Poterasu, V. F. , Miha.lache, N. : BEM in Elastic Torison Problems Using
Green's Generalized Theorem and R-Functions. Bull. Polyt. Inst. Jassy,32,s.5,f.l4(1986)53-60.

The Sweeping Primitive used for 3D Automatic Boundary


Element Mesh Generation
Noriaki SASAKI, Toshihiko KUWAHARA and Tsuyoshi TAKEDA
Department of Electrical Engineering,
Faculty of Engineering, Hokkaido University.
Kitaku Kita 13 Nishi 8, Sapporo, 060, JAPAN.

ABSTRACT
In this paper, an automatic mesh generator is proposed for the Boundary
Element Method (BEM). The mesh generator can make a complicated shape from a
few data on a personal computer by adopting the Constructive Solid
Geometry(CSG) method [1]. Four primitives (a rectangular parallelepiped, an
ellipSOid, a circular cone and a cylinder) have already been defined [2].
The objective shape can be constructed by logical operations <UNION, JOIN
AND SUBTRACTION) among primitives having a simple mesh on it. Every
primitive has an in-out function which judges whether some point is inside
the primitive or not, because this is required to perform logical operations. However, figures such as spiral springs could not be produced from the
previous primi tives. A new added "Sweeping Primi tive" is made by sweeping a
two-dimensional (2D) figure and it enables us to make a great variety of
figures.

Fig. 1 preVious primitives

Fig. 2 Logical operation

309

2D PRIMITIVE

L::::::nJ

First of all, we must obtain a


2D figure to make the sweeping
primitive. Three 2D primitives
(a po I ygon, an ell ipse and a
spline curve) are prepared
for making the 2D figure as
shown in Fig. 3.
The spl ine curve is a smooth
curve connecting given points.
I t is interpolated by the
following formula.
x( t) =ax t 3+bx t 2+Cx t+dx
y(t>=aytl+byt2+cyt+dy

Fig. 3

This equation is defined on


every curved
line between
two given points.
The caefficients( ax~x ,ay~dy )
are given by coordinates of
given points.

IN-OUT FUNCTION OF THE 2D


PRIMITIVE
The

in-out function of the


2D primitive is also defined
on every line. In order to
judge if Ji pOint is inside
or outside, we consider the
line which contains the x-value
of the point in its x-domain of
definition. Then, as in Fig. 4
LINE 1 and LINE 2 are used to

2D

primi Uves

V2

Ya
VI

LINE 4

~------~----~~X

Xa

Fig. 4

.&iven point

310

judge whether point "a" is inside or not. YI and Y2 are y-valuesof


LINE 1 and LINE 2. respectively corresponding to Xa. They are compared with
Va. When the number of the pOints whose y-values are bigger than Ya is
odd, point "a" is inside the primi tive. When it is even. the point is
outside the primitive. Since it is odd in Fig. 4, point "a" is inside .

Fig. 5 logical operation between 2D primitives

SWEEPING PRIMITIVE
In order to make a sweeping primitive from a 2D figure. there are basically
two methods of sweeping. One is an extrusive sweep as shownin Fig. 6. The
other is a rotational sweep as shown in Fig. 7. There are several functions
such as parallel translation. magnification. reduction and spin as shown in
Fig. 8. They can be decided by interpolated functions as shown in Fig. 9
and Fig. 10.

Fig. 6

Fig. 7

Fig. 8

311

Fig. 9

Fig. 10

IN-OUT FUNCTION OF THE SWEEPING PRIMITIVE


The 3D in-out function of the sweeping primi tive is obtained by
performing inverse coordinate transformation onto the original 2D figure.
after which the 2D in-out function is used. The logical operation can be
carried out between the sweeping primitive and the other 3D primitives.

Fig. 11 Logical operation with the other primitive

312

MESH GENERATION ON SIDES OF THE SWEEPING PRIMITIVE


The mesh of side face is generated automatically by using pOints of the 2D
figure and the number of division along a sweeping path.
When the front and rear of a swept 2D figure remain wi th no mesh, the
mesh must be generated on it as shown in Fig. 12 .

Fig. 12

We use a simple 2D mesh generator as follows. The node which has the
smallest angle is removed by forming a triangle element from the boundary
inwards as shown in Fig. 12. When the side length of the triangle is longer
than a cri terion, the side and the triangle are divided equally (Fig. 14).
This process is repeated on the newly formed boundary. After the inside is
fi I led wi th triangles, the mesh is justi fied automatically as shown in
Fig. 15 [2].

Fig. 12

Fig. 13

Fig. 14

Fig. 15

313

SAMPLES

NODES
ELEMENTS
CPU TIME

248
492
37(sec)

NODES
ELEMENTS
CPU TIME

915
1826
170(sec)

The computation is performed on a 32-bi't personal computer.


CONCLUSION
In this paper we have described a mesh generator system and a sweeping
primitive. The sweeping primitive enables us to make a complicated
figure easily.

REFERENCES

1 N. Okino: Methodology for Computer Aided Design, Youkendou, 1982


2 T. Honma, T. Kuwahara and T. Takeda: Development of the Automatic Mesh
Generation for the 3D Boundary Element Method, Boundary Elements Xl I,
vol. I, p.87 97, Computational Mechanics Publications, Southampton, 1990.

Direct Differentation Approach to Boundary Element Method


for Nonlinear Water Wave Problems
R.Sugino and N.Tosaka.
Department of Mathematical Engineering,
College of Industrial Technology, Nihon University,
Narashino, Chiba, 275,Japan

Summary
This paper discusses an application of direct differentiation method to boundary element
analysis for perfect fluid flow with free surface in a two-dimensional fluid region. The numerical solution procedure is essentially based on the boundary element method with the
so-called mixed Eulerian-Lagrangian approach. And, we introduce the direct differentiation method to get the Lagrangian velocity of fluid particles on free surface. The constant
boundary elements and the Euler time integration scheme in conjunction with smoothing
and relocation techniques are adopted in our new solution procedure. Numerical results
of solitary wave motion in a rectangular water tank with wave generator are shown.
Introduction
The motion of fluid with a free surface is observed in many fields of engineering. The
mathematical model of this motion can be given by the Laplace equation. A numerical
approximation procedure for computing the unsteady irrotational motion of an inviscid
and incompressible fluid with a free surface is developted within the boundary element approach. Boundary element procedure by the so-called mixed Eulerian-Lagrangian method
is very effective for highly nonlinear water wave problems. And, there exist many interesting studies[1]-[7].
In order to track motion of fluid particles on the free surface which satisfies the Laplace
equation and nonlinear boundary conditions, the Lagrangian manner is introduced effectively. In the solution method, we need the two spatial derivatives on x - y orthogonal
coordinates system. In the boundary element procedure, we can get directly normal
derivatives on the free surface. However, the tangential derivatives must be computed
by numerical differentiation. These dervatives are transformed to the derivatives with
respect to x and y by the geometrical relation.

315

In this investigation, we introduce the new boundary element scheme which is based
on the direct differentiation procedure instead of numerical differentiation. So that, a
hyper-singular boundary integral equation of potential gradients is used. The x and y
direction derivatives can be determined directly by this hyper-singular equation without
the numerical differentiation.
The solution method proposed in this study is applied to several examples on motion of
nonlinear water wave (i.e., solitary wave) in a rectangular water tank. We can show the
illustrations of three phases which are the generation, the propagation and run-up, and
the reflection.

Mathematical Formulation
Let us consider two-dimensional motion of a fluid in water tank. A fluid region
bounded by the boundary
wall

r wl

and

which is composed of a free surface part

rf

n is

and a wetted

r w2 of the water tank shown in Fig.I.

rw2

Fig.I.

Configuration of mathematical model

We assume that the fluid is inviscid and incompressible, and the flow is irrotational. Under
these assumptions we can get the following governing equations for the velocity vector

u = (u, v) and the velocity potential rf; of the fluid flow:

m
u
where \7

= \7rf;

in

= (fJ/fJx,a/ay) and \7 2 denotes the two-dimensional Laplacian.

(1)

(2)

316

There are two kinds of the boundary conditions to be prescribed. The first one is the wall
boundary condition given by

{ q

on
on

r ",1
r,.,2

(3)

where n denotes the outward unit normal vector on the boundary. The other is the
so-called free surface conditions on the free surface

r f.

They are the kinematic and

dynamic conditions. As the mathematical expressions of these conditions, we introduce


its Lagrangian description in terms of the Lagrangian coordinates (, ry) for a marked
particle on the free surface. Consequently, the kinematic conditions for a surface particle
are given by

D
Dt

= u == 04>
ax

Dry

04>

on

rj

(4)

-==v=-

Dt
oy
where DI Dt denotes the substantial differentiation. Next, we can express also the dy-

namic condition derived from Bernoulli's equation as the following equation on rate of
change of 4> :

D4>
Di

= ~ {(04))2 + (04))2}
2

ox

oy

_gry

(5)

in which we assume that the pressure applied at the surface is zero, and 9 is the acceleration of gravity.
In this paper, we consider the mathematical model given by equations (1)-(5) as the
coupled problem of the boundary-value problem of Laplace equation (1) and the initialvalue problem of the system of evolutional equations (4) and (5).
Numerical Solution Procedure
1. Boundary Element Method for Boundary-Value Problem

By taking into consideration of the linearity of Laplace equation (1), we can easily transform the field equation (1) into the following boundary integral equation:

-214>(Y)

= Jr,
( un
~4>(z)w(z,y)dr(z) + ( qw(z,y)dr(z) Jr..

( 4>(z) 0:* (z,y)dI'(z) (6)

Jr

un

in which w* is the well-known fundamental solution given by

w(z,y)=~ln~
211"

, q*(z,y) =

ow
an =~~(ln~)
211" an
r

,r=!lz-y!l

(7)

317

If the velocity potential on the free surface r f is known, then we can determine the value

of > on r tv and its derivative on r f with solution of the above boundary integral equation.
In order to solve approximately equation (6), we use the boundary element method. We
use the simplest element (Le.,constant element) in solving approximately (6) and then we
can get the following discretized expression of (6):
(8)

H~=GQ

where the influence matrices and the unknown nodal vectors are defined by, respectively,

,
l{O(
= [Hij]
= 21r
Jej (In In II a: _1)
a:, II dr(a:)

= [G ij ] = .Jrr:

1;

In

II a: ~ a:i IIdf(a:)

~=[I'2""'nY )
Q

(9)

= [qI, q2,""

(10)

qnjT

2. New Boundary Element Scheme by Direct Differentiation Method


Here, we introduce our new boundary element scheme to get the two spatial derivatives
on x - y orthogonal coordinates system. Boundary integral equation (6) is differetiated
with respect to Lagrange coordinates

~,

." which are defined on fluid particle. Then, we

can get next two hyper-singular equations as follows


1

{ o

{ 0</1

2'</I,e(Y) = pI Jr on (a:)w~(a:, y)df(a:) - pf Jr (:r::)q~(:r::, y)dr(:r::)

'i<P,'1(Y) = pi Jr on (a:)w~(:z:, y)dr(a:) -

pf Jr >(:r::)q~(:z:, y)df(a:)

(11)
(12)

in which, pf indicates that the integral is carried out in the sense of finite part.
We give the potential and flux values on whole boundary of the fluid region to these two
hyper-singular equations (11),(12), when we can get potential gradients o<P/o~, &>/&." on
the boundary. In using these values, we can construct directly the orthogonal components
of velocity vector (u, v) = (o<Pjox,o>/&y) on the fluid particle without the numerical
differentiation.

318

3. Euler Method for Initial-Value Problems


We can rewrite the kinematic and dynamic conditions expressed in terms of the rate of
change of the Lagrangian coordinates and of <jJ for fluid particles on the free surface as
follows:

o<jJ
ox

D
Dt

o<jJ
oy

1)

<jJ

(13)

! {(Mf + (Uf} -g1)

This system to be considered as the one of first-order ordinary differential equations can be
solved approximately by using the time integration scheme. Although there exist various
kinds of schemes, we adopt the Euler scheme known as the simplest time appropriate
integration scheme. Applying the Euler scheme to the above system, we can determine
the new value of {,
{Hi

1)

and <jJ at the (k

+ l}th time step as follows:

= {k + b.t (~r
(14)

where b.t denotes the short time interval and the superscript" k" indicates the k-th time
step. By using the above numerical integration scheme in time, we can determine the
time-dependant configuration of the free surface.
Numerical Computation Techniques
1. Smoothing Technique

In order to continue stably numerical performance to a sufficiently long time we introduce


some effective solution procedures in addition to the above-mentioned solution method.
First of all, the "saw-tooth" numerical instability discussed in [1] is removed by using some
kind of smoothing technique. To remove this type of instability, we adopt the following

319

5-points smoothing formula given in [8]:


-

h = 35 (-3h-2 + 12h-1 + 17 h + 12fj+1 The formula is applied for the calcurated values of

e,

1)

3h+2)

(15)

and if> at every 10 time-steps.

When we can not take the above 5 points at right or left end, the following modified
formula is applied:
At right side,

(16)
At left side,
(17)

Adapting these formula, we can attain not only a smooth profile with no appearance of
the saw-tooth but also the stabilization of numerical computation.
2. Relocation Technique
Next, we adopt a relocation technique to avoid the failure of computation by the concentration of fluid particles on free surface. If we find a very smaller segment than an
average-sized segment of the free surface part, then the fluid particles on the edge of
segment are relocated to the suitable new positions.

Numerical Examples
In order to examine applicability of our method proposed in this paper, we show the
obtained numerical results. We simulate the motion of a solitary wave in a water tank of
uniform depth shown in Fig.2. The tank has a width L and is filled with water to a height

h in its stationary state. A piston plate as a wave generator is located at the left-hand
end of the tank.

fF

fw.
0

n
fw.

11
x

Fig.2.

The water tank and wave generater

320

The boudary condition of boundary

r wI

is given as follows

r wI

on

(18)

And, we assume that its forward movement follows the function Xp(t) given by
for

O~t~T

(19)

in which Xo is equal to the semistroke of the piston plate, wand tc are parameters
characterizing the movement of the piston and the time T denotes the one at which the
piston plate stopped
Numerical calculations are carried out for the case given by the following parameters of the
wave generator: L/h
and t:.t

= 0.02.

= 18.0,

= 0.5, wjhi'; = 004, tc.fiih = 6.0, T.fiih = 15.0

Xo/h

We divide the fluid region into 70 constant elements in which the free

surface part is composed by 30 elements.


In Fig.3, our results of the non-dimensional maximum height, R/ h of run-up on the
vertical wall are compared with the analytical solution derived by Byatt-Smith [9] and
the experimental data obtained by Camfield and Street [10]. Figures.4-11 show the illustrations of three phases which are the generation, the propagation and run-up, and
the reflection. We can find the velocity vector fields, the pressure distributions, and the
potential distributions in each figure.
1.0

+
O.

:I:

0.6

.....
cc

0.4

0.2
0.0
0.0

Fig.3.

Od

0.2

A/H

0.3

0.4

0.5

Maximum height R/h of run-up with the wave height A/h


o:present solution
+:experimental data
--:analytical solution

321

Velocity vector

Velocity vector

IIR_~I E3!
Pressure distribution

Pressure distribution

Potentia.! distribution

Potentia.! distribution

FigA

Generation phase (t

= 6.00)

Velocity vector

Fig.5

Generation phase (t

= 9.00)

Velocity vector

_s=====&I1, EI=II
i i i

Pressure distribution

Pressure dist ribu tion

Potentia.! distribution

Potential distribution

Fig.6

Propagation phase (t

= 15.00)

Fig.7

Run-up phase (t = 18.00)

322
..... - ,- /

:...
;:
...
! : ":

:::~~~
___ J,'

--_ .... ,

Velocity vector

Velocity vedor

~1I1
Pressure distribution

l...--.-.--L-LLrm~
1 \\\ 1 I
Potential distribution

Fig.8

Potential distribution

Run-up phase (t

= 21.00)

Fig.9

_".1

Velocity vector

1/ IIillTI

Reflection phase (t

= 24.00)

Velocity 'vector

i
l

Pressure distribution

i
.

Pressure distribution

{ (1111\\1

1111 111\\ J

Potential dislIi bu tion

Potential distribution

Fig.l0

II

Pressure distribution

Reflection phase (t

= 27.00)

Fig.ll

Propagation phase (t = 30.00)

323

Concluding Remarks
The boundary element method with use of our new boundary element scheme has been
applied to numerical simulation of two-dimensional non-linear water wave motions. The
method is verified its effectiveness with the motion of a solitary wave in a rectangular water
tank. In this study, the illustrations of three phases of the wave motion can be obtained
by using the boundary element scheme with direct differentiation method. These are
the boundary element configurations, the velocity vector fields, the pressure distributions
and velocity potential distributions at each time. Furthermore, we show the maximum
height of run-up with the various wave height. Consequently, we can simulate numerically
motion of non-linear water wave with a free surface by using the new boundary element
scheme which is based on the direct differentiation method.
References
1. Longet-Higgins,M.S and Cokelet,E.D.: The deformation of steep surface waves on
water, 1. A numerical method of computation, Proc.R.Soc.Lond.A.350,( 1976} ,pp.126

2. Dold,J.W and Peregrine,D.H.: Steep unsteady water waves - An efficient computational scheme, In:Proc.19th.Coastral.Engng.Conf,Huston, Texas, Vol.l,( 1984},
pp.955-967
3. Nakayama,T.: A comllutational method for simulatin~ transient motions of an in~ompressible inviscid fluid with a free-surface, Int.J.Num.Methods in Fluids,Vol.lO,
(1990) ,pp.683-695
4. Sugino,R and Tosaka,N.: Solution procedure for nonlinear free surface problems by
boundary element approach, Theoretical and Applied Mechanics,38(1989},University
of Tokyo Press,pp.53-59
5. TosakalN j Sugino,R. and Kawabata,H.: Boundary element Lagrangian solution
methoa or nonlinear free surface problems, In:Boundary Element Methods in
Engineering (Eds.B.S.Annigeri and K.Tseng),(1989), Springer-Verla.g,pp.131-139
6. Sugino,R and Tosaka,N.: Large amplitude sloshing analysis in a container with
multi-slopped wall by the bounaary element method, In:Advances in Boundary Elements Methods in Japan and USA (Eds.M.Tanaka,C.A.Brebbia and R.Shaw), Topics in Engineering, 7,(1990),Computational Mechanics Publications,pp.307-316
7. Sugino,R and Tosaka,N.: Boundary element analysis of unsteady nonlinear surface
waves on water, In:Boundary Elements XII,Vo1.2(1990), Computational Mechanics
Pablications, Springer-Verlag,pp.107-117
8. Savitzky,A. and Golay,M.J.E.: Smoothing and differetiation of data by simplified
least squares procedures, Analytical Chemistry,Vo1.36,No.8,(1964),pp.1627-1639
9. Byatt-Smith.J.G.B: An integral equation for unsteady surface waves and a comment
on the Boussinesq equation, J.Fluid.Mech.,49,(1971),pp.625-633
10. Camfield.F.E and R.L.Street: An investigation of the deformation and breaking of
solitary waves, Dept.Civil Engineering Technical Report,81, Stanford Univ,(1967)

A Fonnula for the Boundary Integral in the Potential Problem


of1Wo-Dimensional Anisotropic Materials and its Evaluation

Tomohiro TABATA, Toshihiko KUWAHARA, Tsuyoshi TAKEDA


Department of Electrical Engineering, Hokkaido University.
Kita-ku Kita-13 Nishi-B, Sapporo, 060, JAPAN.

Swnmary
Recently, analytical formulas (1,2) of integration for the boundary element
method (BEM) have been shown to be effective for the analysis of electric
fields. However, these formulas can be applied only to isotoropic materials,
not to anisotoropic materials. In practice, as with a liquid crystal, it is
necessary to analyze anisotropic materials. But until now, to analyze
anisotropic materials, we have had to use a nwnerical integration.
In this paper, we describe the analytical formulas of integration for a twodimensional pseudo-Laplace equation which governs the phenomena in
anisotropic materials. These fomulas are formed recursively from the zero
order to a higher-order interpolation function on boundary elements. By
using higher-order elements, the nwnber of elements of the nwnerical model
is decreased without lowering the accuracy and the computing time is
shortened.
We will show the efficiency of these formulas in some sample problems with
anisotropic materials.
Basic Relationships
The electrostatic phenomena in anisotropic materials are governed by the
pseudo-Laplace equation as,
(1)

V'[e]Vu=O,

0]
[e]=eo [ e"
0 e.

(2)

where V, t." u, [e], eo, e and e. denote Hamilton I s operator, the


scalar product of vectors, the electric potential, the tensor of the
dielectric constant, which is supposed to be diagonal, as Eq. (2), the
dielectric constant in vacuwn and the specific inductive capacity,
respectively.
Applying Green's theorem, Eq. (1) is transformed into,
%

325

(3)

where j' is the boundary of the domain, and n is the unit vector of the
outward normal direction. u' is designated as the fundamental solution (3)
in two-dimensional anisotropic materials. It is represented as,

r'=j-~+
ex
e.

(4)

In the first term of the right hand side of Eq. (3), we introduced a new
variable d, which is the Qutward normal component of dielectric flux
density, defined by the following equation.
(5)

d=[e]Vuon

Integral Formula
The interpolation polynomial, u, and d at any point on the element can be
expressed by their nodal values. Using the dimensionless coordinate ~, the
coefficient matrix B, the nodal value u" and de, the solution u and d
are written as follows.

(E)=[

d(E)=[

,"][B][uo"

u,,"y ,

f2

f2 ~n][BJ[do" dI 2 d .. e y

1 2

(6)
(7)

where n and superscript T denote the order' of the approximation polynomial


and the transposition of the matrix to which is assigned, respectively.
Hence, Eq. (3) can be written for the number of elements IV e as,

He

J =\

[H

(0)

II )

H (.. ) ][ U 0 e U

Ie

u ,,/ ]T ,
(8)

where m denotes the number of nodes and G (p) and H (' I ( p = 0- m) are
called influence coefficients, which represent integration between the
source point i and the node p of the element.

326

We then have
[C(o. C(I) ...... C( .. )]=[g

[H

where g

'O )

g Im)][B] ,

(9)

H(J)HI .. )]=[h 'O ) h(1)h'm)][B] ,

(10)

and h

IP)

(P)

(0)

(l)

mean integrations defined by the following equation.


(11 )

( 12)

The above integration can be calculated recursively using the geometrical


data of the triangle formed by the element r j and the source point i, as
shown in Fig. (1).

.f

-+----------------------------------------------. ~
Fig. (1) Definition of geometrical data
In order to simplify the calculation, we normalize r J and let the
normalized coordinate system (the x,y axis in Fig. (I coincide with the
principal axes of anisotropy.

327

Here we apply the coordinate transformation to these parameters.


The definition of this transformation is shown in Fig. (2). And the
relationships between the original system (a) and the compressed system (b)
are difined by Eq. (13) and Eq. (14) (4).
y

X
X

(b)

(a)

Fig. (2) Transformation of the coordidate system from the original (a)
into the compressed system (b)

, I
= r

2
2
- cos
- - -e- +sin
- -e-

ex

(13)

ey

(14)
By this transformation, integral formulas of anisotropic materials can be
expressed simply and similarly to formulas of isotropic one.
h

(Q

and h (') can be wri tten in the form ,


h

(O)= __

(t ) __ _

~
27r
'

.fe:e: f
27rA

(15)

J.

{
I l-ln --1 n ---

f.'

fl"

e .' cot a ' }

(16)

328

where A is a constant which is determined by the geometrical data. The


variables with .'. denote the ones transformed from the original coordinate
system into the compressed coordinate system.
For any order of h

(p

(P)

>1),

we can use the following recurSl ve formula.

( 17)
where Band C are constants in the same manner as A.
Furthermore, applying the integration by parts to Eq. (11),
expressed in any order by h (P I as,

(PI

can be

..... -

2,r(p+1)
-cota' h

(P+I)

( 18)

Sample Problem
To verify the efficiency of these formulas, we analyze the electric
potential field where exact solutions are known. The model is shown in Fig.
(3). Fig. (4) shows the discretization of the boundary. We gave the exact
solutions on all boundaries.

Prob1em domain

..B~;i.;
. '~t:.:;c",,'"'I:"'
I
I

! / s a m p 1e
!

,;
A . _ ..

",

.' t
ry '
'

'"

point

_ t

...

: " ':

. . . . . . .

Number of nodes: 48
AB, CD: Dirichlet B. C.
BC, DA: Neumann B. C.
Numerical integral:
Gauss-Legendre Method
(4 points)
Exact solution:
r'

Fig. (3) Problem model

u(r',O')

(CXcos Ie 0' +CYsinle 0')


Ie =0,770805122685
CX=-0.9687948007
CY=-1.8216068234

329

1st Order
Elements: 48
Nodes: 48

3rd Order
Elements: 16
Nodes: 48

2nd Order
-Elements: 24
Nodes: 48

Fig. (4) Discretization of the Boundary

101~------~----~------~-------r------~

Relative Error at
Sample Point

-0-:
~:

10 -

NUMERICAL
ANALYTIC

Average Error on
the Boundary
-~--:

NUMERICAL

--+-- : ANALYTIC
10 - 8~------~----~------~------~------~

ORDER

Fig. (5) Comparison of error

4th Order
Elements: 12
Nodes: 48

330

30
25

-0-: NUMERICAL
ANALYTIC

-+-:

'":3

.... 20

gJ 15

.Q

10
5

ORDER
Fig. (6) Comparison of CPU time

Comparing the analytical me t hod with the numerical one, we can confirm that
higher the order of the approximation polynomial is , the more accurate the
solution is, in a short calculating time .
The equipotential l ines and the equierror lines ar e shown in Figs.

(a)

(7)~(9 ).

(b)

Fig. (7) Equipotential lines for 3rd order elements using analytical
integr ation (a) and numerical integration (b)

331

______~~____~__~D
(b)

(a)

Fig. (8) Equierror lines (%) for 1st order elements using analytical
integration (a) and numerical integration (b)

c
10-'

A~

____________~________~

(a)

(b)

Fig. (9) Equierror lines (%) for 3rd order elements using analytical
integration (a) and numerical integration (b)

332

Conclusions
This paper presents analytical formulas of integration for two-dimensional
anisotropic materials and efficiency. Since these formulas include the
dielectric constant, the treatment of the equilibrium condition on the
boundary will be easier for analyzing compound dielectric problems. Next we
will try to simulate compound dielectric materials and to derive integral
formulas for three-dimensional anisotropic materials.
Reference
(1] T. Kuwahara and T. Takeda: "On a Formula of Boundary Integral for the
Higher order Element in Boundary Element Hethod and a Consideration of
Error of Solution", Electrical Engineering in Japan, Vol. 105A, No.1,
pp1-8(1985)
(2] T. Kuwahara and T. Takeda: "A Formula of Boundary Integral For Potential
Problem and Its Consideration" , Proceedings of the 1st Japan-China
Symposium on Boundary Element Hethod.
(3] C. A. Brebbia &S. Walker: "Boundary Element Techniques in Enginnering",
Butterworth &Co. (Publishers) Ltd. (1980)
[4] T. Hiyata, T. Takeda and T. Kuwahara: "Electric Fields on Edges of
Anisotropic Dielectric Materials", Electronics Letters, p. 1185-1186,
vol. 25, No. 17(Aug.1989)

BEM Analysis Applied to the Electric Field near the Edge of


Compound Dielectric Materials

Tsuyoshi TAKEDA and Toshihiko KUWAHARA


Department of Electrical Engineering,
Facul ty of Engineering, Hokkaido University,
Kitaku Kita 13 Nishi 8, Sapporo, 060, JAPAN

ABSTRACT
Electronic devices such as capacitors, quartz oscill ators and sensors
are becoming smaller in recent electronic apparatuses. Controlling the
electric characteristics of a device by its geometry is based on an
accurate calculation of the electric-field where the device has many
edges and material s of different natures in the body. This paper
presents an example calculation of the electric field near the edges of
dielectric materials. The field has singUlarity on the edge. In order to
reduce the error of the analysis as much as possible, numerical
techniques such as the multiplicity of flux variable[l], higher order
elements and a performance index of potential error[2] are applied to
this boundary element calculation.

PROBLEM DESCRIPTION

AND THE MODEL FOR COMPUTATION

Figure 1 shows a schematic diagram of the problem under consideration.


Three

dielectric

material s' have

materials
different

come into

specific

contact

diel ectric

on an edge where the

constants:

\ =8,

~2

=1 and

1:3=4. The electric phenomena in the problem region are supposed to be

governed by the Laplace equation with respect to potential.


According to a
the

two-dimensional study on the electric field[3], either

singular(infinite)

potential

or

zero

field

near the edge giving a

diel ectrics, is shown in Fig. 2(a).

can

appear

on

the

edge.

The

singular field, in this geometry of

334

Field evaluating side

2 elements

18 elements
3rd order

(18 elements, 1st order) 3rd order

,.......-

Dirichlet

~Neumann

S.C.
B.C.

Fig.} Model for Computation

(a)

(b)

Fig.2 Exact Potential


(a) having singular fl ux, (b) having non-singlar fl ux
One of the potential patterns which gives a non-singular electric-field
is also

shown in Fig. 2(b). The formula of the potential distribution,

u, in the region of
as

follows:

=8, is represented by the polar coordinate system

335
u(r,e)

= r A.

{A sinA.

+ B COSA.

(1)

where the origin of the polar coordinate system is located on the edge,
and the parameters are approximated as follows;
A. =0.783,

A=-2.701 and

B=-5.132

for the singular field, while

A. =1.194,

A=-0.299 and

B=0.0659

for t he non-singular fiel d.

The boundary el ement anal ysis here concerns these potential s as the
exact

solution of

the

problem and the

features

of

the

method

are

investigated. The potential problem having the singular field represents


one where the flux theoretically becomes infinite on the edge. In order
to

obtain

an

accurate

solution

of

this

problem,

great

deal

of

computing effort, such as calculating the larger unknowns proportional


to the finer mesh near the singular point, is necessary.
For comparison to the singular field problem, a calculation that was not
singular

but had

the

same geometrical

configuration was carried out

using the same mesh. The results of computation are evaluated on the
oblique side of the body having the material constant of

1>1

=8, where the

side is denoted with a boundary element mesh as shown in Fig. 1.

ERROR INDEX (2]


The

error

index

7J

(x),

for

the

boundary

potential,

is

defined

follows:
7J

where x
UI(X)

(2)

(x)= udx)-u,,.(x)
is the error-evaluating point on the boundary,

is the interpolated potential on th boundary

element

incl uding the point, and


u'" (x) is the limit potential of some inner point of the
region when the point has moved to the boundary.

as

336
DISCUSSION 1: Singular field
Figure 3, 4 and 5 show the exact potential, the error and the error
index, respectively, where they are J:lormalized by the average absolute
value

of

the

nodal

solution.

Potential

Fig.3 Exact Potential


The error index is defined as the difference between the interpolated
potential at some point on the boundary element and the limit potential
on the point from the inner side of the region. As seen in the diagram,
the error and the error index are well correlated with each other in
t his singular fiel d problem.

1.0 (%)

(18 elmt., 1 at)

1.0

(2 elmt., Jrd)

Fig,4 Error of Potential

(18 elmt., Jrd)

337
1.0(0/0)

1.0

1.0

-1.0
(18 elmt., 1st)

(2 elmt., 3rd)

-1.0
(18 elmt., 3 rd )

Fig.5 Error Index of Potential


Figures 6, 7 and 8 show the exact boundary f1 ux, the numerical sol ution
and its relative error to the exact flux, respectively. The accuracy of
the solution is improved by the mesh refinement, however, an irregular
'waving' variation of the sol uti on is compressed toward the edge where
the field

becomes singular. The 'waving' length cannot depend on the

order of the interpolation polynomial but rather on the intervals of the


nodal points as seen in Fig. 7. Since the error index is also strongly
related to the error of flux in these diagrams, it is concluded that the
index can be a good guide to the nature of the numerical solution in
practice.

Fig.6 Exact Fl ux

338

(a)

( b)

(c)

Fig.7 Numerical Solution of Flux

In these fi,ur e s ,
(a) 18 elements ,
1st order .
(b) 2 element.s,

3rd order.

(e) Ie element.s,

3rd order.

Fig.8 Error of Flux

DISCUSSION 2: Non-singular fiel d


By changing the boundary conditions, we can obtain a new sol ution having

a non-singular field on the edge. Figures 9 and 10 show the error and
the error index of the potential respectively, where the both diagrams
have a good correlation as in the previous calculation. The fl ux error
shown in Fig. 12 also has the same features as the singular case.

339

5.0

5.0(%)

5.0

5.0
(18 elmt., 1 st)

(2 elmt., 3rd)

(18 elmt"

3rd)

Fig.9 Error of Potential

5.0

5.0 (0J0)

5.0

5.0
(18 elmt., 1st)

(2 elmt., 3rd)

(18 elmt., 3rd)

Fig.l0 Error Index of Potential


The reason why both cases have the same error features can be understood
by comparing the exact potential

and flux

distributions, as shown in

Fig. 13. Whether the field is singular on the edge or not, the potential
and flux distributions have the similar profiles over the boundary where
the

error

is

evaluated.

The

flux

distributions

represent

sharper

340
6.0

(b)

(a)

(c)

Exact solution

Fig.ll Exact and Numerical Sol ution of Flux

50.0 (%)

(a)

(b)

In these fi,ures,
(a) 18 elements,
let order .
(bl 2 elements,
3rd order .
(e) 18 elements,
3rd order.

-50.0

Fig.12 Error of Flux


descent or ascent toward the edge than that of potential. Thus, a mesh
having an irregular pattern rather than a fine but regular mesh will be
necessary to obtain a more accurate sol uti on. FOl' these reasons, it can
be

stressed

that

the

existence

of

the

fl ux

singularity

is

insuperable difficulty in sol ving problems having a sharp edge.

not

an

341

Singular

Non-singular

Potential

Flux

Fig.13 Summary of Exact Solutions

CONCLUSION

This paper described the application of BEM to a singular electric field


near an edge where many dielectric materials came into contact. It was
shown that

the proposed error index is useful for investigating the

improvement of electric field analysis.

342
REFERENCES
(I} T. Takeda, T. Kuwahara, and T. Miyata: A Treatment for the Multiplicity of Edge in 3D BEM Analysis using Higher Order Conforming
Elements, Boundary Elements XII, Vo1.2, pp.307-316, Computational
Mechanics Publications, (1990).
[2] T. Takeda, and T. Kuwahara: The Error Estimation of the Boundary
Element Method of the Multi-regional and Non-linear
Problems, Boundary Element Method

Potential

in Applied Mechanics,

pp.37-46, Pergamon Press, (1988).


[3] T. Takeda: Electric Field Pattern on the Common Vertex of Multiple
Diel ectric Media, Trans.I.E.E. of Japan, Vol.107E, No.7/8,
pp.121-128, (1987).

Analysis of Certain Inverse Problems in Transient Heat


Conduction by the Boundary Element Method
M. Tanaka*, M. Nakamura* and H.lshikawa**

Faculty of Engineering, Shinshu University


500 Wakasato, Nagano 380, Japan
** Graduate School of Shinshu University
500 Wakasato, Nagano 380, Japan

Summary
Some of the inverse problems in transient heat conduction are investigated by the boundary
element method. It is assumed that the time-dependent values of temperature and heat flux on
some part of the boundary are not known and hence should be identified by using the given
measured data. For the inverse problem, this paper presents an analysis procedure using the
boundary element method in space and time for transient heat conduction problems together with
the standard optimization technique. Numerical experiment is carried out for several sample
problems, whereby the usefulness of the proposed procedure is demonstrated.

Formulation of the inverse problems


A large number of investigations have been done on the inverse heat conduction problems by
many other researchers[I-S].ln this study, we consider the inverse problem in which unknown
boundary values in transient heat conduction problems should be estimated by inverse analysis.
It is assumed that the measured data of time-dependent variables at some points on the boundary
are available for inverse analysis. The whole geometry and the thermophysical properties of the
thermal conductive field are given in advance. Furthermore, we assume that the unknown
boundary values are uniform on the boundary and vary only in time.
This paper is concerned with the inverse analysis method in which the unknown boundary
values are estimated directly or approximated by a function with parameters, which are to be
determined by the standard optimization procedure. The former method is called here the direct
method. In the latter method, called the method of approximate functions, we use the Fourier
series expansion or the spline functions[6,7]. The problem under consideration can result in an
optimum problem in which the parameters of the function describing the unknown boundary
values in time should be determined by minimizing an objective function.

344
The objective function can be chosen as the square sum of residuals between the measured
temperatures at selected points on the boundary and the corresponding values computed by the
boundary element method assuming the unknown parameters. The objective function can be
written as follows:
(I)

where

up,q

denotes the measured temperature of a spatial point p at time

t=tq

and

up,q

the

computed ones at the same point in space and time which can be obtained from boundary
element analysis of the transient heat conduction problems[8). P is the number of measuring
points on the boundary, while Q is the number of measuring points in the analyzing time. The
conjugate gradient method[9] is employed in this study to determine an optimal set of
parameters.
In the method of approximate functions, the parameters to be determined are as follows. If the
Fourier series expansion is applied to the time interval [a. b], we can approximate u(t) as

(2)
where ak and bk are the Fourier coefficients which are in this case the parameters to be
determined by the optimization procedure. On the other hand, if we approximate u(t) by the mth
order B-spline functions, we have
N

u(t) =

L CjMmj(t)

(3)

i=1

where Mmi(t) denotes the mth order B-spline functions. In this case, the number of optimum
parameters is N coefficients of C j
In the direct method, however, the values of u(t) at some points in the time axis can be directly
the parameters of the optimum problem.

Estimation procedure
The algorithm of parameter estimation using the boundary element analysis together with the
optimization procedure consists of the following steps:

345

STEP I. Input the data concerning the shape, dimensions, thermophysical properties of the
thermal conductive field and the measured data of temperature and/or heat flux.

STEP 2. Assume the initial parameters of the unknown boundary values.


STEP 3. Approximate the unknown boundary values by using the direct method or the

method of approximate functions.


STEP 4. Carry out boundary element analysis of the transient heat conduction problem and

calculate the objective function. Modify the parameters by using the optimization
technique.
STEP 5. Check convergence. If the solution converges, then stop iterations. Otherwise, go

to STEP 3.
In this study, as the convergence criterion of iterative computations, we use the following
inequality:

z" <
where

1]

11

(4)

is a given tolerance limit, and Zk is the non-dimensional expression of the objective

INPUT DATA FOR BEM ANALYSIS


AND MEASURED VALUES

MODIFY PARAMETERS BY USING


OPTIMIZATION PROCEDURE

NO

Fig. 1 Row chart of inverse analysis

346
function at the kth iteration, which can be expressed as
Zk

= llog (w k I w O )

(5)

Numerical results
Numerical experiment is now carried out for the estimation of unknown boundary values to
demonstrate the usefulness of the proposed analysis method. In this numerical experiment we
assume that the rectangular model ABCD with the size of 20mm X IOmm as shown in Fig. 2.
The longer sides AD and BC are assumed to be insulated, while on the side CD heat flux q(t) is
known. The time-dependent values of temperature and heat flux at the side AB are assumed to

be uniform over the side AB, but not known. It is assumed that the values of temperature at a
measuring point E on the side CD are known at any time intervals. The thermophysical
properties are assumed as shown in Table I. The initial temperature of the rectangular model is
zero degree over the whole domain.

q(t) = 0
q(t)

q(t)

Wlknown

u(O)

u(t)

known

=0

u(t)

unknown

measured

q(t) = 0
Fig.2 Numerical experiment model
Table 1 Thermophysical properties

Thermal Conductivity

12.0

W/mt

Specific Heat

0.88

kJ/kgt

Mass Density

2.77 X 103 kg/m 3

Heat Transfer Coefficient

10.0

W/m 2oC

347

Node

I---l Element

Fig. 3 Boundary element discretization by constant elements


In this numerical experiment the values of temperature at the measuring point E are computed by

the boundary element method for transient heat conduction problems under the above-mentioned
assumptions. Figure 3 shows the boundary element discretization for this computation, where
30 constant elements with equal size are used. The same discretization is employed for the
iterative computations of optimization process. The boundary element analysis coderS] has been
developed with the basis of the boundary integral equation in space and time. In this numerical
experiment, the time axis is divided into constant time elements.
Now, we show some of the results obtained by the present inverse analysis for the following
two cases: The first one is such that the time-dependent values of temperature at side AB are
given by
U(/) = {

150( 1 - exp[ -t /6] )

(0 ~ t ~ 30)

150( l-exp[ -5 1) exp[ (30 -I )/6)

(30<tS60)

(6)

In the second case, the above values are expressed by the step-wise function
u(t) = {

150

(O~t ~30)

( 30 < t S 60)

(7)

It is noted once again that the values of temperature are assumed to be uniformly distributed at
the side AB. Figures 4 and 5 show the temperature variations at the measuring point E obtained
from the boundary element analysis when the temperatures at the side AB are given by Eqs. (6)
and (7), respectively.
Figures 6 and 7 show the estimated values of temperature when the nodal values with the total
number of 21 are directly put into the unknown parameters. In these figures the square symbol
denotes the results obtained by the inverse analysis and the solid line the exact values of
temperature given by Eqs. (6) and (7). It can be seen that the inverse analysis provides a
satisfactory estimation.

348
200.0

...-.

175.0

~ 150.0
'-'
;:::

125.0

"i
.....

860

00

0
0

0
0

100.0

75.0

50.0

25.0

0
0

0.0
0.0

10.0

20.0

30.0

40.0

50.0

60.0

Time t (s)
Fig. 4 Temperature responses at the measuring point E, obtained by BE analysis, when the
boundary conditions on the surface AB are given by Eq. (6)

200.0

...-.

[;..)

'-'

;:::

175.0 I150.0

"

C!)

C!)

C'I

0.

125.0

.....

....=
~

100.0

.....
0

C.

60

E-

75.0

50.0
0

25.0 I-

0.0
0.0

10.0

20.0

30.0

Time

o
40.0

<P '"

50.0

60.0

(s)

Fig. 5 Temperature responses at the measuring point E, obtained by BE analysis, when the
boundary conditions on the surface AB are given by Eq. (7)

349
200.0

--Exact

175.0
,-..,

150.0

Estimated

'-'
~

...
...

Q)

....CIS='

Q)

125.0
100.0
75.0
50.0
25.0
0.0
-25.0

0.0

10.0

20.0

40.0

30.0

Time

t (s)

50.0

60.0

Fig.6 Estimation results by using the direct method in case of Eq.(6) with 21 parameters
200.0

--Exact

175.0

""'
~

150.0

Estimated

'-'

125.0 I-

100.0

8-

75.0

='
~

E
~

50.0
25.0
0.0
-25.0
0.0

10.0

20.0

30.0

Time

40.0

50.0

60.0

t (s)

Fig.7 Estimation results by using the direct method in case of Eq.(7) with 21 parameters

350
Figures 8 and 9 show the estimated results of temperature obtained by the method of
approximate functions using the Fourier series expansion. The dotted line denotes the estimation
results obtained for the case N=16. Figures 10 and II show the estimated values of temperature
obtained by the method of approximate functions using the third-order B-spline functions. In
both the cases the total number of parameters is assumed to be 16.

If the temperature variation is smooth, accurate results can be obtained using either the Fourier
series expansion or the B-spline functions. For the step-wise change of temperature, oscillating
results are obtained by using the both of the approximate functions. However, it can be
concluded that the direct method provides accurate results for all the cases considered.

Conclusion
In this paper a computational procedure has been proposed for the analysis of the inverse
problems in transient heat conduction. From the numerical experiments for a few example
problems it is revealed that a successful estimation of the unknown transient temperature can be
made by using the method of approximate functions for the case where the change in the
unknown values is smooth in time. The direct method can be successfully applied even to the
case with a step-wise change of the boundary temperature. We carried out numerical
experiments for the case where the measured values include no errors. In most of inverse
analyses, estimation results are much influenced by errors included in the measured data, and
sometimes analysis is impossible. In our future work we will investigate the case where the
given temperatures include some measurement errors.

References

1. Busby, H.R.; Trujillo, D.M.: Numerical solution to a two dimensional inverse heat
conduction problem, Int. J. Num. Meth. Eng., 21 (1985) 349-359.
2. Trandy, D.F.; Trujillo, D.M.; Busby, H.R.: Solution of inverse heat conduction
problems using an eigenvalue reduction technique, Numerical Heal Transfer, 10 (1986)
597-617.
3. Zabaras, N.; Liu, J.C.: An analysis of two-dimensional linear inverse heat transfer
problems using an integral method, Numerical Heat Transfer, 13 (1988) 527-533.
4. Flach, G.P.; Ozisik, M.N.: Inverse heat conduction problem of simultaneously
estimating spatially varying thermal conductivity and heat capacity per unit volume,
Numerical Heat Transfer, ParIA, 16 (1989) 249-266.
5. Reinhardt, H.J.: A numerical method for the solution of two-dimensional inverse heat
conduction problems, Int. J. Num. Meth. Eng., 32 (1991) 363-383.
6. Rivlin, J.T.: An Introduction to the Approximation of Functions, Dover Publications,
(1969).
7. Rice, J.R.: The Approximation of Functions, Addison-Wesley Publishing Co., (1964).
8. Tanaka, M.; Togoh, H.; Kikuta, M.: Fracture mechanics application in thermoelastic
states, Topics in Boundary Element Research, Ed. by Brebbia, C.A., Springer-Verlag, 1
(1984) 59-77.
9. Fox, R.L.: Optimization Methodsfor Engineering Design, Addison-Wesley Publishing
Co., (1971) 38-116.

351
200.0

--Exact
........ Estimated

175.0
,-....

150.0

;::::

125.0

'-"

100.0

f!

75.0

;:3

<t.l

0..

S
~

50.0
25.0
0.0
-25.0
10.0

0.0

20.0

30.0

40.0

50.0

60.0

t (s)

Time

Fig.8 Estimation results by using Fourier series expansion in case of Eq.(6) with 16

parameters
200.0
175.0
,-....

150.0

;::::

125.0

100.0

~\

'-'

e
;:3

<t.l

r'

\.;1

1'-,

\
\_~

'.'

"\

'

.
0

!.

t-

75.0

0..

S
~

--Exact
........ Estimated

0
0

SO.O

1
i

25.0

.,

0.0

/"\

\ ~I'

"\

\/

'.,

/"\

-25.0
0.0

10.0

20.0

30.0

Time

40.0

50.0

\ I

60.0

t (s)

Fig. 9 Estimation results by using Fourier series expansion in case of Eq.(7) with 16

parameters

352
200.0

--Exact
........ Estimated

175.0

--..

P' - '

150.0

:::s

125.0

100.0

::s

e
~

75.0

0..

r--

50.0
25.0
0.0
-25.0
0.0

10.0

20.0

30.0

40.0

50.0

60.0

Time t (s)
Fig. 10 Estimation results by using spline functions in case of Eq.(6) with 16 parameters

200.0

--Exact
........ Estimated

175.0

--..

'-'

I',

150.0

:::s

125.0

100.0

::s
:i
....
~

0..

~-,

,.'

.
\.
.
\

.
..

75.0

50.0
25.0

., ," ,
"

0.0

\v i

-25.0
0.0

10.0

20.0

30.0

Time

--'

40.0

50.0

60.0

(s)

Fig. 11 Estimation results by using spline functions in case of Eq.(7) with 16 parameters

Grid Generation of Arbitrary 3D Surface


T. Taniguchi* and Y. Fukuoka**
* Eng.Science Dept.,

Okayama Univ., Okayama 700, Japan

** Dept. of Civil Eng., Okayama Univ., Okayama 700, Japan

Summary
The modelling of arbitrary 3 dimensional body is a tedious work
for

the

user of BEM.

method

for

prepare

triangular

based

on

the

the

This

surface
and

paper includes a

of

arbi trary

quadrilateral

Delaunay

3D

grid

domain,

elements.

triangulation,

generation

and,

geometry of elements is optimum for a set of

which

The

can

method

therefore,

is
the

given nodes.

1. INTRODUCTION
How to define the configuration of 3D domain is presently
one

of

important

analysis.

problems

in

the

field

of

the

numerical

In case of the surface with simple geometry the con-

figuration can be easily defined, but it becomes very difficult


to define

the boundary of 3D domain with complex geometry.

We

assume that nodes are set on the boundary of 3D domain so that


they can sufficiently express its characteristic.

That is,

in

the portion with large curvature nodes are distributed densely.


Then,

the grid generation should work to generate grids

using

the coordinates of these nodes.


Another
erated

requirement

elements

must

for

the

be

appropriate

grid

generator
for

is

the

that

gen-

numerical

analysis, since distorted elements become one of the reasons of


the numerical and discritization errors.
At
surface.

present we find
Considering

the authors introduce

the

several tools to prepare grids for


input

data

the Delaunay

for

the

grid

triangulation as the

of the grid generation [1,2,3J,

and

grid

element method

generator

for

3D boundary

they

3D

generation,
basis

propose an effective
in

this

paper.

354

The

method

can

not

only

prepare

triangular

or

quadrilateral

elements on the boundary of arbitrary 3D domain but also define


the surface, itself, using nodes on the surface.

2. GENERATION OF BOUNDARY AND ITS TRIANGULATION


Assume a

3D domain covered with a smooth surface on which

sufficient number of nodes are set so that they can express its
characteristic. Firstly we propose a method to generate triangular elements which cover whole surface.
The Delaunay triangulation is a method to subdivide any 3D
domain occupied by nodes into a set of tetrahedra. The circumsphere of a generated tetrahedron never include any other nodes
except four nodes which form the tetrahedron. This characteristic of generated tetrahedra is satisfactory as the geometry of
the boundary elements, and our aim is to introduce the Delaunay
triangulation as the grid generation for 3D BEM.
actual use of the Delaunay
method,

triangulation as

But,

grid

for the

generation

it must be modified so that it can be applied for any

non-convex

domain.

It

should

be

noted

that

the

original

Delaunay triangulation is valid only for convex 3D domain.


The Delaunay triangulation is already modified in [4], and
the

modified

Delaunay

triangulation

domain into tetrahedra.


the

surface

surface

The method

triangulation

triangulation is

and

to

can

subdivide

is divided

the

volume

generate

into

non-convex
two parts;

triangulation.

The

triangles using nodes on

the surface so that they can cover whole surface, and we introduce it as the grid generator for BEM.
Let N be a set of distinct nodes on a surface. Our aim is
to generate triangles using these nodes and
face

by

the generated triangles.

covered by

triangles.

Then,

we

to cover

the

sur-

Assume that a surface can be


find

following

two

facts:

(1)

Three nodes forming a triangle locate in neighborhood each


other.

(2) The intersection angle of two adjacent triangles is

small for any smooth surface. These two facts suggest that the
distance and the smoothness are important factors to construct
triangles.
Assume a triangle on the surface is given. Then,
structing new triangle adjacent to

this

triangle,

for con-

we use

fol-

355
lowing criterion to determine new node;

+ b*log

f = max [ a*log 8
In above
Fig.l,
the

expression,

and

and two parameters,

distance,

parameter

respectively.

"a"

should

be

are

a and b,

set

If

the

larger

two angles as

shown

in

are for the smoothness and


surface

is

than lib".

smooth,

the

Fig.2 shows

the

result of the test of the criterion. Two parameters are equated


for

case

(a)

generation.

and
On

the result shows

the other hand,

the failure

of

the

surface

the strict surface can be gen-

erated for a=2 and b=1 as shown in the case (b).


The grid generation of the surface of arbitrary 3D domain
is briefly expressed as followings:
Step 1. Input of all coordinates of nodes, and one triangle.
Step 2. Set two parameters, a and b, appropriately.
Step 3. Select an edge of a triangle.
Step 4. Select new node using above criterion.
Step 5. Form new triangle using the edge and the selected
node.
Step 6. Select an edge locating the boundary of generated
triangles.
Step 7. Go to Step 4.
Steps 4 through 7 are repeated untill all nodes are used.
Above

procedure

generates

triangles

one

by

one

from

the

boundary of the polygon formed by triangles already generated.


Two examples are illustrated in Fig.3, in which smooth and
rough surfaces are used for

the examination.

The

figures

that the proposed method works well as the grid generator.

(a)

Fig.l Determination
of Triangle

(b)

Fig.2 Test of Surface


Generation

show

356

Fig.3 Examples of Surface Generation

3. GENERATION OF QUADRILATERAL ELEMENTS


Assume

that

the

surface

of

any

3D

domain

is

subdivided

into a set of triangles using the method in Section 2.

Now,

we

propose the method to generate quadrilateral elements from the


triangles.
[Method

I]

A quadrilateral

element

can

be constructed

by

removing a common edge of two triangles adjacent each other. As


the condition to form a quadrilateral element there may exist a
number of criteria, and we use following condition:
"Join two triangles to form a quadrilateral,
if their common edge is the longest for each of them."
A

number

of

triangles

can

be

transformed

drilateral elements using above method.


there
and

exist

(2)

odd

there

number

are

of

triangles

triangles

of triangles. Let consider

II]

Any

after

which cann' t

quadrilaterals using above method,


[Method

But,

triangle

into

qua-

in cases that (1)


the

be

triangulation.

transformed

there still remain a

into

number

the transformation of them.


can

be divided

into

three

qua-

drilaterals by adding a midnode on each edge and also a node at


the center of gravity as shown in Fig.4.

Fig.4 Treatment of Triangle

Fig.S Treatment of Quadrilateral

357
Method

II

can

modify

all

triangles

into

quadrilaterals,

length of any quadrilateral element by Method I I

the edge

but
is

just half of the edge length of the one obtained by Method

I.

Then,

into

four

we subdivide all quadrilaterals obtained by Method I


smaller

ones

by

adding a

midnode

on

each

and

edge

also one node at the center of gravity as shown in Fig.S.


The

modification of

triangles

to

quadrilaterals

requires

the setting of additional nodes on the surface of 3D domain. If


the

surface

to

be

subdivided

locates

on

plane,

additional

nodes can be strictly and easily placed using the method mentioned above.
ation

of

But,

arbitrary

present discussion
3D

surface,

and

is

for

the

additional

grid

nodes

genermust

be

placed on the surface. Here it should be noticed that the input


data is not the surface,

itself,

but the coordinates of

nodes

on the surface which is not defined. That is, the surface where
nodes locate must be

firstly

determined,

and additional nodes

can be placed on the surface.


We assume that nodes in neibourhood locate on a 3D surface
, F, defined by a complete quadric polynomial. Then, we firstly
have

to determine the polynomial,

and secondly

find

the

loca-

node

which

tion of the additional nodes on the surface.


Let

determine

the

location

of

the

additional

should be set on center of gravity of a

quadrilateral element

that is obtained by Method I. Firstly, we calculate the center


of gravity of the quadrilateral using
nodes forming

the element,

the coordinates of

four

and new node is temporarily set at

the position. Successively, select the nearest 9 nodes from the


new node,

and determine a

these selected nodes.

surface F using

Find a

the

line

crosses

coordinates

of

line which is normal to the tri-

angle from the center of the gravity,


where

the

the surface F.

and calculate
This crossing

the

point

point is

new position of the additional node.


Same procedure is used for the determination of other additional nodes which are
triangles

to

required for

quadrilaterals.

the

Summarizing

transformation
above

from

consideration,

quadrilateral elements are determined as following:


Step 1. Join two triangles adjacent each other using above
criterion.

358

Step 2. Subdivide the quadrilaterals into four smaller ones


by adding nodes.
Step 3. Find all residual triangles, and subdivide them into
quadrilaterals by adding nodes.
Several examples of the grid generation are shown in Fig.6.

Fig.6 Examples of Grid Generation

4. NUMERICAL EXPERIMENTS
The aims of this section are to examine the grid generator
anJ to survey the error due to the boundary element models.
Assume a spherical cavity in an infinite domain subjecting
a

shock wave as

subdivided

into

shown in Fig. 7.
quadrilateral

elements

Constant elements are used for


time domain boundary

The surface of
using

the cavity is

proposed

method.

the modelling of elements,

element method

and

is used for the analysis.

Refer to [5] on the details of the boundary element method used


in this section. Strict solution of this test problem is shown
in Fig.8,

in which the displacements at

two

point A and B of

Fig.7 are presented. The axes of ordinate and abscissa of Fig.8


show the displacement (d) and the time (t),

respectively.

Both

axes show nondimensional values.


[ON NUMERICAL MODELS]
The models in Fig's 9 and 10 show the quadrilateral elements on

the hemisphere of

the

front

and rear

sides.

Table

shows the numbers of nodes at the beginning and after the grid
generation, and also the number of generated quadrilateral elements. The numbers in the brackets show the number of nodes set
on each

latitude circle of

the sphere.

The nodes on the

latitude circle are placed with the same spacing.

same

359
Proposed method can generate quadrilateral elements for all
cases,

but some cases show that elements between two

circles are not regularily arranged.


Delaunay

latitude

The reason is due

triangulation and also the criterion for

to the

joining two

triangles into one quadrilateral.


Elements

Nodes

Model

Total

Input
15

(1+4+5+4+1 )

66

64

78

76

18

(1+5+6+5+1)

26

(1+4+5+6+5+4+1)

118

116

25

(1+4+5+5+5+4+1)

120

118

27

(1+4+5+7+5+4+1)

124

122

30

(1+4+6+8+6+4+1)

138

136

(1+5+6+7+6+5+1 )

31

142

140

38 (1+4+5+6+6+6+5+4+1)

182

180

45 (1+4+6+7+8+7+6+4+1)

198

196

Table 1

Details of Test Models

[ON NUMERICAL RESULTS]


All models in Table 1 are used for the numerical examinations, and Fig.9 shows only the results for Modell,
The time step for

the

tests is set as

0.28

for

3, and 9.

all

cases.

Fig.10 shows the influence of the time step to the same model,
and four kind of

time steps,

0.14,

0.18,

0.22,

and 0.28,

are

applied for Model 4.


Comparing Fig.

8 with Fig's 9 and

10,

models can exactly express the behaviour at


for

the behaviour of Point B.

we find
Point

that
A,

but

all
not

Comparing all figures in Fig.9,

it seems that the number of nodes or elements can give slight


significance to

the solution at Point B,

because Model 1 with

minimum number of elements can give good


problem.

But,

the

result

shown

in

solution

Fig.10 clarifies

solution is governed not only by the model but the


This

suggests

that

the

result

for

of Model

1 can

be

that

this
the

time step.
largely

in-

fluenced by the time step. Summarizing above consideration, the


time step should be determined
to be used.

for

the boundary element model

360

Fig.7 Test Problem

Fig.8 Strict Solution

Model 1
front

rear

front

rear

Model 3

--.

H"

~~
I'"

V.

'\,

'7

Model 9

front

Fig.9 Models and Results

(~t=O.28)

rear

361

front

rear

lit=O . 18

lit= O.14

lit=O.22

lit=O .28

Fig.IO Models and Results (Model 4)

362
5. CONCLUDING REMARKS
The grid generator in this paper works for generating quadrilateral

elements

smooth

also

but

of

rough

arbitrary

surfaces.

3D

surface

with

Input data for

not

only

the method is

the node coordinates and some additional data concerning nodes.


The improvement of the geometry of elements can

be done

by

changing the node coordinates, since the method is based on the


Delaunay triangulation.
The numerical experiments clarified that the determination
of the time step is one of serious problems at the application
of the boundary element method to dynamic problems. For finding
the

determination

number of

method

of

the

optimum

time

parametric analyses are required,

step,

numerous

and proposed grid

generator may be effective tool for this purpose.


Finally,

the authors would like to

thank Dr.

Soh-ichi

Hirose and Mr. Masahiro Ono of Okayama University for preparing


the numerical results in this paper. At the same time they also
thank to them for their discussion for preparing this paper.

REFERENCES
[1]

Boyer,A."Computing

Dirichlet

tesselations",

The

Computer

Journal,Vol.24, No.2, 162-166 (1981)


[2]
wi th

Watson, D. F., "Computing

n-dimensional

application to Voronoi

polytopes",

Delaunay

tesselation

The Computer Journal,

Vol.24, No.2, 167-172 (1981)


[3]

Taniguchi,T.

&

Ohta,

C. ,"Automatic

generation

of

tetrahedral finite elements for convex 3D body", Proc. of JSCE,


Vol.431 (1991) Written in Japanese
[4] Taniguchi,T.

& Ohta, C.,"Delaunay-based grid generation for

3D body with complex boundary geometry",


Numer.

Grid Gen.

3rd Int.

for Fluid Dynamics & Related

Fields,

Conf.

on

533-543

(1991 )
[5] Hirose, S.,"Boundary integral equation method for transient
analysis

of

3-D cavities

and

inclusions",

Eng.

Boundary Elements, Vol.8, No.3, pp.146-154 (1991)

Analysis

with

Integral Equation Analysis of Viscous Flow Between Two


Rotating Coaxial Disks
N.Tosaka1 , T.Honma2 and T.Manabe!
l.Department of Mathematical Engineering, College of Industrial Technology,
Nihon University, Narashino , Chiba , 275 , Japan
2.Technical Research Institute, Fujita Corp. , Kohoku-ku , Kanagawa , 223 , Japan

Summary
In this paper, application of the generalized boundary element method to the problem
of axially symmetric viscous flow between two rotating coaxial disks is presented. The
nonlinear governing differential equations are linearized in each subdomain , and the
boundary integral equations are derived by the use of fundamental solution tensor for
the linearized differential operator on each subdomain. The global matrix equation
is constructed from the local matrix obtained on each subdomain. The final system
of equations can be solved by using the simple iterative procedure. Numerical results
demonstrate the applicability and efficient of the present method.
Introduction
In 1921 , the flow in an infinite rotating disk was investigated by Von Karman

[1]

In

his study , he showed that the Navier-Stokes equations in cylindrical coordinates are
reducible to a pair of coupled nonlinear ordinary differential equations. For the axially
symmetric steady flow , he pointed out that the axial velocity is independent of radial
distance and the radial and azimuthal velocities are in proportion to the radial distance.
Batchelor[2] extended Karman's work to two infinite disks with rotating different angular
velocities , and showed that solution of this problem depends on both the Reynolds
number and rotating rate. He investigated the flow in the disks rotating in the same and
opposite directions for large Reynolds number. He described that the main body of fluid
is rotating in almost all case. Moreover, Stewartson [3] reexamined Batchelor's results
and studied the flows for large and small Reynolds

numb~rs.

However he showed that

main body of fluid is nonrotating with the disks rotating in the opposite direction. The
controversial results thus have become major subject of discussion in various publications.
Many investigators

[4-10]

have studied this problem by using various kinds of methods.

364

The generalized boundary element method[llJ (GBEM) has been proposed as the new
method in order to extend application of boundary element method to nonlinear problems.
The basic idea of this method is based on the derivation of boundary integral equation on
each subdomain by the use offundamental solution for the linearized differential operator.
The efficient of this method has been shown through several numerical examples [12J-[14J.
In this paper , we apply the above-mentioned GBEM to symmetric steady viscous flow
between two rotating coaxial disks. By modifying the nonlinear govening ordinary differential equations presented by Lance and Rogers [4J , we derive the fundamental equations
to be used in this study. As the boundary conditions on both disks, the non-slip conditions are also adopted. Moreover, we derive the boundary integral equation on each
subdomain using the fundamental solution tensor which is obtained by means of the
Hormander's method(15) . The final system of equations can be solved implicitly by using
the simple iterative procedure . In order to show the efficient of the present approach,
numerical examples are demonstrated.
Statement of problem
Let us consider viscous fluid between two infinite parallel disks. The disks are rotating
with angular velocity n(lower disk) and sn (upper disk) about the common axis perpendicular to both disks .

----+-----':::>'" sO

u
d

c ______-

... 0

Fig.1 Problem sta.tement

365
In Fig.1 , s is the rotating rate, d is the distance between two rotating disks, p is the
pressure and p is the density of fluid. In the case of this problem, the Reynolds number
Re is defined as follows :
(1)

Moreover ,we introduce the non-dimensional quantity

= Z/d.

The continuity equation

and steady Navier-Stokes equations in cylindrical polar coordinates (r,e,Z) governing an


incompressible viscous flow are given as follows:

(2)
(3)
(4)
(5)
where v is the kinematic viscosty , and u, v and ware the components of velocity in the
direction of increacing r, e and Z , respectively.

Following Lance and Rogers[~J , the velocity components u, v and ware supposed the
following forms :

= rOF(~)

v= rOGW
w

(6)

= .;;;nH(~)

And , we assum that the pressure has the following form:

(7)
Here

K,

denotes the integral constant. Substituting equation (6) into equations (2)-(5) ,

we can obtain the nonlinear ordinary differential equations with respect to F , G and H
as follows:
1- F" _ -1- HF' + F2 .;Re
ffe

1"

-R G
e

G2

= vRe
r;,-HG + 2FG

+K,

(8)
(9)

366
1
2F+ ~H =0
vRe
I

where

0' ::= d/d~ .

(10)

In addition to the above set of equations, the non-slip boundary conditions are given as
follows:
F(O) = H(O) = F(I)

= H(I) =

0 }

(11)

G(O)=I, G(I)=s
Moreover, taking the derivative with respect to
the integral constant

K ,

in equation (8) in order to eliminate

we can obtain the following fundamental differential equations

to be used in this study:


Fill

= ../ReH F" -

2ReGd

(12)

0" = 2ReGF +../Red H


H'

(13)

= -2../ReF

(14)

The generalized boundary element analysis


First of all , we divide the whole domain into several subdomains X.

= [~., ';.+1]

. The

nonlinear differential equations (12)-(14) are linearized on each subdomain. So that the
resulting equations can be written compactly as follows:
(i,j

= 1,2,3)

(15)

where

u= [F

HY

= [0

o
r

D2 -hD

(16)

2gD
L=

of

-2/

(17)

in which, the quantities j, g, hand r are defined as


h::=ffeiI j::=ReF }
g::=ReG
r::=2../Re

(18)

367

Here, the linearized quantity () is defined as follows:

(19)
Moreover, D denotes the derivative with respect to

In this stage, we consider the following integral identity of the set of equation (15) :
(20)
where

Vii

denotes the fundamental solution tensor for the adjoint operator

(,ij

of

L ij .

Integrating equation (20) by parts over the subdomain , we can obtain the following
boundary integral equations:

UIc(l'})

[F'V;;]t+l - [F"Vli.1t+l - [FV;~*]~t - [hFV;;]t+ 1+ [hFtVli.]~:+1


[2gGV,.,,]~:+1

+ [a'V;i.]t+1 -

[GV;;]t~1

+ [hGV;i.ll:+1 -

[H~i.]~:+1

(21)

Next,we derive the fundamental solution tensor by using of the Harmander's method.
Th,~

fundamental solution tensor Vik for this problem has to satisfy the following singular

differential equations :
(22)
where 6({ - l'}) denotes the Dirac delta function and

{,ij

is the adjoint operator of

Lij

as

follows:

-2gD

D2 + hD - 2/

-D

From equation (22) , the fundamental solution tensor

- Jr*dx

[l-ji.]

= ](r* o

CT)

Vii

(23)

are given explicitly as follows:

-r J J rdxdx

CT

TJ(r - CT*)dx

(24)

in which r and CT satisfy the following singular differential equations, respectively:

(D2

+ hD - 2f)CT = t5({ -l'})

(25)

368

(D2

Here, introducing parameter Q'

+ hD)7' = O(~ -1])

(26)

= It + 2f , differntial equation (25) is rewritten as follow:

(27)
where

d' is defined as
(28)

It is notice that

17*

of parameters

and h . These functions are given explicitly as follows:

Q'

and

7*

are given as the functions which depend deeply on the value

(Q'

= 0)
(29)

(h
(h

= 0)
(30)

i= 0)

Let us mention briefly the discretization of formulation. The unknown quantities on i-th
subdomain X, are shown in Fig.2.

F, , F: , F:'

Fi +1 ,F:+l F:'

G. , G:

G'+ 1 , G;+l

H,

H,+l

Fig.2

Unknown quantities on i-th subdomain Xi

Arranging the unknown quantities on the i-th subdomain Xi ,the system of equation
(21) can be rewritten as the following local matrix form:

,A,X=o

(31)

369

where oX is the unknown vector in the i-th subdomain. Setting up equation (30) for all
the subdomains , we can obtain the final system of equations as follows:

(32)

AX=B

in which A is the coefficient matrix with a bandness and B is the known vector whose
components are the prescribed boundary values of the problem.

Numerical Examples
In the following , numerical examples for several parameters sand Re are presented in
order to show the applicability and efficient of our approach . In our numerical performance, the initial values of F, G and H are assumed to be zero everywhere in the interior
domain. We show the radial velocity F and axial velocity H in Figs. 3 , 4 , 5 and 6. For
each the Reynolds numbers ( Re

= 1.0,5.0 and 6.0.)

, the results in the case s

= -0.7

and -0.8 are shown graphically in Fig. 3 and 4 , respectively. Figures. 5 and 6 illustrate

F -

eand H - ecurves for s = -0.9 and -1.0 ,respectively.

In particular, when the

disks rotate with the same speeds in opposite direction , it is appear in Fig.6 that F is
symmetric but H is anti symmetric with respect to
1. 0 , . . - - - , - - - , - - - , - - - - - - ,

......Re=6.0
-Re=5.0
_Re=1.0

0.5

0.5

o~--~--~~-~-~

-0. 10

1.0~----,-----~------r_--~

s=-0.7

...... Re=6.0
-Re=5.0
_Re=1.0

e= 0.5 .

-0.05

F(f)

0.05

0.10

O~----~---~--~---~

-0. 10

-0.05

(a)

Fig.3 Numerical resluts for s=-0.7 ) Re=1.0,5.0 and 6.0 ;a) F-

0.05

0.10

H(f)

(b)

eCurves, b) H- eCurves

370
1. 0 ..---....---".,,----r----.

1. 0 , - - - - , . . - - - . . - - - - , . . - - - - ,

,>

5=-0.8

5=-0.8
--- Re=6. 0
...... Re=5.0

~
,

..
..
/

.'.'

"'~

..... .... .
' ... ~

0.05

O. 10

.....

.:..:~

'.~

O~--L---~--~-~

-0. 10

-0.05

1.0----~------~--~
'~.~ ...

'.: ,,
: ,

Curves, b) H-

'

5=-0.9

.:,'

."

,~.r

..

,,~ '

0.5

,,,

:
:
:

..

,, '.'.
"

..
'

'

'

:
:
:

;."

0.5

(a)

,,:.,'
....

/ ' ....

/'

F(f)

.,

,~.'

.. ..::..~,

0.05

......
.,

;,

.~.

Curves

..,

--- Re=7. 0
...... Re=5.0
-Re=l.O

"".

-0.05

1. 0 .---....---.....---~----.

0.10

(b)

Fig.4 Numerical resluts for 5=-0.8, Re=1.0,5.0 and 6.0 ;a) F-

0.05

H(f)

(a)

5=-0.9

--- Re=6.0
...... Re=5.0
-- Re=l.O

, ::
,,, :::
, :
, '.
"

..

/,~ ,

0.5

F(f)

/0'

0.5

-0.05

,.

/.'

0.10

--- Re=7.0
.. ..... Re=5.0
- Re=l.O

O~--~--~----L----~

-0. 10

-0.05

H(f)

0.05

0.10

(b)

Fig.5 Numerical resluts for s=-0.9 I Re=1.0,5.0 and 7.0 ;a) F- ~ Curves, b) H- ~ Curves

371
LO.-----.----.-----~----_,
."

'.,
'"

"

s==1.0

"

-,-

,. ,~,,:.,

..
, ..
,, .-:
",~

0,5

.,,

Re=7,O
Re=5.0
Re=1.0

..-

...
"

"

"..:

..

...: / '

" , ~. '.
' . ~ ..

,,

".

-;~~,. ,

0,5
,
,,

..,
,,

..~~.. ,

.....<0; ..

"

"

.:

;.:;::..,,'

.-- Re=7, 0
.. .... Re=5.0
-- Re=l.O

..-

'

0'-----'------'"<.:......---'-----'
-0, 10
-0.05
0.05
0.10
o

... ,~.~.,

, , '.'.'.
, ,.'.

,
I '". ",

,,

,'.

"

O~----~----~---~----J

-0, 10

-0.05

F(f)

H(f)

(a)

(b)

0,05

0,10

Fig.6 Numerical resluts for s=-1.0 , Re=1.0,5.0 and 7.0 ja) F- ~ Curves, b) H- ~ Curves

Conculding remarks
We presented approximation of the flow problems of viscous fluid between two rotating
coaxial disks by using the generalized boundary element method. We derived the six kinds
of funddJIlental solution tensors of the problem depending on parameters. Numerical
examples for the disks rotating in the opposite direction by using a simple iterative solution
procedure were shown. Finally we have to examine applicability of this method to higher
Reynolds number than the number presented in this study.

References
[1] Karman .T,Von (1921) , Uber laminare und turbulente Reibung , Z . angew Math
. Meeh, 1 , pp.232-252
[2} Batchelor,G.K (1951) , Note on a class of solution of the Navier Stokes equations
representing steady rotationally symmetric flow, Q.J.Mech.AppI.Math.,429,pp.2941
[3] Stewartson.K (1953) , On the flow between two rotating coaxial disks, Proc. Camb.
Phil. Soc. , 49,pp.333-341
[4] Lance.G.K and Rogers.M,H (1962) , The a....dally symmetric flow of a viscous fluid
between two infinite rotating disks, Proc.Ray.Soc.,A206 ,pp.109-121

372
[5] Greenspan.D (1972) , Numerical studies of flow between rotating coaxial disks ,
J.Inst . Matha . Applies .,9 , pp.370-377
[6] Holodniok,M.,Kubicek,M. and HlavacekV.,(1981) , Computation ofthe flow between
two rotating disks multiplicity of steady-state solution, J.Fluid Meeh ., 108 , pp.227240
[7] Holodniok,M.,Kubicek,M. and Hlavacek,V.(1977) , Computation ofthe flow between
two coaxial rotating disks, J.Fluid Meeh.,Sl, pp.689-699
[8] Nguyen,N .D.,Ribault,J .P. and Florent.P. (1975) , Multiple solutions for flow between
coaxial disks, J.Fluid Meeh., 68, pp.369-388
[9] Mellor.G.L.,Chapple.P.J.and Stokes.V.K. (1968) , On the flow between a rotating
and stationry disk, J.Fluid Meeh.,31,pp.95-112
[10] Roberts.S.M. and Shipman.J.S.(1976) , Computation of the flow between a rotating
and astationary disk, J.Fluid Mech.,71,pp.53-63

[H] Tosaka.N (1988) , The generalized boundary element method for nonlinear problems
, Boundary Element 10 ,(Eds.C.A.Brebbia) 1 , pp.3-17 , Springer-Verlag
[12] Tosaka.N (1990) , Generalized boundary element analysis of nonlinear problems ,
In:Boundary Element Method (Principles and Applications},(Eds.Q.Du and M.Tanaka)
, pp.113-122 , Pergamon Press
[13] K'l.kuda.K. and Tosaka.N.(1990) , The generalized boundary element approach to
Berger's equation, Int.J.Num.Meth.Eng. , 29, pp . 245-261
[14] Honma.T. and Tosaka.N. (1990) , Large deformation analysis for elastic rings using
the generalized boundary element technique, (in Japanese) In:Proeeedings of the
Seventh Japan National Symposium on Boundary Element Method, 7,pp.195-200
[15] L.Hormander (1955) , On the theory of general partial differential operators, Acta
Mathematica, 94, pp.161-248

The Coupling Approach of Boundary and Finite Element


Methods to Incompressible Viscous Flow Problems
N. Tosaka1), K. Kakuda1 ),
H. Yoshikawa2 ) and A. Anjyu2 )
1) Department of Mathematical Engineering, College of Industrial Technology,
Nihon University, Narashino, Chiba 275, Japan
2) Kozo Keikaku Engineering Inc., Nakano-ku, Tokyo, Japan

Summary
A coupling approach of the boundary element method and the finite element method for
the incompressible viscous flow problems is presented. A domain involving an obstacle
is divided into two subdomains. The subdomain involving an obstacle is assumed as an
incompressible viscous flow, and the finite element method is applied to simulate the flow.
The other is assumed to be a potential flow, and the boundary element method is applied
to the flow field. Numerical results of the flow past a step demonstrate the applicability
and effectiveness of our approach.

Introduction
The finite element methods have been successfully applied to not only structural mechanics but also fluid mechanics involving non-linear problems [1]. However, the singularities
such as arise in fracture mechanics as well as the exterior problems are not easily dealt
with. On the other hand, the boundary element methods have been introduced as only
the boundary discretization process which is more economical and possesses some advantages for potential problems and solid problems [2,3]. In the formulation this method
needs an explicit fundamental solution of the linear differential equations. Hence, in the
case of non-linear problems the above-mentioned merits are far less. Therefore it seems
to be most efficient to combine the two methods for solving the non-linear problems with
an exterior domain [4,5].
The main aim in this paper is to apply a coupling approach of the boundary element
method and the finite element method to unsteady incompressible viscous flow problems.
A flow domain involving .an obstacle is divided into two subdomains. The flow in the
subdomain involving an obstacle is assumed as an incompressible viscous flow governed

374
by the Navier-Stokes equations, and a suitable finite element procedure [6] is applied
to solve accurately and in a stable manner the Navier-Stokes equations. The other can
be assumed as a potential flow governed by the Laplace equation because the obstacle
and the viscosity for the flow are of no effect if the subdomain is established enough far
from the obstacle, and the boundary element method is applied to the flow field. The
two subdomains have an interface boundary which holds in common some points on the
boundary. In this approach the treatment of the interface boundary is of importance, and
an attempt is achieved. We present numerical results for the flow past a step to show the
efficient of our approach.
Throughout this formulation, the summation convention on repeated indices is employed.
A comma following a variable is used to denote partial differentiation with respect to a
space variable.

Statement of Problems
Let 0 be a bounded domain in Euclidean space R2 with a piecewise smooth boundary
The unit outward normal vector to

r.

is denoted by n .

The domain 0 is divided into two subdomains 0 1


domains have a common boundary part

O2 as shown in Fig.I. The two sub-

which is called the interface boundary. In this

stage, the flow in the subdomain 0 1 involving an obstacle OB is assumed as an incompressible viscous flow governed by the unsteady Navier-Stokes equations. The other O2 is
assumed as a potential flow governed by the Laplace equation. The governing equations
of these subdomains are given in the following nondimensional form, respectively:
In the subdomain 0 1 (Le., viscous field),

at + u u;J == -P,i + Re1


aUi

tli,i

== 0

tI,,)]

(1)
(2)

In the subdomain O2 (i.e., potential field),


<P,i;

in which Re is the Reynolds number,


and <p denotes the velocity potential.

Uj

=0

(3)

is the velocity vector component, p is the pressure,

375

Fig.1 Problem statement

Here, we apply the semi-explicit scheme to time derivative of equations (1) and (2) governing the sub domain

Q1'

Moreover, defining an auxiliary variable

u"

and applying the

fractional step scheme to the obtained equations, we obtain the following two types of
problem:
(a)Advection-diffusion-type problem
Ui-ui
nn
- -Llt
- + u )u.,)

= -Re1un',J]

(4)

(b)Linear Euler-type problem


(5)
where Llt is the time step, and n denotes the number of time step.

If for equation (5) we put

(6)
in which denotes the reduced velocity potential, and the relation between velocity vector
at {n + 1)-th time step and

is given as follows:
(7)

Moreover, by taking the divergence of equation (7) and taking the continuity equation (5)
into consideration, we obtain the following Poisson equation in terms of if> :

(8)

376

Coupling Approach of BEM and FEM


In this section we shall apply the Petrov-Galer"kin finite element method [6] using exponential functions to solve accurately and in a stable manner the equations (4) and (8).
On the other hand, the boundary element method [3] can be employed for solving the
Laplace equation (3).
Finite element model of 0 1
By applying the divergence theorem to the weighted residual statement of equation (4),
we obtain the following weak form :

in which

nj

denotes the subdomain in

nl ,

and

Ti

= ui,jn)/ Re.

Here, we employ the

following weighting function using exponential one [6J for Mo. in equation (9) :
M o (x 1, x)
2 -- or;;-'
L...J Na (x b x 2 )e-{al(NT"'i-"'n+a'(NT",j-X~)}
al

= vf"& ,

-r

a2

= v:;Re'

where No. is the shape function, vi( i

= 1,2) is

the velocity vector prescribed in

(10)

ni , "&

and Re are the suitable constants.


By applying the finite element discretization to equation (9), we obtain the following local
matrix form:

- {u?}il
Mail {fijh .6.t

"}
F { "}
+ K ail (n){
Uj
Uj il = ail To il

(11)

Moreover, equation (11) is also reduced as follows:


(12)
where

Mail (= F;;/ M'Iil)

is the lumped mass matrix.

Taking the equilibrium conditions of {Ttla on each subd~main into consideration and
setting up equation (12) for all the subdomains, we obtain the final system of equations
as follows:

(13)
where C is the diagonal coefficient matrix, and F" denotes the known vector which
consists of the velocity at n-th time step and the given boundary values.

377

On the other hand, by applying Galerkin-type finite element method to equation (8), the
weak form is given as follows:
(14)

The finite element model of equation (14) is also given as follows:

(15)
Setting up equation (15) for all the subdomains, the final system of equations is obtained
as follows:

(16)
where B is the sparse coefficient matrix, and F denotes the known vector with respect
to U a.nd the given boundary values.
Boundary element model of !h
We apply the boundary element method to solve the Laplace equation (3). The solution
4> has the following integral representation:

(17)
in which c is the shape coefficient, and the fundamental solution

1f"(X,~)

is given for two

dimensional case as follows :

1f"(X,~) = 211l" In.!.r

(18)

The boundary element model of equa.tion (17) is given as follows:


(i,j= 1,2,,N)

(19)

where N is the number of boundary nodes.


Solution procedure
In the following, let us mention the solution procedure of this approach.
1.

Give an initial velocity vector

un and calculate the auxiliary vector U

from equation

(13).
2.

Calculate the potential function


method to equation (16).

1> by applying SCG

( Scaled Conjugate Gradient)

378
3. Calculate p.. +1 and ui+1 at (n+ 1)-th time step from the weighted residual statement
of equations (6) and (7), respectively.
4.

Solve equation (19) by making use of the velocity values on the boundary

r. obtained

from step 3.
5.

Calculate the velocity vector and the pressure in the potential flow field, and go to
step 1.

Numerical Examples
In order to show the applicability and effectiveness of the present approach, numerical
results of the flow past a step are demonstrated. Fig.2 shows the geometrical configuration
of the problem. In our numerical performance we adopt the lowest interpolation functions
in which the velocity, the traction and the scalar potential are piecewise linear, and the
pressure is constant over each element. The initial velocities are assumed to be zero
everywhere in the interior domain. In the following, we show the numerical solutions for

Re

= 200 and a fixed time step b.t = 0.1.

The interface boundary

ri

has to be essentially given as the free boundary conditions. As

the first attempt of the treatment on the interface boundary, we adopt the free boundary
condition from the first time step to the final time step. The velocity vector fields obtained
herein are shown in Fig.3. As can be seen from this figure, the flow patterns are unnatural.
Hence we establish the boundary conditions as shown in FigA.

Potential flow

(V2~=O)

interface boundary
Viscous flow
(Incompressible N-S equations)

h
777777

II

yl,
I

7777;<:""\:\,,

14

Fig.2 Geometrical configuration

379

~.- - -- - ---- -------


-------------------

(b) t

(a)t=0.2
Fig.3 Velocity vector fields (Re

.. -.
,.~.CI

.....,

~
I.

..

... . . 0

~I

= 200 , t::.t = 0.1)

..

..

!! .0

~.-I

= 1.0

.. ..
~

I.'

,.

.....,
!!

1'. . .

!!
.0
I.

~,

!t. _.

~
o.

t,. t, .

..

--.
t

......

.j

J
'. _', .0

a '

/ / / /.

(a) Boundary conditions at 1st time step

(b) Boundary conditions after 2nd time step

Fig.4 Boundary conditions


Fig.5 shows the comparisons of the velocity vector fields by varying the ratios of h to a and
the one obtained by using only the finite element method. The corresponding pressure
fields are shown in Fig.6. As illustrated in Fig.5 (a), the unnatural wake phenomena
arises dearly on the interface boundary when the ratio
shown in Fig.5 (b), (c) the numerical solutions for hla

hi a is equal to 3.

However, as

= 4 and 5 do not exhibit the wake

phenomena, and their flow patterns are in good agreement with the one obtained by only
the finite element scheme as shown in Fig.5 (d). On the other hand, the pressure fields for
the various ratios hla are close in comparison with the finite element solutions as shown
in Fig.6 (d) .

380

11111Fiu~IJI!!!1111111111111
t

=5

(a) Velocity vector fields for h/a

= 10

=3

------------------- -- ------------

illlllll"IIIIIIII;

t = 10

t=5
(b) Velocity vector fields for h/a

=4

11111fFi@~ IIIII111111;IJIIIII ~;;;; ~' ~/![Di2iii JIIIIIIIIIIIIIIIIII


t = 10

t=5

(c) Velocity vector fields for h/a

=5

====:=::::======:::

!!Ii! ~i[[fL!7~i!!!!!!! I!I!IIII!!I! II! 11~~; ;;!![[nHF ~i ~~ ~ ~ ~ ~ ~~~ ii~ ii ii~


---.-

::::::::::==========

t=5

t = 10

(d) Velocity vector fields obtained by only the FEM


Fig.5 Velocity vector fields (Re

= 200 , L::.t = 0.1)

381

t=5

= 10

(a) Pressure fields for hla = 3

t = 10

t=5
(b) Pressure fields for

hI a = 4

(c) Pressure fields for hla

=5

(d) Pressure fields obtained by only the FEM


Fig.6 Pressure fields (Re

= 200 , t:.t = 0.1)

382

Conclusions
We have presented a coupling approach of the boundary element method and the finite
element method to unsteady incompressible viscous flow problems. The original domain
involving an obstacle was divided into two subdomains. The flow in the subdomain
involving the obstacle was assumed to be an incompressible viscous flow governed by the
Navier-Stokes equations, and the Petrov-Galerkin finite element method using exponential
functions was applied to solve the Navier-Stokes equations. The other could be assumed to
be a potential flow governed by the Laplace equation, and the boundary element method
was applied to solve the equation.
As the numerical example, we demonstrated the numerical solutions for the flow past a
step through a comparison with the one obtained by only the finite element method. It
was shown that the numerical solutions were seen to be satisfactorily in agreement with
the finite element solutions.
References

[1] Thomasset,F.(1981}. Implementation of Finite Element Methods for Navier-Stokes


Equations, Springer-Verlag.
[2] Banerjee,P.K. and Butterfield,R.(1981). Boundary Element Methods in Engineering
Science, McGraw-Hill, London.
[3] Brebbia, C.A., Telles, T.C.F. and Wrobel ,L.C. (1984). Boundary Element Techniques , Springer-Verlag.

[41 Zienkiewicz,O.C., Kelly,D.W. and Bettess,P.(1977}. The Coupling of the Finite


Element Method and Boundary Solution Procedures, Int. J. Num. Meths. Engng.,
Vol.U, 355-375.
(5] Wendland,W.L.(1988). On Asymptotic Error Estimates for Combined BEM and
FEM, (Eds., Stein,E./Wendland,W.L.) Finite Element and Boundary Element Techniques from Mathematical and Engineering Point of View , Springer-Verlag.

[61 Kakuda,K and Tosaka,N.(1991}. Numerical Analysis oflncompressible Viscous Flow


Problems by the Petrov-Galerkin Finite Element Method Using Exponential Functions, (in Japanese) Proc. Symp. on Computational Methods in Structural Engineering and Related Fields, VoLlS, 11-16.

Boundary Element Method for Couple Stress Theory


of Elasticity
S. Yamashita
Department of Mechanical Engineering.
Oyama Technical College. Oyama. Tochigi 323. Japan

N. Tosaka
Department of Mathematical Engineering.
College of Industrial Technology.
Nihon University.Narashino.Chiba 275. Japan

Introduction
The

concept

of

characterize

stress

in continuum

the internal forces.

mechanics

was

introduced

to

In classical elasticity the stress

tensor is symmetric. However. there occur large stress gradients in the


stress concentration in the vicinity of holes. near notches and cracks.
In this case we must take consideration of the influence of the micro structure of the

continuum in

improved or generalized theory


(1]-[5]

structure.

Such theory

question.
based on

is called

Consequently.
the influence

we need the

of the micro -

the Cosserat theory

of elasticity

theory. asymmetric elasticity or couple stress elasticity.


In this paper we would like to show the applicability
element method

to

couple stress

integral equation
value problems
solution derived
d e t 31. I( 6]

and

are

theory

the related fundamental solution

derived explicitly.

in our paper

is

of the boundary

of elasticity.
Especially.

compared with

The boundary
of boundary -

the fundamental

the existing

one in

384

Fundamental Equations
Let us consider the two-dimensional boundary value problems

of linear

n be a
bounded domain in R2
and have a piecewise smooth boundary r . The
uni t outward normal vector component to r is denoted by n-i'
couple stress theory

for an isotropic,

In the following discussion,


convection for repeat index.

elastic medium.

we would

like

A comma

to adopt
is used

Let

the

summation

to denote partial

differentiation with respect to a space variable.


The local balance equations governing
under the nonexistence of body force

micropolar linear
and

body

torque

elastic body
are

given

as

follows:

1:j-<.,j = 0

( in

n )

(l)

in

n )

(2)

where -<.jk is the Eddington's epsilon,


component of stress and couple stress

and 1:1..j and \1-<.j are the


tensor, respectively. The

constitutive equations are given by


1:

-Lj

= Ad -LJ
,

+(

U I.. I..
11
r<, r~"

+ CL

uj ,-L

( in

n )

(3)

(4)

in which

u-<.

is the component of

component of rotation vector.

displacement

vector.

001..

is the

is the Kronecker's del tao and A

v . e . are independent elast ie constants.


On substituting equations (3) and (4) into equations (1) and (2).

.)J

CL ,

we

can obtain the balance equations expressed in terms of displacement and


rotation as follows:

Adj-<. Uk,kj + (

II

+ CL

u1..,jj+ (

\l -

a ) Uj,-<.j - 2aj1..k 0011.,]

=0

(5)

385

(6)

Equations (5) and (6) must be solved in conjunction with the following
appropriate
and

rq

boundary

conditions

on

ru . r-r

each boundary

. roo

ru )

( on

r-r

( on

( on rq )

For derivation of the integral equation formulation.

(7)

it is convenient

to rewrite the above into the following matrix form

L
'-<.} U
}

( ',j = 1,2,3 )

in n )

(8)

U3=

For two - dimensional


w3= W
= ua (a=1 ,2) and
case the above form can be expressed concretely as follows

Ua

in which we put

(~+a)6+(A+~-a)Of

(A +l1-a) 0201

(A+~-a}Ol 02

(~+a}6+(A+11-a}D~

- 2a 02

2aD 1

2a02
- 2aD l

u1
u2

(v+Il)6-4a

=0
(9)

where D, = () /'dx.(. and 6 = D.(.D.(. denotes the two - dimensional Laplacian.

Integral Equation Formulations


In the first

place.

we

start with the following

weighted

residual

statement of the differential equation set (8) for the weighting tensor
function

v1;.

386

i, j, t = 1, 2, 3 )

Integrating by parts and taking the


consideration.

~ r

Vjl

operator

Lij

into

we can obtain a set of integral equations of the form

..., ( (1)( \I + 13 )

where

defferential

(10 )

V;l

n dr

'

(l

= l,2,3,P. = 1.2

(11 )

must be determined as the fundamental solution tensor which

satisfies the differential equation for the adjoint operator

L~j

of

ltj

such that
(I 2)

with the adjoint operator

'"j

J
l

*
and rpt

(ll+a)b.+(A+ll-a)oi

(A+ll-a) 0201

(t..+ll-a) 01 02

(ll+a)M(A+ll-a)02

-2aO?
the traction

corresponding

-2aDl
(\I+S)l;-4a

2aD1

l..

is

2aD2

to

(13)

the traction vector 'P

which is defined as

+ [(

jJ

a )

V;l, ~ + (

II - a )

(
From the equation (11).

VEt, p]
p,~,k

n~

= 1,2,l

= 1,2,3

(14)

we can obtain the following boundary integral

equation set in terms of the unknown functions:

387

Clj is

where

properties on

the

coefficient

matrix

which

both the location of a boundary node

geometry at the source point

is

determined

with

and the local

Fundamental Solution
Let us consider the constraction of fundamental solution tensor
(7]

by using the Hormander's approach which was

*
Vj'l

well - known in the theory

of partial differential equation.


The fundamental solut ion tensor

Vjl

can be expressed as follows


(16 )

is the transposed co-factor matrix of

where

+ 4a 2 00.o 13

L..
-<.j

given by

(a,S = 1,2)

M13

= - 2 a

D2~

+ 2jJ

M31

2 a

D2Ll ( A

+ 2].1 )

M23

2 a

Dl~

+ 2].1

M32=-2a

Dl~ ( A

+ 2].1 )

)~2{ A

+ 2lJ

~~3 = ( ].1

+a

(17)

(I8)

4l *
is the fundamental solution for the differential operator
L = det ( L ) which satisfies:

and

-<.j

(19 )

388

From the adjoint operator

l.[j of L.[j , we can obtain

immedeately

as follows :
(20)

in which K'

and yare given as

= ( A + 2~

)(

+e

)(

( 21)

+a )

4 a ~
=---~----

( v +

Consequently,

,*

e )(

+a )

(22)

is the scalar function which satisfies next equation


(23)

Moreover,

we can rewrite the equation (23) into the following form

,* =

1 (
K,y2

A -7)(-O(X-Y

A- Y

(24)

From the expression (24), we can find that the fundamental solution $*
is composed of the summation

of

three fundamental

solutions

modified Helmholtz operator, Laplace operator and biharmonic

for the
differen-

tial operator.

The concrete expression of the fundamental solution $ *

is

given

as

follows:
4> *

_1 (

where r = If x _ y

Ko( )
zero.

(25)

K
~

denotes the distance between

and

,m =

ry,

is the modified Bessel functions of the second kind of order


Here the constant

K= 2

rr y (

A+

K
2~

is defined as follows:

)( v + S )(

+a

(26)

389

The

fundamental

solution

tensor

can be

given

relation (16) with the aid of (17). (18) and (25).

completely

by

the

The concrete expres-

sion of the fundamental solution tensors are as follows:

*
V32

* = '1'4*
V33

= '1'3 r, 1

( 27)

where

'1'1 =

("'lm 2

+ m2

2 m4

+ ( -2 m3- -

m2

) Ko(mr)

(-m m3 -

_m4

1
) - Kl(mr)

- r2 + -2m4

r
(28)

in which

m,
m3

= ( A + 2U )( v +

= - ( A+ U -

ms = (

a )/K

a )( v +

a )/K

m2 = - 4 a

m4

=4 a

(A +

2u

( A + U )/K

A + 2u )( u + a )/K

In reference [6] we can find the fundamental solution

(29)

for

elastodynamics by using the Laplace transform with respect


variable.

)/K

micropolar
to the time

390

Discretization
In order

to

get

equation set (15).

approximate

solutions

let the boundary

of

the

boundary

r be divided into N

elements. In tbe case of the constant element.

the

unknown

involved in equation (15) may be represented approximately

integral
constant
functions

in terms of

the nodal vector as follows

(30)
where Xj

denote the central point on the boundary element.

Substituting

of

the discretizations

of

unknown functions (30) into

equation (15), we can obtain the following equation.

Moreover. equations

(31)

can be expressed as follows:


(32)

where .i.

"iP

is as follows

(33)

In the equation (32), the expression of matrix for all nodes result in
the following equation.

391

HU + R w

(34 )

Conclusions
For two - dimensional couple stress theory of elasticity.

we

derived

boundary integral equation and its boundary element discretization.


Moreover. we showed the concrete expression of fundamental solution.
In order to

find

the numerical solutions.

some

numerical

examples

could be performed as near future theme.

References
[1] R. D. Mindlin

Influence of
tions -

couple - stress on stress concentra -

Experimental Mechanics.

Vol. 3. p.l -

1963. Society for Experimental Stress Analysis.


[2] R.D.Mindlin

Micro - structure in linear elasticity' . Archives


of Rational Mechanics and Analysis. Vol. IS. p.51-78.
1964. Springer - Verlag.

[3] R. A. Toupi n

Theories of elasticity with couple - stress" . Archives of

Rational Mechanics and Analysis.

Vol. 17.

p. 85-112.1964. Springer - Verlag.


[4] A. E. Green

" Micro - materials and multipolar continuum mechan ics

Int. J. Engng

Sci.. Vol. 3. p.533-537. 1965 .

Pergamon Press.
[5] A. C. Eringen

"Theory of
(ed.) "
Press.

Micropolar Elastici ty.

in H. Liebowi tz

Fracture. Vol. 2. p. 621-729. 1968. Academic

392
[6) J. Balas, J. SI adek and V. Sladek:" Stress Analysi s by Boundary Ele ment Methods ", Studies in Applied Mechanics, Vol. 23,
p. 441-497,1989, ELSEVIER.
[7] L. Hormander

"Linear partial differential operator (2nd


printing )" , 1964, Springer - Verlag.

revised

Analysis of Wave Motion over Submerged Plate by Boundary


Element Method

Xiping Yu
Technical Research Institute, Toa Corporation, Anzen-cho 1-3, Tsurumi, Yokohama 230, Japan
Masahiko Isobe and Akira Watanabe
Department of Civil Engineering, University of Tokyo
Kazuhiko Sakai
Division of Technological Development, Toa. Corporation

SUMMARY
Wave transformation over and wave force on a submerged plate are computed by the
boundary element method. Two mstinct forms of the fundamental solution, referred as the
logarithmic form and the series form, are employed in the computation and their results
are compared with each other so that the principle for selecting the fundamental solution
in

eng~neering

applications is suggested. Improvement in the numerical processing of the

series form of fundamental solution is also achieved.

INTRODUCTION
It is known that the boundary element method is of practical advantage to solve the

problems of wave transformation over and wave force on submerged barriers. The method
may also be applicable when the submerged barrier is a plate. When the boundary value
of the velocity potential is directly solved, the water surface elevation and the pressure
distribution on the plate can be obtained straightforwardly if the linearized kinematic free
surface boundary condition and the Bernoulli equation are considered. Then, the wave
reflection, transmission and force can be determined.
Application of the boundary element method to the computation of the wave transformation and force related to a submerged plate was carried out by Liu and Abbaspour
{1982}. Liu and Abbaspour's study was based on the logarithmic form of fundamental
solution. On the other hand, a fundamental solution which is oriented to the wave proh-

394

lems and satisfies the linearized free surface, bottom and lateral boundary conditions is
also available according to Wehausen and Laitone (1960). The present paper is to demonstrate the utilization of such a boundary-condition-satisfying fundamental solution to the
submerged plate problem. Comparison is also made to Liu and Abbaspour's approach
and the principle for properly choosing the fundamental solution is finally proposed.

FORMULATION
With the assumption of irrotationality, the wave motion over a plate submerged in
an ideal fluid, as shown in Figure 1, is described by the Laplace equation in terms of the

Figure 1: Definition sketch of wave motion over submerged plate


velocity potential <P. As our interest is only in the periodic solution, <P may be considered
sinusoidal with respect to time. By expressing the velocity potential by the sum of two
parts representing the incident and scattered waves, respectively, the time-independent
part of the scattered potential rjJ is then obtained. Naturally, rjJ is also governed by the
Laplace equation:
(1)

The boundary conditions for rjJ are summarized as


8rjJ
-+I<rjJ+v=O

(2)

8n

where n stands for the normal direction of the boundary and the values of I< and v are
given in Table 1 for various kinds of boundaries, where the parameter

(J

is the angular

frequency, k the wave number and 9 the gravity acceleration; the incident wave potential

395

Table 1: Values of K and v for various boundaries


boundary

free surface

_(J2/g

bottom

lateral

-ik

plate surface

84>0/8n

4>0 is expressed by
4>0

= gHocoshk(z + h)e ikx


2(J

cosh kh

with h representing the water depth, Ho the incident wave height, x the horizontal and z
the vertical coordinate.
Let a fundamental solution be denoted by Gi

== G(x, z; Xi, z.), where

(Xi,

z;) are the

coordinates of the source. Green's identity, therefore, gives the following integral equation:
(3)
where

is the boundary along which Eq. (2) is not satisfied by G,; the factor

C(i)

depends

on the geometry of the boundary where the source is located and takes unity when the
source is put at an inner point of the domain concerned.
Numerical processing of Eq. (3) is usually necessary. Thus, we may have to discretize
the boundary

into elements so that


Ne

r = Lr(e)

(4)

e=l

where N e is the total number of the elements. Under this circumstance, Eq. (3) may be
rewritten as

+ e=1
L Jf
Ne

C(i)i

r(e)

8G.
4>(e)-8' dr n

Ne

L Jf

.=1

r(e)

84>(0)
-8-G; dr
n

=0

(5)

where the superscript (e) represents a boundary element with a and f3 its two nodal
numbers. We introduce the following linear interpolation function L~e) and L ~e) in the
element to express the coordinates, the potential and its derivative in terms of their nodal
values:
x(e)

=L
k=",/3

LkXk

z<el = L

k=a,/3

Lkzk

396

4/ e ) =

I:

k=a,/3

Lk<Pk

where
L~e)

= (1 -

L~e)

~)/2

= (1 + ~)/2

and ~ E [-1,1]. The integration in Eq. (5) may thus be led to

f <p(e) oG;
Jr)
on

dr

= <p~e) jl
=
-

jr)

o",(e)

_'I'-Gi

on

1 - ~ oG i
2 on

-1

dr

j<e)

d~ + <p(e)
/3

jl

-1

1 + ~ oG; J(e) d~
2 on

M(e)",

(6)

'J 'I'}

o",(e) jl
= -'1'-"-

on -\

0"')

~G;J(e) d( + _'1'_/3_

jl

on -\

1+ c
2

- - < , GiJ(e)

d~

=N(e)O<P]
-

where

J(e)

= V(X/3 -

xa)2

'J

(7)

on

+ (Z/3 -

z,,)2 and the summation is taken for j to include all

the boundary nodes. In Eqs. (6) and (7), the integration with respect to

is numeri-

cally carried out by Gaussian quadrature under normal conditions. When the source is
included in the element, G; and

oGdon become singular and special consideration is then

necessa:y in order to ensure the accuracy. For the logarithmic fundamental solution, these
integrations can be analytically performed (Brebbia, 1978). If the fundamental solution
takes the modified series form as mentioned in the next section, the method proposed by
Kutt (1974) may have to be referred.
Substituting Eqs. (6) and (7) into Eq. (5), we get
(8)
When t,he subscript i takes from 1 to N e, that is, when Eq. (8) is applied to all the
boundary nodes, a complete set of linear algebraic equations in terms of all the nodal
potential values may then be formed, Solving the equations by the method of Gaussian
reduction, we can finally obtain the distribution of the potential on the boundary

r.

397

FUNDAMENTAL SOLUTIONS
The fundamental solution may be selected on the basis of two different concepts. One
is to use a simple form which does not satisfy any boundary condition. In this case, a
relatively long path of integration including the free surface, the bottom, the left and
right lateral boundaries which must be separated far enough from the submerged plate to
ensure the Sommerfeld condition satisfied, must be dealt with as the cost of the simplicity
of the fundamental solution. On the contrary, it is also possible to apply a function which
satisfies, for instance, the free surface, the bottom and the lateral boundary conditions
at infinity. The expression of such a function becomes consequently rather complicated.
The benefit is that we need only to consider the boundary of the plate surface in the
numerical computation.
A typical simple form of the fundamental solution is well known as the logarithmic
function:
G 1 ==
I

~ln
411"

{x -

Xi)2

+ (z -

(9)

Z;)2

A number of wave transformation and wave force problems have been successfully solved
by using this function (Bird and Shepherd, 1982; Liu and Abbaspour, 1982).

The

boundary-condition-satisfying fundamental solution, on the other hand, is also widely applied (Macaskill, 1979; N aftzger and Chakrabarti, 1979). The function can be expressed
as a series form (Wehausen and Laitone, 1960):

G2 ==_
I

2
1: k1n k~hP+v
+ v2h _ V cosk
00

=1

-i

(h+z)cosk (h+ z)e-knlx-x,1


I

1
~-~
coshk(h + Zi) cosh k(h
k Ph - v 2 h + v

.
+ z)e,k1x-x,1

(10)
11.=1

where v =

0- 2 /

9 and k n are the real roots of the equation

v + k" tan knh == 0

(11)

Equation (10) includes an infinite series. It may be found that the series is slowly
convergent when x is close to Xi. Under this condition, numerical calculation will be very
inefficient. However, the series may be improved if Gt is expressed as
00

G; == - 1:(Pn - Qn) n=l

1: Qn 00

n=l

Po

( 12)

398

while
(13)
Since k n -+ mr/h as n -+

00,

Qn is obviously the asymptotic expression of Pn as n

increases. The newly-created series

E::"=l Qn,

on the other hand, is able to be expressed

by an analytical function (Moriguchi et al, 1957).

1
1--0;1
L Qn =- -In(l2e-"
k
00

~1

z - z

COS1l"--'

+ e-2r

10-%;1
- -1 In (1 - 2e- .. .,.cos 11" 2h + z + Zi
411"
h

Therefore, the series

E::"=l Pn

10-%;1
h

+ e _2.. 1',:;1)

which must be dealt with in Eq.

(5)

(14)

is replaced by

E::"=l(Pn - Qn) in Eq. (12). The series E::"=l(Pn - Qn) always vanishes within a few
terms because Pn tends to Qn very rapidly.

SAMPLE COMPUTATION
The wave reflection coefficient K, and transmission coefficient K t are defined by the
ratios of the reflected and transmitted wave heights to the incident wave height, respectively, that is,

= H,/ Ho = 2u\<p\(x=-oo;z=ol/(gHo)

(15)

= Hel Ho = 2u\<po + <P\(,,=oo;z=ol/(gHo)

(16)

K,
KI

The wave force, on the other hand, is described by the following nondimensional parameter
in the present study:

KJ

= F/(pgHol.)

(17)

where I. is the length of the plate and F the magnitude of the total force on the plate
which can be obtained through integrating the dynamic pressure along the plate surface.
Figure 2 shows the variation of K" KI and K J due to the change of the wavelength L".
The plate length I. in the computation equals to the depth h while the relative submerged
depth of the plate d/ h is changed from 0.2 to 0.4. The length of each element is taken
1/20 of the depth and the lateral boundaries are separated by 2h from the edges of the
plate when the logarithmic fundamental solution is applied. Though the results from
the two distinct forms of the fundamental solution are rather close to each other, some
differences are still observable. Comparing with the exact solution obtained by matching

399
1.0

]{r

O. B
O. 6
0.4
0.2
0

O. 1 0.2

0.3

0.4

0.5

0.6

O. 7

llLh

]{,

10

0.8
0.6
0.4

0.2
0

]{,

O. 1 0.2

0.3

0.4

0.5

0.6

0.1

0.1

0.3

0.4

0.5

0.6

0.7

e; Lh

O. 8

0.6
0.4

O. 2
0.2

llLh

Logarithmic form of fundamental solution


Series form of fundamental solution
Figure 2: Computed wave reflection, transmission and force coefficients

[f./h= 1.0, d/h= (a) 0.2, (b) 0.3 and (c) 0.4]

400

1<,

0.8
0.6
0.4

O. 2
O~~--~~--~--~~~

O. I

O. 2 O. 3 O. 4 O. 5 O. 6 O. 7
llLh

1. 0 r--'--~[(,

0.8

0.6
0.4
0.2
O~~--~--~--~~~~--~

O. 1 O. 2 O. 3 O. 4 O. 5 O. 6 O. 7
llLh

[(I

0.6
0.4

0.2
O. 1 O. 2 O. 3 O. 4 O. 5 O. 6 O. 7
llLh

Logarithmic form of fundamental solution


Series form of fundamental solution
Figure 3: Computed wave reflection, transmission and force coefficients

[dlh= 0.3, ilh= (a) 1.0, (b) 1.5 and (c) 2.0J

401

the asymptotic potential expressions in the subdivided domains, the series form is found
to be more accurate.
The same indication can also be given if we investigate into another computational
example, shown in Figure 3, where the relative submerged depth of the plate is fixed by
0.3 and the plate length to water depth ratio is varied from 1.0 to 2.0.
The computation shows that the series form of fundamental solution is of advantages
not only in terms of the accuracy but also the required computer capacity and the computational time. Detailed study on the ratio of the required CPU time, T2 , by applying
the series form to that, Tl , by applying the logarithmic function is carried out. The result
is given in Figure 4 against f2/ f l , the ratio of the length of the discretized boundary corresponding to the two cases. It may be noticed that using the series form of fundamental
solution may cause only 1/3 of the computer time needed by using the logarithmic form
even when the plate surface accounts for almost half of the total boundary length. The
series form is, therefore, predominant in reducing the computational time.
O.4.-----~----~----~----_.

T2 /Tl

0.3

0.2
O. 1
O. 1

0.2

0.3

0.4

O. 5

Figure 4: CPU time required by using different fundamental solutions


However, we can also reason from Figure 4 that instead of the plate, if the submerged
barrier has a fairly complicated configuration, the major number of nodes may then have
to be distributed along the surface of the barrier. In such a case, the simple logarithmic
fundamental solution may be more preferable.

402

CONCLUDING REMARKS
Wave reflection, transmission

an<~

force coefficients are computed by the boundary

element method. The computation is carried out on the basis of two distinct fundamental
solutions, the simple logarithmic form and the series form. Numerical processing of the
series form is improved through introducing an analytical function. Computational results
from the two forms of the fundamental solution are compared with each other and it is
found that both forms give fairly satisfactory results but the series form is of some advantages in terms of the accuracy, the required computer capacity and the computational
time especially when the plate is short.

REFERENCES
[1] Bird, H. W. K. and Shepherd, R. (1982): Wave interaction with large submerged
structures, J. Waterway, Port, Coastal and Ocean Div., ASCE, Vol. 108, No. WW2,
146-162.
[2] Brebbia, C. A. (1978): The Boundary Element Method for Engineering, Pentech Press,
46-103.
[3] Kutt, H. R. (1974): The numerical evaluation of principal value integration by finitepart integration, Numer. Math., Vol. 24, 205-210.
[4] Liu, P. L-F. and Abbaspour, M. (1982): Wave scattering by a rigid thin barrier, J.
Waterway, Port, Coastal and Ocean Div., ASCE, Vol. 108, No. WW4, 479-490.

[5] Macaskill, C. (1979): Reflection of water waves by a permeable barrier, J. Fluid M ech.,
Vol. 95, 141-157.
[61 Moriguchi, S., Udagawa, K. and Hitotsumatsu, S. (1957): Mathematical Formulas,
Vol. 2, Iwanami Press, 340p (in Japanese).
[7] Naftzger, R. A. and Chakrabarti, S. K. (1979): Scattering of waves by two-dimensional
circular obstacles in finite water depths, J. Ship Res., Vol. 23, No.1, 32-42.
[8] Wehausen, J. V. and Laitone, E. V. (1960): Surface wave, in Handbuck der physik
Vol. 9, Springer-Verlag, 446-778.

Accurate BEM Elastostatic Analysis for Very Slender Body


and Thin Plate
Ryoji

YUUKI~.

Guo Qiang CAO"'* and Katsumi MAEKAWA "**

Institute of Industrial Science. Univ. of Tokyo. JAPAN


CRC Research Institute. Inc . JAPAN
U"Sony Corp . JAPAN

Introduction
Recently it becomes very important to analyze the bonded structures
with thIn plates, the dissimilar materIals, the adhesive joints, and the
laminate plates. But in the usual boundary element elastostatic analysis
(DSM) of very slender body and very thin plates. it is known that the
unstable and inaccurate solutions are sometimes encountered. In the present
paper. we point out the factors of this inaccuracy and provide the
strategies for each factor to obtain an accurate solution. From this study.
the highly-accurate solutions can be obtained even If the aspect ratio of
slender beams and thin plates is In thousands and hundreds.
Problems of Slender Body and Thin Plate
Consider the problem of the slender beam or thin plate with a fixed end
as shown In Fig.!. It is known that the solution of the usual boundary
element elastostatic analysis Is unstable and Inaccurate when the aspect
ratio becomes large. This problem causes extreme dIfficulty in solvIng the
structure consisting of slender beam or layer and thin plates. As a sImple

" fixed

end

(a) Slender beam

fixed end
(b) Thin plate

Fig.l Problems of the thin plate or the slender beam with a fixed end

404
and critical example, the results of BEM analysis for the bending problem of
a cantilever beam are given in Fig.2. When the aspect ratio of the beam Llh
is greater than 10, the solutions using usual BEM analysis (DSM. Direct
Singular Method) become inaccurate. Here the DSM solutions are calculated
including the improvement of numerical integral accuracy. especially
developed for the analysis of the slender body as mentioned later. It is
noted that the fixed end condition is the main factor causes this
inaccuracy. Whereas. by using the DRM (Direct Regular Method) in which no
singular integral exists by locating all source points of fundamental
solution outside the analysis domain, the accurate solutions can be obtained
and the usefulness of DRM analysis for this problem is also discussed later.
Error(~)

8.0

of deflection at A

r-----------------,
'\ORM

-8.0

~cQ

-16 . 0

1- \

-24.0
-32.0

':

~~OSM
I

-40 . 0

10

210

l t
I
410

I.

At~ Ty

610

810

L/h
1010

Fig.2 BendIng of the cantilever beam


Direct Singular Method (DSM) and Direct Regular Method (DRM)

For an elastostatic boundary value problem shown in Fig.3. the boundary


integral equation Is obtained from Betti's reciprocal theorem as follows.

(P,Q E r)

(1)

where Ui j Ti; are the fundamental solutions and Ci j are the constant which
can be obtained automatically from rigid body motion condition. The usual

Flg.3 Boundary value problem for an elastIc domain

405
BEM discretizes the equation 0). taking the unknown displacements u;(Q) and
the unknown tractions t](Q) on the boundary. Here the integral is singular
and so this method is called Direct Singular Method (DSM). The source point
P is usually to be located on each node of the elements (1)
If we locate the source point out of the elastic domain as point p'"
(Fig.3). the equation (1) reduces to

(p* E O",Q E r)

(2)

and this method to dlscretlze the equ. (2) is called DIrect Regular Method

(DRM). Since the singular integral can be avoided. the DR~ attracted much
concern to analyze elastic problems effectively. However. authors have
investigated the usefulness and limit of this method through many numerical
experiments. It is found that the DRM is useful only for analysis of the
simple problems {:;'" .
Improvement of NumerIcal Integral Accuracy
In BEM elastostatlc analysis for slender bodies and thin plates. it is
very important to consider the improvement of numerical integral accuracy
as shown in FlgA. because the distance d between the source point P and the
element Q to be integrated becomes very small comparing the element length
S. In this case the integral between P and Q becomes near to a singular
integral and we call it a quasi-singular integral. It is found that a
highly-accurate scheme to improve the accuracy of this numerical
quasi-singular integral is required (3-4'. In this study. we employ the
sub-element division scheme for this purpose when the ratio diS is less than
0.25. Moreover. the slender elements in the 3-D BEM analysis of thin plates
usually exist as shown in Fig.! and the highly-accurate integral scheme for
these elements is also necessary. In Fig.5 we show a sub-element division
scheme to improve the numerical integral accuracy in quasi-singular
integral for the highly-accurate analysis of slender beam and thin plate.

a
Fig.4 Quasi-singular integral in analysis for slender beam

406
P : Sauces Point

near corner node

!Xl

near middle node

9P

d= R,:o; 0.258

LIS 1 ~ O. 25 or

d= R,:o; 0.258

Min. of 51:
element lengtn
S2:
. _
Max. of element length

d = min[R" R, R,J

L/S2~

O. 05

near center

(a) Analysis for slender beam


(b) Analysis for thIn plate
Fig.5 Sub-element scheme used in the quasi-singular Integral
P : Source Point

Singular Integral:

/-a---1

N~Z

Quasi-singular Integral:
I

(al

il!

/
(b) Slender E! emen t

Ordinary

(alb;': 10)

Element(a/b<IOl

Fig.S Sub-element scheme used in the slender element


respectIvely. Fig.S shows a sub-element division scheme used for the
numerical integral in the slender element. It is confirmed that these
sub-element divIsion methods are very useful and efficient for the
improvement of numerical integral accuracy.
Fig.7 shows the results of a slender beam under pure bending by using
the usual BEM analysis and our hIghly-accurate BEM analysIs with the
mentioned sub-element scheme as well as DRl'l analysis. When the aspect ratio
Error(%) of deflection at A

1.5

r----------------,
OSM _ _

1.0

wi thout sub-

....... .. ORM./

0.0

-0.5

OSM with

element

-1. 0

-1.5

.......

.... Go ..

element

0.5

SUb-C

bL

\I

__

400

__

___

800

~)

~ OA_

M
~

.............

L
~

1200

____

M
~

1800

_ _ _ _- J

Llh

2000

Fig.7 BEM results of the slender beam under pure bending

407
Llh is greater than 200. the solution of the deflection at point A becomes
unstable and inaccurate in the usual BEM analysis. Using our
highly-accurate BEM analysis. the deflection Is obtained correctly even if
the aspect ratio Llh is in 2000. Since the source point of the fundamental
solution is located out of the boundary in DRM. the singular integral and
the quasi-singular integral are avoided and the highly-accurate solution is
also obtained by using DRM. Fig.S shows the 3-D results of the thin plate
under pure bending. By using the DSM with the sub-element scheme to improve
the numerical integration accuracy. the accurate solution of the deflection
at point A is obtained when the aspect ratio W/H is in 200. But in the usual
BEM analysis the accurate solution can be obtained when the aspect ratio
W/H is only in 50.
Errore:) of defle ct io n at A

10.0,---=-__- - - - - ' . /, - - - --

~
L --- ------ -M)
fM/

w=50.0

. . ....

.
:

-,

.,

0.0 _ _ _ ""

\" DSM without


\ sub-element

"""'"

DSMwith ~
sub-element

-100-""'0I'
---;:O::---'7r.::----:-r:::----r-:---~
50
100
150
200
250
W/H

Fig.S BEM results of the thin plate under pure bending


Utilization of DRM Analysis for a Beam or Plate with Fixed End
It is well known in the elasticity theory that the boundary condition on
the fixed end of a cantilever beam or plate is usually replaced by satisfying
the resultant force and moment condition. based on St. Venant prlnciple(!S) .
This replacement is also important in the BEM elastostatic analysis. In Fig.9
we give the moment distribution along a cantilever beam calculated by DSM
M/{Ty*L}

1.00

1"'<:::----------------,
' (;) ' 0

0 . 75

M/(1Y"L) -l-x/L
1S).

~TY

./"

G....

0.50

Beam Theory

0.25

X
'1:0).

DR#

,y
"" . .

L/h=50

ODS ."

x/L
0 . 00

0 . 20

0 . 40

0.60

0 . 80

1.00

Fig.S Bending moment along the cantilever beam

408
and DRM, respectIvely. It Is found that the sIgnificant error of the moment on
the fixed end causes so that the DSM solution becomes inaccurate. But the
correct moment distribution along the beam is obtained by using DRM. The
numerical results demonstrate that the DRM Is particularly suItable for the
correct treatment or replacement of the boundary condition on the fixed end.
We InvestIgated theoretically the reason why DRM is useful for this problem
by using Taylor's series expansIon of equation (1) and (2) in each small
elements, respectively'3. 0) . Here, since the fundamental solution is
analytical In DRM analysis we can express the fundamental solution in the
form of Taylor's serIes in each elements and it is found that the resultant
force and moment are adjusted automatically In the DRM if the source point is
located apart from the boundary. On the other hand, In DSM such a Taylor's
serIes expansion is ImpossIble and the solution is affected by the boundary
condItion very sensitively.
The solutions of the deflection of a cantilever beam have been gIven in
Fig.2. The DSM analysis gives the accurate solution only in Llh ~ Ie. But the
DRM analysis can gIve the solution accurately even if aspect ratio Llh is in
Ieee. For the problem of a cantilever plate shown in Fig.le the accurate
solution is obtained by using DSM when the aspect ratio Llh is only in 15. By
using DRM the accurate solution can be obtained when the aspect ratio is in
125.
Error( : ) of deflection at A

10.0-.--------------------,
w=50.0
DRM

~: \:~ ;SM ~t. I:v


.-A,

.~

:~
'\

BAY

,;.

XT

-20.0-L,6--.--.----,...:..:..:....,....-,0
...0-.,.---,..---.,.--.-::20=0
W/H

Fig .1e Bending of the plate with fixed end


Utilization of Sub-Region Division Scheae
The BEM analysis with the sub-region division is very useful for solving
the more complicate problems of the slender body and the thin plate with a
crack or notch. The analysis strategy is to divide the domain to the
sub-regions where the SUb-region with the complicate stress distribution is
solved by using usual BEM (DSM). while the other sub-region such as the
slender beam and the plate is solved by the highly-accurate BEM mentioned

409

(a) Uniform tension

(b) Pure bending


Fig.ll Beam with an edge crack

~
crack

t ~

REG/ONA

f-- l !

"

REG/ON
I ..

l2

(a) Two sub-regions

REGIDN A
( lJlw=2)

DE]

(b) BEM meshing


Fig.12 Analysis of the beam with an edge crack
previously.
As an example, we consider the problem of the slender beam with an edge
crack as shown in Fig.1I. As shown in Fig.l2, this problem is divided Into two
sub-regions A and B. The sub-regIon A is solved by usual BEM and the
sub-regIon B is solved by either DSM with the sub-element division scheme or
DRM. The stress Intensity factor is calculated and the results are given In
Table 1 and Table 2 for the cases of uniform tension and pure bending,
respectively. The usual BEM analysis without sub-region division scheme is
possible to obtain the accurate solution only when the aspect ratio of the
beam is less than 25(3). But the sub-regIon dIvision method gives the
highly-accurate solution even If the aspect ratio is In 2000.

410

Table. 1 Nondimensional stress intensity factor for the beam with


an edge crack under uniform tension
LfIV

F,

Hl

213

lee

21313

see

Ieee

l5ee

2ee0

DSM+DRM

2.8137

2.8136

2.8138

2.8137

2.8137

2.8138

2.8138

2.8140

DSM+DSM

2.8137

2.8137

2.8138

2.8137

2.8136

2.8136

2.8137

2.8137

Note: The solution F ,=2.824 (L --

00

is given by reference (7)

Table.2 Nondimensional stress intensity factor for the beam with


an edge crack under pure bending

L/iV

F,

10

20

DSM+DRM

1.5013

1.5012

1.5014 1.5013

1.5013 1.5013

I DSM+DSM

1.51315

1.5015

1.5014 1.51312

1.5012

Note: The solution F,=1.495 (L .....

le0

00

200

S00

1000

1500

200e

1.5010

1.5e15

1.5012 1.5012 1.5012

is given by reference (8)

Summary
In this study. the strategies for accurately solvIng the problems of very
slender body and thin plate are proposed and the various applications are
presented. The results obtained in this study are summarized in Fig.13 and
Fig. 14 as compared with using the usual BEM analysis. It is found that the
problem of the infinite plate with a slender hole and the problem of a curved
beam or a curved shell can be also accurately solved by using presented
strategies. As the conclusions. the strategies proposed in this paper for the
highly-accurate analysis of very slender body and thin plate are summarized
as follows:
1). The accuracy improvement of the numerical integral in the quasi-singular

integral between the upper plane and lower plane as well as the edge of
the slender beam and thin plate is very important and the sub-element
division scheme is useful for this purpose.

411

2-D analysis

-I
N

Or dinary BEM

hl--

L/h

hi )

L/ h

hI *p

L/h

~ MI

.... P

__ I

hl--

~ Mr===v==JiJ

[&-J

b/h

L/h

200
200
10
10

L/h

b/a

Acc ur ate BEM

L/h -

>200 0

L/h >2000
L/h -

1000

b/h -

1000

L/h
25

>2000

L/h
25

> 2000

b/a

o. 01

0. 001

Fig .13 Lim! ts of the aspect ratio in highly-accurate BEM analysis as


compared with usual BEM analysis for two dimensional problems
3-D analy sis

.:-~

(I
I

b ~g~)

t(QP-d:hat

Ord i nary BEM

L/h

L/h

L/h

;~,

,
'.:...
I

r i/!

Accurate BEM

L/h -

lOa

100 0

- so

L/h -

L/h -

125

15

- SO

200

r i/t -

2000

It
I

Fig.14 Limits of the aspect ratio in highly-accurate BEM analysis as


compared with usual BEM analysis for three dimensional problems

412

2). The slender elements in the 3-D BEM analysis of the thin plate or shell
usually exist and the sub-element division scheme Is also used
successfully to improve the accuracy of the numerical integral In the
slender elements.
3).The fixed end condition of the cantilever beam and plate Is correctly
treated by using DRM.
4).For more complicated problem such as the slender body and thin plate with
a crack or notch, the sub-region division method Is useful and the
highly-accurate solution can be obtained.

References
LYuuki,R.;Kisu,H.: Elastostatic Analysis by Boundary Element Method (1987),
Baifukan.
2.Yuuki,R.;Matsumoto,T.;Kisu,H.: The usefulness and limit of the direct
regular method In boundary element elastostatlc analysis. JSME Int.J.
30-260(1987), 227-233.
3.Yuuki,R.;Cao,G.Q.; Accurate BEM elastostatic analysis for very slender
body, Proc. Japan National Sympo. on Boundary Element Method, JASCOME
5(1988), 155-160
4.Yuuki,R.;Maekawa,K.;Cao,G.Q.; Highly accurate BEM elastostatic analysis
of thin plate, Trans. JSME, 57-543A(l991).
5.Dym,C.;Shames,I.; Solid Mechanics: A Variational Approach(l973), McGrawHill.
6.Cao,G.Q.: Ph.D Thesis, University of Tokyo.
7.Nlsltani,H.;Chen,D.H.: Body Force Method(l987), Baifukan.
8.Brown,W.F., et al.: Plane strain crack toughness testing of high strength
metallic materials, ASTM STP, 410(1966), 1-65.

You might also like