(Boundary Elements 3) George T. Symm (auth.), Dr. Carlos A. Brebbia (eds.) - Boundary Element Methods_ Proceedings of the Third International Seminar, Irvine, California, July 1981-Springer-Verlag Ber.pdf

Boundary
Element Methods
Proceedings of the Third International Seminar,
Irvine, California, July 1981
Editor: C. A. Brebbia
Seminar sponsored by the International Society for Computational

Methods in Engineering
With 232 Figures
CML PUBLICATIONS
Springer-Verlag Berlin Heidelberg GmbH 1981

Dr. CARLOS A. BREBBIA
Computational Mechanical Centre
125 High Street
Southampton
England, SOl OAA
ISBN 978-3-662-11272-4 ISBN 978-3-662-11270-0 (eBook)

DOI 10.1007/978-3-662-11270-0
This work is subjekt to copyright. All rights are reserved, whether the whole or part of the material is oonoerned,
specifically those of translation, reprinting, re-use of illustrations, broadQlS!ing. reproduction by photocopying machine
or similar means, and storage in data banks. Under § 54 of the German Cqlyright Law where copies are rrnde for
other than private use a fee is payable to 'Verwertungsgesellschaft Wort', Munich.
© Springer-Verlag Berlin Heidelberg 1981

Originally published by Springer-Verlag Berlin Heidelberg New York in 1981
Softcover reprint of the hardcover I st edition 1981
The use of 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 forge-
neral use.
2061/3020 - 543210
v
CONTENTS
Introductory Remarks IX
C.A. Brebbia
SECTION I POTENTIAL AND FLUID FLOI-1 PROBLEMS
Two Methods for Computing the Capacitance of a 3

Quadrilateral
G.T. Symm
Boundary Element Solutions to the Eddy Current Problem 14
S.J. Salon, J.M. Sahneider and S. Uda
Numerical Solution of the Diffusion Equation by a 26
Potential Method
D.A.S. Curran, B.A. Lewis and M. Cross
The Application of Boundary Elements to Steady and 37
Unsteady Potential Fluid Flow Problems in Two and Three
Dimensions
P.H.L. Groenenboom
Computing Strategy in the Integral Equation Solution of 53
Limiting Gravity Waves in Water
J.M. Williams
SECTION II ELASTICITY PROBLEMS 69
On the Construction of the Boundary Integral 71

Representation and Connected Integral Equations for
Homogeneous Problems of Plane Linear Elastostatics
M. Pignole
Regular Boundary Integral Equations for Stress 85
Analysis
C. Patterson and M.A. Sheikh
A Boundary Element Formulation of Problems in Linear 105
Isotropic Elasticity with Body Forces
D.J. Danson
VI
A Comparative Study of Several Boundary Elements in 123

Elasticity
M.F. Seabra Pereira, C.A. Mota Soares and
L.M. Oliveira Faria
Non-Conforming Boundary Elements for Stress Analysis 137
The Displacement Discontinuity Method in Three 153
Dimensions
W. Scott Dunbar and D.L. Anderson
A Unified Second Order Boundary Element Method for 174
Structures Analysis
M. Dubois
Method of Boundary Integral Equations for Analysis of 183
Three Dimensional Crack Problems
J. Balas and J. Sladek
Cyclic Symmetry and Sliding Between Structures by the 206
Boundary Integral Equation Method
A. Chaudouet
On Boundary Integral Equations for Circular 224
Cylindrical Shells
H. Antes
The Boundary Element Method applied to Two- 239
Dimensional Contact Problems with Friction
T. Andersson
Quasistatic Indentation of a Rubber covered Roll by a 259
Rigid Roll - The Boundary Element Solution
R.C. Batra
An Improved Boundary Element Method for Plate 272
Vibrations
G.K.K. Wong and J.R. Hutchinson
SECTION III GEOMECHANICS 291
Boundary Elements applied to Soil-Foundation 293

Interaction
M. Ottenstreuer and G. Schmid
The Implementation of Boundary Element Codes in 310
Geotechnical Engineering
L.A. Wood
Boundary Integral Method for Porous Media 325
M. Predeleanu
VII
SECTION IV MATERIAL PROBLEMS 335
Numerical Analysis of Cyclic Plasticity using the 337

Boundary Integral Equation Method
M. Brunet
New Developments in Elastoplastic Analysis 350
J.C.F. Telles and C.A. Brebbia
The Boundary Element Method for the Solution of 371
No-Tension Materials
W.S. Venturini and C. Brebbia
SECTION V NUMERICAL TECHNIQUES AND MATHEMATICAL 397
PRINCIPLES
Some Theoretical Aspects of Boundary Integral 399

Equations
M.A. Jaswon
On The Asymptotic Convergence of Boundary Integral 412
Methods
W.L. Wendland
Boundary Methods. C-Complete Systems for the 431
Biharmonic Equations
H. Gourgeon and I. Herrera
The Effect of Mesh Refinement in the Boundary Element 442
Solution of Laplace's Equation with Singularities
H.L.G. Pina, J.L.M. Fernandes and C.A. Brebbia
Boundary Element and Linear Programming Method in 457
Optimization of Partial Differential Equation
Systems
T. Futagami
Approximate Fundamental Solutions and Alternative 472
Formulations
S. Walker
An Efficient Algorithm for the Numerical Evaluation 489
of Boundary Integral Equations
M. Vahle and D.L. Sikarskie
Solution of the Dirichlet Problem using the Reduction 504
to Fredholm Integral Equations
J. Caldwell
BEMSTAT -A New Type of Boundary Element Program for 518
Two-Dimensional Elasticity Problems
L. Bolteus and 0. Tullberg
VIII
SECTION VI COUPLING OF BOUNDARY AND FINITE ELEMENT 539

METHODS
Interfacing Finite Element and Boundary Element 541

Discretizations
C.A. Felippa
The Derivation of Stiffness Matrices from Integral 552
Equations
F. Hartmann
Three Dimensional Super-Element by the Boundary 567
Integral Equation Method for Elastostatics
F. Volait
The Coupling of Boundary and Finite Element Methods 575
for Infinite Domain Problems in Elasto-Plasticity
G. Beer and J.L. Meek
A Finite Element - Boundary Integral Scheme to 592
Simulate Rock-Effects on the Liner of an Underground
Intersection
B.A. Dendrou and S.A. Dendrou
The Use of Green's Functions in the Numerical Analysis 609
of Potential, Elastic and Plate Bending Problems
C. Katz
IX
INTRODUCTORY REMARKS
C.A. Brebbia
I. HISTORICAL NOTE
The great interest that finite element methods have

attracted in the engineering community since the beginning of
1960's has had two important consequences. I) it stimulated
an impressive amount of work in computational techniques and
efficient engineering software; 2) substantial research into
basic physical principles such as variational techniques,
weighted residuals, etc. was originated.
The first of the above points came as a natural conse-

quence of the emergence of new and powerful computers which
were able to solve engineering problems involving large
amounts of numerical storage and manipulation. The development
of mathematical techniques and basic principles came instead
in response to the need of extending finite element modelling,
assessing convergence and accuracy and understanding the
relationship of finite elements to more classical variational
principles and weighted residual techniques.
These techniques can be traced to pre-computer times [1 ,2]

and involve different ways of solving the governing equations
of a problem, i.e, Galerkin, collocation, least-squares, line
techniques, matrix progression or transfer, the combination of
different techniques, etc. Fortunately they were not forgotten
and they reappeared in the finite element literature, sometimes
with different names, such as Galerkin finite elements, finite
element strip-method, some time integration schemes, etc.
Another important development iu approximate analysis was the
investigation of mixed principles and the realization that
physical problems c&< be expressed and solved in many different
ways in accordance with the part or equations of the problem
that we need to approximate. These approximations are of
fundamental importance for the computer implementation of the
different numerical techniques. Mixed methods can be traced
to Reissner [3] and more specifically for finite elements to
X
Pian [4]. An excellent exposition of mixed methods in

structural mechanics can be found in the book by Washizu [s].
[ntegral equation techniques were until recently considered
to be a different type of analytical method, somewhat unrelated
to approximate methods. They became popular in Westem Ettrope
through the work of a series of Russian authors, such as
Muskhelishvili [6], Mikhlin [7], Kupradze [8], and Smi mov [9]
but were not very popular with engineers. A predecessor of
some of this work was Kellogg [10] who applied integral equations
for the solution of Laplace's type problems. Integral equation
techniques were mainly used in fluid mechanics and general
potential problems and known as the ' source' method which is
an ' indirect' method of analysis; i.e. the unknowns are not
the physical variables of the problem. Work on this method
continued throughout the 1960s and 1970s in the pioneering work
of Jaswon ITU and Symm [12], Massonet [13], Hess [14] and many
others.
It is difficult to point out precisely who was the first

one to propose the 'direct' method of anal~sis. It is found
in a different form, in Kupradze' s book IB
J • It seems fair
however from the engineering point of view to consider that
the method originated in the work of Cruse and Rizzo [Is] in
elastostatics. ·
Since the early 1960s a small research group started

working at Southampton University, on the applications of
integral e!}uations to solve stress analysis problems. Hadid's
thesis [l6j partly based on Hadjin's [!7] work was published
in 1964 and dealt with the use of integral equations in shell
analysis. Unfortunately the presentation of the problem, the
difficulty of defining the appropriate Green's functions and
the parallel emergence of the finite element method all con-
tributed to minimize the importance of this work. This work
was continued through a series of theses dealing mainly with
elastostatics problems. Recent developments in finite elements
had started to find their way into the formulation of boundary
integral equations, specially in the idea of using general
curved elements. Finally the question of how to effectively
relate the boundary integral equations to other ap~roximate
techniques was solved using weighted residuals [18J • The
work at the Southampton University group culminated with the
first book in 1978 for which the title "Boundary Elements" was
used Qe].
More recently this work has been e~anded to encompass

time dependent and non-linear problems L20]. Two important
International Conferences were held at Southampton University
in 19 78 and 1980. The edited Proceedings of these conferences
- the only ones so far on this topic - are now standard
references [21 ,22].
XI
One of the most important recent applications of boundary

elements has been for the solution of time dependent problems.
This has now opened the way for solving more complex problems,
such as wave propagation, Navier-Stokes flow, elastodynamics,etc.
The first boundary integral solution for the diffusion equation
was due to Rizzo and Shippy [23], who used the Laplace transform
to remove the time dependence. This approach was discussed by
Liggett and Liu f24] where some of its disadvantages are pointed
out. Chang et al-: [zs]in 1973 solved the problem in the time
domain using space and time piecewise constant values for the
variables .. This approach has been extended by Wrobel and
Brebbia [~6] who applied it to solve complex temperature
problems. Brebbia and Walker ljo] proposed a step by step
finite difference scheme for the time derivatives. The use of
a time dependent fundamental solution implies that one does
not need to perform any finite differences on the time deriva-
tive and produces an accurate and efficient solution specially
when hi~er order time interpolation functions are employed
[27]' [f8J.
The first applications of boundary elements in elasto-
dynamics can be traced to Rizzo and Cruse [30], [31] and the
technique is now widely accepted for solving soil-structure
interaction problems. Cruse's work has been of fundamental
importance for the development of boundary elements as he has
contributed to its application not only in elasto~tatics and
elas todynami cs but also in fracture mechanics [31] where
boundary elements are much more accurate than finite elements.
The boundary element method generally offers the advant-

age over more classical finite elements, of working with
smaller systems of equations and obtaining more accurate
results. Because of these advantages the technique is of
interest for solving a range of non-linear problems as well as
linear ones. The first formulation of boundary elements for
plasticity was presented in 1971 by Swendlow and Cruse [32]
in a purely theoretical paper which laid the foundations for
all subsequent work. However, they did not present the
expressions for initial stresses or strains which are needed
in stepwise plasticity analysis. These expressions were
presented by Mendelson [33] in 1973 for two- and three-
dimensional bodies although the plane strain formulation was
correctly presented by Mukherjee [}4] only in 1977. None of
the above authors gave the proper expressions for strains and
stresses at internal points. It was Bui I3s] who in 1978
publishad the correct expressions for strains and stresses for
three dimensional bodies and finally in 1979, Telles and
Brebbia [36] presented the complete formulation for boundary
elements applied to two- and three-dimensional problems.
Applications relevant to engineering practice are presented
in reference [37} and more can be found in this book.
XII
It may be convenient in many applications to combine

boundary elements with finite elements, as some subregions of
the problem may be more amenable to solution using the latter.
This may be achieved in either of two ways: (I) by considering
the whole problem using an equivalent BEM approach or (2) con-
verting the BEM subregions into equivalent FEM. The two
~proaches are described in detail in Brebbia and Georgiou
L38]. As the FEM is very well established, with computer codes
readily available for its implementation, the consideration of
the BE subregion as an additional equivalent finite element
matrix seems most attractive. The formulation of this
'equivalent' matrix used to model the BE regions presents
however certain problems; at sharp geometric discontinuities
there are also discontinuities of surface tractions which
require special attention and the equivalent FE matrix formed
is not inherently symmetric, unlike the classical FE approach.
A technique which overcomes these problems and provides an
acceptable FE type formulation using the BEM method has been
suggested by Georgiou and Brebbia [4o]. The technique is
based on an idea advanced by Chaudonneret [39] in 1978.
Ways of s~trizing the BE matrices have been suggested

by Mustoe et al. [4Ij, Margulies [42] and others and are
thoroughly discussed in the present book.
2. REVIEW OF THE PAPERS PRESENTED
The 3rd International Conference on Boundary Element Methods

held in California during July 1981 constituted a forum for
the presentation of the most advanced research in BEM. When
comparing the Proceedings - published in this book - with
those of previous meetings it is evident that rapid advances
have been made in important fields. Regarding applications of
BEM, this book contains several important papers dealing with
time dependent and vibrations problems, fluid flow and
material non-linearities. Several types of non-linearity are
discussed, i.e. elastoplasticity, viscoelasticity and no-
tension type materials. A substantial number of papers deals
with stress analysis, including treatment of body forces,
higher order models, crack problems, cyclic symmetry and
sliding and other types of conditions. One of the papers
even extended the BEM to deal with shell analysis opening a
prom1s1ng new area of research. Geomechanical applications
are also well represented and for these problems the advantages
of using BEM which can include radiation as well as material
damping are obvious.
One of the most interesting set of papeiS is that dealing

with the systematic formulation of boundary element method
and mathematical properties such as accuracy, convergence
versus order and number of elements. The topic of numerical
integration is discussed by several authors and some promising
XIII
ideas have been put forward.
Many papers deal with the coupling of BE and FE solutions.

This is a particularly important field in view of the large
number of available FE codes and the convenience of combi~ing
both techniques in many practical applications.
Several of the papers presented here refer to potential

and fluid flow problems. Symm ]2.3] discusses different ways of
finding the capacitance of a quadrilateral based on the direct
boundary element formulation. The boundary element solution
to the eddy current problem ~4] produces a Helmholtz type
equation which can be transformed into integral equations.
The problem is assumed to be steady state and the results
compared against the iinite element solution. The BEM is
capable of achieving high accuracy with fewer equations than
the FEM.
The paper by Curran et al. [45] discusses the solution of

time dependent diffusion problem using integration on time
as well as in space. Although the method is only applied to
one dimensional space dependent problems, the authors conclude
that it can be expanded to solve two and three dimensional
problems as well. Groenenboom [46] explains in detail how
the boundary element in time and space can be applied to solve
the wave equation. He discusses the problem of radiating
boundary conditions and presents results for a three dimension-
al rectangular box. His work appears to be the first
successful attempt to use BEM to solve the transient wave
propagation problem using time as well as space dependent
fundamental solutions. Williams [4'B uses an integral
equation technique to solve finite-amplitude waves and
presents several computational techniques to determine the
unknown position of the free surface.
Pignole [4~ dealing with linear elastic problems, dis-

cusses how direct and indirect methods relate to each other.
The basic formulation of the BEM is also discussed in the
paper by Patterson and Sheikh [49] , who proposes moving the
singularities to points outside the domain of the problem.
In this way the resulting system of equations tolerates higher
order singularities in the solution. A series of numerical
results demonstrates that the technique gives accurate results.
The paper by Danson [So] describes in detail, and in a
consistent way, how b-ody forces can be taken to the boundary.
The work is particularly important for the efficient imple-
mentation of stress boundary element programs and different
types of body forces are discussed. The problem of comparison
between different BEM and FEM solutions for practical problems
is discussed by Seabra Pereira et al. [51]. The authors present
the computer performance of different BEM models including
CPU times using a general purpose BEM program which has
ccnstant, linear and parabolic 2-D and 3-D elements. The
XIV
advantages of using BEM against FEM are also pointed out and
the use of parabolic BEM for stress concentration problems is
advocated. Patterson and Sheikh lszl propose using non-
conforming BEM other than constants, to avoid the problem of
defining the normal at corner points or having errors propaga-
ting into the system due to a change in the nature of the
boundary conditions. They conclude that interelement continuity
is not necessary with BEM and demonstrate this through several
numerical examples.
Another integral formulation called by the authors the

displacement discontinuity method (DDM) is presented by Dunbar
and Anderson [}3]. The DDM is a type of boundary integral
method wherein the fundamental solution used is the static dis-
placement discontinuity on a finite segment in an infinite or
semi-infinite medium. The technique is explained in detail in
the paper together with its possible linkage with finite
elements. Dubois (;i4] presents a generalized way of obtaining
second order BE models for structural analysis. The method
relates displacements, ~otations, stresses and couple-stresses
and has the advantage of ensuring complete compatibility in
complex structural analysis.
The important field of using BEM to solve crack problems

is discussed in paper (}5]. Several element types are dis-
cussed and the BEM is found to be an efficient and accurate
tool to solve the problem. Chaudouet rs6] presents a way to
include cyclic symmetry and sliding betWeen structures in BEM
codes. The paper presents some interesting industrial appli-
cations.
Antes [57] extends the BEM for the solution of circular

cylindrical -shells. He develop!! a new reciprocal theorem and
deduces an original boundary element formulation in what
appears to be the first application of BEM to shell analysis.
BEM has been particularly successful to study contact

problems for which tney offer over the advantage over FEM of
easily and accurately coupling normal and tangential tractions.
The paper by Andersson [sa] describes how BEM can be applied
to two-dimensional contact problems with friction. Several
BEM examples are presented and results compared against other
solutions. The problem of indentation of a rubberlike material
bonded to a rigid cylinder and indented by another rigid
cylinder is studied by Batra ]~9] using BEM. Results are
presented for the pressure distribution over the contact
surface, the shape of the indented surface and the stress dis-
tribution at the bond surface.
Wong and Hutchinson j}o] use BEM to determine the natural

frequencies and mode shapes of thin elastic plates of uniform
thickness with arbitrary boundary conditions and arbitrarily
shaped edges. The complete theory is presented in the paper
XV
but numerical results are not givEn as the computer code is

not yet fully operational.
Part III of this book .deals with geomechanic problems.

Ottenstreuer and Schmid f6Il apply BEM to study soil-
foundation interaction to- obtain a better representation of
radiation damping. The soil is assumed to be an elastic
homogeneous space and the classical fundamental solution for
the steady state case is used. Material damping behaviour is
taken into account by introducin..z A complex modulus of
elasticity. The paper by Wood ~2J places particular emphasis
upon the coupling of FEM to soil BEM to solve soil-structure
interaction problems. He describes two computer codes which
incorporate BEM in the areas of raft foundation and pile/wall
analysis and design. Several interesting practical applica-
tions are described. Predeleanu 1§31 appears to be the first
to apply Biot consolidation theory to BEM. The theory is
extended to the general class of viscoelastic bodies defined
by convolutions. The fundamental singular solution is similar
to the one used in the coupled theory of thermoviscoelasticity.
BEM is increasingly becoming an accepted method of solu-

tion for non-linear material problems. The paper by Brunet 1§4]
deals with its application to cyclic plasticity and presents
a computational procedure for the solution of such classes of
non-linear problems employing different types of constitutive
relations. This work is closely related to the paper by Telles
and Brebbia ~5] which presents further developments on the
application of BEM to plasticity, namely the extension of the
theory to deal with half plane problems. They also show a
series of examples of practical importance. The other paper
in this section of the book written by Venturini and Brebbia
[6~ deals with the extension of BEM to no-tension materials
such as those present in some applications in geomechanics.
The no-tension solution is achieved using an iterative process
which consists of applying at each step a series of initial
stresses to compensate the tensile stresses, much in the same
way as it is done with the FEM. Finally some problems are
solved to illustrate the application of BEM to no-tension
materials. The examples were selected to compare the BEM
solution with already published finite element results and
point out the advantage of using the former.
Several important papers are presented in the section

dealing with numerical techniques and mathematical principles.
Jaswon f67] points out how the scalar and vector field theories
used in-potential and elastostatics boundary elements can be
unified and the analogy between the Green's formula of
potential theorY and Somigliana 's formula of elasticity.
Wendland's ~8] paper discusses the important topic of con-
vergence of boundary integral methods and finds an expression
for the asymptotic error estimate. The work by Gourgeon and
XVI
Herrera [69] tries to establish a systematic approach to the

theoretical foundation of the boundary element method with
special reference to the bi-harmonic equation. The problem of
convergence of the boundary element solution is again discussed
in the paper by Pina, Fernandes and Brebbia [70]. A series of
numerical tests are carried out using linear and parabolic
elements for two different problems and the BEM results are
compared against several FEM solutions. Several meshes are
tried to assess the order of convergence and the effect of
mesh grading on the magnitude and order of error convergence
is also assessed.
The work of Futagami [7!] presents the first attempt to

combine the BEM with linear programming to develop a new
optimization technique. The method is applied to optimal con-
trol problems in heat conduction phenomena. He recommends
the extension of the technique to transient problems and
developing stochastic boundary elements as well as non-linear
programming. The paper by Walker [72] discusses the use of
approximate fundamental solutions and alternative formulations
He advocates using reference points outside the domain to
eliminate having to evaluate singular integrals. Walker also
discusses the effect of taking approximations to the fundament-
al solutions. It is interesting to notice that all his
numerical work was carried out using a microprocessor. The
problem of numerical evaluation of the boundary integral
equations is addressed again by Vahle and Sikarskie [?3] who
developed a special algorithm in which the unknown vector is
represented as a combination of a Fourier series and piece
wise linear function. The authors point out the numerical
advantages of their technique which requires much less solution
time and could be very efficient for certain non-linear
problems. The paper by Caldwell [?4] discusses the difficulties
associated with the numerical solution of integral equations
and their relationship with Gauss quadrature formulae. This
is of course an important topic in BEM research. A new type
of BE program for two dimensional elasticity is described by
Bol teus and Tull berg in reference [?s] . They allow for dis-
continuity in the tractions and propose a method for coupling
BE and FEM solutions, dealing in detail with the integration
needed to find the influence coefficients and producing some
guidelines for 'optimal' integration.
Section VI of these proceedings deal with the topic of

coupling the BEM and FEM. This is a topic of practical
importance in view of the large number of existing FE codes.
The paper by Felippa ~6] reviews critically several tech-
niques for coupling FE and BE solutions. He particularly
assesses three different techniques relevant to his physical
problem, i.e. a three-dimensional structure submerged in an
acoustic field and impinged by a pressure shock wave. He
stresses the usefulness of BE techniques for discretizing
unbounded homogeneous domains governed by linear equations.
XVII
Hartmann [77] reviews in a very clear manner the differ-

ence between BE and FE matrices and the reasons for the lack
of symmetry of the former. He then proposes simple ways of
drawing stiffness matrices from BE and coupling them to FE
solutions. The paper by Volait [7s] describes how stiffness
matrices can be obtained using BE, based on an energy minimiza-
tion technique and introducing possible discontinuity of
tractions simply by choosing nodes for the model tractions
inside the element. The topic of numerical integration is
discussed as well. Beer and Meek [79] propose a coupling of
the two methods to analyse circular excavations in infinite
domain with the region of plasticity confined to the FE zone.
Comparisons are also carried out between coupling FE with BE
and with the so called 'Infinite' Finite Elements. The paper
by Dendrou and Dendrou [soJ also is concerned with the applica-
tions of FE/BE techniques in geomechanics. Finite elements
are used to describe the liner of an underground excavation
and BE to represent the soil. They conclude that the proposed
methodology offers both accuracy and cost efficiency. The
paper by Katz [81] explains the reasons why it is desirable to
link the FE and BE techniques, review several approaches and
puUl forward the idea of using a method proposed in the Second
International Seminar on Boundary Elements, Southampton, 1980.
The paper also discusses the numerical integration of singular
equations.
3. SUGGESTIONS FOR FURTHER RESEARCH
Several important topics related to boundary solutions or more

specifically the boundary element method require further
investigation. Boundary element method implies an approximate
technique by which the problem dimensions have been reduced by
one or in the case of time dependent problems by more than one.
The problem becomes a boundary problem and the surface of the
domain can be divided into elements over which some approximate
technique can be applied. What differentiates boundary element
of classical integral equations is its emphasis on elements
and compatibility as well as its reliance on general principles
such as weighted residuals. This is important as some classical
boundary integral solutions suffer from lack of accuracy and
convergence due to poor representation of the geometrical
shapes. These problems appear not to have been sufficiently
investigated.
It is also evidentto finite element practitioners that the

degree of complexity assured for the geometry and the one for
the functions representing the boundary variables are inter-
related. We know from variational techniques research that
trying to represent the geometry with higher order functions
while assuring simple functions for the variables will
introduce considerable errors in the solution. This is an
important point which seems to have escaped the attention of
many researchers applying integral equations.
XVIII
Although current boundary element practice is heavily

dependent on using fundamental solutions based on solving the
governing equations under concentrated source and no boundary
conditions, it is of the utmost importance to point out that
other solutions can be used as well. These other solutions
are illustrated by those used in such methods as hybrid finite
elements and trigonometric type solutions. Analytical
functions which have been obtained over a close domain with
certain boundary conditions could also be used as fundamental
solutions provided that our domain of interest is within the
domain of the analytical solutions. This of course can be
easily arranged by a change of dimensions. Once these ideas
are better understood we will be able to solve many complex
plate and shell and other problems using the wealth of solu-
tions already available as approximate fundamental solutions.
The general field of non-linear problem solving using BEM
requires further investigation.
Time dependent problems may be one of the most fruitful

research topics in BEM. The method seems well ·suited to solve
problems in time and space. FEM solutions in time and space
are on the contrary, very uneconomical and have been largely
abandoned. BE is now accepted as a valid and efficient way of
solving parabolic problems but little is known about its
applications to hyperbolic problems such as the transient wave
equation. More research should be carried out to determine
its accuracy and efficiency for these problems.
Finally, BEM are also well suited for the small computers
mainly microprocessors, which are nowadays being used in many
engineering offices. BEM solutions tend to require less
matrices and data storage than FE solutions whiie having a
large number of repetitive operations such as numerical
integration which can be economically carried out by a micro-
computer. BE appears ideally suited for computer aided design
activities due to the smaller amount of data generally required.
This is an important advantage specially when dealing with
three dimensional problems and more emphasis should be put into
developing efficient CAD systems based on BE.
REFERENCES
I. Kantorovich,L.V. and Krylov, V.I., Approximate Methods of

Higher Analysis, Interscience, New York (I958).
2. Courant, R. and Hilbert, D. Methods of Mathematical

Physics, Interscience (.1953).
3. Reissner, E." A Note on Variational Principles in Elasticity

Int. J. Solids and Structures, I, pp.93-95 and 357 (I965).
4. Pian, T.H.H. and Tong, P., Basis of Finite Element Method

for Solid Continua, Int. J. Numerical Methods Engng., I,
pp. 3-28 (I 969) •
XIX
5. Washizu, K., Variational Methods in Elasticity and Plastic-

ity, 2nd edn, Pergamon Press, New York (1975).
6. Muskhelishvili, N.I., Some Basic Problems of the Mathemat-

ical Theory of Elasticity, P. Noordhoff, Ltd., Groningen
(1953).
7. Mikhlin, S.G., Integral Equations, Pergamon Press, New

York (I 95 7) •
8. Kupradze, O.D. Potential Methods in the Theory of

Elasticity, Daniel Davey & Co., New York (1965).
9. Smirnov, V.J. Integral Equations and Partial Differential

Equations, in A Course in Higher Mathematics, Vol. IV,
Addison-wesley (1964).
10. Kellogg, P.D.,Foundations of Potential Theory, Dover,

New York ( 1953).
II. Jaswon, M.A., Integral Equation Methods in Potential

Theory, I, Proc. R. Soc., Ser. A, p.273 (1963).
12. Symm, G.T., Integral Equation Methods in Potential Theory,

II, Proc. R. Soc., Ser. A, p.275 (1963).
13. Massonnet, C.E., Numerical Use of Integral Procedures in

Stress Analysis, Stress Analysis, Zienkiewicz, O.C. and
Holister, G.S. (eds), Wiley (1966).
14. Hess, J.L. and Smith, A.M.O., Calculation of Potential

Flow about Arbitrary Bodies, Progress in Aeronautical
Sciences, Vol.8,Kuchemann, D. (Ed), Pergamon Press (1967).
15. Cruse, T.A. and Rizzo, F.J., A Direct Formulation and

Numerical Solution of the General Transient Elasto-Dynamic
Problem, I, J. Math. Analysis Applic., 22 (1968).
16. Hadid, H.A., An Analytical and Experimental Investigation

into the Bending Theory of Elastic Conoidal Shells, Ph.D.
Thesis, University of Southampton (1964).
17. Hajdin, N., A Method for Numerical Solutions of Boundary

Value Problems and its Application to Certain Problems in
the Theory of Elasticity, Belgrade University Publication
( 1958).
18. Brebbia, C.A., Weighted Residual Classification of

Approximate Methods, Applied Mathematical Modelling
Journal, Vol.2, No.3, September (1978).
19. Brebbia, C.A., The Boundary Element Method for Engineers,

Pentech Press, London, Halstead, New York, (1978).
XX
20. Brebbia, C.A. and S. Walker, Boundary Element Techniques

in Engineering, Butterworths, London,(I979).
21. Brebbia, C.A. (Ed.), Recent Advances in Boundary Element

Methods, Proc. 1st Int. Conference Boundary Element
Methods, Southampton University, 1978, Pentech Press,
London ( 1978).
22. Brebbia, C.A. (Ed.), New Developments in Boundary Element

Methods, Proc. 2nd Int. Conference Boundary Element
Methods, Southampton University, 1980, CML Publications,
Southampton (1980).
23. Rizzo, F.J. and Shippy, D.J. , A Method of Solution for

Certain Problems of Transient Heat Conduction, AIAA J.,
8, No. II, pp.2004-2009 (1970).
24. Liggett, J.A. and Liu, P.L., Unsteady Flow in Confined

Aquifers: A Comparison of Two Boundary Integral Methods,
Water Resources Res. 15, No. 4, pp.861-866 (1979).
25. Chang, Y.P., Kang, C.S. and Chen, D.J., The Use of Funda-
,~entalGreen's Function for the Solution of Problems of
~at Conduction in Anisotropic Media, Int. J. Heat Mass
Tr:msfer, 16, pp.I905-1918 (1973).
26. Wrobel, L.C. and Brebbia, C., The Boundary Element Method
for Steady-State and Transient Heat Conduction, 1st Int.
Conf. Numerical Methods in Thermal Problems, Swansea,
Taylor, C. and Morgan, K. (Eds), Pine ridge Press (1979).
27. Wrobel, L.C. and Brebbia, C., A Formulation of the

Boundary Element Method for Axisymmetric Transient Heat
Conduction, Int. J. Heat Mass Transfer,
28. Wrobel, L.C. and Brebbia, c., .. Boundary Elements in

Thermal Problems, Numerical Methods in Heat Transfer,
Lewis, R.W. (ed.), Wiley, Chichester, England (1981).

Numerical Solution of the General Elastodynamic Problem',
I, J. Math. Analysis Applic., 22 (1968).

Numerical Solution of the General Transient Elastodynamic
Problem, II, J. Math. Analysis Applic., 22 (1968).
31. Cruse, T.A. and Vanduren, W., Three Dimensional Elastic

Stress Analysis of a Fracture Specimen with Edge Crack,
Int. J. Fract. Mech., 7, pp.I-15 (1971).
XXI
32. Swedlow, J.L. and Cruse, T.A., Formulation of Boundary

Integral Equations for Three-Dimensional Elasto-Plastic
Flow, Int. J. Solids Struct., 7, 1673-1683, 1971.
33. Mendelson, A., Boundary Integral Methods in Elasticity

and Plasticity, Report No. NASA TN D-7418 (1973).
34. Mukherjee, S., Corrected Boundary Integral Equation in

Planar Thermoelasticity, Int. J. Solids Struct., 13,
pp.331-335 (1977).
35. Bui, H.D., Some Remarks about the Formulation of Three-

Dimensional Thermoelastoplastic Problems by Integral
Equations, Int. J. Solids Struct., No.l4, pp.935-939
(1978).
36. Telles, J.C.F. and Brebbia, C.A., On the Application of

the Boundary Element Method to Plasticity, Appl. Math.
Modelling, 3, pp.446-470 (1979).
37. Telles, J.C.F. and Brebbia, C.A., The Boundary Element

Method in Plasticity, in New Developments in Boundary
Element Methods, Brebbia, C.A. (Ed.), CML Publications,
England (1980).
38. Brebbia, C.A. and Georgiou, P., Combination of Boundary

and Finite Elements in Elastostatics, Applied Math.
Modelling, 3. pp.213-220 (1978).
39. Chaudonneret, fvl., On the Discontinuity of the Stress

Vector in the Boundary Integral Equation Method for
Elastic Analysis, Recent Advances in Boundary Element
Methods, Brebbia, C.A. (Ed.), Pentech Press, London (1978).
40. Georgiou, P. and Brebbia, C., On the Combination of

Boundary and Finite Element Solution, Applied Math.
Modelling, To be published.
41. Mustoe, G., A Symmetric Direct Boundary Integral Equation

Method for Two-Dimensional Elastostatics, Paper presented
at the 2nd Int. Seminar on Boundary Element Methods,
Southampton University, March (1980).
42. Margulies, M., Exact Treatment of the Exterior Problem

in the Combined FEM/BEM, Paper presented at the 2nd Int.
Seminar on Boundary Element Methods, Southampton University,
March, 1980, (Ed. C. Brebbia), CML Publications (1980).
43. Symm, G.T., Two Methods for Computing the Capacitance of

a Quadrilateral, in this book.
44. Salon, S.J., J.M. Schneider and S. Uda, Boundary Element

Solutions to the Eddy Current Problems, in this book.
XXII
45. Curran, D.A.S., B.A. Lewis and M. Cross, Numerical

Solution of the Diffusion Equation by a Potential Method,
in this book.
46. Groenenbook, P.H.L., The Application of Boundary Elements

to Steady and Unsteady Potential Fluid Flow Problems in
Two and Three Dimensions, in this book.
47. Williams, J.M., Computing Strategy in the Integral

Equation Solution of Limiting Gravity Waves in Water,
in this book.
48. Pignole, M., On the Construction of the Boundary Integral

Representation and Connected Integral Equations for
Homogeneous Problems of Plane Linear Elastostatics, in
this book.
49. Patterson, C. and M.A. Sheikh, Regular Boundary Integral

Equations for Stress Analysis, in this book.
50. Danson, D., A Boundary Element Formulation of Problems in

Linear Isotropic Elasticity with Body Forces, in this
book.
51. Seabra Pereira, M.F., C.A. Mota Soares and L.M. Oliveira
Faria, A Comparative Study of Several Boundary Elements
in Elasticity, in this book.
52. Patterson, C. and M.A. Sheikh, Non-conforming Boundary

Elements for Stress Analysis, in this book.
53. Dunbar, W.S. and D.L. Anderson, The Displacement Discon-

tinuity Method in Three Dimensions, in this book.
54. Dubois, M., A Unified Second Order Boundary Element

Method for Structures Analysis, in this book.
55. Balas, J. and J. Sladek, Method of Boundary Integral

Equations for Analysis of Three Dimensional Crack
Problems, in this book.
56. Chaudouet, A., Cyclic Symmetry and Sliding between

Structures by the Boundary Integral Equation Method,
in this book.
57. Antes, H., On Boundary Integral Equations for Circular

Cylindrical Shells, in this book.
58. Andersson, T., The Boundary Element Method applied to

Two-Dimensional Contact Problems with Friction, in this
book.
XXIII
59. Batra, R.C., Quasistatic Indentation of a Rubber Covered

Roll by a Rigid Roll - The Boundary Element Solution,
in this book.
60. Wong, G,K.K. and J.R. Hutchinson, An Improved Boundary

Element Method for Plate Vibrations, in this book.
61. Ottenstreuer, M. and G. Schmid, Boundary Elements Applied

to Soil-Foundation Interaction, in this book.
62. Wood, L.A., The Implementation of Boundary Element Codes

in Geotechnical Engineering, in this book.
63. Predeleanu, M., Boundary Integral Method for Porous Media,

in this book.
64. Brunet, M., Numerical Analysis of Cyclic Plasticity

using the Boundary Integral Equation Method, in this book.
65. Telles, J.C.F. and C.A. Brebbia, New Developments in

Elastoplastic Analysis, in this book.
66. Venturini, W.S. and C.A. Brebbia, The Boundary Element

Method for the Solution of No-tension Material, in this
book.
67. Jaswon, M.A., Some Theoretical Aspects of Boundary Integral

Equations, in this book.
68. Wendland, W.L., On the Asymptotic Convergence of Boundary

Integral Methods, in this book.
69. Gourgeon, H. and I. Herrera, Boundary Methods. C-complete

Systems for the Biharmonic Equations, in this book.
70. Pina, H.L.G., J.L.M. Fernandes and C.A. Brebbia, The

Effect of Mesh Refinement in the Boundary Element Solution
of Laplace's Equation with Singularities, in this book.
71. Futagami, T., Boundary Element and Linear Programming

Method in Optimization of Partial Differential Systems,
in this book.
72. Walker, S., Approximate Fundamental Solutions and

Alternative Formulations, in this book.
73. Vable, M. and D.L. Sikarskie, An Efficient Algorithm for

the Numerical Evaluation of Boundary Integral Equations,
in this book.
74. Caldwell, J., Solution of the Dirichlet Problem using

the reduction to Fredholm Integral Equations, in this
book.
XXIV
75. Bolteus, L. and 0. Tullberg, BEMSTAT -A New Type of

Boundary Element Program for Two-Dimensional Elasticity
Problems, in this book.
76. Felippa, C.A., Interfacing Finite Elements and Boundary

Element Discretizations, in this book.
77. Hartmann, F., The Derivation of Stiffness Matrices from

Integral Equations, in this book.
78. Volait, F., Three Dimensional Super-Element by the

Boundary Integral Equation Method for Elastostatics,
in this book.
79. Beer, G. and J.L. Meek, The Coupling of Boundary and

Finite Element Methods for Infinite Domain Problems in
Elasto-Plasticity, in this book.
80. Dendrou, B.A. and S.A. Dendrou, A Finite Element-Boundary

Integral Scheme to Simulate Rock Effects on the Liner of
an Underground Intersection , in this book.
81. Katz, C., The Use of Green's Function in the Numerical

Analysis of Potential, Elastic and Plate Bending Problems,
in this book.
SECTIONS I - VI
Section I
Potential and Fluid Flow
Problems
3
TWO METHODS FOR COMPUTING THE CAPACITANCE OF A QUADRILATERAL

George T. Symm
National Physical Laboratory, Teddington, Middlesex, U.K.
ABSTRACT
The problem of finding the capacitance of a (curvilinear)

quadrilateral is formulated in two distinct ways, each
involving the solution of a boundary value problem for
Laplace's equation. In each case the boundary value problem is
solved by the direct boundary element method based upon Green's
third identity. Some numerical results are presented.
INTRODUCTION
Given four points A, B, C, D on a closed contour L bounding a

simply-connected plane domain G, the capacitance K between AB
and CD may be defined (Campbell, 1975, p.27) by
K = ~~ J 0'(q)dq, (1)
AB
where 6 satisfies Laplace's equation
(2)
in the domain G subject to the boundary conditions
6 =1 on CD, 0 = 0 on DA, AB and BC, (3)
q is a vector variable specifying a point on L (in this case on

the segment AB), dq is the differential increment of L at q and
the prime denotes differentiation along the normal to the
boundary L directed into the domain G at the point q. More
precisely, K represents the direct capacitance per unit length,
between the faces AB and CD, of an infinite cylinder, of
cross-section G (Lampard, 1957). Solution of the boundary-value
problem defined by Equations (2) and (3) by the boundary
4
element (integral equation) method based upon Green's third

identity yields ~· on L directly, whence K may be determined
from Equation (1) by integration. We shall refer to this
procedure as Method 1.
Alternatively (Gaier, 1979)' the capacitance between AB and

CD is given by
-1
K = K(M) =g
2 2 [n sinh(mTM)] , (4)
1T n=1 , 3, 5 •••
where M is the conformal module of the quadrilateral ABCD,

defined by
M = b/a (5)
when the domain G is mapped conformally onto the rectangle R

with corners 0, a, a+ib, ib (a>O, b>O) in such a way that AB
and CD correspond to the horizontal sides of R. In this case,
defining
m = a/b, (6)
so that
M = 1/m, (7)
we observe, from Gaier (1972), that m minimises the Dirichlet

integral
(8)
which is invariant under conformal transformation, over the set

of functions 6, continuous in G and piecewise smooth in G, such
that ~ = 0 on AB and ~ = 1 on CD. Indeed, by Dirichlet's
Principle (Courant, 1950),
m = DG(~), (9)
where 6 satisfies Laplace's equation (2) subject to the
boundary conditions
~ = 0 on AB, 0 = 1 on CD and ~· = 0 on BC and DA. (10)
By the divergence theorem, Dirichlet's integral (8) may be

written, in our notation, as
( 11 )
whence, from Equation (9),

5
m =-feD o'(q)dq, ( 12)
where 0 satisfies the boundary-value problem defined by

Equations (2) and (10). Solution of this boundary-value
problem by means of Green's third identity again yields 0' on L
directly, whence m may be determined from Equation (12) by
integration; M follows immediately from Equation (7), whence
the capacitance K may be determined from Equation (4). We
shall refer to this procedure as Method 2.
It is clear from Equation (7) that m is the conformal

module of the "conjugate" quadrilateral BCDA. It follows that
if we solve Laplace's equation (2) subject to the boundary
conditions
0 =0 on BC, 0 1 on DA and 0' = 0 on AB and CD. (13)
then
M =-IDA 0'(q)dq, ( 14)
whence the capacitance K may again be determined from

Equation (4). This provides an alternative form of Method 2,
and a further alternative to Method 1, to which we shall refer
as Method 2a.
We note at this point that the capacitance between AB and CD

is the same as the capacitance between CD and AB, so that each
of the methods above may also be applied to the rotated
quadrilateral CDAB. In general, therefore, we have a total of
six possible ways of computing the capacitance K. We shall see
that the optimal choice between these six formulations depends,
in general, upon the geometry of the contour Land the
positions of the four points A, B, C, D.
INTEGRAL EQUATION FORMULATION AND SOLUTION
Given that 0 satisfies Laplace's equation (2) in the domain G,

Green's third identity may be written at a point p on the
boundary L in the form
I 0'(q)loglq-pldq- J 0(q)log'lq-pldq = ~0(p), (15)

L L
whenever the boundary L is smooth at the point p (Jaswon and

Symm, 1977). Thus, if 0 is prescribed everywhere on L, as by
conditions (3) in Method 1, Equation (15) becomes an integral
equation of the first kind (with a weakly singular kernel) for
the boundary normal derivative 0'. Similarly, if 0 is
prescribed on part of the boundary and 0' on the remainder, as
by conditions (10) in Method 2 (or conditions (13) in Method
6
2a), Equation (15) yields a pair of coupled integral equations

for the complementary unknowns, 0 where ~, is given and vice
versa. In either case, solution of the integral equations
yields 0' everywhere on L and in particular on AB in Method
and on CD in Method 2 (orDAin Method 2a). The required
results may then be obtained by quadrature.
This formulation is implemented numerically by dividing the

boundary L into N smooth intervals (elements) in each of which
~ and 0' are approximated by constants. We denote these
constants by
~-and
1
0:,
1
i = 1,2, ••• ,N, (16)
and apply Equation (15) at one "nodal" point qi in each

interval of L to obtain
N
} 0'.
j=1 J
J( J.) loglq-qildq- N
~ 0.
j=1 J
J( .)
J log'lq-qildq- 11"0.
1
= 0,
i = 1,2, •.. ,N, (17)
where J(j) denotes integration over the jth interval of L.

Prescribing one of the constants (16) in each interval, by
applying the boundary conditions (3) or (10) (or (13)) at the
nodal points, we thus obtain a system of N simultaneous linear
algebraic equations, in N unknowns, whose solution includes
values of~, on AB and CD (or DA).
In order to evaluate the coefficients in Equations (17), we

approximate (if necessary) each interval of L by the two chords
which join its end points to the nodal point within it. Then
all the integrations can be carried out analytically as
described in detail by Symm and Pitfield (1974) and by Jaswon
and Symm (1977). Indeed, the setting up and solution of
Equations (17) is all carried out here by Fortran subroutines
(Symm, 1980) based upon this discretisation.
Finally, if there are S intervals on AB and our interval

numbering starts from A, we approximate K, as given by
Equation (1), by
K( 1 ) = ~1T ~ 0'.h .• ( 18)

j=1 J J
where h denotes the approximate length of an interval as given

by its two chords. In Method 2, we approximate K, via Equations
(4), (7) and (12), by
= 1/m( 2 ) an d m( 2 ) = -L' 0'h

j j' (19)
7
where the summation extends over those intervals of L which

form CD. Similarly, in Method 2a, we approximate K by
K( 2a) = K(M( 2 a)) with M( 2 a) =- 2: !IS'.h ., (20)
J J
where the summation extends over those intervals of L which
form DA.
EXAMPLES
The methods presented above have been applied to a number of

examples as follows.
Problem 1
To compute the capacitance K between opposite sides AB and CD
of a unit square (Figure 1). In this case, it is immediately
evident that M = 1.0 whence it follows from Equation (4) that
K = 0.01756. Indeed, it is obvious from Method 2 that K will
have this value for any quadrilateral with a similar degree of
symmetry, as proved by Lampard (1957), who derived the analytic
formula
K =log 2/41T2 • (21)

e
D, . . . . - - - - - - - - - , C
G L
A'----------'B ~
Figure 1. Domain of Problem 1
Computed values of K and M, for various subdivisions of the

boundary L, are presented in Table 1.
8
Table 1
Capacitance K and Module M for a unit square (Figure 1)
N K( 1) M(2) K(2) M(2a) K(2a)

16 0.01744 0.98068 0.01866 1.01970 0.01650
32 0.01755 0.99360 0.01792 1.00644 0.01720
64 0.01756 0.99793 0.01767 1.00208 0.01744
128 0.01756 0.99934 0.01759 1.00066 0.01752
64+ 0.99985 0.01757 1.00015 0.01755
128+ 0.99996 0.01756 1.00004 0.01756
The first four sets of results in Table 1 correspond to

subdivisions of the boundary L into equal intervals and show
that, for this example, Method 1 converges most rapidly. For
these boundary subdivisions,
(22)
by symmetry, whence, from Equation (7),

M(2)M(2a) =1 ( 23)
and K( 2 ) and K( 2 a) bound K (from above and below respectively).
Improved results for Methods 2 and 2a may be obtained by

noting that the corresponding boundary value problems have
analytic solutions c = y and 0 = 1 - x respectively, relative
to the coordinate axes shown in Figure 1. Thus both 0 and c'
are constants on AB and CD in Method 2 and on BC and DA in
Method 2a. The results marked * in Table 1 correspond to
subdivisions which reflect this behaviour by treating these
sides of the square as single intervals and dividing the other
two sides into equal intervals. (Of course, in practical
problems, analytic solutions of the boundary value problems for
Methods 2 and 2a are not generally known. However, examination
of results for successive values of N with uniform intervals
will often reveal sections of the boundary L along which c and
¢' vary relatively slowly and may therefore be approximated by
constants over longer intervals; cf. Problem 3 below.)
In view of the symmetry of the square, no different results

are obtainable by rotating the quadrilateral in this problem.
Problem 2
To compute the capacitance K between opposite sides AB and CD
of a 2 by 1 rectangle (Figure 2). In this case, it is
immediately evident that M = 0.5 whence, from Equation (4),
K = 0.08930.
9
y r
D~----------------------------~c
G L
A'------------------'B~
X

boundary L, are presented in Table 2.
Table 2
Capacitance K and Module M for a rectangle (Figure 2)
N K(1) M(2) K(2) M(2a) K(2a)

16 0.09332 0.49673 0.09034 0.52638 0.08146
32 0.09070 0.49888 0.08965 0.50852 0.08668
64 0.08983 0.49963 0.08942 0.50273 0.08845
128 0.08954 0.49988 0.08934 0.50087 0.08903
96E 0,08930 0.49950 0.08946 0.50070 0.08908
96* 0.49999 0.08930 0.50010 0.08927
In Table 2, the first four sets of results correspond to

boundary subdivisions with N/4 equal intervals on each side of
the rectangle. In this case, Method 2 converges most rapidly.
The results marked E in Table 2 correspond to a subdivision

of the boundary L into equal intervals all round. Whilst
Method 1 is much improved, the results of Method 2 are worse
than before since, as in the previous example, 0 and 0' are
constants on CD; extra intervals here lead to oscillatory
solutions for 0' near C and D - a typical example of "Gibbs
phenomenon" (Hartree, 1952). These results may be improved, as
in Problem 1, by treating AB and CD as single intervals. Such
results and results of a similar treatment for Method 2a are
given in the line marked* in Table 2. Here the results of
Method 2 are better than those of Method 2a since the variation
in 0, from 0 to 1, takes place over the shorter side of the
rectangle in Method 2.
As in the previous example, no different results are

obtainable by rotating this quadrilateral in view of its
symmetry.
10
Problem 3
To compute the capacitance K between sides AB and CD of the
L-shaped domain, formed by removing a unit square from a square
of side 2, shown in Figure 3. In this case, m 0.577350 =
=
(Gaier, 1972), whence M 1.7321 from Equation (7) and
=
K 0.001756 from Equation (4).
Ar-------------~
'--------------,D
B'-------------------------------'c~
X

boundary L are presented in Table 3.
Table 3
Capacitance K and Module M for an L-shape (Figure 3)
N
K( 1) M(2) K(2) M(2a) K(2a)
48 0.001534 1. 7224 0.001810 1.7348 0.001741
96 0.001621 1.7295 0.001771 1. 7327 0.001753
128E 0.001764 1.7326 0.001753 1. 7320 0,001756
48 0.001892 1.7278 0.001780 1. 7358 0.001736
96 0.001851 1.7312 0.001761 1.7330 0.001751
128E 0.001763 1. 7333 0.001750 1. 7323 0.001755
128* 0.001757 1. 7321 0.001756 1.7320 0,001756
In Table 3, the first two sets of results correspond to

equal numbers of intervals on each side of the boundary and the
third set marked E, to equal intervals all round (i.e. with
extra intervals on AB and BC). These results show that
11
Method 2a converges most rapidly in this example.
The remaining sets of results are for the rotated

quadrilateral, where the roles of AB and CD are interchanged.
These results are very similar to the previous ones for the
same data. The improved results, in the line marked *•
correspond to selected boundary subdivisions which take account
of the behaviour of the boundary solution. In Method 1, the
intervals on sides adjacent to points A and B (where ~ is
discontinuous in the rotated problem) are graded in size -
reducing in length towards these points using a feature of the
library software (Symm, 1980). In Method 2, sides AB and CD are
each divided into 16 equal intervals and the remaining sides
into 24 equal intervals each. In Method 2a, sides AB and CD are
each divided into 24 equal intervals, BC into 32 equal
intervals and the remaining sides into 16 equal intervals each.
Problem 4
To compute the capacitance K between sections AB and CD of the
boundary of the L-shaped domain, formed by removing a unit
square from a square of side 2, shown in Figure 4. In this
case, m = 1.279262 (Gaier, 1972), whence M = 0.7817 from
Equation (7) and K = 0.0351 from Equation (4).
y r
r-------------~c
B~-------------,A
~--------------~~--------------~~
D X

boundary L are presented in Table 4.
12
Table 4
Capacitance K and Module M for an L-shape (Figure 4)
N K( 1) M(2) K(2) M(2a) K(2a)

56 0.0347 0.8650 0.0269 0.7053 0.0449
112 0.0349 0.8326 0.0299 0.7334 0.0410
168 0.0350 0.8199 0.0311 0.7449 0.0395
56 0.0349 0.8651 0.0269 0.7050 0.0449
112 0.0350 0.8325 0.0299 0.7333 0.0410
168 0.0351 0.8199 0.0311 0.7448 0.0395
168* 0.7856 0.0347 0.7778 0.0356
In Table 4, the first three sets of results correspond to

equal numbers of uniform intervals on each of the seven
boundary sections, the base of the domain G, on the x-axis,
being divided into two sections at D. These results show that
Method 1 converges most rapidly in this example.
The remaining sets of results are for the rotated

quadrilateral, where the roles of AB and CD are interchanged.
These results are very similar to the previous ones for the
same data. The results of Methods 2 and 2a are slow to converge
because of boundary singularities at the points B and D. These
results may be considerably improved by introducing graded
intervals, as in the previous example, in the neighbourhood of
these points. In particular, the results marked * in Table 4
correspond to modifications of the preceding boundary
subdivision to include 12 graded intervals over a length 0.3
and 12 equal intervals over the remaining length 0.7 of each of
the boundary sections BA, BC and the base section either to the
left of D (Method 2) or to the right of D (Method 2a).
CONCLUSION
We have described here two methods with variations, totalling

six methods in all, for computing the capacitance of a
quadrilateral on a contour L. It is evident from the results
obtained that which method is best will generally depend upon
the geometry of L and upon the positions of the four points on
it. In particular, Method 1 appears to be by far the best when,
as in Problem 4, Methods 2 and 2a introduce boundary
singularities at points of the quadrilateral. However, in other
cases, as Problems 2 and 3 show, Method 2 or Method 2a may be
the best.
Whilst the considered examples all have known analytic

solutions, it is also evident that in a general problem the
accuracy of the results of the various methods may be judged
from the measure of agreement between them. Bearing this in
mind, we observe finally that the equivalence of the various
formulations may be useful in the solution of boundary value
problems other than the capacitance problem, e.g. in problems
of steady-state heat conduction.
13
ACKNOWLEDGMENT
The author wishes to thank his colleague G.F. Miller for some
useful comments on the first draft of this work.
REFERENCES
Campbell, J.B. (1975) Finite difference techniques for ring

capacitors. J. Eng. Maths., 2: 21-28.
Courant, R. (1950) Dirichlet's Principle, Conformal Mapping,

and Minimal Surfaces. Interscience, New York.
Gaier, D. (1972) Ermittlung des konformen Moduls von Vierecken

mit Differenzenmethoden. Numer. Math., j2: 179-194.
Gaier, D. (1979) Capacitance and the conformal module of

quadrilaterals. J. Math. Anal. Appls., ~: 236-239.
Hartree, D.R. (1952) Numerical Analysis. Clarendon Press,

~fu~.
Jaswon, M.A. and Symm, G.T. (1977) Integral Equation Methods in

Potential Theory and Elastostatics. Academic Press, London.
Lampard, D.G. (1957) A new theorem in electrostatics with

applications to calculable standards of capacitance. Proc.
I.E.E., Monograph No. 216M, 104C: 271-280.
Symm, G.T. and Pitfield, R.A. (1974) Solution of Laplace's

Equation in Two Dimensions. NPL Report NAC 44.
Symm, G.T. (1980) The Robin problem for Laplace's equation. NPL
Report DNACS 32/80.
14
BOUNDARY ELEMENT SOLUTIONS TO THE EDDY CURRENT PROBLEM
S. J. Salon, J.M. Schneider, S. Uda

Rensselaer Polytechnic Institute
ABSTRACT
The two-dimensional eddy current problem is formulated in

terms of an electric vector potential function. The resulting
Helmholtz equation is transformed into a F~edholm Integral
equation which involves the unknown vector potential and its
normal derivative at the boundary. A set of simultaneous equa-
tions is found which approximates the integral. Results are
compared to a finite element solution for a sample problem.
I. Introduction
As manufacturers of electrical equipment strive for higher ef-
ficiency, more effort has been put into obtaining efficient
and accurate numerical means to predict stray losses. A large
class of problems involves a slowly time-varying magnetic field
in a conducting medium. In these problems, the displacement
currents can be ignored when compared with the conduction cur-
rent. There is extensive literature on finite difference
(Roberts (1959) and King (1966)) and Finite Element (Carpenter
(1975), Chari (1973), Carpenter (1977), Sato (1977), Salon
(1979)) solutions to these problems. In this paper, the eddy
current problem is formulated using an electric vector poten-
tial function. Only the two-dimensional case is considered,
so that the eddy currents will have two components, say x and
y, and be constant in the third dimension z. With the restric-
tions that the displacement currents be ignored and that the
material is isotropic and linear, the governing differential
equation is derived in section II. By means of Green's identi-
ty, this differential equation is then transformed into an in-
tegral equation. With certain restrictions placed on the forc-
ing function term, the potential can be written entirely as a
line integral on the boundary. This is shown in section III.
In section IV, a set of s~multaneous equations is obtained and
in section V, a sample problem is worked and the results are
compared to a finite element solution.
15
II. Electric Vector Potential

The Maxwell's Equations which apply are
'i/ X H = J (1)
and
'i/ E
oB
X =-at (2)
along with the continuity equation
'i/•J = 0 (3)
and the constitutive relationships
J crE (4)
and
B j.JH (5)
We now define the electric vector potential T as
'i/ x T = J (6)
Thus T is a function similar to H and can differ from H only by
the gradient of a scalar potential,
T =H + 1/¢ (7)
Substituting Equation (6) into Equation (2) and using the con-
stitutive relationships
'i/ X 'i/ X T = aH
-j.Jaa-t (8)
Assuming steady state sinusoidal conditions, Equation (8) be-
comes
'i/ X 'i/ x T = -jWj.JOH (9)
We now choose 'i/•T to equal zero. H can be thought of as being
composed of two components, an applied field, H0 , and the eddy
current reaction field T. Using a vector identity, we obtain
(10)
Defining
cl = jWj.JO (11)
we"obtain Helmholtz's equation,
(12)
III. Boundary Element Formulation of the Electric Vector

Potential
In this section, the Helmholtz equation for the two-dimensional
eddy current problem will be expressed in terms of a boundary
integral. We begin with Equation (12),
'i/ 2 T - a 2 T ~a 2 H (12)
0
Consider the Green's function, G(~,n;x,y) which satisfies the
equation
(13)
where x, y and ~.n are the field and source points respective-
ly. Multiplying Equation (13) by T and Equation (12) by G,
subtracting the two equations and integrating over the region
R
16
JJTo(s-x,n-y)dsdn a. 2 IJH 0 Gdsdn

R R
+ JJ<TV 2 G-GV 2 T)dsdn (14)

R
Using Green's theorem on the second integral on the right side
of Equation (14) results in
T(x,y) = a. 2(J
J H0 Gdsdn + J(-dn
T~G - GvT
dn)dB (15)
R B
In order to evaluate the Green's function, we return to Equa-
tion (13), which reduces to
2 ~ o(~-x,~-y)
(3 G +
a?" r1 eG
dr -
a.2G
<, 0
I
(16)
where
r = v'(s-x) 2 + <n-y)"- (17)
The solution of Equation (16) , is
G(s,n;x,y) = CI 0 (a.r) + DK (a.r)
0
(18)
where I and K are modified Bessel functions of the first and
0 0
second kinds order zero, respectively. From the behavior of
this function at infinity we deduce that C must equal zero.
1
The constant D is found to be 2n by integrating Equation (16)
aG
over a small disk located at r = o. The term dh in Equation
(15) is evaluated as
ClG ()G Clr
a,1 = ar an (19)
where
ar-
ClG - - £.... K (a.r)
2n 1
(20)
and
ar
dn ~ -cos 1jJ (21)
where K1 is the modified Bessel function of the first kind or-
der one. The terms are defined in Figure 1.
17
Figure 1 Section of boundary
Using Equations (15), (18), (20), and (21), we obtain
T(x,y) = ~; fJ H K
0 0 (ar) d~dn + 2 ~ fTK1 (ar) cosljJ dB
R B
1
- 2 1T Jar 0
~ (ar)dB (22)
B
The surface integral involving the forcing function can be
transformed into a line integral if H is limited to the class
0
of functions which satisfy Laplace's Equation,
\7 2 H
0
=0 (23)
Defining a new function g, such that
17 2 g = K (ar) (24)
0
The first integral in Equation (22) can be written
ffa K(ar)dR = JJ{H \7 g-g\7 H}dR

0 0 0
2 2
0 (25)
R R
Applying Green's Theorem
Jfa0 K0 (ar)dR = J{H 0 ~~- (26)

R B
Solving Equation (24) for g gives
-1
g = ~ (ar) (2 7)
a o
and
ClG K1 (ar)cosljJ
Cln = - a <28 )
Substituting Equation (27) and Equation (28) into Equation (26)
yields
18
JJH 0 K0 (ar)dR = -~JH 0K 1 (ar)cos dB

R B
+ -z 1 JaR
~ (ar)dB (29)
a on o
B
Using this result in Equation (22) gives
T(x,y) = -afH 0 K 1 (ar)cos~dB +~-f~H°K (ar)dB

2trr 2·,r n °
B B
+ azf i3
2n TK 1 (ar)cos~dB - -27flfClT
o .... 0
(ar) dB (30)
B B
Thus the potential at any point within the region is complete-
ly expressed in terms of an integral on the boundary. To re-
move the singularities which occur on the boundary, we will al-
low (x,y) to approach the point(x' ,y') on the boundary. The
singularity will be integrated around on a semi-circular path,
~B, as shown in Figure 2.
R
Figure 2 Integrating around singularities
Therefore T may be written as
T(x' ,y') =~ JH 0 K1 (ar)coswdB ~H 0K 1 (ar)cos~dB

B-~B ~B
+
1
2n
f Clno-Ko(ar)dB
ClH
+
1
2n
fdH
dnoKo(ar)dB
B-~B ~B
+ ~ JTK 1 (ar)cos~dB + 2 ~JTK 1 (ar)cos~dB

B-~B ~B
19
(31)
Any singularities have now been isolated into the three inte-
grals around ~B. Using the asymptotic property of K0 that
K (ar) ~· -in r as r ~ o (32)
0
in the integrals along ~B, the integrals are evaluated as fol-
lows:
For the first integral,
lim
e:~o
~H
21T o
K 1 (ar)cos~dB~fHo ~
on o
(ar)dB
~xB
~B
=1 J
21TH "(-inr)dB
3
~B o olt
=1 f
21T H
1T
0
1
-(e:d8)
e:
- H
0
(33)
0
For the second integral 2
lim
e:~o
1
(aH
Jp
n o
(ar)dB 1 f Oe:.tne:) = 0 (34)
6B M
For the third integral
lim
e:~o 2 ~ f TK1 (ar')cos~dB = 2! {: Tt(e:d8)
~B
T(x' ,y')
2 (35)
And for the fourth integral
lim - _!
e:~o 21T
f~
6B
an
(ar)dB =
0
I
~B
Q<e:tne:)dB 0 (36)
An expression for T at the boundary is therefore
T(x',y') H0 -afH 0 K 1 (ar)cos~dB + *]TK 1 (ar)cos~dB

B B
lf()T (ar)dB
- -~
1T on o (37)
IV. Development of a Set of Simultaneous Equations

In order to evaluate the integrals in Equation ~7), the boun-
dary B is approxim?ted by N straight line segments. The values
20
aT
of the potential T and its normal derivative, an' are assumed
to be constant on each segment and equal to its value in the
center of the segment. The values of r in Equation(J7) are dis-
tances between the field points and points on the boundary.
Approximating the integral by a summation
N N
T1• -nH -aL H K 1 (arij)cos~i.~s. + ~L T.K1 (ar )
0 j=l 0 J J l~l J ij
j#i j#i
1N dT.
-:;L ..,..JK (ari.)
" on o J
~s .
J
j=l
j#i
- !T0
TiiK (ar)ds.
I- o
(38)
an
i
The two integrals in Equation (38) contain the singularities.
The first of these integrals is zero, since the direction co-
sine is zero. The second integral has been evaluated in Luke,
and shown to be
2 2
!TI IK 0 (ar)ds ~\ d_
= =tt.. a2K+l -
L<~-'
n a\
"'n~--'1.. ck(3 2K+ 1} (39)
ank=o k=o
i
where aMi
6 = -4- (40)
where the ~ and ck values are found in Table I.
Table I
K
0 2.000 0.8456
1 0.6667 0.5041
2 0.100 0.1123
Using Equation (38) at each node on the boundary gives the fol-
lowing matrix equation.
aT
[J][TJ + [Kl[ann] = [F]
where
Jii 1 (41)
Jij - ~ 1 (ar .. )cos~i.Mj (42)

TI l.J J
Kii (See Equation (38)). (43)
21
(44)
-a. )'
F = -1T ·L H K (a.r .. )coslj! .. LIS. + H (45)
j=l 0 1 ~J ~J J 0
i~j
V. Sample Problem
The formulation was tested on the configuration shown in Figure
3. This is a square conductor 10 mm on each side and extending
to infinity in the z direction. A uniform applied magnetic fi-
eld, H0 is impinging on the conductor in the z direction. Due
to symmetry, only one-fourth of the conductor must be modeled.
For the boundary conditions we set T = 0 on the conductor sur-
face and ~! = 0 on the two lines of symmetry. The resulting
set of equations can be written as
[J] [T] + [K] [ ~!] = [F] (46)
In Equation (46), there would normally be two unknowns at each
boundary element. In this example, either T or ~! is specified
at every point and the system of equations may be solved.
The boundary of the quarter conductor was divided into 72
equal boundary elements. Equation (46) was solved using a
Gaussian elimination technique. The results were compared to
a finite element solution of the same problem and the results
are shown in Figure 4 for both the real and imaginary parts of
the potential. The results are very close, with a small error
occurring at the corner.
VI. Conclusions
A boundary element method has been presented for the slowly
varying magnetic field. The results have been validated by
comparison with a finite element solution of a sample problem.
Very little attempt was made to optimize the number of nodes in
either the finite element or the boundary element problem, but
it seems clear that the boundary element method is capable of
achieving high accuracy with fewer equations than the differen-
tial methods.
22
I
I
I T=O
I
I
II_ _____ _
J.l =J.lo
fJ =5.0xl0 4mho/mm
H = 1. 6 Amps/rom
0
Figure 3
Sample Problem, 10 mm x 10 mm
0 on dotted lines
.----------------·------
~.; ........--... I
r---------·---------·- '~,. _ . . ,\. I
·~, ...
r-------··---·~-,-,. ...,\ \ I
·------------, ---,_\ -\ \ I
\, \ \ !I I I
\
I
•\
\
i
1
iI
I
t
!
I
I
t
'
~
I
I i
I
•, i I 1
: \ ~ ! I i
:............................................~ ...............i ..........! ........L. .... .I
: ' ' • I I
Figure 4-a
Real Part of Electric Vector Potential
23
:---·-· -----·--------,
i-M-·-----·---..--------.. ---..
:! ·-.,.
..... ,
r··-----~------. ~ .. , :
. ·-.. ~--....... ·~
\ '
i
I .... , \
r. ----------------....... "·\. \,
·......... , \
I
r---~---. "-.. . '
'... \
•\
'
' ·-., \ i i
\ '\ i
' \ \ i
i
i
\
\ \·, i '\
l i
.
\
i
\
l I
i
1
l
i..........,..............................i .................l .............~.............i........._i
Figure 4-b
Imaginary Part of Electric Vector Potential
0.5
0.4
T
0.3
0.2
0.1
0.0
Node Number
Figure 4-c
Real Part of Potential T on the boundary
24
1.0
BIM
FEM
0.8
0.6
T
0.4
0.2
0
Node Number
Figure 4-d
Imaginary Part of Potential T on the Boundary
0.5
0
0.4
0.3
0.2
0.1
o.oo-~L-~---L--~--L-~---L--~--4-~
1 2 3 4
5 6 7 8 9 10 11
Node Number
Figure 4-e Real Part of Potential T on the Diagonal
25
References
1. Carpenter, C.J. (1977) Comparison of Alternative Formula-

tions of Three Dimensional Magnetic Field and Eddy Current
Problems at Power Frequencies. Proceedings lEE, Vol. 124,
No. 11.
2. Carpenter, C.J. (1975) Finite Element Models and Their Ap-
plication to Eddy Current Problems. Proceedings lEE, Vol.
122, No. 4.
3. Chari, M.V.K. (1973) Finite Element Solution of the Eddy
Current Problem in Magnetic Structures, IEEE Paper,
T-73-320-9.
4. King, E.I. (1966) Equivalent Circuits for the Magnetic Fi-
eld; 11-The Sinusoidally Time Varying Field. IEEE Trans-
actions PAS-85 No. 9 pp 936-945.
5. Roberts, J. (1959) Analog Treatment of Eddy Current Prob-
lems Involving Two Dimensional Fields, IEEE, Lond. Mono-
graph 341 M.
6. Salon, S.J. and Schneider, J.M. (1979) The Application of
a Finite Element Forraulation of the Electric Vector Poten-
tial to Eddy Current Losses, IEEE Paper A79-545-5.
7. Sato, T. and Sarto, S. (1977) Solution of Magnetic Field,
Eddy Current and Circulating Current Problems, Taking Mag-
netic Saturation and Effect of Eddy Current and Circula-
ting Current Paths into Account, IEEE Paper A-77-168-8.
26
NUMERICAL SOLUTION OF THE DIFFUSION EQUATION BY A

POTENTIAL METHOD
D.A.S. Curran*, B.A. Lewis* and M. Cross+
* Dept of Mathematics & Computer Studies, Sunderland Polytechnic
+ Mineral Resources Research Center University of Minnesota,
U.S.A.
ABSTRACT
The main advantage of using potential methods for generating

numerical solutions to elliptic partial differential equations
is that the dimension of the problem is reduced by one. This
feature reduces, not only the number of algebraic equations
to be solved, but also alleviates the problems of mesh
generation. Both of these factors combine to reduce conside-
rably the cost of computing acceptable numerical solutions.
In this paper a potential method is developed which
essentially extends to the diffusion equation most of the
advantages currently afforded to elliptic equations.
INTRODUCTION
The numerical solution of elliptic partial differential

equations by potential methods has received much attention in
recent yearsl-4 as they have some distinct advantages over
traditional finite difference of element methods for certain
classes of problems. Essentially, the potential method
transforms the solution of the partial differential equation
to that of a Fredholm integral equation over the boundary of
the solution region. This transformation effectively reduces
the dimension of the problem by one. As such, a smaller
number of algebraic equations results and the costs of mesh
generation are reduced since the solution is obtained
entirely in terms of the boundary values. Useful accounts of
these methods may be found in Jaswon and Symm 1 and Brebbia 4 .
During the last ten years various extensions to the potential

method have been proposed so that solutions to linear
parabolic differential equations may also be solved. 5 -8
However, in each case the 'parabolic' method loses the
advantage of a reduction in dimension. Furthermore, they
have generated their own problems with regard to numerical
27
stability and accuracy9. In the present paper, a method is

developed for the numerical solution of the diffusion equation
in one space dimension, which retains most of the advantages
of the potential methods available for elliptic partial
differential equations. The solution to the diffusion
equation is generated in two parts. The first is an integral
over the whole solution region which involves the initial
conditions of the problem only, whilst the second part is a
solution of a system of Volterra integral equations involving
the time coordinate only around the boundary. The solution
in the region's interior is then generated from the boundary
values. The method is applied to several problems and the
results favourably compared with those obtained by finite
difference techniques.
DEVELOPMENT OF THE METHOD
Essentially, the method involves a transformation of the

diffusion equation to a standard form followed by a series
of procedures to evaluate its solution both on the boundary
and within the integration region.
Transformation to standard form

Consider the diffusion equation in non-dimensional form,
Clu a2 u
Clt w 0 < X < R, (l)
subject to the boundary conditions
(2)
and the initial condition,
u(x,O) = f(x) 0 < X < R, (3)

A solution of equation (1) which also satisfies the initial
condition defined by equation (3) is given by
u 0 (x,t)
1
2/irt J'"' "x• C':;r' ), .,
0
(4)
If a new function is defined so that
Ul(x,t) = u(x,t) - UQ(X,t) (5)
then it satisfies equation (1) with initial condition,
ul (x,O) = 0 0 < X < R, (6)

and the following boundary conditions,
28
Ul(O,t) w1 <t> (7)
(8)
r,,,,
u1(£,t) w2 <t>
where
W1(t)
1
wl <t> - 2rrrt exp (- ~~ )dy (9)
w2 <t>
1
w2 <t> - 2/rrt
r,,,,
0
exp (- ~2)
4t dy (10)
Thus, the solution to equations (1) - (3) is given by

u(x,t) = uo(x,t) + ul{x,t) where uo(x,t) is given by equation
(4) and u1(x,O) satisfies the homogenous condition defined
by equation (6) together with boundary equations (7) and (8}.
The solution to equation (4) is easily obtained using a

suitable product integration formula, i.e.,
(11)
where£= mn and fi = f(ih). This leads to expressions for

the boundary values W1(t) and W2(t) also involving product
integration formulae, i.e.,
(12)
(13)
Determination of u 1 (x,t)
The solution of u(x,t) leads to the need for an evluation of
u 1 (x,t). For one dimensional problems Smirnov 10 has shown
that
1
ul(x,t)
2rrrt
(14)
_L (£-x)2) dT
+ 2rrrt (£-x) exp (- - --
4(t-T)
where ~(t) and x(t) are the source intensities of two dipoles
situated at x=O and x=£, respectively. These functions may
be determined by the following weakly singular Volterra
integral equations of the second kind,
29
<j>(t) - 9_ r .,,, K(t,1:)d1: W1 (t) (15)

0
-x(t) + 9- r;<,, K(t,1:)d1: (16)

0
exp(- £ 2 /4(t-1:)}
where the kernel, K(t,1:) (17)
2(il (t-1:) 3 2
In order to solve equations (15) and (16) numerically the

integrals on the left hand side of each equation must be
approximated by product integration formulae. As such,
r.,,.,, <j>(1:)d1:
n
i~O
A~n)
l. <Pi
0 (18)
r.,,.,,x,,,., n
i~O
A. (n)
l. xi
0
where the interval (o,t) has been divided into n subintervals
(ik,(i+l)k) (i=O, ..... ,n-l,k=constant), A~n) are the series
of integrals involving the kernel K(t,1:) ~nd <j>. = <j>(ik),
Xi= X(ik). Using the above integration formulae, equations
(15) and (16) lead to solutions at time nk for each of the
point intensities as,
n-1 ( )
"
<Pn =Wj(nk) + 9- i~O Aon
1- A(n)
n
In fact, for small enough values of k, the integrals A(n)~ 0,

n
so that <Pn and Xn may be obtained as recurrence relations,
ie. n-1
"
<j>n = Wj(nk) + ~n i~O Ao(n) Xi (19)
(20)
The evaluation ofui(x,t) is then given by:

30
where Ai(x) involves an integral of the kernel.
PROBLEMS AND SOLUTION PROCEDURES
In order to provide a realistic test of the basic method

several problems were posed and solved. For each problem
the solution for u (x,t) was obtained either by analytic
integration or by using a piecewise linear approximation to
f(x) on (0,1). If the latter procedure was used then the
interval was divided into m subintervals of constant width
l
h so that xi=ih. Then
x·+1
1
u 0 (x,t) f(y) exp ( - (x-y)2) d
4t y
xi
1 x-x·) +
erf ( 2Ttl. (22)
2
It - (x-x · >2)
+- exp ( ~ -
Tf
where
(23)
with regard to the evaluation of u (x,t), the source

intensities were approximated by piecewise constants so that,
n-2
~n = W (nk) + j~O An-j Xj+ 1 (24)
(25)
where
(26)
and
A . = erf ' !1, . ) (27)
n-J ( 21(n-J-1)k
The solution in the interior then follows as

31
n-1 (28)
u 1 (x,t) I:
j=O
where,
A . (x) (29)
n-J
Evaluation of equations (12), (13), (22), to (29) then defines

a numerical solution to the posed problem as summarized by
equations (1) to (3). Note that the solution is not required
within the whole region before progressing to the next time
step. The information is all carried forward in the
intensity functions at the boundaries.
Problem 1 - discontinuity at the boundary

The problem here is defined by equation (1) subject to the
boundary conditions,
u(O,t) = u(l,t) = 0
and the initial condition,
u(x,O) =1 (0 ~ x ~ 1)
The discontinuity between the initial and boundary conditions

makes it difficult for most methods to obtain accurate
solutions close to the boundary. This problem is most
easily solved by using the substitution v(x,t) = 1-u(x,t)
which leaves the form of the differential equation the same,
but changes the boundary conditions to
v(O,t) = v(l,t) = 1
whilst the initial condition becomes
v(x,O) = 0
A series of results for this problem are set out in Table 1
where comparisons are shown with the Crank-Nicholson finite
difference method using comparable time steps and a variety
of distance steps.
Problem 2 - initial condition piecewise linear

The boundary conditio~in this problem are specified as
u(O,t) u(l,t) =0
subject to the initial condition
2x 0 < X ~ 0.5
u(x,O)
2(1-x) 0.5 < X < 1
This problem was solved using an analytic evaluation of
uo(x,t) and a relevant set of results are shown in Table 2.
For the Crank-Nicolson calculations a small distance step
32
(ie. h = 0.01) was used to minimise space discretization

errors.
Problem 3 - initial condition non-linear

The boundary condition in his problem is defined as
u(O,t) u(l,t) =0
subject to the initial condition
u(x,O) x(l-x) 0 ~ X ~ 1
This problem was solved using both analytic and numerical

evaluations of uo(x,t) and a relevant set of results are
displayed in Table 3.
Problem 4 - sinusoidal initial condition

The boundary conditions hereare as specified above but
the initial condition becomes
u(x,O) = sinrrx 0 ~ x ~ 1
This problem was solved using numerical evaluations only for
u 0 (x,t); a set of relevant results are shown in Table 4.
CONCLUSIONS
The present methodcompares favourably with the C.N. method for

Problems 1 and 2. In Problem 1 where considerable difficulty
is experienced in obtaining accurate numerical solutions
because of the discontinuity between initial and boundary
conditions the method produces significantly more accurate
results than the C.N. method. Results for Problem 3 indicate
that the use of numerical integration in space does not
significantly affect the accuracy of the results, satisfactory
results being obtained for relatively large integration steps.
Only the results for Problem 4 are disappointing. In this
problem using small space and time steps does not significantly
improve the results.
Although the method discussed in the paper has only been

applied to problems in one space dimension it can be easily
extended to higher dimensional problems. The results from
the one-dimensional problems suggest that the present method
is capable of producing results at least comparable with
finite difference method while retaining most of the
advantages of the B.E.M.
33
REFERENCES
Brebbia, C.A. (1978) Bounday element method for engineers

Pentech Press London.
Brebbia, C.A. (Ed) (1980) New developments in boundary element

methods, CML Publications, Southampton.
Chang, Y.P. et al (1973) Int. J. Heat Mass Transfer 16, 1905
Curran, D.A.S. et al, (1980) Applied Math Modelling. ~. 398
Jaswon, M.A. & Symm, G.T. (1977) Integral equation methods in

potential theory and elastostatics, Academic Press London.
Brebbia, C.A. (Ed) (1979) Recent advances in boundary element

methods, Pentech Press London.
Rizzo, F.J. & Shippy, D.J. (1970) AAIAJ. ~. No.ll.
Smirnov, (1964) A course of higher mathematics Vol. IV,

Pergamon Press, Oxford.
Wrobel, L.C. & Brebbia, C.A. (1979) In numerical methods in

thermal problems (Eds R.W. Lewis & K. Morgan) Pineridge Press
Swansea, 58.
34
PROBLEM ONE
t 0.005 llt = 0.001

X 0.1 0.2 0.3 0.4 0.5
Exact 0.68269 0.95450 0.99730 0.99994 1.00000
I.E.M. 0.68269 0.95450 0.99730 0.99994 1.00000
C.N. h=O.l 0.67311 0.93179 0.98966 0.99874 0.99975
C.N. h=0.025 0.68169 0.95363 0.99668 0.99985 0.99999
t = 0.01 llt= 0.001

X 0.1 0.2 0.3 0.4 0.5
Exact 0.52050 0.84270 0.96610 0.99530 0.99919
I.E.M. 0.52050 0.84270 0.96611 0.99530 0.99919
C.N. h=O.l 0.52343 0.83233 0.95437 0.98970 0.99632
C.N. h=0.025 0.52030 0.84220 0.96547 0.99491 0.99897
t 0.1 I.E.M. Solution for decreasing Lit

X 0.1 0.2 0.3 0.4 0.5
Exact 0.1467 0.2790 0.3839 0.4513 0.4745
O.Gl 0.1488 0.2806 0.3849 0.4519 0.4750
0.005 0.1479 0.2798 0.3844 0.4516 0. 4747
0.0025 0.1473 0.2794 0.3842 0.4514 0. 4746
0.001 0.1469 0.2791 0.3840 0.4513 0.4745
PROBLEM TWO
t = 0.1 Llt 0.01
X 0.1 0.2 0.3 0.4 0.5
Exact 0.0934 0.1776 0.2444 0.2873 0.3021
C.N. 0.0948 0.1803 0.2482 0.2918 0.3069 0.01
I.E.M. 0.917 0.1743 0.2340 0.2822 0.2968
t 0.1 I.E.M. Solution for decreasing L\t

X 0.1 0.2 0.3 0.4 0.5
Exact 0.0934 0.1776 0.2444 0.2873 0.3021
0.005 0.0925 0.1758 0.2420 0.2846 0.2993
0.0025 0.0929 0.1766 0.2431 0.2859 0. 3006
0.001 0.0931 0.1772 0.2439 0.2867 0.3015
Solution at X = 0.5 L\t = 0.01
Exact C.N. I.E.M.
t=O.Ol 0.7743 0.7691 0.7743
t=0.02 0.6809 0.6921 0.6796
t=O.lO 0.3021 0.3069 0.2968
35
PROBLEM THREJ<.:
Solution for decreasing t using exact integration

t = 0.05
X 0.1 0.2 0.3 0.4 0.5
0.01 0.0470 0.0906 0.1255 0.1479 0.1557
0.005 0.0478 0.0916 0.1264 0.1487 0.1564
0.0025 0.0482 0.0921 0.1269 0.1492 0.1569
0.001 0.0485 0.0924 0.1272 0.1495 0.1572
t = 0.1
X 0.1 0.2 0.3 0.4 0.5
0.01 0.0293 0.0556 0.0765 0.0900 0.0946
0.005 0. 0295 0.0560 0.0771 0.0906 0.0953
0.0025 0. 0296 0.0562 0.0774 0.0910 0.0957
0.001 0.0297 0.0564 0.0776 0.0913 0.0960
Solution for decreasing t using numerical integration
t = 0.05 L\t 0.005

X 0.1 0.2 0.3 0.4 0.5
h=0.1 0.0474 0.0908 0.1253 0.1475 9,1551
h=0.05 0.0477 0.0914 0.1261 -.1484 0.1561
t = 0.05 L\t = 0.0025

X 0.1 0.2 0.3 0.4 0.5
h=0.1 0.0478 0.0913 0.1258 0.1480 0.1556
h=0.05 0.0481 0.0919 0.1266 0.1489 0.1566
t 0.05 L\t 0.001

X 0.1 0.2 0.3 0.4 0.5
h=0.1 0.0481 0.0917 0.1262 0.1483 0.1559
h=0.05 0.0484 0.0922 0.1270 0.1492 0.1569
36
PROBLEM FOUR
Solution for decreasing Lit and varying integration steps h
t 0.1 Lit 0.01

X 0.1 0.2 0.3 0.4 0.5
Exact 0.11517 0.21907 0.30153 0.35447 0.37271

h=O.l 0.11250 0.21357 0.29389 0.34556 0.36339
h=0.05 0.11319 0.21489 0.29571 0. 34770 0.36564
h=0.025 0.11337 0.21522 0.29617 0.34824 0.36620
t 0.1 Lit = 0.005

X 0.1 0.2 0.3 0.4 0.5
h=0.1 0.1133 0.21521 0.29622 0.34828 0.36623
h=0.05 0.11398 0.21655 0.29805 0.35044 0.36396
h=0.025 0.11443 0.21752 0.29940 0.35199 0.37012
t 0.1 Lit = 0.001
X 0.1 0.2 0.3 0.4 0.5

h=0.1 0.11401 0.21781 0.29842 0.35083 0.36889
h=0.05 0.11472 0.21816 0.30027 0.35301 0.36661
37
THE APPLICATION OF BOUNDARY ELEMENTS TO STEADY AND UNSTEADY

POTENTIAL FLUID FLOW PROBLEMS IN TWO AND THREE DIMENSIONS
Paul H.L. Groenenboom
B.V.Neratoom, The Hague, Netherlands
ABSTRACT
The Boundary Element Method is applied to both steady and un-

steady potential flow of a compressible liquid inside a fixed
volume. For steady flow the velocity potential satisfies the
Laplace or Poisson equation. This potential can be found by
solving the linear equations resulting from a discretisation
of the boundary integral equation.
The velocity potential for unsteady fluid flow can be found by
solving the three-dimensional wave equation. The potential at
a certain point and moment can be expressed in an integral over
the boundary of the values of the potential at the retarded
time plus a (inhomogeneous) source term.
The computer code BEREPOT (Boundary Element Retarded Potential
Technique) has been developed to calculate the potential
successively. This code can handle boundaries consisting of a
rigid structure and non-reflecting openings representing
infinitely long pipes.
This method has been applied to the flow of liquid sodium in

cooling components of Liquid-Metal Fast Breeder Reactors.
Special attention will be paid to the pressure wave propagation
resulting from a sodium-water reaction following a tube rupture
in a LMFBR steam generator. The boundary is given by (parts of)
the s.g. walls and the connections with the piping system.
The validity of the potential flow model and of the solution
method, the boundary conditions and other applications will be
discussed. Some results will be compared with the results from
analytical solutions and from other numerical methods.
INTRODUCTION
One of the main problems in the delivery of the heat transfer

components to the Liquid Metal Fast Breeder Reactor (LMFBR)
SNR-300 at Kalkar (F.R.G.) is the proven integrity of the
38
components to withstand the pressure loads that arise from

sodium-water reactions in the steam generator (s.g.).
Usually these loads are determined by transient pressure
gradients calculated by simple one-dimensional methods as the
characteristic method for water hammer phenomena. If one takes
a look at a straight tube steam generator (fig.1), one may
guess that such a calculation might be not too bad. However,
for a helical coil tube steam generator, as in fig.2, the
assumption of one-dimensional pressure wave propagation becomes
highly questionable. Since these conservative calculations, in
the worst case, give rise to problems in the proof of the
structural integrity of the s.g., we have to look for more
realistic numerical methods.
Two effects, not included in the simple methods, are expected
to give a significant reduction of the structural loads:
-1
the r -dependence of the pressure for spherical waves as
compared to the constant pressure wave propagation in one
dimension,
the dispersion and absorbtion of pressure waves due to the
elastic-plastic behaviour of the internal structures and
the walls.
The boundary element method (BEM) applied to potential fluid
flow may present a feasible method to perform the 3D-calcula-
tions and, in a later stage, also fluid-structure interaction
may be included.
MATHEMATICAL FORMULATION
For the calculation of pressure waves and other fast transients

one can derive the wave equation for the velocity potential.
(See for instance:Chorin and Marsden, 1979)
The conservation equations are:
ap +
at div(p~) = 0 ( 1)
av grad p
-+ (v. grad)v = - (Euler)
at - - p
(2)
The effects of viscosity have been neglected since at the time

scale of the Na-H 2o reaction, viscous boundary layers remain
very thin and the bulk viscosity is also negligible.
The Euler equations may be rewritten as
av 2
- + ~grad 1~1 + ~ Av=-~
at p
(3)
By taking the curl of this equation, it is found that the

vorticity, ~=curl ~· is governed by an evolution equation
aw
at+ curl(~ A~) (grad p A grad p)/p 2 (4)
39
bv neratoom
PRESSUR
RELIEF
PRESSURE
RELIEF
Fig. 1. Model of a straight tube LMFBR steam generator

40
bv neratoom
f WATER/STEAM
Fig. 2. Model of a helical coil LMFBR steam generator

41
The density, p, is a function of the pressure, p, and of the

entropy, S. If we assume that the variation in the entropy is
so small that the explicit dependence of p on Smay be
neglected, the gradients of p and p will have the same direc-
tion and the r.h.s. of eq.4 will vanish.
So,if there is no initial vorticity, no vorticity will be
present. Consequently, the flow is irrotational and can thus
be described by a velocity potential ~(~,t) with
-:!.. = grad ~ (5)
By combining the conservation equations, it is straight-
forward to show that for velocities small compared to the
speed of sound, which is very high in liquid sodium, the
velocity potential satisfies the three-dimensional wave equation
'V
2
~(~, t) -
1
2 -2
a 2~ = 0 (6)
c at
The pressure is related to the potential by the Bernouilli
equation
(7)
For the sodium-water reaction a source term representing the

expanding reaction bubble has to be included in the wave
equation
(8)
The solutions represent the travelling waves in the sodium.

Using the Kirchhoff-method (Stratton, 1941) for the scalar
electrodynamical potential, this can be transformed into an
explicit boundary integral equation. Since that relation is the
basis of this method, a short sketch of the elegant derivation
will be given.
The fundamental solution or Green's function of the wave
equation is defined by
2 I
'V G(~,tlr 0 ,t 0 )-
2G a
2 - 2 = o(r-r) o(t-t) (9)
c at 0 0 -
Multiplication of eq.8 by G and of eq.9 by~ and integration
over the volume of interest and the timevariable yields
+00
fdt
_oo
fdr{GV 2~-G.~ a
V- C at
i 4P'V 2 G~2 ~}=
2
at 2
c
2
+oo ( 10)
= fdt fdr{y(r,t)G-~(r,t)O(r-r )o(t-t )}
-00 v-- - -~ o
42
The Green's theorem for volume-integration reads:
fdr{GV 2~~V 2 G}= fdrV•{GV~VG}=fdA{~}, (II)

v- v- ~ n n
where n is defined as the outward normal at the surface ~.

By a sTpilar dzrivation one finds that the time integration
a~ ac .
over G-=--Z -~ - 2 vanJ.shes.
Applic~tion o~tthis and of identity II to the l.h.s. of eq.IO
yields:
+oo
_
fdt ~fdA(~~)
an an
(I 2)
00
whereas the r.h.s. of eq.IO becomes

+oo
fdr fy(r,t)G(r,tlr ,t )-c. ~(r ,t ) (13)
v --00 - - _£ 0 l. _£ 0
with ci = +I for~ inside V

= +~ for ~ on the boundary ~
= 0 for ~ outside V
Since, for potential problems, the field outside a spherically
symmetric source is the same as that of a central point source
(at~), y(~,t), may be replaced by ?(t) 8(~-rs).
The fundamental solution of the 3D-wave equation is

8(t-tret)
G(r,tlr ,t ) ( 14)
- 0 0 4TIR
with R= ~~-Ea. I
and tret = tQ - R/c (retardation)
Expressl.on IJ then becomes
I y(tret)
4n lr -r I - ci~(ro,to) ( 15)
.....§. _£
and from expression 12 one obtains:
+00 ~,n O(t-t ) ~ O(t-t +R/c)

fdt QfdA {~
_ 00
a~
an
ret
41T R -~(
a
an
o
4TIR )}
I a~ I +oo aR O(t-t )
= fdA 4TIR[an]t -41T fdt fdA~{- 2ret
Q ret -oo Q R
1 a
+ Rc at(o(t-tret)}
,L fdA{.!.[.a!J
R ~n t=t
~ -(~(r,t
aR R
)~(r t ))} (16)
ret R2 an
411
~ a ret c -' ret
43
The final expression for the potential at place ~ and time

ta is subsequently found
This is the basic equation for BEREPOT. A two-dimensional

representation of the geometry is sketched in fig.3.
Fig. 3. Sketch of the geometry used by BEREPOT
Since the potential~(~, to) is expressed solely in quant~t~es

at previous time steps, this method is explicit and requires
no matrix solution. Note furthermore that neither time nor
volume-integration has to be performed. The only integration
involves the boundary of V.
APPLICATION TO THE NA-H 20 REACTION
According to the German licensing authorities, the present

definition of the design base accident for a Na-H 2o reaction
in a steam generator is an instantaneous, complete tube
rupture. Also the chemical reaction,
Na + H2o ~ NaOH + i H2 + energy,
is assumed to take place immediately. Under these assumptions

narrow pressure peaks will be formed. After that, a quasi-static
pressure build-up will occur in the reaction zone and in the
component. The pressure waves will propagate through the steam
generator and into the connecting secondary sodium loop.
44
Further assumptions about the reactions are:

-isothermal reaction with ~1150 K
- the volume of the reaction zone is determined by the formed
hydrogen (ideal gas)
Since the velocities of the sodium are small compared to the
soundspeed, the pressure wave propagation and the subsequent
flow may be described by a velocity potential which obeys the
3D wave equation. Therefore the theory of the previous chapter
may be applied for which we will discuss a number of aspects.
Boundary conditions
The first consideration concerning the boundary is the choice
of the geometry. A reasonable choice for the boundary is (a
part of) the inner flow shroud enclosing the rod bundle or the
structural limits of the header, depending on the assumed
reaction site. For the moment these structures are assumed to
be rigid, so the corresponding boundary condition is that the
velocity normal to the surface vanishes;
d(j) =0 ( 18)
<ln
In a later stage we hope to replace this condition by a more
realistic condition derived from the elastic-plastic response
of the structure.
The enclosure of the reaction volume is not complete; openings
have to be considered. These openings give entrance to other
parts of the steam generator and the connecting piping system.
The corresponding boundary condition is that of an infinitely
long pipe represented by the radiation condition
(19)
Coupling to one-dimensional codes is a future option.

Reaction zone
The source of the pressure waves and of the flow of liquid
sodium is the expanding reaction bubble which is assumed to
remain in spherical shape. In that case,it is easy to prove
that the condition, that the velocity of the sodium at the
Na-~ 2 interface is equal to the growth rate of the bubble
rad1us, may be replaced by a sourceterm, y, as introduced in
eq.S. The sourceterm is equal to the opposite of the volume-
growth of the bubble.
2 .
-4nR (t) R(t) o(E-~)
-'Q'(t) o(E-~) (20)
45
Treatment of pressure discontinuities

The very high pressure peaks immediately after the tube rupture
have a duration in the order of 10 ~s only. The corresponding
size of the boundary elements is so small (a few centimeters)
that a complete partitioning of the s.g. would require more
data than any existing computer could store. Therefore, it is
preferable to treat this as a discontinuous wave front which
reflects at the rigid boundaries [Neilson et al, 1978].
Numerical aspects
Next to the choice of the source function and of the geometry,
the discretisation of the boundary will primarily determine
the success of the BEREPOT calculations. The time-step of the
problem will have to be chosen in accordance with the time-
development of the source function. Subsequently, the element
size has to be equal to or greater than the distance traversed
by a wave in a single time-step.
Numerical problems will arise from the fact that the retarda-
tion within a single element is not a constant and that this
retardation will not be equal to an integer number of time-
steps.
The second problem is solved by linear interpolation between
the two nearest-by values. For the solution of the basic
equation (17) it is necessary to evaluate the time derivative
of the potential at the boundary. This evaluation can be done
by a three-point backward difference formula (Mitzner, 1967)
At sharp edges or corners of the boundary the normal is not

defined uniquely, so if integration points are situated at
these edges, the factor ci of eq.13 remains undetermined.
A solution of this problem has been suggested by Hrobel and
Brebbia (I 9 79).
Finally we have to check the numerical stability of the solution
by applying small changes in boundary conditions, initial condi-
tions, discretisation scheme and time-step.
TEST CASES
To demonstrate the feasibility of the boundary element method

to potential flow problems, we have examined two examples.
Two-dimensional, stationary potential flow

Without sources, the velocity potential satisfies
0 (22)
46
The correspondi ng fundamental solution is

I I
G(E_,r 0 ) = - 2TI ln (~) (23)
- - 0
Proceeding in the same way as for the three-dimen sional, time

dependent case, it is straightforw ard to derive:
(r-r ).n
- 0 -
2 (j) (24)
l!.-roj
with ci=+I,O,~ for~ resp. inside V, outside V or on the
boundary rl.
By division of the boundary in constant elements, this integral
equation becomes a matrix equation which can be solved by
standard subroutines . The boundary conditions are determined
by the specific case (fig.4): homogeneous in- and outflow
through openings at the top or the bottom of one wall of a
rectangular (2D) box. We assume that the normal derivative of
the potential at the boundary is zero everywhere except at
the openings where it is assumed to be a finite constant.
The boundary has been discretised into two meshes
I) 4 elements in the x-direction , 24 for y
2) 8 elements in the x-direction , 48 for y.
The results of the simple BE-code for this case are presented
in fig.S. There the values of the potential along the boundary
(four walls) have been plotted. For comparison, results of a
least-squar es finite element method, developed in our group
(De Bruijn,l980 ), for this case, are also given in this figure.
Since the potential is defined apart from an arbitrary constant,
these constants have been chosen in such a way that the curves
are close to each other but do not overlap. Note the nice
agreement between the three results.
Wave propagation for a point source in a 3D rectangular box

The governing equations is the one treated in the second
chapter (eq.l7).
The specific example of a point source in a box has been chosen
because the potential can be solved analyticall y by the method
of mirror sources. The condition of zero normal velocity at a
wall near a source is satisfied by placing a source of the same
(time-depen dent) strength on the rear side. For two parallel
walls with a source in between, one obtains a whole series of
mirror sources and for a rectangular box one finds series of
sources in all three directions.
Due to the finite velocity of the pressur~ waves, the number of
sources that influence the boY. at a certain moment remains
finite and can be found by a computer code, named SOMMIS.
47
(O,q) (p,q)
._j
d\P
a;=+u ---'
s
--+
-.
--.,
p m
q 6 m
s = m
u = m/s
Nx= 8
C!~P_o
C!n ~-o Ny= 48
a = .0625 m
b .0625 m
d\P
~-u
C!n
-
_ _.!
__.! s
--' 2b
..-!
X I >( I
(0,0) (p,O) 2a
box elements
Fig. 4. Geometry for the testcase of potential

flow in a 2D box
48
.......
- I I I
""" ._,.a
~\~ :>.
""
elf
--~~ "0
l:l
;::J
"'~
0
.0
"'~ -
00
N l:l
0
......
~
elf
Q)
~
C)
~
.....
~ 0)
•.-l
~""'-""-
-
"0
0
"0"'
'0"-
' ..."-
\\\
I\ I \
!II 00
II 1
If)
II 1
1;1
;//
// /
/ /.
//
///
//
/ /,
/'l
/'i
//
f1{/
//- -:1"0000
N-:1"-:t
//~/ -~ ~ ~
II1 E-t!-1
00
--,,,
P-oP<:>::
I I I ~~~
~~....:1
1/J II II II
I I I
-NM
///
Fig. 5. Velocity potential along the boundary of the

2D box
49
A test version of BEREPOT has been made for this problem.

The boundaries of the box, of dimensions 3x4x5, have been
divided into 376 elements of dimensions 0,5 x 0,5. The poten-
tial is assumed to be constant in a single element. For the
source function a short pulsed Gaussian has been taken and at
two boundary points the response of the potential has been
evaluated, both by the exact method of mirror sources (SOMMIS),
and by the boundary element method.
In fig.6 the original pulse at the source is given and the
direct influence which is calculated exactly by SOMMIS.
The calculation by the BEREPOT-code has been performed with
timesteps of 2 (units of time) and with 0,5. For the largest
time-step the reproduction of the exact potential is, of
course, not very accurate for such a fast transient, and a
possibly small delay in the response may be observed. For the
shorter time-step the reproduction of the secondary and the
third reflection is quite good, but at later times the solution
seems to become unstable.
Still, for such a crude mesh, the reproduction of this strong
transient is encouraging.
In fig.7 the results of the potential history at a boundary
point closer to the source is given and a similar behaviour as
in fig.6 may be observed.
CONCLUSIONS AND PROSPECTS
Although effects of viscosity and convection may cause

strongly non-linear behaviour in steady fluid flow, in many
cases, the potential flow model may present some indicational
results (De Bruijn and Zijl,1981).
For pressure wave propagation the potential flow model is
justifiable. Therefore, the application of boundary elements
to the solution of the wave equation for the velocity potential,
is primarily expected to be of interest for the calculation of
the consequences of explosions and violent reactions like the
Na-H 2o reaction.
Spec1fic advantages of the use of the Boundary Element Method
for flow problems are the explicitly time-dependent formulation
and the reduction of the volume integration to the boundary.
A disadvantage of the method is that at every time-step the
mutual influences between all elements have to be evaluated.
So, if one would like to know the pressure field inside the
volume, a Finite Element Method would probably be preferable.
In the future we intend to use this method for more realistic
cases such as the growth of the reaction bubble inside realis-
tically modelled parts of the steam generators. To this end an
order has been given to the Computational Mechanics Centre in
Southampton to develop a BEM-computer-code for the 3-D wave
propagation with more possibilities than in the presented test-
cases. That code will be able to handle arbitrary geometry with
both rigid walls and radiation boundary conditions and,
possibly, Fluid Structure Interaction (FSI).
~
"":!
....
()Q
?' - · - - · SOMM!S (EXACT)

p.. - - - - BERE OT ~t=2
.... 0'"d
srt> ... - - - - - BERE OT ~t=0,5
rt>
::s ...::s
"'....
0
....
Ill
::s ,...
Ill
,... 0
tT
.....
0 Ill
~
'1j
....::s0 -
/_.'I
... I \~---,-\
"'c0 I X \'
l"j
n
rt>
....::s
Ill
...::r'
l"j
rt>
rt>
I f-j-\ ' \ \
v/ ·\. 'II '._ / '-::.:..1
10 20 30 \ 40 50
time (a.u.)
'>:I
,....
(JQ
.....
rt '"d
::ro
ti rt - - - f--- sm jMrs (E ~CT)
(b (b
(b ::s - - - -BEl ~POT !:J. =2
I rt
0. ,.... ----BEl ~POT !:J. =0,5
a,.... "',. . .
(b
::sC/l 0 ('
,..,.,
,.... I \
::s0 "'
"' '"0
...... 0
r-~-t
,....
<T ::s ,r"\, (\ (\ j \
0 rt
:><
i
C/l ~_!_
0
~ .p~~·--{\-1-\
\./ ·v· \ ~-'-11-'t/-
j . ,
ti
n . I ,
(b
,.... j \./'
::s source funct ·on f ;r"·[\ 1 v'\
'\- !\-:- ,.
"' ~·I " ~ ·' \ 1I \ '1 •~ ' v'
,,_ !-!' ' 1,l - , :~- i'-~-·'- v~ ,-1 I
' ' 11 Ji\ :'9...~ :: : ·vf \j
IIII ' ' 1~·~·"'"''
~ ~-'i-t' ' I-
I I ~ ' ' '-"
r-'-.-i-,':\~--(.-!i-:: \;
I I I. ' ., :: I'
I " i '~;X:f·~ ,_, •:
I' I
1 ,\/t '" I"' I
20 40 60 80 100
time (a.u.)
~
52
Both the 3D-effects and the FSI might give a significant

reduction in pressure loads on steam generator structures due
to hypothetical instant tube rupture. If this method proves
successful for this difficult problem, other numerical problems
which can be described by the Poisson or the wave equation for
a vast range of applications also come in hand.
REFERENCES
Bruijn, J.G.M. de (1980), private communication.
Bruijn, J.G.M. de and Zijl, W. (1981), Least Squares Finite

Element Solution of several Thermal-Hydraulic Problems in a
Heat Exchanger, Proc. of the 2nd Int. Conf. on numerical methods
in Thermal Problems, Venice.
Chorin, A.J. and Marsden, J.E., (1979). A mathematical

Introduction to Fluid Mechanics, Springer Verlag, New York.
Mitzner, K.M. (1967), Numerical Solution for transient

scattering from a hard surface of arbitrary shape - retarded
potential technique, J.Acoust. Soc.Am.42: 391.
Neilson, H.C.; Lu, Y.P. and Wang Y.F. (1978)

Transient scattering by arbitrary axisymmetric surfaces,
J.Acoust.Soc. Am. 63:1719.
Stratton, J.A.: (1941). Electromagnetic Theory,

McGraw-Hill Co.
Wrobel, L.C. and Brebbia, C.A. (1979) in:

Numerical Methods in Thermal Problems ed's: Lewis, R.W. and
Morgan, K.; Pineridge Press, Swansea, p58.
53
COMPUTING STRATEGY IN THE INTEGRAL EQUATION SOLUTION

OF LIMITING GRAVITY WAVES IN WATER
J.M. Williams
Hydraulics Research Station

Wallingford, Oxfordshire
U.K.
ABSTRACT
Progressive irrotational gravity waves in water have a limiting

form in which, in their steady flow representation, the crest
reaches the total energy line. The crest is then angled rather
than rounded, with an included angle of 120°, implying a
singularity of order 2/3 in the complex potential plane.
An integral equation technique, employing two leading

terms to represent the crest, has been successfully used by
the author to generate definitive solutions of limiting waves
over the whole range of depth : wavelength ratios. The paper
presents a specimen solution and discusses various aspects of
computing strategy, including the choice of number and position
of nodal points. Conditions under which convergence fails are
also briefly considered.
INTRODUCTION
Inviscid irrotational water waves of infinitesimal amplitude

form part of classical hydrodynamics (Lamb, 1932). The
solution depends on the depth : wavelength ratio which can
vary from zero (the solitary wave) to infinity (the deep-water
wave). The extension to finite wave amplitudes, when the non-
linear surface condition arising from Bernoulli's equation
can no longer be linearised, was begun by Stokes (1847).
Stokes developed a series expansion, with successive terms
g1v1ng successive Fourier components of the wave form. Third
order solutions by Stokes's method have been tabulated, for
example by Skjelbreia (1959),and used extensively in engineer-
ing design. More recently, and with the aid of more advanced
computers, Dean (1965) has developed his so-called stream
function wave theory which has also been published in tabular
form (Dean, 1974).
54
Whereas in Stokes's method the coefficients of the first,

second and third order solutions are found in sequence, each
from its predecessors, Dean finds the coefficients of his
component terms simultaneously by minimising the root mean
square error in the free-surface condition; this is, in effect,
a type of integral equation technique. Dean's tabulated
solutions range from 3rd to 19th order; the higher orders
extend the range of useful accuracy to larger wave amplitudes
compared with the Stokes 3rd order tabulations.
A uniform wave motion may be brought to a steady state by

superimposing a velocity opposed to the wave celerity. This
steady motion has a total energy line above the surface to
which the Bernoulli condition at the surface is referred. The
highest possible wave is that whose crest just reaches the
total energy line to give a stagnation point there. Stokes
(1880) showed that the crest of this limiting wave would be
no longer rounded but angled, with an included angle of 1200.
It follows that waves nearing the limiting form will need a
large number of Fourier components to describe the profile
accurately, and consequently a high-order solution whether by
Stokes's, Dean's or other comparable method.
The problem has received much attention in recent years,

prompted on one hand by practical design needs in offshore
engineering and on the other by the availability of more
powerful computers. We may mention in particular the work of
Schwartz (1974) and Cokelet (1977) each of whom has developed
a series expansion technique extending for some solutions to
orders beyond 100. Although their results offer an important
advance in accuracy over previously tabulated solutions there
remains room for improvement for waves nearing the limiting
height. In common with Stokes's original method, the accurary
also falls away towards the shallow-water end of the range
(depth : wavelength -+ 0) •
The present author developed an integral equation solution

for finite-amplitude waves in 1970 and has recently extended it
to the limiting case. The method used and the results obtained
for limiting waves have been set out in detail in a recent
paper (Williams, 1981). The accuracy is now well within almost
all practical requirements, with many field and integral
properties being defined to within 1 or 2 parts in 106.
Furthermore, this accuracy has been maintained throughout the
depth : wavelength range, so that the results include a defini-
tive description of the maximum solitary wave.
In this paper the method of solution will be described in

outline only, accompanied by some specimen results for non-
limiting and limiting waves. The main part of the discussion
will describe several alternative computing methods for
limiting waves which were considered before the final scheme
55
was established, as well as some important features of the

final scheme itself. It is hoped that this exposition will
encourage others to consider similar techniques for related
problems involving singularities and also to advance the
associated formal numerical analysis.
INTEGRAL EQUATION METHOD FOR NON-LIMITING WAVES
The method is best introduced briefly by reference to a

solution. Figure 2 presents a solution to 21st order computed
by the author's method of 1970. The figure shows, in steady-
flow form, a half-wavelength of the irrotational motion with
symmetry about both the crest and the trough and a rigid
boundary at the bed. At the free surface the flow satisfies,
within the working limits of accuracy which are to be
discussed, Bernoulli's equation, referred to the total energy
line. The flow takes place (for a full wavelength) within
the domain of the complex potential x = ¢ + i~ given by
0~¢~-A, 0~~~-2; the ratio 4TI/A is defined as d. The space
origin and parameter F 2 are chosen, as explained in Williams
(1981), by preliminary consideration of the corresponding
infinitesimal wave. It is sufficient to note here that
F2 = (1/d) tanh d (1)
and that l/2F 2 is in the present notation a normalised form of
the acceleration due to gravity. The remaining physical
dimensions defining the motion are wavelength L, mean depth h
and wave height H.
To remove the difficulty of the unknown position of the

free surface the complex potential X is taken as the
independent variable and is transformed in turn to the
T-plane given by
X = i(2/d)ln T, T = pexp(i8). (2)
With reference to Figure 1, the flow field is bounded in the
T-plane by the concentric circles p = exp(-d) = R (bed) and
p = 1 (surface), while the wave crest occurs at T = 1 and the
trough at T = -1. Limiting cases are given by R = 0 (deep-
water wave) and R = 1 (solitary wave).
It is shown in Williams (1981) that the solution of the

problem may be written as
z = x + iy = -x + iF 2 + s.
leaving s ~ + in as a holomorphic function, with appropriate
symmetry, to be found. The boundary conditions for the
imaginary part n are the bed condition
n = o, p =R (3)
56
and the surface condition
n
p2 (l _~d ~~)2
1 HE(6) O (4 )
----pz- ,
+a- + (~d ~~)2
where a is an initially unknown parameter.
Ps and E are alternative forms for expressing the non-zero

surface error which will inevitably occur in the solutions.
Ps denotes a ratio of pressure to density so that Ps(6)
indicates the distribution of surface pressure which would
make the solution exact. We shall also consider in the
discussion the maximum modulus of Ps• over a wavelength,
denoted by Ps· HE expresses the error as a form of displace~
ment, E being the ratio of displacement error to wave height.
It is also of interest to evaluate<:*, the root mean square of
<:(6) over a wavelength, to compare our solutions with earlier
results.
The solving technique consists of constructing s from a

set of component solutions,each of which has the necessary
symmetry and satisfies the bed condition, Equation 3. Thus
+ (5)
bed, P=R
crest
Figure 1. The T-plane

57
where
r;o -i(l/d)ln(T/R), (6)
(7)
The coefficients are computed iteratively, to satisfy

Equation 4 at N + 1 nodal points, p = 1, 8 = k~/N, k = 0,1,2,
•.• ,N. (The alternative set of nodes 8 = (2k-l)~/2N,
k = 1,2, ... ,N was also used for some earlier runs.) To
determine the nominal amplitude of the wave to be computed
a 1 is chosen and the last unknown is then a, which governs
the level of the total energy line (given by n = -aF2).
The iteration proceeds,from a trial set of coefficients,

by evaluating £ at each nodal point, and hence an error
vector E. Each coefficient in turn is then perturbed by a
small quantity and the errors are recalculated and used to
construct a rate-of-charge matrix ~. with elements
d£k
aa-• k = 1,2, .... , N + 1. (8)
N
Corrections to the coefficients are estimated from -~-lE and

the cycle is repeated until convergence has been achieved to
an acceptable tolerance.
This algorithm converged well and yielded useful

solutions for waves up to about half of maximum amplitude
over a wide range of R, with N taking values up to 21, the
maximum feasible on the small computer (ICL 1901A) available
at the time.
The solution shown in Figure 2, for which Ps and £ * are

each 2.5 x lo- 5 , has been chosen to correspond with a published
5th order s~lution by Dean for which Ps is 2.2 x lo-s and c*
is 1.7 x 10 s. The two solutions are thus of similar accuracy
and the wave profiles are graphically indistinguishable.
Dean's solution (developed in the physical rather than complex
potential plane) minimises the root mean square error sampled
over a large number of points, probably more than the 22 nodal
points used in the author's solution. The computing effort
needed in the two cases may thus be comparable but Dean's
strategy has achieved a solution which is very compactly
defined, with only 5 terms, after what is in effect an
economization process. This topic will not be pursued in the
present paper which presents compact solutions for limiting
waves obtained by a different technique.
58
When in an integral equation solution the coefficients of

the component terms have been found any other property of the
flow may readily be calculated. For illustration Figure 2
includes several streamlines of the steady flow and also a
corresponding set of particle paths which arise when a wave
celerity is superimposed. The celerity is arbitrary unless
other factors enter the problem to define it; we have used in
Figure 2 the important celerity c = A/L which makes the space-
mean bed velocity over a wavelength equal to zero. The
particle paths show clearly the forward motion experienced by
each particle during a half wave period. The remaining half
of each particle path is a mirror image of that shown, doubling
the net forward displacement in a full period.
EXTENSION OF THE METHOD FOR LIMITING WAVES
The 120° angle of the crest of a limiting wave transforms in

the T-plane to a singularity of order 2/3. A first step in the
computation was therefore to include a leading term of this
form. This has been done before by Yamada (1957 a and b), for
both the deep-water and solitary waves, and Lenau (1966) for
the solitary wave.
It was soon found, however, that this leading term,

although valuable, was not sufficient to ensure a really high
accuracy without an unreasonable overall number of terms. The
next development was suggested by the work of Grant (1973) who
investigated an expansion for the flow near the crest and found
that the index of the second term was governed by the roots of
the equation
tan ~~~ = -(4+3~)/313~. (9)
For present purposes the index of the second term in the
T-plane is the first root of Equation 9 > 2/3, namely
~ = 1.469345741. The series is developed further by Norman
(1974) and Williams (1981).
Component ~ functions for use in the computations were

therefore defined by the general form
i ~ [~ - ,m] v ~ - ( ;2tJ v f
- [
(10)
[~ _ R2m]v-[i _ 1 Jv
For the present work with a singularity at the crest, T = 1, we
take m = A = 1 and v = 2/3 or ~. As the singularity moves away
from the crest, for near-limiting waves, A decreases from 1 and
ultimately to zero, when ~m 0 v degenerates to ~m, defined in
Equation 7. The remaining features of Equation 10 ensure that
~m,A,v satisfies the bed boundary condition, Equation 3, and is
normalised in the same way as ~m' with n = -1 at T = 1.
59
Table 1. Tests with leading terms and second crest node; d=2.5
No. of subsequent 6
Run N s q coefficients, ec 10 Ps
aO,al, •..•
1 15 so - 15 - 5344
2 15 floating - 14 - 5534
3 15 so yes 14 - 84
4 15 floating yes 13 - 400
5 15 so yes 15 1T/420 300
6 15 floating yes 14 1T/420 55
7 15 floating yes 14 1T/l40 55
8 21 floating yes 20 1T/420 18
9 21 floating yes 20 1T/140 >14
10 19 floating yes 18 1T/280 12
The trial solutions for iteration thus now contained one

or both of the new terms
s ~1,1,2/3 and q ~1,1,~· (11)
It is shown in Williams (1981) that s has a theoretical value

of
(12)
which follows from Stokes's (1880) analysis of the flow at the

crest.
A series of tests was carried out in which these new terms

were tried in computations of the limiting case for d=2.5,
approaching the deep-water wave. The basic set of nodal points
were taken as before at 8 = k1T/N, k = O,l,2, ..•• ,N. In
contrast with the non-limiting case the control of amplitude
is now automatic so that no coefficient needs to be fixed at
the outset. The option does exist, however, to allow the
leading coefficient either to "float" or to be fixed at its
theoretical value s 0 • a remains as an unknown to be found.
An important innovation which developed during these tests

was to introduce a further nodal point at a small distance 8
from the crest. In this way the two leading coefficients c
intended to define the crest accurately were provided with two
associated nodal points.
The tests are summarised in Table 1, whose final column

shows the resulting value of p , as an indication of overall
s
60
accuracy. The last entry, Run 10, is the solution finally

accepted.
Table 1 shows the marked improvement of accuracy from

Run 3 onwards, after the introduction of the second leading
coefficient q. Runs 3 to 6 then show that introducing the
second crest node has a comparable restraining effect on the
solution to the fixing of s to s 0 ; attempts to do both,
however, nullify the benefit (Run 5). The marginally better
accuracy of Run 6, with floating s, was preferred to the
theoretically superior formulation of Run 3. It is shown in
Williams (1981) that the resulting solution has a locally
incorrect crest acceleration but that this has no significant
consequences. Run 7 then shows that the result is comparative-
ly insensitive to the value of 6c while Runs 8 and 9 show the
effect of increasing N. The significance of the final choice
of N = 19 (Run 10) will be discussed in the next section.
The computing algorithm remained essentially as before

with the program simply being enhanced to accommodate the more
general form of s (Equation 10) and the more flexible treat-
ment of fixed and floating coefficients and nodal point
distribution. Convergence to a solution remained generally
good, although with some important exceptions, to be discussed
later.
The computations for this phase of the work were done on

the CDC 7600 at the University of London Computer Centre. It
was possible, within a reasonable run, for N to be increased
up to 79. Most runs needed from 5 to 30 iterations depending
on the starting values available and the rate of convergence.
The time per iteration varied approximately as N2 • 84 and was
about 5 seconds for N = 35.
THE CHOICE OF N AND 8

c
The algorithm generated accurate solutions for the whole range
of d, which are set out in detail in Williams (1981). There
were, however, important features which dictated somewhat
differing treatments for small, moderate and large values of d.
0.2~d~0.5
This constitutes the shallow-water end of the range, which
causes greatest difficulty in most methods and where, for a
given accuracy, N must be largest. Fortunately, however, the
algorithm was well behaved except for the important proviso
that even values of N exceeding about 40 failed to converge.
N was therefore restricted to odd values which gave progress-
ively greater accuracy as they increased. Table 2 illustrates
the typical behaviour of a sequence of solutions at d = 0.2,
the limit of the range computed. It is shown in Williams
(1981) that this wave is, to present computing accuracy,
61
Table 2. Behaviour of solutions for d o. 2, ec 11/420
106p
s
for e = kTI/200
App. 106
N k=O 1 2 3 4 5 6 7 8 9 10 11 12 13 14
It'ns Ps
49 12 0 22 31 18 1 -7 -7 -4 0 2 2 1 0 0 -1 32
53 14 0 16 21 9 -1 -5 -4 -1 0 1 1 0 0 0 0 23
57 16 0 13 15 4 -2 -4 -2 0 1 1 0 0 0 0 0 18
63 20 0 8 9 1 -2 -2 0 0 0 0 0 0 0 0 0 11
79 30 0 3 1 -1 -1 0 0 0 0 0 0 0 0 0 0 4
Table 3. Behaviour of solutions for d TI/280
106p for 8 = kTI/200

s
106
N It'ns k=O 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 ps
29 8 0 0 3 7 7 5 2 0 -2 -3 -3 -2 -1 0 0 0 1 1 8
30 15 0 -1 1 5 5 4 1 0 -2 -2 -2 -2 -1 0 0 0 0 0 8
31 10 0 -1 0 3 4 3 0 0 -2 -2 -2 -1 0 0 0 0 0 0 8
32 21 0 -1 0 2 3 2 0 -1 -1 -2 -1 -1 0 0 0 0 0 0 8
-
33 11 0 -2 -1 1 2 1 0 -1 -1 -1 -1 0 0 0 0 0 0 0 9
identical with the maximum solitary wave. The residual surface

pressure Ps has been tabulated at intervals of TI/200 from the
crest; beyond the range of 8 shown Ps is in all cases zero to
the accuracy indicated. The actual maximum, Ps• generally
occurs between tabulated points and is shown on the right of
the table.
It was found in this range of d that the behaviour was

not sensitive to the choice of 8 •
c
o.6:;dO.O
This range needed moderate values of N. It was found that N
could now be even without convergence being destroyed although
more iterations were always needed than for odd N. A new
feature was that accuracy no longer improved indefinitely with
increasing N but reached a peak at a particular optimum value.
Table 3 shows the behaviour of a sequence of solutions for
d = 0.7. The position of Ps changes in this sequence from
3<k<4 for N = 29 to a point between the two crest nodes in the
remaining solutions.
62
Table 4. Coefficients of tne limiting wave computed

for d = 1. 3, N = 22, e 1T/280
c
s 1. 82517163 q 0.52996266
ao -0.12288679 a 1.06815531
al -1.66040564 all 0.00000833

a2 0.12164711 al2 0.00000441
a3 0.01081024 al3 0.00000232
a4 0.00241700 al4 0.00000120
as 0.00077108 alS 0.00000060
a6 0.00029856 al6 0.00000028
a7 0.00013036 al7 0.00000012
as 0.00006161 alB 0.00000004
a9 0.00003070 al9 0.00000001
alO 0.00001583 a20 0
h 1. 94852 l/2F 2 0.754303
Apart from the value of p , certain integrals over the

wavelength which involve p ar~ also important in maintaining
overall accuracy, as explafned in Williams (1981). The chosen
solution, N = 32, was preferred on the basis of a near-minimun
Ps coupled with minimum values of these integrals.
The solutions were "fine-tuned" in this way throughout

the intermediate range of d. ec again seemed to have no
critical effects.
2.5~d::lo.o
Within this deep-water end of the range the general character-
istics of the mi.ddle range remained. However, for even N
convergence became very '"slow or sometimes failed altogether.
Furthermore, with N now small the process of finding an optimum
N, which had worked successfully in the middle range, no longer
provided a sufficiently fine control. It was then found that
ec could be used instead to achieve the best accuracy and in
this way ec = 1T/280 was determined to be most suitable. This
value was subsequently adopted throughout the range of d since
it had been found elsewhere not to be critical. The solution
chosen ford= 10.0 used N = 19, with Ps = 12 x 10- 6 . As shown
in Williams (1981), this wave is identical, to working accuracy,
with the theoretical extreme of the maximum infinite-depth wave.
63
The critical N giving best accuracy which has been found

in the second and third ranges of d probably exists also in the
first range. It is, however, beyond the limit of a reasonable
computer run (N = 79) and would also correspond to a higher
accuracy than has been generally achieved.
The algorithm has thus yielded solutions which are defini-

tive for all practical purposes but does not in its present
form seem able to give arbitrarily high accuracy. To investi-
gate whether and how this could be achieved presents the
challenging problem of reconciling the properties of the crest
singularity with the needs of the numerical procedure. Some
brief attempts were made to include further terms in Grant's
series but it was found either that the improvement was
negligible or that the iteration failed to converge.
STUDIES OF CONVERGENCE
The failure of convergence at large, and occasionally small,

even values of N was discovered at a late stage of the work.
The development of the algorithm had been pursued over the
middle range of d, or in the deep-water range with a fortuitous-
ly odd value of N, as in Table 1. The failure does not result
from a near-singular iteration matrix ~. The algorithm had
always used a perturbation of 10- 4 for constructing the rate-
of-charge matrix ~. Some recent tests have shown that converg-
ence can be affected by the perturbation used. The most
dramatic example is for d = 10.0, N = 18, 6c = n/420, which
failed to converge at all with a perturbation of 10- 4 but
converged strongly when it was reduced to 10- 9 . A similar
test was tried for d = 0.2, N = 54, ec = n/420. The reduced
perturbation produced a hesitant, intermittent convergence but
further reduction to lo-ll gave no improvement.
These results suggest that the discretised problem which

the algorithm is set to solve does not have a unique solution.
It seems that, for certain even values of N, two solutions are
sufficiently close to each other for the perturbation scheme
to jump between them and cause an oscillating non-converging
behaviour. In practical terms the strategy adopted seems
completely to avoid the difficulty over the whole range of d,
such that a solution is always found within the close computing
tolerance which has been used. Provided the solution of the
wave problem itself is unique, as is generally accepted, it
should not matter which of alternative discrete close approx-
imations is arrived at. Nevertheless this aspect of the
numerical analysis would benefit from a more thorough study.
EXAMPLE OF A COMPUTED LIMITING WAVE
Figure 3 and Table 4showthe limiting wave computed ford 1.3

with N 22, ec = n/280. The depth: wavelength ratio is
~
r-------- -- Total en,!!9Y line, y=(l-a.)F2
------ ---1y
mean level
I I
III II I
rD
~
L ,
:
: I
:I
i
I /I !
1 I I I ?
tr~ugh line, bed, y = 2 + F2, ljJ = - 2 crest line,
x=~L
=-~ X
x==O
t
Steady motion - stream lines and equipotentials Wave motion- particle paths
Figure 2. Computed non-limiting wave; 4w/A = d = 1.3263; N = 21;

h/L = 0.2148, H/h = 0.3125, c/(2F 2 /h) = 0.8306
___ - ---- ~
- -energy
-------------Total - - line
I
I
L___ m_!an lev~--- _ __

I
I
I
rP
Steady motion
lWave motion
Figure 3. Computed limiting wave; 4~/A = d = 1.3000, N = 22;

h/L = 0.214772, H/h = 0.55257, c/(2F 2 /h) = 0.87885
0)
01
66
0.2148, as for the non-limiting wave of Figure 2. Ps is

1.0 x lO-s and £* is 1.3 x 10-s. Despite the even N this run
converged strongly in only 12 iterations.
Figure 3 shows again streamlines of the steady motion and

particle paths in the unsteady motion. Note that the forward
drift is now very strong at the surface.
CONCLUSIONS
Definitive solutions for limiting gravity waves have been found

from an integral equation approach in a relatively compact form
which permits all required properties of the flow to be
calculated. The algorithm is generally well-behaved although
it is not apparently able to give arbitrarily fine accuracy in
its present form. The full implications of the failure of
convergence under certain conditions with even N remain to be
investigated. With these provisos the method constitutes a
possible approach for other problems involving singularities
and perhaps non-linear boundary conditions.
ACKNOWLEDGEMENTS
The work described in this paper was carried out as part of

the research programme of the Hydraulics Research Station and
is published with the permission of the Director of Hydraulics
Research. Most of the computations were done at the University
of London Computer Centre through the Institute of Computation-
al Mathematics, Brunel University, where the author is
registered for a Ph.D. degree.
REFERENCES
Cokelet, E.D. (1977) Steep Gravity Waves in Water of Arbitrary

Uniform Depth. Phil. Trans. R. Soc. Lend. A, 286, 183-230.
Dean, R.G. (1965) Stream Function Representation of Nonlinear

Ocean Waves. J. Geophys. Res., 70, 4561-4572.
Dean, R.G. (1974) Evaluation and Development of Water Wave

Theories for Engineering Application. (2 Vol.). U.S. Army
Coastal Engineering Research Center, Special Report No. 1.
Grant, M.A. (1973) The Singularity at the Crest of a Finite

Amplitude Progressive Stokes Wave. J. Fluid Mech., 59,
257-262.
Lamb, H. (1932) Hydrodynamics, 6th ed., Cambridge University

Press.
Lenau, c.w. (1966) The Solitary wave of Maximum Amplitude.

J. Fluid Mech., 26, 309-320.
67
Schwartz, L.W. (1974) Computer Extension and Analytic

Continuation of Stokes' Expansion for Gravity Waves. J. Fluid
Mech., 62, 553-578.
Skjelbreia, L. (1959) Gravity waves; Stokes Third Order

Approximation; Tables of Functions. Council on Wave Research -
The Engineering Foundation.
Stokes, G.G. (1847) On the Theory of Oscillatory Waves.

Trans. Carob. Phil, Soc.,~' 441-455.
Stokes, G.G. (1880) Considerations Relative to the Greatest

Height of Oscillatory Waves which can be Propagated without
Change of Form. Mathematical and Physical Papers, Vol. 1,
225-228, Cambridge University Press.
Williams, J.M. (1981) Limiting Gravity Waves in Water of

Finite Depth. Phil. Trans. R. Soc. Land. A, in Press.
Yamada, H. (1957a) Highest Waves of Permanent Type on the

Surface of Deep Water. Rep. Res. Inst. Appl. Mech. Kyushu
Univ., ~' 37-52.
Yamada, H. (1957b) On the Highest Solitary Wave. Rep. Res.

Inst. Appl. Mech. Kyushu Univ., ~' 53-67.
Section II
Elasticity Problems
71
ON THE CONSTRUCTION OF THE BOUNDARY INTEGRAL REPRESENTATION AND

CONNECTED INTEGRAL EQUATIONS FOR HOMOGENEOUS PROBLEMS OF PLANE
LINEAR ELASTOSTATICS
M. Pignole
Laboratoire de Mecanique Physique

Universite de Bordeaux I, France
ABSTRACT
Four fundamental identities are established. They rest on

addition to the interior problem (resp. exterior) of an
exterior problem (resp. interior) and consideration of displa-
cement and tension jumps across the boundary, connected with
these two problems. Two of these identities express, according
to these jumps, displacements and stresses everywhere in both
the interior and exterior domains. The two others respectively
connect the previous jumps to a boundary displacement or ten-
sion linear combination.
We obtain the true boundary integral equations by intro-

ducing boundary conditions into these identities.
Two ways of discretisation are shown,
INTRODUCTION
Boundary element methods are usually classified into direct

and indirect. In the direct method, the boundary node unknowns
are directly obtained by solving a boundary integral equation.
Then, domain unknowns can be computed everywhere with the aid
of boundary node values. In indirect methods, boundary node
parameters are used. A boundary integral system has to be
solved first to compute them, allowing then the calculation of
boundary and domain unknowns.
The direct method [Lachat, J. C. (1975)] and indirect

methods [Banerjee, P. K. (1976)] have been simultaneously
developed for elastostatics. They have allowed an effective
treatment of practical size engineering problems in the last
few years. The equivalence between each category has been
72
established [Brebbia. C.A. and Walker. S. (1978)]. The

parameters used in indirect methods have been often called
fictitious density distributions but they can be given a
physical meaning [Hartmann. F. (1980)].
The main aim of this paper is to show how direct and

indirect formulations ensue from four fundamental identities
and how the indirect method parameters have a mathematical
meaning. They are displacement and tension jumps when a com-
plementary exterior boundary value problem (resp. interior)
is associated with the initial interior one (resp. exterior).
We limit our study to practical two dimensional problems. The
displacements are continuous and the tensions almost every-
where continuous on the boundary. Study is analogous for three
dimensional case. It may be extended to other linear operators.
A BASIC IDENTITY FOR REGULAR FUNCTIONS
Basic equations
The displacement vector u. the strain tensor E and the stress
tensor a are functions of the point. Later on. we shall use
the well known relations.
'V
E = 'V u (I)
a = A0 (tr E) I + 2 \l E (Hooke) (2)
a = A• (div u) I + 2 \l ~ u (Lame) (3)
-div a = f (equilibrium) (4)
- [ll 6. u + ( A'f + ll) 'V di v u J = f (5)

(Navier)
where the operators and the coefficients mean as follows
'V gradient
'V
'V symmetric gradient
div divergence
tr trace (with the meaning tr E E )

mm
I identity matrix
f volume forces assumed to be zero in homogeneous

problems
A~ ll constant coefficients
73
A
.= A for plane strain
(6)
2 \.1 A
for plane stress
A + 2 \.1
A and \.1 are the Lame coefficients.
Connected system
In spite of the common use of Equation 5, we prefer consti-
tuting a system with Equation 3 and Equation 4. For a homoge-
neous case, we write
jA~(div u) I+ 2 \.1 ~ u- a 0
(7)
1div a = 0
or
Sys (u, a) 0 (8)
Variational identity
We call w a bounded open set
y its smooth boundary
(~, ~) some couple of a regular vector ~ and a regular
symmetric second order tensor ~
(u, a) is a solution of Equation 8. Then
~ [A~ (div u) I + 2 \.1 ~ u - a] • ~ dw + L div a • ~ dw =0 (9)
denotes the scalar product of two vectors or the inner

product of two tensors.
Integrating by parts, we find
[A" V (tr ~) I + 2 \.1 div ~] • u dw - L(~ + ~ ~) • a dw
{[Ak (tr ~)I+ 2\.1 ~] n} • u dy +h ~ • (a n) dy 0 (10)
Where n is the outward normal
Now, let us consider a as a function
u ~ a (u)
Equation (10) may be rewritten in a shortened form
-LA <~> • u dw -LB (~. ~) • a ( u) dw
+Jy a'~ ( ~) n • u dy + Jy ~ • t (u) dw =0 (II)

74
with
A (q,) = ·;.:'f· 'V (tr q,) I + 2 ]..l div q, (12)
(13)
t = cr n (14)
t is the tension vector on y.
A PARTICULAR 2-UPLE OF COUPLES (~, q,)
We seek a 2-uple of (~(j), q,(j)) couples, solution, in the

distribution meaning, of
(IS)
(16)
j = 1 '2
Where oy is the Dirac distribution in y, and
[~]
Calculation of ~(j~·)
Substituting for q, J from Equation 16 into Equation 15, we
obtain
N ~(j) = o I(j) (17)
y
where N is the Navier operator
N = - []..l 1:1 + ct + ]..l) " di v J (18)
~(j) is the jth column-vector of a matrix U, an elementary

solution of the N operator. U is a matrix of distributions that
may be represented by functions everywhere except in x = y and
aresummable on a disk around y. Theoretical justification may
be found in [Schwartz, L. (1973)], details of cll.lculation in
[Pignole, M. (1980)].
,,,(j) - U ~ "n _1
= U1 uij"' + (19)
"'(y)i- (y)ij r U2 r,i r,j
with
r = lx - Yl (20)
r,i
3r
= ax. = - ar (21)
~
ayi
75
(22)
A."' + u
(23)
U has two properties of symmetry
u(y}ji (x) = u(y}ij (x) (24)
u( ) . . (y) u( ) . . (x) (25}

X 1J y 1J
Calculation of ~(j)
Using Equations 16 and 19, we obtain
+ 4 u2 r,i r,j r,k] (26)
Where oim is Kronecker 's symbol
The 2-uple of ~(j} constitutes a third order tensor.
A PARTICULAR 2x2-UPLE OF COUPLES (~, ~)
Now, we seek a 2x2-uple of (~(jk), ~(jk)) couples, solution,

in the distribution meaning, of
(27)
(28)
j, k = 1,2
with
I (II)
[6 ~] "!(12)
[~ 6]
1(21)
[~ ~] i22)
[~ ~]
Calculation of ~(jk)
System 27 28 may be rewritten
= v(jk) (29)
r··. ,
N ~(jk)
].I)
with v(II) = - (30}
'y' ']
:>..*oy'2
v02) = v<2I) r ].I

0
y'2]
oy' I
(31)
76
v(22) =- r:Y.oy'l J (32)

Lu,"' + 2 )l) oY, 2
Now, we know an elementary solution U of N. Therefore

~(jk) = u ~ v<jk) (33)
(y) (y)
~(jk) = _!_ [D (o.k r,. + o .. r,k- oJ.k r, 1. )
(y) i r I 1. J l.J
+ D2 r,i r,j r,k] (34)
with
(35)
(36)
The 2 x 2-uple of ~(jk) constitutes a third order tensor D

D = (jk) (37)
(y)jki ~(y)i
D has the following properties
D(y)kji (x) D(y)jki (x) (38)
D(x) jki (y) - D(y)jki (x) (39)
Calculation of ¢(jk)
Using Equations 28 and 34, we obtain everywhere except in y
¢(jk)
(y)H =rz1 [- Dl (oik oj9. + 6l.J
.. oH - ojk oH)
D2
+ (Dl - T) <6ik r,j r, 9. + 0 •• r,k r,.Q.
l.J
+ ok.Q. r' i r' j + 0j 9. r,i r,k)
- 2 D1 ojk r,i r,.Q. - D2 oi.Q. r,j r,k
+ 4 D2 r,i r, j r,kr,.Q.J (40)
The 2 x 2-uple of ¢(jk) constitutes a fourth order tensor.
FOUR FUNDAMENTAL IDENTITIES
Let us consider a bounded open set n with a smooth boundary r

and the complementary !1' of n.
Let us suppose that u may be
twice differentiated in n u n' and the interior and exterior
traces uiir' u!Er' t!Ir' tiEr exist. (u, o) is a solution of
77
Equation 8 and u and a decrease as 1/r and l/r 2 at infinity.
The displacement and tension jumps are given by
[u] ulrr - uiEr (41)
[t] t lrr -tiEr (42)
n is the outward normal for n.

S is a y centred disk the radius of which is assumed to
n
vanishyand SR a disk such as C SR, the radius of which is
assumed to tend to infinity. By wr1ting Identity 11 for (u, a)
and a particular (ljJ, ) in n U n' n BR' 6 , we want to
obtain, when £ vanishes and R tends to infi~ity, four funda-
mental identities between
a and ~(tlrr' tiEr) with the 2x2-uple of (ljJ(jk), q,(jk))

~has the meaning of linear combination.
Calculation of a (ljJ(j)), a (ljJ(j)) n, a (ljJ(jk)), a (ljJ(jk)) n
Using Equations 3 19 and 34, we find
'" (j ) )
crik ( "(y) = rI rT
LI
(, , ,
ujk r,i + uij r,k - uik r, j
)
+ T2 r,i r,j r,k] (43)
with
(44)
(45)
n 1 and n2 are given by Equations 35 and 36.
T(j)
(y)i
- T 1 (n. r,. - n. r,J (46)

1 J J ~
The T(j) constitute a third order tensor.

78
+ 0ki r,i r,j + 0ji r,i r,k) (47)
+ s4 <ou r,j r,k + ojk r,i r, i)

+ ss r,i r,j r,k r, R.J
with
s, - 2 ~ n, ~2 s (48)
0
s2 - 2 (A¥- ~) n, = - (A* - ~) ~ s0 (49)
s3 = ~ (2 n1 - n2) = - A¥ ~ s 0 (SO)
s4 - 4 ~ n, = 2 s, (51)
ss 8 ~ D2 8 ~ (A¥+ ~) s0 (52)
(53)
5o TI (A¥ + 2 ~)
s(jk)
(y)i OH
( ,,(jk)) ni
'~'(y) = ~2 f ~~ [s3 (oik r,j + oij r,k)
S4 8 j k r, i + S5 r, i r, j r, k] (54)
+ s 1 (oik nj + oij ~) + s 2 ojk ni
+ s 4 ni r,j r,k
The S(jk) constitute a third order tensor.
Identity between u, [u] and [t] for ·Y Q Q'

Using the previous procedure, we find
fr, T(j) • [u] dr + h 1/J(j) • [t] dy + h E

T(j) • u dr
+frr
E
1/J(j) • t dr + ( T(j) • u dr +
JrR rR
1/J(j) • t dr J o (Ss)
It is easy to pr~ve that the integrals on rR andrE

vanish, except~ T(J) • u dr, when e: vanishes and R tends
to infinity. E
79
u
j
(y) = Jr
( T(j) (x) • [u (x)] dr (x)
(y)
r
+ u (j) (x) • It (x)] dr (x)
Jr (y) L!
(56)
noting U(j) as 1jl (j).
Writing Identity 11 for (u, cr) and the particular (1, O)

couple (ljl, ~). inn' and nun' we find that the hypothesis
of decreasing at infinity for u and cr is equivalent to
h tlrr dr =h tiEr dy (57)
Identity between cr, [u] and [t]

Using the previous procedure or using Equations 3 and 56,
we find
crJ.k (y) = r s(jk)

Jr (y)
(x) • [u (x)] dr (x)
+( D (jk) (x) • [E (x)] dr (x) (58)

Jr (y)
noting D(jk) as ljl(jk).
Identity between (u ITf~Er), [u] and [t]

For a smooth boundary-
i (uj (y) lrr + uj (y) IEr) = hT~~) (x) • ~ (x)] dr (x)
k ~;~
+ U (x) • [!: (x)] dr (x) (59)
interpreting the first integral in the meaning of its principal

value.
Identity between (tlu.!-.!IEr), [u] and [t]

Assuming that the tens1on is continuous and that the displa-
cement satisfies a Holder condition at y, using the method of
[]:ruse, T. A. (1977)], we find
= r s(jk)
Jr tyl (x) ~ (y) • [u (x~ dr (x)
r n<jk) (x) nk (y) • [!: (xi]

(60)
dr (x)I
Jr tyl
The two integrals must be interpreted in the meaning of their

principal value.
OBTAINING THE INTEGRAL REPRESENTATIONS AND INTEGRAL EQUATIONS
We consider, an initial interior boundary value problem IP,

80
or an exterior one EP, with boundary conditions
= t
ol
EP lulr
tlr
u
t
=
(61)
such as indicated on figure I.

r
figure
First, we obtain particular identities from the four funda-

mental identities, according to the choice of an associate
problem, then integral equations by introducing boundary
conditions •
The direct method

We choose
for IP tiEr o (62)
for EP t lu o (63)
From identities 56 58 and 59 and Equations 62 and 63, we

obtain the direct particular identities
( T(j) (x) • u (x) dr (x)

jr (y)
+( u<j) (x) t (x) dr (x)

Jr (y)
I
2 uj (y) y €f J for IP (64)
u. (y) YES1 (65)

J
I (66)
- 2 uj (y) y€r }
for EP
- u. (y) Y Est (67)
J
81
Ir s(jk) (x) • u (x) dr (x)

(y)
D(jk) (x) • t (x) dr (x)
+Jr (y)
f_
ajk (y) y€. Q for IP (68)
ajk (y) y€Q' for EP (69)
By introducing the boundary conditions into the identities

for r, we obtain the B. I. Equation. After resolving it we
know all the values of u and t on r and we may compute u and
a everywhere in the domain withthe aid of Equations 65 and 68
for IP or 67 and 69 for EP.
Indirect methods
The particular identities contain [uJ and I or [t]. We may
construct an infinite number of indirect formulations, each
according to a complementary problem, using Identities 56 58
59 and 60. First, a system has to be solved on r, then we may
compute u and I or t on r and u and a in the domain. We give
some examples.
Tension jump formulation For this formulation, we choose,

for IP and EP
[u]r = 0 (70)
For Y£ r, Identities 59 and 60 become
u. (y) =(
J Jr U Y(j) (x) • [t (x)] dr (x) (71)
=h D~~~)
I
[t.
~ for IP
+- (y)]
2 J
tj (y) (x) (y) • [t (x)] dr (x) 1 (72)
--z [t.
J
(y)] for EP
For y in the initial domain, we may write, from Identity 56

Identity 71 and from Identity 58 as follows
ajk (y) = kD~~~) (x) • [t (x)] dr (x) (73)
The [t] system to be solved is obtained by using Equation 71

on ru and Equation 72 on rt (for IP or EP respectively).
Next,
Identity 71 => u on r t and in the domain

82
Identity 72 => t on r
Identity 73 => o in d~e domain
Displacement jump formulation For this formulation, we choose,

for IP and EP
(74)
In the same way, for yE r

I
+ - [u. (y)] for IP
2 J
u. (y) =fr T(j) (x) • [u (x)] dr (75)
J r cyl
- -} [uj (y)] for EP
t. (y) =
J
i s(jk) (x)
r (yl
~ (y) • ~ (x)] dr (x) (76)
For y in the initial domain, we may write
uj (y) =Jr Tg~ (x) • [u (x)] dr (x) (77)
ojk (y) = f sg~) (x) • ~ (x)] dr (x) (78)
The [u] system is obtained by using Equations 75 on r and

Equation 76 on rt (for IP and EP respectively). u
Next,
Identity 75 => u on rt
Identity 76 => t on ru
Identity 77 => u in the domain
Identity 78 => 0 in the domain.
A mixed formulation Now, we choose, for IP and EP
(79)
and
(80)
We obtain, for y6 r
Ir t T ~;~ (x) • [u (x)] dr (x) + Ir uU ~;~ • [!: (x)] dx = uj (y)

(81)
yE.r
u
83
kt S~;~) (x) ~ (y) o {E (x)] dr (x) + Jr u D ~;~) (x) ~ (y)

(82)
o [t (x)] dr (x) t. (y)
J
Equations 81 and 82 together constitute a [t (x)] r and
~ (x)]r system which has to be solved at first. u
t
Then,
u. (y)
J =lrt
T (j) (x)
(y)
o lii (xj] dr (x) +Jrr u uU>
(y)
[_t (xj] dr (x)
+..!.2 [u. (y)]

J
for IP
Y E. rt (83)
--z 1
[E. (y)]
J
for EP
t€ ru is given by an analo~us expression using the kernels

Snand Dltand the jump [t. (y)j.
J
u is given in the domain by Identity 81 and a by
a jk (y) = [r t S ~;~) (x} o ~ (x)] dr (x) +jr

u
D(jk} (x)
(y)
(84)
o [!: (x)] dr (x)
A FEW REMARKS AS A CONCLUSION
The four fundamental identities we have written must be regar-

ded as formal identities. Actually, u (€ nor n'), a (€ nor
n'), ulr• tlr belong to suitable vectorial spaces and r -
integrals are dualities between trace spaces. Basic elements
can be found in [Nedelec, J. C. (1977)}.
Identity 60, for example, must not be written in a point

where t is not continuous, on the given form, but in an
"integrated form", using a duality.
In this way, a variational formulation with an associated

bilinear form may be written, i~ the place of the boundary
i_dentities between [u], [t] and."{;(ulu' uiEr)and I or between
[u], [t] and.'f(t IIr, tl Er).
Using the identities in the previously written form, the
discretisation leads us to a boundary collocation method (BCM).
Link equations must be added [Chaudonneret, M. (1978)]. Nume-
rical examples of plane elastostatic problems may be seen in
[Pignole, M. (1980)].
Using a variational formulation, it is logical to use

9Pundary finite elemen~s (BF~M). For theoretical approach see
LNedelec, J. C. (1977)j and LBaldino, R. R. (1979 and 1980)}.
84
ACKNOWLEDGEMENTS
For the helpful theoretical discussions, the author greatly

thanks Professor Y. Haugazeau.
REFERENCES
Baldino, R. R. (1979 and 1980) An integral equation solution

of the mixed problem for the Laplacian in R3 • Part I and 2.
Rapports internes 48 et 54, Centre de Math. Appl., Ecole
Polytechnique, Palaiseau - France.
Banerjee, P. K. (1976) Integral equation methods for analysis

of piece - wise non - homogeneous three dimensional elastic
solids of arbitrary shape. Int. J. Mech. Sci., 18: 293-303.
Brebbia, C. A. and Walker, S. (1978) Introduction to boundary

element methods. Recent advances in boundary element methods,
Pentech Press, Plymouth and London, 293-303.
Chaudonneret, M. (1978) Calcul des concentrations de con-

traintes en elastoviscoplasticite. Publication O.N.E.R.A.
1978-1, Chatillon- France.
Cruse, T. A. (1977) Mathematical foundations of the boundary-

integral equation method in solid mechanics. AFOSR-TR 77-1002,
Pratt and Whitney Aircraft Group, United Technologies Corpo-
ration, East Hartford, Connecticut 06108.
Hartmann, F. (1980) The complementary problem of fli1ite elas-

tic bodies. New developments in boundary element methods, CML
Publications, Southampton, 229-246.
Lachat, J. C. (1975) A further development of the boundary

integral technique for elastostatics. Thesis, University of
Southampton.
Nedelec, J. C. (1977) Cours de l'ecole d'ete d'analyse nume-

rique CEA-IRIA-EDF.
Pignole, M. (1980) Une methode integrale directe pour la reso-

lution de problemes d'elasticite. Applications aux problemes
plans. These, Universite de Bordeaux I.
Schwartz, L. (1973) Theorie des distributions. Hermann, Paris.

85
REGULAR BOUNDARY INTEGRAL EQUATIONS FOR STRESS ANALYSIS
c. Patterson and M.A. Sheikh
Dept. of Mechanical Eng., University of Sheffield, U.K.
INTRODUCTION
THE boundary integral method is now well established as a

general numerical technique available for the solution of
field problems. In contrast with the finite element method
freedoms need only.be defined on the boundary of the domain
of the problem. Once these are determined the solution
within the domain is obtained using appropriate surface
integrals of the boundary solution.
Central to the method is the generation of Boundary

Integral Equations which properly state the problem to be
solved in terms of unknown field functions on the boundary
only. These equations are usually obtained using the Fundam-
ental Solution of the given problem with the singular point
located on the boundary. (The equations for the interior
solution are obtained similarly, by locating the singular
point within the domain of the problem). There ensues an
infinite system of singular surface equations, one for each
boundary point (being generated by moving the singularity
around the boundary) . The system is discretized by defining
boundary elements, after the manner of finite elements, and
the resulting finite system of singular integrals are evalua-
ted, thereby giving a system of algebraic equations.
Two discomforting features are apparent in this, convent-

ional, approach. Firstly, accurate evaluation of the singular
integrals requires special and careful treatment in the neigh-
bourhood of the singular point. Secondly, the class of prob-
lems for which the method is well defined may be unduly rest-
trictive because of divergence of the equations.
In this paper it is shown that 'Regular Boundary Integral

Equations can quite readily be derived which also properly
state the given problem. These are obtained by the simple
device of moving the singularity of the fundamental solution
86
outside the domain of the problem. The resulting system of

equations tolerates higher order singularities in the solution
than previously and requires no special attention to a singular
integrand.
The practicality of the method is demonstrated in two-

dimensional elastostatics. A critical comparison is made of
the results obtained using the new and conventional approaches
for constant and linear elements.
THEORY
The governing equations of elastostatics in terms of the stress

field and in the absence of body forces can be written as:
3 aa ..
l: ~J 0 i 1, 21 3 (1)
'1~
J= J
The stresses and strains are related by the constitutive

relation for an isotropic body as:
auk
a .. A.o .. + 2]..1£ij ( 2)
~J ~J <lxk
where au. au.

~ J
£ ..
~J ox. + ox.
J ~
and o .. is the unit tensor known as kronecker delta and A., ]..1
are ~J Lame's constants given as:
A. = E I ( 1 + v) ( 1 - 2v) ; ll
where E is the Modulus of Elasticity and v is the Poisson's

Ratio.
The fundamental solution which satisfies the homogenous

equation of elasticity theory (Equation.l) is:
u~.
~J -
_ 1
l61T].l(l-a)
[ (3-4a) 1
r
oij + (xi-yi) (xj-~ '] (3)
r
where the so called Poisson's Coefficient a= A.j2(]..1 +A.) and r

is the distance between the points x andy i.e. the point of
application of load to the point under consideration.
Let the displacement field u.(y) be twice continuously

differentiable in some finite regfon Q and continuously
differentiable in the closed region IT= Q + S. We choose a
point x' inside Q and remove from Q a ball of radius £ with
center at'x~ A~plying the third Betti formula [Mikhlin,l965]
to u. (y) and u .. (y) in the region Q with the ball removed
J ~J
87
and letting E + 0, we get
(4)
If equation (1) is to be satisfied the first integral

on the right hand side of equation (4) vanishes and we get,
u.
~
(x)
I
s
u~~,].(x,y)t.(y)ds
J y
-Jt~.(x,y)u.(y)ds
~J
s
J y (5)
Equation (5) is known as Somigliana's identity which in

simplified notations take the form,
ci ui = Is u*tds - Is t* uds (7)
This equation can be used to determine the interior solution

once the boundary integral equation is solved. However, if
the singularity of the fundamental solution is located out-
side the domain U, r does not vanish in Q so that ci = o and
so,
tds = uds
(8)
Equation (8) forms the basis of 'Regular Boundary

Integral Method' in elastostatics.
If the given boundary conditions are:

u=u on s1
where the total boundary is S=S~, we can write an extended

weighted residual equation of the form
Jut~s + I ut*ds J tu*ds + J tu*ds (9)

sl s2 sl s2
For the two dimensional case, equation (9) gives a

88
system of equations, two for each singular point (one for

each direction) corresponding to the boundary node under the
boundary element discretization and located outside the
domain of the problem. Once this system of regular integral
equations is solved, interior solution can be generated
routinely, using equation (7).
Constant Boundary Elements

The boundary is discretized into N elements (say) and the
values of traction and displacement are assumed to be constant
on each element and equal to the value at the mid-node of the
element. The singular point corresponding to this mid-node
'i' is located at an arbitrary distance (taken as L/2 in our
case studies, where Lis the length of the element) from 'i'
and along the positive normal (Figure.!).
L
-= '\. Mid-node
L/2
Singular Point
~l+x2+y2-yl, xl-x2+yl+y2)
2 2
Fig.l. Location of Singular point for Constant element
Equation (8) for this singular point corresponding to the

boundary node 'i' becomes in discretized form,
N
~
(10)
j=l
The integrals in equation (10) can be evaluated numeric-

ally for all segments over the boundary including the one
containing the node 'i'. 2N such equations (one for each
direction at N points) are obtained and solved for the bound-
ary unknowns. Once the solution over the whole boundary is
obtained, interior solution can be generated using equation
(7) in which case ci equals unity.
Linear Elements
The variation of t and u is assumed to be linear within each
element. For N elements equation (8) can be written as
~
j=l
j u* tds
N
~
j=l (11)
Sj
The values of u and t at any point of an element can be

89
written in terms of their nodal values and interpolation

functions F 1 and F 2 as:
(12)
where s is the dimensionless coordinate s x (Figure.2).

L/2
(13)
The singular point corresponding to a node 'i' is located at

an arbitrary distance (taken as L/2 in the present study,
where Lis the length of the element) from 'i' and along the
positive normal (Figure.3).
Equation (11) for any singular point corresponding to

the boundary node 'i' can now be written in two parts (one
for each direction) as:
N
l:
j=l J [·, •,] u * ds { t1x }
t2x
N
l:
j=l
r [ F 1F 2
Js.
Jt*ds{ :!x}
2x
sj J
(14)
N
l:
j=l JC·,
s.
F,] u * ds {
t
t2y
!y I
f
N
l:
j=l
J [ .,., }·d~
sj
u1y }
u2y
J
A system of 2N 'Regular Boundary Integral Equations' can

now be obtained and solved as for constant elements.
Fig. 2. Linear Element

90
Node
'j+l'
Node
'j - 1'
Fig.3. Location of Singular Point for Linear Element
APPLICATIONS
TWO two dimensional stress concentration problems are analysed

using both the conventional and the proposed regular boundary
integral methods for constant and linear boundary elements
and a critical comparison of the results is made.
Notched Circular Hole in a Rectangular Sheet under

Uniform Tension
The problem was first analysed without notch at the circular
hole in the 150 mm X 110 mm rectangular plate as shown in
Figure,4. The radius of the hole was taken as 15 mm, i.e.
less than one-third of half width of the plate to produce high
stress concentration at point XX. Only a quarter of the domain
needs to be taken for analysis due to symmetry, (Figure.5).
Boundary conditions for the problem are:
y = 0
t 0 u along AB (i)
X
tX 120 N, ty = 0 along BC (ii)
t ty 0 along CD (iii)
X
u t 0 along DE (iv)
X y
t t 0 along EA (v)
X y
91
!x
~- .X
llOmm
l50mm -------t
Fig.4. Rect.Plate with a circular hole under uniform tension
D "--t =t =0
X y
u =t =0
X y
y
=0 B
Fig.5. Boundary conditions
for Quarter domain
Conditions u = 0 along DE and u = 0 along AB are specified

to avoid rig~d body motion. Theymaterial properties are
taken as E = 2.1 x 10 5 N/mm 2 and= 0.3.
The finite element and constant and linear boundary

element discretizations used for the solution of the problem
are shown in Figure. 6.,and the stress distribution along ED
thus obtained for conventional and regular integral methods
is compared against finite element results [Patterson and
Bagdatlioglu, 1981] in Figures.7a. and 7b., respectively for
constnnt and linear elements. Although, as expected, the
stress distribution obtained by using finite element method
is better, the results for conventional and regular boundary
integral methods both for constant and linear elements stay
marginally close.
The problem was then made to yield even higher stress

concentration by taking a semi-circular notch of radius 1.5 mm
at points XX of the plate as shown in Figure.8. Again, only a
quarter of the plate was considered with the boundary condit-
ions remaining unchanged (Figure.9). The discretizations
using finite element and constant and linear boundary elements
are shown in Figure. 10., and the stress distribution along ED
obtained for constant and linear elements using conventional
and regular methods is compared with finite element results
[Patterson and Bagdatlioglu, 1981] in Figures.lla. and llb.,
respectively. The stress levels at points XX is as much as
three times than in the previous case of circular hole
without notch. Again, the stress distribution given by finite
elements is better but it is very similar for the conventional
and regular methods in case of both constant and linear elements.
92
Fig.6C. Linear boundary element

discretization
Fig.6a. Finite Element

discretization
9.
Fig.6C. Linear boundary Element
8. discretization
ax 7.
6.
• • • Finite Element
Solution
5. Conventional
boundary elmt.
Solution
4.
Regular boundary
element solut-
3. ion
2.
1.
o.
15 30 45 60
Distance along ED
Fig.7a. Stress distribution along the edge ED of the domain
using constant boundary elements
93
8.
Solution
7.
Conventional
boundary elmt.
solution
ax ------- Regular boundary
6. element solut-
ion
5.
4.
3.
2.
1.
0.
15 30 45 60
-------<~Distance along
ED
Fig. 7b. Stress distribution along the edge ED of the domain

using Linear Boundary elements
94
~--------15omm--------~
Fig.B. Rect.Plate with a notched
circular hole under uniform D
t =O
c
tension y
u =t =0
X y t =120N
X
Fig.9. Boundary Conditions for B

Quarter domain
A Lug loaded by Free Fitting Pin
This is an actual design problem which has already been
analysed both using photoelastic testing and as a finite
element contact problem [Patterson, Wearing and Arnell, 1980].
The geometry of the Lug is described in Figure.l2. Due to
symmetry only half of the domain is considered and the bound-
ary conditions are such that all the nodes on the axis of
symmetry are restrained in y direction and the nodes on the
edge EA are restrained in x direction as shown in Figure.l3.
The load distribution (Figure.l4) that the pin puts on the lug
was calculated by the finite element program [Patterson,
Wearing and Arnell, 1980] and corresponds to 16000 lbs load on
the pin of diameter 1.25 inches and diametrical clearance of
0.001 inches. The properties of the lug material are
E = 26.5 x l0 6 lbsjin 2 and poisson's ratio = 0.3.
Fig.l2. Geometry of the Lug

95
Fig.lOb. Constant boundary elmt.

discretization
Fig.lOa. Finite Element

discretization
Fig.lOc. Linear boundary element

discretization
••--~•~--••Finite elmt.solution
---------- conventional bound-

ary element solution
Regular boundary
element solution
15 30 45 60
--------~ Distance along
ED
Fig.lla. Stress distribution along the edge ED of the domain
96
24.
21.
0
X
18.
Solution
Conventio nal
boundary element solution
15. ------- Regular boundary
solution
12.
9.
6.
3.
o.
15 30 45 60
------~~istance a1ong
ED
Fig.llb. Stress distributi on along the edge ED of the domain

using linear boundary elements
97
The boundary of the domain of the problem was discretized

using 66 constant elements (Figure.l5a) and 71 linear elements
(Figure.lSb) in which case two nodes (one corresponding to
each side) were taken close to each other at all the five
corners of the boundary to avoid any freedom constraint.
Comparison of surface stress distribution calculated by

using conventional boundary integral method with that obtained
from regular boundary integral method is shown in Figures.l6a.
and 16h, respectively for constant and linear boundary element
discretizations. The stresses are calculated at various
internal points along the critical section X- Xwhere maximum
stress occurs and the stress distribution is shown in Figures.
17a. and 17b., for constant and linear elements respectively.
It is clear that although the boundary solution obtained

by using conventional and regular integral methods shows a
little difference, the solution in the interior is in almost
complete agreement. Moreover, it should be noted that the
results obtained for the constant and linear boundary element
meshes used in the present s.tudy are less accurate than the
ones ob-tained by employing photoelastic testing or a refined
finite element mesh, [Patterson, Wearing and Arnell, 1980],
however, they have been found to be much improved by refining
these meshes. But despite of the incompleteness of the meshes
used, the main purpose of comparing the conventional boundary
integral method with the proposed regular boundary integral
method for two different types of elements is well served.
Fig.l3. Boundary Conditions

(0
(X)
Scale lmm 50 Ths-f
Fig. 14. Nodal Load distribution acting on bore turface at 16,000 lbs.f total pin load
99
..._-+---4----1-""'-·-·-·-·-·-·-J.--·-·-·-·-·-·-'-,......-+--~
!
Fig. lSa. Constant boundary elemant discretization
..______.,._ ·-·-·-. -·-· +--·-. -· -· _, __....______,

I
Fig. lSb. Linear boundary element~ discretization

Conventional B.I.M. 8
lmm 2 tons/in 2
Regular B. I. M.
Compression
I
I II ~x
Fig. 16a. Comparison of surface stress distributio~ calculated by using conventional and
Regular boundary integral method for conttant elements.
Conventional B.I.M. 2 tons/in 2
Regular B.I.M.
I jl •X
Fig. l6b. Comparison of surface stress distribut~on cala1lated by using Conventional and regular
boundary integral methods for linear e~ements
~
102
80.
l
60.
Stress
variation
along
XX' 40.
20.
0. Radial distance from inner boundary

along XX'
-10. Fig. 17a. Stress variation along

X critical section XX' using
constant boundary elements
x'
80.
60.
Stress
variation
alonq XX'
40.
Conventional B.I.M.
1 20.
Regular B.I.M.
Radial distance from

0.
Fig. 1 7b. Stress variation along critic 1 1

X section XX' using linear elernenys X
103
DISCUSSION AND CONCLUSIONS

THE test problems analysed were chosen to present a menu of
ascending difficulty to the numerical methods leading to a
typical design problem. The boundary element meshes were
taken to correspond reasonably to what were known satisfactory
finite element meshes but, otherwise, convergence studies
were not regarded as a central issue of the study.
The simplest test was the plate perforated with a cir-

cular hole. Here the normal stresses along the loa& bearing
ligament, Figs.7a,7b, as given by the singular and regular
methods for constant and linear elements were both in good
mutual agreement and agreed well with the finite element
result. It is noted that the boundary element stresses are
consistently below the finite element values. The effect of
adding the notch to the circular hole is to enhance the stress
concentration by about two and a half times. Again, the
boundary element results all are in good qualitative agree-
ment with the finite element values, Figs.lla,llb, and again
they lie beneath. The free fitting pin in lug problem arose
as a live design problem due to unsuspected fatigue failure
of the lug. Earlier photoelastic and contact-finite element
analyses had revealed a severe stress concentration on the
inner face of the lug at about 106° behind the centre load
point. The contact analysis revealed the surface load dis-
tribution given in Fig.l4. This load distribution was used
in the boundary element analyses giving the surface stress
values indicated in Figs.l6a,l6b, for constant and linear
elements respectively. For each element type the boundary
method results are in good mutual quantitative agreement and
show good qualitative agreement with previous values. The
constant elements do not show the well developed compressive
region on the outer face and the tensile zone ahead of the
pin, on the outer surface, varies significantly from the
linear element results. Previous values support the linear
element values. On the inner surface the strong stress con-
centration is manifest in all cases. The linear elements
again give better results than the constant variants. Using
linear elements the stress concentration is better described
by the conventional method being located at about 106°.
Nevertheless, with the meshes used the discrepancy is in the
order of a single element and both methods underestimate the
finite element/photoelastic values. The interior stresses
normal to the critical section XX' are given in Figs.l7a,l7b.
This section is taken at 92° for constant elements and 97°
for linear; to correspond with peak surface stress given by
the regular method. The computed values are as expected,
given the disrepancies in surface values already seen.
The problems analysed show that under similar conditions

the regular and singular methods yield similar results.
~fuilst the perforated plate problem did not discriminate
constant and linear elements the pin-lug problem yielded
104
better results with linear elements. The meshes used were

suggested by finite element experience and gave values which
were generally marginally low in value but trends were well
presented. The likely reason for this is inadequate mesh
refinement which was supported by plate analyses with refined
meshes.
In conclusion: An alternative to the traditional, sing-

ular boundary element method has been presented which gives
regular ke.rnel functions in the boundary element equations.
In the test problems analysed results computed using the
singular and regular methods were in close mutual agreement
and in good general agreement with known values. The advan-
tages uf the reql!l.lar method are that the familiar singular
kernel of the usual method is avoided and that higher singul-
arities can be tolerated in the solution without leading to a
divergent system.
REFERENCES
Mikhlin, S.G. (1965) The Problem of the Minimum of a

Quadratic Functional, Holden-Day, San Fransisco.
Patterson, C., Wearing, J.L. and Arnell, P.P. (1980). A Study

of stress distribution in lugs loaded by free fitting pins.
In: Stress Analysis in Fabricated Structures, Eds.C.Patterson
and T.H. Richards, Applied Science Publishers, London.
Patterson, C. and Bagdatlioglu, N.I. (To be published) A

Nesting Technique for Efficient Determination of Localized
Stress Concentrations. In: Mathematics of Finite Elements
and Applications, Ed. J.R. vfuiteman, Academic Press, London.
105
A BOUNDARY ELEMENT FORMULATION OF PROBLEMS.IN LINEAR ISOTROPIC

ELASTICITY WITH BODY FORCES
D.J. Danson
Computational Mechanics Centre, Southampton, U.K.
I. INTRODUCTION
Traditional formulations of the Boundary Element Method (BEM)app-

lied to elastostatia{l] [2] are extremely convenient when the
loading on the body under analysis is limited to surface load-
ing, since it is necessary to discretize only the boundary of
the body and not the whole domain as must be done when using a
technique such as the Finite Element Method (FEM). When body
forces are present these have usually been handled by evaluat-
ing a domain integral r3J. Unfortunately this requires the
domain of the problem to be divided into integration cells
since for any practical problem the domain integral must be
evaluated numerically. This greatly increases the amount of
data preparation required and causes the BEM to lose much of
its advantage over domain t~e methods. However, Cruse r4J
and Cruse, Snow and Wilson ~1 have shown that for certa1n
types of commonly encountered body forces the domain integral
may be transformed to a boundary integral or boundary integrals
which may be evaluated at the same time as the boundary integrals
involving the surface displacements and tractions. Ref. Is]
is concerned exclusively with axisymmetric geometry; however,
the authors' use of the Galerkin vector to achieve the required
transformation from domain to boundary integrals provides the
key to the present paper which relates to two and three
dimensional geometry. The three dimensional case was derived
in Ref. lj] without resort to the Galerkin vector. However, the
Galerkin vector approach is used below both to demonstrate the
power of the technique and to present the results of [4] in a
slightly more general form. The two dimensional formulation is
also presented.
The following commonly encountered body forces will be con-

sidered.
(i) The body force due to a body being placed in a constant

gravitational field.
106
(ii) The body force due to constant rigi.d body rotation

about a fixed axis.
(iii) Thermal stresses may be calculated by applying a

body force proportional to the temperature gradient
["7]. The domain integral resulting from this body
force may be transformed to a boundary integral
provided that the temperature distribution is a
"steady state" one.
2. BOUNDARY INTEGRAL FORMULATION
The Somigliana identity for the displacements inside an elastic

body may be derived using Maxwell-Betti's reciprocal theorem
to give (Ref, [I])
~(x) J uki(x,y)ti(y)dSY-
s
I
s
Tki(x,y)ui(y)dSY
+ J uk.(x,y)b.(y)dV
:L :L y
(I)
v
Where ~(x) is the displacement at an internal point x in
the k direction
uki(x,y) is the displacement in the i direction at y due
to a unit point load in the k direction at x
Tki(x,y) is the traction in the i direction at y due to
a unit point load in the k direction at x
t.(y) is the traction at y
u:(y) is the displacement at y
b~(y) is the intensity of the body force at y
s:L represents the boundary of body
V represents the domain of the body
The coordinates of the point x are Xj
The coordinates of the point y are yj
The subscript y on dSy, dVy is to indicate that the
coordinates y. and not the coordinates x. are the
integration ~riables. J
Applying Maxwell-Betti's reciprocal theorem to the thermo-

elastic case gives (Ref. [4])
~(x) Tk. (x,y) u. (y)dS

:L :L y
s s
aE r
+ I-2v J Uki,i (x,y) 6(y)dVY (2)
v
107
where a is the coefficient of linear expansion

E is Young's modulus
vis Poisson's ratio
8(y) is the temperature at y
,i denotes differentiation with respect to y.
l
Calling the domain integral in both Eqns (I) and (2) Bk (x),
so that for the gravitational or rotational cases
J uk.(x,y)
l
b.(y)dV
l y
(3)
v
and for the thermoelastic case
Bk(x) = -1
aE
- 2v
Jr Uk l.,.l (x,y) 8(y)dVy (4)
v
the aim of the following is to express Bk(x) as a boundary

integral instead of a domain integral.
A generalized form of Gauss' theorem will be used which may

be expressed as follows.
If A.kn is a general Cartesian tensor field then

J "···
J A.k
J JL ••
. dV
,l y J AjkJI.. • • n. (y)dS
l y
(5)
v s
where n.(y) is the unit outward normal at y.
l
3. TRANSFORMATION TO BOUNDARY INTEGRALS
We define a tensor ~i(x,y), called the Galerkin tensor

such that the displacement kernel Uki(x,y) is given by
Uk1. (x,y) = Gk l,JJ

. . . (x,y) - Jz1~ -v )
Gk. . . (x,y)
(6)
It is worth noting that in the same way that Uk.(x,y) may

be regarded as three (in 3D) or two (in 2D) displac~ment
vectors each corresponding to the direction k in which the unit
load is applied, similarly Gki(x,y) may be regarded as three
(in 3D) or two (in 2D) Galerkin vectors each corresponding to
the direction k in which the unit load is applied.
Substituting Eq. (6) into Eq. ( 3) gives

108
r Gk . . . (x,y))
J lGki,jj (x,y ) -
J 'J ~
2(1-v)
b ( ) dV
i y y (7)
v
Although Eq. (7) looks a lot more complicated than Eq. (3)
it is much easier to arrange Eq. (7) in a form where it can be
transformed to a boundary integral.
eravitational loads (se1f weight)

A body of constant mass density ~n a constant gravitational
field experiences a constant body force b.(y).
~
Thus b.(y) may be taken outside the integral in Eq. (7) and
so using ~Eq. (5) we get
dS (8)
y
which is the required boundary integral form.
Rotational Inertia
Let us consider a body of uniform mass density p rotating about
some axis with angular velocity wi. We may without loss of
generality consider the axis of rotation as passing through the
origin of our coordinate system. The acceleration at a point
yi is given in vector notation as
f = W X (W X y) (9)
By d'Alembert's principle we may replace this acceleration

by a body force b and treat the dynamic problem as a static one,
hence
b = - p W X (W X y) (I 0)
where p is the mass density of the body.
We may write Eq. (10) in tensor notation as
(I I)
where c .. k is the permutation tensor.

~J
Defining
(12)
109
it is easily shown that
w~+w~ -wlw2 -w3wl
.. J
[n ~J = p -WIW2 w~+wf -w2w3 (13)
-w3wl -w2w3 wf+w~
and Eq. (II) becomes
b. (y) n.. y. (14)

~ ~J J
Substitution of Eq. (14) into Eq. (7) gives
Gk . (x,y)}
Bk(x) = !1 .. ·J( y. {G_ . (x,y) - m,~ dV (15)
~J J -k~ ,mm 2 (1-v) y
v
which after some manipulation becomes
.. Jr oy
n~J -J-- {Y.J Gk.
~,m
(x,y)} - Gk . . (x,y)
~.J
V m
(16)
This is now in a form which can be transformed to a boundary

integral using Eq. (5) and using the symmetry of !&ij to effect
some slight simplification we have
~ .(x,y)}
~(x) = 11 ij J yj {~i,m(x,y) - ~(~-v) nm
s
dS ( 17)
y
Steady state thermal loading

Different~at~ng Eq. (6) w~th respect to yi and substituting in
Eq. (4) gives
Bk (x) = Z(~~v) J 6(y) G_ . . . . (x,y)dV

-k~.~JJ y
( 18)
v
110
For steady state heat conduction 6 .. (y) 0 and we may

write ,JJ
Bk(x) = 2(a!E)
-v I 6(y)G . . . . (x,y)- G . . (x,y) e .. (y)dV
-k~.~JJ ik~,~ ,JJ y
(19)
v
Manipulating as before gives
--f-
yj
{ck . . (x,y)
~.~
e . (y)} dV
,j y
(20)
which is now in a form which can be transformed to a boundary

integral using Eq. (5).
Thus
Bk(x) = 2(~~v) I {e(y)Gki,ij(x,y)

s
- Gk . . (x,y) e .(y)} n. dS (21)

~.~ ,j J y
4. 2D BODY FORCES
The Galerkin Tensor corresponding to the 2D Fundamental Solution

is
_ I+v
~i(x,y)- 4TIE 0ki r
2
ln rI (22)
where r is the distance between the force point x and the field
point y
and oki is the Kronecker delta.
Differentiating Eq. (22) twice, performing the appropriate

contractions and substituting into Eq. (6) gives the Fundamental
Solution
uki(x,y)
111
where r. is the unit vector from the force point x to the field
point ~ so that----
y. - x.
r. J J
J r
We note that Eq. (23) differs from the fundamental solution

usually given by a constant. This is not of any great importance
as a constant is equivalent only to rigid body motion.
Gravitational Loads (2D)

For gravitational loads we differentiate Eq. (22) and substitute
into Eq. (8) to get
B ( ) )+V J (2 ln (24)
k x = 4nE r r
s
Differentiating Eq. (I) (with the boundary integral I'eplacing the

domain integral) and substitution in the stress displacement
relations gives the stress at an internal point x. Thus
a .. (x)
l.J
s s
+ S*.. (x,y)dS (25)

J l.J y
s
where Dk .. (x,y) and sk .. (x,y) are given in Ref.

l.J l.J [3] and
*.. (x,y)
s l.J 2 ~" [l(l -ln.!_) {n r (b.r. + b.r. + -1v b r o.. )
r mm L J J l. -v s s l.J
- bmrm (n r + n.r.)}+ - -1- - [ln.!_-.!_)

2(1-v) i j J 1. 2(1-v) r 2 )
}] (26)
Rotational Inertia (2D)

For rotational loads we note that to keep the problem two
dimensional the axis of rotation must be at right angles to the
plane containing the problem. Thus if the angular speed of
rotation is w then from Eq. (13)
112
11 •• p w2 o.. (27)
~J ~J
Substitutio n of Eqns (27) and (22) into Eq. (17) gives
Bk (x,y)
l-2v I }
- 2(1-v) r ln r ~ dSY
(28)
The boundary integral equation for stresses is once again Eq.(25)

with this time
*
sij (x,y) =
pw2
811 LV
r oij
{2nmrmysrs
1-v
+ (2 ln..!..- I) (
r
3- 2v n y +2r n r
(l-v)(l-2v) m m mm
J)}
ymrm
+ 2n r (y.r. + r.y.) - ---(n.r. + r.n.)
mm ~ J ~ J I -v ~ J ~ J
I I
2 ( 1-v ) (2 ln--
+ ------- 1){(3-2v)(n .y. + n.y.)
J ~
r ~ J
+ (l-2v)r(n.r . + n.r.)}J (29)

~ J J ~
Steady state thermal loading (2D)

For steady state thermal loads we write Eq. (21) as
J Pk(x,y) 9(y)dSY - J Qk(x,y) e ,m (y)nm dS y (30)
s s
where
a.E
Pk(x,y) -(l-)
= -2 Gk . . . (x,y)n.J
-v ~.~J
(31)
and a.E
Qk(x,y) 2(1-v) Gki,i (x,y)
113
Differentiating Eq. (22) and substituting in Eq. (31) gives
a.( l+v) { }
Pk (x,y) (ln ..!. - ..!.) n. - n r r
4 'It ( 1-v) r 2 k. mmk
(32)
a. (I +v)
Qk(x,y) = 4n(l-v) r
k
r (ln ..!.r - ..!.)
2
Differentiating Eq.(4) (with the boundary integrals replacing

the domain integral) and substition in the stress displacement
relations gives
oij(x) = J Dkij(x,y) tk(y)dSY- I skij(x,y) ~(y)dSY

s s
+ J s~.(x,y)
~J
e(y)dS
y
- J v~.(x,y)
~J
e
,m
(y) n (y)ds
m y
s s
a.E
- -1 2 e(x)o .. (33)
- v ~J
where Dk .. (x,y) and Sk .. (x,y) are as before and

~J ~J
S~.(x,y)
~J
= 4 n (~E)
-v r mm - v
0
{n r ( 1i 2j - 2r.r.) + n.r. + n.r.}
~ J ~ J J ~
(34)
5. 3D BODY FORCES
The Galerkin Tensor corresponding to the 3D Fundamental Solution

is
(l+v)R oki
Gki(x,y) 4nE (35)
where R is the distance between the force point x and the field
point y. The symbol R is used when discussing the 3D case so
that equations involving the 3D case are easily distinguished
from the 2D case where the symbol r is used.
Differentiating Eq. (35) twice, performing the appropriate

contractions and substituting into Eq. (6) gives the fundamental
solution
114
Gravitational loads (3D)

For gravitational loads we differentiate Eq. (35) and substitute
into Eq. (8) to get
I +v r dS
Bk(x,y) (37)
411E J y
s
The boundary integral equation for stress is Eq. (25) where

for the 3D case Dk .. (x,y) and Sk .. (x,y) are given in Ref. [3]
and S~.(x,y) is ootlined in a 1 Jsimilar manner as for the 2D
case, 1 J whence
S~.(x,y)
1J
= -8-R
11
rn r (b.r.
L mm 1 J
+ b.r.)
J 1
J=vtf vo 1J.. (nmr mb s r s

+ 1 11 - b n )
mm
( 38)
Rotational inertia (3D)

Differentiation of Eq.(35) and substitution into Eq.(J7) gives
~(x,y) = ~;~ f Gkjy/mnm- 2(l~v) {riQi/j~

s
+ (l-2v)RS1k. y.rn ))] dS (39)
J J mm y
The boundary integral equation for stresses is once again

Eq. (25) with
S:j (x,y) 1
= 811 [{nsrs y: + ~(~~v)nJ(Qimrj + Qjmri)
ym ym
+ vo 1J
.. (rks:<km -R n s r s - n s Qsm -R + r s Qsmnm)
ym r:
- 2R (r Q l.!l.r. + n.r.J + D-2v' rn.Q. + nj 11 iml) }]
s sm 1 J J 1 ~ ~1 JID
(40)
115
Steady State thermal loading (3D)

For steady state thermal loading Eqns (30) and (31) apply in 3D
as in 2D.
Differentia ting Eq. (35) and substitutin g in Eqns (31)

yields
and (41)
The boundary integral expression for stresses is Eq. (33)

where this time
s~.(x,y)
l.J
= aE
B1r( 1-v)R2
{n r (~-
m m 1-2v
3r.r.) + n.r.+r.n.}
1. J 1. J 1. J
(42)
6. DIRECT COMPUTATION OF 3D BODY FORCE KERNELS
Cruse [4] has indicated how to transform the domain integral

due to body forces to a boundary integral directly, without
involving the Galerkin Tensor.
Gravitation al loads
We have that
(I +v) b. r (3-4v)oki rkri

l.
Bk (x) 87TE ( 1-v) + - - dV (43)
J R R y
v
ark 0ki - rkri
Noting that ay.- R (44)
l.
this may be written

(1+v)bi J 4(1-~)oki ark
Bk(x) = 87TE(1-v) - ay: dVY (45)
v
We now make use of the following property of the Laplacian
operator on the three dimensional function R(x,y)
2
R (46)
,mm R
116
And remembering that R r we may write R r

and so Eq.(45) becomes ,m m ,mm m,m
(4 7)
This may now be reduced to an equivalent boundary integral by

using Eq. (5)
rkn.b.
Bk(x) = ~;~ f bk nmrm - --~-~-
2 ( 1-v)
dS
y
(48)
s
It should be noted that the integrand (or kernel) of Eq.
(47) is not the same as that derived previously in Eq. (37).
The second term is different and clearly b r ~ is not in
mml<
general equal to bmnmrk.
However it is necessary only that the integral of the

kernel round a closed boundary has a unique value.
Consider the integral
Ik = f bmrm~ - bmnmrk dSY (49)

s
Transforming to a domain integral using Eq. (5) this becomes
bm f r
m,k - r k,m dVy =0 (50)
v
Thus we have shown that both Eq. (37) and Eq. (48) are valid
integral transformations of the body force domain integral. In
other words the form of the body force kernel is not unique.
We note that if we had written
(51)
instead of Eq. (44) and followed the reasoning through as before

we would have arrived at Eq. (37).
The equivalent stress equation kernel S*

.. (x,y) is obtained
by differentiating Eq. (48) and substituting~in the stress dis-
placement relations to get
117
2vb r o.. }
S*
.. (x,y) n r { s s ~J
G + b + b
~J mm l-2v irj jri
Comparison with Eq. (38) shows that the form of S*

.. (x,y) is
not unique either. ~J
Rotational inertia
A similar process of algebraic manipulation enables the body
force kernels to be computed directly without using the Galerkin
tensor. The results are, for the displacement equation
Bk (x,y) 1+v
4nE
J {nmrmstkJ• YJ• - R st
'km m
n
s
(53)
and for the stress equation
s~.~J = -4
7T
r[1-2\1v {nmr ml(rs st ~-
sq R
2(stss)) + 2._ n st
1-v
ys
1-v m ms R
+ r
s
st n }
sm m
..
o~J +2!Ifn r (r. st. YRs + r. st. YRsJ
2\
m m ~ JS J ~s
ym ym)
- [n.~ st.Jm -R + n.J st.~m -R + r.~ stJ·m nm + r.J st.~m nm
+_II (n st
-\) m ms
YRs
~~~J
ro .. - r.r:]-
~ J.
_21st J:.r.
mmr~ J
+ nJ.rJJ }]
~
(54)
Once again comparison with Eqns (39) and (40) indicates

that the form of the body force kernels is not unique.
Steady state thermal loads

Cruse [4] has derived in detail the 3D body force kernels for
the steady state thermal case without making use of the Galerkin
tensor. This time the resulting kernels are identical to those
deri Vl'!d using the Galerkin Tensor.
118
7. EXAMPLES
To verify all the above theory, a fairly considerable test

programme is necessary. We have three different types of
body forces each in two different geometries (2D and 3D). For
the 3D geometry we have two different versions of the body
force kernels to test. For all these different cases both
the displacement and the stress equation kernels need testing.
Time has not permitted a comprehensive test of every kernel.
However, the ability to handle body forces without recourse to
integration over the domain is gradually being introduced into
BEASY which is a general boundary element system for the solu-
tion of problems in potential theory and in linear isotropic
elasticity in 2D, axisymmetric and 3D geometries. The tests
on the body force kernels which have been carried out have all
used constant boundary elements.
Example I. Cylinder under thermal stresses

This example consists of a long hollow cylinder with the inner
surface heated to J00°C and the outer surface held at 0°C. The
details of the problem are given below
Inner radius 30 mm
Outer radius 80 mm
Coefficient of Linear Expansion JQ X J0-6 /°C
Young 1 s Modulus 210 kN/mm2
Poisson 1 s Ratio 0.3
The infinitely long cylinder was modelled by a qu~rter

cylinder of length 50 mm with an appropriate normal displace-
ment boundary condition applied at either end to model the
infinite length. A solution to this problem is given in Ref.
[6] and Figure I plots values of hoop and axial stresses for
the analytical solution compared with those obtained using
BEASY 1 s 3D constant element. A distinction is made between
stresses obtained in the boundary solution which tests the
kernels in the displacement equation, Eqns (41), and the
internal point solution which tests the kernels in the stress
equation, Eqns (42). The same problem was run using BEASY 1 s
axisymmetric constant element which implements the theory
given by Cruse, Snow and Wilson [s]. The figures show good
agreement with the analytical sofution except for an internal
point close to the boundary. This last error is due to
inaccuracies in the integration scheme employed. It is notice-
able that the axisymmetric element gives better results than
the 3D element. This is no surprise as one of the integrations
has been done analytically in the axisymmetric scheme.
Example 2. Rotating Disc as a 2D problem

The details of this example are:-
119
100
(IIDII)
0
40 80
(a) Axial Stresses

-100
-----Analytical solution
-200
6 Boundary results } 3D element
0 Internal results
• AxisJ~~D~~etric element
100
radius (IIDII)
60 80
(b) Hoop Stresses

-100
-200
Figure Cylinder under Thermal Load

120
Analytical solution
6 Boundary results
20
10
0 20 40 60 80
radius (nnn)
Figure 2 Hoop Stresses in Rotating Disc (2D)
--- Analytical solution

6 Boundary results
o Internal results
20
0
10
0 20 40 60 80
radius (nnn)
Figure 3 Hoop Stresses in Rotating Disc (3D)

121
Radius 80 nm
Mass Density 7 • 8 tonne s /m 3
Rotation Speed 10,000 r.~.m.
Young's Modulus 210 kN/nm
Poisson's Ratio 0.3
An analytical solution to this problem is given in Ref. [6].

Figure 2 plots the values of the hoop stresses for the analyt-
ical solution compared with those obtained using BEASY's 2D
constant element. Only the boundary solution is plotted as
the kernel for the stress equation has not yet been incorpora-
ted into BEASY. Thus the plotted points are a test of Eq. (28)
only. Reasonable agreement has been obtained between the
Boundary Element and analytical solutions.
Example 3. Rotating Disc as a 3D ~ro~le~

The details of this example are:-
Radius 78.2 nm
Thickness 50 nm
Mass Density 7. 8 tonnes/m3
Rotation Speed 10,000 r.p.m.
Young's Modulus 210 kN/nm 2
Poisson's Ratio 0.3
Once again Ref. [6] supplies an analytical solution to this

problem. Figure 3 plots the hoop stresses for the analytical
solution along a radius which meets the axis of the cylinder
at its centre (z = 0). The boundary element results were
obtained using BEASY's 3D constant element. The boundary
results plotted are actually obtained at various axial posi-
tions (-25 < z < 25). However,inspection of the analytical
solution shows that the axial variation of the hoop stresses
is very small. The internal results were obtained on the
plane z = 0, The boundary and internal results verify Eqns.
(53) and (54) respectively. As before, reasonable agreement
between the analytical and boundary element solutions has been
obtained.
8. CONCLUSIONS
The ability to handle three of the most commonly encountered

body forces by transforming the body force domain integral to
an equivalent boundary integral greatly increases the attract-
iveness of the Boundary Element Method for use in stress
analysis. The limited tests of the method outlined above are
sufficient to verify the validity of the method, although a
complete test programme to test all the kernels presented here
is still necessary. The disagreements obtained between
analytic solutions and solutions obtained from BEASY are of
similar magnitude to those obtained for problems without body
forces and are due to
122
i) The approximations inherent in the use of constant

elements.
ii) The relatively simple numerical integration schemes
at present used in BEASY.
Linear and quadratic elements are in the process of being

introduced into BEASY and much work has also been done to
develop sophisticated integration schemes to handle the singular
nature of the displacement and traction kernels. Preliminary
results indicate that these developments will turn the Boundary
Element Method into an extremely accurate analysis tool.
REFERENCES
I. T.A. Cruse, Numerical solutions in three-dimensional

elastostatics, Int. J. Solids Struct. 5, 1259-1274 (1969).
2. T.A. Cruse, An improved boundary-integral equation method

for three-dimensional elastic stress analysis. Comput.
Struct. 4, 741-754 (1974).
3. C.A. Brebbia, The Boundary Element Method for Engineers.

Pentech Press 1978.
4. T.A. Cruse, Boundary-integral equation method for three-

dimensional elastic fracture mechanics analysis. AFOSR-TR-
75-D813 May 1975.
5. T.A. Cruse, D.W. Snow and R.B. Wilson. Numerical Solutions

in Axisymmetric Elasticity. Computers and Structures 1977
pp.445.
6. S.P. Timoshenko and J.N. Goodier, Theory of Elasticity

Third Edition, McGraw-Hill, 1970.
7. A.E.H. Love, A Treatise on the Mathematical Theory of

Elasticity. Fourth Edition, 1926.
123
A COMPARATIVE STUDY OF SEVERAL BOUNDARY ELEMENTS IN ELASTICITY
M.F.Seabra Pereira, C.A.Mota Soares and L.M.Oliveira Faria
CEMUL-Centro de Mecan i ca e Materia is da Un i vers i dade Teen i ca de

Lisboa, lnstituto Superior Tecnico, Av. Rovisco Pais
1096 Lisboa Codex - PORTUGAL
INTRODUCTION
Although many elasticity problems have been solved by

both analytical and numerical techniques, solutions for actual
problems involving complex geometries are still relatively
scarce.
Forpractical structural geometries, numerical methods are
advantageous, but have some limitations in representing high
stress gradients.
A general purpose Boundary Element program has been
developed for elasticity problems, in which different elements
can be easily implemented.
In this paper several boundary element solutions for
elasticity problems are compared with finite element resultsand
analytical solutions. The advantages and disadvantages of the
boundary element method and of the different elements used are
discussed with reference to the applications. Also, the principal
features of the computer program are described.
ANALYTICAL BACKGROUND
For the sake of notation and compleeness it is recalled

that the boundary element method is based on the Somigl iana's
boundary identity, as presented by Brebbia and Walker (1980),
c .. (P)u.(P)+Ju.(Q)T .. (P,Q)ds = Jt {Q)U .. (P,Q)ds +
IJ I s J IJ S j IJ
+ fv b . ( q) u.. (P , Q) dV ( q ) ( 1)
J IJ
where P,Q are points of the surface s of the domain V. Uij and
Tij are the fundamental Kelvin solutions, Cij are coefficients
depending on the geometry of the boundary at the point P, and q
is a point in the domain. Tensor notation is used in this
expression and the indices have the range 1 ,2,3. When body forces
124
are not considered equation (1) is only dependent on the suriace

displacements and tractions, Ui and ti respectively.
The surface can be discretized into elements with a
particular shape and a certain number of nodes, sayn. Any field
variable within each element is assumed to be given by
8(~) = Ma(S)ea a=1,2, ... n
or (2)
e(~.n) = Ma(s,n)ea a=1,2, ... n
for 20 and 3D cases respectively,Ma are shape functions of local
coordinates ~and n,ea are the values of field variables at nodal
points. lf the surface displacements and tractions are introdu-
ced as field variables equation (1) becomes,for 20 problems
c .. (P )u.(P) + ~l J Ma(~)T .. (P ,Q(~))J(~)d~uJr;
1J n 1 n k= 1 s k 1J n
(3)
n,t
= l: f Ma(S)U .. (P ,Q(~))J(~)d~tJ
k=l sk IJ n
where n,t is the number of boundary elements
uc:t is the value of uj at local node a
J.
tg is the value of tj at loca 1 node a
J
J is the Jacobian of transformation of coordinates.
Noting that Pn refers to a particular node, then for all nodes
equations (3) can be expressed in matrix form as follows:
[c]{U} + [Hj{U} [G) {T} (4)
or
[H]{U} [G] {T} (5)
where {U} and {T} are the surface nodal displacements and
tractions respectively. The elements of [RJ and [G] can be
obtained from the integrals
Hna J Ma(S)T.. (P ,Q(~) )J (~)d~ (6)
ij Sk IJ n
G~C: Ma(~)U .. (P ,Q(~))J(~)d~

f (7)
I SkJ IJ n
These integrals are normally integrated by standard Gaussian
quadrature. When these integrands are singular they are
calculated by rigid body considerations or by the technique
described in the Appendix.
Applying the boundary conditions, equation (5) becomes
[A] {x} = {f} (8)
where {x} and {f} are the vectors of unknown and known
quantities respectively.
The displacements at interior points can be calculated
using the following expression
125
u. (P ) nfe f Ma.(~. )U .. (P ,Q(~) )J (~)d~ ta.J.

1 n k=l S k IJ n
I
(9)
nft J Ma.(~.)T .. (P ,Q(O)J(~)d~ a.
k= 1 Sk I IJ n Uj
and the stresses at interior points are given by

a .. (P) = ~i (f Ma.DfiJ·J(~)d~ta.- Js Ma.sfiJ.J(~)d~u~) (10)
IJ n k=1 Sk f k .[
where~· -and s •.. are given by Brebbia and Walker (1980).
-<.11 -t-IJ
BRIEF DESCRIPTION OF THE PROGRAM

Basically, the computer program consists of four main
subroutines and the element subroutines. A brief flow-chart of
the program is illustrated in Figure 1.
Subroutine BEM 001
This subroutine reads all the input data and sets up all
the relevant array dimensions for the analysis.
Subroutine BEM 010
This subroutine performs the overall assembly of matrix
[A] and vector {f}. This subroutine calls the element subrouti-
nes which calculate the integrals of expressions (6) and (7)
for the boundary element with reference to a particular node.
These element subroutines can be called from the element library
or supplied by the user.
Subroutine BEM 030
This subroutine is an equation solver, using the Gaussian
elimination technique.
Subroutine BEM 020
This subroutine calculates the displacements and stresses
at internal points according to equations (9) and (10). This
subroutine calls the element subroutines, which calculate the
integrals of equations (9) and (10). Again, these element
subroutines can be called from the element library or supplied
by the user.
Element Subroutines
For each boundary element three subroutines have to be
developed. The first subroutine calculates all the necessary
information for each element (Jacobian, normals,etc.) at the
Gaussian points. The second subroutinecalculates the integrals
of expression (6) and (7) for the boundary element with
reference to a particular boundary node or internal point. The
third subroutine calculates the integral of expressions(9) and
(10) for the boundary element with reference to a particular
point.
126
H.. - coefficients of matrix

IJ
[H]
G.. - coefficients of matrix [G]
IJ
Sij' Dij -See expression (10)
SUB1, SUB2, SUB3- Element subroutines which are in
library or supplied by the user
Figure 1. Primary flow-chart of program BEM

127
Element Library
The following boundary elements have been developed:
- constant, 2D element
- linear, 2D element
-parabolic, isoparametric, 2D element
- triangular, constant, 3D element
-quadrilateral, isoparametric, linear, 3D element
- quadrilateral, isoparametric, parabolic, 3D element
APPLICATIONS
The different test cases which have been chosen have, in
general, severe stress gradients either because of the geometry
or due to the type of loading.
Different boundary discretizations have been idealized and
the following criteria were adopted: for each case a number m of
nodal points is specified, yielding m/2 parabolic elements or m
linear elements. Coarser meshes have been defined from the
previous cases by simply supressing every other node.
The corresponding meshes for finite element analysis have
been such that the outside nodes coincide with the boundary
element nodes. In the finite element analysis, the isoparametric
parabolic element for plain strain/stress problems has been
used.
The first three case studies are 2D analyses and the last
case is a 3D example. All the analyses were performed in double
precision on an IBM 360/44 with a core memory of 256 kbytes.
Hole in a flat plate
The geometry of the problem is illustrated in figure 2a.
·In the present case the ratio a/b is 5. A constant stress Oo
is applied to the plate at faces x=±a. Figure 2b represents the
mesh used in the finite element analysis (FEM). Due to symmetry
only one quarter of the structure has been modelled.
In Figure 3 the values of the o stresses along the yaxis
are compared with the analytical sol~~ion of the infinite plate
case. The solution corrected for the finite width of the present
proportionsyields a value of 3.93 for the stress concentration
factor Kt = oxx/o 0 •
In Table 1 a comparison of results is done jointly with an
indication of the computer times involved in all analyses. A
good measure of the accuracy of the various analyses is made on
the basis of the o stresses at the point A near the hole. The
discrepancy of thexx results is evaluated in comparison with
the value Kt = 3.93
128
4--L-~------------~--+
O"o
a l Hole in a square b) FEM mesh
flat plate
Figure 2
..Qu. ~Roark and Young ( 1975 l

O"o
4. DOF ( degrees of freedom l
o- BEM parabolic 120
• - BEM parabolic 56
c - BEM linear 120
3. • - BEM linear 56
' - FE M parabolic 314
- Analytical solution- infinite plate
2.
1.
•
c
1. 2. 3. 4. 5.
Figure 3. O"xx stresses near the hole

129
Table 1. Comparison of results for the hole in a flat plate

case
ANALYSIS DOF Discrepancy CPU

% (seconds)
BEM Parabolic 120 1.3 437
BEM Parabolic 56 4,7 91
BEM Linear 120 - 345
BEM Linear 56 - 68
FEM Parabolic 314 13 342
The discrepancies for the linear elements arenot indicated,

as the corresponding analysis yielded rather poor results. This
fact is associated with the limitations of the BEM in modelling
properly the stress fields at corner regions. The results show
that the linear elements are more sensitive to this effect.
However, in regions away from the corners all the analyses show
a reasonable agreement.
In all cases a relatively small number of degrees of
freedom have been used. The discrepancy of 4.7% using the para-
bolic element is very satisfactory regarding the rather coarse
discretization of the boundary.
Circular disc compressed by two diametrically opposed forces
The geometry and loading assumed in this example is
shown in Figure 4a and the FEM idealization is represented in
figure 4b. Due to symmetry only one half of the structure has
been considered in the BEM idealization.
Again four analyses have been made using parabolic and
1 inear boundary elements with two different refinements of the
boundary. In Figure 5 a graph of Oyy stresses along the diameter
perpendicular to the direction of loading is indicated. The
results are compared with an analytical solution of Timoshenko
and Goodier (1970).
In Table 2 a comparison of results, is presented. The
computer times involved in each analysis is also indicated. A
measure of the accuracy of the results is done by comparing the
maximum stress Oyy for x=O for the different analyses with the
analytical solution.
Stresses at interior points have been calculated in the
vicinity of the loading point A. The distribution of compressi-
ve stresses in a small neigho.ourhood of the load point approaches
that predicted by Flamant~ problem which is governed by the
expression Oyy=-2P/Tid where P is the applied load and d is the
distance from the load point to an interior point. In figure 6
the results obtained from the BEM and the FEM analyses are
compared with the analytical solution .The typical element size
near the point of application of the load is about 0.01D where
D is the diameter of the disc. This element size is small enough
130
y p
p
X
a) Circular disc b) FEM mesh
Figure 4
<JYyD
2P DOF
o - BEM parabolic 120
• - BEM parabolic 60
0.8 c - BEM linear 120
• - BEM linear 60
• - FEM parabolic 344
0.6 - Analitical solution
0.4
[]
0.2 •
0.1 0.2 0.3 0.4 o.5 x1 0

(-0.16 +•)
+[]
-0.37
Figure 5. Stress distribution in the
diameter perpendicular to loading
131
to allow the evaluation of stresses at interior points relative

ly close to the surface. The distances to which the stresses are
referred are always larger than 5 times the typical element size
in that region.
Table 2. Comparison of results for the disc case
ANALYSIS DOF Discrepancy cpu

% (seconds)
BEM Linear 120 19,0 350
BEM Linear 60 17 77
FEM Parabolic 344 4 100
In this figure it can be observed the inability of the

FEM to model such severe stress gradients. The FEM results near
the point load clearly depart from the analytical solution as
well as from the BEM values. Again the parabolic boundary
elements behave quite satisfactory and with an error of less
than 1% for distances greater than 0.05 D from the point load.
Compact Tension Specimen (CTS)

Another example analysed is a CTS with an edge crack as
illustrated in figure 7. In this case H=W and the crack length
A is .25W. Due to symmetry only one half of the structure is
idealised. Because of the singular stress and strain fields in
the vicinity of the crack tip only parabolic boundary elements
have been used.
In both, BEM and FEM discretizations elements near the
crack tip have an average size of about 0.05 of the crack
length. Plane strain conditions were assumed.
Two methods of deriving the values of crack tip stress
intensity factors, K1 , have been used, firstly, the extrapola-
tion of computed displacements near the crack tip, and second,
an energy method based on the Fracture Mechanics plain strain
relation G 1 =K~(l-v2)JE, where G1 is the strain energy release
rate, K1 is the mode I stress intensity factor and E and v are
the Elasticity Modulus and the Poisson's ratio respectively.
In this case the energy available for extension of the
crack length dA is provided by the work done by the nodal point
tractions applied in the top faces of the specimen. Variations
of crack length dA of 0.05 times the initial crack length have
been considered. A significant improvement on the results is
obtained by displacing in the BEM and FEM elements, themid-side
node of the near tip elements to 1/4 position as described by
Henshell and Shaw (1975).
The present results have been compared with a solution of
K1/Ko=1.5 presented by Rooke and Cartwright (1976). Table 3 shows
132
the results and the CPU times of the various analyses

Table 3. Comparison of results for the CTS
Discrepancy CPU
ANALYSIS DOF Kl /Ko % (seconds)
( 1) BEM Parabolic 120 1 '51 0,67 415
( 1) BEM Parabol i c 60 1 ,49 0,67 103
(2) BEM Parabolic 120 1,495 0,33 850
( 1) FEM Parabolic 530 1,48 1 '3 922
K =aliA; a is the applied remote stress.
0
(1) Extrapolation method
(2) Energy Method
The value of the stress intensity factor obtained with the
energy method involves, at the present stage, the execution of
two analyses with two different crack lengths, thus increasing
considerably the computing costs.
Beam subjected to an axial force

The BEM discretization of the beam is illustrated in
Figure 8. Two analyses have been carried out using linear and
quadratic elements. The results for the linear element have
errors of 2.8% and 3.6% in the axial stresses and maximum
displacements respectively. The corresponding errors for the
parabolic element are 20.2% and 3.5%.
CONCLUSIONS
A general purpose Boundary Element program has been deve-
loped using several 2D and 3D boundary elements.
Results for particular elastostatics situations without
body forces have been presented and the usage of different
element types have been considered.
In general, the results show that the parabolic 2D element
is much more efficient than the 2D linear element. The linear
boundary element is particularly sensitive to corner regions
where the stress results tend to degradate considerably.
The relative computing costs for the various analyses
have been shown in the Tables. In situations of severe stress
gradients, particularly in the vicinity of point loads the BEM
method appears to offer a reasonable accuracy, of the order of
5%, for relatively modest computing facilities.
FEM analyses with similar refinement on the boundaries to
the BEM discretizations involve rather larger numbers ofnodes
without any benefit in the accuracy of the results, yet more
information is generally provided about the stress and displa-
cement fields on the domain.
133
O"yy D ;1614 p
2P1
5.
DOF
o BEM parabolic 120
• BEM parabolic 60
c BEM linear 120
• BEM linear 60
10. X
A FEM parabolic 344
-Analytical solution
Timoshenko and
Goodier ( 1970)
5.
0.05 0.10 0.15 0.20 0.25 dJo
Figure 6. O"yy stresses at interior points
w b) linear
Figure 7. C T Specimen Figure 8. 30 Beam

134
REFERENCES
Brebbia, C.A. and Walker,S. 1980, Boundary Element Techniques in
Engineering. Newnes - Butterworths, London.
Cristescu,M. and Loubignac,G. 1978, Gaussian Quadrature Formulas
for Functions with Singularities in 1/R Over Triangles and
Quadrangles. Recent Advances in Boundary Elements Methods,
Editor Brebbia,C.A., Pentech Press, London.
Fairweather,G., Rizzo,F.J. and Shippy,D.J. 1979, On the Numeri-
cal Solution of Two - Dimensional Potential Problems by an
Improved Boundary Integral Equation Method, J.Comp. Phys., Vol.
31 ' pg. 96-112.
Henshell ,R.D. and Shaw,K.G. 1975, Crack Tip Finite Elements Are
Unnecessary. Int. J. Num.Meth. in Engng. Vol. 9, p 495.
Roark, R.J. and Young,W. 1975, Formulas for Stress and Strain,
5th Ed. MacGraw-Hill, Tokyo.
Rooke,D.P. and Cartwright,D.J. 1976, Compendium of Stress
Intensity Factors. Her Majesty Stationery Officer, London.
Saada,A.S. 1974, Elasticity, Theory and Applications. Pergamon
Press, Nel"i York.
Stroud,A.H. and Secrest,D. 1966, Gaussian Quadrature Formulas,
Prentice Hall, Int. Inc., London.
Timoshenko and Goodier 1970, Theory of Elasticity, McGraw-Hill,
Tokyo.
135
APPENDIX: INTEGRATION OF KERNEL-SHAPE FUNCTION PRODUCTS

For 2D problems the evaluation of the singular integrals in
equation (7) leads to integrals of the following type
<~ln7f(i;)di; (11)
where r is the distance between the points P and Q and f(i;) is
a product of the shape functions and the Jacobian of transforma
tion of coordinates. This integral can be transformed as follows
! 1 ln l f(i;)di; = ;+ 1 ln( 12+i;)f(i;)di; + 2~1 ln ~ f(c;)dc; (12)
-1 r -1 r "
where c; = 21 (/;+1)
The first integral on the right-hand side is evaluated
using standard Gaussian quadrature while the second integral is
calculated numerically by formulas given by Stroud and Secrest
( 1965).
For 3D problems the integrals corresponding to equation (7)
present a 1/r singularity and are integrated using a formula of
the following type
1 1 1 m
f f
-1 -1
- f ( i; , n) di; dn
r
= l: Ai f ( i; i , n i )
i= 1
( 1 3)
The weights A. and the integration point coordinates i;.

and TJ. depend on the geometry of the region of integration ~nd
m is 1 the number of integration points. These points can be
calculated following the method presented by Cristescu and
Loubignac (1978). For 3D linear and parabolic quadrilateral
elements Gauss formulas have been developed to integrateexactly
polynomials of 5th order multiplied by the weight function
1/~ in the regions illustrated in Figure 9.
~
~ REGION A 2 REGION B
2~o------,
0 2 r
Figure 9. Regions of integration
-1 0 •1 r
If the singularity is located in a corner node or in ami~
-side node, the integration region is transformed in the squares
A orB, respectively, by applying simple linear transformation
techniques to the domain of integration.
The integral (13) can be transformed as follows
136
ff - -1- - 1~ +n
2 2 f(~.n)d~ dn ( 1 4)
~ r
where the function (/~ 2 +n 2 /r) f(Cn) is suitably represented by
polynomials.
The values of ~. and n. have been found by calculating the
common roots of orthbgonal polynomials in ~ and n. Using a least
square technique the corresponding weights Ai have been evalua-
ted. The Gauss point coordinates and the corresponding weights
are indicated in Table 4.
Table 4. Gauss points and corresponding weights

REGION A
REGION B
~.
I I n.I I A.
I
1 -0.74772761 0.84529169 0.49133682
2 0.0 0.87119449 0.84924859
3 0.74772761 0.84529169 0.49133682
4 -0.77589199 1. 7321786 0.19429025
5 0.0 1.71964092 0.34350839
6 0.77589199 1. 7321786 0.19429025
7 -0.72502542 0.15427398 0.36769444
8 0.0 0.14090886 1.5127182
9 0.72502542 0.15427398 0.36769444
137
NON-CONFORMING BOUNDARY ELEMENTS FOR STRESS ANALYSIS
Dept. of Mechanical Eng., University of Sheffield, U.K.
INTRODUCTION
IN the boundary element method the familiar infinite system

of integral equations is discretized by the introduction of
boundary elements. As with Finite Elements one can define a
heirarchy of elements here ranging from constant through
linear, quadratic etc.
Hitherto, in the case of linear and higher order

elements, interelement continuity of the unknown functions
has been imposed in applications, undoubtedly in analogy with
finite elements practice. This continuity gives rise to at
least three problems [Lachat,l975]. First, at a point where
surface is not smooth the normal is not defined but the
freedom there (traction say) demands a valid normal. Second,
at an interface where there is a change in the nature of
boundary conditions (say between displacement and traction)
apparently both types of freedom are constrained. Third,
when the problem is partitioned into sub-regions, a conven-
ient device on grounds of computational efficiency, there
can be excessive constraint where several surfaces meet.
Accommodation to the second and third problems have been dev-
ised by appropriate suppression of degrees of freedom and to
the first by introduction of two freedom nodes close to the
geometric singularity, on either side. The last approach is
objectionable since if the nodes are not closely spaced the
boundary integrals are not well discretized, while if they
are the resulting algebraic equations are indefinite. This
is because proximity of freedom nodes implies linear depen-
dence in the algebraic equations. The root of the problem is
imposition of interelement continuity. In the Finite Element
Method adequate continuity is necessary to maintain positive
definiteness of the system [Patterson,l973]. In the Boundary
Element Method there is no such requirement; so that inter-
element continuity is simply not necessary. This point is
138
amply borne out by the apparent success of constant elements.
In this paper a selection of two dimensional elasto-

static test problems is analysed using conventional and non-
conforming boundary elements and the advantages of non-
conforming elements particularly at boundary discontinuities
are discussed.
THEORY
THE general boundary integral equation for two dimensional

elasticity observing interelement continuity and with sing-
ular point 'i' located on the boundary of the domain S, of
the given problem is:
N
i i
c u + 2: (1)
j=l
i
where c is the unknown coefficient at 'i'
N = Number of boundary elements in which the domain is
divided,
u and t are displacement and traction respectively,
u* is the fundamental solution for two dimensional
elasticity given as,
1 + v (xi-yi) (xfyj)}
u*
87TE(l-\l)r
~l (3-4\1) 6l.J
.. + (2)
r2
where \1 = Poisson's Ratio

E = Modulus of Elasticity
6 .. is the Kronecker delta
l.J
and r is the distance between points x and y, that is, the
point of application of the load to the point under consider-
ation.
Also, the Kernel t * for plane stress and plane strain

problem is:
t *
(3)
where n., n. are direction cosines and n is the normal to the

surface 1 of the body so that: 3r/3n = ns<xs- ys)/r
139
The displacements and tractions for a linear element

'j' (Figure.l.) are:
0
u
0
(4)
0
t
and the integrals in Equation (1) can be written as:
0
dS
:,] t'
(5)
Is.
[pl
0
0
dS
J
in which u 1 1 u 1 are the displacements at node 1 in x andy

directionsxrespectively 1
2 2
u 1 u are the displacements at node 2 in x and y
dire~tioXs 1
1 1
t 1 t are the tractions at node 1 in x and y
dire~tioXs I
140
t 2 , t 2 are the tractions at node 2 in x andy

direcdons~
F 1 and F 2 are the interpolation functions given by,
(1-~) /2 (1+~)/2 (6)
where ~ is the dimensionless coordinate ~ = x/L/2
The values of displacement and traction in x and y

directions at any point on the element are:
1 2
u (~) Fl u X + F2 u X
X
1 2
u (~) Fl u y + F2 u y
y
1 (7)
t t2
X
(~) = Fl t X + F2 X
1
t t2
y
(~) = F 1 t y + F2 y
For a node 'i' at the intersection of elements 'j-1' and

' j ' , Equation (1) can be written in two parts for simplicity
(for each direction) as:
N N
i i
c u +
X
E [h +h ]u = E [g +g ]t
j=l xi,j-1 xi,j xi j=l xi,j-1 xi,j X,~
N N (8)
i
c u i+ E [h +h ]u E [g +g ]t
y y, . yi j=l Yi,j-1 Yi,j yi
j=l y i,j-1 ~,J
where
h
xi,j-1
= JcF 1 t*Jx dS h
X,
~,j
JcF 2 t*Jx dS
s.J- 1 S,
J
g
xi,j-1
JcF 1 u*Jx dS gx.
~,j
f [F 2 u*]x dS
s.J- 1 s.
J (9)
h
yi 1 j -1
J[F 1 t*Jy dS h
y, . = f [F 2 t
*
Jy dS
s.J- 1
~,J
s.
J
g *
: J[F 1 u Jy dS gy, . = J[F 2 u*]y dS
Yi,j-1 ~,J
Sj-1 Sj
141
That is two equations are set up for each node (one for
each direction) taking into account the contributions made by
the adjoining elements. This approach gives rise to problems
mentioned earlier in the 'Introduction' section. Also, the
determination of ci in Equation (8) is cumbersome. To over-
come these problems 'Regular Boundary Integrals for Non-
Conforming Elements' are employed here. In this method, free-
domsof an element are not defined at the geometric nodes
i.e.~= ±1 but at~= ±1/2 (Figure.2). Also, the singular
point is relocated outside the domain and along positive nor-
mal to the boundary (taken as L/2 in the present case study
where Lis the length of the element) at any node 'i'
[Patterson and Sheikh,l981]. The resulting integrals are
then regular over the whole domain and no special treatment is
required for the evaluation of singular integrands (ci = 0) •
The interpolation functions take the form
I; + 1
Fig. 1. Linear Element
F l = (~-1/2) F 2 = (~+1/2) (10)
There are 2N freedom points defined over the whole boun-

dary of the domain of the problem and the equations for any
point 'i' similar to Equation (8) can be written as:
Fig. 2. Discontinuous Linear Element

142
N N
L: [h ]u L: [g ]t
X,
j=l xi,2j-l ~ j=l xi,2j-l X.~
N N
L: [h ]u L: [g Jt
j=l Yi,2j-l yi j=l Yi,2j-l yi
N N (11)
L: [h ]u L: [g ]t
j=l xi,2j X.~ j=l xi,2j X,~
N N
L: [h ]u L: [g ]t
j=l Yi,2j yi j=l Yi,2j yi
where
*
h ]X dS h
xi,2j-l J [Flt xi,2j
JcF 2 t*Jx ds
h
*
Yi,2j-l
f [F 1 t ]y dS h
Yi,2j
JcF 2 t*Jy dS
(12)
g *
xi,2j-l f [Flu ]X dS g
xi,2j
JcF 2u*]x dS
*
g
Yi,2j-l J [F 1u ]y dS g
Yi,2j f
*
[F 2 u ]y dS
Equations (12) produce a 4N x 4N system of equations

which can be solved to give the solution at the boundary.
Once the values of displacements and tractions in both dir-
ections are known over the whole boundary, the solution in the
interior can be generated routinely.
APPLICATIONS
THREE two dimensional elastostatic problems are analysed using

both the conventional and non-conforming linear elements and
a critical comparison of the results is made.
Uniform Cylinder under Internal Pressure

The exact solution of the problem is well known and given by
Lame's equations. The cylinder with inner radius r = a and
outer radius r = b is considered to have free ends and sub-
jected to uniform internal pressure P, with no pressure
143
applied on the outer surface. Only a quarter of the cylinder

is considered for analysis due to symmetry. The boundary
conditions are specified in a way to avoid rigid body motion
i.e. along AB zero displacement is prescribed in y direction
and CD is fixed in x direction, as shown in Figure.3. The
numerical values for the problem are assumed to be:
a = 1 em b = 2 em P = 1000 N/cm 2
The Modulus of Elast:icity is taken as 210 ,00000 N/cm 2 and

Poisson's Ratio as 0.3.
The boundary of the domain was divided into 26 linear

elements. No special care was taken in concentrating the
integration points around the singularities to show that this
is not a particular requirement when non-conforming elements
are used. The problem was solved using conventional boundary
integral method and regular boundary integral method for non-
conforming elements. Radial displacements were calculated at
five different radii across the thickness of the cylinder for
four different values of 8 = 0,10,30,45 degrees. As shown in
Table.l., the best results are obtained at 8 = 45° in the
case of conventional method whereas when regular integrals are
used over non-conforming elements, not only the values are
improved but there is no considerable variation in radial
displacement for different values of 8. This is because the
way boundary conditions are specified at corners A,B,C and D,
using conventional method. We can specify either traction or
displacement at A and B in X direction and at C and D in y
direction, depending on the side under consideration and one
generally chooses to prescribe displacement. No such problem
arises due to ambiguity in the specification of boundary
conditions in case of non-conforming elements being used.
Results for stresses in both cases at 8 = 45° are compared in
Figure.4.
~igid Die on Elastic Half Plane

A load Q is applied through a rigid flat die as shown in
Figure.5. The solution to the problem is discussed in
[Saada,l974]. The pressure distribution on the die is given
by,
Q
p (13)
where R = Half length of the die

x = distance from point of application of load Q
through the die, along OX direction.
Expression (13) gives p = Q/TIR as the pressure under

point of application of load i.e. at x = 0 and infinite
144
Fig. 3. Boundary Conditions and discretization co£ the

cylinder using conventional and non-conforming linear
elements
i
Ia r I I • •
- - - - • Exact Solution
i Conventional
loel Boundary Elmt.
Solution
Solution using
Non-Conforming
Linear Boundary
Elements
r=a r=b
Fig. 4. Comparison of Exact Solution with Solutions obtained
using Conventional and non-conforming linear elements
145
Table .1. Comparison of exact solution

with the calculated values of radial
displacement using conventional and
non-conforming linear boundary elmts.
Function Radius Exact Theta Conventional Non-Conf.

Sol. (6) Lin.Elmts Lin.Elmts.
(em) (X lo- 4 cm) (X l0- 4 cm) (X l0- 4 cm)
Radial
Disp. 1.1 0.872 00 0.801 0.825
10° 0.802 0.825
30° 0.810 0.828
45° 0.828 0.830
1.3 0.779 00 0. 712 o. 745

10° o. 716 0.745
30° 0.724 0. 746
45° 0.740 0. 750
1.5 0. 717 00 0.657 0.689

10° 0.660 0.690
30° 0.667 0.695
45° 0.681 0.696
1.7 0.674 00 0.620 0.651

10° 0.621 0.651
30° 0.627 0.652
45° 0.640 0.653
1.9 0.645 00 0.595 0.618

10° 0.598 0.618
30° 0.601 0.622
45° 0.612 0.625
146
pressure at x = R i.e. the edge of the die, which makes the

solution singular at that point.
Due to symmetry only half of the die is considered and

discretized as a part of a large rectangle (Figure.6.) such
that the edges AB and BC are sufficiently far from the die
to remain undeformed. The vertical displacement under the
die is constant and taken as unity, whereas zero lateral
displacement is prescribed at all points along OY. The
boundary of the rectangle is divided into 40 linear elements
taking a refined mesh along the die and its nearest bounda-
ries. The problem was analysed using both the conventional
and non-conforming linear elements. Figure.?., shows the
pressure distribution along the die obtained for the two cases.
Figure.8., displays the variation of horizontal and vertical
displacements along ODC. The clarity of these results at the
singular point D in case of non-conforming elements being
used, is to be noted.
Fig.S. Rigid Die on Elastic

Half Plane
Fig.6. Boundarv Element

DJ.scret1zatl.on
Stress Concentration in a Plate under Uniform Tension due to
a small Circular Hole
The plate with a circular hole in the center is subjected to
a uniform tensile load as shown in Figure.9. The diameter of
the hole is taken as less than one fifth of the width of the
plate to produce high stress concentration at points XX.
Only a quarter of the plate needs to be considered because of
symmetry, and is subjected to boundary conditions shown in
Figure.lO.
The problem was solved by dividing the boundary into 36

linear elements and using conventional and regular boundary
integral methods. In the latter case non-conforming elements
were used with appropriate boundary conditions specified at
the 'freedom points' of the element and not at the geometric
147
8.
• • • Exact Solution
Conventional Boundary Element

Solution
7. Non-conforming Boundary
Element Solution
6.
5.
4.
3.
2.
Fig.7. Pressure distribution along the die

CXI
-""'
•(' Conventional Linear Elements
'I~ ' \
0 \ 0
X
,o y
',Y Non-Conforming Linear Elements
', ------
''
'
''
....
''
I I
0
{'-....X
....., .............
' ',
"'~
~'
~ .....
II ..... .....,
I ...........
..... __ .....
I I - ' ..........
/ - ....... ----
.../ .
0 D c
Distance along the die
Fig. 8. Vertical and Lateral displacements along the die

149
nodes. The curves showin g stress distrib ution along

DE are
plotted in Figure .ll., for the two cases used.
-ct- .x
I
Fig. 9. Plate with a circul ar hole under uniform tension
D c
u t 0
y y
t (given)
y
t 0
y
A B
Fig.lO . Bounda ry Condit ions for quarte r domain
DISCUSSION AND CONCLUSIONS
THE main object ive in this study is to compar e the

behavi our
of conform ing (i.e. continu ous across an elemen t bounda
ry)
and non-co nformin g bounda ry elemen ts. The linear elemen
t in
two dimens ion was selecte d as the simple st case. Becaus
e
of modell ing diffic ulties at singul ar bounda ry points
(corner s) and at points where abrupt change s in bounda
ry con-
dition s occur it is expecte d that the solutio ns obtain
ed,
using conform ing elemen ts, will be poorer in the neighb
our-
hood of these anomal ous points than their counte rparts.
It is
of intere st to see if this degrad ation is local or
more wide-
spread . Althou gh a reason able quality of converg ence
was
sought in each case analyse d no attemp t was made to
assess
mesh densit ies require d for a solutio n of prescr ibed
accura cy.
On examin ing the calcul ated displac ements for the thick
cylind er problem , Table 1, it is eviden t that the non-co
nform-
ing elemen ts give results which respec t the axial symmet
ry
~
0
f-'·
Ul
rt Convent ional Linear Elements
§
(l
ro Non-Conf orming Linear Elements
~
1-'
0
::l
LO
ro
£'ro
t>j
0
0
H\
rt
::r
ro
'0
1-'
~
rt
ro =====-":..~=--= =~-::.-=-==,::== =-= ---
=== =-=--------=-= ...=..=.-=~-=-= ~-=--
--------;.~"':.----=-===-~-~..:....:~-= =--=---=---..: -:--::--: "'=-""":;.""::."="=:---- -
... -=-~~-."';;:' '=
= == =~~ ='= -==---..;:::-:=
E
a
y
Fig. 11. Stress distribu tion along edge ED of the plate
151
to around 1% whereas the conventional elements show a

pronounced degradation at the radial edge of the model which
is felt even on the symmetry line (45°) . The non-conforming
results are about 5% low on the inner and outer edges with
some improvement in the interior. The best conforming
results; at 45°, are only marginally lower than their counter-
parts but there is a further degradation of around 5% as the
radial boundary is approached. The degradation of the con-
tinous element results is due to the modelling problem at the
points where the radial edges meet the inner and outer edges
of the tube. On cutting the mesh at the corners and introd-
ucing two adjacent nodes, separated by 10% of element length,
at each side Of the corners the continuous elements gave
better results, comparable with the non-conforming elements.
Computed internal stresses, at 45°, are given in Fig.4, these
again give similar low values with worst performance on the
boundary.
The rigid die problem was selected as a severe test

because of the known singular normal stress at the edge of
the die with zero normal stress immediately beyond. The com-
puted and known pressure distribution across the die are
given in Fig.7. Away from the singular region both elements
agree with the exact solution adequately, within 5%. At the
edge of the die the continuous elements gave a normalized
pressure 3.1 while the non-conforming elements gave 4.9 at the
node nearest the edge. The exact value at that node is 7.1.
Clearly the non-conforming results are substantially better.
In Fig.B. are given the displacements at the surface under and
away from the die. The horizontal motions are qualitatively
satisfactory for both elements.
Neglecting the oscillation apparent at the edge of the die

in both cases, the vertical motions are qualitatively satis-
factory for the non-conforming elements but not for the con-
tinuous elements. A developing trough with a step at the die
is not correct. At the time the analyses were performed a
displacement solution was not known by the authors, however
one was subsequently found [Sneddon,l951]. This shows that
the oscillation at the die edge is a spurious numerical feat-
ure; this is not really surprising in view of the severe
singularity. Otherwise it confirms the non-conforming results.
The perforated plate problem is a routine test. Here

again, the non-conforming results are superior at a corner
and give a satisfactory stress concentration factor.
In conclusion: It has already been observed that inter-

element continuity with boundary eleMents gives problems at
geometric singularities, at interfaces between different
load types and, in subdomain models, where several surfaces
meet. In this paper it has been observed that interelement
152
continuity is not necessary with boundary elements. In the

test cases analyzed the continuous elements gave poorer
results where the model was ambiguous, the degradation being
felt well away from the anomaly. The non-conforming elements
did not suffer this defect and gave results which were gener-
ally acceptable.
REFERENCES
Lachat, J.C. (1975) A Further Development of the Boundary

Integral Technique for Elastostatics, Ph.D. Thesis,
Southampton University.
Patterson, C. (1973) Sufficient Conditions for Convergence

in the Finite Element Method for any Solution of Finite
Energy. In: Mathematics of Finite Elements and Applications,
Ed. J.R. vfuiteman, Academic Press, London.
Patterson, C and Sheikh, M.A. (In print) Regular Boundary

Integral Equations for Fluid Flow. In: Numerical Methods
in Laminar and Turbulent Flow, Eds. C. Taylor and
B.A. Schrefler, Pineridge Press, Swansea.
Saada, A.S. (1974) Elasticity: Theory and Applications,

Pergamon Press, New York.
Sneddon, I.N. (1951) Fourier Transforms, Mcgra1v Hill, New

Jersey.
153
THE DISPLACEMENT DISCONTINUITY METHOD IN THREE DIMENSIONS

W. Scott Dunbar 1 and Don L. Anderson 2
ABSTRACT
The displacement discontinuity method is a type of bound-
ary integral method wherein the fundamental solution used is
the static displacement due to a displacement discontinuity on
a finite segment in an infinite or semi-infinite medium. The
extension of the method to three-dimensional problems is pre-
sented together with some examples of its application. It is
shown that the use of the Volterra integral to derive the fun-
damental solution not only clearly identifies the method as a
standard boundary integral procedure with a different influence
function, but allows the development of several useful exten-
sions of the method in a straightforward manner, and also pro-
vides a natural method of linking the method to a finite ele-
ment computer code.
INTRODUCTION
The displacement discontinuity method (DDM) is a type of
boundary integral method wherein the fundamental solution used
is the displacement due to a constant displacement discontinu-
ity on a finite segment in an infinite or semi-infinite elastic
medium. Boundary value problems in mechanics are solved by
distributing a series of such displacement discontinuity seg-
ments over a boundary on which the displacement or stress is
known. The solution is found by adjusting the magnitude of
each displacement discontinuity to match the boundary condi-
tions.
The DDM has been developed in two dimensions by Crouch
[1973, 1976a,b]. Although it is generally applicable to most
elastic boundary value problems, the method has mainly been ap-
plied to geomechanical problems such as the analysis of stress-
es and displacements near underground excavations. In this
regard, one of the principal advantages of the method is its
ability to model cracks and faults in a realistic fashion. A
1senior Research Engineer, Weidlinger Associates, MenloPark,CA
2Professor, Department of Civil Engineering, University of
British Columbia, Vancouver, Canada
154
numerical advantage of the method is that there is no singular-

ity in displacement at the point of application of the boundary
force or stress. Rather, because of the assumption of a con-
stant displacement discontinuity, stress singularities occur at
the edges of the boundary segments.
This paper discusses the development of the DDM in three
dimensions. The present formulation is for infinite media, but
it will be shown that the extension to semi-infinite media is
relatively easy. The primary motivation for this development
is the potential application of a three-dimensional formulation
to problems in mining engineering.
DERIVATION OF FUNDAMENTAL SOLUTIONS
The fundamental solutions required for the DDM are the
displacements due to a constant displacement discontinuity over
an arbitrarily-oriented segment in a three-dimensional elastic
medium. The displacement discontinuities may be normal to the
plane (mode 1) or parallel to the plane (modes 2 and 3).
Many authors have derived such equations. Much of this
work has appeared in the geophysical literature [Steketee, 1958;
Chinnery, 1961; Maruyama, 1964, 1966; Press, 1965; Mansinha and
Smylie, 1971] and relates to faulting by shear failure. By
means of a Galerkin vector fomulation Rongved [1957] and Rong-
ved and Frasier [1958] derived the displacements due to con-
stant displacement discontinuities as an arbitrarily-oriented
rectangular segment in infinite and semi-infinite media. Berry
and Sales [1962] derived the displacements in infinite andsemi-
infinite transversely isotropic media for a constant mode 1
discontinuity over a rectangular segment parallel to one axis.
Chen [1964] completed the derivation for mode 2 and 3 discon-
tinuities.
For reasons which will become apparent later, the deriva-
tion of the displacements due to an arbitrarily-oriented rec-
tangular segment in infinite media will be done by means of the
Volterra integral [Hirth and Lothe, 1968, p. 93] rather than a
Galerkin vector formulation. The coordinate system used is
shown in Fig. 1. The origin of the local coordinate system
1 = (11 , l2, 13 ) is at the center of the rectangular segment.
The source coorainates c2 and c3 are coincident with 12
and l3 . c1 = 0 . The dimensions of the segment are -a $ c2
$ a , -b's c3 s b . The displacements u~ in the local co-
ordinate system 1 are given by
ui (l) =fa Jb
-a -b
155
(1)
where
dj(c) =the displacement discontinuity in the lj = cj
direct1on at c corresponding to a mode j displacement dis-
contuity;
nk(c) the kth component of the unit normal vector to the
segment at c;
U~(l,c) =the displacement in the li direction due to a
force 1n the lj direction at c ; i.e., the Green's function
of the medium; and
A,~= Lame's constant and the shear modulus, respectively.
The notation u~·~eans the derivative of U~ with respect to
ck . Expressio~s for U may be found in Love [1944, p. 185],
Mindlin and Cheng [1950] and Press [1965]. The displacement
discontinuity dj is reckoned positive if it acts in the posi-
tive cj direct1on on the side of the segment with the positive
normal n(c) For the planar segment n(c) = n = (1,0,0) .
To derive the fundamental equations for the DDM, the dis-

placement discontinuity d is assumed constant over the seg-
ment. The remaining part of the integrand in eq. (1) may then
be integrated analytically without too much algebraic diffi-
culty. The evaluation of the resulting function at the four
corners of the segment was done within a computer program. The
displacements in the local coordinate system may be transformed
to the global system g by means of the matrix equation
u = ctu~
where u is the vector of displacements in the global coordi-
nate system and C is the transformation matrix which trans-
forms a vector in the global coordinate system to one in the
local coordinate system. The superscript t denotes thematrix
transpose. Thus, the displacements at any point due to a dis-
placement discontinuity d on a segment may be written
3
u=Ud= L:u .. d. (2)
j=l lJ J
156
where Uij is the displacement in the gi direction due to

the mode j displacement discontinuity dj on the segment,
i.e., an influence coefficient.
By differentiation, the strains may be derived in the lo-
cal coordinate system and, by substitution into Hooke's law,
the stresses. If s~ is the stress tensor in the local co-
ordinate system, the stress tensor, S , in the global coordi-
nate system is given by
s = cts~c
Given S and the normal to any segment, the mutually ortho-
gonal tractions, t , on the segment due to a displacement dis-
continuity, d , on another segment may be computed by the
product
3
t = Td = L TiJ.dJ. (3)
j=l
where Ti · is the traction in the 1i direction due to the

mode j rlisplacement discontinuity dj on the segment.
For stress boundary conditions, the DDM is formulated as
follows: N displacement discontinuity segments are distri-
buted over a boundary on which the stress is known. The trac-
tion vector t(i) at the centroid of the ith boundary segment
due to a discontinuity d(j) at the jth boundary segment is
given by the superposition
N
t( i) = L T(i,j)d(j}
j=l
where T(i ,j) is the 3x3 matrix relating tractions at the ith
boundary segment to displacement discontinuities at the jth
segment. By writing a similar equation for 1 s 1 S N , a sys-
tem of equations results which may be solved for d(j) , 1 s j
s N . Given the latter, the displacements and stresses may be
computed anywhere in the medium using eqs. (2) and (3).
Displacement and mixed boundary conditions may also be
handled by selecting the appropriate parts of eqs. (2) and (3)
when constructing the system of equations. The details may be
found in Crouch [1976a,b]. The problems dealt with in the ex-
amples have stress boundary conditions only.
DEVELOPMENT OF COMPUTER PROGRAM
Based on the latter formulation, a computer program en-
titled DDM3 was written in FORTRAN to solve three-dimensional
stress boundary value problems. The central element of this
157
program is 27 function subroutines representing the nine dis-

placement derivatives for each of the three modes of displace-
ment discontinuity. Each routine is the FORTRAN rendition of
the result of integrating eq. (1) for a particular i and j
and differentiating with respect to a local coordinate lk .
The displacement derivative is computed by evaluating the func-
tion at each of the four corners of the segment. Each function
was tested by numerical differentiation of the appropriate one
of the nine displacement functions. The displacement functions
were tested by numerical integration of eq. (1).
The remainder of the program consists of subroutines to
compute the matrix of influence coefficients, an equation solv-
ing algorithm, and several matrix and vector multiplication
routines. A flow diagram is presented in Fig. 2. The program
is not in production form. However, none of the examples pre-
sented in the next section required more than five seconds of
execution time on the CDC 7600. The bulk of the computational
effort lies in the calculation and decomposition of the matrix
of influence coefficients.
One problem which immediately presented itself was the
large size of the system of equations to be solved. Since
there are three degrees of freedom on each segment, a system of
3N equations results. The system is typically fully populated
and unsymmetric. If the number of segments, N , approaches 40,
computer storage may be a problem and one must resort to out
of core equation solvers for such systems. One such algorithm
was developed by Hofmeister [1978] and was implemented in the
program.
Other devices to reduce the size of the system of equa-
tions were also implemented. For some problems, there exist
planes of symmetry. On either side of a plane of symmetry, the
shear displacement discontinuities will be equal in magnitude
but opposite in sign [Crouch, 1976b]. The normal displacement
discontinuities will be equal. Given the geometry and location
of the segments on one side of a plane of symmetry, the effect
of the segments on the other side may be incorporated by super-
position after multiplication by the appropriate sign. Thus,
the size of the system of equations may be reduced by a factor
of 2 for each plane of symmetry.
Depending on the boundary conditions, the size of the sys-
tem of equations may be reduced. For example, purely hydro-
static stress boundary conditions do not induce shear displace-
ment discontinuities. The members of the system of equations
corresponding to such discontinuities may thus be eliminated,
resulting in a significant saving of storage. This is particu-
larly relevant to deep mining problems where the boundary condi-
tions are almost hydrostatic.
By far, the most difficult problem is the visualization of
the three-dimensional model and the definition of its location
158
and geometry. The ideal solution would be to link the problem

to an interactive graphics system where a light pen would des-
cribe the geometry on a screen. A possibly less expensive al-
ternative would be to design a simple physical modeling system
with more angular flexibility than TinkerToy or Meccano.
EXAMPLES
The computer program was extensively tested and checked
against known analytical solutions for three- and two-dimen-
sional problems.
The simplest available three-dimensional problem, which is
amenable to an approximate discretization into rectangular seg-
ments, is that of a penny-shaped crack opened by a constant
pressure parallel to the crack axis. The discretization is
shown in Fig. 3a. Nineteen segments were used to model one-
quarter of the crack. The remainder of the crack was modeled
by symmetry conditions about the axes g2 = 0 and g3 = 0 .
Since the boundary condition is a normal pressure, there is
only a mode 1 displacement discontinuity at the crack. The
other two modes are zero. Consequently, the order of the sys-
tem of equations was reduced to 19 instead of 19x3 =57.
The radial and axial stresses along the g1 axis are
shown in Fig. 4a. They are compared with the analytical solu-
tion given by Sneddon [1946]. For this particular discretiza-
tion, the agreement near the crack surface is not particularly
good. As pointed out by Crouch [1976a, p. 307] in relation to
another problem, this is due to the assumption of a constant
displacement discontinuity over the segments. At larger dis-
tances from the crack, the agreement is within 1-2%, which is
more than adequate for most geomechanical applications.
The two-dimensional problem of a circular hole in an in-
finite medium under a uniaxial stress at infinity was also
analyzed by means of the three-dimensional DDM. The discreti-
zation is shown in Fig. 3b. In this case, the length of the
segments in the g3 direction was made very long in order to
simulate the two-dimensional plane strain conditions. Fifty
segments were used to discretize one-quarter of the hole, the
remainder being modeled by symmetry conditions about the axes
gl = 0 and g2 = 0 . To create a free surface at the boundary,
the boundary tractions were set equal to the negative of the
tractions applied at infinity. Since a plane strain condition
is simulated, there are no mode 3 displacement discontinuities.
The order of the system of equations was therefore reduced to
100 from 150, and could therefore be solved in core.
The numerical results along the gl axis are shown in
Fig. 4b where they are compared with the analytical solution
given in Jaeger and Cook [1976, p. 249]. The agreement in this
case is excellent, even near the boundary of the h9le. This
159
latter problem was used as a more stringent test of the coordi-

nate transformations used to calculate the stresses and trac-
tions.
As an example of the type of geomechanical problem which
may be solved by the three-dimensional DDM, the mining layout
shown in Fig. 5 was modeled. Two ore seams, known as the East
and West limbs, are to be mined by a series of 60 by 80 metre
panels, each separated by a 20 metre pillar. For reasons of
economic access to each limb, it is desirable to locate the
mineshaft at the location shown. The problem is to determine
values of displacements due to the mining operation along the
shaft for use in design of the shaft.
Each panel was modeled by 32 segments as shown in Fig. 5.
The initial stress on each panel was assumed to be hydrostatic
and related to depth by
The Young's modulus and Poisson's ratio of the medium were

assumed to be 60 GPa and 0.25, respectively. The coordinates
of the points a,b,c,d and e shown in Fig. 5 are sufficient to
define the location and geometry of the limbs and shaft. The
mining operation was modeled by applying the negative of the
initial (hydrostatic) tractions to each segment. This produced
a free surface at the panels. Each of the panels shown in
Fig. 5 was mined individually and the resulting displacements
along the shaft were stored on a disk file. The displacements
due to any sequence of panel mining were then calculated by
superposition.
The horizontal (ul and u3) displacements along the shaft
are shown in Figs. 6a and 6b. After mining the panels shown on
the East limb, the u1 displacements are very small and posi-
tive. However, they become larger and change sign once the
panels shown on the West limb are mined. The u3 displace-
ments are greater in magnitude than the u1 displacementsafter
mining the East limb and do not change appreciably after mining
the West limb. This may be expected considering the geometry
and sequence of mining. The vertical (u2) displacements were
relatively unaffected by the mining sequence.
One could continue mining the panels in any order and com-
pute the resulting displacements. However, the displacements
are nearly negligible and the induced stresses may be of more
importance to design.
160
FUTURE DEVELOPMENTS
The use of the Volterra integral readily allows a number
of developments of the theory which would be useful in appli-
cations. These are discussed below.
1. Extension to semi-infinite media
To derive the fundamental equations for the displacements
and stresses in a semi-infinite medium, the Green's functions
for such a medium would be substituted into eq. (1) and in-
tegrated, assuming a constant displacement discontinuity, as
before. The resulting expresssions, which are additive to
those of the infinite medium, could be incorporated into the
function subroutines of the computer program in a straightfor-
ward manner.
2. Integration over general quadrilaterals
With rectangular segments, it is difficult to model sur-
faces whose boundaries are not at right angles to one another.
An example is a boundary segmenf whose edge lies along the line
bd in Fig. 5. To model such a boundary, it would be necessary
to integrate eq. (1) over a general quadrilateral or a triangle.
By means of suitable transformations, a general quadri-
lateral may be transformed into a square region over which the
integration may be performed. Consider the quadrilateral shown
in Fig. 7 whose area is denoted R~ . The transformation to
R~ from the standard region R is
4
c2 = ~ hi (x2,x3)c2i = p(x2,x3)
1=1
(4)
where c2i and c3i are the local coordinates of the corners
of the quadrilateral and the functions hi are given by
h3 = t(1-x 2) (I-x 3)
[Bathe and Wilson, 1976, p. 131]. An integral of a function
f over R~ is then given by
where J(xz,x3) is the Jacobian of the transformation (4).

Substitution of the above transformation in eq. (1) results in
161
a very complicated integral which must be evaluated numerically.

However, the use of Gaussian integration of order two along the
x2 and x3 axes requires a total of four function evaluations,
which is the same as that required to evaluate the displacement
or its derivative due to a rectangular segment. Consequently,
the time required to evaluate a displacement or its derivative
due to a general quadrilateral segment should not be much
greater than that required for a rectangular segment. An ex-
periment was carried out to test this and the times required
were nearly equal. Order 2 Gaussian integration is probably
not sufficient for complete accuracy, considering the order of
the function to be evaluated (which is approximately six), but
is likely to be sufficient for geomechanical problems. This is
currently under study.
3. Representation of the displacement discontinuity on segment
In some applications, such as those involving fracture, it
may be desirable to represent the displacement discontinuity,
d , by some function other than a constant. A suitable repre-
sentation would avoid stress singularities at the edges of a
segment. If the coordinate transformation functions, h , used
in the last section are also used to interpolate d in terms
of its values at the corners of a segment when integrating
eq. (1), then a linear isoparametric boundary segment would re-
sult. Compatibility at the edges between segments would be
guaranteed by such a formulation [Bathe and Wilson, 1976,
p. 144]. However, it is possible that the amount of numerical
integration required may be prohibitively expensive. Lachat
and Watson [1976] discussed a successful implementation of such
a procedure in a conventional boundary integral code.
4. Linkage with finite elements
In the finite element analysis of geomechanical problems,
it is often necessary to model fractures or, more commonly,
boundaries at infinity. Conversely, there may be portions of a
model which are not amenable to analysis by the DDM because of
inhomogeneities or inelastic behavior. The ideal solution to
this problem is to find a method of coupling a DDM analysis to
a finite element analysis. A method of accomplishing this was
outlined by Zienkiewicz [1977] and Zienkiewicz et al [1977] in
terms of conventional boundary integral procedures. Their ob-
jectives were to derive a symmetric stiffness matrix of a bound-
ary-type element which could be assembled into the conventional
finite element stiffness matrix in a standard fashion. The
following outlines a very similar development for the DDM.
Referring to Fig. 8, the potential energy of the region H
may be written
(5)
162
where u and t are the displacements due to the displacement

discontinuity segments along the boundary B and G are the
displacements along B due to the finite element mesh. The
latter displacements are given in terms of the nodal displace-
ments w by the finite element shape function N :
u = Nw
Substituting this and eqs. (2) and (3) into eq. (5) and mini-
mizing with respect to d results in
or
(6)
Minimizing with respect to w results in
Kt d = 0 (7)
dw
Eliminating d between eqs. (6) and (7) yields the required
stiffness matrix for the region H :
which is symmetric and can be directly assembled into the fi-

nite element stiffness matrix. However, notice that as the
matrix KQw is calculated by integration along the boundary,
contribut1ons from each boundary node will be introduced. This
means that the bandwidth of KH will generally be large and
will increase the bandwidth of the finite element stiffness
matrix.
To ensure continuity along the interface B , the inter-
polation of the vector d of displacement discontinuities
should be the same as that used in the adjacent finite elements.
If the finite elements are of the isoparametric type, the iso-
parametric boundary segment discussed in the last section would
be a natural choice of interpolation and would likely lead to a
compact computer code as well as a versatile method of modeling
different boundary shapes.
A program to implement the above element in a three-dimen-
sional finite element code is currently being developed.
163
CONCLUSIONS
The development of the DDM for the analysis of three-
dimensional elastic media is relatively straightforward if the
Volterra integral is used to derive the fundamental solution.
The use of the Volterra integral permits the identification of
the DDM as a boundary integral procedure with a different in-
fluence function. It also allows:
1) a relatively easier means of extending the technique
to semi-infinite media,
2) more flexible representation of boundary surfaces in
three dimensions by means of integration over general
quadrilaterals, and
3) a natural linkage with finite element analysis.
A working computer code has been developed and tested.
Two major difficulties in the computer implementation of the
technique are the large size of the system of equations to be
solved and the definition of the location and geometry of three-
dimensional models. The large size of the system of equations
can be decreased by either the use of symmetry or the elimina-
tion of non-existent modes of displacement discontinuity, or
both. Alternatively, the equations can be solved by an out of
core solution algorithm. The ability to visualize and prepare
the input data for three-dimensional models would be greatly
enhanced by the use of an interactive graphics system.
ACKNOWLEDGMENTS
The authors would like to thank Dr. Ken Matthews and
Mr. Byron Stewart of Golder Associates in Vancouver, B.C., for
suggesting the mining problem shown in this paper. The first
author would like to gratefully acknowledge the assistance pro-
vided by Dr. Tom Hughes of Stanford University with regard to
the problem of linking the DDM with finite elements.
REFERENCES
BATHE, K.J. and E.L. WILSON [1976] Numerical Methods in Finite
Element Analysis, Prentice-Hall, Inc., Englewood Cliffs, NJ.
BERRY, D.S. and T.W. SALES [1962] An Elastic Treatment of
Ground Movement due to Mining- III. Three Dimensional Problem,
Transversely Isotropic Ground, J. Mech. Phys. Solids, lQ, 73-83.
CHEN, W.T. [1964] Displacement Discontinuity Over a Transversely
Isotropic Halfspace, IBM J. Res. and Dev., ~, 435-442.
CHINNERY, M.A. [1961] Deformation of the Ground Around Surface
Faults, Bull. Seis. Soc. Amer., ~, 355-372.
164
CROUCH, S.L. [1973] Two Dimensional Analysis of Near Surface,

Single Seam Extraction, Int. J. Rock Mech. Min. Sci. & Geomech.
Abstr., J.Q, 85-96.
[1976a] Solution of Plane Elasticity Problems by
the Displacement Discontinuity Method, Int. J. Num. Meth. Eng.,
J.Q, 301-343.
[1976b] Analysis of Stresses and Displacements
Around Underground Excavations: An Application of the Displace-
ment Discontinuity Method, University of Minnesota Geomechanics
Report, Dept. of Civil and Mineral Engineering, University of
Minnesota.
HIRTH, J.P. and J. LOTHE [1968] Theory of Dislocations, McGraw-
Hi 11, Inc., New York.
HOFMEISTER, L.D. [1978] An Out of Core Equation Solver for Full,
Unsymmetric Matrices, Int. J. Num. Meth. Eng., ]l, 721-731.
JAEGER, J.C. and N.G.W. COOK [1976] Fundamental of Rock Mech-
anics, John Wiley &Sons, Inc., New York.
LACHAT, J.C. and J.O. WATSON [1976] Effective Numerical Treat-
ment of Boundary Integral Equations: A Formulation for Three
Dimensional Elasto-Statics, Int. J. Num. Meth. Eng., J.Q, 991-
1006.
LOVE, A.E.H. [1944] A Treatise on the Mathematical Theory of
Elasticity, Dover Publications, New York.
MANSINHA, L. and D.E. SMYLIE [1971] The Displacement Fields of
Inclined Faults, Bull. Seis. Soc. Amer., ~. 1433-1440.
MARUYAMA, T. [1964] Statical Elastic Dislocations in aninfinite
and Semi-Infinite Medium, Bull. Earthg. Res. Inst., 42, 289-368.
[1966] On Two Dimensional Elastic Dislocations in
an Infinite and Semi-Infinite Medium, Bull. Earthg. Res. Inst.,
44, 811-871.
MINDLIN, R. and D. CHENG [1950] Nuclei of Strain in the Semi-
Infinite Solid, J. Appl. Phys., 11, 926-930.
PRESS, F. [1965] Displacements, Strains and Tilts at Teleseis-
mic Distances, J. Geophys. Res., 70, 2395-2412.
RONGVED, L. [1957] Dislocation Over a Bounded Plane Area in an
Infinite Solid, J. Appl. Mech., 79, 252-254.
and J.T. FRASIER [1958] Displacement Discontinuity
'i~n't~h-e'E~l~a-sticHalf-Space, J. Appl. Mech., ~. 125-128.
165
SNEDDON, I.N. [1946] The Distribution of Stress in the Neigh-

bourhood of a Crack in an Elastic Solid, Proc. Roy. Soc. A.,
187, 229-260.
STEKETEE, J.A. [1958] Some Geophysical Applications of the
Elasticity Theory of Dislocations, Can. J. Phys., 36,1168-1198.
ZIENKIEWICZ, O.C. [1977] The Finite Element Method, 3rd ed.,
McGraw-Hill, New York.
~-o7~~· D.W. KELLY and P. BETTESS [1977] The Coupling of the
Finite Element Method and Boundary Solution Procedures, Int. J.
Num. Meth. Eng., Jl, 355-375.
166
g Global Coordinates
Local Coordinates
c Source Coordinates
Figure 1. Global and local coordinate systems of

dislocation segment
167
INPUT
r1odel geometry
Boundary conditions
Compute matrix of
influence coefficients
Solve system of
equations
OUTPUT
Displacements and
stresses
Figure 2. Organization of program DD~13

168
a
r--.....
~
"\
i\ p
\
a
Figure 3a. Discretization of penny-shaped crack
j p
Divided into 50 segments

200a units long tn g3
a
direction
g3 into page
Figure 3b. Discretization of circular hole in infinite

medium two-dimensional
169
Ana lytica 1
Numerical 91
1.0 93 =r
V1
V1
(!)
s..
....,
V1
0.8
7Sll/P
p
92 =r
"0 0.6
(!)
N
<tJ
E 0.4
s..
0
z
0.2
0.0
0.0 0.5 1.0 1.5 2.0
91/a
Figure 4a. Analysis of penny shaped crack
3.0
V1
V1
g 3 into page
s.. 2.0
(!)
....,
V1
"0
(!)
N
<tJ
E
s.. 1.0
0
z
sll;P
7
0.0
0.0 1.5 2.0 2.5 3.0
9/a
Figure 4b. Two-dimensionalized analysis of circular

hole in infinite media
.....
0
~~
I I I I 180m
c
I I I I 11
Detail of pane 1
discretization
shaft/
a : ( 300, 1000, 0)
0 100 m
b: (0, 1000, 0)
c: (-200tan15°, 1000+200tan 215°, 300cosa) a o sin- 1 (zt~n15°)
d: (200tan15°, 1200, 200tan15°)
e: (150, 1000, 150)
Figure 5. Orebody configuration and mining plan

171
Horizontal (u 1 ) Displacement (mm)

0 0.1 0.2 0 -1.0 -2.0
1000
1020
1040
1060
1080
..s 1100
....,
..<:::
c.
a
OJ 1120
1140
1160
1180
1200
After East 1imb After East and West limbs
Figure 6a. Horizontal (u 1 ) displacements of shaft due to

mining sequence shown in Figure 5
172
Horizontal (u 3 ) Displacement (mm)
1000 ° -1.0 -2.0 -3.0 0 -1.0 -2.0 -3.0
1020
1040
1060
1080
..5 llOO
...,
.c:
c.
ll20
"'
Cl
ll40
1160
ll80
1200
After East 1imb After East and West limbs
Figure 6b. Horizontal (u 3 ) displacements of shaft due to

mining sequence shown in Figure 5
173
x3
+1
R
-1 +1
x2
3
-1
Figure 7. Transformation of general quadrilateral to a standard

region
FE Mesh
or
Figure 8. Linkage of DDM with finite element method

174
A UNIFIED SECOND ORDER BOUNDARY ELEMENT METHOD FOR STRUCTURES

ANALYSIS
Michel Dubois
Research Engineer
CETIM - France
ABSTRACT
The classical modelizations used in structures analysis fail

for singular situations giving high gradients. All numerical
methods lead then to inaccurate or instable results, due to
the geometrical and/or loading singularities.
High gradients zones (even very localized) are precisely

damaging areas, especially responsible for fatigue collapses.
Therefore, it seems particularly interesting to introduce a
better modelization in order to eliminate the mentioned dif-
ficulties, and insuring by the same way a complete compatibi-
lity in complex structures analysis.
From the mathematical point of view, we are lead to construct

a second order modelization with free rotations : kinematics
of continua are defined by the first gradient of displacements
(classical strains) and, further, by second order terms,
homogeneous to second gradient components. In addition to
classical (symmetrical) stress, we have to take into account
skew-symmetric components and couple-stress which are second
order terms. The second order quantities are negligible com-
pared to the first order ones in low gradients zones, but
they cannot be omitted in singular high gradients zones.
The main advantage of the new modelization is the existence

of a common variables system used for all kinds of structures:
three dimensional thick structures, shells, beams
From the numerical point of view, the existence of a common

variational principle give~ after discretizatio~ a common
system of algebEic equations, which is automatically reduced
in function of geometrical data and in function of a test
related to the mesh fineness, according to singular zones.
175
For the isotropic elastostatic case presented above, the four

additional elasticity coefficients are defined in terms of
the two classical ones and in function of a mesh size para-
meter, in order to recover the classical formulation in re-
gular zones.
By using distribution's theory, the field equations are

transformed into two integral equations relating displace-
ments, rotations, stress and couple-stress. Taking limits at
collocation points on these relations, we obtain the boundary
integral equations to be numerically solved.
The integral equations kernels are the elementary solutions

corresponding to unit concentrated forces and couples. These
elementary solutions are calculated by FOURIER's transform
and HORMANDER's method. The numerical treatment is then the
same as used for standard B.E. programs of CETIM.
NOTATION
Let V be a (normed) vectorial space of ~ admissible elements

defined on a S-structure [If . We introduce a bilinear
symmetrical positive definite and continuous application of
V x V in IR • If ~I , ~ 2 ~ V , if v 1 is a definition subdo-
main of V with boundary s 1 ( I = 3, 2, I according to the
sub1omai~ is three, tw~ or ~ne dimensional), we designate as
< VI , v2 > Y
and < v1 , v2 > sr the real numbers that the
bilinear app !cation associates to ~~ , ~2 on v 1 and on s 1 .
If I = 3, v 3 is a volumic subdomain and s 3 is the boundary
+ -+
¥
<VI, v2> 3 =
s
surface of v 3 ; the bilinear application is then defined by :
+ -+ + + j +
V3v1. v2 dV3, < vl ,v2 > S 3= ~v1.v2dS3
-+ (!)
where ~1 , v 2 is the scalar product, dV 3 is the differential

elementary volume, dS 3 is the differential elementary area.
If I = 2, V2 is the reference surface of a shell and s 2 is
J J
the boundary curve of V2 ; the bilinear application is then
define1 by : + + + -+ -+ +
< VI,Vz > V2 = V2v1.v2dV2, < vl,v2)s 2= S 2v1.v2 dS 2 (2)
where dV 2 is the differential elementary area on the referen-
ce surface and dS 2 is the differential elementary arc on the
boundary curve.
If I= I, v 1 is the reference fibre of a beam, and s 1 repre-
sents the boundary points of v 1 ; the bilinear application
is then defined by :
<~1.~2>v =Jv~1·~2dV1, <~1.~2>s
1 ~1·~2 Is
1 1 ..... 1 .....
where dV 1 is the differential elementary arc and vl .v 2 S1
I
(3)
designates the value of the scalar product to be taken at

boundary points sl •
176
Assume S' and S" be the complementary parts of SI where

kinematical condittons and efforts are respectively pres-
cribed.
We use a curvilinear system of coordinates si (i = 1,2,3)

which are intrinsic coordinates for the numerical discreti-
zation. We note ~ the position vector on VI with regard
to a global cartesian reference frame and are the g.
basis-vectors of the local natural reference~ frame:
-+ -+
g.
~
a.
~
x ( 4)
In a general way, di is the derivative with respect to si '

so that :
-+ -+
(l. v g. (5)
~ J
where vjl i is the covariant derivate. At last, ri will

be the exterior normal on the boundary SI of VI.
GENERALIZED MODELIZATIONS
The kinematical quantities are defined on each substructure

VI by displacements I~ and free rotations 1 ~ as by genera-
l~zed strains I;i and Iki •
The efforts are defined by stresses and couple stresses
I-r· I-+i .
cr ~ and ].l .Presyr~bed efforts are forces and moments
densities 1 ! and t given on VI as forces and couples
I+ I+ - -
cr and .1:!... given on S"I
- If I = 3, the above fields are volumic distributions on

V3 Indexes i take then values I , 2, 3. The {}
symmetrical components of 3;i in the natural reference
frame reduce to classical strain E; ••
~J
€· • e. . with i, j = I , 2, 3 (6)
~J ~J
-+'
The corresponding symmetrtcal components of 3cr~ reduce
to classical stress cr~J
ij
(} with i, j = 1,2,3 (7)
But the generalized ~~delization involves skew;symmet~~cal

components of 3;i • 3cr~ and components of 3 ki ., 3 ].l ~
which do not appear in classical theory.
177
- If I = 2, the fields are surface distributions on refe-

rence surfaces v2 of shells. v 2 being defined by s 3 = 0
indexes i take values a 1,2. The 2 ~a are membrane
strains Ea and 2 k a are strains Ka due to shells
curvatures changes
+
= Ea a
K
a
a= 1,2 ( 8)
Components of' 2"t a along = % i

(sl, s 2• s 3 = 0)
with S I ,2 are membrane eftorts N c(3; components of
2+ + + -+
o a along g3 orthogonal to g1 , g2 represent
transversal sharing forces Na 3 :
(9)
Components of 2 ji a along t + are bending moments M ~

and components of 2ji a along g3 represent twisting moments
M a 3 :
(10)
-If I= I, the fields are linear distributions on referen-

ce fibres v 1 of beams. Indexes i have then the single
value I, where we define v1 by s2 s3 = 0 and get
+ +
g1 g2 <s 1 , s 2 = s3 = 0) with g2 g3 orthogonal to
-+
g1
The 1 ~ 1 and 1k 1 are then reduction elements E and K
of strain torsor at s1
(II)
while 1(p and are reduction elements and M

of contact effort at s 1
1"t1 = N (12)
Vector (1"t1 , g 1 ) g 1 is a traction or compression at s 1

according to · 1Jl.g1 is positive or negative ;
t
1 1 - t .i1 >
( 1 1 i1 is the shearing force ;
(1ji"1 .g1) g1 is the bending moment and 1ii" 1 ( 1 1 .g1 )g1 t
is the twisting moment.
178
BASIC EQUATIONS
The generalized strains have covariant components I E·.
I l.J
kij in the natural reference frame
I
E• • (13)
l.J
where the generalized strain vectors are defined by compa-
tibility equations :
I+ lit.1. =Cl 1..Ie
E • (14)
1. '
with x being the vectorial cross product. Equation(l4)are

symbolically written :
( I+E • lit. I+ I+8
= c ( u, (IS)
1. ' 1.
where C is the linear first order differential operator

defined by (14).
Kinematical boundary conditions are

II) I+
= u on S' I (16)
I+ I+
where __!:!.and J2_ are prescribed distributions on parts
Generalized efforts have contravariant components I 0 ij

I ij
]1 in the natural reference frame
I+ r_,_i I ij + I j +
0 on. 0 nigj 0 n gj
1.
I+ I+ i n· I ij + I j -> (17)
]1 11
1. ]1 nigj ]1 n g·J
Transformation of I + i I -+ i I+ I+
(J ]1 into on ' Jln is
symbolically written
I + I+ l+ i I + i
(J
Jl n
N ( (J ]1 (18)
n '
I +i I+ i
Vectors a and Jl verify equilibrium equations
for the quasistatic case here to be considered :
+ I+ i I+ i I+ (Jg)
+ 0 g. X a +a i Jl + c =O
1
179
which are symbolically written

D ( I~ i, It i ) (20)
Boundary conditions on efforts are

I ~ I ~ I~= I~ on SI" (21)
(J ~ )J _!:__
I~ • I~ • . . .
where o ana )J are prescr~bed d~str1but1ons on parts
of SI. At last, we have to define constitutive relations
the linear elastic case to be considered:
I i. 3 WI I i. 0 WI ( 22 )
(J J )J ]=---
diE:·· oik··
~J' ~J
which are symbolically written

(I ~i I Vi ) R (I (23)
Strain energy densities w1 are quadratic positive defi-

nite forms of principal invariants of :
I EiJ. Ik.. , with GRI coefficients
~J
(24)
The generalized modelizations defined by the above equa-
tions are called IDGM according to whether~ = 3 2 1.
The generating model is 3DGM while 2DGM and IDGM may be
considered as simplified representations.
The mesh size parameter is defined as e/e 0 where e is

a typical measure of local mesh size whereas e 0 is
a prescribed (adjustable data) critical dimension.
If e ~ e 0 , GR3 reduce to the classical R3 coefficients
so that 3DGM reduces to classical 3DM elasticity.
If e < e 0 , the generating model coefficients are obtai-
ned as functions of R3 and e/e 0
- If e -~e 0 , GR3_.R3 = (A,JJ)
- If e < e0 , GR3 = GR3 (A ,)J ,e/e 0 ) (25)
The GR2 and GRI are then obtained from GR3 by the same way
as in classical 2MD shells and IMD beams theories [2],[1].
180
FUNCTIONAL FORMULATION
Transformation of basic equations into integral relations is
performed by a former presented method [4], L5], making use
of distributions theory {8/. In a first step, functions
equations are transf~rmed into linear continuous functionals
equations. The Iti, e are prolonged by zero if ~ VI ,t
by using the characteristic function :
.....
if ~ ~ VI
{ 01
if x f. VI (26)
and we apply in distribution sense at (H (VI)Iu,H(VI)Ie)
the composed second order linear differential operator :
L =D o R o C (27)
where D, R, C, are respectively defined by (20), (23),(15),
We use distributions derivation formulas like
v cS (SI )
<li (H (VI) -+v) = H (VI ) <l i v-+ - ni-+ (28)
and :
Clj <li (H (VI) :;j ) = H(VI) <lj <li :;j- <lj (ni:;j o (SI))
- nj Cli~ cS (SI) (29)
where n.:;jo (S 1 ) is the distribution on SI defined on each

admissibl~ ~ oy :
(30)
whose value is given by (!), (2) or (3) according to

I= 3, 2 or I.
As result we have the functional equation :
I+ I+ I+ I+ I+ I+
L ( H(VI) u, H (VI) 6 = H (VI)L( u, 6 )- T( u, 8) o (SI)
tT((I~, Ie) o (SI)) (31)
where T is the composed first order linear operator

T =N o R o C (32)
with N defined by (18), and where tT is the transposed
operator in the sense of partial differential equations
theory.
Making use of (16), (19) and (21), we obtain finally for(31):
I+ ( )I -+6 ) ( ) (I-r I+) (I+ I+
L (H( VI ) u, H VI = -H VI ~· ~ - an, ~n) o (SI)
- tT ((I~. Ie ) o (S 1 )) (33)
which is the functional formulation of equations (13)to (23).
181
INTEGRAL RELATIONS
Let IE (Iu , I H ) be the tempered elementary solution
of L The (6 x 6) Iu and I H matrix has components
I I-+ -+ I H
Uij = ui.gj (34)
ij
where Iu. and I H
are respectively displacements and rot~
tions cofresponding to concentrated forces and couples at
origin in direction
~
Iu.
~
g. :
and I H. are tempered
~
solutions of
I-+ I =t- -+I -+I
L ( Ui , H i) (gi o (0), gi o (O)) (35)
with i o (0) being the I dimensional DIRAC distribution.
The elementary solution IE is obtained by HORMANDER'S
method ru
By convolution product of IE with (33), we obtain then
I I -+ ~ I~ I-+ -+
<( U, H) (x -0, {_..£., _£) (t; )>
vi
I -+ I
-+ I-+ I-+ -+
H)(x- t;),( an, ~n)(t; )>S
+ <( U,
I I -+ -+ I-+ I -+ -+ I
- <(T ( U, H))( x-t;), ( u, 8) (l:)>s
I
(36)
which are integral relations available for any ~ E VI
(I= 3,2,1),! being an integration variable.
BOUNDARY INTEGRAL EQUATIONS
Transformation of (36) into boundary integral equations is

performed by taking the limit on (36) of ~G. VI to ~G.SI
We obtain then integral equations

a (-+)
x (I-+
u, I-+
8 ) (-+x) __ < (I U, I H) (-+
x - -+) ..!_, I-+)(-+)
t; , (I~ ....£ t; >
VI
I I -+ -+ I-+ I-+ -+
+ <( U, H)(x -t;),( a , ~ )(t;)> S
n n I
I I -+ -+ I-+ I-+ -+
- <(T( u, H))(x- t;),( u, 8)(t;)>s
(37)
I
. -+
w~th x E SI and where a (~) is a coefficients matrix
depending on the local geometry of SI •
The numerical treatment of (37) is quite the same as the
one used in standard programs of CETIM [6} for classical
elasticity. Boundaries s1 are discretized by finite elements
s~ ' the unknowns :
I-+ I -+ I -+ I-+
u , 8 , an , ~n are isoparametrically inter-
182
polated by cl~ssic~l shape functions, and (37) is solved at

collocation po~nts xc.
The obtained algebraic system has coefficie~ts which are

integrals on S~ of (Iu, I H) and T( Iu, I H). These in-
~egrals are computed by ordinary GAUSS quadrature rules if
!; ¢s~ containing ;c . When t
E. s~ containing ;c '
the most singular coefficients (arising from T (IU, IH ))
are obtained by the unicity condition
a (I~, Ie ) + < T (IU, I H ) , (I ;r, I e) >
0
I-*
with u b where ~' b are arbitrary
constant vectors.
REFERENCES
/1/ Dubois, M. (1978) Methodes numeriques de calcul des

structures. Part. 2. N.T.I. - Calcul, CETIM, Senlis,
France.
/2/ Dubois, M. (1980) Bases d'un systeme compatible de
calcul des structures. Second International Symposium
on Innovative Numerical Analysis in Applied Enginee-
ring Science, Montreal, Canada.
/3/ Dubois, M. (1981) Etude d'un systeme compatible de
calcul des structures. Recherche sous contrat D.G.R.S.T.
n° 80.E.0885, France.
/4/ Dubois, M.,Lachat, J.C., (1972) The Integral formula-
tion of boundary value problems. Variational Methods
in Engineering, University of Southampton.
/5/ Dubois, M. , Lachat, J.C., Watson, J.O. (1973) Le Cal-
cul des Structures par la methode des equations inte-
grales. C.R. quatrieme Congres Canadien de Mecanique
Appliquee, Montreal, Canada.
/6/ Dubois, M. , Lange, D., Serres, D., Chaudouet, A.,
Boissenot, J.M., Loubignac, G., Percie du Sert, B.(l978)
Application de la methode des equations integrales a
la mecanique. N.T- Calcul, CETIM, Senlis, France.
/7/ Hornander, L. (1965) Lineinie diferentialnie operatori s
ciastimi proizvedn~mi, Moskva, C.C.C.P.
/8/ Schwartz, L. (1966) Theorie des distributions. Hermann,
Paris, France.
183
METHOD OF BOUNDARY INTEGRAL EQUATIONS FOR ANALYSIS

OF THREE DIMENSIONAL CRACK PROBLEMS
J. Balas and J. Sladek
Institute of Construction and Architecture, Slovak

Academy of Sciences, 885 46 Bratislava, CSSR
1. INTRODUCTION
The solution of spatial problems for finite

region with cracks often encounters difficulties
which results in the fact that there is a incompa-
rably smaller number of solved spatial problems
for crack than that of plane problems. There are
almost no exact solutions of this kind at all.
Those prevailing are numerical solutions moetly
using the finite elements method (FEM). Among the
earlier papers these by Tracey [1] and Benzley [2]
should be mentioned. The calculations made by this
method are rather inefficient and require large
computers. It has been shown that the method of
boundary integral equations (BIE) is more efficient
for these problems not claiming such storage capa-
cities as the finite element method. The BIE method
is very well suited to the solution of three-dimen-
sional stress eoncentration problems reducing the
latter to boundary solutions, i.e. requiring only
1~
the elements on the boundary to be defined. In con-

tradistinction to them, the finite elements require
many interior nodes and approximations involving
discontinuity of stresses. The feet that the BIE
method allows direct solutions for boundary
displacements with no modelling ef the internal
stresses end stress intensity factor analysis, re-
quiring only crack opening displacements, implies
that there is no need to determine the stress in-
side the body, near the creek tip [3].
In spite of the foregoing advantages, there

will be a large number of surface elements where
more complicated problems ere treated. For this
reason, the problem solution is split in two auxi-
liary problems. In the first part, a finite region
without crack will be solved by the BIE method. In
the second part (using the so-called perturbence
method) the solution will be concerned with the
state of stress and strain of a crack in infinite
space. Using the Schwarz algorithm we obtain then
the final solution.
2. BOUNDARY-INTEGRAL EQUATION ANALYSIS
The theoretical deduction of integral equa-

tions being known [4], only the fundamental equa-
tions and their numerical solutionare presented.
For the deduction of the integral equations we use
the reciprocity theorem (the Green formula in the
theory of potentials, the Betti's theorem concer-
ning reciprocity of virtual displacements) as well
as the corresponding singular solution of diffe-
rential equations the so called Kelvin problem of
185
elasticity. The combination of both these partial

solutions is well-known from the classical theory
of elasticity as the Somigliano's identity. The
above mentioned method includes only physical quan-
tities i.e. loads and displacements on the surface
of the region and which permits also a stttble nume-
rical solution for the unsmooth geometries (corners
end edges) • The studies region may be simply
or multiply connected this feature by no means
complicating the numerical procedure.
The set of integral equations for the solu-

tion of elasto-static problems due to surface loads
relates the boundary values of tractions ti(x)
and displacements ui(x).
Cij(x)uj(x) + JTij(x,y)uj(y)ds = (1)

s
= J uij(x,y)tj(y)de
s
where Cij is a constant, in the case of conti-
nuous tangent plane at x, c1 j(x) = oij/2. The ker-
nel functions Tij(x,y) and Uij(x,y) are the
tractions and displacements in the xi direction
at y(y1 , y2 , y3 ) due to orthogonal unit loads in
the xj directions at x(x 1 , x 2 , x 3 ).
Uij(x,y) = Sn;(~~~)Er { (3- 4v)oij + r,i r,j} (2)
TiJ" (x,y) = -
8n;(1-v)r
2 { (1-2v) fn oij + (3)
+ 3 ~ r,i r,j + (1-2v) (ni r,i- nj r,j>}

186
where
or Yi - xi
r,i = Clyi = r
v and E are elastic constants, r is the distan-
ce between pointe x and y, 5ij is the delta-
-function, ui is a component of the unit outward-
-normal of the surface. The repeaten inrlex in a
procuct is to be understood as a summation over a II
the indices.
The given boundary conditione are the trac-

tions tj on a pert of the boundary and the dis-
placements uj and/or the combination of both tj
and uj on the remainder of the boundary. The so-
lution of BIE [4] gives the other unknown parte of
the functions t j and uj over a II the surface.
However, the BIE are exactly solvable only for a
few types of regions and must, therefore, be solved
numerically. In the last case we use ieoparametri-
cal shape functions for the representation of both
the geometry and the boundary data [5]. The ap-
proach is known from the FEM.
The Cartesian coordinates of an arbitrary

point of an element are expressed in terms of nodal
coordinates Y! and shape functions Ne (g) of lo-
cal coordinates g.
yi(g) = Ne(g} Y~ (4)
where
N1(g) = t <g,+1) (g 2+1) <g,+g2-1)
N5(g) =~ <g,+1} ( 1-g~) etc. [6J
187
~1 ~1
2 4
2
3
Fig. 1. Surface elements: a) linear approximation

b) quadratic approxim.
By differentiating equation (4) with res-

pect to g1 we obtain the tangent vectors to the
coordinate curves by means of which the Jacobean
and the normal to the element surface is obtained.
Similarly, the displacements ui(x) and the

tractions ti(x) on the surface are expressed in
terms of their values at the nodal points and the
shape functions.
<t>(g) = Mb(g) <t>b (5)
where Jlb(g)stand for linear approximation
M1(g) = t (g1 + 1) (g2 + 1)
M2(g) . t (g2 + 1) (1 - g,)
MJ(g) = l4 (1 - g,> (1 - g 2)
M4 (g) "' t (1 + g,) (1 - g2)
For quadratic approximation the following

formula is true
188
(6)
The points x, at which the integral equations are
expressed are internal points of the surface ele-
ment not identical with the nodal points in terms
of which the boundary coordinates, the displace-
ments and/or the tractions are expressed. General-
ly the number of the equations after the discreti-
zation of BIE and the number of unknown values of
the nodal displacements and/or tractions are dif-
ferent, i.e. in the matrix form
[ Tij] {uj} - [uij J {tj} = { o} i
j
= 1 ,2 ••• n
= 1,2 ••• p
(7)
J
where [ Tij and [ Uij J
are matrices of the coef-
ficients determined from the numerical solution of
the BIE {1), {uj} and {tj} are the displacement
vector and the traction vector, respectively, of
the nodal points.
In the form {6) the problem is not solvable,

it is, however, possible to minimize the error
function by the least squares method [7J
(8)
n n
L
i=1
{Ei} 2 +L
i=1
([Tij] {uj}- [uij] {tj}) 2 =min
The resulting system of equations is given in the

form
(9)
The matrix of this system of equations

(n x n) is symmetric, fully populated. The non-sin-
gularity condition of the system is n~p except
the static determination, i.e. the number of equa-
tions in (7) should be greater or equal to the num-
189
ber of the unknown parameters.
To obtain the matrix coefficients Tij' Uij

in (9) we solve numerically the integrals of the
product of the kernel functions Tij (3) and Uij
(2) and the shape functions Nu(~) of the displa-
cements and Nt(~) of the tractions with the use
of the Gaussian quadrature formulas.
I If <
1 1
11 1' 112> d 111 d1]2 = (10)
-1 -1
n1 n2
LL
i==1 j=1
+ E: '
where e: is the value of the error and n 1 , n 2

are the orders of the used quadrature formula.
In the case of singular elements the trans-

formation of the local coordinates to local polar
coordinates ~, IP having their origin at the sin-
gularity point [1] leads to a reduction of the or-
der of the singularity by one degree. It means
that in this transformation the integrals with a
weak singularity are reduced to integrals over
smooth functions. The strong singularity (1/r2 )
cannot be eliminated from the integrals. For their
evaluation, however, the solution of the following
homogeneous integral equation may be used
cij(x) uj(x) + Is Tij(x,y) uj(y) de= 0 (11)
The solution of this equation yields the function

ui(x) = ai + Eijk xj bk, (12)
the physical interpretation of which are rigid dis-
190
placements end rotations of the body; ei' bk (i,k=

1, 2, 3) ere arbitrary consteats, Eijk is an anti-
symmetric unit tensor.
From equations (11) end (12) by the choice of

constant aj = uj(x) there followsthe relation
Cij (x) uj (x) + I Tij (x,y) uj (x) ds • (13)

so
= -
I
s-s0
and when this is used we receive
(14)
=- I
s-s 0
[ uj (y) - uj (x) J ds
where S0 denotes the area of the element with a
singularity point.
The first of the right hand side integrals in

equation (14) is obtained as a sum of integrals
over all the elements but for the singularity (hen-
ce they need not be evaluated once more) and the
second one contains at most a week singularity and
may, therefore, be evaluated numerically in terms
of the local polar coordinates.
The stresses in the interior of the studies

region ere obtained by differentiating u at the in-
191
tarnal points [s]

crij(x) = J Dkij(x,y) tk(y) ds - ( 15)
s
- J skij(x,y) uk(y) ds
s
where
Dkij = : 2 { ( 1-2 v) [ Bki r , j + B kj r , i - Bij r, k ] +
+ Jr,i r,j r,k} 8TL(~-v)
skij = rJ~ { 3
2 ar [ (1-2v)
an 5ij r,k +
+ v( 5ik r,j + ojk r,i> - 5r,i r,j r,k J

+ Jv (ni r,j r,k + nj r,i r,k) + (1-2v)
(Jnk r,i r,j + nj 5ik + n 1 5jk) -
- (1-4v) nk oij} 8n:(~-v)

3. CRACK OF ARBITRARY CONTOUR IN INFINITE SPACE
In this part of the paper we shall discuss

the perturbation method of solution departing from
the work of Weaver [9] who has derived the system
of integral equations for unknown dislocations of
a spatial crack of arbitrary contour in an infinite
space. All the other qumtities describing the state
of stress and strain near the crack may be expres-
sed in terms of unknown dislocations will be calcu-
lated after the transformation of the system of in-
tegral equations to a system of linear equations.
192
We shall study an infinite elastic space

using Cartesian coordinates xi. According to the
preceding paragraph, the j-th stress component due
to a unit force in an i-th direction at point sk
is expressed in terms of equation (3)
3r r
Tij(x,s) .. - K2 [ ~~ (6ij + , ·~ 2~j) - (16)
r
- nj r ,i + ni r ,j J
where
1 - 2v
K = 8n:(1-v)
We shall assume that region D consists of two
parts mutually complementary to the infinite space.
SC"r
s----------------
Fig. 2. Crack of arbitrary contour in infinite
space
The internal region is infinite and denoted

by n+. The external D- is its complement. The
common surface of n+ and D- is denoted by s±.
The plane part S± (crack) is denoted by S~r·
193
According to Kupradze DoJ it holds for xED+
ui (x) • I
s+
tj (y) uij (y,x) de - (17)
- J uj (y) Tij (y,x) de

s+
( 18)
o = I
s-
tj (y) uij (y,x) de -
s-
I uj (y)
Let us assume the crack loads to be symmetric with

respect to x 3 = 0 i.e. tj+ = -tj- • Similarly
+
Tij= -
-Tij + = Uij
and Uij - • We obtain by summing
equations (17) an~ (18)
ui (x) = -
+
J t. uj (y) T!j (y,x) de (19)
8 cr
where
+
t. uj = uj - uj
The stress components ere obtained frem equation
( 19)
= !J.(u~,J + uj,a.> for a.= 1 ,2
(20)
+
= 21J. [ v u a., a. + (1-v) uJ,J J _L_
1-2v
and by integration by par~ we get
ui,a. =- J t.uj(y) aT!j<y,x)

de •
s+ a xa.
cr
= J t. uj ( y) aT!j(y,x)
'* ds =
s+ oya.
cr
• - ) •uj,a T!j(y,x) ds
or
194
Because on the crack there is a vector of normal

+
ni = - 0 i3
T
ij
,. - !L[ 2
r2 r
Hence, form the preceding relationships and from
equation (20) we receive
oa. 3 = E <1 -v)

+ ~a.
16n:(1-~ 2 )
J[o r
,y
-5
a.y
r
'~
+
8 cr
3r ,a. r 8 r ,y
(21)
+ 1 - ~v
0 33 E 2
• 8n:(1-v J
+ )
8 cr
for a., ~,y .. 1, 2
The integrals in equation (21) are understood in
the sense of Cauchy's principal values. Now, we can
take up the numerical solution of the system of
integral equations of first order (21) where dis-
placements on the surface of the crack to which the
known loads o33 and/or oa. 3 are applied. The
crack surface is approximated by meaas of surface
elements. Then, the Cartesian coordinates of any
point of the element are expressed in terms of co-
ordinates of nodal points and shape functions in
terms of local coordinates. For this purpose, we
can use immediately the elements shown in Fig. 1
with the corresponding shape functions. As it will
be shown later for penny-shaped cracks or ellipti-
cal cracks it is advantageous to use a transition
element, Such an element permits matching of one ele-
ment with two elements side-by-side [1]. For the
195
sake of simplicity, a formulation of the transition

element with the introduction of transition inter-
mediate nodes is presented here. However, the me-
thod is general and can be easily applied to other
and higher-order elements.
In Fig. 3 the edge AB of element I has a

diecontinuity in the displacement slope at node c.
The variation of any displacement u along edge AB
can be expressed as
U= t (1-0 UA + t (1+~) UB + (1- ~~~) b.UC (22)
I
A c B
A(-1) C(O) B(l)

II III
Fig. 3. The transition quadrilateral element and

displacement field
If we express b.u0 in terms of displacement

u at the nodal pointe we obtain from equation (22)
u = t (1~1 -g)uA + t (1~1 +~)uB + (1-1~1 )u0 (23)
The procedure used to obtain the shape functions
for the transition elements may be summarized as
follows: Write the shape functions for the corner
nodes as if there were no intermediate transition
node. Write the shape function for the intermediate
196
node treating the displaceaent at the intermediate

node as a departure ~llc from the linear displa-
cement variation defined by the displacements at
the corner nodes. Modify the shape function at the
corner nodes in such a way that the displacement
at the intermediate node becomes the total displa-
cement at the node Uc· Note that the shape function
at the intermediate node does not change
if N6 • N7 • Na = 0
N1 = 41 (1-0 ( 1 -1])
1
- 4 (1-1~1) ( 1 -1])
1
N2 .. 4 (1-~) ( 1+1])
1 1
N3 '"'4 ( 1+ ~) ( 1-1]) -4 (1-1~1) ( 1-1])
1
N4 ""4 (1 + ~) (1+1])
1
N5 =2 (1-l~i> ( 1-1])
'7)
Fig. 4. The transition quadrilateral element and

the shape function
In the present case the integrand of equation (10)

is discontinuous at ~ = 0 and the integral is
split into two continuous integrals.
197
The gradient of the shape functions can now

be written as
aNi
ax-
(24)
it)
where
[J] .. ( ~t
..Q..z ~
01] 01]
By substi ti ng (24) into integral equation (21)

and after performing the corresponding numerical
integrations we get a system of linear equations
for the unknown displacements in the crack.
The strong singularity integrals will be cal-

culated by erecting a normal at each nodal point
on which the auxiliary points P~ (l = 1,2,3,4)
lie for which the integral equation (21) will be
written.
The subintegral expressions at these points

will be regular and the value of the integral for
point Pjlying on the crack surface (nodal point)
will be found by extrapolation of integrals for the
auxiliary points (12). A beam including an eccen-
trically situated crack of rectangular shape with
respect to the neutral axis is subjected to a ben-
ding moment. The stresses at the place of the crack
are then
198
Po = Mh1
-y- (1 - L)
h
1
p1 = M I
I - moment of inertia of the cross-section
area of the beam
Fig. 5. Creek in a beam subjected to a bending

moment
The stress intensity factor Kr is related

to the displacements near the edge of the crack.
Kassir and Sih have shown that the two-dimensional
equations for the plane strain crack hold for the
three-dimensional smoothed crack, too 8J].
Thus for the opening mode

1 /2
uJ = ~~ (
trt )
sin~[ 4(1-v)-2cos 2 ~ J (25)
where (r,~) is a local polar coordinate system in

a plane perpendicular to the edge.
199
Then a stress intensity factor is computeo

from equation (25) for 4' = 1t by means of the extra-
polation method.
kr f p0 frr b p0 =0125MPa ---- for Vp;; load

P1 /Po = 016 - - f o r p load
116
112
x1 •010Sm
Q4
Fig. 6. The stress intensity factor on the shorter

crack sioes
It is noted that at point x 2 = b KI ~ 0 because

at this point the contourof the crack is not smooth
and the stress singularity is not equal to 1/2.
Further details of the singularity calculation at
this point may be found in [14] •
4. INFLUENCE OF FREE EDGE ON VARIATION OF STRESS

INTENSITY FACTOR
The free edge near the crack engenders an

200
increase of the local stress due to the crack. It

means that the stress intensity factor will be
greater than in the case of creek in en infinite
space. The solution of such problems is of a great
practical importance owing to the fact that cracks
occur very often near free edges or on their sur-
face.
Smith [15] studied a semi-penny-shapeo surfa-

ce crack in a half-space by combining the Levy's
solution of a half-space with the solution for a
penny-shaped crack in an infinite space. The men-
tioned method cannot be used for finite regions
only for semi-infinite ones. More general numerical
methods lay high claims on the computers. In oroer
not to ex~gerate the claims on the computer the
problem of the state of stress for a crack in a fi-
nite region is split into two problems described
in the prece~ing paragraphs.
To illustrate the suggested method, let us

study a tensioned beam including a penny-shaped
crack and an elliptical creek in its interior.
As the first step let us calculate the stres-

ses at the place of the crack for a tensioned beam.
Hence, these stresses, only with opposite sign,
will be used as loads of the crack in the infinite
space and accoroing to the perturbation solution
we shall calculate the stresses on the beam surfa-
ce. These stresses with the opposite sign will be
loading the beam end again the stresses at the pla-
ce of the crack will be calculated. This procedure
will be repeated several times. The number of ite-
201
rations used will depend on how far the crack is.

from the external boundary of the finite region.
For the given region it is sufficient to cal-

culate only once the left side of the system of li-
near equations for different loads. Using this as
basis we can calculate integrals with a strong sin-
gularity. Once we know the displacements for the
given geometry under a certain load (e.g. constantl
we can calculate the diagonal coefficients of the
left side matrix.
We know analytical solutions for the penny-

-shaped crack G2J and the elliptical crack [13]
in en infinite space. The calculation of displace-
ments for constant loads on the basis of an analy-
tical solution is rather simple. It is not so sim-
ple in the case of a general load. For this reason
we calculate first the diagonal coefficients of the
left side matrix for a consta•t load and then for
the known coefficients of the matrix the displace-
ments due to the general load will be calculated.
The advantage of this procedure consists in elimi-
nating the sources of error consisting in the
calculations on singular elements.
The stress intensity factor for an elliptic

crack with ratio a/b = 2,5 and the free edge in-
fluence ratio b/h = 0,75 is normed by stress-in-
tensity factor for a penny-shaped crack for an
infinite space loaded by constant pressure cr0 •
The BIE method has been foun~ to offer an

efficient and accurate tool of fracture mechanics
202
analysis using a numerical solution of a set of

boundary constraint equations.
Kz /Kxe K1e- 2ao ffi

X
1.20
h /a -1,1
h
1,15
Fig. 7. Influence of free edge on stress-inte nsi-

ty factor for penny-shape d crack
203
....... ......
...,
' ''
''
' ',
'' ' a numerical results [3]
for a/b .. 2; b/h=O
\j)
0.7 L__-----'----~-----:L:-----=-:::------::!::-------;;!=--
30 60 90 120 150 180
Fig. a. Influence of free edge on stress intensi-

ty factor KI for elliptica l crack
Reference s:
[1] D.M. Tracey: 3-D elastic singulari ty element
for evaluatio n of K along an arbitrary crack
front; Int. J. Fracture 9, 1973, 340-343
[2] S.E. Benzley: Represen tation of singulari ties
with isoparame tric finite elements, Int. J.
Num. Meth. Eng. a, 1974, 537-545
[3] T.A. Cruse, G.J. Meyers: Three dimension al
fracture mechanics analysis, Journal of the
structura l division, Proceedin gs of the Ame-
rican Society of Civil Engineers , 103, 1977,
309-320
204
[4] T.A. Cruse: Application of the boundary in-

tegral equation method to three dimensional
stress analysis, Computers and Structures,
3, 1973, 509-527
[5] J.C. Lachat, J.o. Watson: Effective numerical
treatment of boundary integral equations: A
formulation for three dimensional elastosta-
tics, Int. J. Num. Meth. in Engng. 10, 1976,
991-1005
[6] o.c. Zienkiewicz: The Finite Method in Engi-
neering Science, Me Grew-Hill, London, 1971
[7] V. Kompis: Boundary integral equation method
for three dimensional elastostatic problems
and a formulation of the problem for large
displacements, Prog. of 14-th Jugosl. kongres
racion.: primenjene mechanike c, 1978, 113-
120
[8] C.A. Brebbia, R. Nakaguma: Boundary elements
in stress analysis, Journal of the Enginee-
ring mechanics division, Proceedings of the
American society of Civil Engineers, 105
No EM1, 1979, 55-69
J. Weaver: Three dimensional crack analysis,
Int. J. Solids and Structures, 13, 1977,
321-330
V.D. Kupradze: Metody potenoiala v teorii
uprugosti, Moskva, Fiz.-mat. izd., 1963
A.K. Gupta: A finite element for transition
from a fine to a coarse arid, Int. J. Num.
Meth. in Eng., 12, 1978, 35-45
~2] J. Sladek: Numerical methods in Fracture me-
chanics, Thesis, Slovak Academy of Science,
Bratislava, 1980 (in Slovak)
205
~J] M.K. Kassir, G.C. Sih: Three dimensional

stress distribution around an elliptical
crack under arbitrary loadings, Journal of
Appl. Mech., Trans. ASME ser. E, JJ, 1966,
601-611
[1~ V.L. Rvacev, v.s. Procenko: Kontaktnyje za-
daci teorii uprugosti deja neklasiceskieh
oblastej, Naukova dumka 1977, Kijev
~5] F.W. Smith et. el.: Stress intensity factors
for penny-shaped crack, Part I Semi-infinite
solid, Journal of Appl. Mech., Trans. ASME
ser. E, 34, 1967, 953-959
206
CYCLIC SYMMETRY AND SLIDING BETWEEN STRUCTURES BY THE

BOUNDARY INTEGRAL EQUATION METHOD
Anne CHAUDOUET
CETIM - France
INTRODUCTION
Seven years ago, CETIM developed a computer program based on
the Boundary Integral Equation method to analyse three dimen-
sional structures under mechanical loading. At that time, the
boundary conditions that could be taken into account by the
program were :
- given displacements on nodes, lines or elements,

- given tensions on lines or elements, and concentrated force~
- gravitation and centrifugal forces

The symmetries with respect to the surfaces x=o, y=o and
z=o could also be taken into account, and a subregioning
process allowed to analyse bodies made of different
materials, or to analyse contact between structures when the
contact zone was known and when the contact was perfect,
The advantages of this method are well known :

great saving of time in input and output operations since
only the surface of the structure has to be represented, and
results are given only on the surface and in areas described
by the user,
great saving in computing time since unknowns are only on
the surface thus leading to smaller systems than the ones
that are generated by the use of other methods such as Finite
Difference or Finite Element ones.
This is the reason why, since 1975, all the analy~s of three
dimensional comjact bodies have been performed by using this
program, I I1. 12 , 121.
I
207
Then, in order to be able to take into account thermal loading

the program was developed in such a way that elastic or
steady state thermal analysis could be performed by the same
program. The results of the steady state thermal analysis
being then used as data for elastic analysis under thermal
loading. 131 , lsi , 161.
In this paper, we present two of the latest models developed

at CETIM on its program based on the Boundary Integral
Equation Method to analyse three dimensional structures
cyclic symmetry and perfect sliding between structures.
SHORT REVIEW OF THE METHOD

Any linear partial differential equation with constant
coefficients can be transformed into a Boundary Integral
Equation.
Once the fundamental solution is known, by applying a reci-

procity theorem, a relation is obtained between the quantity
that appears as unknown in the linear partial differential
equation and other quantities on the surface. Then, by a
limiting process a relation between unknowns that lie only
on the surface is obtained (fig.!) 111.
Two examples are given below, the first one concerns elasti-
city problems whose Governing Equation is the Navier's
Equation, (fig.2), the second one concerns steady state heat
transfer problems whose governing equation is the Laplace
Equation (fig.3).
208
Governing Navier's Laplace's

Equation Equation Equation
Boundary
Integral B.I.E. B. I.E.
Equation
General process Process applied to Process applied

to obtain BIE Elasticity to steady state
fig.! fig.2 heat transfer
problems
fig.3
with Dirac function (fig.l and 3)

u displacement (fig.2)
8 temperature (fig.3)
The final equations obtained by this process are :

- For Navier's Equation :
Cij(x) uj(x)+ fs Tij(x,y)u/y)dsy= f

s
Ui/x,y)tj(y)dsy (I)
with : i 1,2,3 ; j I ,2 ,3
t tension
- For Laplace's Equation

C(x)8(x) + J T(x,y)8(y)dsy j U(x,y) '!(y)dsy (2)
s s
209
with : l£ = Flux
In those equations, the Kernels T .. and Tare varying as
1/R 2 and the Kernels Uij and U ar~Jvarying as 1/R in the
three dimensional case (fig.4).
141 '191 'I JOI 'III I' 1131'

s 1141
fig.4
To solve a particular problem three kin~of data are

needed :
- the geometry,
- the clBracteristicsof the material
- the boundary conditions (fig.S and 6)
u=u on Su , t=t on St 8=6 on sa ' '! = 'i on s'!

Elastic boundary Thermal boundary
Conditions Conditions
fig.S fig.6
with u, t, 6, i = Known quantities
The geometry is represented by B~nodes quadrangular elements

and by 6.nodes triangular elements, (fig.7). On each element
the unknown may vary linearly or quadratically I lSI.
210
D
triangular element
fig.7
quadrangular element
In the quadratic case, equations (I) or (2) are written at

each node of the elements describing the structure, where
the unknowns are located. In the linear case, this is
restricted to corner nodes. The number of equations written
at each node and the number of unknowns at those nodes is 3
in the elastic case, and I in the thermal case.
When using a subregioning process, (fig.S) equations (I) or

(2) are written for every subregion, with, on the interface :
lsi
fig.8
Finally, in the elastic case, equations (I) are written in

the direction of given displacements at each node.
CYCLIC SYMMETRY
When analysing a body whose. geometry is the same from one
sector to another around a given axis, and when the loading
is radial, it is possible to restrain the analysis of the
behaviour of the total structure to the one of only one
sector by imposing a normal displacement equal to zero on
every element of the radial surfaces representing the sector.
But when the loading is not radial (for example, torsion
problem), it is no longer possible to study the behaviour
of a sector alone with the following boundary conditions :
given tension or given displacement on the surface, because
on surfaces "inside the volume" (fig.9), tension and displa-
cements are both unknowns.
211
surfaces
inside the
volume
Representing the whole structure would lead to a very big

linear system of equations. If the boundary conditions
(tension and displacement) are the same from one sector to
another, by representing each sector by a subregion, it would
be possible to save computing time when building the linear
system of equations by writing that the matrix corresponding
to subregion I is the same as the one corresponding to the
first subregion provided all the subregions are the same and
the unknowns and equations at each corresponding node of
every subregion are the same up to a rotation. But during the
reduction of the system, the matrix would become fully popu-
lated and computing cost would remain high. A more powerful
way to proceed is presented here.
Let consider the structure of figure 9.
Only one sector is modelized (fig.IO).
On surfaces (A), which correspond to real boundary surfaces

of the body1 one group of equations is written at each node,
the unknown being either the displacement or the tension.
212
On surface (B), which correspon~to the interface between

subregion I and subregion 2 both displacement and tension are
unknown, On points on this surface a group of equations is
written considering the point as belonging to subregion I,
and another group of equations is written considering the
point as belonging to subregion 2, the number of equations
written at each node being then equal to the number of
unknowns.
On surface (C), both tension and displacement are unknown, but

only one group of equationsis written at each node, In case of
cyclic symmetry, the following relations between two corres-
ponding points (fig.IO) may be written.
'\,
ua ub (3)
'\,
ta tb (4)
with ua and ta displacement and tension at point a in a
,... 'V
given basis
and ub and tb displacement and tension at point b in a
basis deduced from the basis at point a by
a rotation of angle a around the z axis.
Using these relations, there is an equal number of unknowns
and equations :
- number of unknowns : 6 at point a
6 at point b
- number of equations: 3 equations ( I) at point a
3 equations ( I) at point b
3 relations (3)
3 relations (4)
Relations (3) and (4) can be taken into account in two ways :
- they can either be added to the initial system of equations
in which the unknowns are ua, ub' ta, tb on the cyclic
surfaces
- or they can be taken into account implicitly in the system
of equations by writing that unknowns at point a (point
belonging to the "master surface") are displacement and that
unknowns at point b (point belonging to the "slave surface")
are tensions ; the unknowns being expressed in the local
basis of the master surface, If we consider only the displa-
cement, in the first case, for any equation there is matrix
term Ui· for the unknown ua expressed in the bases Ba,
and thefe is a matrix term Uij for the unknown ub expressed
in the bases Bb • In the second case, for any equation the
matrix term is Uu· + tiij forth~ unknown ua, Uij being
expressed in the lasis B and aij being expressed in the
basis Bb deduced from Ba ~y a rotation of a around the
z axis.
213
The second method leads to a smaller system of equations than

the first one and it is this method that was chosen to be
included in the program. Furthermore, if the corresponding
points are in the same subregions, the structure of the matr~
remains the same as the one without cyclic symmetry and so
computing time spent to solve the linear system of equations
does not increase. Finally, a special treatment must be
performed for points on the z-axis. For these points, only
3 equations (for the elastic case) are written at each node
(the "master" node and the "slave" one being the same), and
there should be 6 unknowns. But on the z-axis we have :
- ur 0
0 (5)
0 or tz =0 on the "cyclic surfaces"
The only unknowns that remain are then : uz - t 8 - tr

with t 8 and tr : tensions on the cyclic surface, or tz - ~8 -
tz when the d~splacement along the z-axis is known, tz be~ng
a tension on an element perpendicular to the z-axis.
For the thermal case, the analysis is performed in almost the

same way as the one described above . . The differences are :
- no rotation is taken into account (the equations and
unknowns being scalars)
- the condition on the z-axis which aiminates one unknown
is : no flux on the cyclic surface on this axis.
Industrial example : Analysis of a head of Jack

Consider the structure shown in figure 12.
Fig.12
214
It is visible that, from a geometrical point of view, it was

possible to restrict the analysis of the structure to the
analysis of a sector of 90 degrees. As the loading of the
structure was not radial, it was not possible to give boundary
conditions on the surfaces x = o and y = o.
But as the loading was the same on each pawn, the analysis of
the structure could be restricted to the analysis of only one
sector by using cyclic symmetry.
The structure that was modelled is shown in figure 13.
z
on
surface
c
fig.l3
The mesh was composed of 2 subregions, 116 elements, 319 nodes

leading to a system of about 1200 unknowns.
As far as the meshing of the cyclic surfaces is concerned,

only the surface x = o was meshed ; the meshing of the
surface y = o being made automatically by the program in
order to have the most precise correspondence possible between
unknowns on these surfaces.
Zero normal displacement was given on elements corresponding

to contact zones between the head of Jack and the exterior
body (fig.l3).
Fives loadcases were taken into account, each one correspon-

ding to a different position of the head of Jack with respect
to the exterior body.
Each loadcase was composed of

- Internal pressure,
- Pressure on some elements of the cylindrical surface of the
pawn, this pressure (zone and intensity) was determined
earlierby a 2D calculation by the Finite Element Method
taking into account contact with friction phenomena.
215
- Traction on the same elements to take friction into accoun~
- Pressure on some elements on surfaces A and B in order to

have no momentum applied to the structure (those elements
were corresponding also to contact zones),
The conclusion of the study was that the stresses were too
important at the basis of the pawn for the force that had to
be transmitted. In order to come back to a reasonable level
of stresses in this area it was decided to increase the
diameter of the pawn,
The computation time needed for this analysis was 210

seconds (CP) on a CDC 7600.
SLIDING ON INTERFACE
When there is contact between two bodies, three kin~of
contact may exist :
- perfect,
- contact with friction,
- perfect sliding,
When the contact zone is known and when the contact is

perfect, the global structure may be analysed by writing
conditions of continuity on the displacement field and of
equilibrium on the tension field, in the contact zone.
When the contact zone is known and when the kind of contact
is perfect sliding, the conditions are (fig.J4).
In the normal direction : una = unb and tna - tnb

- In the tangential directions : tta = 0 and ttb =0
Subregion a
Subregion b
fig,l4
216
The six unknowns at each node on the interface between

subregion a and subregion b, in the contact zone are then
with n normal direction, and t1 and t 2 two tangential

directions.
When the displacement is known in a tangential direction then

the corresponding unknown is the tension on the surface in
that direction.
Finally, when the surface of the contact zone is not smooth

(fig.IS), then for nodes on line L sliding can occur only in
the direction d the unknowns are then :
displacement in the d direction in subregion a

and b
- u • displacement and tension in the normal direction
n1 of element I
t displacement and tension in the normal direction
nz of element 2
with t unknown tension on element I and t unknown

n1 tension on element 2 (t on elemeRt 2 and t
on element I beeing equ~1 to zero). nz
Subregion a
Subregion b
fig. 15
Theoretical example : cylindrical shaft sliding into a hollow

cylinder
To test the algorithm, we have compared numerical results to
analytical ones. Condsidering a shaft sliding into a hollow
cylinder (fig.l6) the upper surface of the shaft being
loaded with a pressure p ; the analytical solution is
217
- In the shaft :
_ vpr R2 2-R1 2
u r (r,z)
E
[ --2-
1-v
Rz 1
u z (r,z) .E
E
[ I - v2 Rz 2 -R1 2
Rz 1
(r,z) v p Rz -Rl
2 2
0 rr oee<r,z) =2 Rz
a
zz
(r,z) = p
arz = a ez = a re 0 ue 0
- In the hollow cylinder

r 2 2
u (r,z) -~ [(l-v)Rl 2 +(I+V)Rz;Rl]
r 2E R?
u (r,z)
z
= nE R1 2
R? z
v
arr (r,z) -2 p R1 2 [ R!z - ~J
v
oee<r,z) - 2 p R1 2 ( R> + ~J
with E : Young Modulus and v : Poisson Ratio
I
.!Lj
I
h p_2
r 111
I
r
fig.l6
218
The test was done with h = 10 mm

R1 10 mm
R2 20 mm
p - 30 daN/mm 2
E 21000 daN/mm2
v .3
The modelized structure is shown in figure 17.
- ....,,,
- - - 1',.......
~
'
'
Sector of 45 degrees
\
\
X
fig.17
The elements on surfaces e = 0 and e = 45° had zero normal

displacement. The comparison between analytical and numerical
results is given in table I for the displacements and in
table 2 for the stresses.
On those tables, we can see that the difference between
theoretical and numerical results is usually under 10 %
of the theoretical result. Furthermore, the biggest dif-
ferences can be noticed for midside nodes, the numerical
results at corner nodes beeing usually very near the theo-
retical ones.
Shaft Hollow cylinder
Theoretical Numerical result Theoretical Numerical result

Result Result
unn Min Max unn Min Max I
ur(r=5unn) 1.58 10- 3 1.53 10- 3 1.55 10- 3
ur(r=IOunn) 3.16 10- 3 3.13 10- 3 3.14 10- 3 3.16 10- 3 3.13 10- 3 3.14 10- 3
ur(r=l5unn) 2.42 10- 3 2.39 10- 3 2.40 10- 3
ur(r=20unn) 2.14 10- 3 2.11 10- 3 2.13 10- 3
uz(z=2.5unn) -3.33 10- 3 -3.31 10- 3 -3.33 10- 3 -8.03 10- 5 -7.97 10- 5 -8.41 10- 5
-3
u (z=5unn) -6.66 10- 3 -6.63 10- 3 -6.69 10 -1.61 10- 4 -1.55 10- 4 -1.62 10- 4
z
u (z=7.5unn) -9.99 10- 3 -9.37 10- 3 -1.00 10- 2 -2.41 10- 4 -2.27 10- 4 -2.35 10- 4
z
-2
uz(z=IOunn) -1.33 10- 2 -1.33 10- 2 -1.35 10 -3.21 10- 4 -2.99 10- 4 -3.25 10- 4
Table
~
co
0
"'"'
Shaft (*) Hollow cylinder i
I
Theoretical Numerical result Theoretical Numerical result
Result Result
daN/mm 2 Min Max dan/mm2 Min Max
cr r -3.375 -3.19 -3.47
cre -3.375 -3. 13 -3.63
cr r - 30 -29.7 - 30
crr(r=IOmm) -3.375 -3.15 -3.47
cr 9 (r=l0mm) 5.625 5,53 5.72
crr(r=l5mm) -.875 -1.21 -1.28
cr 9 (r=l5mm) 3. 125 3.07 3. 12
cr (r=20mm) 0 10-9 10-1
r
cr 9 (r=20mm) 2.250 2. 19 2.25
cr z 0 10-3 10- 1
x Results on the axis of the shaft are not taken into account.
Table 2
221
CONCLUSION
These two last developments enable us either to analyse the

behaviour of structures at lower cost (cyclic symmetry) or
to treat new kinds of problems (sliding on interface),
When analysing contact problems when the contact zone is not
known, it is obvious that the method described above cannot
be used right away, In that case an incremental method should
be used to determine the contact zone, But since up to now,
all inquiries for which sliding had be taken into account
were restricted to cases where the contact zone was known,
we have decided not to include the complete algorithm in our
program. We consider it more important to give new possibi-
lities to our program such as those to take into account
elastic support in elastic analysis, or thermal radiations
in thermal analysis.
222
BIBLIOGRAPHY
I 11 Boissenot, J.M., Lachat, J.C. and Watson, J,O. (1974)

Etude par equations integrales d'une eprouvette C.T.
15. Revue de Physique Appliquee, Vol.9, pp. 611-615.
121 Boissenot, J,M,, Gazagne, L. and Lange, D, (Mai 1977)

Some industrial applications of the boundary integral
technique in the field of 3-D elastostatics.
International symposium on innovative numerical
analysis in applied engineering science Versailles,
131 Chaudouet, A. (Juin 1980), Les Equations Integrales

de Frontiere : outil d 1 analyse mecanique de pieces
industrielles, 2eme colloque international sur les
developpements nouveaux dans les methodes numeriques
de l'ingenieur- Montreal.
141 Chaudouet, A. and Loubignac, G, (July 1979). Analysis

of bodies made of incompressible material under ther-
mal loading by the boundary integral equation method.
Numerical Method in Thermal Problems, Swansea.
lsi Chaudouet, A. and Loubignac, G. Boundary Integral

Equations used to solve thermoelastic problems
application to standard and incompressible materials.
Numerical Methods in thermal problems - Ed. John Wiley
& Sons Ltd,
161 Chaudouet, A, Loubignac, G, and Serres, D,(Octobre 80)
Utilisation de la methode des Equations Integrales
pour l'analyse elastique de pieces tridimensionnelles
soumises a des chargements thermiques. Congres AFIAP -
Paris.
171 Dubois, M, and Lachat, J.C. (1972). The integral

formulation of boundary value problems. Variational
Methods in Engineering. University of Southampton.
lsi Lachat, J,C. (1975). A further development of the

boundary integral technique for elastostatics,
Ph. D. thesis. University Southampton.
191 Lachat, J,C. and Watson, J.O, (March 1977). Progress

in the use of boundary integral equations illustrated
by examples, Computer methods in applied mechanics
and engineering. Volume 10 n° 3.
I 101 Lachat, J.C. and Watson, J,O, (1975). A second gene-

ration boundary integral equation program for three
dimensional elastic analysis, (Appl. Mech. Div.
ASME II, New-York,00-00,
223
BIBLIOGRAPHY
I 111 Lachat, J.C. and Watson, J.O. (1976). Effective nume-

rical treatment of boundary integral equation
Int. J. Numer. Metho. Eng. 00.
I 121 Lange, D. (13 Janv. 1978). Fracture analysis using

the boundary integral equation method. Com. Interna-
tion. Conf. Num. Methods Fracture Mech., Swansea, 9.
I 131 Rizzo, F.J. and Shippy, D.J. (1977). An advanced

Boundary Integral Equation method for 3D thermoelasti-
city. Int. J. Num. Meth. Eng. Vol. 10 p. 1753.
I 141 Shaw, R.P. (1974). An integral equation approach to

diffusion. Int. J. Heat Transfer. Vol. 17 p. 693.
I lsi Zienkiewicz, O.Z. (1978). The finite element method

in engineering science. Third edition Me Graw Hill,
London.
224
ON BOUNDARY INTEGRAL EQUATIONS FOR CIRCULAR CYLINDRICAL SHELLS
H. Antes
Ruhr-University, Bochum, West-Germany
ABSTRACT
The so-called direct approach establishes boundary integral

equations with the aid of Somigliana's boundary identity. This
identity is based on the Betti's reciprocal work theorem and
the so-called fundamental solutions of the basic equations. For
elasticity problems in general these singular solutions are due
to point loads.
In this work for the linear theory of thin shells a new

reciprocal theorem is deduced. There the interaction energy of
two elastic states is expressed by derivatives of the displace-
ments and the stress functions.
The knowledge of stress functions, which solve the equi-

librium equations for the cases of point loads directed along
the three coordinate lines (Antes 1980), gives the possibility
to derive from this new reciprocal theorem boundary integral
formulations for the solution as well of the geometrical as of
the statical boundary value problem. Here this derivation is
realized for the case of circular cylindrical shells.
BASIC EQUATIONS
Let ~ be the simply-connected interior of the middle surface

of a shell and r its boundary. In an operator-matrix formula-
tion the equilibrium conditions for the stress resultant o,
the vector of the symmetric membrane forces naS and bendi;g
moments m<aS) are (Antes 1980)
T K* T
~ · (£ + £ ) = D • o + £* 0 ( 1)
in the interior ~. and

225
(2)
on the boundary r , the part of r where static terms are pre-

scribed. PNote that on the boundary r only four quan-
tities can be given independently. These shall be indicated
by asterisks.
These and the following equations are written using vec-

tor and matrix notations to get as short and clearly arranged
expressions as possible. The explicit forms are given for the
operator-matrices in an appendix and for the vector quantities
in the table.
The relations between the displacement components ua and

the stretching and bending strains a(aS) and S(aS)' respective-
ly, are almost identical
D • N • u = £
- ,...., ,...., -
+ E:Q* (3)
The so-called extra-strain E:Q is produced by an inhomogeneity,

an incompatibility ~* of th~ compatibility equations:
T Q* T
12 • ~ • (£ + £ ) = 12 • ~ • £ + ~* = 0 (4)
The equations (4) are identically satisfied by each u if the

relations (3) are used. Besides, ~must fulfil ~n ru the
prescribed geometric conditions
u = g* • (5)
-G • ,....,
R
,....,
A conclusion from analogy gives
(6)
a rule to determine stresses out of functions ~' the so-called

stressfunctions ~' which solve the homogeneous~equilibrium
equations (1) id~ntically.
These differential equation systems (1) and (4) with

their pertinent rules (5) and (3), respectively, to gain homo-
geneous solutions, are connected by the constitutive equations
K. £ (7)
~
-1
K £ (8)
~
It is easy to show that it is possible to describe the whole

problem by only one equation system.
If we use the relation (3), which solves the equations (4)
identically, and the constitutive equations (8) we can substi-
tute the stress 2 in the equation system (1)
226
T -1 Q*
D
- -
• K
- -
• (D • N • u - E
....... .......
) p* = .......
+ ....... 0 (9)
This equation (9), the reduction of the two systems (1) and (4),
can be called the Navier's equation of the linear shell theory.
Alternatively we can use the relation (6) to fulfil the

equilibrium equations (1). Then, expressing the strain E with
the help of the constitutive equation (7) and this relation
(6) by stress functions ~' the only equation system to solve
will be ~
(10)
In three-dimensional elasticity theory an equivalent system

is called Beltrami equation.
n. 11
a ( 11)
n.22
a (22)
n.12
2a ( 12)
m (11)
8 (11)
m(22)
8 (22)
m ( 12)
28 ( 12)
Schematic presentation of the basic equa-

Table
tions in a linear theory of thin shells
227
This table gives a review about connections between

these basic equations, similar to the diagrams given from
Tonti (1972) for a large class of physical theories.
RECIPROCAL THEOREM
The foundation of the above-mentioned direct approach to inte-

gral equations is Betti's reciprocal theorem combining dis-
placements and internal as well as external loadings of two
different elastic states. The original basis is the equality
of the double internal deformation energy and the external
work (Brebbia 1978).
For shells this can be expressed in the vector-matrix

notation as follows
j E:T · a drl = juT · p* drl + j (R · u) T · VT · (R · a) df • (11)

n~ ~ n~ ~ r -G ~ - -s ~
Introducing so-called modified strains
( 12)
E = B • E:
- ~
and modified stresses (Antes 1980)

T -1
= (~ ) • £ or (13)
the double internal deformation energy may be rewritten as
(14)
Using the relation (6)

K* T
£ + £ =B • <I + I K* ) = ~ T . !2 • ~ • j!_ ( 15)
it is possible to transform the surface integral (14) by

Stokes' theorem (Antes 1979)
j[ (DT . E) T . TjJ - ET IK*]drl +

n - ~ ~ ~
+ Ir (R-s • E) T •
~
v . (~G • ;J!,l df
-
( 16)
If we have problems without incompatibilities, that is, with

homogeneous compatibility equations (4)
DT · B · E:
- - f"'oJ
= DT · E
- I"Y
= 0 in n
~
(17)
the equation (16) will be reduced to

228
J~T • <!, + !_Kl drl =

f <~s • ~l
T
. y . <~G • ;t> dr . ( 18)
n r
If we consider two different states the interaction energy is
given equally by
1 T 2 1T 2K*
fr <~s • ~l · v • (R • 1/Jldr-
-G ~
JE •~r
n~
dn . (19)
Since the constitutive equations (7) and (8) are uniquely

invertible the equation (19) can be written as the following
reciprocal theorem:
1 T 2 1T 2K*
f( (R-S • ~) • V • (R
-G
• 1/1) df -
~
jE
Q~
·!:
~
drl
2 T 1 2T 1K*
= j(R
r -s
·E) · V • (R
~ - -G
·1/i)df-
~
jE
n~
•!:
~
dQ. (20)
Here only modified strains E and stress functions 1/1 are used
to express the unknown quantities. This new formulation of
the reciprocal theorem can be combined with the well known
Betti's reciprocal theorem, for shells in the here introduced
vector-matrix notation given by
1 T T 2 1T 2*
Jr (R-G • u) · V • (R
~ - -s • cr)df +
~
fu · ~p
n~
drl
2 T T 1 2T 1*
= f<R ·ul
r -G ~
·V
-
• <~ 8 -~)dr+J~
n ·,e drl. (21)
If we apply the equation (11) the interaction energy formula-

ted in equation (19) can be given in the following well known
form
JE1T
n~
2
•<Jdrl =
~
j(R-G
r
•U)
~
1 T
•V
-
T
· (R
-s
2
~
1T 2*
·cr)df+}u •p dQ.
n~ ~
(22)
This yields immediately a mixed formulation of the reciproca~

theorem:
1 T T
Jr (R-G • u)
~
• V
. 2 T 1 2T 1K*
j(R ·E) ·V· (R ·'l')df-jE ·!: drl. (23)
r -s ~ -G ~ n~ ~
229
BOUNDARY INTEGRAL EQUATION
The above derived reciprocal theorem (23) can be the basis of

a boundary integral equation for the unknown state "1". With
it the necessary assumptions are weaker than for Betti's reci-
procal theorem. It is only to postulate from the singular
state "2", the fundamental state caused by singular unit
loads in the point x
"'P
2*
p a 6(x- ,....., x )ek
,..,p,....., k = 1,2,3 , (24)
2 2
that the modified strains ~ and the stress 2k are known in
the interior S"l and on the whole boundary f. This is much less
than the knowledge of the displacement field i·
For each unit load in the point ~p we get from equation
(23) an integral representation of the displacement component
~k in this point x
"'P
l l T T 2
u (X
k "'P
) =
rf[ (R • 1jJ (y) )
-G ~ ~
• V • (R 5 · Ek(x ,y))
- - ~ "'P~
-
(25)
Collecting the three vectors ~ and Zk• respectively, in ma-

trices
~* (~p·,¥) [E 1 (x ,y), ,.....,

E 2 (x ,y), ,....,
......p,...., E 3 (x ,y)]
......,p,...., (26)
~ "'P ~
s* (x ,y) [a 1 (x ,y), a 2 (x ,y), a 3 (x , y) ] (27)
- "'P ~ ,...., "'P,...., ,...., "'P,...., ~ "'P ~
the three equations (25) can be combined to one vector integral
equation
u
T (x ) =
f [ (R • 1jJ (y))
T • v T • (~s • ~* (~,x_l)-
~ "'P r -G ~ ~ - "'
-(R ·u(y))T.vT. (R 5 ·s*(x ,y))]dr-

-G ~ ~ - - - "'P ~ y
(28)
This equation states an integral relation between the unknown

displacement field in the interior and the unknown boundary
reactions.
If the boundary integral disappears, for example, because

all boundary values of the fundamental state are zero - such
a fundamental state is often called Green's state -, the dis-
placement field 2 can be calculated directly by equation (28).
230
But in general it is necessary to determine the unknown

boundary reactions ~G · ;t on r and/ or ~G · ~ on r u. A boundary
integral equation can solve this problem. For that purpose in
the equation (28) the singular point ~p is to move towards the
boundary f: xp E D ~ E r.
~
y
~
Dependent on the discontinuity of RS · s* (y ,y) and

~s . ~* (z.zl for z
~ y and on the smoothness-ort 'he boundary r
(Hartmann 1980) the Yimiting values of the boundary integrals
yield
-(R · lj!(y)) T · VT · (R · G*(y,y)) ]df

-G ,...., "' - -s - ,...., ,...., y
Here III
denotes the Cauchy principal value of the boundary
integral. The characteristic matrices ~ 1 (X) and ~ 2 (il are
calculatedas the limiting values (Hartmann 1980)
I V- T * ,.- -,. . , ) df Y
. , ,y. . , = ,y. . , + pw)
lim · (R · S (y, (30)
-s -
p ~or
p
lim IY.T. (~s. ~*(z,z=x + p~))dry (31)

P-+ or P
With these limiting values and the relation (29) we get from
the integral relations (28) the following three boundary inte-
gral equations, here given as one vector equation
(R ·u(y)l+III(R ·u(y))T·VT (R_8 ·s_*(Y.•Y)ldfy

-G ~ ~ f G ~ ~ - ·- ·-
- -
IQ-G* (y,x) ·I K* (x)dD . (32)
,....,,...., ,...._, ,...., X
In the linear shell theory at any point of the boundary four

quantities are given and four reactions are unknown. The
essential geometric components are u 1 , u 2 , u 3 and au 3 ;av.
They determine uniquely the boundary values ~G · ~- Formally
the same is valid for the static boundary values ~G · ~. Thus
it is necessary to obtain a fourth equation.
For that purpose we form the directional derivative of

the third component equation of the integral relation (25) in
231
a fixed direction z, with respect to the coordinates of ~p

(Bezine 1978): ~
3u 3 3E*
T
·az-
~3
dZ
(x )
~p
f[ (~G • _t (x_) ) T ·V <~s (X
"'P~
,y))
r
3cr*
- (R · u(y)) T . v
-G ~ ~
T
<~s . az- ~3
(x ,y)) ]dr
"'P ~ y
3E 3
- J _::_ (x ,x)· rK* (x) dQ (33)
Q dZ "'P~ ~ X
When x tends to a point y of r we put z = where is the v v

unit ~tward normal vector at the point~y. Analogou~ to the
above given derivations limiting values are calculated.
The result, the fourth boundary integral equation
3u 3 3cr*
T ~3
d\1 <fl +C3(fl~G. ~(fl + ifi(R ·u{y))T·V

r -G ~ ~ - <~s av (y,y)df=
~~ y
3E*
T
·av-
~3
=C (y)R · l)J(Y) + IJI<R ·l)J(y))T·v <~s (y,y)df
4 ~ -G ~ ~ r -G ~ ~ - ~~ y
3E 3
- JQ _::_
d\1
<y_,?:_{) • LK* (x) drl
·- ·- ~ ~ X
(34)
can contain kernels which behave like 1/r 2 . In this case we

revert to the integral relation (33) and integrate by parts
with respect to the curvilinear abscissa s tangential to the
boundary f(Bezine 1978). This procedure gives new kernels
like 1/r. Thus these integrals can be interpreted as Cauchy
principal values.
Note that, in general, fundamental state des-

criptions, where the modified strains ~ are not derived
from the pertinent displacement field );l,k, \vill not satisfy
the equation (17). Then the corresponding integral term (see
equation (16)) must be supplemented in the reciprocal theo-
rems (20) and (23), in the integral relations (25) and (28),
and moreover in the boundary integral equations (32) and (34).
FUNDAHENTAL STATES FOR CIRCULAR CYLINDRICAL SHELLS
If the above given integral formulation of boundary value

problems is used, at least the modified strains ~ and
the stresses £k must be known. It is not necessary to have
the total description of the fundamental state. Both these
quantities are determined if singular stress functions .tk
have been found which solve the equilibrium equations (1) for
the cases of unit point-loads.
232
It is known (Chernyshev 1963) that the behaviour of

shells in the neighbourhood of singular load points is almost
like the behaviour of plates and disks, respectively. Thus
the stress functions ~a of the plate and the Airy's stress
function ~ 3 of the disk must be appropriate starting-points.
For circular cylindrical shells the basic equations given

before are simplified because coordinate lines can be
chosen such that the metric tensor aas is the unit tensor caB
and the only non-zero component of the curvature baS is
b22 = 1/R.
The necessary condition for the fundamental state "2"
l
uk (x ) +
f (R l T
• ~) • vT
2
<~s · f!..ldf (35)
~p r -G
2
becomes substituting thy stress function ~ by equation (6) and
the displacement field ~ by the relation (3)
T l
I (D. N. ~)l T (~ • !? • ~ • ;J:.l d
2
uk(x ) +
n- - ~p
+ f (R • u)
l T
• V
T
• (~S •
2
f!._) df (36)
r -G ~ -
Considering the special case of a circular cylindrical shell

and assuming homogeneous boundary conditions this equation
(36) is explicitly (Antes 1980):
fn[u1,1 <~3,22 +
1
R ~2,2l - u1,2<~3,12 +
1
R ~2,1l
1 1
+ u2,2(~3,11 - R ~1,1) - u2,1 (~3,12 - R ~1,2)
(37)
For singular normal point-load p 3 Eo(~- ~p)~ 3 the correspon-

ding stress functions of a thin plate can be used. We assume
free coefficients in order to have the possibility to consider
all shell influences (x x1 - x1p• y = x2 - x2pl
2
a 1 'l" arctan yX + b(xlnr + c1x (38)
a 2 x · arctan y + biy ln r 2 + c 2 y
.?!.
. (39)
233
The condition (37) determines a 2 1/2n and requires
w< 3 > ,;_.!.

3 R
Iw<2 3 > (x,y)dy =
1 2 2 2
=-R [a 2 xy • arctan yX + (b 31 x + b 32 y ) ln r ] . (40)
Analysing the equilibrium of a small neighbourhood of the

singular point ~p (Antes 1976), for example in the direction
~1
-11 -11 a -12 12

Ia[n <a,y> - n <-a,y> ]dy + I [n <x,b> - n: <x,-b> Jdx
y=-b x=-a
b1
= !b[R <w2,2(a,y) -w2,2(-a,y))+(1jl3,22(a,y) -w3,22(-a,y))]dy+
a 1
+ f[R <w 2 , 1 (x,-b)- w2 , 1 (x,b))+ <w 3 , 12 <x,-bl-w 3 , 12 (x,b))]dx=o,
-a
(41)
we get a1 = -a2. The other coefficients are not determined by

these statical considerations. Thus the behaviour of the
corresponding strains must be studied. In order to get real
boundary integral equations, the compatibility equations (17)
have to be satisfied in their homogeneous form
T T
D ·~=D -~·£ 0 (42)
For circular cylindrical shells these conditions are explicitly
1
0 (43)
'R a<22l ,1 + 13 <22> ,1 - 13 <12 > ,2
0 (44)
= 0 (45)
The free coefficients of the stress functions should be chosen

in such a way that at least no essential singular terms rest
in DT • E. If this is not possible completely, the remaining
terms ~* can be considered as corning from an extra-strain £Q*
(46)
Then, integrating by parts these terms we get

234
For this case instead of the reciprocal theorem (23) the

following formulation is to be used as the basis for the inte-
gral equations:
1 T T 2 lT 2
Jr (R-G • u)
~
· v- · (R-s • cr)dr
~
+ fu · ~p*
rl~
drl =
f(R
~ -s
(~~ -EQ*llT
~
· v · (R
-G
• ,T,ldr +
~
+j[EQ*T · f + (EQ*-
n,.....; . -. .; ""-J
~) T
f"oJ
· fK* ]drl
,..._,
(48)
Here we have found the following three stress functions as

the best possible ones, provided we use the approximations
(38), (39) and (40):
1/J (3) X ( l+V) 2 ( 1-v) ]

[- y. arctan -4-- X. ln r + -- 4- x , (49)
1 21T y
1/J (3) X ( 1 +V) 2 (1-v) ]

X• arctan - 4 - y · ln r + --4-y I (50)
2 21T y
( 3) 1 1 ( 3-v) 2 2 2
1/J
3
=---
R 21l
[xy ·arctan ~ + 1 6 (x - y ) ln r ]
y
(51)
By adequate considerations and calculations the stress func-

tions ;t(l) and ;t( 2 l for the unit-point loads in the direction
of the other two coordinate lines can be found too.
CONCLUSIONS
In this paper new formulations of reciprocal theorems for the

linear theory of thin non-shallow shells are given. These can
be used, as hasalso been show~ to formulate boundary integral
equation if the fundamental states have been found. The re-
search has not yet been finished. Thus it was not possible to
check the practicability by numerical example-s.
235
APPENDIX
The operator matrices, which are used to get short expressions

for the basic equations of the linear shell theory, are given
explicitly in the following. There Da signify the covariant
differentiation Ia along the coordinate lines ea. As usual aaS
and b~ mean the covariant metric tensor and the curvature
tensor, respectively. va denotes the components of the outward
unit normal vector.
1
0 -Vlbl 0 0
0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
F
-1
K K
0 0
236
2
(1+V) a12
F syrnm.
Et
(1-v) (a12) 2
( a 11 l 2 12 11
a a
11 22
+va a
12 22
-1 Et syrnm. a a
~ =~
1-v 11 22
- 2- a a
0 0 -1 0
b2-b1 0 -1 0 0
2 1
b1 0 0 0
2 2
a
0 0 0 0
0 0 0 0
1
0 0 0 0
2
0 0 l
N 0
0 0
0
-1
j
237
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
:= 0 0 0 0 0
~s
0 0 0 0 0
0 0 0 D1 0 D2
0 0 0 0 D2 D1
0 0 )
0 0
R := 0 0
-G
-b1 -b2 -D1

1 1
-b1 -b2
2 2 -D2J
REFERENCES
Antes, H. (1976) tiber singulare Lastfalle in einer linearen

Schalentheorie und ihre finite Behandlung. Ingenieur-Archiv
45, 99- 114.
Antes, H. (1979)0n Dual-Complementary Variational Principles

Generated from Dual Functionals in Linear Shell Theory. Acta
Mechanica, 33, 55-67 (German, English summary).
Antes, H. (1980) tiber Fehler und Moglichkeiten ihrer Abschat-

zung bei numerischen Berechnungen von Schalentragwerken (Habi-
litation thesis). Mitteilungen Institut fur Mechanik 19, Ruhr-
Universitat Bochum, W.-Germany.
Bezine, G. (1978) Boundary Integral Formulation for Plate

Flexure with Arbitrary Boundary Conditions.Mech. Res. Comm. 5,
197-206.
Brebbia, C. A. (1978) The Boundary Element Method for

Engineers. Pentech Press, Plymouth, London.
238
Chernyshev, G. N. (1963) On the action of concentrated forces

and moments on an elastic thin shell of arbitrary shape.
J. Appl. Math.Mech. 27, 172- 184.
Hartmann, F. (1980) Elastische Potentiale in Gebieten mit

Ecken. Dissertation, University Dortmund, W.-Germany.
Tonti, W. (1972) On the Mathematical Structure of a Large

Class of Physical Theories. Rend. Accad. Naz. Lincei. Class.
Sci. fis. mat. nat. 52, 48-56.
239
THE BOUNDARY ELEMENT METHOD APPLIED TO TWO-DIMENSIONAL

CONTACT PROBLEMS WITH FRICTION
Torbjorn Andersson
Linkoping Institute of Technology, Dept of Mech Eng,

S-581 83 Linkoping, Sweden
INTRODUCTION
Contact problems and the study of load transfer in mechanical

assemblages are of great importance in mechanical engineering.
For nearly one century, since H Hertz (1881) published his
famous work on normal contact between elastic bodies, much
research has been performed in this area, both theoretical and
experimental work. An interesting survey of the mechanics of
contact between solid bodies is given by Kalker (1977). He
gives an account for the classical formulation of the contact
problem as well as for the variational one. The latter has been
used especially for numerical calculations, for instance with
the finite element method, (FEM). In most contact problems the
contact area is a function of the external forces. When fric-
tion has to be taken into account, the whole load history has
to be followed. Thus when using numerical methods, contact
problems have to be solved by iteration and in the frictional
case also with incremental technique. A lot of computer time
has to be spent which makes it important for the system matrix
to be small. When sliding occurs in the contact zone, the nor-
mal and tangential forces have to be coupled with a friction
parameter. The possibility of coupling the normal- and tangen-
tial tractions in the matrix system is an advantage for the
Boundary Element Method (BEM) over the displacement FEM. This
makes it fruitful to use BEM for solving contact problems with
friction.
This paper describes a technique for applying BEM to two-

dimensional contact problems with friction. The pressure and
shear stress distribution between two elastic bodies in contact
and the resulting stress distribution for the case of no body
forces are calculated. By use of a standard boundary element
program the relations between increments for displacements and
tractions at the boundaries of the bodies are determined.
Considering the friction properties inside the contact area,
the two systems of equations are linked together and solved
240
numerically. To get the correct contact area consistent with

the external forces the load is applied in increments corre-
sponding to the element model of the contact. In each increment
the correct adhesion and slire zone are calculated with iterative
technique. The solution from the contact program can be treated
by a standard BEM-program for calculating the stresses and
strains of the two bodies.
Applications are made to some practical problems and the

BEM solutiomare compared to solutions from other methods.
The influence of body forces are easy to take into con-

sideration. The general equations can be extended to three-
dimensional problems.
BEM-FORMULATION OF THE TWO-DIMENSIONAL CONTACT PROBLEM

WITH FRICTION
Basic relationship A B
Consider two linearly elastic bodies n and n bounded by cy-
lindrical surfaces. The traces gf the goundary surfaces in the
(x 1 ,x 2 )-plane are the curves f and f respectively. The
geometry and all variable quantities are independent of the
x 3-coordinate. The problem is thus considered as a plane pro-
blem in the (x 1 ,x 2 )-system, Figure 1.
Figure 1 Definition of the problem. Two bodies in

contact along the boundary rc
The tractions acting on the boundaries are denoted t. and

the di~placement by u .• The contribution due to body forc~s is
neglected in the follBwing presentation. On the boundary fc
241
the two bodies are in contact with each other. The contact
region can be separated into two parts. On one part res sliding
occurs i.e. the surfaces in contact move relative to each
other. This part is called the slip region. On the other part
rca' called the adhesion region, no sliding occurs. Thus
rA rA + rA
c ca cs
rA and rB are the contact boundaries in the uniefo:wed state.

Agcordin~ to small displacement theory, r ~ r % f .
c c c
The relation between increments of boundary values for
displacements and tractions in the local coordinate system,
defined in Figure 1, may be written Andersson (1980)
f Hk.. k
(P,Q)~u.(Q)df(Q) = f Uk.. (P,Q)~t.(Q)dT(Q),
k k=A,B (1)
k l.J J k l.J J
r r"
From the previous load state a normal gap between the contact
boundaries can be calculated, Figure 2. This gap is denoted
u~'n where n is the load step under consideration.
Figure 2 Definition of local coordinate system, the gap

o n
u 1 ' and the parameter a
k
Four new displacement components v. (k A,B) in the local
coordinate system can now be definea.
242
(2a)
(2b)
a is a function of the location of the contact boundary and

is defined in Figure 2.
In the slip region the relation between the normal- and

tangential traction is defined by the coefficient of friction
]1.
(3)
The sign in (3) depends on the relative displacements so that

the tangential force works against the relative displacement
and energy is dissipated.
The friction coefficient )1 is assumed to depend on the

total effective slip ve, Fredriksson (1976)
Jli
]1( ve ) = ]1m {1 - (1 - - ) . e-hve } (4a)
Jlm
v b.v (4b)
e e
The parameter h is the hardening coefficient. In the case when

the initial friction ]1. and the limit friction ]lm are equal
the ideal Coulomb slip 1 is assumed, Figure 3.
]1
ve
Figu.e 3 Friction coefficient as a function of
effective slip
When the friction coefficient is a function of the slip, it is

also a function of the load. Hence, ]1 is continuously changing
during the loading. In the BEM model the load is changed in
finite increment. The new )1 values for each element pair are
then computed after each load step. If the "exact" ]1 value has
to be used in each load step, iterations should be performed to
get those values.
243
Since this iterative technique needs a lot of computer

time the ~ values are computed approximately from the pre-
vious load step without iterations. An effect of this technique
is that an element which changes the contact status from adhe-
sion to slip or slip to slip with changed ~ value has a residual
tangential traction
AtE,n = ! n-1 n-1 n-1 (5)

'"" ~ tl - t2
which in the slip to slip case also can be stated, Figure 4
,___
+ (,n-1 _ n-2) n-1 (6)
.. ~ tl
t2
slipsurface
::...___
t3
3 2
t2 + !::,tE,3
2 2
t2
2
tl + !::,tE,2
2 2
tl
2
0 tl
Figure 4 Contatt conditions for a pair of contact

elements
Second
_,.,_
First load step 0-1: adhesion
1-2: slip
Third - II - 2-3: slip
In slip state the increment in tangential traction is
(7)
This ensures that after each load step the condition for slip,
t 2 = : ~t 1 is valid, see Figure 4.
The contact conditions are then stated

244
r
A B
/1v1 + L'lv1 0, /1tA - L'lt1B 0,
A
t1 < 0, t1B<O (Sa)
c 1
A B L'ltA - MB= 0
r 11v 2 + 11v 2 0, (Sb)
ca 2 2
k + k E: k
r
cs
/1tA - L'ltB=
2 2
o, M2= -l.l11tl + 11t 2 • (k = A,B) (Sc)
The contact variables for body B in equation (1) are

now eliminated by use of the contact conditions, equation (8).
If these contact variables are written without superscript A
the following system of integral equations is obtained
= f u~. 11t~dr + f u~1 ./1t. df + f (u~1 1 ~ l.l ~1• 2 )M 1dr +

A A 1J J A J J A
r -r c r ca r cs
f NTA.. o,n df
~ 1J ul +
fA Ui2
A
u
•ts,nar
2 (9a)
fA T
c cs
+ B
f u~. /1t.df +
J
f u~.
1]
M.dr +
J
f B(U~l l.lUi2) M 1dr +
BB J 1
rB
1
r -r r
c ca cs
B o,n
- f (1-a)Til ul df + f Bu~1 2 11l·rar
2 (9b)
rB r
c cs
In equations (9) there are for every point on the

boundary one pair of variables
rk- rk k k
( L'lui' L'lti) (lOa)
c
rk (11v. , /1t.) (lOb)
ca 1 1
rk k +
cs (/1vl, /1tl), (11v2 , l.J/1tl) (lOc)
245
Outside the contact region there is cne relation for one un-
known quantity and inside the contact region there are two
relations for two unknown quantities.
The problem is now solvable since, for every boundary

point, there is one relation for every unknown quantity.
The parameter a in equations (9) represent the correct

contact boundary and can be calculated if ~v 1 are set to zero
(u~'n # 0). The solution is however performed practically by
giving trail values to a, for example 0.0, 0.5 or 1.0 and
solving for the misfit ~v 1 in the contact area.
The problem is linear if the contact-, adhesion- and

slip areas are assumed to be constant during the load incre-
ment. Hence, the principal of superposition is valid and the
problem can be separated into two parts. One part depends on
the two last integrals in (9a) and (9b) and the other on the
load increment. The last two integrals represent the work needed
to get the new contact area forced together and the residual
tangential traction to complete the total traction to perform
t 2 = ±vt 1 • With this separation the solution can be done for
a load increment up to the total load and scaled afterwards.
In the discrete case the scaling is made up to the point where
two new elements are contacting each other or where an element
pair gets positive normal traction, Figure 5.
Figure 5 Scaling of the load increment. The scaling

parameter 8 is calculated from the conditions
for positive normal traction inside the con-
tact region or geometrical incompatibility
outside the contact region
246
Until now it has been assumed that the adhesion and slip
zones are the correct ones and that the tangential traction in
the slip zone has the right sign. One way to find the correct
solution in each increment is to get a solution for an assumed
contact condition that is not violating itself. In the assumed
adhesion zone the total tangential traction has to satisfy
Jt 2 J < ~Jt 1 J. In the slip zone the friction force should work
against the relativeAdisplHcement increments. The condition
sign (t 2) # sign (~u 2 + ~u 2 ) has thus to be fulfilled.
If the contact conditions are not violated the load can

be incremented and the displacement and tractions succesively
calculated as
E:,n n n 1 (lla)
~t + S~t + t -
n E:,n n n-1
u ~u + S~u + u (llb)
In equation (11) subscript i is omitted.
Boundary element discretization

For the numerical solution of equations (9) the integrals are
divided into a sum of integrals over line segments or elements
of the boundary. Over each element the geometry, displa<:~ments
and tractions are assumed to vary according to interpolation
functions Xm, ~m and ~m respectively. The interpolation func-
tions give the variation over the element in terms of values
in discrete points or nodes. If ~u. x. ,
and ~t. are the
nodal values for the element the ~~riatt~n can b~mwritten
x.
l.
I'm.
l.
~
m
~u.
l.m
, ~t.
l.m
(12)
From
o,n equations
. (2) and (12) it can be concluded that
l. and
~v.
u1 are l.nterpolated by ~m·
The interpolation functions are now substituted into equations

(9). Integration is performed over the boundary elements. If
two variables from different elements act in the same node the
results h~ve to be adged (assembled). The total number of nodes
is N on r= and M on r . The summation is performed over N-N
and M-M nodes outside the contact area and over N and M godes
inside the contact area. Note that N M and thaf thesecnodes
have to correspond to each other. c c
It is now more convenient to work with matrix notations.

If the point P successively coincide with every N+~f node
2 x (N+M) relations can be established. Collecting the contact
variables in the two left hand side vectors the whole system of
equations can be written in matrix notation as
A [x, Sy] = B [c, Sd] (13)

or written out in extensio
247
~a T~sl 0 -~a(-Uls~~~s) -~a 0 ]~uA 'SuA

-B
- 1 -B - B-+ - B
-- B B
-
T2a 0 I TB-ula(-Uls-~U2s) -u2a T2s ~vlaiSvla
~vlslsvls
[~
0
UB
-TA
0
1
0, 0
UBs -TB
J 0
0
GtA
StB
~v2alsv2a
2s 1
A I
A
o,n ~v2s Sv2s
au 1 0
~ E:,n
~uB !SuB
0
t2
I
M1 St 1
(1--a) u~' n 0 aj a
Mlsj Stls
M2al St2a
B B
~v2sl 8v2s
(14)
These are two linear systems of equations, each one with

2 x (M+N) relations for 2 x (M+N) unknown nodal quantities.
Thus it is solvable if rigid body motion is suppressed.
NUMERICAL SOLUTION AND THE COMPUTER PROGRAM SYSTEM
Forming and solving the system of equations

The numerical solution technique is described for two-dimen-
sional contact problems with friction. The boundary element
used is the constant displacement and traction element with one
node at the center of the element, Figure 6.
n_j~
i&1VV1~V<RkkRxd
\ element limit
Figure 6 Boundary element used

248
Even if the formulation is made for a special element it can

easily be extended to elements with shape functions of higher
order and also to three-dimensional contact problems. The fact
that using shape' functions with higher order than linear may
imply geometrical incompatibilities near the contact boundaries.
The same effect is found in FEM-fracture mechanics, Backlund
and Aronsson (1978). This limits the advantage of such elements.
The two bodies under consideration are divided into boundary
elements in the usual way, see for example Brebbia (1978). In
the expected contact area the pair of contact elements opposite
to each other should be modeled in the same length.
With a standard boundary element program the relation

of displacement versus traction in the (x 1 ,x 2 )-system can be
formed for the two bodies A and B. For each pair of contact
elements we define local coordinate systems (~,n). The ~-axis
is determined from the normal vectors of the two elements, ~A
and fiB respectively, (see Figure 7) •
...,A
,.A n - "B
n
(15)
n~
,~ - n . . B,
Figure 7 Definition of local coordinate systems for

a pair of contact elements
The initial gap ~uo,n between the elements referred to the

(~,n)-system is ca~culated from the vectors between the element
nodes in the last increment.
-
o,n = (r - u
-A - -B
u ) (16)
u~ A-B
r -B is the vector between the nodes in undeformed state.

Tfte variables which belong to the contact area are transformed
to the (~,n)-system. The two matrix systems are then linked
together in accordance with the contact conditions, equation
(8), to from the equation (14). Outside the contact area
249
traction or displacement are known. The variables can be re-

arranged so that the unknown variables are placed in the left
hand side vectors and the known quantities in the right hand
side vectors. The right hand side vectors which arise after
matrix multiplication are called [b 1 , 8b 2 ] and the unknown
vectors are called [y 1 , 8y 2 ]. Hence, if the system matrix is
called A, the system of equations simply becomes
(17)
If the rigid body motion is suppressed, the equation (17) can

be solved for [y 1 , 8y 2 ].
Scaling of the contact solution

By use of the equations (2) and (8) all increments in the
boundary conditions are known for the two bodies. The total
displacement and traction are calculated from equations (11).
For each contact pair, 8 can be calculated to give zero contact
pressure E,n n-1
/}t +t
8 (18)
For the element pairs in close vicinity to the contact zone,

8 are calculated for zero gap, i.e. these new elements are
contacting each other.
[ -rA-B - L:
('uE,n +
u
k=A,B (19)
8
Depending onwhether the load is increased or decreased the

smallest of the 8-values from equation (18) or the largest
S-value from equation(l9) are selected and the total state
is calculated for that 8-value.
Validity of assumed contact conditions

After scaling the solution the assumed condition for adhesion,
slip and sign for the frictional tractions has to be verified.
The condition for the adhesion area is
(20)
and for the slip area where energy dissipates
sign (t 2) f sign (l}u~ + /}u~) (21)
If these conditions are satisfied the correct contact solution

is found. If not, new contact conditions have to be chosen and
the solution procedure repeated once again. After some itera-
tions the correct contact solution is found for the load state
250
and the problem is thereby solved. It remains to be proved that

the solution found with this technique is unique.
Stresses and strains inside the two bodies

The BEM solution for each increment found from the friction
program is written to disc store. From this solution stresses
and strains inside and on the boundary of the two bodies may
be calculated using an ordinary BEM-program, see for example
Brebbia (1978) or Andersson (1979).
Computer program
A computer program system for solving elastostatic contact
problems with friction using the Boundary Element Method is
developed. The program system is built up of three programs
and data are transformed between the program by use of disc
store, Figure 8. The programs are written for two-dimensional
plane strain or plane stress problems and linear elastic
material properties. The shape functions for describing dis-
placement and traction variation over each element are assumed
to be constant.
The programs are:

ISOTROPIC A conventional BEM-program
GENERATOR A computer program for generating the T-
and U-matrices
FRICTION A computer program for solving two-dimensional
elastostatic contact problems with friction
using incremental and iterative technique.
Input
Input
Displacement Displacement
Traction Traction
Stress
Figure 8 Configuration of the computer program system

for solution of elastostatic contact problems
with friction
251
APPLICATION
In order to examine the accuracy of the described technique

three problems have been solved. The first two solutions are
compared to analytical solutions for the frictionless case.
The third problem is a typical engineering problem where no
other solutions are available.
Problem 1
An elastic roller on an elastic foundation is considered. The
BEM model of the two bodies is shown in Figure 9.
50 element
E 4000 [N/mm 2 ]
v = .35
R = 50 mm
tany = .20
50 element
E 4000 [N/mm 2 ]
v = .35
H II I II':' I l l ' ' ' I lilt:
Figure 9 Elastic roller on an elastic foundation,

BEM models
The non-linear behavior between contact length and applied

force P1 is compared to Hertz solution in Figure 10. The con-
tact pressure and shear stress distribution for two different
load cases with varying ~-values are shown in Figure 11. The
actual parameters for the BEM solution are compared to Hertz
solution in Table 1. For the BEM solution of the frictionless
case, see Andersson(l980). The contact pressure seems to be very
little influenced by the friction properties.
252
~ 10.0
.a 8. 0
:r:
~ 6 .o
z BEM solution
~ 4 • 0 1------.,,ff--+-
E-< Hertz:
2
~ 2 • 0 1-.c'f----+- p = 'ITEb
~ 32(1- v 2 )R
8 0.0 L-________. _________. _______~
0 250 500 750
FORCE P [N/mm]
Figure 10 Contact length versus applied force.

Comparison between BEM and Hertz solutions
100~---------~---------~---------~----------,
Shear stress
-100L-.--.-----~---------~---------~---------~
-5.0 -2.5 0 2.5 5.0
CONTACT LENGTH x [.nun]
Figure 11 Elastic roller on elastic foundation, contact
pressureandshear stress distribution on the
foundation
253
Load
case CD 0 G) ®
load
step 8 8 9 9 Nr
)1 .05 .01 .25 .30
pl 591 591 0 0 N/mm
p2 0 0 549 549 II
2
pmax 95 95 90 90 N/mm
Pmax 93 93 88 88 II
Hertz
b 8.0 8.0 8.1 8.1 mm
bHertz 8.1 8.1 7.8 7.8 II
Table 1 Hertz problem with friction. Comparison

of the BEM and Hertz solutions (Hertz: )1 0)
Problem 2
A circular disc is placed in a cylindrical cutout of an infinite
plate. The radii of the hole of the plate and the disc are
approximately the same. In the BEM calculations the clearance
is assumed to be ~R/R = 0.003. The BEM oodel is shown in Figure
12.
Circular disc with Infinite domain

center hole with hole
50 elements 2 40 elements 2
E = 4000 N/mm E = 4000 N/mm
'J = .35 'J = .35
R1 = 50 mm R2= 50.015 mm
Ri/R1 = 0.05
Figure 12 Circular disc in a circular hole of an

infinite domain, BEM models
254
In the center of the discthere has to be a hole where the load is

applied. The disc is loaded by a force Q and a moment M. The
contact pressure and shear stress distribution for three differ-
ent cases with varying ~-value and moment are shown in Figure
13.
-'rr"'
.6
I...-..-~ r--·~1
,.....-
/.I
......•\-t•~-
•1• r-:~."."
<3?,\-
3
.4
v~r:~r ~/
Ar•<·~,_iJ- /
l•/
r
0
.2 ~~ ,(
()I
()I
;::l.
........ ........
.o
l IPI@
0 ~
~
@
-e- -e-
d
t-e
.
0
0..
~. ~I
1-'
p:: p::
~
p
Ul
Ul Ul
-.2
~~ ~
~d>
M/~QR
-
l'~o
~
CY
Ul
"' 'o--o
p.. ~ -.4
.0 .0
-
8 Ul
I :}
0 .4 .0
z~ ril~
• Contact pressure
0 Shear stress
G) .4 .5
0 ::r:
0 Ul -.6
-90 -60 -30 0 30 60 90
CONTACT ANGLE ¢0
Figure 13 Contact pressure and shear stress distribu-

tion for the circular disc
The analytical solution without friction is given by Persson

(1964) but also calculated with BEM by Andersson(l980). The BEM
solution for ~ = 0 is in good accordance with those solutions.
The friction has in this problem a clear influence on the maxi-
mum contact pressure since the frictional forces take a part
of the load.
Problem 3
Load transfer in mechanical assemblage is of practical interest
255
since high stress gradients often occur and the stiffness pro-
perties often determine the load transfer. A problem where a
part K of the load P is transferredbetween the sheets by a steel
rivet is shown in Figure 14.
Steel rivet
Al-sheet
(l-K)P ~
Figure 14 Load transfer in a mechanical assemblage
The middle sheet is assumed to be made of aluminium. No bending

effects in the rivet or the sheet are considered. In Figure 15
the BEM model of the problem is shown.
Dl_··_ ·~·-· 0 ·_· I~

(l-K)P '\symmetry line._-- t<P
Al-sheet Steel rivet

50 elements 2 30 elements
E 69000 [N/mm ] E 2lOOOO[N/mm 2 ]
\) = .3 \) = .3
R = 5 mm R = 4.995 mm
length = 8R R./R = .2
width b = 4R
~
Figure 15 Steel rivet in an aluminium sheet, BEM model

256
The contact pressure and shear stress distribution for two

~-values and different load transfer K are shown in Figure 16.
2.5
.-·
I I
~ = 0.0 ~ = 0.2
./
K=l.O y
/
2.0 ~---r-1-~~---+----+---~----,
•
• K = 1.0
•
I 0.8 ·--- ·-·
~-·-.............
ill . . . .-·--·-
• / 0 8
~I \
1.5
r--t.·4~ ·.----·~-·,\-
·-·-
II I ·---......._
·-·-
0. 6
.
__ ......
~
0.6 /.yl·\~
r 'I"
!'{,...
0.4
·-·
./
•
.r·-·~
1.0 -t~·~~\ -·--v·/;-;. . .~
~~·IY
1-
\ ·"·r.._ ·-·
.
0.
\n.z •'•...... /
\
-t---\ •,
/.
0 0
;t:~~8:_tf_~ i
b b
........ ........ 0.5
0.. 1-'
·y
1\
"
I • ·;;-·-·-•c:.;:::~,·
I ·-·-.".J...i:::
- ........ J.....•-
11 l/ ~r·--·1 •
& '• _&'//; -~ .4/
0.0 ~--~-----+----~~~~~~~--~
.-~ ·-~,............
'·-·/ ./. 2
•
Contact pressure
• Shear stress
a0
p
=b = lOO[N/nmh
I I I
-0.5
90 60 30 0 30 60 90
CONTACT ANGLE ¢
Figure 16 Contact pressure and shear stress distribu-

tion for different values of the load
transfer
The friction properties influence the contact pressure in

accordance with the load transfer by the friction forces. In
all cases the clearance is assumed to be ~R/R = 0.001 and the
contact area starts growing from ¢ = 0.
257
CONCLUSIONS
In the paper two boundary integral equations for elastic con-

tact including friction between two-dimensional plane struc-
tures have been presented. The initial normal gap in the con-
tact zone can be treated in both of the equations using a para-
meter a. The influence of applied load as well as the initial
gap and residue tractions are treated as two different loads
resulting in two linear equation systems. The solution is after-
wards found by superposition and a scaling technique determin-
ed by the element model of the contact region.
The discretization procedure with a simple boundary

element and the numerical solution gives good agreement to
analytical solutions. The choice of a in the numerical solution
has no influence in the presented problems. In the case of an
illconditioned matrix system for one of the bodies it might,
however, have a numerical effect. The parameter a could then
be used to avoid illconditioning.
The incremental and iterative procedure works well for

the presented problems. For a more general contact problem
where the contact, slip and adhesion zones are changing discon-
tinuously,it would be necessary to restart from a previous
load step with an intelligent guess.
REFERENCES
Andersson, T. (1979) The Boundary Element Method with Appli-

cation to Contact Mechanics, Linkoping Institute of Technology,
Department of Mechanical Engineering, Report R-135, Linkoping.
(In Swedish)
Andersson, T. (1979) MARCUS*ISOTROPIC, TA*GENERATOR and

TA*CONTACT. Computer package for Analysis of Elastostatic,
Frictionless Contacts Between Plane St~uctures - Users Manual,
Linkop'ing Institute of Technology, Department of Mechanical
Engineering, Report R-136, Linkoping, (In Swedish)
Andersson, T. (1980) "The Boundary Element MethodApplied to

Two-dimensional Contact Problems". Proceedings of the Second
International Seminar on Recent Advances in Boundary Element
Methods, Southampton March 1980, Ed. C.A. Brebbia, CML
Publications
Argyris, J. H., et al. (1977) ASKA User's Reference Manual,

ISD Report No. 73, Rev. D, Stuttgart
Brebbia, C. A. (1978) The Boundary Element Method for Engineers,

Pentech Press, London
258
Backlund, J. and Aronsson, C.G. (1978) Effects of Geometrical

Incompatibilities on Stress Intensity Factors Calculated by
the Finite Element Method. Proc. First International Conference
on Numerical Methods in Fracture Mechanics (eds A. R. Luxrnore
and D. R. J. Owen), Swansea 9-13 Jan., 1979, pp 281-291
Cruse, T. A. (1977) Mathematical Foundation of the Boundary

Integral Equation Method in Solid Mechanics, AFOSR-TR-77-1002,
Pratt and Whitney Aircraft Group, United Technologies Corpo-
ration, East Hartford, Connecticut 06108
Fredriksson, B. (1976) Finite Element Solution of Surface Non-

linearities in Structural Mechanics, Computers & Structures,
6: 281-290
Kalker, J. J. (1977) A Survey of the Mechanics of Contact Be-

tween Solid Bodies. ZAMM75, T3-Tl7
Persson, B. G. A. (1964) On the Stress Distribution of

Cylindrical Elastic Bodies in Contact. Diss. Chalmers Inst.
of Technology, Gothenburg, Sweden
Rizzo, F. J. (1967) An Integral Equation Approach to

Boundary Value Problems of Classical Elastostatics. Q. Appl.
Math. 25, pp 83-95
259
QUASISTATIC INDENTATION OF A RUBBER COVERED ROLL BY A RIGID

ROLL - THE BOUNDARY ELEMENT SOLUTION
R. C. Batra
Department of Engineering Mechanics, Univ. of Missouri-Rolla
ABSTRACT
The linear elastic problem involving the indentation of a

compressible rubberlike layer bonded to a rigid cylinder and
indented by another rigid cylinder is analyzed by the boundary
element method. For the same contact width, the pressure at
the contact surface is found to depend noticeably upon the
thickness of the layer and the Poisson ratio of the material of
the layer. Results computed and presented graphically include
the pressure distribution over the contact surface, the shape
of the indented surface and the stress distribution at the bond
surface.
INTRODUCTION
Traction in vehicles, nip action in cylindrical rolls in the

paper-making process and in the textile industry, and friction
drives are examples of the kind of problem studied here-
in. Each of these problems involves indentation, by a steel or
granite cylinder, of a rubberlike layer bonded to a cylindrical
core made also of steel or granite. Hertz's (1881) solution
for determining stresses in two bodies in contact is applicable
when both bodies are homogeneous and linear elastic and the
dimensions of the contact region are small compared with the
dimensions of the bodies. The contact problem involving two
rollers, one of which is assumed perfectly hard, in contact
with their axes parallel is a special case of Hertz's problem
and has been investigated by Thomas and Hoersch (1930) on the
assumption that plane strain state of deformation prevails.
Solutions of the problem when the deformable roller is not
homogeneous but consists of a thin elastic layer bonded to a
hard supporting core has been given by Hannah (1951), and Hahn
and Levinson (1974). Whereas Hannah's approach involves the
numerical solution of an integral equation, Hahn and Levinson
260
use an Airy stress function and obtain the solution in the

form of a double series one of which converges rather slowly.
In the aforementioned studies the rubberlike layer is

taken to be linear elastic and the deformations involved are
presumed to be infinitesimal so that the linear theory applies.
Also for the case of infinitesimal deformations, Batra et al.
(1976) solved the problem by taking the rubberlike layer to be
made of a thermorheologically simple material. In order to
explore the effects of material and geometric nonlinearities,
Batra (1980, 1981) recently studied the problem in which the
rubberlike layer is modeled as a homogeneous and incompress-
ible or nearly incompressible Mooney-Rivlin material. We note
that Batra used the finite element method to solve the problem.
Experimental work involving varying thicknesses of the

rubberlike layer and different combinations of the diameters
of mating cylinders has been carried out by Spengos (1965).
However, the only material property listed for the rubber is
the durometer hardness. For the linear elastic problem involv-
ing incompressible material, the durometer hardness is suffi-
cient to find Young's modulus and hence solve the problem.
Since the length to diameter ratio for the rolls used by
Spengos is of the order of one, the assumption of plane strain
state of deformation made in the work referenced above is not
valid. For such geometrical configurations one needs to solve
the three dimensional problem.
It seems that the Boundary Element Method has an advantage

over the Finite Element Method for three dimensional problems
since at least the data preparation for the former method is
easier. With the ultimate aim of solving the three dime~sional
problem, in this paper, we solve the two dimensional (plane
strain) problem by the Boundary Element Method. Results
obtained from the use of the boundary element method are found
to compare very favorably with those obtained by Hahn and
Levinson (1974).
Formulation of the problem

The system to be studied and the location of rectangular
Cartesian axes are shown in Fig. 1. Since the length of rolls
perpendicular to the plane of paper is large as compared to
their diameters we assume that plane strain state of deforma-
tion prevails. Also the deformations of the rubberlike layer
are presumed to be infinitesimal so that a linear theory
applies. The rubberlike layer is assumed to be made of a homo-
geneous elastic material. In the absence of friction at the
contact surface, the deformations of the rubberlike layer are
symmetrical about the line joining the centers of two rolls.
Neglecting the effect of body forces such as gravity,

equations governing the quasistatic deformations of the rubber-
261
Elastic Layer Rigid Material
Fig. 1 System to be studied
like layer are
p = p0 (1 - u. .), (1)
J.,l.
(i,j 1,2) (2)
Ev
(3)
at the inner surface s 1 (xixi = R1), (4)

at the outer surface s 2 (xixi= R0 );
e 1 o1 jnj 0, (5)
n.o .. n. 0, x1 ~ b (6)
l. l.J J
nioijnj p(x1' x2)' x2 < b, (7)
lim p(x 1 , x2) = 0, x2 -+- b, (8)
and at the plane surface s 3 (x 2 = 0);
u2 = 0, (9)
012 = 0, (10)
where
(11)
262
In these equations, p is the mass density of the material

particle in the deformed state, Po is its mass density in the
undeformed state, ~ is the displacement of a material particle,
E and v are respectively Young's modulus and Poisson's ratio
for the material of the rubberlike layer, a comma followed by
an index i indicates partial differentiation with respect to
xi, the summation convention is used, § is a unit vector
tangent to the surface, n is an outward unit normal to the
surface, oij is the Kronecker delta, cri. is the Cauchy stress
tensor and the function p gives the pre~sure distribution at
the contact surface. The condition (8) implies that the normal
pressure at the boundary of the contact surface vanishes. This
ensures that a contact problem rather than a punch problem is
being solved. We note that the semicontact width b and the
pressure p (x 1 ,x 2 ) at the contact surface are not known a
priori but are to be determined as a part of the solution. As
has been employed by Hahn and Levinson (1974), the boundary
condition (7) can be replaced by a displacement type boundary
condition. We find it convenient to use the boundary condition
(7) rather than the equivalent displacement type boundary
condition.
BOUNDARY ELEMENT FORMULATION OF THE PROBLEM
Taking the inner product of both sides of Equation (2) with wi,
integrating the resulting equation over the domain n consisting
of the region occupied by the layer above the plane x 2 = 0 and
using the divergence theorem, we obtain
nJ (L.1J·u w. . ),t uk dQ-

1 .J an J (gkuk- f.w.) ds
1 1
(12)
where
f. LijU uk,t nj' (13)

1
gk Lijkt w1,]
. . nt. (14)
Details of deriving Equation (12) from Equation (2) are given

by Brebbia (1978). We now choose wi to be a solution of
( ) - ~(m)= 0 (15)
Lijkt wi,j 't k
where ~~m) is the Dirac delta function and represents a unit

load at point m in the k direction. With this choice of wi,
Equation (12) becomes
u~m) + J f. w. ds = J gkuk ds (16)

1 an 1 1 an
Note that wi and gi are displacements and tractions due to a
unit concentrated load at the point m in the k direction.
Considering unit forces acting in the three directions, Equa-
tion (14) can be written as
263
u~m) + J w.~ f~ ds = J G.k uk ds (17)

1 an 1 an 1
where Wi~ and Gi~ represent the displacements and tractions in

the ~ direction due to a unit force acting in the i direction.
Equation (17) is valid for the particular point m where these
forces are applied. Expressions for Wi~ and Gi£ are given in
Brebbia's book (1978). In Equation (17) the integrations are
over the boundary of the domain. At points of the boundary,
either surface tractions fi or displacements ui or a suitable
combination of the two are known. Equation (17) enables us to
solve for the unknown surface tractions or unknown surface
displacements at each point of the boundary.
COMPUTATION AND DISCUSSION OF RESULTS
In order to solve the problem by the boundary element method,

the boundary of the rubberlike layer in the first quadrant is
divided into a large number of segments as shown in Fig. 2.
Displacements and surface tractions within each segment
(boundary element) are assumed to be constant. The mesh is
finer within approximately twice the semicontact width. As has
been shown in earlier studies (Batra, 1980, 1981), the boundary
conditions on the vertical plane have essentially no effect on
the pressure profile at the contact surface and the deforma-
tions of the rubberlike layer within the vicinity of the
contact region. In the results presented herein the vertical
plane is taken to be traction free. The computer program given
in Brebbia's (1978) book for constant boundary elements has
been modified to solve the present contact problem.
Half nip width b and the form of the pressure function p

are presumed. With this additional data the problem is well
defined and can be solved. Having solved the problem a check
is made to insure that the deformed surface in the assumed
contact zone matches, within a prescribed tolerance, with the
circular profile of the indentor and that the nodal point just
outside the contact area has not penetrated into the indentor.
If the second condition is not satisfied implying thereby that
the nodal point just outside the presumed contact width has
penetrated into the indentor, either the value of b is
increased or the total load is decreased. However, if the
second condition is satisfied and the first is not, the values
of p at various nodal points are adjusted so that the deformed
shape of the assumed contact area conforms to the circular
profile of the indentor. The deformed surface of the roll
cover is taken to match with the profile of the indentor if
the distance of each nodal point on the contact surface from
the indentor is less than 1.5 percent.of the indentation uo
(see Fig. 1). Usually, with a little experience one can make
pretty good estimates of b and p (x 1 ,x 2 ) so that the entire
process converges in four or five iterations. The total load
P is obtained by integrating p over the contact area.
Test cases To ensure that the discretization of the boundary
was adequate, the contact problem with R0 = R = 454.7 mm, R1
441 mm, b = 22.9 mm, E = 8.96 x 10 5 N/m 2 , and v = .3 was
solved. The pressure on the contact surface matched exactly
with that obtained by Hahn and Levinson (1974). Stresses on
the bond surface could not be compared since such information
is not given in Hahn and Levinson's work for v = .3 and our
formulation is not valid for v = .5.
Next we solved the problem for a different set of material

and geometric parameters, namely that Ro = 61 mm, R = 76.2 mm,
R1 = 46.5 mm, b = 12.7 mm, v = .45 and E = 1 x 106 N/m2. Batra
(1981) has solved this problem by the finite element method.
The pressure profile at the contact surface came out to be
virtually the same by the two methods. For this case we plot,
in Fig. 3, the variation of the normal stress orr and the shear
stress ore at the bond surface with the angular distance from
the center line of rollers. These stresses decay to essen-
tially 0 as e1e 0 approaches 3. Thus the assumption that the
vertical plane is traction free is verified to be true. In
this and all other figures eo = b/Ro·
Effect of thickness of the layer Keeping Ro = 61 mm, R = 76.2

mm, b = 12.7 mm, v = .3 and E = 1 x 10 6 N/m2 fixed, the value
of R1 was varied. The total load required and the resultant
indentation for various thicknesses of the layer are listed in
Table 1. As expected, the load required to keep the contact
width constant increases as the thickness of the layer
decreases. For thinner layer the indentation is less even
though the load is more. However, the indentation/thickness
increases with the decrease in the thickness of the layer.
Indentation
Thickness/R 0 Load P/(Gb) Indentation/b Thickness
.2343 2.162 .2784 .2486
.2092 2.182 .2550 .2550
.1883 2.216 .2368 .2631
.1674 2.272 .2190 .2738
.1454 2.303 .1974 .2820
.1172 2.633 .1802 .3218
Table 1. Load and indentation for different thicknesses

of the layer
Fig. 4 depicts the pressure profile at the contact surface

for two different thicknesses of the layer. Note the change in
the shape of the curve near the ends of the contact width as
the thickness of the layer is reduced. With the decrease in
the thickness of the layer it seems that the pressure at the
center increases more than the decrease in the pressure near
the ends so that the total load increases. Such a change
265
in the curvature of the curve near the ends of the contact

width is also apparent in Meijer's (1968) work which studied the
indentation, by a rigid cylinder, of a layer fixed to a rigid
plane surface.
In Fig. 5 is shown the deformed surface of the rubberlike

layer for two different thicknesses of the layer. Note that
the scales along the vertical and horizontal axes are differ-
ent. Therefore, the undeformed position of the rubberlike
layer plots as an ellipse rather than as a circle. It seems
that the change in the curvature of the deformed surface near
the ends of the contact zone is heavily dependent upon the
thickness of the rubberlike layer.
Effect of Poisson's ratio To investigate the effect of

Poisson's ratio on the load required to cause the same contact
width the values of Ro = 61 mm, R = 76.2 mm, R1 = 46.5 mm, b =
12.7 mm, and the shear modulus G = 3.45 x 10 5 N/m 2 are kept
fixed. The total load required and the resulting indentation
for various values of v are listed in Table 2. These data
show that
Poisson's Ratio Load/(Gb) Indentation/b
.41 1.878 .2850

.42 1.951 .2844
.43 2.030 .2840
.45 2.162 .2784
.47 2.154 .2599
.48 2.152 .2506
.49 2.147 .2414
Table 2. Load and indentation for various

values of Poisson's ratio
the total load increases with the increase in the value of

Poisson's ratio upto a certain value of v. As v is increased
beyond .45, the total load changes little but the pressure
distribution on the contact surface differs in that it is more
at the center and less near the ends. This becomes clear from
Fig. 6 in which we plot the pressure distribution at the
contact surface for two different values of v.
Remarks
We should emphasize that the results presented above and the
conclusions drawn are applicable only to the geometric config-
uration studied herein. For other values of geometric para-
meters, one will expect similar qualitative and not necessarily
quantitative results.
For the present problem the CPU time and the core require-
266
ment for the finite element solution and the boundary element
solution were virtually the same. In the boundary element
solution of the problem the grid used had 10 elements across
the thickness of the layer and 35 elements across the circum-
ference. The boundary element solution perhaps gave slightly
better values of tractions at the bond surface. The grid data
in both caseswere generated internally in the computer program
and required the same effort.
Even though it is usually said that the finite element

solution generates a lot of unneeded information about stresses
and strains within a body, to us, it seems to be an advantage
especially in a problem like the one studied here. In such a
problem one does not know a priori where the maximum principal
stress or the maximum principal strain will occur. It is more
expensive to compute stresses and strains at internal points
from the boundary element solution than it is from the finite
element solution.
REFERENCES
Batra, R. C., Levinson, M and Betz, E. (1976) Rubber Covered

Rolls--The Thermoviscoelastic Problem. A Finite Element
Solution, Int. J. Num. Meth. Engng. lQ: 767-785.
Batra, R. C. (1980) Rubber Covered Rolls--The Nonlinear Elastic

Problem, Trans. ASME- J. Appl. Mechs., ~: 82-86.
Batra, R. C. (1981) Quasistatic Indentation of a Rubber Covered

Roll by a Rigid Roll, Int. J. Num. Meth. Engng. ~: (to appear).
Brebbia, C. A. (1978) the Boundary Element Method for

Engineers, John Wiley & Sons, New York.
Hahn, H. T. and Levinson, M (1974) Indentation of an Elastic

Layer Bonded to a Rigid Cylinder--I. Quasistatic Case Without
Friction. II - Unidirectional Slipping with Coulomb Friction,
16: 489-5.
Hannah, Margaret (1951) Contact Stress and Deformation in a

Thin Elastic Layer, Quart. Journ. Mech. and Applied Math. ~:
94-105.
Hertz, H. (1881) J. F. Math. ~: 156-171.
Meijers, P. (1968) The Contact Problem of a Rigid Cylinder

on an Elastic Layer, Appl. Sci. Res. ~: 353-383.
Spengos, A. C. (1965) Experimental Investigation of Rolling

Contact, Trans. ASME- J. Appl. Mechs., ~: 859-864.
Thomas, H. R. and Hoersch, V. A. (1930) University of Ill. Eng.

Exp. Sta. Bull. 212.
267
IO
g r id
e le m e n t
B o u n d a ry
F ig . 2
268
Qj
<J
"'
.....
"";:J
Ol
'1:1
U) en en
U)
"': 2v 0
1::
,.0
II II II Qj
D .c:
VI ctl;lo .j..J
.j..J
"'
Ol
Qj
Ol
ct~IQ) Ol
Qj
""
.j..J
<ll
<"")
00
..-i
""'
.4 8o =.209
71 =.45
r=
10::: .3
.....
0
0:::
:!:.
~
.....
1.&.1 .2
0:::
::::>
en
en
1.&.1
0:::
Q.
.I
0 .25 .5 .75
e
8o
Fig. 4 Pressure distribution at the contact surface; effect of layer thickness 1\)
!B
2 "'-..s
0
1.5
y
ROBO
UNDEFORMED
DEFORMED SURFACE
(S; =.766)
.5
DEFORMED
SURFACE (S;= .883
.92 .96 .98

X/Ro
Fig. 5 Deformed surface of the rubberlike layer

.9
,.,
~
Ia:
0
'a:
.6
..!
-C)
..._,
Si= .766
Ql .209
"i .3~
.... Bo=
a.
0 .25 .5 .75 1.0

J}_
Bo
Fig. 6 Pressure distribution at the contact surface; effect of Poisson's ratio
rv
-..J
272
AN IMPROVED BOUNDARY ELEMENT METHOD FOR PLATE

VIBRATIONS
George K.K. Wong and James R. Hutchinson
Graduate Student, Mechanical Engineering Department, Stanford
University
Professor, Civil Engineering, University of California, Davis
ABSTRACT
The boundary element method is used to determine the natural

frequencies and mode shapes of thin elastic plates of uniform thickness
with arbitrary boundary conditions and arbitrarily shaped edges. The
boundary element formulation leads to a pair of integral equations
that contain the boundary condition variables of displacement, normal
slope, bending moment, and effective shear, providing an effective
means of solving plate problems with arbitrary boundary conditions.
The boundary elements are taken as straight lines or as circular arcs,
and unknown boundary functions are assumed to vary linearly along
the elements.
INTRODUCTION
In a 1979 paper by the authors [ 1], a formulation was presented by

which the natural frequencies of thin uniform elastic plates could be
found. The formulation was limited to simply supported and clamped
edges only, and was also limited to straight line segments. The
numerical results from that research, however, were very encouraging
and this present research is an attempt to remove many of the
limitations from the previous solution method. The problem has,
therefore, been reformulated for boundary condition variables of
displacement, normal slope, bending moment, effective shear, and
twisting moment discontinuity at corners. Also, in addition to straight
line segments, one has the option of using circular arc elements.
The formulation of integral equations for static loads on plates

was presented by Bezine and Gamby [2] in 1978, and independently
by Stern [3) in 1979. The present formulation for the vibration
problem closely follows Stern's development. Previous research
employing boundary elements for plate vibration include a recent
paper by B~zine [4], in which the problem is treated by a mixed
boundary element-finite element approach. In a 1974 paper by Vivo1i
and Filippi [5], the vibration problem is formulated in a similar way
273
to that used in this paper, but there are differences in both the
formulation and application of the integral equations.
At the present time, the computer program which we have

developed is not yet fully operational, but preliminary results indicate
that accurate frequencies can be found with a small number of
elements.
FORMULATION
The governing partial differential equation for the free vibration of

a homogenous, isotropic, linear elastic thin plate of uniform thickness
is
(1)
where v4 is the biharmonic operator, D is the flexural stiffness, w

is the vertical displacement, p is the mass per unit area, t is the
time and n is the area of the plate bounded by curve r.
The time dependence of Equation (1) is removed by assuming

that w varies sinusoidally with time; thus
w(x, y, t) = w(x, y) sin (wt) (2)
Substituting Equation (2) into Equation (1), the final time-independent

differential equation is
4
Vw-A.w=O
4 for w in n (3)
where A. is the frequency parameter defined as

/..4 = w2p/D (4)
and w is the circular frequency of the plate. Equation (3) is the

governing differential equation for the natural frequencies and mode
shapes of the vibrating plate. Let E be a singular solution that
satisfies Equation (3) for any point inside of n. Hence,
v4E- t.. 4E = 0 forE inn (5)
Multiplying Equations (3) and (5) by E and w respectively, then

subtracting and integrating over the region n gives
JJ(EV 4w-w'IJ4E)~ = 0 (6)

n
Equation (3) can be reduced to a boundary integral by making use of
the Rayleigh-Green indentity with K corners [2,3] as:
274
fJ(EV 4w-wV 4E)dA

n
= orn
.!qy (E)w- M (E) dwd + ddE M (w) - E Vn(w)}ds
n n nn
K
L < Mt(w) > E - < Mt(E) > w =0 (7)
k=l
where V (w), M (w), M (w), ~wand w are the effective shear, normal
bending nmomenq, twisting ml6ment, normal slope and displacement
respectively, and < •> represents the discontinuity jump in the
direction of increasing arc length.
Using polar coordinates (r,e) and letting S be the angle from the
radial direction to the outer normal yields
M (w) = D{
- -O+v)'V 2w + 0-v) ~( - aw + -
I - I -a2w - a2w ) cos 2S
n 2 r ar r2 ae2 ar2
- 2 .L
ar
aw )
( .!r ae sin 2aJ t
r (8-a)
Mt(w) = - D(l-v) {< .! aw + .!_ a2w- a2w ) sin 2S

2 r ar r2 ae2 ar2
+ 2 ~r < i~ > cos 2a} (8-b)
- [ -D -a 'V 2w + -a M (w)~ (8-c)

r ae ar t
where R is the radius curvature at a regular boundary point with a

sign convention that indicates a negative curvature when the center
of curvature is on the outward normal, and v is Poisson's ratio.
The integral equation (7) contains a singular point when r equals

zero. In order to exclude this point a circular region bounded by
curve f with a radius e: is deleted from the region n. Hence, the
integratfun is taken over two curves, r and r , and letting the limit
of e: approach zero. Capital letter P and sm~H letter p denote the
singular point inside Q and on r respectively, while capital letter Q
and sma111etter q denote any point on the curve r and r respectively,
as shown in Figure I. e:
275
FIGURE 1
Hence, the integral equation (7) for any point P inside of n

becomes
1 dw dE
5[{Vn(E) lwl - Mn(E)i dnl + dnl Mn(w)i- El Vn(w)i} dsl
r Pq q Pq q Pq q Pq q q
K
+ ~- L <Mt(w)i>EI- <Mt(E)i> wl 0 (10-a)
k=1 qk Pqk Pqk qk
where
~=lim J
e+O r
b{Vn(E)iwl-
Pq q
Mn(E)i:l
Pq q
+ ~~~
Pq
Mn(w)i- EIVn(w)i}dsl
q Pq q q
E
{10-b)
By placing the ongm of the polar coordinate at point P the
singular function E that satisfies Equation (5) is
( 11)
where J is the zero order Bessel Function of the first kind, Y is

the zero0 order Bessel Function of the second kind, K is the £ero
order modified Bessel Function of the second kind, i::{.91 and
d = _1_ (12)
2
4-A 1T
276
Making use of Equations (11 ), (8-a) and (8-c) and separating in real
and imaginary parts gives
Ei(Ar) =i c J (Ar)
0
03-a)
~(Ar) = c Y (Ar) + d
0
K0 (Ar) ( 13-b)
i
~ (Ar) = Ni = - i AC J 1 (Ar) cos8 ( 13-c)
f (Ar) = rf' = -A[c Y 1 (Ar) + d K 1 (Ar)] cos8 ( 13-d)
Mn(E)i =Mi = i { c ~ [(1+v) + (1-v) cos28] A2 J 0 (Ar)
J 1 ( Ar)
- c D A (1-v) -r - cos28 } (13-e)
(13-f)
Mt(E)R = TR = - 0 (i-v) [A 2 (cY0 (Ar) - d K 0 (Ar))
- 2; (cY 1(Ar) + d K 1 (Ar))] sin28 ( 13-h)
· { [ 3 A3 (1-V)
.
Vn(E) 1 = V 1. = - 1 c D J 1(Ar) A cos8 + 2
(13-i}
_ co~38)}
277
Vn(E)R =.,.. =- D A3 (cY1(Ar) - d K1 (Ar)) cosa
+ D 0-v) [A: (cY0 (Ar) - d K0 (Ar)) - :~ (cY 1(Ar)
+ d K1 (Ar))] cos3S- D 0-v) [A~ (cY

0 (Ar)- d K0 (Ar))
- ~~ (cY 1(Ar) + d K1(Ar))] cos2S- 0 (~-v) A3 [cY1 (Ar)
03- j)
On the circular curve r t.' the following observations can be made.

r PQ = t. ; S = 1T ; R = -t. (14)
Using these results, Eqs. (13) reduce to thP following:
E~ = i c J 0 (At.) 05-a)
fQ
Ef = c Y0 (At.) + d K0 (At.) {15-b)

fQ
= 05-c)
rJj = A [c Y1(Ad + d K1 (At.)] {15-d)

fQ
(15-e)
{15-f)
=0 {15-g)
278
( 15-h)
{15- i)
Also, the limit of Eqs. (15) as e: approaches to zero is
= ic {16-a)
lim(~ c-d) logAe: =0 {16-b)

e:+O n
2
' i ' .cAe:
1 lml-2-= 0
lJmNI {16-c)
e:-+o fQ e:+O
2 1
lim r-Jf 1 im - (- c-d) - =0 ( 16-d)
e:+O fQ e:+O n e:
1 im M~ {16-e)
e:+O fQ
lim Mj lim D A2
e:+O
(~ c+d) logAe:
n
(16-f)
e:+O fQ
1 im vj ( 16-g)
e:+O fQ
lim Vj 2 2
= 1 im - D A (- C+d) -
e:+O n
1
e:
{16-h)
e:+O fQ
The results of Eqs. (16) are substituted into Eq. (lO-b) and we have
279
~=lim J2n l j-idD (l+V) dw £del + 1' J2n 1

e:-+OO D1 2 dn ' ~~O 5
{- D >.. 2 (~ C+d) i w- D >.. 2 (~ c+d) logA.£ : } £ d 6
- >.. 2 (~ C+d) w (P) 2n = -w (P) (17)

n
Hence, a boundary integral equation valid for any point P inside of

the region fl is found. Writing the equation in both imaginary and
real parts yields
- E~ Vn(w)l} dsl
Pq q q
K
- L [ <Mt(w)l> 0 (18-a)
k=l qk
and
~ { 1 WI - Mf
Pq q Pq q
R
Mn(w)l- El Vn (w)l
Pq q
K
- L[ < Mt(w) I>
R
El - < rr
...R
> wl] -WI = 0 08-b)
k=l qk Pqk Pqk qk p
where the kernals are defined by Eqs. (13-a to 13-j).
A second independent integral equation is obtained by taking a

directional derivative of Eqs. (18) in a fixed direction ~ that makes
an angle \jJ with the x axis. A local ~-n coordinate system i,s introduced
at point P and the coordinates are rotated by an angle \jJ with respect
to the x axis. A new polar coordinate (r,(jl) with respect to ~-n is
formed. Using this new polar coordinate system, the appropriate
singular function that satisfies Eq. (5) is
Again, making use of Eqs. (8a to 8c) and writing the kernals in
imaginary and real parts gives
(20-a)
280
(20-b)
. J1(Ar)
Ny1 (Ar) = i c [A J o (Ar) cos4> cosS - -r - cos(4>+S)] (20-c)
~(Ar)
y
= A(c Y0 (Ar) - d K0 (Ar)) cos4> cosS
(c Y1(Ar) + d K1(Ar))
- r
· cos(4>+S) (20-d)
. A2
M~(Ar) = i c D {J 1(Ar) [~ cos4> (1+V+(1-v) cos2S)
J (Ar)
- 2 ;~-v) cos(4>-2S)] + (1-v) A~ cos(4>-2S)} (20-e)
rv\<Ar) = 0~ 2 (c Y1(Ar)- d K1(Ar)) cos4> [1+V+{l-v) cos2S]

+D{l-v) lf (cY (Ar)- dK (Ar))- 22 (cY1(Ar)
0 0
r
+ d K1(Ar)) ] cos(4>-2S) (20-f)
Ti(Ar) =- i c D{l-v) [J 1(Ar)

y
(~ 2 cos4> sin2S + 22 sin(4>-2S))
r
J (Ar)
- --0- - -
r
sin(4>-2S) ] (20-g)
~(Ar) =- D(~-v) A2 (c Y1(Ar)- d K1(A.r))cos4> sin2S

+ D(1-v) lf (c Y (Ar)0 - d K0 (Ar)) - 22 (c Y1(Ar)
r
+ d K 1 (Ar~] sin(4>- 2S) (20-h)
V~(Ar) = i c D {J0 (Ar) [A 3 (cos4> cosS + ( 12v) cos4> sin2S sinS)
+ {1-v) ( 3 ~ cos(4>-3S) - ;~ cos(4>-2S))1

r
2 (1 v)
+ J 1(Ar) lr-
A
(-cos(4>-S) + ~ (cos(4>-3S) + sin(4>-2S) sinS
+ cos3S cos$))- ~2 (1-v) cos$ cos2S- G( 1jv) cos($-3S)

r
4
+ r~ (1-v) cos($-2S)]} (20-i)
281
"("Ar) = D A. 3 (c Y0 ("Ar) + d K0 ("Ar >) [cosq, cosS

y
+ <1;v> cosq, sin28 sinS] +

2
o;
(c Y1 ("Ar)
- d K 1 (A.r)) [-cos(<J>-8) + (lzv) (cos(<j>-38) + sin(<j>-28) sinS
~ ( c Y 1 ("Ar) - d K1 ("Ar) ) [(1-v) cosq, cos28]

+ cos38 cosq,)] - DA2
+ 3D (1-v) ["A 2 (c Y0 ("Ar) - d K0 ("Ar)) - 2 3 (c Y1 ("Ar)

r r
+ d K1 (A.r))1 cos(<j>-38) - 2 D{l-v) [;R (c Y0 ("Ar) - d K0 (A.r))
- r~ (c Y 1 ("Ar) + d K 1 (A.r))J cos(<J>-28) ( 20- j )
Again, making use of the observations of Eq. (14) along r e:' we

have the following
(21-a)
R
EYI = [c Y 1 (A.e:) + d K 1(A.e:)] cosq, (21-b)
~
i Jl(A.e:)
NY I = - i c ["A J 0 (A.e:) - -e:-1 cosq, (21-c)
NY~ = -A.(c Y0 ("Ae:) - d K0 (A.e:)) cosq,

~ c Y 1 ( "Ae: ) + d K l ( "Ae: )
+ [ e: ] cosq, (21-d)
Mi
Yl
~
282
MY~ = D >.. 2 (c Y1 (A.E:)- d K1 (>..d) coscj> + D 0-v) [~ (c Y0 (A.e:)
~- d K0 (A.e:)) - 22 (c Y1 (A.e:) + d K1 (>..d)l coscj> (21-0

e:
2 J ( A.e:)
T i i c D (1-v) {J 1 (A.e:) e: 2 -A.~} sincj> (21-g)
rl
~
TR = DO-v) {~ 2
(c Y0 (A.E:) - d K0 (>..e:)) -2 (c Yl (A.e:)
rl e:
~
+ d K 1 (A. E)) }s i ncj> (21-h)
J 1 (A.e:)
vi A.
= i c D {-J 0 ( A.e:) [A.3 + 0-v) --:zl coscp + - e : - [>..2
rl e:
~
+ 2(1-v)] coscj>} (21-i)
2
e:
- d K1 (A.e:)] coscj>- D (1-v) [). 2 (c Y0 (A.e:)- d K0 (A.e:))

e:
- 23 ( c Yl ( A.e:) + d K l ( A.e:)) ] cos cp (21-j)

e:
Taking the limit of e: to zero, Equations (21) reduce to:
1 im E
e:+O y~
l
= 1 im i c
e:+O
~e: coscp = 0 (22-a)
1 im E Rl = 1 im -(~ c-d) ie: coscp = 0 (22-b)

e:+O y~ e:+O
limN
e:+O y~
l
= lim- i c [A.-
e:+O
~~] coscp =- I~A coscp (22-c)
limN Rl =lim -A.(~ c+d) logA.e: coscj>- (~ c-d) --1-- coscp

e:+O y~ e:+O 11 11 A.e: 2
= 1 im -A. (~ c+d) logA.e: coscj> (22-d)

11
283
(22-e)
limMRI =lim[- DA(~c+d) + D(l-V)A(~c+d) logA€] cos$ (22-f)

€+0 Y €+0 € n n €
~
limTil =0 (22-g)
€+0 y
~
lim T Rl = lim {D(l-v) ~ (~+d) logAE} sin$ (22-h)

€+0 Y E+O 8 n
~
. A3
limV 1I
€+0 y
=-z- (22-i)
~
limVRI
€+0 y
~
- lim (A(~c+d) logA€ + (~c-d) _1_] cos$ D (l-v)(22-j)

E+O n 82 n 84
Again, the results of Eqs. (22-a to 22-j) can be substituted in Eq. (10-6)
in order to determine q;. However, we see that the limit does not
10-b) by~ = w - w I r
converge. This can be corrected by replacing w of Eqs. (lO-a and
Note, this assumption does not affect its
derivatives, also note hat along r 8 we have
(23-a)
Neglecting the quadratic term and noting that
(23-b)
Equation (23-a) can, therefore, be rewritten as

284
dw
wl = wl - wl =- £ dnl (24-a)
fQ Q p \.,(
dwl .
Also, -d
np IS
~~ - (~~ cos~ + ~~ sin~) (24-b)
When the results of Eqs. (24-a), (24-b) and (22-a) to (22-j) are
substituted in Eq. (10-b), ¢i is found to be
Hence, the second boundary integral equation at point P is
1 . . dw . .
J5 {\t (wl- wl)- Myt dnl + Nyt Mn(w)l- Eyt Vn(w)l} dsl
r Pq q q Pq q Pq q Pq q q
K
- I [ < Mt (w) I> EYl - < \ l > (wl - wl )] =0 (26-a)
k=l
pk Pqk Pqk qk p
and
Jr D.! {V R
YI w I
(
- wI ) - MR
YI
dw N R M ( )
dn I + YI n w I - EYRI Vn (w )I}d s I
Pq q P Pq q Pq q Pq q q
K R .
I [ < Mt (w) I
-k=1 >E I
y
- < T
y
I >
qk Pqk Pqk
1 aw (26-b)
- 2>.. a~l =o
p
where the kernals are defined by Eqs. (20-a to 20-j).

285
The integral equations (18) and (26) are for a singular point P
inside of the region n. The equations involve only the boundary
unknowns of effective shear, bending moment, normal slope and
displacement.
If the singular point p is on the boundary r instead of in the

interior then the curve r t would be around a portion of the arc,
hence the integrations in Eqs. (17) and (25) would be over a semicircle
for a regular boundary point, and over the interior angle for a corner
point. The boundary integral equations valid for any point p on the
boundary curve r thus become,
~~ = f,
p
PVfpq wlq - Mjpq :1q + ~ ~(w) I - El
pq q pq
Vn(w) l}dsl
q q
(27-R)
1 K R
- 5 kL [ <Mt (w) I > EI - <rl
...R
> wl 1
qk pqk pqk qk
Kt" ~~ + K
.. a.,p n ~npI
a
= _Dl
rf{~l
y
(wl-wl )- r.Jyl ddw +
nl
r-f-yl M (w) I
n
pq q p pq q pq q
-Ifyl V (w)l dsl

n
(28-R)
pq q q
1 K R ...R
--D k=l
L [ <Mt (w)l >E
yl
- q-Y-1 > (wl - wl)
qk pqk pqk qk pk
0 f{vj wl- Mj : + Nj Mn(w)l- Ej Vn(w)l} dsl

pq q pq ~ pq q pq q q
K . i
- kL [ <Mt (w) I > El - <T I > (wl )] (27-I)
qk pqk pqk qk
and
286
0 J{vi (wl - WI) - Mi dw +

yl dn I
N~l ~(w)l Ei Vnn(w) I }ds I
r Yl p
yl
pq q pq q pq q pq q q
(28-1}
K
I [ <Mt (w) I > Ei - <Ti > (w- w )
k=l yl yl I I
qk pqk pqk qk p
where a is a coefficient which relates to the interior angle of a

corner as:
a = interior angle
n
if p is on a regular point (not a corner point) then a equals unity.

The coefficient K~ and Kn of Eq. (2-R) are defined as
K~ = ~A + 2 ~A [sin 2o + sin 2 (an - o)] and
Kn = 4 ~A [cos 2o - cos 2 (an -o )]
where the angle o is measured from the boundary to the ~ axis in

a counterclockwise sense.
At a regular point (not a corner point) of the boundary r we

have 4 independent variables (w, ~~, Mn(w) and Vn(w)); however, on
the corner point, we have a total of 8 independent variables since
two distinct limiting values of ddw , M (w) and V (w) exist. The 8
d d n n n
variables are w, d~l , d~ I, Mnl, Mnl , Vn I , Vnl, and <Mt(w)>.
+ - + - + -
At a corner point, Eq. (28-R) can be rewritten in two independent
representations.
1) let o1 = (2+a>; (i.e. ~ bisects the exterior angle)
and Eq. (28-R) becomes

287
and Eq. (28-R) is

k
a - v/n sin an {dw
-:1p- } = 6Hv! ... }ds 1 I [ ... ]
4 COSA(an) dnl
D k=l
2 p+ (28-R-2)
NUMERICAL SOLUTION
For the numerical solution, it was decided to assume a linear variation

of the unknown boundary variables along the arc length between
nodes. Node points are chosen at each corner point, as well as at
selected regular points along the boundary. It was further decided
to consider only the real parts of Eqs. (27) and (28).
At a regular point, there are two known and two unknown

boundary conditions and hence, two unknowns correspond to each
regular point. At a corner point, five of the eight boundary point
variables can be determined from considerations of the boundary
conditionson each side of the corner, leaving three unknowns at each
corner point.
The integral equations are applied at each node point on the

boundary. At a regular point, Eqs. (27-R) and (28-R) are applied;
whereas, at a corner point, Eqs. (27-R), (28-R-0 and (28-R-2) are
applied. This leads to a homogeneous system of 2NR + 3NC
simultaneous equations with 2NR + 3NC unknowns, where NR is the
number of regular nodes and NC is the number of corner nodes.
For a given plate geometry and physical properties, the

coefficients of the simultaneous equations are a function of the
natural frequency alone. The solution process, then, consists of
searching for values of the frequency which makes the determinant
of the coefficients approach zero. On finding such a frequency, the
relative values of the boundary point unknowns are evaluated. This
eigenvector contains the information necessary to deduce the complete
mode shape from Eqs. (18) and (26).
COMPUTER PROGRAM
In the present computer programs for the solution of this problem,

only the real parts of Eqs. (27) and (28) are searched to find the
roots. This procedure yields roughly the same number of false roots
as true roots. If both the real and imaginary parts of Eqs. (27) and
(28) are searched, then the true roots would be those in which both
real and imaginary roots coincide. Searching both sets of equations,
288
however, requires twice as much computer time. The true and false
roots found from the real parts of Eqs. (27) and (28) are easily
separated, if one looks at corresponding mode shapes.
The computer program is set up with automatic node generation

along either straight line segments or circular arcs. At corner points
the interior angle must also be input. The integration in the program
is carried out by Gaussian quadrature, and the option is provided for
using up to seven quadrature points.
At present, the program can only treat homogeneous boundary

conditions. Indexing of the unknown function is performed by assigning
an index number at each regular point for each of the two unknown
functions and a zero for the two known (zero) functions. At corner
points, two indices are assigned for each of the two unknown functions
and an index for the twisting moment discontinuity is also required.
Zeros are input for the known functions.
As of this writing, the computer program is not fully checked

out, so we have omitted numerical results. Preliminary results for
certain cases, however, indicate that the method is convergent and
provides excellent results with a minimum number of boundary
elements.
CONCLUSIONS
The method used in this paper provides a viable and useful means of
determining frequencies and mode shapes of arbitrarily shaped thin
uniform elastic plates. At this time, nothing can be said about the
relative merits of this method as opposed to the finite element method
or the mixed method proposed by Bi!'zine (4). Since all of these
methods produce acceptable results, the criteria for choice must be
computer costs. Comparisons of these methods will be the subject
of future research.
REFERENCES
1. Hutchinson, J. R. and G.K.K. Wong, "The Boundary Element

Method For Plate Vibrations", Proceedings of the ASCE Seventh
Conference on Electronic Computation, St. Louis, Missouri,
pp. 297-311, August 1979.
2. Be'zine, G. P., and D.A. Gamby, "A New Integral Equation

Formulation For Plate Bending Problems", Recent Advances in
Boundary Element Methods, edited by C. A. Brebbia,
Pentech Press, London:Plymouth, pp. 327-342, July 1978.
3. Stern, M., "A General Boundary Integral Formulation for the

Numerical Solution of Plate Bending Problems", International
Journal of Solids and Structures, Vol. 15, pp. 769-782, 1979.
289
4. Bezine, G., "A Mixed Boundary Integral - Finite Element Approach

to Plate Vibration Problems", Mechanics Research Communi-
cations, Vol. 7, pp. 141-150, 1980.
5. Vivoli, J. and P. Filippi, "Eigenfrequencies of Thin Plates and

Layer Potentials", Journal of the Acoustical Society of America,
Vol. 55, No. 3, pp. 562-567, March 1974.
Section III
Geomechanics
293
BOUNDARY ELEMENTS APPLIED TO SOIL-FOUNDATION INTERACTION
H. Ottenstreuer and G. Schmid
Ruhr-Universitat Bochum, W-Germany
INTRODUCTION
The response of a structure to a given loading does not only

depend on its geometric configuration but may also be influenced
by the surrounding media. This holds particularly for dynami-
cally loaded structures whose behaviour is strongly influenced
by soil-structure interaction. In such cases the foundation
environment can not be neglected.
An essential feature exhibited by soil is the "geometric

damping", which is due to the fact that waves travel out to
infinity. This results in energy dissipation even if the soil
were assumed to consist of purely elastic material.
In early attempts to simulate the embedding of a struc-

ture in soil simple spring and dashpot elements, connected to
the mathematical model of the structure were used.
A more refined method for the simulation of soil-struc-

ture interaction takes advantage of the possibilities offered
by finite elements which allow the simultaneous discretization
of the structure and the soil. The method is equally applica-
ble to embedded foundations and inhomogeneous soil. However,
it has serious disadvantages when applied to three dimensional
problems since it requires extensive, complicated and expensive
data management. Horeover geometric damping effects can only
be represented approximately by sem~lnfinite elements (Waas
1972) which are limited in that the infinite extension can
only be simulated in the horizontal direction.
Another possibility to calculate soil-structure inter-

action is the substructure-method which analyses the soil and
the structure separately. Here the soil is assumed to be an
elastic homogeneous halfspace and the influence of the foun-
dation upon the soil is considered by using Green's functions.
294
This method, however, is not applicable to layered soils and

to embedded foundations.
As an alternative we present the boundary element method

which - together with the substructure technique - has several
advantages in the treatment of soil-structure interaction prob-
lems. First it enables one to consider embedded foundations
and foundations on a viscoelastic halfspace, and second, the
coupling effects of several neighbouring foundations may be
taken into account. Single foundations with, and without, em-
bedment have been considered earlier by Dominguez (1978) for
an elastic halfspace.
DYNAMIC STIFFNESS
The dynamic stiffness matrix of the foundations represents the

complex, frequency dependent impedance function which charac-
terizes the soil-structure interaction. Its derivation of the
matrix is shown in the following.
The differential equatiomof motion are ?iven by

2 d U·
(c -c )u . . . +c u . . . +K.=-- 1. (1)
P s J,Jl. s l.,JJ 1. at 2
where c is the velocity of dilatational waves,
p
c is the velocity of distortional waves,
s
u. is the displacement component in the i-direction,
1.
K. is the body force component in the i-direction,
1.
t is time.
The conditions along the boundary are
u. = u. (2 a)
1. 1.
oij nj = ti , (2 b)
where o .. is the stress tensor,

1.]
n. is the component of the normal vector

J
and the initial conditions are taken as
u. (t=o) = u. (3 a)
1. 1.0
dU.
1
-- (t=o) = u. (3 b)
d t 1.0
Equations (1) through (3) may be transformed into the frequency
domain by use of the Laplace transform yielding the differen-
tial equation
2
(c - c ) u ... + c u . . . + K.
p s ],]1. s l.,JJ 1.
s 2 u. - su.
1. 1.0
- u.1.0 (4)
295
-
Writing s = iw and defining the body force to be
P.
l. 1.0
K. +
l.
u.1.0 (5) + iw u.
the frequency dependent differential equation of motion takes
the form
2 ,..,_. 2 ,..,_. 2
(c ) u ... + c - c
u. + en u. + P. = 0. (6)
s J,Jl.
p s l.,JJ l. l.
where w is the circular frequency.
The transformed boundary conditions are
- u . = u.
l. l.
(7 a)
0.l.J.
J
n. = T'il:l. = t.l. (7 b)
where T is a differential operator.
The inverse transformation from the frequency into the

time domain allows the calculation of the response of the
system to any specified loading history.
The fundamental solution for Equation (6) as given by

Cruse and Rizzo (1968) is an essential prerequisite for the
application of the boundary element method. In this context it
corresponds to the response of the system to an impulse in the
form of a Dirac-delta function.
Writing Equation (6) as

- c ) u~ .. + c
2 2 2
(c u*. + w u. + 6. 0 (8 a)
p s J,Jl. s l.,jj l. l.
or in the abbreviated form
Lu + 6 = 0 (8 b)
where L is a differential operator and 6 is a point load,
the fundamental solution of Equation (8) for the three-dimen-
sional situation is
1
u* ..
l.J 2 (lj! 6 ..
l.J -x r,i r' j) (9)
411 pc
2 s -iwr
3c 3c c 3c
s s
(- - 2 - 2 + -.- + 1)
e cs s (- ____£_
X r
---2
2 2
w r l.tllr c wr
p
-iwr
3c ( 10 a)
e Cp
+ ..,..-£. + 1)
1.wr
--
r
- '
-iwr
c c c c
s e cs s ...L.
1jJ (- -2-2 + ~+ 1) ---2 (-
JWr r 2 2
wr c
p
wr
-iwr
( 10 b)
c
e cp
+.-E-)
J.wr r
In Equations (9) and (10) r is the distance between the point
load and a generic field point, and p is the mass density.
296
The application of the method of weighted residuals with

u* as weightLng function, yields the Betti-formula
j(u*Lu- uLu*) d~ ~u*Tu -uTu*) ds (11)
I.: r
where B is the domain,
r is the boundary,
d~ is a volume element and
ds is an area element.
With the help of the fundamental solution the deformation of a
point inside a domain can be expressed in terms of the boundary
values (Brebbia.. 1978):
u = ~(u*Tu - uTu*) ds. ( 12)

r
For a point on a smooth boundary the relation
= ~u*Tu
1
-zu - uTu*) ds ( 13)
is found.
To obtain a numerical solution the boundary is divided

into elements. In our case the tractions and the displacements
are assumed to be constant within the elements. Thus,the inte-
gration can be replaced by a summation over all elements on the
boundary, and Equation (13) becomes
i=1
n
l: J
r.
u*dsi ¥
i=1
ft*
I'.
i
ds.L u . ( 14)
L L
where ti and ui are constant for each element.

Equation (14)can also be written in matrix notation as follows
1
2 - = u t - Tu
-u (15)
where u ri_~* ds ( 16 a)
T (16 b)
rJt*
- ds
or in a more condensed form
T u !::! .t I
( 17 a)
T T + _!_ I (17 b)
2-
Equation ( 17) allows us to establish the dynamic stiff-
ness matrix of one or several foundations. The surface of the
soil represents the boundary of the domain which is discretized
by boundary elements. They may be divided into two parts: One
part, with index i, stems from the foundation, the remaining
297
part, with index a, from the traction-free surface. Then Equa-
[i]
tion (17) may be partitioned as follows:
[
~i~ ~ia]
Ta~ Taa ua
=[ !:!i~
Ua~
!:!ia]
Uaa
[!i]
ta
( 18)
- - - - - -
As all boundary elements belong to the same horizontal plane
some of the coefficients become zero because at the surface
~~ = 0 (19 a)
(19 b)
(19 c)
Correspondingly some of the coefficients of the element-related-

submatrices in ! vanish:
(20)
Force-displacement relation of the foundation

The six degrees of freedom of a rigid foundation consist of
three translations ~ and three rotations p. The kinematic
relations between the rigid body degrees of freedom and the
boundary element displacements are given by the following geo-
metric constraints
u1i u1 - 1jJ3 r2i, (21 a)

u2i u 2 - 1jJ3 r1i, (21 b)
u3i u3 + 1jl1 r2i - 1jl2 r1i (21 c)
which, written in matrix notation, read
-u = -a -u •
(22)
The corresponding forces, (Fig.1) acting in the direction of~

and ~. can be obtained by integrating the tractions over the
area of the boundary elements:
dA. M. (23a,d)
~ ~
298
Figure 1 Forces on a single foundation
K2 L:
].
f
A.
t2 dA.l ' M2 L:
i
f
A.
t3 r1 dA.l (23b,e)
l l
K3 L:
i
f
A.
t3 dA.l ' M3 L:
i fA.
t2 r1 dA.l
l l (23c,f)
- L:
i
f
A.
t1 r 2dAi.
l
In matrix notation Equation (23) reads:
p aT A t (24)
where A is a diagonal matrix and stands for the area of the in-
dividual boundary elements of the foundation.
Depending on the particular assumptions concerning the

motion, two stiffness relations may be derived.
Stiffness matrix with uncoupled boundary conditions (relaxed

boundary conditions) If it is assumed that at the surface
vertical loads result only in vertical displacements and hori-
zontal loads produce only horizontal translations, then the
coefficients in Equation (20) can be set equal to zero. Since
the tractions outside the foundation are zero as well Equation
(18) reduces to
(25 a)
resulting in the decoupled system of equations
1 i
2 .!:! (25 b)
299
for the foundation. The surface deformation outside the founda-

tion can be determined by
I a
-u (25 c)
2 -
Equation (25 b) shows that, within the above mentioned assump-
tions, the displacement-traction-relation of the foundation is
not affected by the traction-free part of the surface. It is
therefore sufficient to introduce boundary elements only at the
location of the foundation itself. This simplification results
in considerably reduced numerical effort.
The stiffness matrix can now be obtained by premulti-

plying Equation (25 b) with (~ii)-l and using Equations (22)
and (24). One obtains
~~=~ (26)
where the complex dynamic stiffness is defined as
(27)
Stiffness matrix with coupled boundary conditions (non-relaxed

boundary conditions) Here, tractions and displacements on the
surface are not assumed to be decoupled. Thus, the traction-
free surface outside the foundation has to be represented by
boundary elements as well. It should be mentioned that for
practical numerical calculation the boundary element discreti-
zation need not extend further than a few times the length of
the foundation.
The stiffness is obtained by premultiplying Equation (18)

by A u-l to give
(28)
where the definition

g = ~ ~-) ! (29)
has been introduced.
From Equation (28) ua can be eliminated so that, with the

assumption of rigid foundations, the relation between the foun-
dation forces and displacements is obtained as
T (gii _ gia (gaa)-1
~ 9ai) a u = aT A t (30)
The dynamic stiffness matrix is now given by

T (gii _ gia (gaa)-1 gai)
K !! !! (31)
300
Force-displacement relation of two foundations (Uncoupled

boundary conditions)
To calculate the influence of a second foundation on the re-
sponse of the first one, the stiffness matrix of two rigid
foundations (Fig. 2) is derived.
Equation (17) is partitioned into three parts, where

parts I and II correspond to foundati~ns I and II and part "a"
to the traction-free surface outside the foundations:
~I II
T
Til II
Ta II
(32)
The assumption of uncoupled tractions and displacements yields

three matrix equations
~I I
T u
I UI I ti + UI II til (33 a)
~II
II II UII I ti UII II til
T u + (33 b)
;:a a a Ua I tl Ua II til
!! +
- - (33 c)
Again, the traction-free part of the surface has no influence
Figure 2 Forces on two foundations

301
on the stiffness of the foundations and need not be discretized

by boundary elements.
The dynamic stiffness matrix for two rigid foundations

is again given in the form
l aT A U-l a = K (34)
2 - - - -
a now being the kinematic transformation matrix for two foun-
dations.
NUMERICAL EXAMPLES
For numerical calculations the stiffness matrices are divided

into real and imaginary parts which, mechanically speaking
represent the spring and damping coefficients of the system
~ = ~R + i ~I (35)
The stiffness matrix for the static case K0 may be factored
out, i.e.
K (36)
with a = w ·b (37)
0 c
s
where b is the half side-length of the foundation,and KR and KI
are dimensionless stiffness coefficients. Inversion of-~ yields
the flexibility matrix
F = FR + i FI
- - - (38)
again divided into real and imaginary part.
With shear modulus G and side-length b they can be written in
dimensionless form:
for the displacements
fR G b FR (39 a)
G (39 b)
and for the rotations and the torsional degrees of freedom
fR G b 3 FR (39 c)
G (39 d)
Dynamic stiffness matrix for a single foundation

As an example for the numerical calculation of the stiffness
matrix a quadratic foundation on a halfspace is chosen. In
addition to geometric damping,material damping behaviour is
taken into account by introducing a complex modulus of elas-
ticity (Achenbach 1975)
E =E (l+in). (40)
302
.. .
i!
0 8 8
lil
.;
Iii
.;
..
Iii
~
"'
..8 .
8
lil
.:~
~
i
0
~
0
oo·t .., .. '>9"8 96"J

"~">I
.... 8
0 OZ'7 BG'C '8'8 S'J '3
Z~\t
B7''J
i! i!
0
.8 0
lil
.;
Iii
.;
.Iii
.~
.
8
Tlil
.: d
~
Iii
0
~
0
8
oo·• BL ·o !IG"O 't'O Bl"O 0 o• ·o so ·o 110 ·o 110 ·o- so ·o-
"~">I .ou Z~">l
Figure 3 Real and imaginary part of the stiffness

coefficients k and k ; constant hyster-
etic damping x z
303
.8
!i!
"'
or·c ,s·z 86"l 21£"'J. ss·o

XWX!f~
.8
oo ·~ et ·o 9S. ·o 'r ·o m ·o oo'-::.,c+c.ac::>:-:·o~os:?cc:.o-+-:z~s·""o-+-:-..,:+:.:-0.........!- ~

XWX~\j ZWZ~\j

coefficients krox and kmz; constant hys-
teretic damping
304
8
~
g
~
g
~
g
N
~
N
8
N
g
~
li!
g
0
~
0
8
oe·o st ·o 31 '0 89'0 '19'0 0 OT ·~ 10'l 9b'O Z&'O 98'0
X~\j ZT'>t
8
~
g
..;
g
~
g
N
~
N
8
N
g
~
f\1
g g
0 0
~
0
8
oo·• oo·o oe·o oro os·o 0 oo~-.~~,t~-~.~~~-o~~="-·o~,~•.~o-.+-4 ~
X~\j Z~\j

coefficients kx and kz; Voigt damping model
305
o,. ·o oc ·o tz ·o 9T ·o eo ·o o,.·o zr·o ~·o s: ·o eo·o

xwx~~ ZWZ~\1
.
8
oo.7..1~~~·o~=••~.o~=~~.o~=~~·o~-4~
xwx~~

coefficients kmx and kmz; Voigt damping
model
306
In Equation (40)n is the damping coefficient which depends on

the chosen damping model.
E.g. for a "Voigt" model one has
.!l
w
= e = const. (41)
whereas for constant hysteretic damping n becomes
n= w ·e con st. (42)
Figures 3 to 6 show the dependency of stiffness coefficients
on the dimensionless frequency a 0 for various values of the
damping coefficients n and 8. It can be shown that the real
parts of the dynamic stiffness matrix decrease with growing a
which corresponds to a reduction of the stiffness of the sprigg
in the model. Simultaneously damping increases in a way which
depends on the chosen damping model.
It is important to point out the difference between the

two models: The model with constant hysteretic damping shows
a significant increase at small values of a and approaches the
damping of the elastic-halfspace model at h~gh frequencies. The
"Voigt" model however shows a completely different behaviour
in that material damping causes a nearly constant increase of
damping over the whole frequency range and that it is almost
independent of a • It could be shown (Ottenstreuer 1981) that
these results arg in good agreement with those found inthe liter-
ature (Wong and Luco, 1976).
Dynamic stiffness matrix for two rigid foundations

As a second example the interaction between two rigid founda-
tions (Fig. 7) is examined. Contrary to a single vibrating
foundation even some offdiagonal coefficients of the stiffness
matrix are present (Fig. 8). This demonstrates the coupling of
the two foundations. Figures 9 to 10 show some of the stiffness
!::1 }(
x2
45°
>xi 2
v
:.____j,_:t-0
xf
xl
Hgure 7 Two foundations Figure 8 Non-zero stiffness

with mesh coefficients
307
g g
~ ~
e e
f!i f!i
~ ~
l
g g
~ ti
e e
.
~
0 .
~
0
l
"

coefficients kzz 11 and kmx my 11
308
8 8
!l !l
~
~
~ ~
~
~
~ ~
!a 5I
g g
iii
'2
~
~ ·~f!
"'
~
~ ~
"'
.; 0
)( t: ~
al ~
ti ti
g
.;
.g
8
0
8 8
!l !l
~ ~
~ ~
~ ~
;i ;i
!a !a
g g
g iii
~ !i
~ f!
"' "g
!i!
.; 0 .; 0
)( )(
~ al
ti ti

coefficients kzz 12 and kymz 12 (cross
coupling)
309
coefficients dependent on their relative distance. The influ-

ence of the offdiagonal terms becomes smaller with growing
distance of the two foundations so that finally the influence
of the coupling disappears. Thus, at large relative distance
the two foundations behave in the same way as two single foun-
dations.
CONCLUSION
From the results presented in this paper it can be concluded

that the boundary element method-combined with the substructure
technique often used in soil dynamics - is a suitable tool to
simulate soil-structure interaction.
Although our results are given here only for rigid, flat
foundations on homogeneous soil the method can easily be ex-
tended to embedded foundations on layered soil.
ACKNOWLEDGEMENT
The programming work was done by Margret Gibhardt and typing

by Christa Hoogterp.
REFERENCES
Achenbach, J.D. "Wave Propagation in Elastic Solids"

Amsterdam: North Publishing Company, 1975
Brebbia, C.A. "The Boundary Element Method for Engineers"

London: Pentech Press, 1978
Cruse, T.A. and Rizzo, F.J. "A Direct Formulation and Numeri-
cal Solution of the General Transient Elastodynamic Problem I"
Journal of Mathematical Analysis and Applications 22, 1968
Dominguez, J. "Dynamic Stiffness of Rectangular Foundations"

Publication Nr. R 78-20, MIT, Department of Civil Eng., 1978
Ottenstreuer, M. "Das Verfahren der Randelemente - Ein Beitrag

zur Darstellung der Wechselwirkung zwischen Bauwerk und Bau-
grtind"
Dissertation, Ruhr-Universitat Bochum, 1981
Waas, G. "Linear Two Dimensional Analysis in Soil Dynamics

Problems in Semi-Infinite Layered Hedia"
Dissertation, University of California, Berkeley, Cal., 1971
Wong, H.L. and Luco, J.E. "Dynamic Response of Rigid Footings

of Arbitrary Shape"
Earthquake Engineering and Structural Dynamics, 4, 1976.
310
THE IMPLEMENTATION OF BOUNDARY ELEMENT CODES IN GEOTECHNICAL

ENGINEERING
L. A. Wood
Lecturer, Department of Civil Engineering, Queen Mary College,
London, U.K.
SYNOPSIS
The application of simple boundary element techniques in the

field of geotechnical engineering is discussed. Particular em-
phasis is placed upon the coupling of standard finite elements
to soil boundary elements in order to solve soil-structure in-
teraction problems. Two computer programs which incorporate
these techniques in the areas of foundation and pile/wall ana-
lysis and design are briefly described. In addition the use of
these programs is demonstrated with respect to several case
histories in which the agreement between predicted and observed
behaviour is most encouraging.
INTRODUCTION
In the field of geotechnical engineering the semi-infinite na-

ture of the soil domain may be efficiently modelled using boun-
dary element techniques. Finite element programs which will
handle the modelling of three~dimensional continua exist, but
for semi-infinite domains the computation involved may become
excessive. Furthermore, in the context of foundation design
the tnree dimensional nature of the distribution of structural
loads is such that the reduction of the problem to a two-dimen-
sional axi-symmetric, plane stress or plane strain model, in
order to facilitate economic analysis, is inappropriate. In
order to exploit the potential offered by the boundary element
approach two computer programs RAFTS and LAWPILE (Wood, 1978a
and 1979a) have been written specifically for the analysis of
foundations and, piles and walls subject to lateral loads.
Both are based upon the coupling of standard structural finite
elements with soil boundary elements.
The soil boundary elements are formulated on the basis of

linear elastic behaviour but incorporate procedures to enable
the effects of non-linear soil response to be assessed. Appro-
311
ximate extensions of the basic element formulation have been

implemented in order to model layered (not neccessarily hori-
zontal layers) and transversely isotropic continua. It is
felt that these approximations based upon the solutions ob-
tained by Mindlin (1936) and Gerrard and Harrison (1970) for
homogeneous elastic continua, are acceptable in the context
of soil where the assumption of elastic behaviour is itself
questionable. Indeed the performance of the simple boundary
element adopted has been shown to produce results which lie
well within the normal tolerances associated with geotechnical
design procedures, and suggest that the soil model is compati-
ble with the level of soils information usually available.
The application of the simple boundary element is discussed

below with respect to two general soil-structure interaction
situations. Consideration is first given to predicting the
behaviour of raft foundations using the program RAFTS and then
to the use of a sister program LAWPILE for determining the
behaviour of piles or walls subject to lateral forces. For
both comparisons are shown between the actual recorded perfor-
mance of structures and that predicted by back-analysis. The
results are most encouraging and confirm the predictive power
of this, albeit approximate, approach.
RAFTS
The development of the basic soil boundary element has taken

place over a number of years and the acceptability of the app-
roximate extension for layered elastic continua has been de-
monstrated by Wood(l977) with respect to a wide variety of
inhomogeneous situations and elastic layers of finite extent.
Subsequently the element has been enhanced (Hooper and Wood,
1977; Wood, 1978b) to include transversely isotropic behaviour.
That is, a continuum within which the stiffness in the vertical
direction differs from that in the horizontal plane. Thus, the
elastic properties of the continuum are defined by five inde-
pendent constants:
Ev, \1 vh• \1 hh• and
where E and v are Young's modulus and Poisson's ratio res-
pectively, G is the shear modulus and suffixes v and h refer to
the vertical and horizontal planes.
The use of a purely elastic analysis with regard to the

design of raft foundations may give rise in some instances to
somewhat unrealistic soil reactions under the edges of the
foundation. This is particularly so in the case of stiff
structures. Although, the adoption of a non-linear soil model
(Wood and Larnach, 1975) produces a more realistic assessment
of the raft behaviour, the increase in computational effort
with respect to three-dimensional situations is rarely justi-
312
fied on economic grounds. Hence, the inclusion in RAFTS of

a procedure which allows the imposition of an upper limit on
the magnitude of the developed soil reactions in order to
assess the effect of local shear failure of the soil. In add-
ition similar routines are incorporated for the assessment of
the effect of new construction on neighbouring structurures
(Wood, 1980) and the effect of heave due to stress relief of
the underlying strata.
Three examples are given below in order to demonstrate

the flexibility of the use of RAFTS with respect to different
soil conditions. In the first example consideration is given
to the performance of a raft subjected to uplift forces due
to swelling of the underlying clay strata. The other examples
relate to two published case histories of rafts founded on
sand and chalk strata respectively and illustrate the predic-
tive power of the program.
Uplift
A 30m square, 2m thick raft resting on a clay soil 40m thick
vJith:
Ev = 10 + l.Oz MN/m 2 , Eh/Ev = 2.3, ~vh = 0.1, Vhh = -0.15
and Gvh/Ev = 0.661;
where z is the depth below the underside of the raft, has been
subjected to a uniform uplift pressure of 200kN/m 2 • The raft
material is of unit weight, 25kN/m~ and elastic moduli E =
28GN/m 2 , v = o. 2.
This example has been chosen to illustrate the ability of

RAFTS to take account of loss of contact between the raft and
the underlying soil. This point is clearly demonstrated in
Fig. 1 where the stiff nature of the raft has caused it to
lift away from the soil surface at the edge; and is shown in
more detail in Fig. 2 where the contours of raft and soil
movements are given.
Raft foundation on Sand

Recorded movements of the raft foundation for the reactors of
Dungeness 'B' Nuclear Power Station, Kent have been presented
by Dunn (1975). The 3.4m thick raft is founded at a depth of
9m below ground level on 30m of beach sand overlying mudstone.
The layout of the reactor building is shown in Fig. 3.
In this case the mudstone has been taken as incompressible

and the sand as exhibiting isotropic elastic properties;
Ev = 85MN/m 2 increasing linearly to 180MN/m 2 , with V = 0.1
The datum for measured settlements was established after the
construction of the raft itself and therefore correspondsto
the imposition of the load from the superstructure of 500MN
for the reactors and 115MN for the central fuel handling unit.
In the analysis the raft concrete has been assigned a Young's
modulus of 28GN/m 2 and Poisson's ratio of 0.15 and the raft
313
200
a~~--~--~~--~~
a 5 1a (m) 15
a~~~~~~--~--~--_,
'(5
Vl
E
'zE 2
->,
'iii
c

c
c
E
0
~
a
Figure 2 Contours of computed raft and soil movements (mm)

314
West
·~~
I
Layout of reactor building
0 10 (m) 20
Fig. 3 Layout of reactor building
has been taken as symmetrical about both axes. Computed and

measured settlements along the longitudinal axis of the raft
are shown in Fig. 4; together with computed soil pressures
and raft bending moments. It sh9uld be noted that the measur-
ed settlements also exhibited symmetry about the raft axes and
that the agreement shown in Fig. 4 is very encouraging.
Silos on Chalk
Settlement observations for a complex of four grain silos
supported on independent 23m diameter, 1.2m thick raft founda-
tions founded on soft chalk have been reported by Burland and
Davidson (1976). The layout of the complex, in which each
silo weighs approx~mately 4lMN and has a storage capacity of
120MN, is shown in Fig. 5.
The results obtained from plate bearing tests on the chalk

are shown in Fig. 6 together with the assumed 11 elasto-plastic 11
curve with a Young's modulus of 120MN/m 2 , Poisson's ratio of
0.25 and a limiting soil stress of 0.7MN/m 2 • The chalk stra-
tum has been taken as homogeneous and isotropic, and to be 60m
thick. Comparison between computed and recorded settlements
is somewhat complicated by the loading sequence adopted in
practice and a more detailed explanation of the analysis is
given in Wood (1979b). However, the results obtained for two
of the silos are shown in Fig. 7 where computed settlements
315
Measured • West} of r.
• East
E 't.
.s
c
••
50
Q)
E
~
~
Qj
(f)
100
or-----.------.-----,.-----~----~
~
z
~0.2 ------------~~~~--------
~
::J
Ill
Ill
Cll
L.
~0.4
'(5
(f)
-10
Hogging
E ' ',, Sagging

'E
z 10 \
~ \
\
~
>. \
'iii \
c \
\
c
Q)
,"' ---
20 \
\
c '
Cll
E
0
............... ' , ,--
~
30
'''v" " Transverse
Figure 4 Computed results along longitudinal axis,

Dungeness 'B' Nuclear Power Station
316
,. 68. 230m ~
~
B$6(:)$-
Silo No.1 Silo No.2 Silo No.3 Silo No.4
Figure 5 Plan of silo complex
Plate dia. (mm)

A 305
y 457
Cll .8 • 610
1-
:::J
IJl
IJl
Cll
f-
a. .4
"0
Cll
0..
a.
<t o~----~----~----~-
o .5 1.0 1.5
Displacement/plate dia. (x m-2)
Figure 6 Results of plate loading tests
for both the elastic and for the 11 elasto-plastic 11 analysis are
shown for comparison. Agreement is tolerably good given the
complex behaviour of chalk, which exhibits creep characteris~
tic·s, and is much improved when the "elasto-plastic' 1 nature of
the material is modelled by the imposition of an upper limit
of O.?MN/fi 2 on the soil pressures.
LAWPILE
Many methods of analysis are available to the designer in

order to quantify the effects of lateral loads on foundation
members. These range from simple classical earth pressure
theory through to elaborate non-linear finite element solu-
317
20 Silo No.2
E
.s •
+-
c
<11
E
...
..9:!
+-
~ 20 ..--S_ilo_N_o_.4----+----------,
60
Elasto-plastic
- - - Elastic
• Measured
Figure 7 Computed and Observed settlement of silos
tions, with the middle ground occupied by methods based upon

the theory of sub-grade reaction (also known as the beam on
elastic foundation approach). The former ignore the interac-
tion between the two basic components of the system, namely
the soil and the structure, and the latter involve expensive
data preparation and high solution costs, whilst the deter-
mination of equivalent spring stiffnesses with any degree of
precision, as required in the sub-grade reaction methods, is
a daunting task. LAWPILE has been developed in order to over-
come the problems associated these approaches whilst still
providing the designer with a knowledge of the overall stabi-
lity of the structure and the magnitude of displacements and
induced stress resultants.
The formulation of the simple boundary element is based

upon the equations developed by Mindlin (1936) with an appro-
ximate extension to inhomogeneous soil desposits similar to
that suggested by Poulos r1973). This approximation is analo-
gous to that used in RAFTS and has been found to produce re-
sults of acceptable accracy for design purposes.
318
Hence, an initial elastic analysis is undertaken yielding the

displacements, rotations, shears, induced moments and soil
reactions. For an earth retaining structure such as a dia-
phragm wall the initial state of stress in the soil is taken
as that corresponding to at rest, Ko conditions. Invariably
the elastic analysis gives rise to some unrealistic soil reac-
tions when consideration is given to the likely ultimate re-
sistance that the soil is capable of sustaining.
In order to take account of local failure of the soil the

computed elastic reactions are compared with limiting reac-
tions obtained from the shear strength parameters of the soil.
Where the elastic reactions exceed these, the soil reaction
is maintained at the limit and a new solution of the displace-
ments, etc., obtained. This procedure is repeated until all
of the soil reactions lie within the limiting reaction enve-
lope. In this manner the elasto-plastic behaviour of the soil
is approximated to without incurring the high costs associated
with more rigorous solutions.
For single piles or pile groups the limiting reactions are

computed on the basis of a wedge type soil failure near the
surface and flow around the pile at depth. In the case of a
wall it is assumed that the soil pressures may not exceed the
passive state or fall below the active condition.
Use of the program is illustrated below with respect to

two examples, covering both of the main applications of the
package.
Diaphragm Wall
In ordef to illustrate the use of the program in predicting
the behaviour of a diaphragm wall consideration has been given
to the case study reported by Burland and Hancock (1977) of
the basement wall to the underground car park constructed at
the Houses of Parliament, London. The wall has a nominal
thickness of l.Om and an over all height of approximately
28.0m, supporting retained soil to a height of about 17.0m
The car park was constructed from the top, down, after the in-
stallation of the wall, by casting the uppermost floor slab
-and then excavating the material below down to the next slab
level and so on. In this manner the floor slabs act as both
temporary and permanent strutts to support the diaphragm wall.
The soil profile and construction stages are shown in Fig. 8,
together with the assumed soil properties used in the analy-
sis.
The computed final wall displacements obtained from the

analysis, taking account of the construction sequence, are
given in Fig. 9, together with results obtained by other re-
searchers employing more refined finite element methods. In
319
Shear strength (kN/m2)

-2 0 100 200 300
~
u: 0
·~.
_o
0 ·-
• ,"0
2 C'=Im2
<?
-t~ -
.~:
0
. .; 0' = 30°
~ 0
·o t.
.o
oil d.
j_
: o. 'a
-o o· 6
c u'o
0
(f) .,
'.:
~--- 8
-_<s:- E
-.:r.::. -w 0
- _--;.;:: 0 0
:s:
'X'.:::: 0 12
_-_-?t 0.
0
..... 1t.
~-..::::
:s:0
_-_><: Qj
0 0
.0 16 0 0
0
.J::.
_'t.= 15.
-* g 18 0
0
0
>. 0 0
0
u x=
c --~ 20 0
0 0
-o 0
c -~; 0
0
0
_J
22
0
0
)';_-=
21.
---~
__ ,.,.
-:r_- 26
0
0
0 0
-;;r-
-_-_g 28
Figure 8 Soil Profile (diaphragm wall)

320
{mm)
20 30
4 2
6
Floor slobs
8
_ 4_
10
12
g _ 5_
0
~
14
0
a.
0
16
~ 18
.D
:Ea. 20
 - -- - - S1mpson et ol {1979)
0
22 - - - - - Burland & Hancock {1977)
- - - - Present
24
·:·:·:·:·:·:·:-:·:·:·:·:· Range of recorded
::::::::::::::::::::::::: movements
26
28
Figure 9 Computed and observed wall movements

addit i on the range of observed wall movements is s hown for
comparison. Fig. 10 shows the computed and observed maximum
movements determined a t each stage of construction as a per-
centage of the maximum final movement.
It is appar ent that t he agreement between the simple boun-

da ry element appr oach adopt ed herein and t he more rigorous
analyses is good and that t he computed movements are in tole-
rable agreement with the obser vations .
Recorded } Burland &
Predicted Hancock (1977)
• Simpson et al (1979)
Present
Figure 10 Computed and observed maximum wall, movements at

each stage expressed as ro of maximum movement at
final stage
Sea-bed Anchor
As part of the development programme of a novel embedded an-
chor for marine applications (see, Harvey and Burley, 1977 and
Nath et al. 1978) a series of lateral load tests have been
conducted on a prototype anchor. The anchor has an unusual
geometry enabling it to develop a large carrying capacity both
with regard to vertical and horizontal applied forces; and is
very efficient with respect to the load carried compared to
its self-weight. The results of one such load test with the
322
200
-- --
- -
.~
.......
Vl
0
ljJ /
-150
z /
~
/
/
Q)
l)
'- 7 - - - - - Observed
.E 100 I
0 I
+-'
c I
0 I
N
I
'-
0 50 I
I
I
I
I
OL-----~----~----~------~--~
0 10 20 30 40 50
Displacement at soil surface (mm)
Displacement (mm) Soil pressure (MN/m2J

175kN 0 20 40 60 2 0 2
.....
~
g-2.0
0
2.5
3.0
Figure 11 Sea-bed anchor

323
anchor embedded in a dense silty sand of unit weight 23kN/m3

with an angle of internal friction, ¢ = 39° and E = 50MN/m 2 at
the soil surface, increasing linearly to ¢ 4f
= and E=200MN/
m2 at a depth of 3.1m; with cohesion, c = 5kN/m 2 and v = 0.1;
are shown in Fig. 11 together with the computer predictions.
It should be noted that for a load of l75kN the computed soil
pressu7es are limiting to a depth of 1.5m.
Bearing in mind the imponderables associated with a test

of this type the agreement obtained between the computed and
observed load-displacement curve is encouraging. The computed
elastic curve is also shown for comparison and it is clear
that the assumption of elastic behaviour would be valid for a
maximum force approximately equal to one-quarter of the fai-
lure load only.
CONCLUSION
The two programs discussed illustrate the power of the boun-

dary element approach in providing economic and realistic so-
lutions to normal everyday geotechnical engineering problems.
Although, the particular boundary element described above is
based on elastic theory the various approximate extension that
has been developed to take account of 11 elasto-plastic 11 soil
response would appear to lead to reasonable agreement between
computed and observed behaviour.
ACKNOWLEDGEMENTS
RAFTS and LAWPILE are available on the bureau network of Unit-

ed Computing Systems, whose help in the preparation of this
paper is gratefully acknowledged.
REFERENCES
Burland, J.B. and Hancock, R.J.R. (1977) Underground car park

at the House of Commons, London : Geotechnical aspects. The
Structural Engineer, 22, 87.
Burland, J.B. and Davidson, W. (1976) A case study of crack-
ing of columns supporting a silo due to differential founda-
tion movement. Performance of Building Structures, Pentech
Press, London.
Dunn, C.S. (1975) Settlement of a large raft foundation on
sand. Settlement of Structures, Pentech Press, London.
Gerrard, C.M., and Harrison, W.J. (1970) Circular loads app-
lied to a cross-anisotropic half space. CSIRO Aust. Div. Appl.
Geomechanics Tech. Pap. No. 8.
Harvey, R.C. and Burley, E. (1977) A proposal for the develop-
ment of a Sea-Bed anchor. Journal of the Society of Under-
water Technology, ~. 9-ll.
324
Hooper, J.A. and Wood, L.A. (1977) Comparative behaviour of

raft and piled foundations. Proc. 9th Int. Conf. Soil Mechan-
ics and Foundation Engineering, Tokyo.
Mindlin, R.A. (1936) Force at a point in the interior of a
semi-infinite solid. Journal of Physics, 12·
Nath, B:, Harvey, R.C., and Burley, E. (1978) The development
of an embedded anchor to provide multi-directional restraint.
Proc. Conf. on Offshore Structures, Oceanology Int. 78, Inst.
Civ. Engrs., London.
Poulos, H.G. (1973) Load deflection prediction for laterally
loaded piles. Aust.Geomechanics Journal, Q2.
Simpson, B., O'Riordan, N.J and Croft D.D. (1979) A computer
model for the analysis of ground movements in London Clay.
Geotechnique, 29, 149.
Wood, L.A., and Larnach, W.J. (1975) The effects of soil-
structure interaction on raft foundations. Settlement of
Structures, Pentech Press, London.
Wood, L.A. (1977) The economic analysis of raft foundations.
Int. Journal Num. Anal. Meth. Geomechanics, 1, 397-405.
Wood, L.A. (1978a) RAFTS - a program for the analysis of
soil-structure interaction. Advances in Engineering Software,
];_, 11-17.
Wood, L.A. (1978b) A simple boundary element approach to the
prediction of the settlement of structures. Recent Advances
in Boundary Element Methods. ed. C.A. Brebbia, Pentech Press,
London.
Wood, L.A. (l979a) LAWPILE - a program for the analysis of
laterally loaded pile groups and propped sheetpile and diaph-
ragm walls. Advances in Engineering Software,];_, 173-179.
Wood, L.A. (l979b) A rat~onal approach to the analysis of
building structures taking full account of foundation move-
ments. Engineering Software, ed. R.A. Adey, Pentech Press,
London.
Wood, L.A. (1980) An unusual soil-structure interaction pro-
blem associated with the Thames Barrier Project. Proc. Instn.
Civ. Engrs. Part 1, 68, 747-758.
325
BOUNDARY INTEGRAL METHOD FOR POROUS MEDIA
M. Predeleanu
Laboratoire de Mecanique et Technologie

Universite Paris VI, E.N.S.E.T., Cachan (France)
INTRODUCTION
The physical non-homogeneity of some materials of technological

interest such as soils, concrete, composite and metals has been
modelled mainly by continuum theories applied to the hi-phasic
media. The concept of porous medium is used at present not only
in geomechanics but also in metal science to study the damage
and rupture phenomena.
The fluid-saturated porous medium has received much atten-

tion due to the applications in hydraulics and mechanics of
soils. The first model describing the uni-dimensional consoli-
dation of soils was proposed by Terzaghi (1923), but Biot
(1961) has developed a three-dimensional consolidation theory.
Further generalizations of this theory have been made by its
author and recently the non-linear effects have been included
(1972).
The interest of the consolidation theory has increased

because of satisfactory results obtained in the calculations
of Venice subsidence (see, for instance Lewis and Schrefler
(1978))and of seismic response of Lower San Fernando Dam (see
Zienkiewicz, et al. (1980)}
From the mathematical point of view, the consolidation

Biot theory is governed by coupled equations system describing
the interaction of fluid flow in a porous deformable skeleton.
The linear rheological behaviour of solid phase and small
strains kinematics leads to linear governing equations.
This paper is concerned with Biot consolidation theory,

extended to a general class of viscoelastic bodies defined
by Riemann-Stieltjes integral convolutions. Integral repre-
sentations of the displacements and pore pressure are obtained
so that they may be usedfor the formulation of the corresponding
326
boundary integral equation method. The analogy with the coupled

theory of the thermo-viscoelasticity - for which the present
author hasobtained some recent results (1979) -permits us to
use the fundamental singular solutions in the theory of the
porous media. This approach can be extended to the non-linear
behaviour and coupled with other numerical methods.
GOVERNING EQUATIONS
Consider a porous deformable body, occupying a region B with

a regular boundary aB, in the three-dimensional Euclidian
space, referred to a fixed rectangular cartesian coordinate
system oxi, i = I, 2, 3. By Xi will be noted the coordinatffi
of the polnt x E B.
Let t be the time variable, denoting by t the present

time moment and by T the values of t forT < t. We assume, for
shortness, that every particle of the body is at rest for T <o
and at T = 0, the body is in a free stress reference configu-
ration.
The material under consideration is a hi-phasic continuum,

composed of a compressible fluid which diffuses through a po-
rous deformable solid. We are dealing with the average displa-
cement of a volume element of material, sufficiently large
with respect to the pore size. The displacement vector compo-
nents of a fluid and solid particle are denoted respectively
by ui and ui·
The small strain tensor Eij of the solid skeleton is

defined by :
2 (u·l,J· + uJ· ,i)

E·. _!_ (I)
lJ
The volumic dilatation of the fluid will be noted by

8 = U11 and the volumic dilatation of the solid skeleton
with s:, ul,l where ui,j = ()ui/dxj.
Following the Biot theory, the measure of total stressest ..

acting on the mixture is decomposed in"aneffective " stress lJ
tensor aij characterizing the solid skeleton and a spherical
stress tensor o0 .. characterizing the fluid interaction. Thus
lJ
lJ
t .. = a ..
lJ lJ
+ao .. (2)
with a = - fp, where f is the porosity of the material and p

the fluid pressure.
Obviously to take into account a more general fluid

influence on the solid phase, a general stress tensor must be
considered. Various constitutive relations can be used for
327
the theory of porous media. To include the visco-elastic

effects of the mixture behaviour, we shall generalize the
Biot theory by using the following linear-stress strain rela-
tions
a ..
l.]
2N x d E: .. +
l.J
I A X dE: + Q X dG J 0 ij (3)
a Q X dE: + R * dG (4)
where by A X d 1jJ is noted the Stieltjes convolution of
two real valued functions A and 1jJ defined by
(Ax d 1)!) (t) = ftA(t-<r) d 1jJ(t) (5)

-co
under appropriate conditions formulated for instance by Gurtin

and Sternberg (1962).
Equations (I) - (4) must be completed by the quasi-static

stress field equations
( a. .
l.J
+ ao l.J
.. ) , J. + F. =
l.
o (6)
a, i i =
..
and by the fluid flow equation of the Darcy type
s C e- s + q ) (7)
where F· are body forces, S a filtration constant, q the

external source of fluid 8 = 38/Clt.
If it is assumed that the body is originally undisturbed,

the initial conditions will be considered zero on B X c-~,o).
The stress boundary conditions are taken as

(a .. +& . . a)ni Pi (x,t), x E. 3B, t> 0 (8)
l.J l.J
or in the form :
a .. n. =(1-f)Pi(x,t), xE;()B, t> 0 (9)

l.J J
ani= fpi = g(x,t)ni' xE(lB, t> 0 (10)
where ni are the component of unit normal vector.
INTEGRAL REPRESENTATION
To deduce the integral representation of the displacement

field u. and fluid prea;ure p a generalized reciprocal theorem
of the ~etti type, deduced by Predeleanu (1968) will be used.
328
Gonsider two loading programs

J(a)= {P.(a), g(a), F.(a), q(a)}a= 1,2 (II)
~ ~
which induce two deformed configurations of the body B, defined

by the fields
$(a)= {u. (a) u. (a), (a) (a) }
~ , ~ oij ,a ,a = I ,2 (12)
Reciprocal theorem
If the porous med~um B is subjected to two loading programs
J(a), = 1,2, then governing fields of the corresponding defor-
med configurations verify the following reciprocity relation
1 2 - t (I) • (2)
L12 (J ,¢ )= !ds(x) f P. (x,t-T)u. (x,T) dT
3B o ~ ~
I t (I) (2)
+ 0 f 3s(x) f g (x,t-T)O,i (x,T)nidT
~ 3B o
+ £ dv(x) £ t
Fi
(I)
(x,t-T)ui
• (2)
(x,T)dT
+ f dv(x) ft q(l)(x,t-T)o( 2)(x,T)dT =

B o
2 I
1 21 (J ' ¢ ) (13)
This theorem has been demonstrated by using the Laplace trans-

form in respect of time variable t and it differs with other
analogous theorems in the coupled theories by eliminating the
volumeintegral terms concerning two unknown field variables.
This reciprocity relation allows a more suitable use for the
boundary integral approach. It is noted in passing that the
dynamics effects can be takminto account and a similar theo-
rem can be deduced as in coupled theory of thermoviscoelasti-
city (see Predeleanu 1979).
Integral representation of the Somigliana type
The use of the fundamental solutions of the governing equations
and the reciprocity theorem permit us to deduce the integral
representation of the Somigliana type for the displacement field
and fluid pressure p.
Firstly, consider
k k k k k
¢={a. ,V. ,A .. , C }, k = 1,2,3, (14)
~ ~ ~J
a fundamental solution of the governing equations, regular at

infinity, with homogeneous initial condition corresponding
to the loading
329
0 (15)
where 0 is Dirac distribution.
By a.k(x,~,t) is noted the i-component of the soliddis-

Placement 1 at x due to a concentrated load in the o~-direc
tion, applied at a point ~ of the porous medium, that occupies
the entir~ space ; v.k is the displacement of the fluid par-
ticle, Aiik• the tot~l stress tensor and ck, the corresponding
solution Ior a.
We obtain
~(~,t)+ fds(x) /A~(x,~,t--r)~(x,T) dT+

aB o nl (16)
t k k
]~ds(x) f C
0
(x,~,t-T)a,.
l
(x,T) n.dT = L(J,~ )
l
where
k k
A . =A .. n.
nl lJ J
L(J,~ k) is a known function defined by Equation (13) where

J ={pi' g, Fi, q} (17)
and~ k is given by Equation (14).

Secondly, consider
(18)
a fundamental solution of the governing equations regular at

infinity, with homogeneous initial conditions, corresponding
to the loading
Fi(x,t) (19)
From the reciprocity relation (13), we obtain :
a ( ~,t) + ~f ds(x) / Bnl. (x, ~,t-,) u.1 (x, T) d-r +

0
1 t (20)
+- f ds(x) f e(x,~,t--r)a ,. (x,T) n.dT=L(J,~)
f3 aB o l l
where
B •
n1
= BlJ
•• n .•
J
J and cp are defined respectively by Equations (17) and (18).
Equations (16) and (20) define a coupled linear integra-
differential equations system for the unknown fields ui anda.
330
Analogous integral representations can be obtained for

mixed boundary value problem.
The Laplace transform can be applied to eliminate the
time dependence and reduce the viscoelastic problem to an asso-
ciated elastic problem. Also, important simplifications of this
system can be obtained if fundamental solutions of Green's type
are used. In this case the homogeneous boundary condition. is
required for the singular solution of the governing equations
A .k B . = ck = e = o, X€ CJB, t > 0 (21)
nl. nl.
From Equations (16) and (20) it is easy to see that the dis-
placement field and the fluid pressure are directly determined
in every interior point of the medium if the boundary and
inside inputs are known
uk(l';,t) = L(J,<Jh,i';;E B, t >0 (22)
o(i';;,t) = L(J,¢) ,i';;E.B, t>O (23)
It sho~ld be noted that, for appropriate definitions of the

Green's functions, it is possible also to express the solution
of the mixed boundary value problem exclusively in terms of
prescribed functions by volume integrals and surface integrals.
Total volumic dilatation formulas

Let us define the total volumic dilatation of solid and fluid
phase respectively by
! E(x,t) dv(x) (24)

v
f 8(x,t) dv(x) (25)

v
By using the reciprocity theorem, the total volumic dilatations
can be obtained in terms of boundary and inside prescribed
data.
Thus, following the results obtained by Predeleanu

and Nan (1968), we can write:
Dl (t)= [ ds(x) /m(t-T)P. (x,T) x.dT +

CJB o l. 1
t
f dv(x) f n(t-T) F.(x,T) x.dT +
B o l. l.
3 f dv(x) !t[m(t-T) +n(t-T)jo(x,T) dT (26)

B o
D2 (t) = -a~ ds(x) ftn(t-T)Pi(x,T) x. dT-

0 1
f dv(x) ft n(t-T) Fi(x,T) xi d T +
B (27)
0
331
+ 3 f dv(x) ft[n(t-T)+ r (t-r)] a(x,T) d T

B 0
where the functions m, n and r are obtained by inverse

Laplace transforms
-I
-
R -1 Q
m(t) = L ------~~--------)n(t)=L
p(2NR + 3 AR - 3Q 2) p(2NR+3AR-3Q2)
-I
r(t) L 2N + 3A (28)
by f(x,p) is noted the direct Laplace transform of a numeri-

cal function f(x,t).
Obviously, the formulas (26) and (27) became very

simple in the case of elastic behaviour of solid skeleton
(see Predeleanu and Nan (1968).
BOUNDARY INTEGRAL EQUATION APPROACH
One of the most fruitful applications of the integral

representations obtained in the previous section is the
numerical approach by using the boundary integral equation
method (BIEM). The BIEM has been utilized successfully in
the non-coupled linearized theories of continua and that
has been extended to some non-linear theories. In this
section the BIEM is formulated for a coupled theory of
fluid-saturated porous media and this method is proposed as
an alternative to the finite element method (FEM) used for
geotechnical calculations. It is proved by numerical appli-
cations that for infinite domains and also for the large
domains the BIEM is more convenient that the FEM.
As is known, the boundary integral equations can be

obtained from integral representation of Somigliana 's or
Green's type if the point ~inside B approaches an arbitrary
point on the boundary () B .
For that let us note formally by integral symbol with

an asterisk the following limit understood in the sense of
its Cauchy principal value :
lim fr~, (x, ~,t) w(x,t)ds(x) (29)

~ -+y ClB
= ~~ rl(x,y,t) w(x,t)ds(x)
where ~ € B and y € () B.
Thus, from Equations (16) and (20), we obtain
• ~ t k .
u(y,t)+; ds(x) f A.(x,y,t-T)u.(x,T)dT-
k dB 0 m. 1
332
.f t k
- f ds(x) f a.(x,y,t-T)t .(x,T) dr +
() B o 1 n1
..!. jds(x) JtCk(x,y,t-T) 0 ,. (x;r:)n.dT
S ()B o • k 1 1
1 dv(x) jtFi(x,t""'[)ai (x,y,T) dr + (30)

B o
f dv(x) Jt q(x,t-T)C k (x,y,T) dT +
B o
~!_ d s ( x ) it g ( x,t-T ) C,ik (x,y;r)nidT

s ()B o
g(y,t) + 1" ds(x) / Bni (x,y,t-T) ui (x,T) dT-

()B o
' ds(x) ft b. (x,y,t-T) tni (x,T) d T +

()B o 1
I
- , . ds(x) fte(x,y,t-T)o ,.(x,T)n.dT (31)
S ()B 0 1 1
f dv(x) jtF.(x,t-T) bi(x,y,T) dT +

B o 1
J dv(x) /q(x,t-T) e (x,t,T) dT +

B o
..!. *'!- ds (x) 0ft g ( x,t-T ) e,i ( x,y,T ) ni d T
S ClB
where t .
n1
= t 1J
.. n.
J
It is worth noting the following:
a. For the stress boundary value problem in which surface
tractions and the fluid pressure are prescribed over the
entire boundary () B for all time, the third terms of the
Equations (30) and (31) are known. On the other hand for the
mixed boundary value problem, these theorems contain the
unknownfields on some parts of the boundary.
b. For certain singular Kernels Q (like A.. and B.. for
instance) and regularity conditions for 1 J the 1 J density
wand the boundary () B some additional terms (jumps) of type
nw(y,t) can be separated from the singular integral defined
by Equations (30)-(3l).The calculations of these limits are
performed after eliminating the time convolution integrals
by Laplace transformation. Otherwise, the regularity condi-
tions must be imposed for the time and spatial integral
operations.
c. The analogy between the consolidations theory of porous
media and coupled theory of thermo-viscoelasticity allows the
use of some fundamental solutions (see Nowacki (1962), Derski
(1965)).If supplementary hypothesis concerning the non-cou-
pled interaction between the two phases, is adopted, then
important simplifications may be introduced. On the other
333
hand, singular fundamental solutions known for the non-

coupled problem can be used for coupled problems by using
the method proposed by Ionescu-Cazimir (1964).
FINAL REMARKS
Another application field of the boundary solution approach

is the coupling of BIEM with other analytical or numerical
methods(the finite elements method or finite difference
method, for instanc~.Forthat a natural or arbitrary boundary
r can be introduced to separate the parts B1 and B2 of B.
If in B2 , the BIEM is used, then the effects of B2 on B1
must be i:ncluded by an"i.nterface "condition on common boundary
r. For instance, if u. (I), a (I) and u. (2), a (2) are the
fields characterizing fespectively the iegions B1 and B2 it
may be asked that
(I) (2)
ui (y,t) ui (y,t) YE r t>O
(32)
a<O(y,t) a ( 2 ) (y, t) yE r t >0
Other interface conditions can be assumed too.
By using the integral representations , a global

integral boundary condition can be obtained for u.(l) and
a (I) to be used for instance in FEM as a natural~boundary
condition for the variational formulation of the problem to
solve in B1 (see Predeleanu 1979).
REFERENCES
Biot M.A. (1961) General Theory of the Three Dimensional

Consolidation.
J. Appl. Phys. ~ : 155-164.
Biot M.A. (1972) Theory of Finite Deformations of Porous

Solids.
Indiana University Mathematics Journal, ~. 7 : 597-620.
Derski W. (1965) Equations of the Consolidation Theory for

The Case of a Source of Fluid.
Bull. Acad. Polen. Sci. Ser. Sci. Tech~. I : 37-43.
Gurtin M.E. and Sternberg E. (1962) On the linear theory of

Viscoelasticity.
Arch. Rational Mech. Anal. ~. 4 : 291-356.
Ionescu-Cazimir V. (1964) Problem of Linear Coupled Thermo-

elasticity. Some Applications of the Theorems of Reciprocity
for the Dynamic Problem of Coupled Thermoelasticity.
Bull. Acad. Pelon. Sci, Ser. Sci. Tech, ~. 12
334
Lewis R.W. and Schrefler B. (1978) A Fully Coupled Consolida-

tion Model of the Subsidence of Venice •
Water Resources Research ~. 2 : 223-230.
Nowacki W.(l966) Dynamic Problems of Thermoelasticity.

Noordhoff International Publishing Leyden, P.W.N. Polish
Scientific Publishers, Warszawa.
Predeleanu M. (1968) Reciprocal Theorem in the Consolidation

Theory of the Porous Media (Roum).
Anal. Univ. Bucuresti, SeriaStiintele Naturu Matematica-
Mecanica l2• 2 : 69-74.
Predeleanu M. and Nan L.(l968) Formul~ for the Variation of

Volume in the Porous Media.(Roum).
Anal. Univ. Bucuresti Seria Stuntele Naturi.i Matematica-Meca-
nica, l2• 2 : 75-79.
Predeleanu M. (1979) On Boundary Solution Approach for Dynamic

Problem of Thermo-Viscoelasticity Theory •
International Conference on Numerical Methods in Thermal
Problems. Swansea 2-6 July 1979. (in press).
Terzaghi K. (1923) Ann. Oster. Akad. Wissensch Wien Kl.

Abt. Ila : 132.
Zienkiewicz O.C. - Hinton E. - Leung K.H. - Taylor R.L. (1980)

Staggered Time Marching Schemes in Dynamic Soil Analysis and
a Selective Explicit Extrapolation Algorithm.
Innovative Numerical Analysis for the Engineering Sciences
Ed. R. Shaw,University Press of Virginia. Charlottesville.
Section IV
Material Problems
337
NUMERICAL ANALYSIS OF CYCLIC PLASTICITY USING THE BOUNDARY

INTEGRAL EQUATION METHOD
M. Brunet
Laboratoire de Mecanique des Solides - INSA Lyon - Bat. 304

69621 Villeurbanne Cedex - France
INTRODUCTION
Reliability criteria for the design of engineering compo-

nents and structures subjected to complex service conditions
often include their resistance to fatigue failure. Low cycle
fatigue may be part of the design criteri8 and a major drawback
in life predictions has been the inaccuracy in relating nominal
loads to local stresses and strains, particularly if local plas-
tic deformation occurs. Finite element methods (FEM) are mostly
being used for such problems and the cyclic plastic analyses
usually require large amounts of computer time. The boundary
element method (BEM) is less used and it is being explored for
this class of non-linear problems. In this paper a computational
procedure is presented for the solution of such problems in pla-
nar bodies subjected to arbitrary loading histories. LinPar and
non-linear constitutive relations are used in the calculations.
PLASTICITY RELATIONS
The plastic behaviour of a material can be described by a

yield function such as :
f(o .. -a .. )-o~=O ( 1)
lJ lJ
a .. is the stress tensor, 0 0 the yield stress and a .. is a

t~tlsor defining the total translations of the yield 1 ~urface. The
flow law is defined by the normality condition of the plastic
strain increment tP . . , to the associate yield surface at the
stress point. lJ
Therefore
•p • ()f
E: •• \-- (2)
lJ dO ..
lJ
338
in which ~ is a positive scalar determined by the condition that

the stress point remains on the yield surface during plastic
flow, such as :
0 ( 3)
It is to be noted from the incompressibility of plastic strains

that
8 .. E•p •• -- 0 (4)
lJ lJ
In Prager's rule, the increment of translation of the yield

surface is in the direction of the plastic strain increment.
Therefore,
a .. = C f.P .. ( 5)
lJ lJ
Ziegler's modification of Prager's model sugges~instead of (5)

the relation :
alJ.. = ~(o lJ
.. -a .. )
lJ
( 6)
With constant coefficients, Equations (5) - (6) describe linear

kinematic hardening and when these models are applied to cases
of cyclic stressing, they predict closed cyclic loops after one
cycle of stress. This prediction is not corroborated by numerous
experimental data and a more realistic description can be obtai-
ned when accumulation effects accompany plastic cycles. Mroz
[1] and others postulate instead of (5) and (6)
~-.
lJ
= C(sp) f_P .. - d (f:P) f:P Ep..
lJ lJ
(7)
with
X ••
lJ
1
a .. - -3 8 . . a
lJ J.J mm
•
and ~p
~ -- [23 •p ij E•p ij
E
J 1/2 (8)
where c(£P) and d(sp) are two material functions which could be
determined from uniaxial cyclic tests.
Marquis [2] modified (7) and proposed the hardening rule :

• _ 2 •p 2 -w£P :p
X· • - -3 S0 [ E • • - -3 (c)> 0 + lj!e ) E X •• ] ( 9)
lJ lJ lJ
o 0 , S0 , cJ>u, 1jJ and ware five parameters characterizing each

material obtained by trial and error from uniaxial experimental
data . This law can account for the cyclic transient behaviour
of most structural materials. When hardening develops, the term
339
-wE;P
~e vanishes in (9) and we get the cyclic steady behaviour
when ~ = w = 0.
In order to complete the constitutive relationships, assume
that the total strain increment can be decomposed into an elas-
tic part and a plastic part :
e
in which Dijkl 1s the elastic material tensor :
D~.kl
lJ
= __ E_
(1+v)
[o.ko.l + o.lo.k + (1:v2 )o .. okl]
1 J 1 J v lJ
(11)
2
where E = Young's modulus and v = Poisson's ratio.
Using equation (5), the following relationships are

obtained
A
• =1-[- 3f
-De •
E
a.+S acrij ijkl kl
J ( 12)
with
a + S (13)
and
•p
E •• ( 14)
lJ Gijkl Ekl
ep •
(J ••
Dijkl Ekl ( 15)
lJ
in which
G
ijkl
= _1_[lf_
a+SL3crij
De _lf_J
ijkl 3crk1
( 16)
ep e _ D~. ( 17)
Dijkl = Dijkl lJmn Gmnkl
We explicitJy formulate equation ( 15) for the special case

of Von Mises yield condition and plane-stress state.
f = ~ S .. s .. - l cr~ = 0 in which S .. is the translated

deviatoric st~e~~. d~fined by lJ
1
s lJ
.. (cr .. -a .. ) - -3 8 .. (a
lJ lJ lJ mm
-a )
mm
( 18)
340
0
8 0
011 -(811+\1822) 1 ~~ E 11
8
0
E
022 -(822+\1811) 1 ~~ E22
Q
012 8YM.
( 19)
where
R and P
Q = R + 2(1-v 2 ) P
with
*A=~ o~ H'/E for Prager's rule (or Ziegler's rule)

wher~ H' is the plastic slope of the uniaxial stress-
strain curve.
* A= 32 [
Bo 2o~ - (~o + ~e
-w£P l
)o 0 Tj
for the complex non-linear hardening rule (9). These cases are
incorporated in the solution procedure analyzed with the boun-
dary element method.
BOUNDARY ELEMENT FORMULATION
The boundary integral formulation for the two dimensional

elasto-plastic problem without body forces are presented in
this section following the results of 8wedlow and Cruse [3]
and Mukherjee [4] . The range of subscript indices in all the
following equations is 1, 2.
C..
lJ J
rUlJ.. (P,Q)i.(Q)df
~.(P) = Jr' J
-
lr'~1 lJ.. (P,Q)~.(Q)df
J
(20)
+ ~n LlJ. .k(P,q)e:.k(q)
op
J
dn
where u. is the displacement vector, t. the traction vector,

n is th~ surface of the planar body, rJits boundary, p and Q
are boundary points, p and q are interior points respectively.
This expression allows us to determine the boundary unknowns
341
and once this is done, the displacement incremenware given by

(20) with C.. = o .. and P = p
1J 1J
The kernels U.. and T .. are known singular solutions due
to a point load in 1 ~ infi~ite elastic solid and are available
in many references such as C.A. Brebbia [5] . The kernel L: •• k
for the case of plane stress is given by (21) with v replac~a
by v
= vI ( 1+v )
L.1J"k 4rr ( ~ -v ) r [ ( 1-2v ) ( 01J r , k + ok 1. r ,J. o.kr .)
J
0 0
,1
+ 2 r .r .r k] ( 21 )
'1 ,J '
The kernelZ is Etrongly singular and in order to perform

the derivative, H.D. Bui [6] has arrived at the correct solu-
tion by employing Mikhlin's concept of derivatives of singular
integrals.
Internal stresses can then be calculated and for the plane

stress state we obtain [7]
(22)
where (23)
- L: ijk
and
S1·J·k = G ~ 2 ~r [(1-2\!) o .. r k
2rr( 1-v)r 2 I on 1J •
+ v(o.kr . + Q.kr .) - 4r .r .r k] + 2\!(n.r .r k

1 ,J J ,1 ,1 ,J ' 1 ,J '
+ n.r .r k)+ (1-2\!) (2nkr .r + n. o.k

J ,1 ' ,1 ,J J 1
0
+ niojk) (1-4\i) nkoijf (24)
The last integral in (22) 1s to be taken in the Cauchy

principal value sense in which
342
G
- - - - { 2( 1-v) (oiJ" r.k r.l + okl r . r . )
27T( 1-v)r 2 •1 .J
(25)
NUMERICAL IMPLEMENTATION
The boundary of the region is discretized into a series

of elements which can be constant or linear in the program deve-
loped. With the linear elements, the tractions must be allowed
to have discontinuities across a node at corner. Additional
equations may be obtained using the symmetry of the stress ten-
sor and the compatibility between strains and displacements. In
both cases, the integrals in equation (20) have been evaluated
in closed form for straight boundary elements. We assume the
incremental values of internal plastic strains to remain cons-
tant over rectangular or triangular internal cells. The inte-
grals in equation (22) are calculated analytically for straight
boundary elements and rectangular internal cells and numerical-
ly for triangular internal cells using three or six Gauss-
Cowper points.
Equation (20) can then be written in matrix form
(26)
After applying the boundary conditions the system can be
reordered and we obtain
(27)
where {~} contains the unknown displacement and traction incre-

ments and {f} is obtained by multiplying the known displacement
and traction increments with the appropriate columns of matrices
[H] and [G].
Similarly, equation (22) gives
(28)
in which [c•J represents the independent terms and [n•] stands
for the plastic strain integral. As it was shown in [7], the
343
principal value can be computed by applying equation (28) to

represent a state of constant plastic strains.
ITERATIVE SOLUTION PROCEDURE
An incremental solution algorithm based on equations (27)

and (28) can be developed as follows :
The first increment is scaled up to magnitude at which

first yield is encountered at the centroid of an internal cell.
Starting with {EP} = 0 at each internal cells, the following
steps are common for each load increment.
1. We apply a loading increment {f} and we solve

(29)
We obtain the incremental values of unknown boundary displace-

ments and tractions and update the boundary vectors.
2. By using equation (28) and noticing that :
(30)
We calculate the str~ss increment {cre} and with the current

stress state {a} + {ae} we compile a list of yielded cells.
For the points in the plastic range, we recalculate the correct
stresses {a} with the aid of the elasto-plastic stress-strain
relation (19) where the total strain increment {E} is imposed
by the elastic calculation using {cre}.
Therefore [ne}-1 {cre} ( 31 )
and [neP] {~} (32)
The current stress state necessary to evaluate [Dep] may

be taken as the average during the increment such that no impor-
tant departure from the true yield surface is observed.
3. We update stresses and strains arid compute the plastic

strain increment as
{E} - [ne] {cr} (33)

and
{cre} - {cr} (34)
The. hardening parameters such as the equivalent plastic

strain {IP} which is a state variable defined by (8) and the
translation of the yield surface {;} are calculated. If necessary
some corrections of the stress level can be made to maintain
the stresses on the true yield surface.
344
4. The non-linear "nodal boundary contribution" of the

plastic-strain increment is computed as
(35)
Then we solve
(36)
and update the boundary displacement vector.
5. The process is continued until some tests of conver-

gence ensure sufficient accuracy for termination. An efficient
criterion formulated on the basis of the equivalent plastic
strain increment has been found. The criterion takes the form
for the (k+1) iteration
If
go to step 1, otherwise go to step 2.
As a result, it hasbeenfound from numerical experience

that this method is convergent and stable for even large size
increments.
NUMERICAL EXAMPLES
The first example is a rectangular thin strip modelled by

20 linear boundary elements (or 40 constant B.E.) with 16 rec-
tangular internal cells as shown in figure 1. Axial loads were
applied at the free end of the strip and the other end was res-
trained against any displacement, such that a two-dimensional
stress distribution exists. The iterative procedure is seen to
converge and for example, 6 iterations in each load increment
are required using the linear B.E. and 8 iterations with the
constant B.E. As shown, good correlation between the B.E. method
and the exact solution [8] is achieved.
Figure 2 shows numerical results of the same two dimensio-

nal strip. The material is stainless steel A 316 L (Nuclear en-
gineering) idealized by the non-linear hardening rule (9). The
figure displays the capability of the present non-linear har-
dening model with the B.E. method in describing the cyclic tran-
sient behaviour and the cyclic steady behaviour.
345
II£Pll~1 -li£PII~
II£PII~1
0.1
~p
o constant BE (40)
• linear BE (20)
E
0.05 E
0
00
/" " / / " / / / / i / /
40mm
0 E = 70000MPa
0 234 567 8 v = 0.33
Rate of convergence ( step 3 ) t = 1.0 mm
H'= 0.3E
do= 300 MPa
daN Applied load P
2000
BE Analysis Kinematic
1000 • Hardening. LlP=120 daN
Exact Solution ref [BJ
0.25 0.5 0.75 1 mm

Vertical deflection at free end
Figure 1
346
MPa
200
I
I
I
I
I
I
-0.2 I
j
I
I
I
I '
I
I
I
I
I
I
I
A
K
I ............. 20 BE Analysis
Cyclic transient behaviour
""'~ -·- Cyclic steady behaviour
2 (}' =±280 MPa

max Mat.: A 316 L stainless steel
complex hardening model
do 200 MPa
~ t:J t:22
1<122 y
E 200000 MPa
0.3
E 8 0 = 54666 MPa
E
0
co
f IIJo 0.00513 MPa-1
~ == 0.011 MPa-1
.Q 65.3
1
...-_.....-/ '%
40mm""
Figure 2
347
M Pa Local stress 0"22
600
Local strain C22 %

I
-0.5 I 0.5
I
I
I
I _.__ BE analysis
I
rf Experimental
--0--
I
data ref [9]
I
I
residual stress
I
-600
Mat.' SAE 4130 normalized

kinematic hardening model
E = 189270 MPa
do= 470 MPa
H' = 0.35 E
Internal cell
or finite element
348
The third example is a notched bar shown in figure 3 used

for the cyclic elasto-plastic analysis of an experimental stu-
dy carried out by Crews [9] . The boundary element model inclu-
des 31 linear B.E. and the finite element model includes 158
degrees of freedom and 127 constant strain triangular elements
which serve as internal cells in the B.E. computation. The
material is steel alloy SAE 4130 idealized by the Prager-Ziegler
hardening model with a hardening slope of 0,35 E. The calcula-
ted results with the B.E.M. and the F.E.M. using the initial-
stress procedure reasonably correlate with the test data.
As an example of the required computer time, the CPU time

using an IRIS 80 computer for the notched bar is 15,2 mn for
the F.E. analysis (32 load increments with 3-5 iterations for
each load increment and 29 yielded elements) and 18.4 mn for
the B.E. analysis (32 load increments with 3-4 iterations for
each load increment and 30 yielded elements). However, with
the F.E.M., during the process of iteration the stiffness
matrix can be kept constant as above (modified Newton-Raphson)
or can be updated as often as desired and in this case we have
obtained a CPU time of 10.3 mn. But, using a domain discretiza-
tion only where plasticity occurs, the B.E.M. requires a CPU
time of 12.6 mn with 62 triangular internal cells instead of
127.
If for the last example, the computational costs are about

20 %higher than the corresponding solutions obtained by using
the F.E.M., it is important to note that for the B.E. and the
F.E. subdivision in figure 3, the stress concentration factor
in the elastic range was obtained to~= 2.10 with the B.E.M.
and 1.97 with the F.E.M. compared with the theoretical value
of~= 2.11, so the B.E.M. seems to be better.
CONCLUSION
An elasto-plastic boundary element approach for cyclic

plasticity has been presented. Two kinds of constitutive equa-
tions, linear and non-linear hardening with strain memory effect,
have been used in the calculation and the computational costs
can be reduced by further careful programming effort. From the
results presented in this paper, it appears that the B.E.M. is
a promising computational procedure for the solution of such
complicated non-linear problems. One of the advantages of the
B.E.M. applied to plasticity problems is that the domain dis-
cretization is necessary only in regions where plasticity
occurs.
~9
REFERENCES
Mroz Z., Shrivastava H.P., Dubey R.N. (1976) A non-linear

hardening model and its applications to cyclic loading
Acta Mechanica 25, 51-61
2 Marquis D. (1979) Modelisation et identification de

l'ecrouissage anisotrope des metaux. These Universite
Paris
3 Swedlow, T.A. Cruse (1971) Formulation of Boundary inte-

gral Equations for three-dimensional Elasto-plastic flow.
Int.. J. Solids Structures- Vol. 7, 1673-1683
4 Mukherjee S. (1977) Corrected Boundary Integral Equations

in Planar thermoelastoplasticity. Int. J. Solids Struc-
tures - Vol. 13. 331-335
5 Brebbia C.A. (1978) Boundary element Method for Engineers

Pentech Press. Plymouth and London.
6 Bui H.D. (1978) Some remarks about the formulation of three

dimensional thermoelastoplastic problems by Integral Equa-
tions. Int. J. Solids Structures. Vol. 14, 935-939
7 Telles J.C.F., Brebbia C.A. (1979) On the application of

the boundary element method to plasticity. Appl. Math.
Modelling - Vol. 3, December
8 Kalev I., Gluck J. (1977) Elasto-plastic finite element

analysis. Int. J. Numerical Meth. Engineering. Vol. 11,
875-881
9 Crews J.H. (1969) Elasto-plastic stress-strain behavior

at notch roots in sheet specimens under constant amplitude
loading. NASA T.N. D. 5253
350
NEW DEVELOPMENTS IN ELASTOPLASTIC ANALYSIS
J.C.F. Telles and C.A. Brebbia
Department of Civil Engineering, University of Southampton,

Southampton, S09 5NH, U.K.
INTRODUCTION
This paper presents further developments on the application

of boundary elements to plasticity on the lines previously
presented by the authors (Telles and Brebbia, 1979, 1980a,
1980b).
The application of boundary elements (or boundary integral

equations) to nonlinear material problems started with
Riccardella (1973) who presented a valid plasticity formulation
but used a not entirely successful numerical procedure. Later
on Mukherjee and Kumar (1978) produced a formulation based on
original constitutive relations which involved plasticity as
well. All the implementations were then restricted to piece-
wise constant interpolation for the plastic strains.
In 1979 Telles and Brebbia produced the complete and

correct formulation for two and three dimensional bodies,
including the implementation of higher order interpolation
functions for the plastic strains. In 1980 (a and b) they
applied this formulation to study two-dimensional plasticity
problems showing how different numerical procedures could be
properly used; namely initial strain, initial stress and
fictitious tractions and body forces. Emphasis was given to
the accuracy and efficiency of the numerical formulation,
firmly establishing the potentialities of boundary elements in
plasticity.
One of the advantages of the implementation which was not

pointed out in the past is that it overcomes the well-known
disadvantage of finite elements first reported by Nagtegaal
et al.(l974); i.e. the impossibility of representing properly
the limit load due to the severe constraints created by the
incompressibility of the plastic strains.
351
This difficulty is always present when internal strains

are computed from the derivatives of the interpolated dis-
placements within the finite elements (or internal cells), and
is likely to occur in the approximate boundary element formula-
tion presented by Banerjee and Cathie (1980).
In th~ present paper the plasticity formulation is extended

to solve half-plane problems using the elastic fundamental
solution of Melan's problem. This author (Melan, 1932) presented
the stresses due to point loads within the semi-infinite plane.
Telles and Brebbia (1981) extended this work to compute dis-
placements, strains and surface tractions, presenting the
complete fundamental solution ne~ded in generalized boundary
elements. This solution is the one used in the last two
examples shown here.
In addition to the known advantages of the Kelvin elasto-

plastic formulation, the main benefit of the half-plane
implementation is that boundary elernenw are not needed over the
traction-fre·e surface of the semi-plane. This property is of
great significance in problems connected with soil-structure
interaction, mining engineering, tunnels, etc. and has not been
presented before for pl asti city.
BASIC EQUATIONS
The incremental equilibrium equations for three-dimensional

bodies in terms of displacement rates (~.) and plastic strain
rates (~~.), are ~
~J
b.
I 2(~~- . + _v_ ·p ) - _1_
u + --u ( I)
j, U l-2v t, Q,j ~J,~ l-2v Ekk,j G
where b. represents the body force rate components, v is

Poisson3 s ratio and G is the shear modulus. Time derivatives
(or simpl~ increments) are indicated by a dot and space
derivatives by a comma.
The traction boundary conditions for equation (I) can be

written as
G(~ . . + ~- . )n. (2)

~.J J,~ J
where p. represents the traction rate components and n. stands

for the~direction cosines of the outward normal to theJboundary
of the body.
352
If plastic strains are considered as initial strains, the

application of Hooke's law to the elastic part of the total
strain rate tensor leads to the following expression for the
stress rate components,
a .. ( 3)
~J
in which o~J
.. is the Kronecker delta.
Alternatively, equation (3) can be written in terms of

initial stresses,
a .. (4)
~J
where a~. stands for the components of the "initial stresses"

given b~J
•p
a .. (5)
~J
Expressions (I) to (5) are valid for plane strain (i,j,~

1,2; k = 1,2,3) and also for plane stress (i,j,k,~ = 1,2) if
vis replaced by~= v/(l+v).
HALF-PLANE FUNDAMENTAL SOLUTION
The complete fundamental solution for half-plane problems can

be obtained by adding to the well known two-dimensional Kelvin
fundamental solution a complementary part. This provides
satisfaction of the traction-free condition over the surface
of the half-plane. In order to save space, only the comple-
mentary part of the solution will be presented, but the complete
fundamental solution is obtained by the following relation,
( ) * (6)
k c
where ( ) and ( ) stand for Kelvin part and complementary
part respectively.
In. the present work the traction-free surface of the half-

plane is assumed to be represented by the axis x 1 = 0.
With reference to figure I, t~e complementary part of the

components of the displacements u .. , corresponding to the dis-
placement in j direction due to a~~it force acting in i
direction will be shown. For plane strains they are given by
353
s load point
q field point
s' omage of s
IP1 1= IP21= 1
R,
s
x,
Figure 1 . Unit point loads applied within the half- plane
(
Kdl- [s(l-v) 2 - (3-4v)] ln R +
[0-4v) R~ - 2c xJ
( 7)
Kd { - [}l(I-v) 2 - (3-4v)] ln R +
[0-4v)ri + 2c xJ
----= -----
R2
354
where the following notation was used,
e arctan(Rz'R 1)
r = (r. r.) !
1 1
I
R = (R. R.) ~
1 1
r. x. (q) - x. (s) (8)

1 1 1
R. =x.(q) -x.(s')
1 1 1
c = x 1 (s) > 0
X = XI (q) > 0
K = I I [81rG( 1-v)]
d
The complementary part of the stresses due to unit forces

in i direction (represented by cr~k.) are written as,
J 1
2R 1(R~ + 2cx) - 4xr~(l-2v)

_ K {(3~ + c)(l-2v)
3 + --~~----------~-
R2
2 [x 2 - 2cx - c2 + 2xR 1 (l-2v)]

-K r {- (l-2v) +
s 2 R2 R4
+ 16cx Rf }
R6
2 [R (r 2 +2c 2) - 2cr 2+2xr 2 (1-2v)]
- K
s
{<x + 3c)(l-2v) + I 2 2 2
R2 R4
(9)
355
(l- 2 v) 2[c 2 - ~ 2 + 6c~- 2iR 1 (l-2v)]

{
- Kr ------ - --------------------~------
s 2 R2 R4
+ 16ci r~ }
R6
2[(2ci + r~)R 1 - 2iRf(l-2v)]

_ Ks {<3i + c)(l-2v) +
R2
where
Ks = 1/[411(1-v)] ( 10)
The corresponding tractions and strains can be computed

from (9) by using the relations,
c c
p .. (I I)
~J
0 jki "Ilk
and
c =...!_ [crJki - v c
Ejki OR.H ojk] (12)
2G
For plane stress v is replaced by v in all the above

formulae.
It is interesting to note that the complementary part of

the fundamental solution does not present any singularities
within the domain XJ ~ 0 when c > 0. Therefore, many express-
ions already obtained for the Kelvin elasto-plastic implementa-
tion, can still be used in the present formulation.
356
BOUNDARY ELEMENT FORMULATION
The extension of the elastic half-plane boundary element formu-

lation (Telles and Brebbia, 1981) to elastoplastic problems,
follows the same procedure as for the Kelvin implementation.
If we consider the plastic strains to be incompressible,

the starting equation for the initial strain formulation is
given by,
r r *
u *.. p. dr - p *.. u. dr + b.J d(l
c .. u.
1J J Jr 1J J J
r•
1J J J
u ..
(l
1J
r h* •p
J d(l ( 13)
+ 0 jki e:jk
(l
where r represents the total boundary of the body, r• the part
of r in which x1 > 0 and n stands for the domain of the body
that has to be always located within the half-plane x1 ~ 0.
Note that the second boundary integral is performed over r•
because the fundamental solution is zero on the surface of the
half-plane and hence the integral vanishes there.
For plane strain problems the complementary part of the

tensor that multiplies the plastic strain rates is (Mukherjee,
1977),
(14)
whereas for plane stress

he c
(15)
0 jki = 0 jki
Equation (13) is valid for any location of the load point

(s e: r or s e: n ) , provided c. . and the boundary integral over
r' are properly interpreted a~J known from the previous elastic
application of the half-plane fundamental solution.
By suitably modifying the plastic strain rate integral in

equation (13), an initial stress formulation without the
condition of incompressibility of the plastic strains (Telles
and Brebbia, 1980b) can be equally obtained,
c .. u.
1J J
* p. dr -
I u1J..
r
J I
r'
p *.. u. dr +
1J J Ju1]*..
(l
b.
J
d(l
* •p
+
I e:jki ojk d(l (16)
(l
357
•p
where ajk was given in expression (5).
It is interesting to note that in contrast with the initial

strain formulation, equation (16) can now be used for plane
strain and plane stress problems with the replacement of v by
v being the only modification.
STRESSES AT INTERNAL POINTS
It is well-known ~hat one of the advantages of boundary elements

over "domain type" techniques in elastic analysis, is that
stresses at internal points can be accurately computed by
using the appropriate integral equation. The elastoplastic
implementation of the technique is by all means no exception,
and from the beginning the present authors have used the correct
integral expression for computing the internal values of the
stress rate tensor (Telles and Brebbia, 1979). In spite of
this, some boundary element researchers have found the internal
stresses by computing displacements at internal points, and
differentiating them numerically as it is done in finite
differences or finite elements (Banerjee and Cathie, 1980
and Cathie, 1980). This procedure possesses some drawbacks
which are going to be commented on in other sections of this
paper.
In order to compute accurately the stress rates at internal

points, the derivatives of (13) are combined to produce the
expressions for the total strain rate and then substituted
into equation (3). Here one notices that, due to the non
singular nature of the complementary tensors, the derivatives
of the plastic strain rate integral create exactly the same
singularities obtained for the single Kelvin implementation.
Hence, for plane strains one has,
.
a ..
~J
(I 7)
in which the plastic strain rate integral is to be computed in

the principal value sense and fij is the same independent term
obtained for the Kelvin formulat~on. i.e.,
f.. ( 18)
1]
358
In addition,
Ac c c
0 ijkQ. = 0 ijU - v 0 ijDml 0U ( 19)
and
c c
G [aakii + aak~J (20)
ax;- ax.
J ~
where the derivatives are taken with reference to the coord-

inates of the load point. These derivatives together with
uljk and pljk are given in (Telles and Brebbia, 1981).
An interesting feature of the half-plane implementation is

that if the problem to be analysed satisfies the traction-free
condition (pk = 0) over some part of the boundary r - r•,
stresses at points located along this part of the boundary can
be computed as if they were internal points. In order to
validate equation (17) for such cases, only the expression of
fij needs to be modified to take into consideration the limit-
ing case c = 0. This expression can be easily obtained as
follows; let us assume a semi-circular free body, of radius
p, whose straight boundary is contained by the surface of the
half-plane. If body forces are not considered, the application
of a uniform plastic strain field(~.) to this body, will
produce only displacements, internal~Jstresses and tractions
remain zero throughout the process.
The application of equation (17) to represent the stresses

at the centre of the semi-circle leads directly to,
I
r'
(21)
moreover, from the condition of existence of the principal

value (Mikhlin, 1962), one can prove that,
A*
fn
:P <ill 0 (22)
Eki 0ijki
hence,
-p *
f. • (E )
~J J pijk ~
r'
df (23)
where the relevant boundary displacements (neglecting rigid

boundary movements) can be computed by (Telles and Brebbia,
1979),
359
~.
~
= P(~.
~J
- v £'Pkk o~J
.. )n.J (24)
in which nj represents the direction cosines of the outward

normal to the curved boundary.
Equation (23) therefore provides the required expression

for f .. when c = 0,
~J
fl2 = 0
G •p •p
2(1-v) (£22 - £11) (25)
For the initial stress formulation the procedure is entirely

similar and the equation equivalent to (17) is of the following
form,
crij J u;jk pk dr - J p:jk ~ dr + J u:jk bk dn

r r' n
(26)
where E~jk~ is obtained from (20) by substituting cr~ki by cjki'
g .. 8
I _ ~-2~p
- _(I-v) •p oij
_ ij + (I-4v)crH J fore> 0 (27)
~J
and
gl2 = 0
I
- 4(1-v) Lcr22 -all
r:P •p J for c = 0 (28)
Plane stress problems can be handled by equations (17) and

(26) if Poisson's ratio is modified as before, a:jk~ = a:jk~
and expression (18) is replaced by
_ G
f .. - - - - -
r, ·p..
!Jc ·p
+ EH o,. J (29)
~J 4(1-v) ~J ~J
Notice that expressions (25), (27) and (28) still remain valid.
Although body forces have been considered in the equations

presented so far, in the following sections this term will not
360
be included for simplicity.
NUMERICAL IMPLEMENTATION
For numerical implementation of the equations presented in the

previous sections, linear boundary elements and linear triang-
ular cells were used in much the same way as described
by Telles and Brebbia (1980a,b) for the initial strain and
initial stress formulations. It is worth mentioning that
integration of che complementary part of the expressions
presents no singularities when c > 0. Consequently, simple
quadrature formulae can be used for the boundary and domain
integrals. The limiting case c = 0 can be shown to produce
singularities of the same order as those present in Kelvin's,
therefore the two parts of the fundamental solution are added
up and integrals are properly computed using the same inte-
gration scheme normally employed for the Kelvin part.
In addition, computation of stress rates at boundary nodes

(S S E: non traction-free part Of f - f I) is included
E: f I and
into the implementation through simple expressions which do not
require any integration and involve the interpolated traction
and displacement rates over the boundary elements.
DIFFERENT SOLUTION TECHNIQUES
At present, the initial strain formulation has been implemented

with the von Mises criterion and employs Mendelson's successive
elastic solutions method (Mendelson, 1968). This solution
technique, also called "elastic predictor-radial corrector
method" by Schreyer et al. ( 1979), has proved to be very effic-
ient and stable with reference to the load increment size.
Therefore it is recommended for von Mises material problems.
The initial stress implementation, on the other hand, is more
general and can handle four different yield criteria (Mises,
Tresca, Mohr-Coulomb and Drucker-Prager), with two different
itera.tive routines. The first is a pure incremental tech-
nique, comparable to what was used by Zienkiewicz et al. (1969)
for finite elements. The second deals with accumulated values
of the initial stresses, in a similar fashion to the initial
strain implementation.
A common feature of the alternative implementations is that

they are all incremental-iterative processes, capable of per-
forming iterations by using a single recursive expression
relating stresses to the plastic strains and the initial elastic
solution. In what follows, it will be seen that the equivalent
relations are not so straightforward when internal displacements
are interpolated for computation of the internal stresses.
Recalling the procedure presented by Cathie (1980), one

notices that in order to implement linear interpolation functions
361
for the initial stresses, triangular cells with quadratic

interpolation for the displacements must be used. Consequently,
the number of internal points to which the cell is connected
(also called internal points by the author), is doubled to six.
By differentiating the interpolated displacements, a linear
variation of the total strains is obtained and used to calcu-
late the initial stresses at three so-called stress points.
All operations are performed in incremental form.
For comparing the number of operations during one iteration

cycle, one can consider, for simplicity, the case of a unique
cell located within the body. Following Cathie's procedure,
the computation of internal displacement increments at the six
internal points would be carried out by using a 12 x 9 matrix
(called tlz in the reference), which multiplies the increments
of initial stress at the three stress points. On top of this
the derivatives still have to be performed to obtain the
corresponding strains.
For the pure incremental procedure of the initial stress

formulation, the equivalent operation will only involve the
product of a 9 x 9 matrix by the initial stress increments,
leading directly to the increments of elastic stress. By
elastic stresses one means the result from the application
of Hooke's law to the total strains, which already saves one
operation when applying the elastoplastic constitutive rela-
tions (Telles and Brebbia, 1980b). Note that for quadrilateral
cells the same argument still applies. Furthermore, as the
number of cells increases, the difference between the two
procedures becomes even more pronounced.
In conclusion, the above demonstrates that for accurate

and efficient elastoplastic BE implementation, where the
influence of the plastic term should be interpolated at least
in linear piecewise form, computation of internal stresses by
using the proper integral equation should always be preferred.
EXAMPLES
In the present section, the results of some applications of

the technique described are compared with numerical and
analytical solutions presented in the literature. The first
example was solved by the elastoplastic implementation of the
Kelvin fundamental solution. It is here included to highlight
the improvement obtained when the integral equation is used
for computation of internal stresses. In the second and third
applications, the half-plane fundamental solution is employed
throughout and the capability of the formulation is demonstrated.
Thick cylinder
In the first example the plane strain expansion of a thick
cylinder subjected to internal pressure is studied. Ideal
plasticity under von Mises yield criterion is assumed with the
362
tI
I
I
\
b= 2a
a= 100mm
\"
--y
Figure 2. Thick cylinder problem. Boundary element and internal

cell discretization
0.8
0.6
Analytical
0.4 • • • B.E.M.
0.2
1.0 2.0 3.0

Figure 3. Outer surface displacements for thick cylinder problem
363
following material parameters,
E 12000. dN/mm 2
a 24. dN/mm 2 (uniaxial yield stress)

0
v 0.3
Boundary element results computed without boundary dis-

cretization of the symmetry axes as depicted in figure 2, are
here compared with the analytical solution produced by Hodge
and White (Prager and Hodge, 1968).
a~
10 00
+
I
I
I
08
!
plast1c elast1c
06
l
04 o/00 =0 755
Analytical !
. . . B.EM !
!
02
I '/a
10 12 14 16 18 20
Figure 4. Circumferencial stress distribution in thick cylinder. Plastic front at r' =1. 6a
Radial displacements over the outer boundary and circum-

ferential stress distribution (plastic front at r' = 1.6a)
exhibit good agreement with the analytical solution as shown
in figures 3 and 4 respectively.
It is worth mentioning that the example presented here has

been solved by Banerjee and Cathie (1980) using constant
boundary elements and linear piecewise displacements over
internal cells. For comparing results, they have employed a
coarse mesh which had the same number and pattern of cells
shown in figure 2 and a fine mesh with 192 cells (twice the
364
number of boundary elements). Although radial displacements

were in good agreement with the analytical solution, the
circumferential stresses presented for the refined mesh still
did not match the accuracy exhibited in figure 4.
Strip footing
In this example, the plane strain analysis of a flexible strip
footing under uniform loading is presented. The finite soil
stratum is discretized taking full advantage of both, symmetry
and free-surface condition, using the reduced number of 14
boundary elements and 42 internal points as shown in figure 5.
Figure 5. Strip footing on elastoplastic soil . Discretization used for B. E. results
The soil was considered to be a perfectly plastic ma terial,

obeying the associated Mohr-coulomb (M-C) c rite rion, with
E 30000 psi
c = 10 psi (cohesion)
$ 20° (angle of internal friction)
'J = 0.3
An alternative solution was also carried out using the

associated Drucker-Prager (D-P) yield criterion (extended von
Mises) given by,
3 tan$ r-- 3c
! p + ~J2 0
(9+12tan 2 ~) (9+12tan 2$)!
where p is the hydrostatic component of the stress tensor and
J 2 is the second invariant of the deviatoric stresses.
365
160
FEM ( M-C)
••• BE M (M-C)
12 0 B E.M (0 - P)
a v
80
l-
4.0
12ua/ l
004 008 012 016
Figure 6. Load- displacement curves for strip footing problem
•
.......
4 6 81 12 2 14 2 14 8
I '!1i l D D D
Figure 7 . Spread of plastic zones at different load levels.
Mohr-Coulomb yield criterion .
366
Ground surface displacements are presented in figure 6.

Also included is the equivalent M-C finite element solution
obtained by Zienkiewicz et al. (1975) using quadvatic isopara-
metric elements with 121 nodal points. The collapse loads
achieved by the boundary element and finite element techniques
(M-C) are q/c = 14.9 and q/c = 15.1 respectively, which agree
well with the Prandtl solution (Chen, 1975) q/c = 14.8. As for
the D-P results, it is seen that although the displacements
were larger, the maximum load obtained was still not far from
the previous ones.
Zones of yielding defined by the M-C solution are shown in

figure 7. These zones compare well with the reported finite
element computations.
Shallow tunnel
In a recent paper the present authors have applied the elasto-
plastic boundary element technique to solve the problem of a
deep circular excavation of radius r' in an infinite medium.
The great advantages of the technique were then pointed out when
comparing results with different finite element solutions.
Herein, an analogous problem is studied by considering the
tunnel to, be shallow, located within the semi-infinite domain
and with its centre at a depth of 5r'.
The rock-like material was assumed to follow the D-P yield

criterion presented in the last example, with the following
characteristics
E 500 ksi
c = 0.28 ksi
<P 30°
\) = 0.2
In order to produce a more realistic analysis, the semi-

infinite medium was assumed to be initially under the in-situ
linearly varying stress field given by the formulae,
av + yh (vertical stress)
0.4o (horizontal stresses)
v
where a is a uniform pressure that may be due to an overburden
of wate¥ or very weak material, y is the unit weight of the rock
and h is the distance from the ground surface.
To simulate the stress state adopted for the deep tunnel

problem (crv = I ksi) at the depth of the excavation axis, the
following values were chosen,
Ov 0. 3 ksi
y 8.9 x 10-2 ;lb/in3
r' 131 ft
367
r
----------- ----~-----
y
.
in - s1tu stress field

I
I
I
I
I
I
5 r·
I
r----
I
,~,.
2 .25r·
Figure 8. Shallow circular tunnel problem. Discretization used for B. E .

results and total spread of plastic zone
368
The plane strain analysis was carried out by applying

increments of external loads, corresponding to the relaxation
of the in-situ stresses, over the boundary of the cavity. The
discretization employed is depicted in figure 8 where the total
extent of the plastic zone is also shown.
-7.5
,,
I \ deep tunnel
shallow tunnel
-5.0
-2.5
r,
/
/ ~r'
2 3 4 5
Figure 9. Final stresses along the horizontal section through the medium
Final stresses along the horizontal section are presented

in figure 9 with the equivalent results from the deep tunnel
case (Telles and Brebbia, 1980b) included for comparison. It
should be noted that stress values outside the internal cell
region were computed at simple internal points.
The above example clearly indicates the powerfulness of

the half-plane implementation. Problems of this sort can only
be satisfactorily solved by using this technique, which requires
neither ground surface nor outer boundary discretization.
CONCLUSIONS
In this paper the complete formulation for the half-plane

boundary element technique applied to plasticity has been
presented. The correct integral equations for stresses at
internal points and stresses at the traction-free surface were
derived following the same procedures previously established by
the authors.
369
The efficiency of computing stresses at internal points

by using the proper integral equation in preference to employing
interpolated displacements over internal cells is demonstrated.
The practical applicability of the formulation is illustra-

ted by examples, which show that semi-infinite or finite sized
problems can be analysed without boundary elements over the
traction-free surface. This is, without doubt, a great advant-
age of the elastoplastic boundary element technique for solving
problems concerning the semi-plane.
REFERENCES
Banerjee, P.K. and Cathie, D.N. (1980) A Direct Formulation and

Numerical Implementation of the Boundary Element Method for
Two-Dimensional Problems of Elastoplasticity, Int. J. Mech.
Sci., 22_: 233-245.
Brebbia, C.A. and Walker, S. (1979) The Boundary Element

Techniques in Engineering. Butterworths, London.
Cathie, D.N. (1980) On the Implementation of Elastoplastic

Boundary Element Analysis, Proc. 2nd Int. Seminar on Recent
Advances in Boundary Element Methods, (edited by C.A. Brebbia),
Univ. of Southampton, U.K.: 318-334.
Chen, W.F. (1975) Limit Analysis and Soil Plasticity. Elsevier

Scientific Publishing Co., Amsterdam.
Melan, E. (1932) Der Spannungszustand der durch eine Einzelkraft

in Innern beanspruchten Halbscheibe, Z. Angew. Math. Mech., 12:
343-346.
Mendelson, A. (1968) Plasticity: Theory and Application.

MacMillan, New York and London.
Mikhlin, S.G. (1962) Singular Integral Equations. Amer. ~~th.

Soc. Trans., Series I, lQ: 84-197.
Mukherjee, S. (1977) Corrected Boundary Integral Equation in

Planar Thermoelastoplasticity, Int. J. Solids Structures, 13:
331-335.
Mukherjee, S. and Kumar, V. (1978) Numerical Analysis of Time-

Dependent Inelastic Deformation in Metallic Media Using the
Boundary Integral Equation Method, trans. ASME, J. Appl. Mech.,
45: 785-790.
Nagtegaal, J.C.; Parks, D.M. and Rice, J.R. (1974) On Numeric-

ally Accurate Finite Element Solutions in the Fully Plastic
Range, Comp. Meth. Appl. Mech. Engng., 4: 153-177.
370
Prager, W. and Hodge, P.G. (1968) Theory of Perfectly Plastic

Solids. Dover, New York.
Riccardella, P.C. (1973) An Implementation of the Boundary

Integral Technique for Planar Problems in Elasticity and Elasto-
plasticity. Report No. SM-73-10, Dept. Mech. Engng., Carnegie
Mellon Univ., Pittsburgh.
Schreyer, H.L.; Kulak, R.F. and Kramer, J.M. (1979) Accurate

Numerical Solutions for Elastic-Plastic Problems, trans.
ASME, J. Pressure V:essel Technology, .!..Q_~: 226-234.
Telles, J.C.F. and Brebbia, C.A. (1979) On the Application

of the Boundary Element Method to Plasticity, Appl. Math.
Modelling, l= 466-470.
Telles, J.C.F. and Brebbia, C.A. (1980a) The Boundary Element

Method in Plasticity, Proc. 2nd Int. Seminar on Recent Advances
in Boundary Element Methods, (edited by C.A. Brebbia), Univ.
of Southampton, U.K.: 295-317.
Telles, J.C.F. and Brebbia, C.A. (1980b) Elastoplastic Boundary

Element Analysis, Proc. U.S.-Europe Workshop on Nonlinear
Finite Element Analysis in Structural Mechanics, Ruhr-University
Bochum, W. Germany.
Telles, J.C.F. and Brebbia, C.A. (1981) Boundary Element

Solution for Half-Plane Problems, Int. J. Solids Structures,
(in press).
Zienkiewicz, O.C.; Valliappan, S. and King, I.P. (1969) Elasto-
plastic Solutions of Engineering Problems, Initial Stress
Finite Element Approach, Int. J. Num. Meth. Engng.,~: 75-100.
Zienkiewicz, O.C.: Humpheson, C. and Lewis, R.W. (1975)

Associated and Non-Associated Visco-Plasticity and Plasticity
in Soil Mechanics, G€otechnique, 25: 671-689.
371
THE BOUNDARY ELEMENT METHOD FOR THE SOLUTION OF

NO- TENSION MATERIALS
W.S. Venturini & C. Brebbia
University of Southampton, England
I. INTRODUCTION
The Boundary Element Method has developed considerably since

its beginning in 1967 for plane elastostatics 17 The technique
has been applied to a large number of problems in engineering
such as plasticity, elastodynamics, viscoelasticity, and others.
The purpose of the present paper is to extend the Boundary

Element Method to no-tension materials such as those present
in underground and surface excavation and discussed in ref.
(II). Unlike Finite Elements, the Boundary Element Method does
not require the imposition of artificial outer boundaries at a
finite distance, since the technique automatically satisfies
the boundary conditions at infinity and prescribed displacements
are not required.
In this paper the boundary of the underground tunnel or

surface excavation is divided into a number of linear elements
over which the displacements and tractions are considered to
vary linearly from node to node. The applied loads are computed
by considering the residual stresses produced by removing the
excavated material.
The no-tension solution is achieved using an iterative

process which consists of applying at each step a series of
initial stresses to compensate the tensile stresses. The
initial stress components are approximated over the domain by
linear interpolation functions.
In the following, the boundary integral equations are

presented for elastostatics and they are extended to cover no-
tension materials using an initial stress formulation. In
addition, the way in which prestressing forces can be applied
is discussed.
372
In Section 3, the necessary steps to formulate the matrix

equations are presented and following that, the iterative tech-
nique required to obtain the no-tension solution is also shown.
Finally, some problems are solved to illustrate the

application of boundary elements for no-tension materials.
The examples were selected to compare the boundary elements
solution with already published finite element, results and
point out the advantages of using the former.
2. BOUNDARY INTEGRAL EQUATIONS FOR ELASTOSTATICS
2.1 Governing Equations

An isotropic linear elastic problem is completely defined by
the following set of governing equations.
i) Equilibrium equations
a'· . . + b.
l.J. J l.
= 0 in n (2. I)
where o .. and b. are the stress and body forces components and
n is th~Jdomain~
ii) Displacement-strain relationships
£ ••
l.J
= -2I (ul.,J
.. + u .. )
J,l.
in n (2.2)
where ui,j and Eij are the displacement derivatives and strain
components respectively.
iii) Constitutive equations, in this case they are the Hooke's

law
2Gv o
0
ij ---
1- 2 \) Ekk ol.J
.. + 2 G £ ••
l.J
-
0
••
l.J
(2. 3)
where G and v are the shear modulus and Poisson's ratio resp-
ectively and oP. ar'e the initial stress componentslO
l.J
iv) Boundary conditions
u. (Q) ;;, (Q) Q £: rl

l. l.
(2.4)
pi (Q) pi (Q) QE. r2
r = rl + r2 is the whole boundary of the structure; ui and pi

are prescribed components of the displacements and tractions
on r.
373
2.2 Boundary integral equations

Using Betti's theorem 2 • 3 or the weighted residual technique, 1
displacements at any point in the domain n can be expressed
in terms of boundary values ~ and pk, and body forces bk, i.e.
J p~k ~ dr + J u~k pk dr + J u~k bk dn (2.5)

r r n
The above equation can be extended to the case of having

an initial stress field as follows.
* * * bk
ui
J pik
r
~ dr +
J uik
r
Pk dr +
J
n
uik dn
+ * crmk dn
Eimk
0
(2.6)
J
n
* uik
pik' * and E*~mk a~e Kelvin's fundamental solutions 4 for
d~splacements, ~ract~ons and strain respectively, and their
values are given by:
[r , ~ r , k- (3-4v) ln r o~k ]/8rrG(I-v)
(2. 7)
~l-2v)[r,k6~m + r,mo~k] + omkr.~- 2r,ir,m:r,k]

8rrG (1-v) r
where 6 .. is the Kronecker delta; r . are the derivatives of r

~J .~
with respect to Xi (which is a coordinate of field point); and
ni is a direction cosine (fig. I )
The expression (2.6) gives the displacements for an interior

point. If the point is considered to be on the boundary the
expression becomes,
(,.)
~2 .....
-1:-
'2 '2
iiK = 6ii
axK
x2
r.K = k = x~-x~
GxK -,-
,,
(PI v,:c;,v ~,
'2
'1
Fig. 2 Triangular coordinates
P (load pomt)
,,
Fig.1 Geometric parameters. Fundamental solutions
a
Ol / /
// LINEARLY ELASTIC / ,;:_.LINEARLY ELASTIC CURVE
/
/
/ ///
/
// /LOADING CURVE ...,
e: (compress1on ) £ ( tractaon ) £ {compression) E ( tract•on)
'-UNLOADING CURVE 6E --UNLOADING CURVE
Fig. 3 No-tension as an elastic material Fig. 4 No-tension as a plastic material

375
+ J (2. 8)
n
the value cik has been determined by different authors 1 • 2 •
The value of stress at an internal point can be obtained

by differentiating the equation (2.6) and applying the Hooke's
law , this gives
r 0
(J • •
~J I
r
s.~J'k ~ dr
J
r
D.~J'k pk dr +
J Dijk bk
n
drl +
J
n
Fijmk o mk drl
(
0 0
- (Jmk ~mk'. dr 1 - CJ •• (2.9)
J ~J l.J
rl
where S .. k and D.. k are presented below; F. 'mk and Jr

~J ~J ~J
~mk·.
~J
dr 1
are given in the Appendix A. r1
{(1-2v) [ok. r . + ok.r . - o .. r k] + 2r .r .r k}

~ 'J J '~ ~J ' '~ 'J '
D.~J'k
47T(I-v)r
3r r.
{ 2-;:;-
-
L(I-Zv)o .. r k + v(o.kr . + o.kr .) - 4r.. r .r k]
s.~J'k on ~.J ' ~ ,J J .~ .~ ,J •
+ 2v(n.r .r k + n.r .r k) + (J-2v)(2nkr .r .

~ ,J ' J .~ ' .~ ,J
(2. 10)
The stress values at boundary points can be calculated from

surface tractions and displacement derivatives as in Appendix
B.
376
Prestressing forces can be introduced as body, line or

boundary forces. For the last case we effectively increase the
boundary of the domain to be analysed. For the other two cases,
the force can be treated as a concentrated load. Notice that at
the load point a singularity will arise.
Considering prestressing forces as concentrated loads the

equation 2.5 can be written as follows,
(2. II)
Notice that when this equation is applied on the boundary

the term ctk must multiply the d~splacement ut.
For stress determination, we also have to add the pre-

stressing term D.. k P in the equation (2.9).
~J vk
3. BOUNDARY ELEMENT METHOD
In the last section we have presented the internal equations for

displacements and and stresses in two-dimensional elasticity.
Now as was presented in ref. (I ), we are going to transform
them into a set of algebraic equations by discretizing the
boundary into elements and expressing the values of displace-
ments and tractions in terms of interpolation functions and
nodal values.
In this case, linear interpolation functions (eq. 3.1) for

displacements and tractions on the boundary are considered. They
can be written in matrix form 6
u = 'PT ua.
on r
p 'PT pa.
b ljJT ba. in Q (3. I)
and
ao ljJT 0 oa. in Q
where Ua., Pa., ba. and ooa. are the nodal values for displacements,
tracti~ns,~body forces and initial stresses respectively. ;pT and
pT are interpolation functions such as those used in finite
elements 7 •
Notice that we need to consider Ua. and Pa. only on the boundary
in order to solve the integral equations, while Ba. and cr 0 a. have
to be defined over the whole domain.
377
The integral equation can be written in discretized form

by expressing each integral as a summation of element and c~ll
integrals, thus one has,
NE NE
(J
)
uc + I p * cpT dr. ju I u* cpT dr.) pa.
j=l J - J - j=l - J
r. r.
J J
NCELL
+ N~ELL( J u* lj!T dO.)ba.+ I [J e: * lj!T dQ. )o oa.
J=l J - j=l - J -
0. n.
J J
NP
+ I u* p (3. 2)
~v.
j=l J
The above integrations over elements and cells can be

carried out and the final results written in matrix form as
follows,
cu + h u G P + D b + E a· 0 + V P (3. 3)
- _u
The integrals over each element can be done analytically

or numerically. Only over elements close to the singular
point, the integral which gives H terms must be done analytic-
ally.
The domain integral (integral of body forces and initial

stress) can be carried out numerically. In what follows the
procedure to obtain the initial stress integral is discussed.
The initial stress integral over any single cell j can

be written as
I (3.4)
n.
J
For triangular cells and linear interpolation the follow-
ing function can be used
(3. 5)
with
378
( 3.6 )
where A is the area of cell and

k - xj
a. XI
l I
k
xj - x2
b. (3.7)
l 2
with i = 1,2, j = 2,3 and k = 3,1
This interpolation function can also be written in a

polar coordinate (fig. 2) i.e.
sa = sa(p) + J_(b cos6 + a sin6)r (3.8)

2A a a
where p is the load point and
(3.9 )
The fundamental solution *

£~mk can be written as,
* I
E~mk(p,S,r) = e~mk(p,S),r (3.10)
and each cell integral becomes,
I I e~mk(p,a)[sa (p) + 2~ (ba cosa + aa sin6)rjdrd6 (3 .II)
n
This can be integrated over r remaining only to carry out
the integral over the angle a. The a integral is performed
numerically using a four or six Gauss integration rule. Once
these integrals are computed over cells one obtains the matrix
E.
The expression (3.12), after integration over r, becomes
I Ia e~mk(p,a) Ga(p)(R2 -R 1) + 21A(ba cos6 + aa sinS)
(3.12)
379
R1 and R2 are given in function of 8 and cell geometry and can

be easily calculated using expression (3.9) over an appropriate
side of the cell.
The final system of matrix equations is obtained by inter-

changing columns in the matrices shown in eq.(3.4)in accordance
with the applied boundary conditions 1 • So the unknowns are
written in a vector X on the left hand side and all known terms
are accumulated on the right hand side.
After grouping together the right hand side matrices, one

can write,
AX (3. 13)
where T contains the effects of all prescribed loads and dis-

placements, the initial stress term has not been added to it
as it will be needed during the iterative process.
Solving equation (3.14) one obtains,
X M + K a0 (3. 14)
where
(3. 15)
The stresses of internal points can also be expressed in

matrix form after integration, i.e.
a = D P - S U + D + WP
v
+ F a0 - o0 (3.16)
We can rearrange the matrices in function of the prescribed

values U and P, i.e. all the unknown values are in~ and the
values prescrtbed on the boundary are given by Y. After
rearranging the appropriate matrices in the above equation,
one obtains,
a D y SX+Db-WP+Fa (3. 17)

- -v
with
F =F - I (3.18)
In a more: compact form

380
a = T - S X+ F a (3.19)
- _o
substituting X value one obtains
a = N + B cr0 (3.20)
where
N T - S M
(3.21)
B F S K
Stresses on the boundary are calculated in the same way

and the matrices N and B can be obtained as in Appendix B.
4. ANALYSIS OF NO-TENSION MATERIALS
In many practical applications, tunnels can be built only by

excavation, and no structural lining is required to guarantee
the cavity stability. Careful analysis of the rock medium is
required in this case. Considering the medium as a linear
elastic material is generally unsafe, as the tensile strength
of rock is usually small and numerous faults and joints are
to be expected.
In the present analysis, the rock will be considered as a

no -tension material. The solution will be iterated until
structural equilibrium is reached with only compressive
stresses. This stage of stress is usually possible for under-
ground structures, and the limit load is reached when the mass
removal is completed.
Problems with no-tension material can be solved in one of

the two following ways, i) by considering the material can not
withstand any traction as shown in fig. 3; loading and unload-
ing occur along a straight line, i.e. its behaviour is assumed
to be elastic. ii) by assuming that the material behaves as a
plastic one, when in tension this will produce a different un-
loading path as shown in fig. 4.
The second case is a special case of plasticity, and can

be solved using load increment small enough to achieve the
required accuracy.
The numerical procedure adopted here to reach the no-

tension solution is similar for the two cases described above.
The load can be divided into a series of increments for the
plastic analysis or applied in one step for the first case.
The procedure consists of the following steps:

381
i) Elastic stress and shown determination. The elastic stress

due to a load increment or an initial stress field can be
calculated as
where a Q corresponds to the elastic response for the load

increme~t. The total strain E can be obtained using the
elastic stress values. -e
ii) The principal directions can then be determined using the

strain values; and the stresses in these directions can be
computed. Letting the tensile stresses be zero in those
directions, the true stresses (~t) can be determined.
iii) Determination of the initial stresses for the next step

which can be given by
where ~to is the previous value of the ':t.
iv) If the values of tensile stresses are not equal to zero

within the required accuracy, the first three steps must be
repeated. When convergence has been achieved another load
increment can be applied.
5. APPLICATIONS
In this section, some practical examples, assuming no-

tension material, i.e. that material cannot withstand any
tensile stress, are presented.
Whenever possible results are compared with published

finite elements solutions.
Steep Valley
The case of a steep valley was solved by Valliappan 11 ,
using finite elements and no-tension material. The load applied
to this structure is only due to the removal of residual com-
pressive stresses which are taken to be equal to -yy and -yK 0 y
in vertical and horizontal directions respectively (fig. 5).
The sloped valley boundary is discretized into a series

of straight elements over which tractions corresponding to the
residual stresses are applied. In figure 6 the mesh used to
solve the problem is presented. The convergence of the results
was tested by using a finer mesh which produced practically the
same values. Notice that for this problem the domain tends
to infinity and no displacement boundary conditions have to be
prescribed.
382
-,
1
40' PARABOLIC CURVE
y =1501b/tt 3
G =1061b/ft2
v = 0 .3
Kill =0.2
Fig. 5 Valley to be excavated
40 5 10 12 18 15 15 35 50 100 200 250 250
~--------------------~~f-f~r-i~~--~
-
I
BOUNDARY NODES
'-BOUNDARY ELEMENTS
Fig. 6 Steep valley. Discretization

y
y
CAl
Fig. 7 Steep valley. Elastic solution Fig. 8 Steep valley. No-tension solution ~
384
Figure 7 shows the elastic initial solution and figure 8

the final no-tension results in function of the principal
stresses. The results agree well with a published finite element
solution 11 , for which a 247 elements and 144 nodes mesh was used.
The boundary element solution only required 24 nodes and 23
boundary elements. For integrating the initial stress term
29 internal cells were used.
This example was solved using the two no-tension solution

criteria previously discussed (see figs 3 and 4). It was found
during the solution that the discretization of the curved
boundary into straight elements produces small tractions forces
which cannot be completely eliminated. This problem occurs with
finite elemenmas well, but in both methods, it can be solved
using curved elements.
Semi-circular tunnel
The semi-circular tunnel, fig. 9 is a comparison example pre-
viously solved by Bath 12 , 13 using an isoparametric quadratic
finite elements program. The loading and the boundary conditions
were chosen to make this example an optimum numerical test
rather than try to reproduce an actual tunnel.
The load consists of a large vertical load P which is

applied to the ground surface. Residual stresses are taken as
constant during the loading process and are calculated using a
specific weight equal to 120 lb/ft 3 and a maximum height equal
to 116 ft.
The results for no-tension can be seen in figure 10. The

no-tension zones are plotted for load corresponding to 28.87.,
40% and 100% of the total load respectively.
Figure II presents the displacements for points A and B

showing their non-linear behaviour. The non-linearities started
at 4% of loading and continued up to 607. of it. From then to
the final load the structure behaves again linearly, once the
final no-tension zone has already been reached.
Deep tunnel
The tunnel example shown in fig. II was used for a hydro-
electric power station and taken from Valliappan 11 • He analysed
it on a no-tension problem using 500 constant elements and 276
nodes mesh with a finite element program. The fig. 12 and fig.
13 show the discretization of this tunnel for the boundary
element solution, both with and without prestressing forces.
A 48 nodes and 39 elements boundary mesh was used for both cases,
the internal cells for initial stress integration are plotted
in the same figures. For the first case the only applied forces
are those corresponding to the removal of the material. Fig. 14
shows the initial elastic and final no-tension solutions. The
results are very similar to those presented in (II) using the
Finite Element Method.
385
r
Ov=H .Y
------ ----
I• -Ou =H .Y
H = 116 - y
Fig. 9 Semi -circular tunnel. Residual stresses,

Loading, Boundary discretization and
internal cells .
I
I
I
""-V""
NO- TENSION ZONES CORRESPONDING

SOME PERCENTAGE OF LOAD
~ 28.8 %
~ 40 . 0 %
• 100%
Fig. 10 No-tension zones

ground surface (.o)
I 8l
*!:m;i~=~=!®!:}.t::!m=~~~=a.W:;mm~mmtW:=w=w.l=~=!~J=i~:Ifl:@imi:;:~ir:l*?:m!:r::!:ill!:~:!*?.i~~!m;®.;;;S}[~
u(lt) I v (It)
y
~ E
,f" <;"
N
#',f"v 6 (Venical displacement at B) II
0 .08 + 0 .4 I
0.06 + 0 .3 u A Horilontal
( displacement at A )
Y.~~~i~~; ;~~~§~;q
:/~... .... : .... :
0 04 + 0 .2 ...... ----
\- ··t
------
-------
-------
0 .02 . ...... ------- X
~~:,....-- ..........
6.o 1 i~~9{¥~:~{;~~:·.~·,.
4% 40% 64 % 100% P(LOAD) . . ~.··r···•·-,..·--.>1'·- ~---'~>
,' \\./,r\ / ' \ ,/ \, ./ \, / '.\:
,/•'! ,'/: ,.,/' 1,',' : ,,•'"' :
t·::-:--'i-.JO.- -+ --JI:,--~- -*·-·---If·--'' :
/~;:~::t~~:~:~~~t!"~,)t.:>
Fig . 11 Displacements at the structure Fig . 12 Boundary discretization and internal cells
grourld surface
~{U<:i?''imii·-·-·
l
·· ,._ RWti¥1U!llh:""'i ·· 'f
E
N
•
:1:
SCALE : 20kg /cm 2 lor lcm

SCALE FOR COMPRESSION STRESSES : 2em ~ ZOkg / cm2
LEGEND : ···--------- ELASTIC TENSION ZONES
LEGEND : ELASTIC TENSION ZONES
NO - TENSION ZONES
NO TENSION ZONES
....______, COMPRESSION
~~~±:;]~~~~~~~ ' FISSURES DIRECTION
-- COMPRESSION
FISSURES DIRECTION
Fig.13 Boundary discretization . internal cells and

prestressing forces at the deep tunnel
Fig. 14 Elastic and No-tension results . Tunnel
-
Fig. 15 Prestressed tunnel. Elastic
without prestressing forces and No-tension results
~
388
Notice the large no-tension zone present over the roof

which requires some additional structural component or pre-
stressing to avoid the rock ralling down. Prestressing using
rock bolts has recently been extensively used in tunnels in
place of lining and this technique has been adopted for the
present example.
The boundary discretization for the analysis of the case

with prestressing is the same as it was shown in fig. I3, but
the internal cell configuration has been changed to take into
account the expected new tension free zones. The prestressing
forces shown in the same figure are applied in 20 boundary
points and 20 internal points which are distributed along a
circular path.
The results obtained with the Boundary Element Method are

shown in fig. IS, where the initial elastic zone and the final
no-tension zone are plotted. As can be seen, the tension free
zones in the roof have been considerably reduced due to pre-
stressing effects. These results are in close agreement with
the finite element solution presented in (II).
6. CONCLUSION
In this paper the Boundary Element Method is applied, for the

first time, to solve no-tension material problems in tunnels
and excavations.
The technique is of considerable advantage for this type

of problem as only the boundary of the domain has to be dis-
cretized, except where the boundary goes to infinity, for which
no discretization is required. Pinite elements, on the other
hand, require the definition of an artificial outer boundary
where displacements must be prescribed, in addition to the dis-
cretization of all the domain. The artificial outer boundary
is by defininition an approximation, and is a source of
inaccuracy which is compounded by the discretization of the
domain. Besides, these approximations, the finite elements
method requires considerable increase in the amount of data
needed to run the problem.
The iterative process used to obtain the no-tension solu-

tion proved to be very stable and has a fast convergence.
Notice that during the whole process the same system of equations
is used and the initial stresses only modify the right hand
side vector. Because of the way the right hand side vector is
formed, the effect of cell discretization is reduced and the
nllffierical solutions are only slightly affected by the coarse-
ness of the mesh.
This paper also presents, for the first time, the way in
which concentrated prestressing forces can be introduced in the
Boundary Element Method.
389
I. REFERENCES
I. BREBBIA, C.A. "The Boundary Element Method for Engineers"

Pentech Press, 1978.
2. CRUSE, T.A. ·~athematical Foundations of the Boundary

Integral Equation Method in Solid Mechanics" Air Force
Office of Scientific Research, Report No. AFOSR-7R-1002,
1977.
3. CRUSE, T.A. "Boundary Integral Equation Method in Solid

Mechanics", Report SM-73-17, 1973.
4. LOVE, A.E.H. Treatise on the Mathematical Theory of

Elasticity, Dover, 1944.
5. TELLES, J.C.F. and BREBBIA, C.A. "On the Application of

the Boundary Element Method to Plasticity" Research Note.
Applied Mathematical Modelling, Vol. 3, 1979.
6. TELLES, J.C.F. and BREBBIA, C.A. "The Boundary Element

Method in Plasticity" Second International Seminar on
Recent Advances in Boundary Element Methods, University of
Southampton, March, 1980.
7. BREBBIA, C.A. and CONNOR, J.J. "Fundamentals of Finite

Element Techniques for Structural Engineers" Butterworths,
(1973).
8. LACHAT, J.O. and WATSON, J.O. "Progress in the use of

Boundary Integral Equations illustrated by Examples",
Computer method in applied mechanics and engineering 10
(1972) 273-289.
9. ZIENKIEWICZ, O.C., VALLIAPPAN, S. and KING, I.P. "Stress

Analysis of Rock as a No-tension Material", Geotechnique,
Vol. 18, 56-66, 1968
10. ZIENKIEWICZ, O.C. and VALLIAPPAN, S. "Analysis of Real

Structures for Creep, Plasticity and other complex
Constitutive Laws" Conference on Materials in_
Civil Eng., University of Southampton, 1969.
II. VALLIAPPAN, S. "Non-linear Stress Analysis of Two

Dimensional Problems with Special Reference to Rock and
Soil Mechanics" Thesis submitted to the University of
Wales, March 1968, Swansea.
12. BATHE, K.J. "An Assessment of Current Finite Element

Analysis of Non-linear Problems in Solid Mechanics"
Symposium on '~umerical solution of differential equations
III" University of Maryland, 1975.
390
13. BATHE, K.J., OZPEMIR, H. and WILSON, E.L. "Static and

Dynamic Geometric and Material Non-linear Analysis"
Report No. UCSESM74-4, Department of Civil Engineering,
University of California, Berkeley, February 1974.
14. RABCEWICZ, L.V. "The New Austrian Tunnelling Method"

Water Power, No.v, Dec. 1964 and January 1965.
IS. RICCARDELLA, P.C. "An Implementation of the Boundary

Integral Technique for Planar Problems of Elasticity and
Elastoplasticity" Ph.D. Thesis presented to Carnegie-
Mellon University 1973.
16. MIKHLIN, S.G. "The Problem of the Minimum of a Quadratic

Functional" Translated by A. Feinstein, Holden-day Inc.
( 1965) •
17. F .J. RIZZO "An Integral Equation Approach to Boundary

Value Problems of Classical Elastostatics" Quart. Appl.
Math. 25, 83-95 (1967).
391
APPENDIX A SINGULAR INTEGRAL IN THE OOMAIN
The initial stress integrals involve multiplication of initial

stress components by the strain fundamental solution. Special
care has to be taken when calculating these integrals due to
the singular characteristics of some of their terms.
Considering equation (2.6) with only initial stress term,

i.e.
r *
u. (p) 0
Eimk(p,q)omk(q) dQ(q) (A. I}
~
J
stresses are related to displacement derivatives as follows
(au. (p)
)
2Gv auJI. (p} au. (p)
a •• (p) = - - cS
~J ax (p + G
1-2v ij ai('p') a:tri;T
p
+ J
ax. (p) (A.2)
J J
Notice that the initial stresses are being interpolated on the cell,
one can write
(A.3)
As shown in equation (2.5) the fundamental solution for

*
Eimk(p,q) can be written as
(A.4)
Hence any polynomial interpolation function ~(r,e) can be

written as
a a a a n
~ =~(e)+ ~ 1 (e)r
o
+ ••• ~
n
(e).r (A.5)
For the case of linear interpolation functions such as those

shown in equations (3.9) one has
,,,a
"' = .,ra = .,ra( p ) + 2A
I (ba cos e + aa sin8)r (A.6)
Substituting these relationships into A.3 and using Hooke's law,

derivatives of the singular integrals arise. The singularity
is due to the first term in equation A.6 which will now be
considered in more detail by analysing the following integral,
392
(A. 7)
One can treat this derivatives in two different ways. The

first way consists of performing the integral and then the
derivatives Assuming triangular cells as shown in fig. A.l
and considering, for instance, the point pat corner I, one
has
(A. B)
and the derivatives become
D = ax~(p) J fimk(p,e) dr d6 (A.9)

J
Integrating over r one can write
(A. I 0)
with
-2A (A.II)
R(6)
b 1 cose + a 1 sine
Substituting the equation A;ll and the value of f. (p,e)

into (A.3) one can carry out the integral and the~mfind the
derivatives. The derivatives on the linear term imply no
singularities and their values can be computed numerically
performing the integration after derivation.
The second way to carry out the value of (A.7) is using

the treatment due to Mikhlin 16 , in which the derivatives can be
introduced into the integral and another term is added, so that
the equation is written as
2n
r
l
a rfimk(p,e)) a a
D =J ax.(p) r )~ (p)<ill- ~ (p) J f.mk(p,6)cos6.d6
1. J
n J 0
(A. 12)
where cosej is cos6 or cos(e- f) for j = I or j =2 respectively.
In this way the derivatives of all terms can be introduced

into the in,tegral and the whole expression for initial stress
can be written as follows,
393
( ) = zcv 6 [ Jr a
0 ij p I-2v ij ax~(p)
rl
r a
G[ J ax. (p) (Eimk(p,q))ljl
* a.
drl(q) +
J ax.a(p) *
(Ejmk(p,q))
.ll J 1
,,,a. drl (q)-lcrmkoa. - 2G':._ o ..

"' ~ I-2v 1J
1-~a.,(.p )
L 2JTT f
~mk p'
( 8) e -J oa.
cos ~ crmk
0
21I
- G ~a.(p) { J ~imk(p,e)cosej + fjmk(p,S)cosei]de}cr:

0
(A. 13)
After carrying out the necessary steps and performing the last
three integrals the whole expression for the stresses (2.9) can
be written as,
r
cr .. (p)
1J f .. Pk dr
Dk 1J - f 8kij ~ dr +
J
.. b k drl +
Dk 1J
r r n
f
n
0 I
F.1J"mk crmk drl + 8(1-v) [<6-Sv)cr~.
1J - (1-v)cr~k oij J
0
- cr .. (A. 14)
1J
where the term F .. mk was achieved after performing all deriva-
tives of t::*(p,q) 1J indicated in (A.l3). Hence, one can write
F .. mk {z(I-2v) [ok.o. + ok.o . - omko .. + 2o .. r r l

1J STT(I-v)r2 J 1m 1 mJ 1J 1J ,m ,m-1
+ 4[o kr kr . + v(o
m , ,J nk r , k r , 1. + o.kr
J , 1.r , m + o.1mr , kr , J.
+ o. kr r . )] - 16r . r . r r k } (A. 15)

1 ,m ,J ,1 , J ,m ,
Notice that the integral of F. "mk has to be performed in

the sense of Cauchy principle and1J the correct value can be
reached only after all the cells adjacent to the point p have
been integrated (see ref. 6).
394
The value obtained in equation (A.I4) after carrying out

the last three integrals at (A.I3) can also be found by inte-
grating the expression (A.9) before any derivation.
If one removes the singularities and includes the initial

stress value, the integral can be written as,
I
ax.(p)
[ f [f.mk(p,a))
r 1
o
omk(q)dn(p) +
1
J <n-n e; )
+ f [ lmkr(p,6))
f.
o~(q)dn(q) (A. IS)
ne;
where n e; is a circular area whose centre is the singular point.
The derivatives can be passed inside the first integral

and one can write
I
f
n-n e;
+ ax .a(p)
J n
f (A. 16)
or
I = 1 Fijmk(p,q)o~(q)dn(q)
n-n e;
The limit when ne; tends to zero can be carried out at the
first integral taking into account that it has to be performed
in the sense of Cauchy principle value.
Considering the density o~(q) satisfies a Lipschitz con-

dition with positive exponent, in the second term of (A.I7)
when ne; + 0, one can write,
I I Fijmk(p,q)o~(q)dn(q) + o (p)
0 mk ax.a(p) I *
e:imk(p,q)dn(q)
n J n
e;
Now the integrals for all components of e;.* k must be carried
out, then the derivatives and the limit when lm n e; + 0
395
can be performed. After these steps one can find the same
constant values as those given in (A.I4).
APPENDIX B CALCULATION OF STRESSES ON THE BOUNDARY
The simplest way to calculate the stresses on the boundary is

by differentiating the displacements.
The displacement and traction at any boundary element can

be written in a local or global system of coordinates (fig.
B.I).
In the local system, fig. (B.I) the tractions and displace-

ments are,
p AP (B .I)
and
u AU (B.2)
where
A
[~:: -cosSl
sinS
(B.3)
On the boundary one of the relationships between stresses and

tractions can be written as,
0 11 PI PI sinS - P2 cosS
(B.4)
0 12 P2 PI cosS - P2 sinS
The third stress can be obtained by using expressions for

strains deduced from the boundary displacements, i.e •
..!..
.R,
(;:;2 - ;:il)
2 2 (B.5)
{ cosS [uf - utJ + sinS [u~ - u~J} (B.6)
Using Hooke's law and including initial stresses, one has

396
(B. 7)
Once the value of cr 0 is known in the global system of coordin-

ates, one has to transform these values to the local system to
a
obtain cr0 , then obtain and finally transform back to the
global system. -
After all these transformations have been performed one is

\able' to assemble the matrices, D, S and F using the p, u and
cr 0 nodal value coefficients respectively:
x,
x,
Fig. B ( 1a) Traction and displacement according
to global system of coordinates
x,
x,
Fig. B ( 1 b) Traction and displacement according
to local system of coordinates
Section V
Numerical Techniques and
Mathematical Principles
399
SDr4E THEORETICAL ASPECTS OF BOUNDARY INTEGRAL EQUATIONS
~1. A. Jaswon ,Department of Mathematics,

The City University, London
1. Introduction
Harmonic functions in a domain may be represented as simple-
layer or double-layer potentials, generated by hypothetical
source density distributions on the boundary which satisfy
Fredholm integral equations. An alternative formulation is
via Green's formula, in which the boundary data themselves
function as source density distributions, so yielding
Fredholm integral equations satisfied by the boundary data.
The two approaches are theoretically equivalent, and we show
how the second provides some insight into the first.
Vector potentials analogous to scalar simple-layer and
double-layer potentials were first introduced by Kupradze.
These provide a representation of linear elastostatic
displacement fields analogous to the representation of
harmonic functions by scalar potentials. In consequence,
the fundamental boundary-value problems of linear elasto-
statics may be formulated by vector integral equations analogous
to the Fredholm integral equations. From this point of
view Somigliana's formula is the exact. vector analogue
of Green's formula, and it yields vector integral equations
analogous to those obtained on the basis of Green's
formula. We show how the scalar and vector theories may
be unified by means of a suitable symbolism.
2. Scalar potential theory
A continuous distribution of simpTe sources extending over
a closed Liapunov surface aB, of surface density a(g) at
g e: aB, generates the simple-layer potential
v(.E_) f g (.e_,_g) d.9.) dq (1)
aB
400
-1
at any point£ in space, where g(p,q) = 1£-ql and dq
denotes the area element at g. It Ts sometimes convenient
to refer to g as the source point and Q as t~e f~eld poi~t.
This potential is continuous everywhere and 1s d1fferent1able
everywhere except at aB. Also it satisfies Laplace's
equation everywhere except at aB, and it is therefore a
harmonic function everywhere except at aB. Its normal
derivative at£ £ aB is given by
~~ (p) = I g' (£,g) 0 {_g_) dq - 2110 (£) £ € aB ( 2)

aB
where g'(£,q) denotes the normal derivative of gat£
keeping .9_ fixed. However, we preferably write
av (
gi (£,g) 0(g) dq - 211cr{£) ; ££ aB {3)
ani J
aB
av
ane
aB
I
g~{£,g)0(g)dq - 2110(£) ; £€ aB (4)
where gi{£,g) , g'{£,.9_} respectively denote the interior

and exterior normal derivatives of g at p. These are
connected by -
gi(£.g} + g~{£,g) = 0' (5)

from which follows the useful relation
{6}
These properties of V may be exploited to provide a

simple, but efficient, formulation of the classical boundary-
value problems of harmonic function theory. First, the
Dirichlet problem requires us to construct a harmonic
function $ inside or outside aB given its values (assumed
continuous) on aB. A straight-forward procedure is to write
$(£} I
aB
g(£.g)o(g)dq p £B.1
-
(7)
$(£} J g(£,.9_)0(g}dq (8}

aB
where 0 appears as a hypothetical simple source density
on aB, B. denotes the interior domain bounded by aB and Be
denotes 1 the infinite exterior domain bounded internally
by aB. Now both the right-hand side and left-hand side
of (7}, {8) are continuous at aB, and therefore each of
401
them yields the boundary relation
$ (£.) J g(_e_,g)cr(g)dq ; E. E aB (9)

aB
which is a Fredholm integral equation of the first kind
for cr in terms of $ on aB. This has a unique solution,
which may be utilised to generate$ in Bi' or in Be subject
to the regular behaviour
$ = I.E.I- 1 f cr(g)dq + o I.E.I- 2 as I.E. I +"' • {lo)
aB
The most important example of the Dirichlet problem is
$ = 1 on aB, which defines the capacitance problem of
electrostat ics. In this case we replace (7) by
= J g(£_,.9_) A(.9_) dq £. E aB , ( 11)
aB
where A> 0 and J A(.9_)dq measures the electrostat ic
aB
capacitance of aB.
The Neumann problem requires us to construct a harmonic
function inside or outside aB given its normal derivatives
(assumed continuous) on aB. As before we write $ in the
form (7) or {8), from which there follows
$i (£.) J gi {£_,.9_)cr(.9_)dq - 27fcr(£.) .E_E aB (12)
aB
J g~(_e_,g_)a(g_)dq - 2na(£.) E. E aB , (13)

aB
where $i or $~ is given on aB. These are Fredholm
integral equations of the second kind for a in terms
of $~1 or $'e , and equation (13) has a unique solution
which generates a regular $ in Be . Equation (12) only
has a solution if
J $i(£)dp = 0 ' ( 14)

aB
i.e. the Gauss condition on harmonic functions, in which
case
a= a + kA ( 15)
0
where $e is a particular solution and k is an arbitrary

402
constant. This non-uniqueness stems essentially from the

fact that$= l on aB implies$= l in Bi, which implies
$e = 0 on aB, which implies that A in {10) satisfies the
homogeneous equation
0 =
aB
I
gi(~.~)A(g_)dq - 2'ITA(£) ; E_£ aB . (16)
Accordingly the capacitance problem may also be formulated

as a Neumann problem, since the non-trivial solution of (16)
may be normalised to satisfy (ll) and then utilised to compute
I A(~)dq
aB
Fredholm integral equations of the first kind generally

present considerable theoretical and numerical difficulties,
though equation {10) has been successfully exploited to
compute the capacitance of various closed surfaces. A
classically preferred formulation of the Dirichlet problem
is to represent$ as a double-layer potential, generated
by a hypothetical double source density on aB, i.e. we write
$(£) =
I
aB
g(E_,~)i )J (~) dq E.£ Bi ( 17)
ls
$ (£) = g(E_.~)i )J {~) dq ; E.£ Be . (18)
where \l is the hypothetical double source density and

g(E_,~)i denotes the inward normal derivative of g(E_,~) at~
keeping£ fixed. Here g(E_,~)~ functions as the dipole
potential at .2_ generated by a 1 unit dipole source at~· this
interpretation being implied by the fact that the normal
derivative operation moves with the source point q , by
contrast with gj(E_,~) in {12) where the normal derivative
operation is taRen at the field point p. It would of
course be perfectly acceptable to replace g{E_,~)! by g(E_,~)'
in (17) and (18). Now the right-hand sides of 1 e
(17),(18) jump by amounts -2'IT\l(E_), 2'IT\l(E_) at .2_£ aB, whilst $
remains continuous, so yielding the boundary relations
¢{£) = I g(E_.~)i
aB
)J (~) dq+ 2'IT)J (£) ; E.£ aB (19)
¢{£) I ll(~)dq-
aB
g(E_d)i 2'IT)J(E_) ££ aB {20)
403
respectively. These are Fredholm integral equations of

the second kind for ~ in terms of ~ on aB, and equation (19)
has a unique solution which generates ~ in B. by virtue
of (17). Equation (20) only has a solution lf
J ~(£)A(£)dp = 0. (21 )
aB
in which case
~ =~ + k
0
(22)
where ~ 0 is a particular solution and h is an arbitrary
constant. Condition (21) essentially arises from the fact
that the representation (18) implies
~ = Oi£1- 2 as (£) + oo • (23)
whereas the classical existence theorem envisages the more
general behaviour (10). Bearing in mind that
J ~(£)A(£)dp = Ja(3_) dq (24)

aB aB
we see that condition (21) eliminates the leading term of

(10) and converts it into (23). As regards (22), k
satisfies the homogeneous equation
0 J 9(£,3_)~ k dq - 2rrk (25)

aB
which is a disguised form of the Gauss flux theorem,
more usually expressed as
(26)
3. Green's formula
All the above formulations are indirect in that they
introduce hypothetical source densities on the boundary.
However, there also exist direct formulations in which the
boundary data themselves function as source densities. Thus,
given a harmonic function ~ in B. which assumes continuous
values~ (3_) and normal derivativ~s ~'(g) at each .9.£ aB, we
have identically
404
Ig(.e_,_g) i <P(_g_)dq - I g(.e_,_g) <Pi(g_)dq 41r<P(E_) ; E.£ Bi. {27)

aB aB
This is Green's formula. It exhibits <P in B. as the

superposition of a double-layer potential, g~nerated by
a source density .p/4TI on aB and of a simple-layer potential
generated by a source density -.p!/4TI on aB. However,
according to the fundamental existence theorems, either .p
alone or .p!1 alone essentially suffice to determine <Pin B1
Thus, the application of Green's formula requires more
boundary information than would be available in any well
posed boundary-value problem. An effective way forward is to
note that the second integral in {27) remains continuous at
aB, whilst the first integral jumps by an amount -2TI{<P/4TI),
i.e. -.p/2 , at aB so yielding the boundary formula
aB
I
g(.e_,_g)i<P(g)dq -
aB
I
g(.e_,~)<Pi(_g_)dq=2TI<P{E.) ;E_£ aB (28)
This differs fundamentally from {27) in that it provides
a functional relation between <P, .p! on aB which ensures
their compatibility as boundary data. If, for instance,
values of <P and .p! are separately prescribed on aB, they
could not be util~sed in (27) unless {28) were satisfied.
Well posed boundary conditions provide partial information
which may be extended by coupling it with (28). Some
examples will now be outlined.
Given <P on aB (Dirichlet problem), relation {28)
becomes a Fredholm integral equation of the first kind for
<Pi in terms of .p • This has a unique solution, which
automatically satisfies the Gauss condition (14). Conversely,
given .p! on aB (Neumann problem), relation {28) becomes
an inte~ral equation of the second kind for .p in terms of .p!
~~!!only has a solution if <Pi satisfies (14), in which 1
 + S<P!1 y (30)
at each point of aB, i.e. a = 1, s = 0 and a= 0, s = 1

respectively, where y is a given continuous function on aB.
An existence-uniqueness theorem is available for the Robin
problem of heat conduction, i.e.
a< 0, S = 1, y given on aB. (31)
405
Finally, an existence-uniqueness theorem is available

for the difficult mixed boundary conditions
a= 1, 8 = 0 on aB 1 and a = 0, 8 =
(32)
aB aB 1 + aB 2 y given on aB
Essentially, in all these cases, we couple the local relation
{30) with the global relation (28) to compute a compatible
~. ~! on aB. From the numerical point of view, hardly any
dist~nction needs to be made between one type of boundary
condition and another.
To each interior problem there corresponds an exterior
problem subject to regular behaviour at infinity. Thus,
Green's exterior formula is
l g(_E_,_g_)~~(_g_)dq
aB aB
- f g(_p_,_g_)~~(_g_)dq = 4n~(_p_) _p_s Be, {33)
which yields the exterior boundary formula
J g(_p_,_g_)~~(_g_)dq - f g{_p_,_g_)~~(_g_)dq 2n~(_p_) ; _p_s aB . {34)

aB aB
Given~ on aB (exterior Dirichlet problem), relation (34)

becomes a Fredholm integral equation of the first kind for
~e in terms of ~ . This has a unique solution, which does
not necessarily satisfy (14). Conversely, given ~· on aB
(exterior Neumann problem), this becomes a Fredholme
integral equation of the second kind for ~ which always has
a unique solution. With a compatible ~ .~· known on aB,
we may generate~ in Be via (33). The ext~rior problems
corresponding to (31), (32) present no special analytical
difficulties compared with the interior problems.
Nothing in potential theory is more fundamental than
Green's formula, and we now prove that it covers the indirect
formulations mentioned in section 2. First, relation (28)
yields
J g (_p_,_g_) i ~ (_g_) dq - J g (_p_,_g_) ~i (_g_) dq = 0 ; _p_ de ( 35)

aB aB
as p crosses from aB into Be . Similarly, relation {34)
yieTds
I g(_p_._g_)~~(_g_)dq - f g(_p_._g_)~i (_g_)dq {36)

aB aB
406
as aB crosses from aB into Bi. Changing~ to fin (36),

where f is an arbitrary regular harmonic function in B ,
and superposing this on (27), we obtain the more gener~l
continuation formula
I g(.e_,g)i [~(g_)-f(_g_)]dq - I g(.e_,g) [~i (g_)+f~(_g_)]dq=4n~(e) ;.E_E Be.
aB aB (37)
Two distinct possibilities for f may now be considered
1. f =~on aB, providing the representation
- h
aB
I
g (.E_,_g_) [~i (_g_) + f ~ (g_)] dq = ~ (.E_) ; .E. E: Bi ' ( 38)
which may be identified as a simple-layer potential generated

by the source density
(39)
This possibility hinges upon the existence of a unique
regular f on Be satisfying f = ~ on aB, as is in fact
ensured by the exterior Dirichlet existence theorem. Accor-
dingly we have recovered the formulation (7) of Dirichlet
problems by an adaption of Green's formula, and we have
also shown that the integral equation (9) has a unique
solution given by (39).
2. f~ = -~i on aB, providing the representation
in I
aB
g(.e_,g_)' [~(_g_) - f(_g_)]dq = ~(£) ; p E Bi (40)
which may be identified as a double-layer potential

generated by the source density
~ = !n (~-f) on aB. (41)
This possibility hinges upon the existence of a unique

regular f in Be satisfying f' = -$. on aB, as is in
fact ensured by the exterior Neuman~ existence theorem.
Accordingly we have recovered the formulation (17) of
Dirichlet problems by an adaptation of Green's formula,
and we have also shown that the integral equation (19) has
a unique solution given by (41).
407
4. Vector potential theory

Classical linear elastostatics may be formulated by a
vector potential theory which closely parallels scalar
potential theory (Kupradze, 1965). Thus the elastostatic
displacement vector parallels the scalar harmonic function.
The traction vector parallels the normal derivative. Well
known global properties of the traction parallel the Gauss
flow theorem. Vector potentials may be constructed which
closely parallel the scalar simple - and double - layer
potentials. Vector boundary integral equations parallel the
scalar boundary integral equations.
Corresponding to g(Q.~) we introduce the fundamental
displacement dyadic of the medium
rg(Qdl)
g(Qz~l)
g(Qdz)
g(Qz~z)
g(Ql9:.3)]
g(Q2~3) (42)
lg(Q3~l) g(Q3~2) g(£_3~3) .
Here g(£a9:.n) signifies the displacementin the a-direction

at£ generated by a unit point force active in the y-
direction at q, i.e. column 1 defines the displacement ·
vector at p generated by a unit point force in the 1-direction
at~· etc.-By virtue of the equality
g(~n) may also be interpreted as the displacement in

the n-direction at q generated by a unit point force in the
a-direction at p ' e.g. row 1 gives the displacement
vector at~ generated by a unit point force in the 1-
direction at p . Corresponding to g' (p,q) we have the fun-
damental traction dyadic of the medium--
~*(Q.~) = [ g* (1'.,~, I g*(Ql~2) g*(Qd3)
g*(Q2~l) g*(Q2~2) g*(Q2~3) (43)
g*(Q3~l) g*(Q3~2) g*(Q3~3)
where column 1 defines the traction vector at p generated

by a unit point force acting in the 1-direction at ~· etc.
Similarly the traction dyadic
408
~(p_,_g)* g(p_l_g)*
g (£2.9.2) * (44)
g (£3.9.2) *
parallels g(p_,_g_)'. Each column of (44) defines a_singula~

elastostatic displacement vector generated by a un1t tract1on
source at q e.g. column 1 refers to a unit traction
source actTng in the 1-direction at .9.• in line with the fact
that g(p_,q)' may function as a unit dipole potential . .
Clearly each row of (43) defines a singular elastostatlC
displacement vector generated by a unit traction source at p_,
and by the same token each row of (44) defines a traction
vector generated by a unit point force at p_. Any
individual component of the dyadics {43),(44) carries two
possible interpretations, i.e. it is the traction component
generated by a unit point force or it is the displacement
component generated by a unit traction source. However,
the interpretation will always be clear from the context.
Finally we note that g*{p_,S) stands for gt(p,_g_) or g*(p_,_g_)
as the case may be, and sim1larly for g(p_,_g_)* . e
Corresponding to a we introduce the vector simple-source
density a , which enables us to write down the vector
integral-equation
1. (p_) = f ~(p_ '.9.) . .2: (.9.) dq p_ e: aB (45)
aB
analogous to (q), with components p_ e: aB
3
= J L g(p .9. ) a (_g_)dq a= 1,2,3 (46)
aB n= l --a n n
where ¢ is a prescribed displacement on aB. By analogy with
{39), this has the unique solution
= - !rr (~*·
1
+ f*)
-e
( 47)
where ~* signifies the traction vector associated with p

in Bi and f* signifies the traction vector associated with
f in Be here f is the unique regular displacement
vector in Be defined by i = p on aB. Also, corresponding
with the electrostatic equation (11), we have the vector
equation
~ +.!?.A£ = f ~(p_,_g_). ~(_g_)dq p_ e: aB (48)
aB
409
where a,b are arbitrary constant vectors defining an

arbitrary rigid-body displacement field. It is convenient
to break this down into the six independent vectors
<1,0,0> ' ~2 <0,1,0> ' ~3 <0 '0' 1> }
(49)
<0,0,1>/\£
which break down (48) into the six independent equations
E.£ aB
~s(£) = J ~ (£,.9_) . ~s (_g_) dq s = 1,2, ... ,6. (50)
aB
The solutions ~s play an analogous role in the vector theory
to that of the electrostatic source density A
Corresponding with (12) we have the vector integral
equation
~i(£) = J gi(£,.9_). E.(_g_)dq - 2Tr E.(£); E.£ aB, (51)
aB
where the traction vector ~~ is given on aB. This equation
only has a solution if 1
J ~i(£) • ~s(£)dp= 0 ; s = 1,2, ... ,6 (52)

aB
i.e. the tractions must be self-equilibrated with respect
to both force and moment resultants on aB. Clearly (52)
is analogous to (14). If (52) holds, then
E. = ~ + k 5 ~5 ; s = 1,2, ... ,6 (53)
where a is a particular solution and the k are arbitrary

coeffictgnts. This non-uniqueness stems ess~ntially from
the fact that -~ = -~son aB implies -~ = -'--5
~ in B.,
1
which implies p* = 0 on aB, which implies that ~ in
(50) satisfy the homogeneous equations ££ aB
0 = J. gi (£,.9_) . ~s (_g_)dq - 2Tr~s (£) ; s = 1 ,2' ... ,6 (54)

aB
analogous to (16). Similarly, corresponding with (20), we
have the vector integral equation
p(£) = J ~(£ •.9.)i . ~(_g_)dq - 2Tr~(£) ; ££ aB (55)
aB
which only has a solution if
410
J .tCe). ~s(p_)dp J Q.(g) • .H.s (q) dq =0 {56)

aB aB
in which case
l!o+ ks.H.s; s=l,2, ... ,6 (57)
where ~ is a particular solution and the ks are arbitrary

coefficients. The homogeneous equations
0 = f p E aB
j(p_,g)i. ~s(g_)dq- 2TI.H.s(p_) ; ~ = 1,2, ... ,6 {58)
aB
are essentially vector generalisations of the Gauss flux
theorem, alternatively expressed by
f i(p_,g) i . .H.s (g) dq 4TI.H.s (.P_) p_ E Bi

aB
2TI.H.s (p_) p_ E aB {59)
0 p_ E Be
for s = 1,2, ... ,6 . The vector analogues of equations
{13), {19) hardly need to be written down explicitly, and
the analogue of (19) has a unique solution given by
{60)
where f signifies the unique regular exterior displacement

field on aB defined by f*-e = -~~:~:,
on aB.
5. Somigliana's formula
Corresponding to Green's formula (27), we have
Somigliana's formula
~ [g_(p_,g)i·.t(g) - ~(p_,g_). ~i(_g_)] dq

which yields the boundary formula
J ~(p_,_g_)i ·_t(_g_)-g(E.•.9.) ·.ti(_g_)]dq = 2Trf(p_) ; p_ E aB. (62)

aB
This provides a functional relation between tractions
and displacements on aB which ensures their compatibility
411
as boundary data. Given ¢ on aB, this becomes a vector

integral equation for i~-on aB , with a unique solution
which automatically sati~fies (52). Conversely,
given ¢~ on aB, this becomes a vector integral for ion aB,
which_, only has a solution if (52) holds.
In this case
-¢ = ¢-o + ks-s
]l s=l,2, ... ,6 (63)
as in (57). There are, of course, corresponding exterior

equations, and reciprocal relations analogous to (35), (36).
Integral equations generated from Somigliana's formula were
formulated and solved numerically by Rizzo, 1967.
List of References
Jaswon, M.A. &Symm, G.T. (1977). Integral Equation Methods

in Potential Theory and Elastostatics. Academic Press :
London & New York.
Kupradze, V. D. (1965) . Potentia 1 r1ethods in the Theory
of Elasticity. Israel Program for Scientific Translations
Jerusalem.
Rizzo, F.J. (1967). An integral equation approach to
boundary value problems of classical elastostatics.
Quart. App. Math. 25 (1), 83-95.
412
ON THE ASYMPTOTIC CONVERGENCE OF BOUNDARY INTEGRAL METHODS
W. L. WENDLAND
TEClilliSCHE HOCHSCHULE DARMSTADT, GERMANY
and UNIVERSITY OF DELAWARE, NEWARK, DEL. 19711 U.S.A.
INTRODUCTION
The numerical treatment and corresponding error analysis of

boundary integral equations hinges on the type of discreti-
zation due to the shape and type of trial functions used for
the approximation of the unknown functions, due to the type of
test functionalsreplacing the integral equations- which hold
everywhere on the boundary - by a finite number of equations
and due to the numerical integration. In reality further errors
accumulate from round off effects. Here we are concerned with an
error analysis which only takes into account the effects of the
first three kinds. Since it seems to be a too pretentious task
to find computable error bounds we consider so called asympto-
tic estimates.
For the trial functions we use finite element functions on the

boundary manifold, i.e. boundary elements. We consider the
boundary manifold to be given by local representations and re-
gular partitions of the parameter domain into subintervals or
triangles, such that they are mapped onto corresponding parti-
tions of the boundary curve or surface r , respectively. To
one partition we assign a meshwidth h , i.e. the longest sub-
interval or side length of the partitioning triangles. On the
partition in the parameter domains we use as trial functions
regular finite elements, e.g. piecewise polynomials. Then the
local representation of r transplants these finite element
functions onto r . For the error analysis we consider not only
one partition but a whole family with corresponding h tending
to zero. For this whole family we choose a regular (m+l ,m)
system in the sense of [6]. Then asymptotic error estimates are
formulated in terms of powers of h if h + 0 •
Here we present such estimates for the two most popular methods
of discretization, the Galerkin procedure and the standard
collocation together with corresponding discrete versions due
413
to numerical integrations.
For Galerkin's procedure, the error analysis can be developed

along the lines of finite element methods for variational pro-
blems. Theoretically Galerkin's method requires a time con-
suming double integration over r for the computation of every
element of the stiffness matrix. Thus we choose the correspon-
ding numerical quadratumformulas with respect to a fast compu-
tation, i.e. by evaluating the kernels of the integral opera-
tors as seldom as possible. To this end we assume that the
principal part of the equation is given by a convolutional in-
tegral, After converting the principal part into a standard
form, the corresponding weights in the stiffness matrix can be
evaluated from two vectors of weights forming a Toeplitz matrix
provided the finite elements are defined by shifting and
stretching subject to a regular uniform grid. The vector of
weights can be evaluated independently of the boundary and in-
dependently of the meshwidth. For all remaining smooth terms we
use numerical integrations with node points connected with the
finite element grid. For n=2 , i.e. boundary elements on a
curve we perform this method in §2. It results in a very simple
and effective modified collocation scheme which we call Galer-
kin-collocation. Numerical experiments with examples from-con=
formal mapping, plate bending and Stokes flows show high accu-
racy and efficiency of this method.
On the other hand almost all computer programs inplementing the

boundary integral method are executing the standard collocation
method numerically. Under the same assumption on the principal
part as before it can be shown that the convergence of the
standard collocation follows from that of the Galerkin proce-
dure. Then the convergence can be transmitted to the discre-
tized versions if the numerical integrations are accurate
enough.
In all these cases the asymptotic error estimates show an op-

timal order t of ht, i.e. the same order as the best appro-
ximation to the actual solution in the sense of (m+l,m) finite
element systems. The class of linear boundary integral equa-
tions providing these results is characterized by a generali-
zation of the positive definiteness of energy forms, i.e. co-
erciveness in the sense of the G&rding inequality. In addition
we assume that the principal parts are convolutional operators.
This class is characterized in §I, It turns out that most pro-
blems of the applications belong to this class. The reason is
that those boundary integral equations that correspond to
boundary value problems are defined via the fundamental solu-
tion which depends on the difference between the integration
point and the observation point only. In §2 we formulate the
Galerkin procedure and the corresponding asymptotic error esti-
mates as well as the Galerkin-collocation. For the latter we
present a numerical example from [30], In §3 we formulate
corresponding estimates for the standard collocation.
414
The whole lecture is essentially extracted from [32] where al-

so the proofs can be found which here are omitted completely.
For brevity we consider here only one single equation although
in the applications one mostly finds systems of integral
equations. The error analysis applied to systems also [32]. The
whole analysis can also be extended to the integral equations
of mixed boundary value problems which correspond to crack and
punch problems [33].
I. STRONGLY ELLIPTIC INTEGRAL EQUATIONS
All boundary element methods can be considered as to be appro-

ximations of corresponding boundary integral equations. Con-
versely, all boundary value problems of mathematical physics
with elliptic differential equations can be transformed into
equivalent boundary integral equations provided the fundamen-
tal solution of the differential operator is available. The
boundary integral equations are by no means uniquely determined
by the boundary value problem. In general there are several
different integral equations being equivalent to the original
problem. E.g. the Dirichlet problem for the Laplacian can be
solvedwith the classical Fredholm integral equation of the
second kind via Gaussians double layer potential [27] or with
the Fredholm integral equation of the first kind ("Symm's
equation' in two dimensions) via Green's identity and the
"direct method" [16]. A third method using complex function
theory leads to a singular integral equation with Cauchy's
kernel [20]. (For singular integral equations see also [19].)
The Neumann problem for the Laplacian can also be solved with
everyone of the above types of boundary integral equations. In
addition, one can even formulate a corresponding integra-dif-
ferential equation with nonintegrable kernel [12].
All these types of equations or corresponding systems appear

in applications as in classical potential theory, scattering
theory and inviscid incompressible ideal flows in conformal
mappings, viscous flows, electrostatics and elastomechanics,
in elasticity, thermoelasticity and electromagnetic field
theory.
Although these types of equations have very different proper-

ties in classical theory of integral equations it turns out
that if they are considered as so called pseudodifferential
operators [28] they have a very strong, common property.
Namely the equations of practical interest are "strongly el-
liptic". In order to formulate this property one needs the
Sobolev spaces Hs(r) of generalized functions on r , their
interpolation spaces and their dual spaces. For the definitions
we refer to [2] (in particular p. 214). Then each of the above
mentioned o~~rators A defines a continuous linear mapping
A : Hs + Hs a for a whole scale of real s (depending on
the smoothness of r ). 2a is called the order of the pseudo-
415
differential operator A [28] • (G. Richter calls -2a in [26]

"smoothing index''.) For our examples we have 2a = 0; -I; I ,
respectively. The boundary integral equation we write in short
Au =f on r . (I • I)
The announced common property is the coerciveness in form of

the G3rding inequality:
Re(Av,v) Re I vAv ds <! Yll vii~ - lk[v,vJI (1.2)

r for all v E Ha(r)
where y >0 is a constant independent of vHa anda where

k[u,v] denotes a compact bilinear form on x H • In some
cases k equals zero, then Equation (1.2) corresponds to
energy estimates.
We further assume that the principal part of A is a convolu-

tional operator and - for brevity - that r is a simple closed
curve. Let r be given by a regular parameter representation
r : x = x ( t) , t € [ o, 1J (I • 3)
with x(t) a !-periodic sufficiently smooth vector valued

function satisfying
~~~~ = p(t) ~ p0 > 0 for all t , (1. 4)
where p denote~ the Jacobian. Then the operator A with a

convolution operator as principal part has the form
p.v. I 1[p 1 (t-T) +loglt-Tip 2 (t-T)]u(t)p(t))dt

It-rl<z (I .5)
+ I 1L(T,t)(u(t)p(t))dt = f(t) .
~-T 1<2
Here p 1 (~) and p 2 (~) for ~ +

0 are homogeneous functions
of degree S 1-2a-n • In this case the coerciveness (Equation
(1.2))is a consequence of the so called strong ellipticity of
the operator A which is defined by means of the Fourier trans-
formed kernel function a 0 (~) of the principal part of A, i.e.
a 0 (0 := F(p 1 (·) + logl·lp 2 (•))1~ (1.6)
A is called strongly elliptic if there exists a complex valued

0 +0 and a constant y 1 such that
Re 0a (0 > y I > 0
0 -
for all I~ I = I. ~ € lR. (1.7)
without loss of generality we may multiply the Equation (1.1)

by 0 then in Equation (1.7) and furtheron it suffices to con-
sider the case 0 = I .
416
For singular integral equations with the Cauchy kernel, the

above definition and in particular the special form (Equation
(1.5)) of A are too restrictive. We leave this detail to [32].
From now on we consider strongly elliptic integral equations of

the form of Equation (1.5)and we further assume that there-
maining terms collected in L(T,t) define a sufficiently smooth
function of T and t • Otherwise we again split into two
terms, where the first contains the singularity and has to be
treated similarly to the principal part. Then strong ellipti-
city implies coerciveness (Equation (1.2)) [17].
To our knowledge, all examples of boundary integral equations

in applications belong to the class of strongly elliptic inte-
gral equations or systems, for references see[ll] and [30].
Since in Equation (1.5) only p depends on r we consider

Equation (1.5) as an integral equation over [0,1] for the
)-periodic new unknown function
v(t) := p (t)u(t) • (I. 8)
Note that the principal part in Equation (1.5) then becomes
independent of the special choice of the curve r .
2. THE GALERKIN METHOD AND THE GALERKIN-COLLOCATION
Now let us assume that the linear boundary integral equation

Au = f on r (2. I)
is strongly elliptic and has no eigensolution,i.e. the solution
of Equation (2.1) is unique and therefore it is uniquely solva-
ble due to Fredholm's alternative which is valid because of the
estimate (1.2). In case of eigensolutions to Equation (2.1)
the problem can easily be modified [32].
By Hh let us denote the finite dimensional space of finite

element functions on r belonging to the meshwidth h > 0
and belonging to the regular (m+l,m) system [6]. Let
~·. j=O,I, ••• ,N denote a basis of Hh. Then the well known
Galerkin procedure for Equation (2.1) is to find the coeffi-
cients Yj of the approximate solution
N
v(x) = I y.~. (x) , x e: r (2.2)
j=O J J
by solving the finite system of linear equations,
N
I (A~ .• ~k)y. = (f,~k) • k=O, ••• ,N • (2.3)
j=O J J
For the convergence of this procedure we have well known re-
sults going back to s. Michlin [18], S. Hildebrandt and
E. Wienholtz [10]. Here we use the version known as Cea's
lemma, [8] p. 104.
417
Theorem 2.1: Let Equation (2.1) with A be~ strongly ellip-

tic equation with unique solution u E lP .!2_ any f E H-a.
~ ~ ensts h0 >0 such that Equations (~.3) ~uniquely
solvable !2.!_ every O<h~t:;:-Moreover there ~ ~ constant
c 1ndependent of h and f such ~
< c in£ II u - X II (2 .4)

xdth a
For convenience, in the following asymptotic error analysis we

are always using c, c', ••• as generic constants which might
change their size and meaning at different places.
It should be mentioned that Theorem 2.1 is not restricted to

our finite element approximations but applies to a rather wide
class of Galerkin methods as e.g. for the projection methods
using trigonometric polynomials as in [25].
For the proof of Theorem 2.1 we refer to [32]. The proof is

based on the coerciveness(estimate (1.2)) which also holds
for v E Hh. That means that A can be decomposed as
A =D+ C
where D is a positive definite operator and C
compact.
Our boundary elements have been defined by the transplantation

of a regular (m+1,m) system in the parameter domains onto r
with local parameter representations of r. For calculations, the
integrals can be evaluated by using the local coordinates. In
those the £1nite elements appear as simple functions over the
parameter domains. This construction of finite elements on r
requires that the parameter representations are fully available.
For the two-dimensional case this is a sensible requirement. In
the space, however, the boundary surface has also to be approxi-
mated [21 ], [22]. The regular (m+1 ,m) systens in the parameter
domains in connection with regular representations of the
boundary curve provide the approximation property:
.!:.!;!. -m-1
~t ~ s ~ m+l , -m ~ s, t ~ m • ~ .!2_ any given
u E Hs (r) and any Hh of ~ family there exists ~ ~
element function uh E % such that C'J'T-
(2.6)
The constant c is independent of u,h and ~ • Moreover,

it provides the inverse assumption:
For -m ~ t ~ s ~ m and any ll E Hh there holds an estimate
!Ill II s ~ ch t- s !Ill II t (2.7)

H H
418
where the constant c is independent of 11 and h [ 23].

Using the regular finite elements for Galerkin's method
(Equations (2.2), (2.3)) and inserting Formula (2.6) into
Formula (2.4) we immediately find optimal order of convergence.
The result can easily be improved with the inverse assumption
(Formula (2.7)). But with all our assumptions, also the
Aubin-Nitsche lemma can be applied and yields super approxi-
mation [15]. Collecting all these results we find the following
convergence theorem [15],[26],
Theorem 2.2: Let A in Equation (2.1) be strongly elliptic

and Equat1on (2.1) be uniquely solvable.~ t ~the
Galerkin solution of Equations (2.2), (2.3) subject~ regu-
lar f1n1te element spaces providing the ineaualities (2.6)
and --rr:rf . Dehne a.' :- min{a.,or- and suppose
2a. - m - I ~ t ~ s ~ m + I , -m ~ a. ~ m , a. ~ s, t ~ m .
Then~ have the asymptotic~ estimate
II u - ~~~ t 5: chs-tll ull (2. 8)

H Hs
In addition, if we consider the discrete equations (2.3) in

L2 (r) then we find for the ~ditioning the asymptotic~
mate
~ 2a. I
[[ v[l 1 (r) ~ ch [[ f [[ 1 (r) • (2.9)
2 2
For t = -m-1+2a. and s = m+l in E~uation (2.8) one finds

super approximation of order h 2m+ 2- a. which yields correspon-
ding super approximation of the desired potentials away from r.
For the numerical implementation of Galerkin's procedure

(Equations (2.3)), the weights
ajk := (Ailj,llk), j,k = O, ... ,N (2.10)
have to be evaluated. Since A is given by an integral opera-

tor (in the usual or the generalized) sense, the computation
of ajk requires a double integration over rxr . If this
is done numerically, the kernels of the integral operators must
be computed at all combinations of grid points on r • In
addition, special care must be taken of the singular integrals.
Therefore we shall adapt numerical integration to the special
integrals in Equation (1.5).
The principal part in the standard form (Equation (1.5)) will
be handled independently of the special boundary r yielding
a Toeplitz matrix whose elements are given by a vector. This
vector can be computed exactly up to the desired accuracy once
for all independent of r as well as of h for any fixed type
of elements, It should be pointed out that the accuracy of the
numerical results depends significantly on how to compute the
approximate principal part.
419
The Galerkin weights due to the smooth rema1n1ng parts will be

treated numerically by appropriate quadrature formulas depen-
ding on the particular finite elements to be used. In them we
use only grid points in a regular grid connected with the
finite elements such that the kernel functions are to be
evaluated as seldom as necessary. This leads to simple modi-
fied collocation formulas and the computation of the corres-
ponding stiffness matrix is extremely fast.
In order to utilize the convolution in the principal part we

use regular finite elements on a uniform grid of [0,1] de-
fined with shifts and stretched variables from one shape
function ~(n) • The latter we define as in [5] Chap. 4 b1
suitable piecewise polynomials of order m with ~ E em-
For m = 0,1,2 e.g. we have
m= 0 m= I m=2 for
I I 2 0 n I
n zn $ <
0 2 - TJ -n2 + 3n- 3/2 I $ n < 2

~ (n) = (2. II)
0 0 I 2 - 3n+ 9/2 2 3
2'1 $ n <
0 0 0 elsewhere
With ~ we define a basis of Hh by
~j(t):= ~(~- j) for hj:>t:>l+hj, j=O, ••• ,N, h=I/(N+I) (2. 12)
and their !-periodic extensions

~.(t+t) := ~.(t) for integer t .
J J
For v in Equation (1.8) we use the approximation
N
vh(t) := L y.~.(t) (2. 13)
j=O J J
Inserting Equations(2,12) into Equations(2.10) we find for the
terms due to the first expression in Equation (1.5)
I
d.k
J •
fp.v. f 1 [p 1 (t-T)+loglt-Tip 2 (t-T)]~j(t)dt~k(T)dT
o ~-TI~
2 m+l m+l
h +B{ f p.v. f [p 1 (t'-T'+(j-k))+p 2 •loglt'-T'+(j-k)iJ
T 1 =0 t'=O
' ~(t 1 )~(T 1 )dt 1 dT 1
m+l m+l
+ log h f p.v. f p 2 (t'-T'+(j-k))~(t')~(T')dt'dT'} ,
T'=O t'=O
420
2+fl
djk = h {Wlp + w2P log h} with p j-k E l • (2. 14)
Here the two vectors of weights

m+l m+l
J p.V. J [p 1 (t'-T 1 +p)+p 2loglt'-T 1 +piJ)J(t 1 ))J(T 1 )dt 1 dT 1
T 1 =0 t'=O
(2. IS)
m+l m+l
J p.v. f p 2 (t'-T'+p))J(t')ll(T')dt'dT',P E z (2. 16)
T'=O t'=O
can be computed once for all independent of r and h • For
more details see [32] and [II]. For all the remaining smooth
terms in the Galerkin equations to Equation (1.5) we use nu-
merical integration.
Since in the corresponding integrals

m+l
J f ( t)\l . ( t) d t = h J f (h (j +cr) )\l (cr) dcr (2. 17)
supp \l· J cr=O
J
the finite element functions appear as factors, the numerical
integrations are chosen accordingly to the respective refe-
rence function \l such that polynomials f up to the order
2M+I are integrated exactly. This leads to formulas like
M
J f(t))J.(t)dt = h I blf(z.l) + R (2.18)
supp \l• J l=-M J
J
where
z k ·=
•
m+l
z(h(k+ --2--)) and zj l := z(h (j + m; I + .t• y l)) (2. 19)
are the gridpoints subject to the boundary elements and, cor-

respondingly subject to the integration formula. R denotes the
error term which is of order 2M+2 • The simplest choice yl =I
yields zjl = zj+l and weights bl = b_l as follows:
m= 0 m= I m= 2
b b) b b) b b)
0 0 0
I
M= 0 : I 0 I 0 0
3
II I 5 I 3 I
M= I :
12 24 6 12 4 8
I
For yl = 2 and M = 2 one has
m=l m= 2
b b) b2 b0 b) b2
0
13 4 I 2 7 I
30 TI 6o 5 30 T5
421
Using Equation (2.18) for the smooth terms of the weights in

Equation (1.5) we obtain
I
J
T=O
(2.20)
2 u
h I
b.b 0 L(zk.,z. 0 ) + R
.t,i=-M l ~ l J~
with the error term

s+2 asL asL
!RI ::; h c{max!-1+ maxi-!}, O::;s::;2H+2 • (2.21)
dTS dtS
Now we are ready to formulate the Galerkin-collocation

equations by using Equations (1.8), (2.2), (2,14) and (2.20),
They read as
N
N 2+S
I
j=O
ah.ky.:= I {h
J J j=O
(w 1 (. k) + log h w2 ,p (J. ,k))
,p J,
2 M
+ h I b.b 0 L(zk.,z. 0 )}y. (2.22)
!,i=-M l ~ l J~ J
M
h I b . f ( zk . ) = : Fk k=O, ••• , N •
i=-M l l
For saving computing time, the values of L and f at the

grid points should be evaluated only ~ at the beginning
and then be stored for further use as to build up the stiff-
ness matrix in Equations(2.22).
I I
This suggests a choice Y£_ = I or Z or 3 etc. in the
numerical integration formulas.
For the examples treated in [11],[12],[13],[30] with

a = - ~ = a' the choice m=2, M=l and Y£_ = I provided
excell~nt numerical results in combination with short compu-
ting times. For the asymptotic error due to the Galerkin-
collocation we shall use the already established error
estimates ( Formula (2.8)) for Galerkin's method, To this end
we abbreviate the Equations (2.22) by
N
I ah.ky. Fk, k= o, ... ,N (2.23)
j=O J J
~
as mappings in If
~·
N
wh I a.lJ.
J J
(2.24)
j=O
then the mapping A associated with Equation (2.23),

422
(2.25)
will be defined by the linear equations for the coefficients

S,e •
k=O,,,, ,N. (2.26)
Since the Gram matrix (~,e.~k) is regular, A in Equation

(2.25) is well defined. Correspondingly we define r E ~ by
(2.27)
Then the Galerkin Equations (2.3) and the Galerkin-collocation

Equations (2.22) take the form
(2. 28)
respectively. One easily obtains the estimate
(2.29)
This estimate shows clearly that we need estimates for stabi-

lity, i.e. II A 1 11 1 1 , consistency, i.e. II (A-PhAPh)~ll 1
and the truncation e~rSr II Phf-FII 1 • Let us begin with th~
consistency, With Formula (2.21) on~ can prove the following:
Theorem 2. 3: .!:.!:,! ~ wei~hts w1P, w2 P ~ accurate _!£..!!!.
~ ha ~ let <a~) 2M+\ and <aat)2M+ 2L .!?!_ continuous.

Then~ have~ consistency
From the estimates (2.30) and (2.9) one easily obtains

stability,
Theorem 2.4: ~~assumptions of Theorem 2.3 be fulfilled

and 1n add1tion let a > 1+2(a-a 1 ) , M > -a 1-l , Then we have
stahiTity, l,e. there~ ho > 0 ~that - - - - -
(2.32)
where c i! independent of h for all

- -
0 < h ~ h
0
423
Finally, the estimation of the error term R in Equation

(2.18) in connection with Equations(2.27) yields for the trun-
cation error:
Theorem 2. 5.: in Equations(2.27)

there holds
with I !> a !> 2M+2 • (2.33)
Collecting the foregoing estimates and using Formulae

(2.29) and (2.8) we find the following estimates for our
Galerkin-collocation.
Theorem 2.6: For a> m+2+2(a'-a) and M >

- ~-
2 N'
~
we
f1nd an error estimate
(2.34)
with I + 2a 1 !> s !> m+l and 0 !> s
For a> 2m+3-2a'

approximation
s-t
II v- vhll t !> ch {II ull + II ull 1 +II£ II s-t- 2a,} (2.35)
H Hs 2 H
provided 2a-m-l !> t !> s !> m+l , s-t ~ l-2a'
As we can see from the foregoing error estimates, it seems that

the Galerkin-collocation (Equations(2.22)) combines the
theoretical advantages of Galerkin's method with the practical
advantages of the collocation methods. For illustration we
present some numerical results from [13] which have been ob-
tained for the computation of slow plane viscous flows around
elliptical obstacles. Let us consider the exterior flow around
the ellipse
r = ~(t) ={cos 2nt. I sin 2nt} = {z 1 ct), z 2 (t)}
governed by the dimensionless Navier Stokes equations

-+
ll•q =0 (2.36)
in the exterior of r with the boundary conditions
........
+
q =0 on r and q(x) - {1,0} + 0 for lxl -+ 00
q(;) denotes the dimensionless velocity, p is a dimensionless

pressure and R the Reynolds number.
Following Hsiao and Mac Carny [14] we computed the first two
terms of the Stokes expansion
424
+ + + + -1 + + -2
q(x) ~ q 1 (x)(log R) + q 2 (x)(log R) (2.37)
by solving the following system of integral equations:

2 (x 0 -z0 (t))(xS-zS)
f logl~-;(t)lg.s(t)p(t)dt + I { I+ +l 2
f J a=J z(t) - X (2.38)
Ji J.pdt
r
=
+
A.
J
j I, 2 .
+ I 3
Here X = {x 1,x 2 } and ~ = 2(2 + log 4-y) with y Euler's
constant,
:A, =-{1,0} and
+
A2 {wl2' -wll} .
Then the velocity terms in the expansion (2.37) are given by
+
q. (x)
J
+
+ c-}- 20) Jdt
(2.39)
The computations for Equation (2.38) have been made by our

Galerkin-collocation with m = 2, i.e. piecewise quadratic
polynomials and with N = 39 , i.e. 40 grid points. A test
computation (like in [12]) showed accuracy of ~j (and
gj(t)p) up to 6 decimal digits. In particular at the points
{1.5,0}, {1.5, 0.5} and {3,0} we found the following velo-
cities:
R ql q2 ql q2 ql q2
0.0025 0.0399 0 0.0641 -0.0330 o. 1232 0
0.005 0.0426 0 0.0655 -0.0353 0.1317 0
0.0075 0.0442 0 0.0711 -0.0366 0. 1368 0
0.01 0.0454 0 0.0729 -0.0376 0.1402 0
{I • 5; O} {I • 5; 0,5} {3;0}
The results show clearly the acceleration of the flow for

increasing R • Further results including plots of the velo-
city field will be presented in [13]. Other numerical results
for modifications of Equation (2.38) belonging to plate ben-
ding and to conformal mapping can be found in [!IJ,[l2] and
[30].
425
§3 STANDARD COLLOCATION AND ITS DISCRETISATION
Although most numerical implementations of boundary integral

methods are done with the standard collocation there are yet
known only a few results on the asymptotic errors except in the
case of Fredholm integral equations of the second kind. (See
references in [32].) For the more general case there are re-
sults available only for the special case of Symm's equation,
i.e. p 1 = 0 and p 2 =I in Equation (1.5) [1],[3] p.ll9 and
[29].
For singular integral equations with the Cauchy kernel

s. Prossdorf and G. Schmidt have proved that the collocation
method with piecewise linear functions (m=l) only converges
if A is strongly elliptic [24].
We announce here that, conversely, for strongly elliptic

equations with convolutional principal part the convergence of
Galerkin's procedure implies convergence of the standard col-
location.
The standard collocation method can be formulated as to find

N
~ = l y.~. € ~ (3.1)
j=O J J
such that the collocation equations at the grid points,

N
l y.A~.(zk) = f(zk) , k=O, ••• ,N (3.2)
j=O J J
are satisfied. In order to formulate the convergence result
let us introduce the interpolation operator Ih mapping con-
tinuous functions onto ~ :
N N
Ihf = La.~. with a.~.(zk) = f(zk), k=O, ••• ,N (3.3)
j=O J J j=O J J
For Ih we have the approximation property [6]
II Ihf- til t ~ chs-tll til for O~t~s~m+l, ~ ~s, t~m.

H Hs
(3.4)
Then Equations(3.2) can also be written as
(3.5)
Based on the special form (Equation(l.5)) and on estimates

for R in Formulae (2.18) with M = 0 and by making use of
the collectively compact operator theory by P. Anselone [4]
we can prove asymptotic error estimates.
The key result is the stability.

426
Theorem 3.1: [31] The Fguations (3.5) in ~ ~stable

wHh
-I 2a'
II (IhAPh) II L L ~ ch (3.6)
2' 2 of h $ h with a suitable
where c is independent 0 - - - - -~~~~
h > 0
0
Estimate (3.6) is in accordance with [I] and [29]. The sta-

bility in connection with convergence (inequalities (2.6) and
(3.4)) and with the inverse assumption (Formula (2.7)) pro-
vides us the following convergence result:
Theorem 3.2 : The collocation equations (3.2) are uniquely

solvable for O~h $ h 0 and the approximate sotUtions ~
verge.!:!:.!
I I
11~-vll $ chs+ 2 (a'-a)llvll if 0$s$m+I, 2 + 2a<s, 2 +2a<m.
12 Hs
(3.7)
For the discretisation of Equations(3.2) we again insert
Equations (1.8), (3.1) into Equation (1.5) and use Equation
(2.18) for the integral of the second term in Equation (1.5).
Neglecting the error term we find the discrete version of the
collocation equations:
I (hi+B{~
j=O lp
+ ~
2p
log h}
M J
+ h L biL(zk,z.i))y~ f(zk) , (3.8)
i=-M J J
k O, ••• ,N •
X X
Here the weights w 1P, w2P with p j-k also can be com-
puted once for all independently of h and r by the vectors
of weights
m+l m+l
wx = p.v.
lp J
t'=O
)..l(t')[pl (t' - -2- + p)
(3. 9)
m+l
+ P2 loglt' --2- + PIJdt',
m+l
wx
2p
p.v. J m+l
)..l(t')[p2(t' - -2- + p) ]dt I . (3. 10)
t'=O
Since the neglected error terms of estimate (2.8) are of the

same order as those in Formula (2.21) for the Galerkin-collo-
cation, one easily finds that under the same assumptions as in
Theorem 2.3 there holds between Equations (3.2) and Equations
(3.8) a consistency estimate like inequality (2.30). Hence one
eventually has:
427
Theorem 3.3: Let w 1 ~, w 2~ be accurate to an order ha

with a > m+2 + 2(a 1 -a) and let 2M> m-1 - 2a' • Let
( a~ ) ZH+~ and ( aat ) 2M+;, be continuous. Then ~ find ~

error estimate
llvx-v!IL s c•hs{llfll _ + llfll -2 ,} (3. II)

2 Hs 2a Hs a
with I f 2a' s s s m+l and 0 s s •
In practice the numerical integrations in Equation (2.18) are

often executed with Gaussian quadrature. Then the bt are
defined by the Gaussian weights multiplied by ~j(zjt) and
the Yi correspond to q Gauss-Lobatto points in each of the
intervals [K,K+I], respectively,with 0 s K s m , K € m
In this case the estimate (3.11) remains valid if onere- 0
places 2M by q in the assumption.
ACKNOWLEDGEMENTS
The author wants to thank Dr. P. Kopp who made the numerical
computations.
This research was supported by the "Deutsche Forschungsgemein-

schaft" under the project number We 659 and also partially
supported by the "Applied Mathematics Institute" and the
Department of Hathematical Sciences at the University of
Delaware.
REFERENCES
[I] Abou El-Seoud, M.S. (1979) Numerische Behandlung von

schwach singularen Integralgleichungen erster Art.
Dissertation, Technische Hochschule Darmstadt, Germany.
[2] Adams, R.A. (1975) Sobolev Spaces. Academic Press,

New York.
[3] Aleksidze, M.A. (1978) The Solution of Boundary Value

Problems with the Method of the Expansion with Respect
to Nonorthonormal Functions. Nauka, Moscow (Russian).
[4] Anselone, P.M. (1971) Collectively Compact Operator

Approximation Theory. London, Prentice Hall.
[5] Aubin, J.P. (1972) Approximation of Elliptic Boundary-

Value Problems. Wiley-Interscience, New York.
428
[6] Babu~ka, I. and Aziz, A.K. (1972) Survey lectures on

the mathematical foundations of the finite element
method. In: "The Mathematical Foundation of the
Finite Element Method with Applications to Partial Dif-
ferential Equations" (Aziz, A.K. Ed.) 3 - 359,
New York, Academic Press.
[7] Bramble, J. and Schatz, A. (1970) Rayleigh-Ritz-Galer-

kin methods for Dirichlet's problem using subspaces
without boundary conditions. Comm. Pure Appl. Math.
23:653-675.
[8] Ciarlet, P.G. (1978) The Finite Element Method for

Elliptic Problems. North Holland, Amsterdam.
[9] Giroire, J. and Nedelec, J.C. (1978) Numerical

solution of an exterior Neumann problem using a double
layer potential. Math. of Comp. 32:973-990.
[10] Hildebrandt, St. and Wienholtz, E. (1964) Constructive

proofs of representation theorems in separable Hilbert
space. Comm. Pure Appl. Math. ll:369-373.
[II] Hsiao, G.C., Kopp, P. and Wendland, W.L. (1980)

A Galerkin collocation method for some integral equa-
tions of the first kind. Computing ~:89-130.
[12] Hsiao, G.C., Kopp, P. and Wendland, W.L. (1980) The

synthesis of the collocation and the Galerkin method
applied to some integral equations of the first kind.
In: C.A. Brebbia (ed.): New Developments in Boundary
Element Methods. CML Publ. Southampton: 122-136.
[13] Hsiao, G.C., Kopp, P. and Wendland, W.L.: Some appli-

cations of a Galerkin-collocation method for integral
equations of the first kind. In preparation.
[14] Hsiao, G.C. and Mac Carny, R.C. (1973) Solution of

boundary value problems by integral equations of the
first kind. SIAM Review 15:687-705.
[IS] Hsiao, G.C. and Wendland, W.L.: The Aubin-Nitsche lemma

for integral equations. To appear.
[16] Jaswon, M.A. and Symm, G.T. (1977) Integral Equation

Methods in Potential Theory and Elastostatics. Academic
Press London.
[17] Kohn, J.J. and Nirenberg, L. (1965) On the algebra of

pseudodifferential operators. Comm. Pure Appl. Math.
18:269-305.
429
[18] Michlin, S.G. (1962) Variationsmethoden der Mathe-

matischen Physik, Akademie-Verlag Berlin.
[19] Michlin, S.G. and ProBdorf, s. (1980) Singulare Inte-

graloperatoren. Akademie-Verlag Berlin.
[20] Muskhelishvili, N.I. (1953) Singular Integral Equations.

Noordhoff Groningen.
[21] Nedelc, J.C. (1976) Curved finite element methods for

the solution of singular integral equations on sur-
faces in JR3. Comp. Math. Appl. Mech. Engin. ~:61-80.
[22] Nedelec, J.C. (1980) Formulations variationelles de

quelques equations integrales faisant intervenir des
parties finies. In: R. Shaw et al. (ed.): Innovative
Numerical Analysis for the Engineering Sciences.
Univ. Press of Virginia, Charlottesville: 517-524.
[23] Nitsche, J.A. (1970) Zur Konvergenz von Naherungsver-

fahren beztiglich verschiedener Normen. Num. ~~th.
15:224-228.
[24] Prossdorf, S. and Schmidt, G.: A finite element collo-

cation method for singular integral equations.
To appear.
[25] Prossdorf, S. and Silbermann, B. (1977) Projektions-

verfahren und die naherungsweise Losung singularer
Gleichungen. E.G. Teubner Leipzig.
[26] Richter, G.R. (1978) Numerical solution of integral

equations of the first kind with nonsmooth kernels.
SIAM J. Numer. Anal. 17:511-522.
[27] Riesz, F. and Nagy, B.Sz. (1956) Vorlesungen tiber

Funktionalanalysis. Dt. Verl. Wiss. Berlin.
[28] Treves, F. (1980) Introduction to Pseudodifferential

and Fourier Integral Operators I. Plenum Press,
New York and London.
[29] Voronin, V.V. and Cecoho, V.A. (1974) An interpolation

method for solving an integral equation of the first
kind with a logarithmic singularity. Dokl. Akad. Nauk
SSR 216; Soviet Math. Dokl. 15:949-952.
[30] Wendland, W.L. (1980) On Galerkin collocation methods

for integral equations of elliptic boundary value pro-
blems. In: J. Albrecht and L. Collatz (ed.): Numerical
Treatment of Integral Equations. Intern. Ser. Num.
Math., Birkhauser Basel, 53:244-275.
430
[31] Wendland, W.L.: On the asymptotic convergence of the

collocation method for a class of strongly elliptic
integral equations. In preparation.
[32] Wendland, W.L.: Asymptotic accuracy and convergence.

In: C. Brebbia (ed.): Progress in Boundary Elements,
Vol. 1, Pentech Press, London, 1981.
[33] Wendland, W.L., Stephan, E. and Hsiao, G.C. (1979)
On the integral equation method for the plane mixed
boundary value problem of the Laplacian. Math. Meth.
in the Appl. Sci. 1:265-321.
431
BOUNDARY METHODS. C-COMPLETE SYSTEMS FOR THE BIHARMONIC

EQUATIONS
Herve Gourgeon and Ismael Herrera
Institute de Investigaciones en Matematicas Aplicadas y en

Sistemas (IIMAS). Universidad Nacional Autonoma de Mexico
(UNAM), Apdo. Postal 20-726, Mexico 20, D. F. MEXICO
ABSTRACT
A boundary method for solving the biharmonic equation is pre-

sented. It is based on the use of systems of solutions of the
homogeneous equations, which are complete. A convenient crite-
rium for the completeness of such systems, is the notion of
c-completeness. Using a convenient representation of solu-
tions for the biharmonic equation a procedure for constructing
c-complete systems for this equation is developed. Examples
of such systems are constructed.
1. INTRODUCTION
In recent years, by a boundary method, it is usually understood,

a numerical procedure in which a subregion or the entire
region, is left out of the numerical treatment, by
use of available analytical solutions (or more
generally, previously computed solutions). Boundary
methods reduce the dimensions involved in the problem leading
toconsiderable economy in the numerical work and constitute a
veryconvenient manner of treating adequately unbounded regions
by numerical means. Generally, the dimensionality of the
problem is reduced by one, but even when part of the region is
treated by finite elements, the size of the discretized domain
is reduced [Zienkiewicz, 1977, Zienkiewicz, et al., 1977].
There are two main approaches for the formulation of

boundary methods; one is based on the use of bound-
ary integral equations and the other one, on the use
of complete systems of solutions. In numerical
applications, the first one of these methods has
received most of the attention [Brebbia, 1978]. This is in
spite of the fact that the use of complete systems of solutions
presents important numerical advantages; e.g., it avoids the
introduction of singular integral equations and it does not
432
require the construction of a fundamental solution. The lat-

ter is especially relevant in connection with complicated
problems, for which, it may be extremely laborious to build up
a fundamental solution. This is illustrated by the fact that
there are methods for synthetizing fundamental solutions star-
ting from plane waves, which can be shown to be a complete
system [Sanchez-Sesrna, Herrera and Aviles, 1981).
One may advance some possible explanations for this situation.

Although, the principle of superposition, is a standard proce-
dure for building up solutions of linear equations, many of
its applications have been based on the method of separation
of variables; this has lead to the frequent, but false, belief
that complete systems of solutions have to be constructed
specifically for a given region. Of course, this is not the
case; indeed,most frequently systems of solutions are complete
independently of the detailed shape of the region considered
[Herrera and Sabina, 1978), and the systems developed here for
the biharmonic equation possess this property.
Also, in some fields of application, procedures which consti-

tute particular cases of the approximation by complete systems
of solutions, have presented severe restrictions and inconve-
niences. For the case of acoustics and electromagnetic field
computations, a survey of such difficulties, was carried out
by Bates [1975). For this kind of studies, the so called
"Rayleigh hypothesis", restricts drastically the applicability
of the method. However, work by Millar [1973), implies that
these difficulties are due, mainly, to lack of clarity, since
he avoided Rayleigh hypothesis, altogether, by adopting a
different point of view.
Motivated by this situation, one of the authors, started a

systematic research of the subject [Herrera and Sabina, 1978;
Herrera, 1977a, 1979b, 1980e), oriented to clarify the theoret-
ical foundations of the method, allowing its systematic and
reliable use. The aims of the research have been satisfacto-
rily achieved to a large extent and have just been reported
[Herrera, 198lb,c). This has been possible due to the progress
that has been made in the understanding of partial differen-
tial equations [Lions and Magenes, 1972; Temam, 1977). The
methodology also owes much to work of Amerio, Fichera, Picone,
Kupradze and Trefftz [Miranda, 1955; Kupradze, 1967; Trefftz,
1926). The systematic development of the procedure, in a
manner which is applicable to any linear problem, was made
possible, however, by an abstract theory that has been devel-
oped by Herrera [1979a,b, 1980b,c,d,e, 198la).
The numerical solution of Stokes and Navier-Stokes equations,

is a problem of great practical interest at present, and it is
not our purpose to review it, since recent surveys are avail-
able [Glowinski and Pironneau,l978; Temam, 1977). Taking this
interest for granted, we explain briefly the method mentioned
433
before, in connection with the biharmonic equation and supply

an efficient procedure for developing c-complete systems for
this equation, starting from c-complete systems for Laplace
equation.
2. THE BOUNDARY METHOD USED
Consider the biharmonic equation
(2.1)
This equation must be satisfied in the sense of distributions
by elements of some spaces of functions. In general, we ask
u to be in a linear subspace DSH 2 (n)~ so that the equation
(2.1) is between elements of H 2 (n).T
On this assumption the biharmonic problem (2.1) is equivalent

to the formulation
u E Ker P , P:H 2 (n) + (H 2)*
defined by
<Pu,v> = f V4 u v dx V uEH 2 (n)
(2.2)
n v vEH 2 (n)
Integration by parts gives
<Pu, v> f V2u V2v dx + f {at>u v av} (2. 3)
an t>u an dx
n an
In (2.3) four different boundary values occur. We note that

[Lions and Magenes, 1972]
u E H 2(n) ~ u E H3/ 2 (an)
au E Hl/z(an)
an
t>u E H- 1 / 2 (an)
at-u E H-3/2(an)
an
Let us associate with the operator P, an antisymmetric opera-
tor A by
A P - P*
<Au,v> <Pu,v> - <Pv,u> (2.4)
V. u, vED <Au, v> f { v ;n t>u - t>u ~: + t>v ~~ - u aa~v} dx

an (2.5)
in which, only boundary values appear. Let us introduce the
Boundary operators Band B':
t We use the usual notation for Sobolev spaces.

434
V u,v ED <Bu,v> = f v-flu dx

a (2.6)
an an
au
 = + J flv an dx (2. 7)
an
Then
A= B + B' - B'* - B* (2.8)
In fact we can directly define A by (2.5) in a different space
D:
D = {uEH 1 12 (n) j
fluE H0 (an)
D is not a Sobolev space, but we note the inclusions:

H7/2(S1) CDC Hl/2(S1) (2.9)
and that C (S1) is dense in D.
00
Define
1 1 = Ker(B + B') and r 2 = Ker(B* + B'*) (2.10)
Then it can easily be shown that {r 1 ,r 2} is a canonical decom-
position of D, in the sense defined in [Herrera, 1980b]; i.e.,
1 1 ,1 2 are completely regular:
<Au,v> 0 V v E r 1 ~ u E 11 (2.11)
<Au,v> 0 V v E r2 ~ u E 12 (2.12)
and
This implies that every uEo can be writte~ as ~=u 1 +u 2 , with

u Er 1 and u 2Er 2 , and this representation lS unlque, except for
1
e ements of Ker A.
Another canonical decomposition would be
Il = Ker (B - B'*) ; Iz = Ker(B* - B') (2.13)
Notice that the boundary values of elements of D, can be

characterized as follows:
with
u
a
=u u
c
= flu (2.14)
on aS1. Now, associated with the canonical decomposition

{I 1 ,r 2}, we have
D/Ker A = r 1 /Ker A ffi 1 2/Ker A
435
so if u E D/Ker A 3 {u 1 ,u 2 }~ u = u1 + u2
u 1 E Ker(B + B') u1 = [ua , uc] (2.15)

u 2 E Ker(B* + B'*) u2 = [ud, ~] (2.16)
with these notations then it is easy to exhibit the identity:
<Au,v> (u 2 , v 1 ) - (u 1 , v 2)
= (2.17)
where it is understood that if [a,b], [ c , d] E [ H ( (lQ) 12
0
([a,b] , [cd1) = f (ac + bd)dx (2.18)
(lQ
Using the other decomposition we would have similarly

D/Ker A = 1i/Ker A e 1z/Ker A
u E D/Ker A u = u'1 + u'
2
u 1' = [ua,ub] u'2 = [u d' -u c 1 (2.19)
and the identity:
(2.20)
Let Np be the subspace of solutions of the biharmonic equation

(2.1) and define
(2.21)
Then 1P is completely regular; i.e.
<Au,v> = 0 V v E 1p ~ u E 1p (2.22)
This comes straight forwardly from some results of existence

of solution of the biharmonic eqMation with compatible bound-
ary conditions, the density of C (Q) in D and results reported
previously [Herrera, 1980b]. Define IP = 1p/Ker A. The
results of uniqueness imply that NP n Ker A = {O} so that Ip
is naturally imbedded in D/Ker A, with the notations intro-
duced. The following definition is relevant, for our discus-
sion.
Definition: A denum~bie ~et B = {w 1 ,w 2 , ••. } NP ~ on

~-~ompiete (~ompiete ~n ~onneet£vity), wzth ~~pect to A ~6
<Au,w > = 0 Va EN~ u E I •
a P
For any canonical decomposition {1 1 ,1 2 }, IP is decomposed as
rP = r1P e r2P' r1P c 11,r2~ c 12
The following can be proved [Herrera, 1980e, 198lb,c1.
Proposition 1. The 3 ~ta.:temenU aJLe equival.ent

436
i) B is c-complete in IP, with respect to A.

ii) B1 {w1a}aEN spans I1P.
iii) B2 {w2a}aEN spans I 2P.
Let us suppose that we have a c-complete system
B = {wa }a=1,... C N
P
and a canonical decomposition {I 1 ,I 2 }. Consider the problem;
find uED such that
'V 4 u 0 in n
u1 uB given on the boundary
Assume u 1EI 1p, in order to have existence of a solution to the

problem. In view of the Proposition, we know that {w }~
spans I 1P so that any element of I 1p can be approximat~d by
linear combinations of {w 1 }; more precisely, one can choose
coefficients {aN} _ 1 N guch that
a a-, ...
N
u1 (2.24)
has the property that
Then N
N
u r. (2.25)
1
is the biharmonic function in n, such that the boundary value
u~ approximates the data. Therefore, uN is an approximation
to a solution of the problem and as N + oo one has uN + u in the
sense (at least) of H1 12 (r.!) [Lions and Magenes, 1972].
If the missing boundary value u 2 is required, and if it is

known to be in [H 0 (an)] 2 in r 2P, we can indeed approximate it
by N
L: (2.26)
but we can also compute it using (2.17). Indeed
If {w } is orthonormal (and if it is not, we can orthonormal-

ize it~y the well known Gram-Schmidt orthonormalization
process), then, u 2 is given by
00 00
(2.27)
437
3. C-COMPLETE SYSTEMS FOR THE BIHARMONIC EQUATION
In this section we give a general procedure for constructing

c-complete systems for the biharmonic equation, whenever a
c-complete system for Laplace's equation is known.
Proposition 2. Let {w ,w , ...

}be hanmoniQ 6unction6 ~uQh
:thCLt :they Me a Q-Qomptet~ ~y~:tem 6ott LapiaQe equa;t.Lon ..tn :the
tteg..(on n. AMume {cp 1 ,cp 2 , ••• } M.e ~a hanmoniQ and ~uQh :thCLt
acpa
~= l)Ja a=1,2, ... (3.1)
Then :the ~y~:tem {l)J 1 ,_~ 2 , ••• } u {xcp 1 ,xcp 2 , ••• } Me b..thanmoniQ and
Q-Qompiete 6ott eq~n (2.1).
Proof. Consider the canonical decomposition (2.15), (2.16),

then
(3.2)
Here, u, ~u refer to the boundary values on an. The biharmon-

ic problem with u1 given, is the biharmonic equation (2.1),
subjected to the ooundary conditions
on an (3. 3a)
on an (3.3b)
where f 1 and f 2 are given functions of H0 (an). Equivalently,

one can solve
~p 0 in n (3.4a)
p f2 on an (3.4b)
and
~u p in n (3.5a)
u = fl on an (3.5b)
In view of Proposition 1, it is enough to prove that the

system of boundary ~alues {[l)J 1 .~w 1 1, [w 2 ~~w 2 1 , ... }2 U
{[xcp 1 ,~x~ 1 J, [xcp 2 ,~xcp 2 J , ... } spans I 1P- [H 0 (an)] • To this
end, not~ce that
a=l,2, ... (3.6)
Theref~re,_given [f 1 ,f 2] E [H 0 (an)] 2, consider the following

approx~mat~ng sequence
N
u = (3. 7)
438
Define
N N
N
p E bN llxcj>N = 2 E bN l)Ja (3.8)
a a a
a= 1 a=1
and choose bN so that (as N-+oo)
a'
(3.9)
This is possible, because {1)! 1 ,1)! 2 , ••• } is c-complete. Relation

(3.9), implies that there ex~sts vEH 5 12 (an) such that as N-+oo,
N
p + v (3.10)
Therefore
N
p + v (3.11)
Choose now aN so that

a
N N
E aa l)Ja + f 1 - v (3.12)
a=l
This is again possible because {1)! 1 ,1)! 2 , ••• } is c-complete.
Hence, clearly
(3. 13)
and the proof of Proposition 2, is complete.
As an example of the application of Proposition 2, we exhibit

a polynomial system which is c-complete for biharmonic equa-
tion in any bounded region n.
ProEosition 3. Le.t (a=1,2, ••• )
l)Ja Re z(a-1)/2 whe.rt a is odd (3.14a)

a/2
l)Ja Im z whe.rt a ,u., e.ve.rt (3.14b)
cpa = l)Ja+2 ( 3. 15)
The.rt {1)! 1 , 1)! 2 '.••• } U {xcj> 1 , x¢ 2 , .•. } ,U., c.-c.ompf.e.te. fian :.the. bihM-
mortic. e.q~ort, irt arty bourtde.d ne.giort n.
Proof. It has been shown [Herrera and Sabina, 1978], that
{1)! 1 ,1)! 2 , •.• } is c-complete for Laplace's equation in any
bounded region. In addition, it is easy to see that equation
(3.1) is satisfied.
We recall finally, that a c-complete can be used to approximate

439
any other boundary value problem prescribed by means of regular

subspace; this, by virtue of Proposition 1.
4. THE EXTERIOR DOMAIN
Let n be the exterior of a bounded domain. A c-complete

system for Laplace's equation, which satisfies a radiation
condition, inn, is given (Herrera and Sabina, 1978], by
{\jJ1,\jJ2, .•. }
\)!1 Re Log z (4.1a)

-(CI.-1)/2
\j!CI. Re z Cl. odd > 3 (4.1b)
-a/2
\j!CI. Im z Cl. even (4.1c)
Applying Proposition 2, it can be seen that system {\j!l,\j!2, .•. }

U {x~ 1 ,x~ 2 , •.• }, where
~1 Re(z log z-z) ; ~2 Im log z (4.2a)
~CI. Cl. > 3 (4.2b)

\j!CI.-2
is a c-complete system for the exterior problem.
REFERENCES
Bates, R.H.T. (1975) Analytic Constraints on Electro-magnetic

Field Computations. IEEE Trans. on Microwave Theory of Tech-
niques. 11_, 605-623.
Brebbia, C.A. (1978) The Boundary Element Method for Engineers.

Pentech, Press. London.
Glowinski, R., and Pironneau, 0. (1978) On Numerical Methods

for the Stokes Problem. Chapter 13 of Energy Methods in Finite
Element Analysis. R. Glowinski, E.Y. Rodin and O.C. Zienkie-
wicz; ed., Wiley and Sons.
Herrera, I. (1977a) General Variational Principles Applicable

to the Hybrid Element Method. Proc. Nat. Acad. Sci. USA, ~.
7:2595-2597.
Herrera, I. (1979a) On the Variational Principles of Mechanics.

Trends in Applications of Pure Mathematics to Mechanics. II,
115-128, H. Zorsky, ed., Pitman Publishing Limited (Invited
general lecture)
Herrera, I. (1979b) Theory of Connectivity: A Systematic Formu

lation of Boundary Element Methods. Applied Mathematical Model
ling. 1, 2:151-156.
440
Herrera, I. (1980b) Variational Principles for Problems with

Linear Constraints. Prescribed Jumps and Continuation Type
Restrictions. Jour. Inst. Maths. Applies. 12. 67-96.
Herrera, I. {1980c) Boundary Methods in Flow Problems. Proc.

Third International Conference on Finite Elements in Flow
Problems, Banff, Canada, 10-13 June, 30-42. (Invited general
lecture).
Herrera, I. (1980d) Boundary Methods in Water Resources. Finite

Elements in Water Resources, S.Y. Wang, et al., ed., The
University of Mississippi, 58-71. (Invited general lecture).
Herrera, I. (1980e) Boundary Methods. A Criterion for Com-

pleteness. Proc. Nat. Acad. Sci. USA, 12. 8:4395-4398.
Herrera, I. (198la) An Algebraic Theory of Boundary Value

Problems. Comunicaciones Tecnicas, IIMAS-UNAM. (In press).
Herrera, I. (198lb) Boundary Methods in Fluids. FiniteElements

in Fluids, Volume IV, R.H. Gallagher, ed., John Wiley and Sons.
{In press).
Herrera, I. (198lc) Boundary Methods. Theoretical Foundations

for Numerical Applications of Complete Systems of Solutions.
Comunicaciones Tecnicas, IIMAS-UNAM.
Herrera, I., and Sabina, F.J. (1978) Connectivity as an Alter-

native to Boundary Integral Equations. Construction of bases.
Proc. Nat. Acad. Sci. USA, ]2, 5:2059-2063.
Kupradze, V.D. (1967) On the Approximate Solution of Problems

in Mathematical Physics. Russian Math. Surveys. 1l• 2:58-108.
(Uspehi Mat. Nauk. 1l• 2:59-107).
Lions, J.L., and Magenes, E. (1972) Non-homogeneous Boundary

Value Problems and Applications. Springer-Verlag, New York.
Millar, R.F. (1973) The Rayleigh Hypothesis and a Related

Least-Squares Solutions to Scattering Problems for Periodic
Surfaces and other Scatterers. Radio Science. ~. 785-796.
Miranda, C. (1970) Partial Differential Equations of Elliptic

Type. 2nd Ed., Springer-Verlag, New York. (Translation of
Equazioni alle Derivate Parziali di Tipo Ellitico, 1955).
Sanchez-Sesma, F.J., Herrera, I., and Aviles, J. (1981)

Boundary Methods for Elastic Wave Diffraction-Application to
Scattering of SH Waves by Surface Irregularities. Comunicacio-
nes Tecnicas, IIMAS-UNAM. (In press).
Temam, R. (1977) Navier Stokes Equations: Theory and Numerical

Analysis. North-Holland, Amsterdam.
441
Trefftz, E. (1926) Ein Gegenstruck zum Ritzschen Vergaren.

Proc. 2nd. Int. Congress Appl. Mech. Zurich.
Zienkiewicz, O.C. (1977) The Finite Element Method in Enginee£

ing Science. Mc-Graw Hill, New York.
Zienkiewicz, O.C., Kelly, D.W., and Bettess P., (1977) The

Coupling of the Finite Element Methods and Boundary Solution
Procedures. Int. J. Num. Math. Eng. 11. 355-377.
442
THE EFFECT OF MESH REFINEMENT IN THE BOUNDARY ELEMENT SOLUTION

OF LAPLACE'S EQUATION WITH SINGULARITIES
H.L.G. Pina, J.L.M. Fernandes & C.A. Brebbia
ABSTRACT
It is known that the non-smoothness of boundary data causes

the order of convergence of the numerical solutions of partial
differential equations [11] to be less than optimal. In this
paper we assess the efect of mesh grading to overcome this
difficulty in the context of the Boundary Element Method (BEM).
As test cases we employed two potential problems proposed by
Schultz [10]. We conclude that the BEM yields for a given mesh
smaller errors than those obtained by the Finite Difference
Methods (FDM) of [10]. but at the expense of a greater comput~
tional effort. Also a judicious choice of mesh grading can im-
prove significantly the actual error and recover the optimal
order of convergence.
INTRODUCTION
It has been well known for some time that potential problems
in a domain ncRn can be formulated via integral equations on
the boundary r of n. Since 1 this reduces the dimensionality of
the problem from Rn to Rn- , it is expected that the numerical
solution of this class of problems will become easier. The
application of finite element concepts to the discretization of
the boundary integral equations has brought renewed interest in
this field leading to the Boundary Element Method (BEM).
A description of the BEM can be found in [s] where nume-
rous references are given. Brebbia (41 shows how to implement
the BEM for several classes of problems found in elasticity and
fluid flow. The paper by Zienkiewicz et al. (1s] discusses the
relative merits of the Finite Element Method (FEM) and BEM and
develops ways of taking advantage of both techniques. In [2,7]
we can find theoretical analysis of the BEM in the context of
singular Fredholm integral equations.
In this paper we examine the performance of the BEM as
443
compared with the Finite Diference Method (FDM) and FEMusing as

test cases two problems proposed b':y Schultz [10]. Several meshes
are tried, to assess the order of convergence and we investiga
te the effect of mesh grading on the magnitude and order of -
convergence of the error.
THE BOUNDARY INTEGRAL EQUATIONS

2
Let n be a domain (i.e., an open connected bounded set) of R
with boundary r. We wish to solve the following problem for
the unknown function u:
lap u = 0 in n' ( la)
u=g onr1, (lb)
avu = h On f2, (lc)
where r1 and r2 are complementary parts of r. The symbol av(.)
denotes the derivative along the exterior unit normal\! tor.
We assume that the given functions g and h and the boundary r
possess the necessary smoothness properties to ensure the exis
tence and uniqueness of a solution u to problem (1).
The method to transform (1) into a boundary integral equa
tions is well known. If we represent by G(x;y) the fundamental
solution of (la); that is, --
lap G(x;y) + o(x;y) = 0 (2)
where o(x;y) Is the Dirac distribution centered at ~· then we
find that T4,8} u must satisfy the integral equation
c(~) u(~) - J G(~;~) av u(~) dfl(~) + J u(~) av

rl r2
G(l;~) dr 2 (~) = - J g(~) av G(l;~) dr 1 (l) +

rl
+ J G(l;~) h(l) dr 2 (l) • (3)

r2
In this equation c(~) is defined by
c(~) u(~) = Jnu(l) o(~;l) dn(l) , (4)
and by the known properties of the Dirac distribution c(x) = 1

if X is an interior point of n, C(X) = 1/2 if X is a SmOOth
point of r and c(x) = w/2IT if x is-a corner point of r with an
interior solid angle of value w
at x. The fundamental solution
G for the Laplace equation in R2 is-
G(~,~) = in ln
1
l-x-yl
_I
(5)
where lx-yl denotes the Euclidean distance between points ~

and y, - -
-Now av u is unknown on rl and u on r2. We introduce func
tions ~and F such that
~(x) u(x) and F(x;y) for xtr2, (6a)
H~) avii(~) and F(~;l) for ~trl. (6b)
444
Denoting by p(x) the right hand side of equation (3) we can

write this expression as
(7)
We notice that F(x;y), like G(x;y), and 3vG(x;y) have a

singularit~ as X+ y. However the integrals in (3} or (7) do
exist [7,9J. Since p(~) is a known function depending on the
boundary values this expression could be used to find u at any
point x of the domain n if ¢(x) were known. Therefore the next
step is to find methods to determine this function.
If we set the point x on the boundary r equation (7)
yields the following integral equations
c(~) ¢(~)-I F(l,~) ~(l) dr(l) = p(~) for ~~rz,(8a)

- I F(l·~) ¢(~) dr(l) = p(~) - c(~) g(~) for l ~ r1•
r (8b)
These are coupled linear integral equations of Fredholm type
of the second kind on rz and of the first kind on r1, respec-
tively.
Now we show how to solve this problem numerically by the
BEM. Let us construct via finite elements a function space
over r which we denote by VE(r) with dim Vh(r) = Nh < oo and
having the set of function ~hi, i = 1, ••• , Nh as a basis. Then
we seek an approximate solution ¢h to (8) belonging to Vh(f),
As
Nh
¢h(~) =~ ¢hi ~hi(~) • (9)
i=l
and introducing this relation in (8) we obtain
Nh
~ ( c(~) ~hi(~)
i=l r
-I
F(l;~) ~hi(l) dr(l))~hi
= p(~), for~ e r 2 (lOa)
To determine the coefficients ~hi we solve these equations by

collocation at the Nh nodes associated with the finite element
mesh. Thus for any such node ~j we have
Nh
~
L-. (c(x·)
_J oiJ. - J F(y;x·)
- _J
i=l r
= p(~j) for ~j ~ rz • (lla)
445
Nh
2:
i=l
for :5j 6 r 1 • (llb)
In this way we have formed a linear system of Nh equations

with Nh unknowns which we can write in matrix notation as
(12)
In spite of the fact that the basis functions ~hi have local
support, the formulation of the BEM leads to fully populated
matrices Ah, contrary to what happens with the standard fi-
nite element method.
THE TWO TEST CASES
To test the error and order of convergence of the BEM we

chose two problems proposed by Schultz [10.[. These consist
of solving the Laplace equation in the square with mixed boun
dary condition, which we state as follows.
Problem 1
lap u = 0 inn= { (x,y) : xE(0,7!), yE.(0,7!) }, (13a)
u = 0 on rJ. = { (x,y): xe[t,1f), y = 0}, (13b)
u = ~ 1 on rJ.' = { (x,y) :x E[0,1f], y = 1f}, (13c)
av u = 0 on rz = { (x,y) : x = 0 or x = 1f and
Y'-[o,1f] }, (13d)
av u = 0 on rz'= { (x,y): xE(o,;J. y = 0}. (13e)

The function ~l was calculated by fitting the numerical data
of [10}.
Problem 2
lap u 0 in n = { .. (x,y) : x'(0,7T), yE.(0,1f) } ' (14a)
u Oonri={ (x,y) : x=O or x = 1f
and Y '(o,1fj) (14b)
u = 0 on r" = {(x,y) :xd;,1f],y=O}, (14c)

1
d u = 0 on r2
1f
{ (x,y) :xt(0,2),y=O }. (14d)
v
The function ~2 was also computed by fitting the numerical data
given in [10].
We note that the derivatives of u show a singularity at
point (; ,0). As Schultz [10] pointed out with regard to the
FDM and we shall confirm for the BEM, this has a detrimental
446
effect on the accuracy of the aproximate solutions.
RESULTS AND COMMENTS
Using the linear and parabolic elements as described in the

appendices we computed the BEM approximated solutions to Pro-
blems 1 and 2 for several meshes. A typical sampl~ of the BEM
results is compiled in Tables 1 and 2. Table 3 shows the re-
sults obtained by the FEM for Problem 2.
The computer program developed employed everywhere single
precision arithmetic (32-bit words). The final system of equa
tions (12) was solved by Gauss condensation with full pivo--
ting using a standard subroutine (GELG in [13}).
Figures la and lb present typical meshes, We note the pre
sence of double nodes that allows one to have u and 3v u si--
multaneously unknowns at some points. This is possible since
either these nodes are placed at corners of r1 where the nor-
mal v suffers a discontinuity and so does 3v u, or they are
placed at the singularity (~,0) where one can form by colloc!
tion one equation of type (lla) and another of type (llb). In
both cases this procedure gives rise to two distinct rows of
matrix Ah in (12),
The-type of singularity is known for the two test problems
and the development of special elements incorporating func-
tions modelling the respective behaviour is therefore possible.
Nevertheless, we used a standard program, so we had to resort
to mesh grading to investigate if this way we could improve
the order of convergence of the error and namely if the opti-
mal order of convergence could be recovered,
Mesh grading was accomplished by decreasing the distance
between consecutive nodes on the face y = 0 (see Figure lb) in
a geometric series of ratior. This is in contrast with the ari
thmetic spacing employed by Symm [12] , the change being motiva
ted by the fact that this la~er type of grading tends to gene~
rate uniform meshes when the number of nodes becomes great.
Since we wanted to determine how mesh grading would improve
the order of convergence of the BEM, the arithmetic spacing
would therefore confuse the issue. On the other hand however,
when the number of nodes increases, the geometric grading leads
to a strongly non-uniform mesh. As an example, for IT/32 and
r = 0.6 the ratio of greatest distance to smallest distance
between consecutive nodes becomes equal to 0.616 = 3500, which
is very large indeed and caused ill-conditioning of matrix
Ah of system (12). Those runs that were unsuccessful due to
this phenomenonare marked with an asterisk (*) in Tables 1 and
2.
Let us analyse briefly the results presented in Tables 1
and 2.
a) Problem 1, linear element

The results for the uniform mesh show a clear o(h) convergence
similar to the FDM. However the magnitude of the error of the
447
BEM is about half of that obtained with the FDM of Schultz [10].
When grading with r = 0.8 is used the order of convergence
improves significantly to o(h2) which is the optimal for line-
ar elements. The magnitude of the error decreases also at all
points shown except for the point(~ ,0) and meshes h =-¥-
and h = ~· This we attribute to the fact that this point is
no longer a node when mesh grading is employed and the respec-
tive values of the unknown have to be obtained by interpola-
tion. For meshes with a large mesh size h this may offset the
benefit from the grading.
The results for grading with r = 0.6 show a similar behav-
iour convergence of o(h2) everywhere except-at a point ( ~ ,0)
where the values obtained do not seem conclusive. We remark
that mesh h = : 2 is strongly non-uniform and ill-conditioning
prevented us from obtaining the respective results.
That pattern is even more evident in the case of r = 0.4.
Now even the convergence order seems to be less than O(h2). As
ill-conditioning must be even greater than before we suspect
that round-off errors now play an important part.
b) Problem 1, parabolic element

The uniform mesh yields an o(h) convergence as before,
with the magnitude of the error marginally smaller than for
the linear element.
When grading r = 0.8 is employed results show a very mar-
ked improvement, with a sizeable decrease in the error and con
vergence order approaching the optimalo(h3).
For stronger grading r = 0.6 and r = 0.4 the results shownare
too scarce to allow conclusions. Nevertheless case r = 0.6
still exhibits an o(h3) convergence except perhaps at point
(~ ,0) and the error is still less than for the other previous
meshes. For grading r = 0.4 we still have gains at some points
but at the expense of greater errors at others and an order of
convergence probably less than optimal.
c) Problem 2, linear element

For the un~form mesh we have also a o(h) order of convergence.
For points away from the singularity however this seemsto in-
crease slightly. The error is smaller than for methods 1 and 2
of Schultz, method 3 performing better near the singularity.
For grading r = 0.8 we get an order of convergence roughly
close to o(h2) with a smaller error magnitude, which is more
marked the finer the mesh.
Results for grading r = 0.6 confirm this behaviour. As be-
fore ill-conditioning prevented us from obtaining results for
II
mesh h = 32 •
Again tne values presented for grading r = 0.4 do show so-
me improvement but the order of convergence seems to be less
448
than optimal.
d) Problem 2, parabolic element

The uniform mesh exhibits a sharp o(h) convergence with an
error magnitude somewhat smaller than that for the correspon-
ding mesh with the linear element.
Grading with r = 0.8 shows an improvement of the order of
convergence but less than optimal, except for points away from
the singularity.
The case r = 0.6 seems to yield an optimal order of con-
vergence of O(h3) with a marked reduction in the magnitude of
the error.
The results for grading r = 0.4 do not show a systematic
behaviour. In fact, at points near the singularity the errors
TI TI
for mesh h = ~ are worse than for the mesh h • -s-;we guess
that round-off due to ill-conditioning is responsible for this
situation.
From the evidence presented in this paper we draw the
following conclusions. The BEM tends to yield smaller errors
than the corresponding FDM schemes, requiring however a grea-
ter computational effort for the same accuracy. Mesh grading,
judiciously employed, can allow us to recover the optimal
order of convergence in cases where singularities are present.
The development of special elements may not be necessary.
As compared with the FEM (Table 3) our results point to a
roughly similar accuracy for the first order elements. For the
second order elements the BEM seems to provide greater preci-
sion. In both cases the FEM required more computing time than
the corresponding BEM runs.
APPENDIX A. The linear Isoparametric Element
In this appendix we sulllillarize briefly the main aspects of the

linear isoparametric element [4,14] implementation used to
obtain the results presented above.
Base functions. The reference element is the interval [-1,1]

equipped with the basis functions
1 1
N1<0 = 2 (1 - ;) , N2 (;) 2 (1 + ;) • (A.la,b)
The coordinate transformation is x = N(s) with
N : (-1,1] + fa E. R2 - - - (A.2a)
and is given explicity by the relations
2 2
x=~N.() xi' y=LNi(s) y., (A.2b,c)
i=l ~ i=l ~
where (xi; Yi) are the global coordinates of node i of element
a. The Jacobian J of transformation (A,l) is
J(;) = 1<~)2
ds
+ <~)211/2
ds
(A.3)
449
The linearity of equations (A.l) imply that this Jacobian is

constant in each element and ra is a segment of a straight
line.
Computation of regular integrals. Let I($) be the integral
!
over ra of a bounded function $. Then
I($) J $(~) dr =
1
¢<;> J(;) d; J Jl ¢(;) d;, (A.4a)
ra -1 -1
with
¢(;) $(x) = $ (N(;)) (A.4b)
The integral on the right hand side of expression (A.4a) can
be evaluated numerically. We employed a two-point Gauss-Lege~
dre quadrature rule [1] to form the linear system of equations
(12). Later to compute values at interior points and in order
to overcome the difficulties already mentioned we resorted to
an adaptive integration scheme based on the following heuris
tic approach. Let ha be length of element a and define oa(~)
by
(A.5)
where x are the coordinates of the interior point under consi
deration and ~l and ~2 are the coordinates of nodes 1 and 2
of element a. In a sense oa(x) measures the mean distance from
point ~ to element a. We compare this distance oa(~) with ha
and determine the number na(x) of Gauss-Legendre points to use
to integrate on ra by the formula
oa(~)
na(~) =min (2 + 5ka(~), 5) with ka(~) = h (A.6)
This criteria resembles that employed by Hess [6J. Remembering
the well known properties of the ellipse it can be easily
shown that the points'satisfying
oa(x)<kha (A.7)
lie within an ellipse with focii on the nodes 1 and 2 of ele-
ment a and semi-axes k ha and (k2 - 1/4)1/2 ha.
Computation of integrals with a logarithmic singularity. As

shown above we need to evaluate integrals of the form
Ia($) = J ln r(~) $(~) dr, (A.8)

ra
wherer(x) is the distance from a node of element to a generic
point x-on the same element. This integral can be transformed
to
Ia($) = Jl ln r(N(;)) ¢(;) J(~) d; , (A. 9)

-1
Now, for the sake of definiteness, suppose that the singulari
ty sits on node 1 of element a. Then (A.9) can be transformed
to
450
Ia(¢) = ! 1
-1
ln 2 1 r~s~ ~(s) J(s) ds- !1
-1
ln( 1 ;s) •
<P<s> J(s) ds • (A.lo)

The integrand in the first integral of this expression can be
easily shown to be bounded and therefore we are in condition
to use the Gauss-Legendre quadrature points as before. The s~
cond integral in (A.lO) has an integral with a logarithmic s1n
gularity at s= -1. Using the transformation
s = l + s (A,ll)
2
it can be put in the form
!1
-1
ln( 1 ; s ) iP<s> J(s) ds = 2!0
1
ln ~ ¢<s> J(~) ds •
We evaluate this integral by applying to the right hand side

(A,l2)
of (A,l2) the quadrature rule of Berthod-Zaborowiski [9] with

two integration points.
If the singularity sits on node 2 of element a we use a si
milar procedure with obvious adaptations.
APPENDIX B. The Parabolic Isoparametric Element
Here we present briefly the main aspects of the parabolic iso

parametric element implementation [4,14] employed to obtain-
the results presented in this paper.
Base functions, The reference element in this interval [ -1 '1]

equipped with the basis functions
1
N1<s> :r (1 - s) s , (B .la)
N2 <s> = 1 - s2 , (B, lb)
N3<s> = :r <1 + s> s •
1
(B .lc)
The coordinate transformation is ~ = ~(f) with
N: [-1,1) + fa&R2 (B .2a)
and Is given explicity by the relations
3
X=~ Ni(s) Xi , (B.2b)
1=1
3
Y = ~ Ni(s) Yi • (B .2c)
i=l
where, as before, (xi, Yi) are the global coordinates of node
i of element a. The Jacobian J of transformation (B.l) is also
given by (A.3) but, contrary to the linear element case, is
no longer a constant, and fa can be in general a segment om a
curve whose parametric equation are (B.2). However the value
of the Jacobian J at a point s depends on the set of nodal
coordinates (xi, Yi), i=l to 3. For some choices of these coor-
451
dinates the Jacobian vanishes and the transformation (B.2)

ceases to be one to one [3 ,5]. This fact has been used in the
FEM to model the behaviour of the singularity.
Computation of regular integrals. The computation of regular

integrals proceeds as in the case of the linear element, but
now we employ three Gauss-Legendre integration points [1] in~
tead of two. To compute values at interior points the adaptive
integration scheme described in the previous appendix was chan
ged slightly to adapt to this type of element as follows. -
0 a(~)= ~ ( ~~-~11 + ~~-~3~ ) • (B.3)
where x1 and x3 are the coordinates of the extreme nodes of
element a. The number of integration points is now chosen
according to
na(x) = min(3 + 5ka(~), 5) • (B.4)
with katx) given by
ka(x)
-
=-
oa. (x)
ha.
(B.5a)
where ha is the length of the straight line segment connecting

nodes 1 and 3.
Computation of integrals with a logarithmic singularity.

The evaluation of integrals with logarithmic singularity pro-
ceeds along the same lines as for the linear element with the
necessary adaptations. We employed now a quadrature rule of
Berthod-Zaborowiskf [9] with three integration points.
AKNOWLEDGEMENT
This work was supported by CTAMFUL, Universidade Tecnica de

Lisboa and by NATO Research Grant No. 1964.
REFERENCES
1. M. ABRAMOVITZ and I.A. STEGUN, Handbook of Mathematical func-

tions, Dover (1968).
2. K.E. ATKINSON, A Survey of Numerical Methods for the Solu-
tion of Fredholm Integral Equations of the Second Kind, SIAM
(1976) •
3. R.S. BARSOUM, On the Use of Isoparametric Finite Elements
in Linear Fracture Mechanics, Int. J. Num. Engng, 10, 25-37
(1976). --
4. C.A. BREBBIA, The Boundary Element Method for Engineers,
Pentech Press, (1978).
5. R,D, HENSCHELL and K.G. SHAW, Crack Tip Elements Are Unne-
cessary, Int. J. Num. Engng., ~. 495-507 (1975).
6. J.L. HESS and A.M,O, SMITH, Potential Flow About Arbitrary
Bodies, in Progress in Aeronautical Sciences, vol. 8, Per-
gamon (1967 •
452
7. Y. IKEBE, The Galerkin Method for the Numerical Solution of

Fredholm Integral Equations of the Second Kind, SIAM Review,
14, 3, 465-491 (1972).
B. M.A. JASWON and G.T. SYMM, Inte ral E uations Methods in
Potential Theory and Elastostat~cs, Academ~c Press 1977).
9. H. MINEUR, Techniques de Calcul Numerique, Dunod (1966).
10. D.H. SCHULTZ, Two Test Cases for the Numerical Solution of
Harmonic Mixed Boundary Value Problems, J. Inst. Maths.
Applies, .!:2_, 1-8 (1975).
11. G. STRANG and G.F. FIX, An Analysis of the Finite Element
~. Prentice-Hall (1973),
12. G.T. SYMM, The Robin Problem for Laplace's Equation, in
New Developments in Boundary Element Method, Proceedings
of the Second International Seminar on Recent Advances in
Boundary Element Methods, ed. by C.A. Brebbia, CML Publi-
cations (1980).
13. System/360 Scientific Package, version III, Programmer's
Manual, IBM (1970).
14. O.C. ZIENKIEWICZ, The Finite Element Method in Engineering
and Science, McGraw-Hill (1971).
15. O.C. ZIENKIEWICZ, D.W. KELLY and P. BETTESS, The Coupling
of the Finite Element Method and Boundary Solution Proce-
dures, Int. J. Num. Meth. Engng., ll• 355-375 (1977).
TABLE 1 - RESULTS FOR PROBLEM 1
(Point) Mesh BEM - Linear Element BEM - Parabolic Element I

Exact Value Size h
r = 1. r = 0.8 r = 0.6 r = 0.4 r = 1. r = 0.8 r = 0.6 r = 0.4,
n/4 0.133 0.152 0.175 0.208 I
- - - -
(n/4,0) n/8 0.077 0.078 0.045 0.074 0.071 0.044 -0.014 0.040
1.5286 n/16 0.042 0.025 0.023 0.054 0.037 0.015 -0.009 0.023
TT/32 0.022 0.003 0.019 § 0.019 0.002 -0.012 §
* *
n/4 0.026 0.021 0.014 0.006 - - - -
( Tf '3TT I 4) n/8 0.012 0.008 0.004 4x1o-4 0.010 0.006 0.002 -0.001
2.9570 n/16 0.006 0.002 -2xlo-4 -0.001 0.005 0.002 -2xlo-4 -0.007
n/32 0.003 lxlo-4 -4x1o-4 § 0.002 lxlo-4 -4xlo-4 §
* *
n/4 0.074 0.064 0.052 0.037 - - - -
(n/2,n/2) n/8 0.036 0.024 0.014 0.009 0.028 0.018 0.008 0.002
2.2530 n/16 0.018 0.007 0.003 0.004 0.014 0.005 0.001 5xlo-4
n/32 0.009 0.001 0.001 § 0.007 0.001 lxlo-4 §
* *
n/4 0.128 0.113 0.093 0.068 - - - -
(n/2,n/4) n/8 0.063 0.043 0.025 0.015 0.050 0.031 0.014 0.004
1.4220 n/16 0.031 0.012 0.004 0.006 0,024 0.009 0.002 -lxlo-4
TT/32 0.015 0.002 0.002 § 0.012 0.001 o.o §
* *
TT/4 0.184 0.158 0.135 0.103 - - - -
(n/2,n/8) n/8 0.103 0.065 0.037 0.019 0.076 0.048 0.022 0.005
0.9452 n/16 0.055 0.018 0.006 0.006 0.036 0.013 0.002 -0.001
n/32 0.023 0.003 0.002 § 0.018 0.002 -lxlo-4 §
* *
§ Solution of system (12) in double-precision.
* Elements too small. Underflow during formation of matrices H and G .I>.
of small elements. ~
.!>.
~
TABLE 2 - RESULTS FOR PROBLEM 2
(Point) Mesh BEM - Linear Element BEM - Parabolic Element

Exact Value Size h
r = l. r = 0.8 r = 0.6 r = 0.4 r = 1. r = 0.8 r = 0,6 r = 0.4
7f/4 0.165 0.150 0.129 0.102 - - - -

(7r/2,7f/8) 7f/8 0.084 0.058 0.033 0.018 0.066 0.040 0.018 0.002
1.0924 7f/16 0.040 0.016 0.005 0.006 0.030 0.011 0.002 -0.003
7f/32 0.019 0,003 0.002 § 0.015 0.002 1xlo-4 §
* *
7f/4 0.109 0.097 0.083 0.065 - - - -
(7r/2,7f/4) 7f/8 0.048 0.033 0.020 0.013 0.034 0.021 0.009 0.001
1.9520 7f/16 0.022 0.099 0.004 0.005 0.016 0.006 0.001 -0.002
7f/32 0.010 0.001 0.002 § 0.008 0.001 0.0
* *
7f/4 0.058 0.052 0.046 0.039 - - - -
(7r/2,7f/2) 7f/8 0.020 0.015 0,010 0.007 0.013 0.008 0.004 0.001
4.7054 7f/16 0.009 0.004 0.002 0.003 0.006 0.002 3x1o-4 -3x1o-4
7f/32 0.004 0,001 0.001 § 0.003 3x1o-4 -lx1o-4 §
* *
7f/4 0,037 0.035 0.032 0.030 - - - -
(7r/2,37f/4) 7f/8 0.010 0,007 0.006 0.005 0.004 0.002 lx1o-4 -0.001
10.5032 7f/16 0.003 0.001 4xlo-4 0,001 0.002 4xlo-4 -3x1o-4 -4x1o-4
7f/32 0.001 -2x1o-4 -1xlo-4 § 0.001 -3x1o-4 -4xlo-4 §
- -
*
------- -- *
TABLE 3 - RESULTS FOR PROBLEM 2 WITH FINITE ELEMENTS
(Point coord,) MESH

EXACT VALUE SIZE
6 D ~ D
(tr/2, tr/8) tr/4 0.0926 0,2269 0.42315 0.3336
1.0924 tr/8 0.1041 0.0695 0.26922 0.1072
tr/16 0.0594 0.0390 0.17758 0.0143
TT /32 0.0302 0.0189 0.12428 0.0211
(tr/2, tr/4) tr/4 -0.0472 0.2211 0.393 0.0726

1.9520 tr/8 0.0195 0.0789 0.2639 0.001
tr/16 0.0212 0.0299 0.1209 0,0239
tr/32 0.0139 0.0126 0.0912 0.008
(tr/2, tr/2) tr/4 -0.2861 0.3739 0.7054 -0.501

4.7054 tr/8 -0,0639 0.1002 0,1101 0.0133
tr/16 0.0097 0,0291 0,055 0.0059
tr/32 0.0009 0.0094 0.042 0.0027
(tr/2, 3tr/4) tr/4 -0.3708 0.4412 0.0318 0.5893

10.5032 tr/8 -0.0928 0.1082 0.0602 -0.0498
TT/16 -0.0218 0.0282 0.0212 2x1o-4
tr/32 -0.0048 0.0072 0.0152 2x1o-4
m
456
(0,11) r-~--~--~--~-+--~--+-~ (11,11)
+ (11/2,311/4)
+ (11/2,11/2)
+ (11/2,11/4)
+ (11/2,11/8)
(0 ,0) ( 11 '0)
a) Uniform mesh (h=11/8)
(0,11) (11,11)
+ (11/2,311/4)
+ (11/2,11/2)
+ (11/2,11/4)
+ (11/2,11/8)
(0,0) ( 1r '0)
b) Mesh (h=11/8) with grading r=0.6
Figure 1. Typical meshes for Problem 2

457
BOUNDARY ELEMENT & LINEAR PROGRAMMING METHOD IN OPTIMIZATION

OF PARTIAL DIFFERENTIAL EQUATION SYSTEMS
T. Futagami
Hiroshima Institute of Technology, Itsukaichi, Hiroshima, Japan
ABSTRACT
By combining boundary element method with linear programming,

a new optimization technique (boundary element & linear
programming method, or, the BE&LP method) is developed and
systematized in order to control partial differential equation
systems with both equality or inequality constraints and an
objective function. The BE&LP method is applied to optimal
control problems jn heat conducti-on phenomena. Minimization of
total of the controllable loads to meet with temperature
requirements is performed by using simplex method of linear
programming. The tractability in both the boundary conditions
and the equality or inequality constraints makes sure that the
method becomes a powerful technique for several new types of
boundary value problems.
INTRODUCTION
By combining finite element method with linear programming,

finite element & linear programming method (the FE&LP method)
has been developed in order to control partial differential
s~ems (Futagami (7, 8, 9)). The combined use of finite
difference method with linear programming has been studied
in the field of ground-water management by Aguado and Remson
( 1) •
In this research, by combining boundary element method with

linear programming, boundary element & linear programming
method (the BE&LP method) is developed and systematized in
order to control partial differential equation systems with
both equality or inequality constraints and an objective
function. Such systems are frequently encountered in various
engineering and scientific problems of control and optimal
design.
458
Boundary element method, increasingly used in Europe, is a

powerful numerical method for the solution of partial
differential equations because of its generality with respect
to geometry, its simplicity of the input data required, and its
numerical accuracy (2, 3, 4). Linear programming is one of the
most frequently used mathematical methods of operations
research (5,10). In the development of the BE&LP method the
concepts of the decision variable and the state variable are
adopted as in Bellman's dynamic programming and Pontryagin's
maximum principle. The BE&LP method utilizes the advantages of
established numerical techniques of both boundary element
method and linear programming.
In order to show the applicability of the BE&LP method, the
method is applied to optimal control problems in heatcondudion
phenomena. Numerical examples are performed in a simple model.
The method may become a useful technique for problems of
control and optimal lesign in various fields of engineering
and science.
BOUNDARY ELEMENT & LINEAR PROGRAMMING METHOD
Systems of Basic Differential Equations

Boundary element method & linear programming method (the BE&LP
method) is developed to solve the following systems of
differential equations with both equality or inequality
constraints and an objective function (see Fig. 1).
Objective Function (throughout the whole domain Q)
( {¢} ' {q}' {9})

z Opt. f ({cj>}, {q}, {8}) ~ax.
8} f (l)
{8}
({cj>}, {q}' {8})
{~}· f
subject to:
Equilibrium Equations
Governing Differential Equation (in the whole domain Q)
D.E. (xk, cj>, ~!k' · • ·, ~~' 8) = 0 (2)
Boundary Conditions ( on the boundaries r)

( 3)
Constraints (in the sub-domains Qr)
r (xk, cj>, q, 8) >< 0 (4)

in which <P = the state variable; 8 = the decision variable;
xk = Cartesian coordinate (x, y, z); Xk = Cartesian coordinate
of the boundary (X, Y, Z);and q =the flux.
459
¢ )=0
D. E.: Governing Equation
0 e Boundary Condition
II
r Constraint
Q:
f Objective Function
Z=Opt.f({¢},{8})
n=N-1
~~
D • E • ( x, y, <jJ '"axay
," ,
.. a~ - '" n '" Th 8 ) -0
<P
e
State Variable
Decision Variable
-- '
oX'ay
n=N rr. ', ........... ....... S1 Whole Domain

"2· < \ /
/
r '
2 (x,y,8)~ \ I f<3" < r Boundary
n=l I J}x,y,¢,8)~ S1r: Sub domain
n=21
Associated with
Constraint
Fig. 1. General Concepts of the BE&LP method
In systems governed by two-dimeLsional second-order

differential equations for example, Eq. 2 is expressed as
follows:
c 13 r- +
32<P ~ 32¢ .21
c 23 x<ly + c 3 3y2- + c 4 3x + c 5 3y + c 6 ¢ +
~
e + b = o ( 2)~
The examples of the boundary conditions are as follows:
<jl(x=X, y=Y) = $ \X,Y), ~~ (x=X y=Y)

3n '
As for the constraints, the following simple inequalities are
frequently encountered.
t:
~.!
.! ~ <P ~ ¢, or,
~"¢"
(4)~
~~
~~e ~ e, or,
{: ~e
in which ¢ = the lower limit of the state variable; "¢" = the
upper limit of of the state variable; 8 = the lower limit of
the decision variable; and e
= the upper limit of the decision
variable.
Formulation of Boundary Element & Linear Programming Method

Boundary element method is used in order to discretize the
above-mentioned systems as systems of linear algebraic
equations. (As for the details, see the next chapter). Then,
the following matrix-vector forms of the BE&LP method are
obtained and the application of linear programming is possible.
460
Objective Function (throughout the whole domain n)

Z =Opt. f ({x }, {8.}) = Z =Opt. f ({<f> },{a },{8.})(5)
{8 } n 1 { 8 .} n 11 1
i 1
subject to:
Equilibrium Equations (N-Eqs.)
[H){<f> _} = [G]{q } + [D){8.} + {p } , or, [A){x } - [D){8.} = {b }(6)
NxN n NxN n NXI 1 n NXN n NXI 1 n
Constraints (L-Eqs.)
[R] {{xn}} ~ {!l } (7)

Lx(N+I) {8i} > z
in which Xn = nth boundary variable; <Pn = nth state on boundary
elements; 8i = ith decision variable; qn = nth flux
on boundary elements; [A] = matrix derived from boundary
element method; [H) = the state-matrix derived from boundary
element method; [G) = the flux-matrix derived from
boundary element method; bn = constant in nth equ1iibrium
equation; P.n = constant in nth equilibrium equation; [R] =
the constraint matrix, sparse matrix; Bz
= constant in lth
constraint; n = 1 ~ N (N: total number of boundary elements);
i = 1 ~I (I: total nu~ber of the decision variables); Z = 1 ~
L (L: total number of the constraints).
Therefore, the BE&LP method is one that optimizes the objective
function under the conditions of the equilibrium equations and
the constraints. In the sense of general linear programming,
the equilibrium equations of the BE&LP method are also the
constraints. The solutions of the BE&LP method are obtained by
the simplex method for linear programming.
HEAT CONDUCTION CONTROL BY THE BE&LP METHOD
Systems of Basic Equations in Heat Conduction Control

Heat conduction control by the BE&LP method is studied in order
to show the applicability of the method (see Fig. 2). The b~
equation systems of two-dimensional heat conduction phenomena
with constraints are as follows:

Z =Opt. f ({<f>}, {q}, {8}) =Min. E 8 (8)
{8 l {8}
subject to:
Equilibrium Equations
Governing Differential Equation
KV 2 <f> + 8 + Q= 0 in n (9)
461
Boundary Conditions
<P = $ on r 1 (10)
q = - K~ = q on r 2 (11)
q y(cp- cps) on r 3 (12)
Constraints
<P 2:,_! in rl1 and on r2 + r3 (13)
8~8 in r{ (14)
2
Non-negative Conditions
<P ,;:,o in rl1 and onr 2 + r3 (15)
8 ~0 in rl2 (16)
in which <P =the state (temperature); q =heat flow (the out-

flux (+) or the in-flux (-)); 8 =the decision variable
(controllable load of heat generato,·); $ = the prescribed
temperature; q = the prescribed out flux; <Ps = the surrounding
temperature; K =the conductivity coefficient; Q =the heat
generated in the body (uncontrollable load); y =the convection
coefficient; ~ = the lower limit of the state (temperature
requirement);-~= the upper limit of the decision variable
(the upper limit of the controllable load); n =the whole
domain; r = r 1 + r 2 + r 3 = the boundaries; n:r = subdomain
associated with the state-constraints; n~ = subdomain
associated with the decision-constraints.
The boundary condition expressed in Eq. 12 is encountered when
heat is lost from the surrounding media by convection. The
heat flux is assumed to depend linearly on the temperature
difference (cp- <P ).
s
Although, the objective function may be composed of the
temperature distribution {cp}, the heat flows {q}, and the
controllable loads {8} in general, the minimization of the
total of the controllable loads L 8 is sought in this study.
As for the constraints, although only the lower limit of the
state ~ and the upper limit of the controllable load are e
imposedinthe above systems, the upper limit of the state and
the low·er limit of the controllable load may occasionally become
necessary with respect to the conditions of the problems.
Formulation of the BE&LP Method in Heat Conduction Control

In order to discretize the above mentioned differential
systems, boundary element method is used. As for the details
of boundary element method, one may follow Brebbia (3).
Boundary element method in heat conduction problems has been
462
studied by many :re3earchers (2, 4, 6). By using the vcightcd

residual method, theequilibrium equations (Eqs. 9-12) are
replaced by the following integral formulation.
{¢ - $}K~ di'
an
+ fr {q - q}¢* ar + fr [q - y{¢ - <P } ]¢* ar = o (17)

2 3 s
~~e weighting function¢* has continuous first derivatives.
Later this function will also be required to satisfy the
governing differential equation (Eq. 9). Integrating the
first term of Eq. 17 by parts twice, we obtain:
Jn KV 2 ¢*¢ drl + Jn (e + Q)¢* drl - Jr1 q¢* ar - fr 2 q¢* ar

- f y(¢ - cp )¢* ar + Jrl $q* ar + J cpq* ar = o (18)
r3 s r2+r3
where
q =- K~,
The weighting function is now taken as the fun~~ental
solution to Laplace's equation, which is the solution
corresponding to a unit point source, i.e.
(19)
where ~n is the Dirac delta function and n is the point where

the source acts. The fundamental solution¢* has the
following property
(20)
where <Pn represents the value of the unknown function ¢ at the

point where the source is applied. The fundamental solution
to two-dimensional Laplace's equation is given by
l l
¢* = - ln- ( 21)
27T r
where r is the distance from the point of application of the
unit source to the point under consideration. Substitution of
Eqs. 20 and 21 into Eq. 18 yields the following equation •
. . ·n r·n 1
K:-2-
r
ar + ;:r 2+r 3 <PK-
r2
df = fr q ln- di'
r
1
1
+ fr q lJ:. ar + ;:r y¢ln;:- df
2 r 3
( 22)
463
where r is the vector from the point where the source is

applied to the point on the boundary and n is outward unit
normal from the point on the boundary.
If 8 (the decision variable, i.e. controllable load) or
Q (constant, i.e., uncontrollable load) is not equal to zero,
then the whole domain n has to be divided into a series of
cells, so a numerical integration can be performed. But in the
case all of the loads are point sources (concentrated loads),
we just have to subtract I;8.lntl/r) andLQ ln(l/r) from the
right hand side of Eq. 22. 1 8. means thescontrollable point
load at internal or boundary ~oint 'i', and Q means the
uncontrollable point load at internal or bounaary point 's'.
From now on in this study we consider only such point loads.
Eq. 22 is now applied on the boundary of the domain under
consideration. Although several types of boundary elements
h~ebeen presented (3), constant boundary element method is
used in this study (see Fig. 2). In the constant boundary
element method the values of ~ and q are assumed to be constant
on each element and equal to the values at the mid-node of the
element. Therefore, Eq. 22 is rewritten by the following
Equation.
- TIK~
n ~r r.n
+ 'f'J·r K - a + ~
fr r.n f
K--::T df + ~ r
r.n - yln-
( K--::T 1) df
1 r2 2 r 3 r r
qfr ln! ar + qfr 1J:.

r
a.r - y~ s !r ln!
r
ar
1 r 2 3
( 23)
Then, Eq. 23 for a given node 'n' becomes in the following

discretized form:
N I
j~l gnjqj + i~l dni 8 i + Pn ( 24)
where
f Kr·n df (for 'j' on r 1 + r2)
finj r r2
e
B. .
r·n 1
fr ( K-=z - yln-) df (for 'j' on r 3)
nJ e r r
gnj fr ln! df (for 'j' on r 1 +r 2)

e r
gnj 0 (for 'j' on r )

3
d . ln-L
r1
n1
!1
s ln.l.. -y~s fr ln! ar
Pn sh Qs rs
n 3 r
464
Therefore, the whole set of the equilibrium equations for the

N nodes on the whole boundary is expressed in the following
matrix-vector form:
[H]{<jl } [G]{q} + [D]{e.} + {p} ( 25)

NxN n NxN n NXI 1 n
where
= nnj - 'TTK (n = j)
= nnJ. (n F j)
Ey noting the fact that N1 values of <P and N2 values of q are

known on the whole boundary, Eq. 25 is rewri~ten in such a way
that all the unknown; are on the left hand side as follows:
[A]{x } - [D]{e.} = {b } ( 26)

NxN n NxN 1 n
where {x } is the vector of unknown <P and q.
n
Then, the discretized form of the equilibrium equations are
obtained :in Eq. 25 or 26.
As for the state-constraints (Eq. 13), the following di.s.creti.zed
inequalities are obtained.
<Pz = 'TTlK (j~l 0 lj<jlj - j~l Pz) ~ .!t

(for 1 Z1 point ( 27)
(for 1 Z I mid-node on r 2+ r 3) ( 28)
As for the de cis ion-constraints ( Eq. 14) , the following

inequalities are obtained for the points fitted for the
~ontrollable loads (the decision variables)
ei ~ei (for 1 i 1 point inn~, i = 1 ~I) (29)
Then, the matrix-vector forms of the BE&LP method in the h8at
conduction control problem are obtained as shown below:

z =Opt. f ({x }, {e 1.}) Opt. f ({<P }, {a}, {e.})
{e.} n {e.} n 11 1
1 1
(30)
subject to:
465
Equilibrium Equations (N-EQs.)
[H]{~ } = [G]{~} + [D]{9.} + {p } , or, ( 31)

NXN n NXN Nxi ~ n
[A]{x } - [D]{e.} = {b } ( 31 )~
NxN n Nxi ~ n
Constraints ( (L=L~+I)-Eqs.)
{~z}=([BljH~j}- [glj]{qj}- [dliHei}- {pz}):K 2:, {!ZJ

(for interior points in n~) (32)
(for mid-nodes on r 2 + r 3 ) ( 33)
(34)
~n ~. 0 (n =1 ~ N), e.
~
>
=
o (i = 1 ~I) ( 35)
NUMERICAL EXAMPLES IN A SIMPLE MODEL
In order to clarify the features of the BE&LP methods, the

systems of the governing equations are written down for a
simple square model (see Figs. 3, 4(a) and 4(b)).
The obtained matrix-vector forms of the BE&LP method in Run 1
are as follows:
Objective Function
2
Z = Min. i~l 6i = Min. (6 1 + 62 ) (36)
{ei}
subject to:
Equilibrium Equations (4-Eqs.)
[-i:i~ -~:it:~:~~ ~:~~]\:~} = ~-8.74

-0. 59 -8.74 0
-0.59 G 0 -10.98
-8.74]{0)
0
0.93 1.11 -2.55 1.11 10.98 ~3 -8.74 0 -8.74 q3
1.11 0.93 9.85 -3.14 ~4 -8.74 -10.98 0 -0.59 0
1.10 1.50] { 2.03 }

+ [ 1.10 0 1.21 {e1 } + 5.49 ( 37)
1.10 o.41 e2 7.52
1.10 1.21 5.49
!, or,
466
l -3.14 l.ll 11.91

l.ll-3.14A9.8 5 0.93
0.93 1.11 -2.55 1.11
1.11 0.93 9.84 -3.14
l.l~j<h}
cj>2
cp3
cp4
_ ~1.10
1.10
1.10
1.5~ {81}= \2.03}
1.10 0 1.21
0.41
1.21
82
5.49 ( 37 r
7.52
5.49
Constraints (4-Eqs.)
[~ :] 1m ( 38)
[ ~ R8 ~ J {~~} (39)
cj> > 0 (n = l ~ 4), 8. > 0 (i =l ~ 2) (40)
n= l =
Computed results for the boundary temperatures (¢>1, cj>2, cj>3, <Jl4),
the heat flux (q,), the controllable loads (8 1 , 8 2 ) and the
optimal objectiv~ (Z) are shown in Figs. 5(a) and 5(b). It
should be noted that the distribution patterns of boundary
temperatures {cj>n} in Figs. 5(a) and 5(b) arise from the
resultant loads composed of the obtained controllable loads
{8i} and the given uncontrollable load Q1 .
CONCLUSIONS
Boundary element & linear programming method (the BE&LP method)

was described through the application to the optimal control of
heat conduction phenomena. The tractability in both the
boundary conditions and the equality or inequality constraints
makes sure that the method becomes a powerful technique for
various problems of optimal control and design.
Finally, the problems to be attacked from now on in the
applications of the BE&LP method should be mentioned.
Extension of the method to the time domain is necessary to solve
transient problems. The related methods such as boundary
element & non-linear programming method, and stochastic
boundary element & linear programming method could be developed.
The developments certainly make it possible to solve more
complicated problems.
REFERENCES
l. Aguado, E. and Remson, I. (1974), "Ground-Water Hydraulics

in Aquifer Management," Journal of the Hydarualics Divis:ion,
ASCE, Vol. 100, No. HYl, pp. 103-118.
2. Bolteus, L. and Tullberg, 0. (1980), "Boundary Element
Method Applied to Two-Dimension al Heat Conduction in Non-
homogeneous Media," Advances in Engineering Software, Vol. 2,
467
No. 3, pp. 131-137.

3. Brebbia, C.A.(l978), "The Boundary Element Method for
Engineers," Pentech Press, London.
4. Brebbia, C.A. (1980), "Fundamental of Boundary Elements,"
New Developments in Boundary Element Methods, Proceedings
of the Second International Seminar on Recent Advances in
Boundary Element Methods, CML Publications, pp. 3-33.
5. Dantzig, B .G. ( 1963), "Linear Programming and Extensions,"
Princeton University Press.
6. Dubois, M. and Buysse, M. (19tl0), "Transient Heat Transfer
Analysis by the Boundary Integral," New Developments in
Boundary Element Methods, Proceedings of the Second
International Seminar on Recent Advances in Boundary
Element Methods, CML Publications, 137-154.
7. Futagami, T. (1975), "Finite Element & Linear Programming
Method and Water Pollution Control," Proceedings of 16th
Congress of the International Association for Hydraulics
Research, Vol. 3, c7, pp. 54-61.
8. Futagami, T., Tarnai, N. and Yatsuzuka, M. (1976), "FEM
Coupled with LP for Water Pollution Control," Journal of
Hydraulic Division, ASCE, Vol. 102, Hy7, pp. 881-897.
9. Futagami, T. ( 1976), "Several Mathematical Methods in Water
Pollution Control- THE FINITE ELEMENT & LINEAR PROGRAMMING
METHOD and the Related Methods," HIT-C-EH-1, Department of
Civil Engineering, Hiroshima Institute of Technology.
10. Gass, I. S. ( 1969) , "Linear Programming," 3rd ed, McGraw-Hill
Koga.kusha.
5)
00
<Ps N $ = Prescribed Temperature
<f>s= Surrounding Temperature
! = Lower Limit of Ten: perature
q =Heat flux (Out-flux (+))
rl(<fl=~l q = Prescribed Heat Flux
B = Controllable Load
Q = Uncontrollable Load
Qs n
n = Whole Domain
s~S K'i7 2 <f> + 8 + Q 0
nf= Sub-domain Associated
Ql with Temperature
Qs
y Requirements
.___rs
s=s n ~= Sub-domain Associated
r2(q=q)
LX with Controllable Loads
f = f1 ~ f 2+ f 3 =Boundaries
K = Conduction Coefficient
0 Source Point Fitted for Controllable Load
y Convection Coefficient
• Source Point Fitted for Uncontrollable Load
r = Distance
~: Regulated Point in Temperature Requirement
Fig. 2. The Boundary Element & Linear Programming Method in Heat Conduction Control
Planning Heat
Generator r-6m
0
K = 1.0 W/m°C
0 ov2 0
I I I y = 1.0 W/m 2 °C
6m I r -,--1 0 I o e, I 0 0 I
Q1 = 5.0 W/m
I I
• •"1 . • • cp = 0.0 °C
s
Area to be Existing Heat

Analyzed Generator
Fig. 3. Simple Model of Heat Conduction Control by Boundary Element & Linear
Programming Method
m
470
n=3
8~8 2 (=20.0W/m)
0 i=I=2
n=N=4
Q1 =5 .0 W/m
•s=S=1
q1 =0 ~1~ !1 ( =10 .0 °C)
Unknowns: ~ 1 .~ 2 .~ 3 .~ 4 ,q 3 ,e 1 ,e 2 and Z
Fig. 4(a). Input Data and UnknQwns in Run 1
q3=y( ~3 -~ s )=1x( ~3-0 ), ~3~$3 ( =3 .0 oc)

A
e~02(=20.0W/m)
0 i=I=2
8 1 ~8]_ (=20 .0 W/m)

0 i=1 n=2
Q1 =5.0W/m
• s=>S=1
q1 =o ~ 1 ~${;!_0.0°C)
Unknowns: ~ 1 .~ 2 .~ 3 .~ 4 ,q 3 ,e 1 ,e 2 and z
Fig. 4(b). Input Data and Unknowns in Run 2
471
e1 =20.00W/m (=81 )
0
<Pl =16. 82 °C
z = e1 + e2 = 20.00 + 8.51 = 28.51 (w/m)
Fig. 5(a). Results in Run 1
¢1= 12.01 °C
z = e1 + e2 = o.oo + 15.16 = 15.16 (W/m)
Fig. 5(b). Results in Run 2

472
APPROXIMATE FUNDAMENTAL SOLUTIONS AND ALTERNATIVE FORMULATIONS
S Walker
Structural Dynamics Limited

Southampton UK
ABSTRACT
The two main problem areas associated with the application of

the Boundary Element Method to the solution of physical
problems are:-
1. The determination of the form of the appropriate

fundamental solution.
2. The evaluation of the integral of this singular function

over the elements needed for the determination of the
diagonal terms in the influence matrices.
In this paper the method of 'refArence points' is described

which avoids the second difficulty and a method for the
construction of the fundamental solution for any linear
partial differential equation is given. These methods are
discussed with examples from two-dimensiona l potential and
diffraction theory in order to demonstrate simply the
principles involved. All the examples were solved using
an Apple II microcomputer, demonstrating the efficiency
and usefulness of the Boundary Element Method.
1 INTRODUCTION
It is not proposed in this paper to present an exhaustive

description of the theory of fundamental solutions, the
interested reader is referred to the theoretical expositions
to be found in [1-6]. In this paper we shall be concerned
with applications, particularly related to potential
and diffraction problems. These examples were chosen
because of their familiarity and in order to bring out the
essential points in the simplest possible way. It is
however necessary to clarify a number of points relating to
the use of fundamental solutions in the Boundary Element
Method at this stage.
473
The Fundamental Solution
Consider a medium in which we have a physical property

represented by the function u which may be a function of
space and time. If we know the governing equation for u and
this equation is linear we can write this equation as
(1)
throughout the medium which we shall for the moment consider

to be infinite in extent.
Here ~is a linear differential operator. The eigenfunctions

¢k corresponding to ! are defined by
where \k is a constant called the eigenvalue corresponding to

the eigenfunction ¢k.
The form of the eigenfunctions will depend on the coordinate

system used. For rectangular Cartesian coordinates we can
choose the eigenfunctions to be the harmonic functions cos(Kx),
sin (Kx), cos (Ky), etc., or using complex notation
e!iKx. The eigenvalues \ are then just polynomials involving
powers of 'K' and 'iK'. k
For infinite regions K is a continuous parameter so we shall

write ¢k more explicitly as ¢(!, ~). These eigenfunctions
are usually orthonormal in the sense that
/¢<!. ~); <!'. ~) d~

= 0<!- !') (3)
all space
where the hat 'A' denotes complex conjugation.o is the Dirac
delta function which is zero when~ r K' but infinite at K = K'.
To define 8 uniquely we require that
r~.! ¢'<!, ~) 8(!- !') dK = ¢<!', ~) (4)
which is the replacement property of the delta function with

respect to our eigenfunctions.>J.oo represents integration over
all space.
From (4):
/a<! !') dK 1 (5)

rl,oo
and
!¢<!. ~) ¢ <!. ..0 Q (~ s.> (6)

>J,oo
474
This expansion enables the fundamental solution to be deter-

mined.
The fundamental solution u* for the linear differential operator

·i:_ may be written
t. <u* <~, f> > = o<~ - .D (7)
and is a function of two points, the source point I and the

observation point ~- (NoteL may be a polynomial of differen
tials with respect to the components of I or ~ . ) This funda-
mental solution represents the value of the field at point
x due to a point source at I (figure 1).
Using the properties of the delta function and the normalis-

ation properties of the eigenfunctions the fundamental solution
may be written
*
u <~. I> dK (8)
The formula given in (8) is a generalised transform.In fa~t

in section 3 we shall use the Cartesian eigenfunctions e-~~.~
and (8) will then be an equation involving Fourier transforms.
2 REFERENCE POINTS
One of the main numerical problems associated with the

fundamental solution is the evaluation of the diagonal terms
of the influence matrices. These diagonal terms often consist
of the evaluation of the integral of the fundamental solution
over the singularity. Typically integrals like
f
9,
Gii = u(~, .f> ds (9)
-9,
and 9,
Hii f au<_!, I> ds (10)
-~-::-an--
are sought, where ds is an elemental length along the element

considered.
These problems may be simply avoided. Consider the general

second order linear partial differential equation:-
a'u + 2b'u + c'u + d'u + e'u + f'u g' (11)

XX XXY yy X y
where the coefficients a', b', c', d', e', f' and g' are
functions of x and y only, and the subscripts x and y denote
differentiation w.r.t. these variables.
475
~ = (x,y)
Observation point
field= u*(x,yj~,n)
r = f2. -
II
Origin
FIGURE 1 NOTATION FOR DEFINITION OF

THE FUNDAMENTAL SOLUTION
476
2
If b -ac > 0 then the equation is elliptic, we shall consider
this case. Equation (11) may always be reduced to the
canonical form (see reference 6 t Chapter 3)
/.. (u) = u
XX
+ u
yy
+ au
X
+ bu
y
+ cu (12)
define the adjoint operator toJ.., i.'+ such that
/...+ (v) = v XX + v
yy
-av
X
-bv
y
+ (c-a -b )v
X y
(13)
Then with n as the outward normal, for u and v any sufficiently

differentiable functions:-
J vi(u) - u '/..+ (v) dV
n
= .![~
r an
av
- u-
3n
+ (a a;
ax + b an) uv ~ JdS (14)
n is the problem region and r is the boundary.
Then if we choose v to be the fundamental solution~*(~, f>

defined by:-
~ <u*<~. I» = a <~-I> (ls>
then we ~y use the selective property of the delta function

to reduce (14) to:-
cu(x) =
-
p
S[ au - u -au*]
u* --
3n
-
3n
+ Ia ax + b
[ 3n 3n
~] uu* dS(_O (16)
we have taken the integrations to be over the variable I
where for smooth boundaries
c =0 ~~ n
c =i ~ £ r
c = 1 ~
€. n
Notice that u* is a function of two variables and that (16) is

an identity for all values of the ~reference point.
Consider the above with a = b = 0 then
c u (~) =Iu*(~,I! ~~(O - u ~n u*(~,I)dS(s_) (17)

Then if we choose x to be a reference point ~i outside n say
then c 0 and ( 17 ) may be written:-

477
ju* (x ,~) au(~) dS(~)

=i- ?n - = f<§)
r r
or approximately for n elements I:.
J
L au(O fu*(~,_Q dS(~_)
.an
J j
=L u(S_) /-/~ u* (~ ,§_) dS(S_) (19)
r
j fj
If we discretise into boundary elements then we can now
choose n such reference points. We have now constructed n
equations for the n unknown u values on the boundary.
WE NOW HAVE NO SINGULAR INTEGRALS TO EVALUATE. One dis-

advantage of this method 1s that the influence matrices
become less diagonally dominant 1 with the resulting numerical
problems 1 as the diagonal terms now correspond to an evaluation
of u* between the ith element and ith reference point. This
difficulty may be overcome by choosing the reference point
close to the corresponding element.
Example 2-D Laplace equation
For the two dimensional Laplace equation
a2 u a2 u = V2u =0 (20)
~ + ay2
-. l
X .. (21)
and
1
u* (S_,xi) (22)
2TT
where
(23)
then (14) becomes

:L;au(O
j an (24)
note that at no point on the boundary element r . does rij

become zero not even when i = j. In the exampl~ the reference
points x. were chosen to be on the normal to element i outside
the problem region (Figure 2) distancea£ from the element
where 9. is the length of the element.
478
o n p o in ts
In te g ra ti
e le m e n t j
......t-r-........- ·
e le m e n t i
p o in t
re 1 'e re n c e
te rm s
D ia g o n a l
o n P o in ts
In te g ra ti
UNDARY
RENCE POINTS IN BO
F REFE
THE USETS O
FIGURE 2 ELEMEN
479
n equations like (24) may be written down,one corresponding to

each reference point position x .
-,
The example considered is shown in Figure 3 and follows the
geometry and discretisations given in [7]. As in [7] the
example was solved on a desktop microcomputer (the Apple II).
The problem was solved for a number of values of the scale

parameter a which determines the distance of the reference
points from their corresponding elements. a= 0 corresponds
to the conventional Boundary Element Formulation. The
results are presented element by element in Figure 4. The
value of a giving the best results was about 1.
For more complicated equations the evaluation of the

diagonal terms will often involve the use of complicated
special functions. For the two dimensional diffraction case
the fundamental solution is given by
(25)
i.e.
___!_ H( 2 ) (Kr) (26)

4i 0
where r = I~ -I l
and H( 2 ) is a Hankel function of order zero of the second
kind ...
The diagonal term is of the form
Gii =- ___!_
4i
~
-£,
H( 2 )
0
(Kr) dS (27)
An integral which when evaluated involves Struve and Bessel

functions. The use of reference points avoids the need for
the use of Struve functions altogether.
Example
Consider the diffraction problem in Figure 5. A wave of the

form ei(kx-wt) is incident on a circular cylinder from the
right.
The zero normal flux on the cylinder determines the boundary

condition on the cylinder to be:-
au (28)
an
480
Heat flow through metal plug
flow 6cm to
au
q = an
q =0
(~,6) (4,6) • Node
(0,6) (6,6) x Element
~
9 8 7 number
10 6
(2,4) (4.4)
(0,4) ~
1t! •
15
1-(6,4)
u = 30 (3,3) u =0
11 1' 5
(2,2) (4,2) (6,2)
(0 '2) • 1:! 1\
12 4
1 . 2
(6,0)
3
(0,0) (2,0)
<'4,0)
q = 0
Discretisation and Boundary Conditions
FIGURE 3 METAL PLUG EXAMPLE

481
(x) u value
(25.2)
--
X q value
' ~
,p5.0)
~
'-~!_4.7)
I
.
' ~
~
I
I
I
I 5.2 5.2 I
I I
I
I (20.0) (9.9) I
I
• • I
I
I (15.0)
•
4.8 I
I 4.8 I
'
I
l
(,0. 0) ~9 .9) '
I
I
I I
! I
'
I 5.2 5.2 I
I
I
-
I
__ -r <4. 7)
-- --- - (15. 0)
(25.2) - ____ Exact Solution
FIGURE 1 METAL PLUG EXAMPLE RESULTS

482
REFERENCE POINTS - COMPARISON OF RESULTS

SCALE FACTOR a
ELEMENT o.t 0.5 1.5 3
1 (25.2) (25.5) (25.8) (25.9)

2 (15.0) (14.9) (14.99) (15.0)
3 (4.77) (4.35) (4.22) (4.11)
4 -5.30 -4.97 -5.09 -4.92
5 -4.88 -4.75 -4.97 -5.21
6 -5.30 -4.97 -5.09 -4.92
7 (4.77) (4.35) (4.22) (4 .11)
8 (15.0) (14.9) (14.99) (15.0)
9 (25.2) (25.5) (25.8) (25.9):
10 5.30 5.16 5.09 4.92
11 4.87 4.80 4.97 5.21
12 5.30 5.16 5.09 4.92
13 (20.0) (19.8) (20.0) (20.0)

14 (9.97) (9.82) (9.96) (9.96)
15 (9.97) (9.82) (9.96) (9.96)
16 (20.0) (19.8) (20. 0) (20.0)
17 (15.0) (14.8) (14.99) (14.99)
t a =0 corresponds to the classical boundary element

method
(x) u value
x q value
TABLE 1
483
e i(wt-kx)
incident Boundary condition

field ui
,- ~ -
an
<lui
-- Cln
/
/
~ \
?1
I
I \
l /..---Sca ttered field
' \
\
I
u =su
u
+ u
~ I S
;t. . . . . . __ ./
2S. 2!\,
18• 19.
27• 3(1
17.
2cf
L}
9
26 • •
16 i"l ~1
~
1~ f2
15 22 •
•
14 !3
~4 l3
DIFFRACTION PROBLEM AND BOUNDARY
FIGURE 5
ELEMENT DISCRETISATION
484
where u = Us + u1
and UI is the incident field

US is the scattered field due to the presence of the
cylinder
The results for this problem for radius = 1 and wavelength

=2 are given in Table 2 . The values given are the complex
moduli of the scattered field representing the amplitude
of the oscillations of this field. This field is calculated
using the reference point method and the classical Boundary
Element Method for comparison.
In the examples considered above it was not strictly necessary

to use reference points.In some cases,however,the integrals
necessary for evaluation of the diagonal terms are very
cumbersome or not known in closed form. This eventuality
usually necessitates the use of an elaborate and slow numer-
ical integration scheme.
3 NUMERICAL AND APPROXIMATE FUNDAMENTAL SOLUTIONS
As an example of the methods involved in the use of numerical

and approximate fundamental solutions consider the two-
dimensional reduced Helmholtz equation.
2
+ K U 0 (29)
using (8) and the normalised eigenfunctions given by
(30)
I"'
then the fundamental solution may be written
eik 1 (~-x)
u* (xyjt;,n) =( 2~)2J
2 2
00 "' (k - Kl
In this case the integral may be evaluated using contour

integration or coordinate transformations to give
u* (x Yl t;,n) =- ~
41
H(~) (Kix-
-
t;l)
-~
(32)
for more complicated linear differential operators (31)

becomes
485
DIFFRACTION PROBLEM MODULI OF SCATTERED POTENTIAL,COMPARISON

OF RESULTS
ELEMENT I! SCALE FACTOR a APPROXIMATE

OR NODE FUNDAMENTAL
0 0.5 1.0 SOLUTION
1 1.23 1.29 1.24 1.28

2 0.516 0.520 0.507 0.546
3 0.482 0.506 0.455 0.480
4 0.815 0.803 0.776 0.808
5 0.891 0.894 0.924 0.917
11 1.15 1.18 1.15 1.17
12 0.326 0.358 0.340 0.346
13 0.134 0.152 0.135 0.145
14 0.446 0.437 0.429 0.432
15 0.630 0.616 0.637 0.613
21 0.975 1.00 0.976 0.996
22 0.308 0.335 0.318 0.317
23 0.137 0.153 0.136 0.146
24 0.431 0.421 0.416 0.418
25 0.516 0.504 0.524 0.504
Wavelength =2
Radius of cylinder = 1
The omitted nodes have symmetric values 5+10 = 1+5 etc
N.B. Only ten boundary elements were used for a wavelength

of only twice the radius
TABLE 2
486
J
00
q* (x,y l;.n> = (2~)2 (33)
where P is a polynomial of ik1, and ik and the following

correspondences may be made between t~e components of the
operator and the polynomial
ik 1 , and (34)
The integral (33) may be evaluated numerically for all relevant

points (x,y) and <~;,n> used in the problem geometry, although
the singularities corresponding to the real roots of P (k, k 1 ,
k 2 ) need some considerable care. Integrals of u* over elements
of the form (9) may be evaluated analytically with respect
to dS by change of the order of integrations.
It now only remains to consider the evaluation of integrals

of type (10) of the normal derivative of u*. Remembering that
au* au* E.£ +

au* ~
an a~; an an an (35)
au* au*
then ~ and -d-- may be evaluated by differentiation of
(33) witli respectntol;and nunderneath the integral signs.
if and ~R are for straight elements just dependent on the
e nl ement anor1entat1on.
· · Th e 1n
· t egra 1 s over dS a 1 ong th e e 1 emen t
may be eliminated as before. In fact
[9.cos0
} -au* ds =J9.sin0 a_
u* (36)
Q_
an Q. •
s1n
0 a I; J-Q.cos0
where 0 is the angle of the element makes with the horizontal.

(For 0 = 0 or nTI (36) is even simpler.)
2
The wholly numerical approach described above may not be
appropriate in all cases. In fact in the example considered
above an approximate fundamental solution may be used. The
justification of this is that the most singular terms of the
fundamental solution arise from the second derivative terms
in the corresponding linear operator. In fact these are the
terms which correspond to the Laplace equation which has a
simple and well known fundamental solution.
Consider the limits for large and small arguments of the

diffraction fundamental solution given in (26).We have:-
487
Small Kr Large Kr
1
- 2 1'l log Kr --1-
1 ("")
4 i H0 <Kr>"' ___!__
4i
j2
1Tkr
e
-i(kr- 11:.
4
(37)
For fixed K and small Kr the diffractionftlndamental solution

hehaves like the potential fundamental solution. For correct
choice of element size the integral
-£
J u* dS
for diffraction problems may be evaluated using the potential

fundamental solution in place of u*. For widely spaced
elements (with respect to the wave length considered) the
approximation corresponding to Large Kr for u* may be used
for the G .. and H. j terms.
~J ~
Consider the diffraction problem in Figure 5. This time

the small Kr approximation has been used in the evaluation
of the diagonal terms, this is a valid approximation for
elements o~ suitable size. The last column in Table 2 shows
the small effect of this approximation on the results.
4 CONCLUSIONS
It has been shown that it is usually possible to avoid the need

to integrate the fundamental solution across the singularity
which occurs when the observation point and source point
coincide, by the use of reference points.
In section 3 a method is outlined for solving second order

partial differential equations when the explicit form of
the fundamental solution is not known or very cumbersome,
by numerical evaluation of specific integral representations
of the solution.
Where numerical integration is not appropriate it is often

possible to use an approximate fundamental solution, consequen-
tly increasing speed and decreasing complexity. This is a
consequence of the fact that the singular part of most
fundamental solutions is a variant of the potential fundamental
solution with the same number of dimensions as the problem
considered.
5 REFERENCES
1. Brebbia, C.A., WalkerS Boundary Element Techniques

in Engineering (1980)
Newnes-Butterworths
2. Walker, S. Fundamental Solutions in 1 Progress in Boundary

Element Methods' (1981)
Ed. C A Brebbia Pentech Press
488
3. Roach G F. Green's Functions-Introductory Theory

with Applications (1970)
Van Nostrand
4. Thome R.C. Multiple Expansions in the Theory of Surface

Waves, (1972)
Camb. Phil. Soc. 17
5. Titchmarsh E.C. Eigenfunction Expansions Associated

with Second-Order Differential Equations (1958)
(2 Vols) Clarendon Press
6. Garabedian P.R. Partial Differential Equations

John Wiley & Sons
7. Waters J. J. and Nelson J.M. A Program for Solving

Potential Problems on a Desktop Computer (1980)
Adv. Eng. Software Vol. 2 No. 2
489
AN EFFICIENT ALGORITHM FOR THE NUMERICAL EVALUATION OF

BOUNDARY INTEGRAL EQUATIONS
Madhukar Vahle and David L. Sikarskie
Lecturer, Chairman
Department of Metallurgy, Mechanics and Materials Science
Michigan State University
ABSTRACT
The usual approach to the solution of boundary integral

equations is to represent the unknown vector by a piecewise
constant, linear, or quadratic function over the given mesh
subdivision. These representations have the advantages of
consistency, ability to integrate the equations for the given
functional approximations, and, in general, improved accuracy
as the degree of approximation is increased. While adequate
for many problems, special requirements arise for certain
nonlinear problems, e.g., plasticity, where the integral
equations must be solved for each load increment. In the
present paper a special numerical algorithm is outlined in
which the unknown vector is represented as a combination of a
Fourier series and piecewise linear function. The piecewise
linear function is used only in high gradient regions of the
unknown vector thus permitting an excellent representation
with relatively few Fourier terms. The algorithm is compared
with a linear representation alone for two problems which
show the effects of multiple connectivity, sharp corners and
discontinuous loading. For comparable accuracy both problems
show a significant improvement in computer time required.
INTRODUCTION
The boundary integral method [3,5,7] is fast becoming a
generally accepted alternative to finite element, finite
difference methods at least for linear problems. The
principal reason is that the problem dimension is reduced by
one, e.g., a two-dimensional problem is reduced to a line
integral equation on the boundary. After discretizing, the
integral equation formulation results in a lower order set of
algebraic equations which implies less computer time. There
are also certain inherent accuracy improvements, particularly
for interior information.
490
The issue is less clear for non-linear problems, e.g.,

plasticity. Since plasticity can be viewed as a linear,
incremental theory, a boundary integral approach can be
formulated. The practical difficulty is that a set of
integral equations must be evaluated numerically for each
load increment. To be competitive, the numerical algorithm
for solving the integral equations must be efficient. The
present paper addresses that question.
Before outlining the details of the numerical algorithm

(section II), it is necessary to give a brief outline of the
pertinent equations. It should be noted that there are many
boundary integral formulations [1,4], some having advantages
over others for particular problems. The ideas outlined
herein could conceivably apply to any formulation. An
indirect formulation, developed previously [2,3], will be
used. One practical advantage of such an indirect
formulation is that the unknown is a single traction vector
around the entire boundary even for mixed boundary value
problems. This permits the representation of unknown vector
in a single series.
Consider a two-dimensional, linearly elastic region R

with boundary B as shown in Figure 1. For prescribed data
on B, i.e., tractions on Bt, displacements on B , the
stresses (and displacements) in the region R arg to be
determined.
An outline of the logic involved in the formulation of

the integral equation is as follows. The region R is
embedded in an infinite (fictitious) plane of the same
material as R for which the influence functions Hij;k(Q,PB),
Iij;k(Q,PB) are known, see Figure 1. Hij;k(Q,PB) is the itn
stress component at a field point Q due to a unit line in
the kth direction at a source point PB and Iij ;k (Q,PB) is
the displacement in the ith direction at Q again due to the
unit line load at PB. Consider now a fictitious traction
f (unknown) acting along the contour B, see Figure 1. The
stress an~ displacement fields both internal and external to
B due to f are: _ (
crij(Q) - JB
Hij;k(Q,PB) fk(PB) ds (la)
ui(Q) = J( Ii;k(Q,PB) fk(PB) ds (lb)

where PB is on B and ds is an element of length along
B, s measured as shown in Figure 1. The functions
Hij;k and Ii;k are given by:
H
xx;k
f
k
=
(2a)
+ r (a 3r 2 - a 2 r 2 )f ]
y X y y
491
y
+ + +
f=f i+f j
X y
+ + -t
p=p i+p J
X y
B=B +B
t u
Figure 1. Region R Embedded in an Infinite Plane
Usual Boundary
m+l
Boundary
Present Boundary Approximation
Figure 2. Boundarv Division and Subdivision

492
2 2
H f 1 [r (a 3r a2rx )fx
yy;k k 4TTr4 X y
(2b)
2 2
+ r y (a 1 r y + a 2 r ) f 1
X y
1 2 2
H f ---4 [r (a 1 r + a 2 r y )f X
xy;k k y X
4TTr
(2c)
2 2
+ r X (a 2 r X + a 1 r y )f y 1
1
I
x;k fk = -
4 TTr 2
1
4TTr 2
+
rx, r are the components of the radius vector r drawn from
pB· toYQ. The ai's are,
3-2v (3-4 v) (l+v)
al 1-v E(l-v)
l-2v l+v
a2 1-v E(l- v)
where v is Poisson's ratio and E is Young's Modulus.

Since equations (1) represent the superposition of funda-
mental solutions, all equations of linear elasticity are satis-
fied. In order to solve the boundary value problem of interest,
the boundary conditions on B are yet to be satisfied. These
conditions are:
crij(P)nj(P) pi (P) (Sa)
ui(P) P on B (Sb)
u
where nj(P) are the direction cosines of the normal at the
point P on B and Pi(P), Oi(P) are the specified traction
and displacement components, respectively. Note that Bt, Bu
of Figure 1 are not mutually exclusive portions of the boundary,
i.e., at a given point of B the mixed condition of specified
traction in one direction and specified displacement in the
other is possible.
A description of the numerical algorithm is given in the

next section. Two examples have been selected to illustrate
the method. The first is a multiply connected region; namely,
an axisymmetric region with an axisymmetric hole loaded with a
uniform external tension. This problem illustrates the
493
advantages of the method for smooth geometries and loadings.

The second problem is the problem of pure shear of a square
region. This example illustrates the effects of discontinuous
loading and corners.
NUMERICAL FORMULATION
Problem solution requires the determination of the fictitious
traction vector fk in equations (5).t With fk known,
stresses and displacements are available from equations (1).
The usual approach to the solution of equations (5) is to
first extract the singularity analytically [2]. The boundary
is then subdivided into M piecewise linear segments as shown
in Figure 2. fk is then approximated by a piecewise constant
[2] or linear [6] function over this linear boundary segment.
After substituting the above approximation into the integral
equations, the integrals can usually be evaluated analytically,
resulting in a set of the usual linear algebraic equations.
A better boundary approximation used by the authors for

both the piecewise linear approximation and the smoothfunction
expansion (yet to come) is shown in Figure 2. The mth boun-
dary segment is further subdivided into Nm segments (only
three are shown in Figure 2 for ease of explanation). fk is
still linear over the mth segment, i.e., this further subdivi-
sion introduces no new unknowns. Note that the boundary con-
ditions will be satisfied in a collocation sense at M points
P which are not the nodal points m, m+l, etc.
The linear approximation for fk is:

d (m+l)_ d (m)
k k (s-S ) + d (m) •
m k '
m 1,2, ••• ,M (6)

k = x,y
where s = running coordinate along the fully segmented boun-
dary,
Sm = value of s at m,
d (m) unknown constant (nodal value for linear
k
approximation only).
Continuity requires that
d (M+l) d (1)
(7)
k k
tThe actual boundary integral equations are generated by sub-

stituting equations (1) into equations (5).
494
Rather than substitute fk into the integral equations

resulting from substituting equations (1) into equations (5), a
different procedure will be followed here. fk, equation (6),
is substituted directly into the stress and displacement equa-
tions (1). Note that these expressions are valid at a field
point Q. Substituting equation (6) into equation (la) and
performing the required integrations, we obtain
M Nm d (m+l)_ d (m)
k k
(Jij(Q) 2: 2:
m=l n=l
{d (m) - S
k m sm+l - sm
(8)
d (m+l)_ d (m)
k k (2)
:\S~~(Q,Sn) +
sm+1 - sm Mij ;k (Q,Sn)
where M( 1 f·;k and M(Z)ij;k are defined in detail in (8].

Note that t~e M(r)ij;k's are dependent on an interior point
Q and subinterval geometry. A similar expression to equation
(8) could be written for the displacements Ui(Q). Only trac-
tion boundary value problems will be used to illustrate the
numerical algorithm, hence all displacement equations will be
dropped. Traction boundary conditions will be satisfied when
equation (8) is substituted into (Sa) with Q + P (boundary
point).
Before boundary conditions are satisfied, the Fourier

series expansion for fk is introduced.
a cos Znqs + . 2nqs

bkq Sl.n S (9)
kq s
where Ql, Qz are the number of retained cosine and sine terms
respectively and S is the total subdivided boundary length.
The boundary is now divided exactly as in the linear approxima-
tion case, i.e., into m intervals with further Nm subinter-
vals. fk, equation (9) is now substituted into equation (la)
for <Jij· This results in a series of integrals which the
authors could not obtain in closed form. To overcome this,
the Fourier components are individually expanded in a Taylor
series about the midpoint of each subinterval. Retaining the
first two terms fk can be written as
fk(PB) Akn + Bkn [s - sn l; sn-< s < 5

n+l
(lOa)
where
Ql Znqs Q2 Znqs
n n
~n ako + 2:
q=l
akq cos s + 2: bkq sin - -
5- (lOb)
q=l
495
Ql Q2 2nqs
n
2:
q=l
)b ~
...... kq s cos
q=l
s
(lOc)
s (Sn + Sn+l)/2 n = 1,2, .•.. ,Nm
n
m 1,2, •... ,M
Substituting equations (10) in the expression for cr ij (Q)

and performing the integrations, we obtain
M Nm
(1) (2)
crij(Q) = 2: 2:
m=l n=l
{Akn Mij ;k(Q,Sn) + Bkn[Mij;k(Q,Sn)
(11)
(1)
- sn Mij;k(Q,Sn))}
(1)
where Mi··k and M( 2 ) are the same quantities as used in
ij ;k
equation ~ ~) •
Note as Ql and Q2 increase, i.e., as more sine and

cosine terms are retained, the size of the nth subdivision must
become smaller if the Taylor expansion approximation is to re-
main valid. Usually we will restrict the number of sine and
cosine terms to 3 or 4. The reason for this restriction is
that the smaller subintervals required lead to larger round-off
errors.
Both approximations for fk (linear and Fourier) have

problems; the linear requiring too many unknowns in regions of
smooth behavior and the Fourier requiring too many unknowns in
regions of large gradients. Thus, by using a combination of
the two approximations, each compensates for the other's weak-
ness.
Let
f = f (l) f (2) (12)
k k +
k
where fk(l) is given by equations (6) and fk( 2 ) by equation

(9). The stresses Oij are then the sum of equations (8) and
(11).
M Nm d(m+l)_ d (m)
cr .. ( Q)
1]
L L
m=l n=l
[{Akn
+ d (m)
k
k
8m+l - 8m
k }
d(m+l) - d(m)
(1) k k (13)
Mij ;k(Q,Sn) + {Bkn + -s=----s=---}
m+l- m
{M(2) ) (1) ( )}
ij;k (Q,Sn - sn Mij;k Q,Sn )
496
To determine the various constants in equation (13), Oij(Q) is

substituted into boundary conditions given by equation (Sa).
I t must be pointed out that even thongh Hij ·k(P ,PB) goes to in-
finity when P-+ PB, the quantities H:ij;k(ij and Hij;k(2) are
bounded. With proper care taken in constructing the boundary
subdivision near the selected boundary condition point, thesin-
gularity contribution can be evaluated numerically (see Figure
2 and ~eference [8]). This greatly simplifies the writing of
the computer code.
The summed representation of fk has considerable flexi-

bility. For example, non-zero values of dk(m) can be per-
mitted only in regions of anticipated high fictitious traction
gradients. This occurs near sharp changes in geometry and/or
loading.
The final linear algebraic equations are generated by sub-

stituting crij• equation (13), into the boundary condition,
equation (Sa). The unknowns ako, akq• bkq, dk(m) are thus
found. Stresses at any point can then be found from equation
(13). The algorithm will now be demonstrated on two examples.
EXM1PLES
A. Thick-walled Cylinder (Lame Problem)
Problem geometry and loading are shown in Figure 3. The
applied boundary traction is:
l
0 for r = 1
p (14a)
X
cos e for r = 10
l
0 for r 1
p (14b)
y
sin e for r = 10
This problem will be solved in two ways and the solution

compared with respect to computer time and accuracy.
In the first solution the unknown traction vector fk

is approximated by a piecewise linear function only. Each
boundary is divided into 24 intervals subtending lS 0 each,
H = 48. Each interval is further subdivided into S sub-
intervals, Nm = S for all m. The total number of unknowns
is 96.
In the second solution, the unknown traction vector is

expanded in a Fourier series, equation (9), (Ql = 3, Q2=4)
on each boundary. Each boundary is divided into 8 intervals,
subtending 4S 0 , H = 16. Each interval is further subdivid-
ed into lS subintervals, Nm = lS, for all m. In fact, the
subinterval division is identical with the first solution.
The total number of unknowns is 32.
497
Figure 3. Lame's Problem
/
/
Figure 4. Pure Shear Problem.

.1:>
(0
CXl
Percentage error in crr Percentage error in ae
r e = 0° e = 22.5° e = oo e = 22.5°
Linear Fourier Linear Fourier Linear Fourier Linear Fourier
1. 05 -6.72 -7.36 4.36 -1.39 -0.61 -0.06 -0.32 0.10
1.10 -2.34 -2.97 1. 59 -0.57 -0.48 0.03 -0.46 -0.02

1. 20 -0.76 -1.05 -0.05 -0.73 -0.46 0.00 -0.44 -0.02
1. 50 -0.48 -0.40 -0.35 -0.39 -0.41 -0.04 -0.41 -0.04
2.00 -0.41 -0.25 -0.46 -0.25 -0.39 -0.07 -0.40 -0.07
5.00 -0.38 -0.15 -0.38 -0.15 -0.38 -0.12 -0.38 -0.12
9.50 -0.44 -0.32 -0.41 -0.05 -0.47 -0.19 -0.24 -0.05
9.80 -0.19 -0.13 -0.77 -0.36 -0.72 -0.46 -0.14 -0.41
9.90 0.00 0.03 -0.85 -0.03 -0.65 -0.41 -0.41 -0.13
9.95 0.05 0.07 -0.65 -0.27 -0.41 -0.19 -0.92 -0.45

L_ _ _ _
-·----·- -- --- -·-
Table 1. Error in Linear and Fourier Approximations for the Lame' Problem
(minus indicates approximation is less than the exact)
499
Linear Fourier
Matrix Generation
8,950 10,118
(milliseconds)
Solution Procedure
21,408 938
(milliseconds)
Total 30,358 11,056
Table II. Computer Time Summary for Lame

Problem
Linear Fourier and Linear

(combination)
m
X y N X y N
m m
1 5.0 -4.9 3 5.0 -4.9 3
2 5.0 -4.1 2 5.0 -4.1 10
3 5.0 -3.1 2 5.0 -0.1 11
4 5.0 -2.1 2 5.0 4.1 3
5 5.0 -1.1 5 5.0 4.9 2
6 5.0 1.1 2 4.9 5.0 3
7 5.0 2.1 2
8 5.0 3.1 2
9 5.0 4.1 3
10 5.0 4.9 2
11 4.9 5.0 3
Table III. Boundary Subdivision for Pure Shear

500
The algebraic analog of the boundary integralequations

for the linear boundary approximation is generated by sub-
stituting equation (8) into equation (Sa) and letting Q
successively take on the M discrete boundary points
(collocation). The set of equations for the Fourier boun-
dary approximation is generated in the same way except
equation (11) replaces (8).
Once the fictitious tractions are found from these

equations, stresses at any field point Q are found by
substituting back into either equation (8) or (11). Table
1 is a comparison of stresses computed from the linear and
Fourier solutions with the exact. What is shown is the
percent. error of each along two representative radial lines
e = 0°, 22.5°.
Note that the errors are comparable, the Fourier tend-
ing to be somewhat better near the boundaries and the linear
tending to be somewhat better in the interior. The high
percentage errors in Or near the inner boundary is ex-
pected since the exact solution approaches zero. The boun-
dary condition is satisfied close to e = 0° but e =22.5°
lies between two boundary condition points for both solu-
tions. The computer time summary is shown in Table II.
Time for matrix generation is somewhat higher in the
Fourier case because of some additional algebra in the
approximation. Both solution procedures were identical.
In spite of the fact that this axisymmetric problem

was solved in Cartesian coordinates, it can be argued that
it is ideal for a Fourier series approach, i.e., both load-
ing and geometry are smooth functions. The next example
has both discontinuous loading and geometric corners.
B. Pure Shear
Problem geometry and loading are shown in Figure 4. The
applied boundary traction is:
-1 X = 5 -5 ~y~ 5
p 1 X -5 -5 ~y~S (15a)
X
0 -5 < X < 5 y = ±_5

( 0 X = +5 -5 ~ y < 5
p 1 -5 < X < 5 y 5 (lSb)
y
[-1 -5 < X < 5 y -5
Again, the problem will be solved two ways. The first

solution, as before, has the fictitious traction represented
by a piecewise linear approximation. The second solution is
501
Percentage Error
X = y
Linear Combination
0 1.87 2.18
1.0 1.57 2.00
4.0 1.12 2.02
4.3 .98 1.10
4.5 .11 -.08
4.6 -1.49 -1.71
4.7 -5.32 -5.47
4.8 -14.11 -14.13
4.9 -37.40 -37.35
Table IV. Percentage Error along x = y

for the Pure Shear Problem
Linear Fourier
Matrix Generation 2,918 4,358

(milliseconds)
Solution Procedure 10,759 1,502

(milliseconds)
Total 13,671 5,860
Table V. Computer Time Summary for Pure

Shear Problem
502
a combination of a Fourier expansion and a piecewise linear

approximation (in the first example the Fourier expansion
was used alone). The main purpose of the piecewise linear
approximation in the combination is to pick up large gra-
dients in the fictitious traction which would require many
Fourier terms alone.
Table III shows the division of the boundary of one of

the sides, namely x = 5. The rest of the sides are divided
similarly. The rows under m are the node numbers where
a non-zero value of dk(m) is permitted. x and y are the
coordinates of the node in the two cases, and Nm is the
number of further subdivisions of the mth interval.
Since the boundary is already piecewise linear, no

further subdivision would be required to approximate the
boundary for the piecewise linear approximation of the
fictitious traction. Nm = 2, however, is used, with one
of the subdivisions very small and used for the numerical
evaluation of the singularity contribution (see Figure 2
and appendix for details). The nodes near the corner are
very close to one another, i.e., node 10 and 11 for the
linear approximation case and nodes 5 and 6 for the combin-
ation case (see Table III). The close proximity of nodes
near the corner is because of anticipated high gradients in
the fictitious traction there.
In the combination approximation, further boundarysub-

division is necessary because of the trigonometric function
approximation (see equation (10)); note the larger values
of Nm in Table III for the combination approximation.
For this example, 4 sines, 4 cosines and 20 nonzero

values of dk(m) were chosen to represent the boundary trac-
tion; a sum total of 56 unknowns.
For the linear case, 40 non-zero values of dk(m) were

chosen to represent the fictitious traction; 80 unknowns.
The stresses were evaluated along x = y and the percentage
errors are reported in Table IV. Table V gives the time
summary. The time for matrix generation was higher for the
combination case as compared to the linear, because of the
additional required boundary subdivision and associated
algebra. Note the sizable time saving over a conventional
linear approximation for comparable accuracies.
REFERENCES
1. Altiero, N.J. and Gavazza, S.D., "On a Unified Boundary-

Integral Equation Method," J. Elast., 10, 1-9 (1980).
503
2. Altiero, N. J. and Sikarskie, D. L., "An Integral Equation

Method Applied to Penetration Problems in Rock Mechanics,"
Boundary-Integral Equation Method: Computational Applica-
tions in Applied Mechanics, ed. T. A. Cruse and F. J. Rizzo,
ASME AMD- Vol. 11, New York (1975).
3. Benjumea, R. and Sikarskie, D. L., "On the Solution of

Plane, Orthotropic Elasticity Problems by an Integral
Method," JAM, 39, No. 3, Series E, 801-808 (1972).
4. Heise, U., "Application of the Singularity Method for the

Formulation of Plane Elastostatical Boundary Value Problems
as Integral Equation," Acta Mechanica, 31, 33-69 (1978).
5. Massonnet, C. E., "Numerical Use of Integral Procedures,"

Stress Analysis - Recent Developments in Numerical and
Experimental Methods, edited by Zienkiewicz, 0. C. and
Holister, G. S., Wiley, 198-235 (1965).
6. Riccardella, P. C., "An Improved Implementation of the

Boundary Integral Technique for Two Dimensional Elasticity
Problems," Rep SM-72-36, Carnegie-Mellon University (1972).
7. Rizzo, F. J., "An Integral Equation Approach to Boundary

Value Problems of Classical Elastostatics," Q. Appl. Math.,
40, 83-95 (1967).
8. Vahle M. and Sikarskie D. L., "An efficient algorithum for

the numerical evaluation of boundary integral equations."
To appear in Computer and Structures.
504
SOLUTION OF THE DIRICHLET PROBLEM USING THE REDUCTION TO

FREDHOLM INTEGRAL EQUATIONS
J. Caldwell
Sunderland Polytechnic, England
ABSTRACT
The numerical solution of potential problems using a method

which reduces them to Fredholm integral equations of the first
and second kinds is considered. The method is illustrated for
potential problems of the Dirichlet type and solutions have
been obtained for the case of a unit sphere. This method
provides an alternative approach to the classical finite-
difference methods which have been widely used in the past.
It overcomes some of the difficulties associated with the
classical approach.
The problem is to solve Laplace's equation
in some closed volume V with the condition
<~> = <~>p <.!:s>

on the boundary surface S, <l>p being prescribed.
By using a Green's function type of solution and replacing the

derivatives by simple finite differences the problem can be
reduced to a Fredholm integral equation. Rather than solving
a Fredholm equation of the first kind it is preferable to
introduce a small parameter E and solve a Fredholm equation of
the second kind. This is done by solving for different small
values of E and then extrapolating to the limitE= 0. It is
possible to obtain reasonable accuracy without resorting to
high order (N) Gaussian quadrature for integral evaluation.
The advantages of extrapolating in N are then considered. It

is possible to use an empirical formula and fit a curve of the
form
 = A + B . NP
505
to the computed results. In this way it is possible to get an

idea of the goodness of fit by comparing with the results
obtained for larger N. For any r, the results can be extra-
polated to E = 0 which is equivalent to extrapolating to N
and then to E = 0 by simple calculations which avoid large
matrix inversion.
One of the difficulties associated with the Fredholm integral

equation is the peaked nature of the kernel. This can be overcome
by rearranging the equation and using an iterative approach in
the solution.
INTRODUCTION
Much work has already been carried out on the solution of two
and three dimensional magnetic field problems. Substantial
effort has been devoted to the computerised calculation of
two and three dimensional magnetic fields which involves the
numerical solution of Laplace's equation. As a result relevant
computerised methods have been devised by Perin and
VanderMeer (1967), Halacsy (1968), Halacsy et al. (1967),
Nelson et al. (1969), Erdelyi et al.(l968), Winslow (1964),
Colonias (1967), Trowbridge (1972) and others.
Many problems in engineering and science involve the solution

of Laplace's or Poisson's equation both for analytic and non-
analytic three dimensional geometries. Among the variety of
numerical techniques, the relaxation method is particularly
suited to a digital computer due to the method'sintrinsic
simplicity. However, since the relaxation method is an
iteration over a three dimensional mesh, the large number of
mesh points necessary to solve typical problems has resulted in
heavy demands on memory space and computer time. This applies
to most of the computerised work by the authors listed above.
For these reasons an alternative is considered in this paper.
A summary of the initial stages of this work is provided by
Caldwell (1980). The numerical solutions of potential problems
in iron-free regions are found using the technique of
reduction to Fredholm integral equations of the first and
second kinds. The motivation of this analysis is to construct
numerical solutions to potential problems of Dirichlet type in
the first instance although the work could be extended to mixed
boundary value type problems. This provides an alternative
approach to the classical methods by the authors mentioned above
One particular line of approach is the axi-symmetric case and,

in particular, the spherical case. By assuming a Green's
function type of solution for Laplace's equation and using
simple difference approximations, the problems can be reduced
to a Fredholm integral equation. This may be solved by a
direct method using Gaussian quadrature and matrix inversion
as described later.
506
REDUCTION OF DIRICHLET TYPE PROBLEMS TO AN INTEGRAL EQUATION
The analysis described below can be used to construct solutions

to potential problems of the Dirichlet type, that is the prob~
of solving Laplace's equation v2 ~ = 0 inside a region where ~
is specified at each point of the boundary (for example, the
boundary curve of a plane region, or the surface of a three
dimensional region). It can also be used for problems of the
Neumann type whose values of the normal derivative of the %%
function are prescribed on the surface and can be further
extended to solve problems of the mixed boundary value type.
We shall now show how the Dirichlet problem can be reduced to

an integral equation. The problem is to solve Laplace's
equation
(1)
in some closed volume V with the condition
(2)
on the boundary surfaceS, ~p being prescribed (see Figure 1).
A Green's function type of solution of Ec,.uation (1) is given by
~dS'
j I~-~· I
(3)
s
where the surface charge density is
O(.!:g) = - _l_ (~]

41T an s
- _L (~+- ~-) (4)
41T an an .
Replacing the derivatives in Equation (4) by simple differences
gives
a <.!:B> = - 1
4 1T£ { ~ (~ + £.!!_) +~ (~ - £.!!_) - 2 (
~P (~8 )}+0 £). ( 5)
Using Equation (3) this then reduces to
41T£0 (~) +
!,"'""'{1-~-,----=~'----E-.!!I+ Irs-~+ Enl} ds'
= 2 ~p(~) + 0(£). (6)
This is a Fredholm integral equation of the second kind and we
are interested in solutions of this equation in the limit as
£ + 0. Proceeding to this limit we obtain
jls ~<-:-S> I
~ ~
dS' = ~p(~) (7)
507
which is a Fredholm integral equation of the first kind.

Having solved Equation (7) for o we may then compute~ using
quadrature from Equation (3).
Equation (7) and its relation to solutions of potential

problems is well known (see Garabedian 1964) and is the basis
for much analysis. We now show how it can be used to obtain
numerical solutions of three dimensional potential problems.
The'difficulties associated with the numerical solution of

Equation (7) are illustrated by considering the one
dimensional equation
J'K(x,y) f(y)dy = g(x) (8)
0
and the corresponding eigenvalue problem
l
J K(x,y) f(y)dy = >f(x). (9)

0
In Equation (8) the function g(x) and the kernel K(x~y) are
prescribed and f(y) is the unknown. When we consider
Equation (9) we find that it possesses a set of eigenfunctions
fi and corresponding eigenvalues Ai which have a limit point
at zero, that is
Ai ~ 0 as i ~ oo.
This means that if we replace Equation (8), in an obvious

notation, by
3 wj Kijfj = gi (10)
where wj are weights corresponding to a numerical quadrature,

then Kij will be singular in the limit. In other words, our
problem is ill-conditioned.
These difficulties do not arise in the solution of Equation (6)

since it is an equation of the second kind and this suggests
a possible line of approach, namely, that we solve Equation (6)
for different values of E and extrapolate to the limit. We
must expect , however, that to obtain accurate results it is
necessary to solve Equation (6) for small values of E.
Furthermore, as E decreases we will find that more and more
points must be included in the quadrature scheme. Note also
that as E ~ 0 the kernel associated with Equation (6) has a
sharp peak near ~ = Es which will further complicate our
numerical quadrature.
It should be pointed out that Bakushinskii (1965) has

considered the numerical solution of Fredholm integral equati~
508
of the first kind, namely
fbK(x,y) <(y)dy • g(x)

a
by examining the limit case of the equation
c <(x) + fbK(x,y) <(y)dy g(x)
a
as E -+ 0.
All the difficulties outlined above are well illustrated by

the problem discussed in the next section.
SOLUTION OF A SPHERICAL PROBLEM
By applying the method outlined in the previous section we can

find a solution for the Dirichlet problem in the special case
where the volume V is a sphere of magnetic material of unit
radius. In the first instance we shall make no allowance for
axi-symmetry and so the full three dimensional problem will be
considered.
This means that we consider the problem

=0 =1
}
v2 ~ inside the unit sphere r
and (11)
~ ~p(6,¢) on the surface.
Now the square of the distance between P(r,6,¢) and

P'(r',6',¢') is given by
~~- ~· 1 2 = r2 + r• 2 - 2rr' {cos6 cos6'- sin6sin6'
(12)
cos(¢ - ¢')}
rF·,
and so Equation (6) becomes
4TIECJ(8,¢) + CJ(8' ,¢') {k(E,8,6' ,¢,¢') +
0 0 (13)
k(-E,8,8',¢,¢')} d8' d¢'
where
k(E,8,8',¢,¢') = [1 + (1-E) 2 - 2(1+E) {cos8cos6' -
(14)
sin8sin8' cos(¢- ¢'>}]-!.
If we now introduce a scheme based on N quadrature points and
associated weights wi(i = 1,2, ..... N) we find, in an obvious
notation, that
(15)
(i,j = 1,2, ....... ,N)
509
In Equation (15) we have written
Aijk£ = wkw£ sin9k {k(£,9i, 9k, ¢j• ¢£) +

(16)
Equation (15) may now be rewritten as

N N
(17)
kJ;1 £h (Aijk£ + 4TT£0ik oj£)ok£ 2 (<!>P)ij
where
oij 0, i 'f j
oii 1.
Note that the matrix equation (17) is of order~ 2 xN 2 ) and thus
to solve problems involving 8 quadrature points in each
direction involves inversion of matrices of order (64 x 64).
This method will now be demonstrated by taking the axi-

symmetric case as a test example. This means that we must
start from Equation (13) with a = o(9) and we then find that
'''''''' •f' 0
l''o(Oj)'inOj f(c,e,,e,,j) dOj dj
0
(18)
2<!>p(9i)
where
(19)
Note that the kernel in Equation (18) is
.,,,,,,, 0 f';,,,,,,,, .• , •• (20)
0
and is therefore independent of ¢ . We now have a one
dimensional integral equation which may be solved in the usual
way by a direct method. The kernel may be written as an
elliptic integral but this would not appear to be useful and
so it is best to use direct quadrature. Hence by introducing
Gaussian quadrature Equation (20) becomes
N
K(8i,8j) = 11 kJ; 1 wk f{£,9i,ej, rr(1+rk)} (21)
where wk, rk are the weights and abscissae of Gaussian

quadrature.
This means that o(8i) fori= 1,2, ....... ,N can be obtained by

solving the integral equation
510
•••• ' ' • ' + /"·····'•' • ,,,, •••,, .,, - 2 ••• ,.~ (22)
0
Again using Gaussian quadrature to evaluate the integral in
Equation (22) we then have a set of N linear equations in N
unknowns, namely
A a = _!!, (23)
where
i,j 1,2 •...•... ,N
i 1,2, . . . . . . . . . ,N (25)
i = 1,2, ......... ,N. (26)
Using a computer program POTQUAD Equation (23) has been solved

for various orders of Gaussian quadrature, N, in order to
obtain approximate values for a(6i). However, since
increasing the number of quadrature points yields information
at a different set of points, it is necessary to add to the
program a numerical integration procedure to calculate the
potential ~. We could, for example, calculate ~ along the
axis using Equation (3). Thus, on the axis we have
I TI
0
2
J :(8')sin8' d8' dp
0
(l+r 2 -2r cos6');
(27)
_ 2 TI Jna(6')sin6' d6' !
- (1+r 2 -2r cos6')
0
which may be evaluated at the points r = 0(0.1)0.9 by Gaussian
quadrature to give
TI
~(r) = n2 ~ w;CJ(6;)sin7 · (1+r;) (28)
i=1 {l+r 2 -2r co~· (1+ri) }i
A useful check on accuracy of the above method can be obtained
by considering the problem in which the potential on the
surface of the sphere r = 1 is given by
~p(6) = Pn(cos6) (29)

which has analytic solution
~(r,6) = rn Pn(cos6). (30)
Results have been generated for the n=1 case, that is,
511
~P(6) = case on r 1
which has exact solution ~ = r on the 6 = 0 axis. In this way

values of ~(r) on the axis have been obtained by using
Gaussian quadrature of order N = 2(2)16 for r = 0(0.1)0.9 for
the cases E = 0.1(0.1)0.4. Typical results for the cases
r = 0.2 are presented in Table 1. Clearly there is good
agreement between these results and the exact values and hence
these results can be used to produce more accurate results.
Table 1 Computed values of ~ for r = 0.2

where ~p = PI(cos6) on r 1
H0.2)
N
E = 0.1 E = 0.2 E = 0.3 E = 0.4
2 .4209 .3629 .3219 .2920
4 .2745 .2329 .2065 .1872
6 .2341 .2079 .1904 .1769
8 .2205 .2004 .1865 .1748
10 .2128 .1965 .1845 .1739
12 .2083 .1944 .1836 .1735
14 .2052 .1931 .1831 .1733
16 .2032 .1923 .1828 .1732
Remembering that analysis has shown the error to be O(E), the

results for the cases E = 0.1, 0.2, 0.3 and 0.4 can be
extrapolated to zero using the formula
~0 = ~i + AEi + Be 2 i + Ce3i
for i = 1,2,3,4. This gives
(31)
The extrapolated results for the cases r = 0.2, 0.5 and 0.8
are presented in Table 2. Similar results apply for other
values of r. Note ~nat as N increases the extrapolated
results tend to some fixed value which is encouraging. This
suggests that we can continue the calculations with larger N
and then try to obtain the results for N = oo by extra-
polating inN.
512
Table 2 Extrapolated results to E 0
<l>o
N
r = 0.2 r = 0.5 r = 0.8
2 .5018 1.2250 1.6464
4 .3394 .7382 .9772
6 .2737 .6477 .9098
8 .2508 .5986 .8732
10 .2363 .5708 .8494
12 .2277 .5536 .8375
14 .2213 .5404 .8281
16 .2170 .5320 .8221
If we continue the calculations with higher order Gaussian

quadrature we find that we get very accurate results. This is
clearly seen by examining the results for the case N = 20,
E = 0.1 which arepresented in Table 3. The maximum error over
the range of values of r is approximately 1%.
Table 3 Computer values of <l> for N = 20, E 0.1
r <l>(r)
0 .0000
0.1 .1004
0.2 .2007
0.3 .3007
0.4 .4004
0.5 .4995
0.6 .5981
0.7 .6960
0.8 .7934
0.9 .8903
In exactly the same way results can be obtained for the cases
<l>p= rnPn(cos8) for higher values of n.
It should be noted that an improvement in accuracy will be

obtained be replacing Equation (5) by the more accurate
difference scheme
1
O(~S) =- 4ITE { 2 <lJ(~S + E~) + 2 <lJ(~S
(32)
513
EXTRAPOLATION IN N
On examining the results presented in Table 1 in more detail

we find that some interesting patterns are forming. Further-
more, we can use these to extrapolate to the limit as E ~ o
and hence save in computer time and storage requirements. One
of the drawbacks of using Gaussian quadrature to evaluate
integrals is that the error term is not helpful in error
estimation. For this reason we use an empirical formula and
fit a curve of the form
<!> = A + BNP . (33)
where A, B and p are the arbitrary constants to be determined,
to the N = 2,4 and 6 results. We can then compare with the
computed results for larger values of N to test the goodness
of fit. This means that we can extrapolate toN = oo and then
to E = 0 by simple calculations which avoid large matrix
inversion (see Bakushinskii 1965).
In the first instance the computed results for the N = 2,4

and 6 cases are curve fitted and compared with the computed
N = 8 case. If sufficient accuracy is not achieved the
process is repeated by fitting the N = 4,6 and 8 results
and comparing with theN= 10 case, and so on. The equations
for this procedure are as follows:
<!> A + B(N+2)p
(N+2)/2
<!> A + B(N+4)p
(N+4)f2
This gives
Np - (N+2)p
k
(N+2)P - (N+4)P
and hence
f(p) = k(N+4)p- (1+k)(N+2)p + Np = 0. (34)
Since Equation (34) is non- linear we use the Newton-Raphson

iterative method to find p. This gives
f(pi)
pi+1 = Pi - f~pi)
k(N+4)Pi_(1+k)(N+2)Pi + NPi
Pi- k(n+4)Pi~n(N+4)-(1+k)~(N~+-2~)~pn1~-~~n~(N~+-2~)+_N_P~i~-n-N- ( 35 )
(i = 0,1,2, 0 0 0 0 0 ,)
514
Then B is found from the equation
¢N; 2 - ¢(N+2)/2
B (36)
NP - (N+2)P
and A fro~ the equation

A = ¢N/2 - BN~ (37)
A computer subroutine POTFIT is available which reads in the

computed values of ¢ obtained previously for the cases
N = 2(2)16 and calculates p to an accuracy of 6D. The process
is repeated until the difference between the computed ¢ from
this program and the corresponding value obtained from the
program POTQUAD is less than 0.0001. In this way values of ¢
for higher orders of quadrature have been predicted.
By feeding in results such as those in Table 1 we find that

the scheme converges at an early stage to give the accuracy
specified and in the cases
r = 0.2, E = 0.1(0.1)0.4
the values of A,B and p relating to Equation (33) are
presented in Table 4. These results have then been extra-
polated to E = 0 to give ¢ = 0.19793when r = 0.2.
Table 4 Values of A,B and p for the case r 0.2
E A = ¢o B p
0.1 .19383 .61678 -1.51106

0.2 .18952 .68083 -1.98765
0.3 .18209 .95039 -2.58805
0.4 .17251 .24498 -2.24123
The problem of slow convergence may arise particularly for

values of r close to 1. In suchcases we may either have to
relax the accuracy requirements or to feed in to the program
values of ¢ for orders of Gaussian quadrature beyond N = 16.
Again this work can be generalised to consider the cases
¢p = rn Pn(cos8) for higher values of n.
AN ITERATIVE METHOD OF SOLUTION
One of the difficulties associated with the direct method of

solution of integral equations of the type (6) is the peaked
nature of the kernel. This means that high order quadrature
must be used and leads to inversion of large matrices. To
overcome this difficulty it is possible to use an alternative
method of solution of Equation (6). Firstly, we can rewrite
Equation (6) in the form
515
4nEa(Es> + a<Eg> Is
g<Es· ES , E~)dS'
(38)
+ 18 {o(~s)- o("s)} g('s, ' " ' <~)dS'

where
(39)
Then we can proceed to solve Equation (38) by iteration;

that is
(40)
I.
where
4'"M+1 (~g) + 0 M+1 ("s) g(!g, ~·· <~)dS'

-I, '"•''"' -"•''"') "''"' ~- .<~)dS'
(41)
= 0,1,2, ...... )
I E~)dS'
(M
and
+ ao<Es> g(Es· Es • (42)

s
Clearly we no longer have the difficulty of a peaked kernel
since the second integral in Equation (38) has no such
behaviour near Eg = E5
and the first integral does not
involve a. It may therefore be computed by high order
integration without affecting the iteration process.
Preliminary results from a numerical investigation of

Equations (40) ,(41) and (42) for the case of a sphere are very
encouraging. It is also possible to analyse the convergence
of this method together with the error accumulation.
CONCLUDING REMARKS
The previous analysis can be used to solve Dirichlet and

Neumann problems for a linear elliptic partial differential
equation by transforming them into integral equations. In
past work the numerical solution of potential problems has
largely been carried out by using the classical approach based
on finite differences and then solving the resulting difference
516
equations by relaxation methods. These methods require great

computer storage capacity and a very fast access. The
storage problem can only be overcome at the expense of compuwr
time. Convergence is also always a problem in iterative
methods using computers. A very large number of iterations is
required and known iterative methods prove useless in
accelerating convergence. Other difficulties are posed in
problems having symmetry and uniformity. For example,
symmetry has the effect of making the inversion matrix
singular.
Many of these problems are overcome by using the technique

of reducing the problem to a Fredholm integral equation of the
second kind. This means that by using the direct method
described we can solve for various small values of E and then
extrapolate to the limit E = 0. Further improvements can be
made by extrapolating in N (the order of Gaussian quadrature)
and using an iterative solution to overcome the peaked
nature of the kernel. The method gives fairly accurate
results for reasonably small N and has the advantage over the
classical method of making smaller demands on computer time
and memory store. In a typical case, when using the program
POTQUAD for the spherical problem values of ~(r) for a
particular E, r = 0(0.1)0.9 and orders of Gaussian quadrature
N = 2(2)16 required only 6 minutes central processor time
on an ICL System 4/50 machine. The time then taken for the
extrapolation inN using the program POTFIT was negligible.
ACKNOWLEDGEMENT
The author would like to thank Professor R.D. Gibson

(Newcastle Upon Tyne Polytechnic, England) for his interest
and assistance in this work.
REFERENCES
Bakushinskii, A.B. (1965) A Numerical Method for Solving
Fredholm Integral Equations of the First Kind. Zh. vychisl.
Mat. Fiz., ~. 4 : 744-749.
Caldwell, J. (1980) An Application of Extrapolation to the
Limit. BIT, Sweden, 20, : 251-253.
Colonias, J. (1967) Calculation of Two Dimensional Magnetic
Fields by Digital Display Techniques. Report No UCRL-17340,
LRL Berkeley, U.S.A.
Erdelyi, E.A., Fuchs, E.F. and Ratti, U. (1968) Acceleration
of Convergence Techniques for the Relaxation Solution of
Nonlinear Laplace's,Poisson's and Diffusion Equations. Proc.
of Discussion on Analysis of Magnetic Fields, University of
Nevada, U.S.A.
517
Garabedian, P.R.(l964) Partial Differential Equations. Wiley.

Halacsy, A.A. (1968) Three Dimensional Calculations of Magnetr
Fields. Proc. of Discussion on Analysis of Magnetic Fields,
University of Nevada, U.S.A. : 109-121.
Halacsy, A.A., Clark, G and Dunks, J. (1967) Computerised
Calculation of Three Dimensional Magnetic Fields. Proc. 2nd
Int Conference on Magnet Technology, Oxford : 61-78.
Nelson, D., Kim,H. and Reiser,M. (1969) Computer Solutions
for Three Dimensional Electromagnetic Field Geometries.
I.E.E.E. Trans on Nuclear Science, NS-16, No.3.
Perin,R.and VanderMeer, S. (1967) The Program MARE for the
Computation of Two Dimensional Static Magnetic Fields.
Report 67-7, C.E.R.N. Geneva.
Trowbridge C.W, (1972) Progress in Magnet Design by Computer.
Proc. 4th Int. Conf on Magnet Technology, Brookhaven, U.S.A.
Winslow, A.M. (1964) Numerical Solution of the Poisson
Equation in a Nonuniform Triangular Mesh. Report No. UCRL
7784-T (Rev.2) TRIM, LRL Berkeley, U.S.A.
Figure 1. Volume V bounded by a surface S

518
BEMSTAT - A NEW TYPE OF BOUNDARY ELEMENT PROGRAM FOR

TWO-DIMENSIONAL ELASTICITY PROBLEMS
Lars Bolteus and Odd Tullberg,

Dept of Structural Design and Dept of Structural Mechanics,
Chalmers University of Technology,
S-412 96 GOteborg
SWEDEN
ABSTRACT
A new type of boundary element program is presented. The equations are derived by
use of integral operators. A natural way of deducing element matrices in BEM is
shown, which leads to a new technique for establishing the· system of equations. A
simple subsidiary condition technique, which makes it possible to take discontinous
tractions into account, is presented. Guidelines are given for the BE-discretization and
the numerical integration. Formulas for the analytical integration of the singular terms
are shown. A non-conventional method for coupling BEM and FEM is proposed.
Numerical studies have been made in order to investigate the performance of different
elements.
INTRODUCTION
A wide range of engineering problems in structural mechanics, fluid mechanics, and

electrical engineering can be formulated with boundary integral equations (Banerjee
and Butterfield, 1979). These equations are most conveniently solved by numerical
methods.
The numerical techniques for integral equations are developing rapidly and many
articles have been written on the subject during the last decade, for references, see
Banerjee and Butterfield (1979) and Brebbia and Walker (1980). The collocation
method is commonly used for the numerical solution (Jaswon and Symm, 1977), but
also variational methods have been presented (Nedelec, 1977; Wendland et at, 1979;
Jeng and Wexler, 1977).
In this paper, the two-dimensional mixed boundary value problem of elastostatics is for-
mulated with the direct integral equations. To the authors' knowledge, no variational
solution exists for these equations. Hence, the collocation method is here used for the
numerical solution. The purpose of this paper is to report on BEMSTAT, a new type of
boundary element program for two-dimensional eleasticity problems. The program in-
cludes a new technique for establishing the system of simultaneous equations, a simple
method for handling discontinuous tractions, and a method for coupling BEM and
FEM. In addition, guidelines for the numerical integration and the element discretizati-
on are given. Several numerical examples are calculated in order to investigate the beha-
viour of different elements.
519
DERIVATION OF THE BOUNDARY ELEMENT EQUATIONS
Boundary integral representations
The displacement field in a two-dimensional elasticity problem can be represented by a

boundary integral formula, the Somigliana identity:
u (x) = Ju(x,~)t(f;)d~- fTT (x,C)u(~)d~ xEQ (I)
r r
where
u ·= [ux uy)T (2)
t [tx t )T (3)
y
are the physical displacement and stress vector (or traction) components in a
Cartesian coordinate system Oxy and the body under consideration is occupying the
domain shown in Figure I.
Figure I. Considered domain
U(x,O is the fundamenfal solution matrix, which represents the displacement field due to
a unit pointload in ~ with the direction vector e(~).
u(x) = U(x,C)e(~)
(4)
and T is a matrix which gives the corresponding traction field (for a given normal direc-
tion),
t(x) = T(x,l;)e(~) (5)
T can be deduced from the fundamental matrix U. (See Appendix A.) The representation
formula, Equation I, can conveniently be written with integral operators:
u(x) = Kt(x) - Lu(x) xEQ (6)
These linear operators are defined as

Kv(x) = JU(x,~)v(l;)dl; (7)
r
Lv(x) = fTT(x,l;)v(l;)dl; (8)
r
When the displacements and tractions are known over the entire boundary, the interior
displacement field can be generated from Equation 6. Moreover, the strains can be
found through the kinematic relation (small strains and deformations)
£ = Vu (9)
where the strain tensor components are represented by the matrix
(10)
520
and the gradient operator is defined as

a;ay] T (II)
a;ay a;ax
Introducing Equation 6 in Equation 9 gives the integral representation of the interior
strains
E(x) = 9Kt(x) - VLu(x) (12)
where
vKt(xJ = fv u(x,cJt(cJdc (13)

r x
(14)
The subscript x indicates on which argument the gradient matrix operates, and as can
be seen, it only operates on the fundamental matrices. These C.(ln be calculated analyti-
cally. Hence, no approximations are involved in Equation 12. VU and VT can be found
in Appendix A. When the strains have been calculated, the stresses can be found by the
generalized Hooke's law
(15)
The stress components in the Cartesian coordinate system are represented by the matrix
a = [a
x
a
y
T
xy
lT (16)
and Sd is the differential stiffness matrix for plane strain.
Boundary integral equations
The preceding formulas have been formulated for points in the domain. By a limiting
procedure, the point can be taken to the boundary. The general boundary integral
equation can then be written as
(E+L)u(x) - Kt(x) = 0 xEr (17)
where the new integral operator is defined as

Ev(x) = fD(c)v(c)o(x,c)dc = D(x)v(x) (18)
r
For a point on a smooth boundary, the matrix Dis diagonal
D= [0 ,5 0 J (19)
0 0,5
but on a sharp corner D is full. The general form of D for two-dimensional problems is
given in Appendix A.
Equation 17 is the integral equation to be solved. In a general boundary value

problem, displacements are prescribed on a part of r, e.g., on r 1 , and tractions on the
rest, called r 2, thus r = r 1+r 2 . The system of integral equations can now formally be
restated as
(E+L) 11 (20)
[ (E+L) 12] [u1] - [K11
(E+L)21 (E+L)22 u2 K21
521
where the first subscript indicates on which part of the boundary the pole is located,
and the second over which part the integration is performed. The bar indicates
prescribed function. Reordering of Equation 20 gives the final system of boundary
integral equations to be solved
xEr 1 (21)
[ K11
-K21
- (E+L) 12J [t1J
(E+L) 22 u2
= [:~] xEr 2
where
f1 (E+L) 11 u 1 - K12t2
(22)
f2 -(E+L)21u1 - K22t2
To the authors' knowledge, the different operators in Equation 21 have not yet been
mathematically investigated, which implies that a correct variational formulation
cannot be performed.
Numerical method
To achieve a numerical solution, the collocation method is used. This can be looked
upon as a weighted residual method with a special choice of weight functions, namely
the Dirac function with origins in the collocation points. Thus the boundary integral
equation will be forced to zero in the collocation points. For convenience, Equation 17
will now be used, and the boundary conditions will be introduced later. Thus,
I ;: : : 1 I 2, ... ,N (23)
where N is the number of collocation points and 1 the coordinates for node I. The x
boundary is now discretized with boundary elements over which displacements and
tractions are approximated. The approximation is of a polynomial shape. In
BEMSTAT, the shape functions can be of constant, linear, or quadratic forms, and the
collocation points are placed in the element function nodes. Thus, the approximation
can be written as
u(x) = ~T(x)u (24)
(25)
:r
where the shape function matrix is defined by
<P (x)
2 (26)
<P 1 (X) 0 <P (x)
2
and where the shape functions are defined as having only local support
- {=1ifi=J (27)
<PI(XJ) = 0 IJ = 0 if I~ J
The nodal quantities are stored in u

- [-1 -1 -2 -2 -N -N)T (28)
u = ux uy ux uy ••• ux uy
The same is valid fort. As shown in Equations 24 and 25, the same shape functions are
used for displacements and tractions. In a consistent formulation, tractions should be
approximated by functions of one order less than the displacements. However, this can
be handled in the program since different shape functions can be mixed. A more severe
error in Equation 25 is the continuity requirements for the tractions. This problem is
circumvented in the program by an efficient subsidiary condition technique, which
makes possible discontinuous tractions. Whenever this occurs, two nodes are placed in
522
the discontinuity point. No element is defined between the nodes. The nodal
displacements are forced to be equal, i.e., one of the nodes is considered to be slave of
the other. The row associated with the slave is deleted in both H and G, and the
associated column in His added to the column associated with the master. Notice that
the column in G associated with the slave is kept. This method is easier to use than the
one proposed by Alarcon et al. (1979), since they define the tractions as discontinuous,
a priori, and then generate the correct continuity requirements by a code system.
The approximation in Equations 24 and 25 is now inserted in Equation 23, which gives
Hu-dt=O (29)
where the elements in the matrices are
(E+L)~J = J[o(x)6(x,i 1 > + TT(x 1 ,x>l~J(x)dx (30)

r
GIJ = K~J = Judx (31)

r
H and G are assymmetric. In addition they are dense, although locally defined shape
functions are used. When the boundary conditions are introduced, the system in
Equation 29 can be reordered in the same manner as in Equation 21 to a solvable
system of equations
Av = E" (32)
where A is a mix of H and G, which implies that it is dense and assymmetric, vis the
unknowns on the boundary, and b includes all known boundary data.
When the boundary variables are known, requested quantities in the domain can be
calculated. Interior displacements can be found by
u(x) = K~ T (x)t- - L~ T -
(x)u (33)
and strains by
e(x) = VK~T<x>t- VL~T<x>u (34)
Establishment of the system of equations
In a traditional BEM-program the system of equations is established collocation

pointwise, i.e., row by row. This method is different from the efficient and systematic
method used in FEM, where matrices for each finite element are formed and assembled
into the system matrices. Here will be shown that this procedure also can be used for
boundary element methods.
The displacement in a point on the boundary is approximated as
u(x) = ~T(x)u (35)

If we define local element matrices (see also Figure 2),
(36)
0 ]'r (37)
0 l()e(x)
2
523
the approximation can be written

u (x) = E~eT (x) ue (38)
e
I 2
Figure 2. Locally defined element quantities
The displacements have to be continuous over the element boundaries; thus, one can
define a topological incident matrix c• for each element, which connects local and
global variables
(39)
or
a = 1 to 2 (3 or more) (40)
I 1 ,N
where a is an index running from I to 2 for linear shape functions, 3 for quadratic, and
so on. Introducing Equation 40 in Equation 38 and comparing this with Equation 35,
yields
(41)
This relation is now used in Equation 30, which leads to
HIJ = l:(E+L)~~(x 1 )C~J (42)

e
and similar for the G matrix
(43)
Hence, element matrices for the collocation boundary element method have been
derived
(44)
(45)
The matrices have the size Nx2 for linear shape functions and Nx3 for quadratic ones
(block matrix formulation). By introducing the boundary conditions for the element,
the final element matrices A• and \J•, can be formed. These can then be assembled into
the global system matrices, in the same manner as in FEM. Formally, this can be
written
A = l:Ae; b = Ebe
e e (46)
where A• is a mixture of H•·and o• and \J• includes the known boundary element data.
By this method, element data (integration point data) are calculated only once. This is
in contrast to a traditional HEM-program, where these calculations are repeated for
each collocation point. Thus, the proposed method is less time consuming.
524
DERIVATION OF OPTIMAL INTEGRATION SCHEME
The element matrices are integrated numerically when the collocation point is located
outside the element. This is here called exterior integration. When the collocation point
is inside the element, the integration is performed analytically or with some special
numerical scheme, since the integrands tends to infinity. This situation is here called
interior integration.
Exterior integration
The exterior integrals in the element matrices are numerically integrated with a Gauss
quadrature formula with unit weight function. Since the construction of the
simultaneous equations may take longer than it does to solve them, an optimal
integration scheme should be developed and used. In this section an integration scheme
is recommended and guidelines for the discretization are given (see also Watson 1980.)
The integration is most conveniently done on a canonical element (-1,1). Hence, both
geometry and functions are represented on this element, and then the geometry is
mapped to the global system. Thus,
(47)
where ~ is the global coordinate vector, 11 the intrinsic coordinate, the shape function
matrix, and x• the nodal coordinate matrix for the element. Hence, the element
matrices can be evaluated as
fTT(x,x)$eT(n)Je(n)dn (48)
-1
1 -
Ju(x,x)$ eT
(n)J e (n)dn
(49)
-1
where the superbar on fi• indicates that the D-matrix is left out, »g.» stands for
integration point, and w is the integration weight. 16 is the Jacobian:
e d~eT_e 2 d~eT_e 2) 1/2 (50)
J (n) = ( (Clilx ) + (Clily )
The element matrices in Equations 48 and 49 are established in the following manner;
I) Calculate necessary data for the current integration point, 2) Calculate contributions
from all collocation points, 3) Take the next integration point.
In BEMSTAT the geometry can be mapped by linear or quadratic shape functions.

This choice is independent of the function approximation chosen. Hence, subparamet-
ric, isoparametric, and superparametric elements can be used.
Gaussian integration is exact for polynomials. However, the integrands consist not
only of polynomials. Present are also functions like 1/r and ln(r), multiplied by
complicated trigonometric functions. These functions become predominant, when the
pole is close to the element considered. This implies that the Gaussian integration no lon-
ger is exact. Therefore, an empirical study has been made here, in order to investigate the
optimal number of integration points necessary. In Table I the performance of the
Gauss qu!!_drature with 2 and 4 integration points has been studied for a characteristic in-
tegral in H 6 (see also Figure 3).
525
Figure 3. Position of pole and element under

consideration
Table 1. Values of a characteristic integral in H" for

distance ratios between 2 and 5
IlL »exact» 4 g.p. 2 g.p.
5 0.051311 0.051311 0.051311

4 0.063331 0.063331 0.063328
3 0.082243 0.082243 0.082235
2 0.115172 0.115172 0.115154
The »exact» values in Tables I and 2 have been calculated using the trapezoidal rule
with 20,000 subintervals. Table I shows that the 2-point formula gives less than 5
correct decimals when 1/L < 4. The 4-point formula is accurate for alli/L > 2, but for
distance ratios less than 2, more integration points are needed. This is shown in Table 2
(see also Figure 3).
Table 2. Values of a characteristic integral in H" for

distance ratios between 0.1 and 1.0
IlL »exact» 10 g.p. 4 g.p.
1.00 0.029552 0.029552 0.029559

0.50 0.066141 0.066141 0.065957
0.25 0.110832 0.110830 0.108785
0.10 0.169263 0.169436 0.202618
Table 2 shows that the 4-point formula gives less than 5 correct decimals when
IlL< !.The 10-point formula gives accurate results for distance ratios greater than 0.25.
Two important questions can now be answered:
• What is the optimal number of integration points for a specific case? (The
program should be able to make this choice automatically.)
• How should the BE-discretization be performed? (Since the program puts
the collocation points at the boundaries or at the middle of the element, one
can give guidelines concerning the element lengths.)
The answer to the first question is:
• Use 10 g.p. when 0.25 < IlL < I

• Use 4 g.p. when I < IlL < 4
• Use 2 g.p. when 4 < IlL
526
If the above integration rules are applied, the following guidelines for the BE-
discretization can be given:
• for linear elements, 0.25 < L1/L 2 < 4

• for quadratic elements, 0.5 < L 1/L 2 < 2
where L 1 and L 2 are lengths of two adjacent elements. Thereby, the second question is
answered. The geometrical configuration in Figure 3 implies that the integrand is
strongly influenced by the trigonometric functions. The formula proposed by Watson
(1980) neglects this influence, since the only term considered is 1/r.
BEMSTAT chooses dynamically the number of integration points, 2, 4, or 10,

depending on the distance between the pole and element, according to the above rules.
BEMSTAT has also an algorithm for automatic refinement or enlargement of the
elements, which make~ it easy for the user to follow the guidelines given concerning
element ratios.
Interior integration
When the collocation point is located in the element under consideration, both !f• and
a• become singular. The a• matrix is, however, only weakly singular and C3J! be
integrated numerically by a Gauss quadrature with logarithmic weight function. He is
strongly singuhr, and the integral exists only in the sense of a Cauchy principle value
and cannot be integrated numerically. In a finite domain with a closed boundary, this is
overcome by the well-known rigid body motion trick. However, when studying
problems with infinite or semi-infinite domains with an open element model, this trick
does not work and the terms must be integrated analytically. This is possible when the
geometry is approximated with straight line elements, and has been done by the
authors' for linear and quadratic shape functions in BEMST AT. Hence, infinite and
semi-infinite domains can be treated in the program by using open element models. The
contribution from the singular integrals to the element matrices for linear shape
functions is shown in Appendix A.
COUPLING OF HEM AND FEM DISCRETIZED STRUCTURES
The structure in Figure 4 is divided into two regions, A and B. A is dicretized with finite
elements and contains N +I number of nodes; B is discretized with boundary elements
and contains M +I nodes; I is the number of interface nodes
QA .-- rl nB .-- r B
------
A B
Figure 4. Structure discretized with finite elements,
region A, and boundary elements, region B
The discr[::ed :~ilte[;~e]m~nf 1~]uations for QA can formally be written

(51)
5rA 5 rr ui lii
(N+I) X (N+I)
where S is the stiffness matrix, u the nodal displacement matrix, and ?A the equivalent
nodal forces acting on A. ?1 is defined as
11 = f H~dS (52)
rr
527
where is the finite element shape function matrix for displacements on the interface
and tt the traction acting when A and Bare uncoupled. The system in Equation 51 can
be condensed to the interface variables by static condensation, which gives
s~ 1 u 1 = I* + I 1 (53)
(Ixi)
where formally
5 ir = 5 11 - 5 rA 5 ~ 5 AI (54)
f* = -S 1 AS~~fA
For region B the boundary element equation can be written as
(55)
(M+I) X (M+I)
When the boundary conditions on region B are introduced in Equation 55, a non-
square system is obtained
[s -G1 HI] 1:: = [o] (56)
(M+I)x(M+2I) u1 (M+I)
v r
where B is a mix of H 8 and 0 8 and 8 is the unknowns on 8 . There are too many
unknowns in this system, and the missing equations have to come from region A. Up to
now, A and B have been looked upon as two uncoupled structures. They are now
coupled together by forcing the displacements at the interface to be equal and the
tractions to be equal in magnitude but with opposite signs, i.e.,
t~ = -t~ (57)
If the same shape functions are used in A and B, this implies that
(58)
The tractions on r 1 are approximated in the usual BE-manner
t~ = ~Ttl (59)
and by use of Equation 57
(60)
Introducing this in Equation 52 and rewriting Equation 53 gives
(61)
where the mass-type matrix is defined as

MII = f ~~TdS (62)
rr
528
Equations 56 and 61 can now be added together to form a square, almost dense, and
non-symmetric system of equations which is solvable:
(63)
(M+2I)X(M+2I)
This method of coupling is easy to implement and no explicit matrix inversion is

needed. Other methods of coupling FEM and BEM are presented in Kelly eta/ (1980).
NUMERICAL STUDIES
BEMST AT is divided into four separate modules, and the user can choose which
module should be executed, by including the proper subgroup(s) of data, as illustrated
in Figure 5. A short description of the modules is given below:
Module 1: Reads and generates the geometry and topology.

Module 2: Reads and generates the boundary conditions, solves the system of
equations, and prints the boundary results. This module can be
executed again with new boundary conditions, without reexecuting
module I.
Module 3: Calculates requested internal quantities such as displacements, strains,
stresses, and principal stresses. This module can be executed
interactively, which enables the user to »zoom in» on regions of
interest.
Module 4: Plots geometry with node and element numbers, boundary
displacements, boundary tractions in x- and y-directions, and
principal stresses.
Shear wall
The plane stress problem shown in Figure 5 is studied.
...........1
SHEAR WALL - 16 QUADRATIC ELEMENTS
0.3333 GENERAL
..:.....:.. 1 9
GEOMETRY
16 ALL 2
1TO 2 0 J 3 0
3TO 83039
9 TO 10 3 9 0 9
11TQ160900
TOPOLOGY Module l
E-modulu., I 1T0231T05
Poi..,,.on'-. rat10 0.3 3 TO 8 3 6 TO 18
thilkm•o;;, I 9 TO 1Q 3 19 TO 23
11 TO 16 3 24 TO 36
SUBSTRUCTURE DATA
3 1 1 •3 1
BOUNDARY CONDITIONS
1TO 2X3Y2
3 X 1 Y 2
4TO 5X3Y2
6T036X3Y4
19 TO 23 X .333333 END
SUBSIDIARY CONDITIONS
5 6
18 19
,'/,
23 24
1 36
PLOT
GEOMETRY NODES
BOUNDARY DISP 0. 1 r:f 1
Figure 5. Plane str,ess problem - shear wall
Typical for the problem is the large rigid body motions, i.e., 1t 1s in principle a
deflection problem. Different types of discretizations have been used, and the results
are compared by studying the horizontal deflection at point A (Figure 5). Subsidiary
529
conditions are used to couple the nodes in the corners. Table 3 shows the results; listed
is also the CPU-time on an IBM 3033N required for the establishment and solution of
the system of equations.
Table 3. FEM- and BEM-results for the shear wall problem
horizontal CPU-time(sec)
deflection establ. solution
at A
48 finite elements 118.7 1.59

96 linear elements 117.4 6.55 3.28
64 linear elements 115.4 2.94 1.06
16 linear elements 83.9 0.22 0,03
48 quadratic elements 118.9 3.76 3.32
32 quadratic elements 118.7 I. 72 1.06
8 quadratic elements 116.2 0.14 0.03
The finite element calculation was made using 48 (4xl2) 8-noded isoparametric
elements, and the time given is for both the establishment and solution. Two conclusions
can be drawn from this study, namely: I) Quadratic elements give better results, and 2)
They also require much less time for the establishment. Constant elements cannot descri-
be this problem, since they are subparametric. Figure 6 shows plots of node numbers,
displacements, and tractions for 16 quadratic elements.
node numbers displacements tre.ctions Y
23 2o€'2 18 19
"
28
29 13
3S t 2 3 4
Figure 6. Plots of node numbers, displacements, and tractions

530
Rectangular plate with a circular hole
The stress concentration near a circular hole in a rectangular plate is studied. Due to
symmetry, the problem can be reduced to that shown in Figure 7.
Figure 7. Rectangular plate with a hole
Table 4 shows the calculated results, horizontal stress at A (Figure 7), and vertical stress
at B.
Table 4. Stresses in a rectangular plate with a hole
horizontal vertical number of

stress at A stress at B equations
69 linear elements 3.25 1.18 148

40 quadratic elements 3.24 1.18 170
24 quadratic elements 3.21 1.16 106
Wennerstrom and
Petersson (1979), FEM 3.31 214
Roark (1965), empirical 3.235
Both linear and quadratic elements gave good results in this typical stress concentration
problem. Also for constant elements the result converged, but many more elements
were needed. However, the quadratic elements are still more attractive, since they
require less time for the establishment. During the study of this example, it became
obvious that the representation of the geometry of the hole was important (as reported
by Wennerstrom and Petersson). The difference in the results between a polygonial
and curved boundary was about 5 O!o. Horizontal tractions are plotted in Figure 8.
traction• X
Figure 8.
:'--'--:
Horizontal tractions
_)-
531
Strip load on a finite layer

Figure 9 shows the problem and the points A and 8, where displacement and traction,
respectively, are studied. The problem has been calculated with Poisson's ratio equal to
0 and 0.5. Quadratic elements have been used.
~<Ar•M>..,.JAr}M£
~
j E·modulus
I Poisson's ratio 0 , 0. 5 2.0
I
I
lx /-8
Figure 9. Strip load on a finite layer
Figure 10 shows the discretization with 18 elements and Table 5 calculated and
analytical results.
element numbers
u:m.
•11 +10+9+8 +•+•+ + +• +
'12 7 s 5 .. 3
6.0
Figure 10. HEM-discretization with 18 quadratic elements
Table 5. Analytical and calculated results for a strip

load on a finite layer
v = 0.5 v = 0
analytical calculated analytical calculated
vertical stress at 8 0.35 0.348 0.35 0.350
vertical displ. at A -0.84 -0.866 -0.38 -0.378
horizontal displ. at A -0.38 -0.370 0.06 0.070
532
The analytical results are given with two decimals only, due to the accuracy of the
diagrams in Poulos and Davis (1974). In Figure II the vertical stress distribution at the
bottom surface and the displacements of the upper surface are shown. Two conclusions
can be drawn from this example, namely: I) Incompressible materials can be treated,
and 2) Semi-infinite domains can be approximated with open element models. The
CPU-time required for the establishment and solution of the system of equations was
0.62 and 0.23 seconds, respectively.
displacements tractions Y
llilJ
Figure 11. Vertical stress and displacements
Strip foundation on a semj-infinite mass
The contact pressure distribution beneath a smooth strip subjected to a uniform

pressure is studied (Figure 12.) This example was calculated using 20 four-node finite
elements and 50 linear boundary elements. The discretization was refined at the strip
ends.
t:t m:~•$~$~~£M~$~~;~~~~'
I b - ~
f '
Figure 12. Strip foundation on a semi-infinite mass
Two calculations, with different relative stiffnesses, have been made. The relative
stiffness is defined below. Figure 13 shows plots of the displacement profile and the
contact pressure distribution for the two calculations. Borowicka (1939) has presented
solutions for the distribution of contact pressure p beneath a strip subjected to a
uniform pressure q, and some of his results are indicated in the plots. No difference
between the plots below and plots of Borowicka' s result can be observed. All
calculations were performed assuming plane strain conditions.
533
K= iJ
-I xfb I ·I xfb
~ ; ~
0.5
1.0
1.5
K= '73 K= "'
/30
2.0
2.5
x Borowicka (1939}
%
Figure 13. Displacements and contact pressure beneath
uniformly loaded smooth strip
The relative stiffness, K, is defined as

2
1 ( 1-vs) E t 3
K =- - - - ...EH (64)
6 ( 1 -v2) E 8 b
p
where EP is the elastic modulus of the strip, vll. Poisson's ratio of the strip, E5 the elastic
modulus of the mass, v5 Poisson's ratio of ot the mass, t the strip thickness, and b the
half width of the strip. This example demonstrates one of the features of coupling BEM
and FEM, namely, the possibility to treat soil-structure interaction problems.
CONCLUSIONS
BEMSTAT, a boundary element program for two-dimensional elasticity problems, has

been presented. The program is easy to use, it requires a small amount of input data,
the output facilities are extensive, and it is built up of well-separated modules (allowing
restart). Due to the new philosophy applied when integrating and establishing element
matrices, the computer time required has been reduced considerably. Infinite and semi-
infinite problems can be treated by open element models, since the strongly singular
integrals are calculated analytically. Non-homogeneous materials can be handled by
dividing the body into a number of homogeneous substructures. Coupling with FEM
can be made, which makes the program well suited to solve soil-structure interaction
problems. Practical rules for discretization and integration are given in the paper.
The examples studied show that typical deflection problems are best solved with
quadratic elements, and few elements are needed. Constant elements cannot describe
these types of problems. In stress concentration problems, quadratic elements are still
superior to the linear ones, but not as much as for the deflection problems. Constant
elements can describe stress concentration problems if a large number of elements are
used.
534
ACKNOWLEDGEMENT
The support received from IBM Sweden, in the form of computer time, is greatly
appreciated.
REFERENCES
Alarcon, E., Martin, A., and Paris, F. (1979), Boundary elements in potential and
elasticity theory, Computers and Structures, 10, pp. 351-362.
Banerjee, R.K. and Butterfield, R. (1979), Developments in Boundary Element
Methods, Appl. Science Publishers Ltd, London.
Borowicka, H. (1939), Druckverteilung unter elastischen Platten, lngenieur Archiv,
vol. X, No. 2, pp. 113-125.
Brebbia, C.A. and Walker, S. (1980), Boundary element techniques in engineering,
Newnes-Butterworths, London.
Hartmann, F. (1980), Computing the C-matrix in non-smooth boundary points, Proc.
of the Second Int. Sem. on Recent Advances in Boundary Element Methods, held
at Univ. of Sothampton, March 1980.
Jaswon, M.A. and Symm, G.T. (1977), Integral Equation Methods in Potential Theory
and Elastostatics, Academic Press, New York.
Jeng, G and Wexler A (1977), Isoparametric, finite element variational solution of
integral equations for three-dimensional fields, Int. J. for Numerical Methods in
Eng., II, pp. 1455-1471.
Kelly, D.W., Mustoe, G.W, and Zienkiewicz, O.C. (1979), Coupling Boundary
Element Methods with Other Numerical Methods, Developments in Boundary
Element Methods - I (editors: Banerjee and Butterfield), Appl. Science Publishers
Ltd, London.
Nedelec, J.C. (1977), Cours de !'Ecole d'Ete' d' Analyse Numerique, C.E.A.,
I.R.I.A., E.P.F ..
Poulos and Davis (1974), Elastic solutions for soil and rock mechanics, John Wiley &
Sons, New York.
Roark, R.S. (1965), Formulas for stress and strain, McGraw Hill, New York.
Watson, J .0. (1980), Advanced implementation of the boundary element method for
two- and three-dimensional elastostatics, Developments in Boundary Element
Methods - I (editors: Banerjee and Butterfield), Appl. Science Publishers Ltd,
London.
Wendland, W.L., Stephan, E., and Hsiao, G.C. (1979), On the integral equation
method for the plane mixed boundary value problem of the Laplacian, Math.
Meth. in Appl. Sci., I, pp. 265-321.
Wennerstrom and Petersson (1979), GENFEM-3, Verification Manual, Pub!. 79:5,
Dept of Structural Mechanics, Chalmers University of Technology, Goteborg,
Sweden.
535
APPENDIX A- DEFINITIONS OF MATRICES USED IN BEM FOR TWO-

DIMENSIONAL ELASTOSTATICS
A. I Fundamental matrices
The fundamental matrices for two-dimensional plane strain elasticity are defined as
1+v-v) [ (3-4v)ln;::1
U(x,~) = 4 1TE( 1 +rx2 rxry ] ( 6 S)
SYM (3-4v)lnl+r 2
r y
T(x,~) =41T(11_v) ~[ ~~((1-2v) +2r;) 2~~rxry+ (1-2v) (rxny-rynx)] (66)
2~~rxry + (1-2v) (rynx-rxny) ~~ ((1-2v) + 2r;)

where the pole is placed in ~. and
(67)
The fundamental matrix for the tractions can be derived in the following manner:
t(x) = nT(x)a(x) = nTSdE = nTSdVxu(x) = nTSdVxU(x,~)e(~) = T(x,~)e(~) (68)
(69)
where the unit normal vector is defined by

n =
(70)
A.2 Gradients of fundamental matrices
When the strains and stresses are to be calculated the gradients of U and T must be
derived. It should be noted that the gradient operator worked on the free argument, not
the pole coordinate, when the T-matrix was derived. Here, the gradient operators
operate on the pole-argument. By using the following identities:
l
'J
X
= -'J
~
(71)
the old definitions of gradients of r, Equation 67, can be kept, so(~ is the pole)
.,
v U
~ -A[
=--
r
r
X
(2(1-r )-B) 2
X
2
r y (1-2r X ) (72)
r
X
(1-2r 2 )
y ry(2(1-r~)-B)
r (1-B-4r 2 ) r ( 1-B-4r 2 )
y X X y
where
A = 1+v B 3-4v (73)
4nE(1-v)
536
and
2 2 ar
(2r -1)0n +2(1-0)r r n + (1-4r )2-ary
X y y X X X n (74)
2 ar 2
(2ry+O)ny + (2-0-4ry)2anry
2 2 ar
(20+ ( 1-0) 2ry) nx + 2 (0+1) ryrxny + ( 1-0-Sry) 2anrx
where
0 1-2v (75)
A.3 Exact calculation of the free term matrix for non-smooth boundaries
The exact calculation of the free term matrix for two-dimensional elastostatics was
shown by Hartman (1980). With the definitions in Figure 14, the matrix is
O(x) = 41r(~-v) [(4rr-2 (a 1 -a 2 )) (1-v) +1(sin2a 1 -sin2a 2 ).
SYM
sin2a1 - sin2a2 ] (76)

(4rr-2 (a 1 -a 2 ) (1-v) -1(sina 1 -sina 2 )
Figure 14. Definition of positive rotation and angles involved
A.4 Exact integration of the singular integrals for straight line elements
When the boundary is approximated with straight line elements, it is possible to

integrate the singular terms exactly. Here, the result of the interior integration of an
element with linear shape functions is shown. The collocation point is in node I, where
I= 1 or 2 for linear shape functions:
A[(OLX) 2 +B(2,5-I+lnL) OXOY
7
OXOY
{> 2 +B(2,5-I+lnL)
<0
7
(OLX) 2 + B (I-0, S+lnL)
OXOY
7
537
lie = C[ 0 1-!+(2-I)lnL 0
(78)
-(1-!+(2-I)lnL) -(2-!+(1-I)lnL)
where
A = 1+v L B 3-4v
-1
C = 4n(1-v)(1-2v) (79)
4nE(1-v) 2
(80)
and L is the element length. Notice that the contribution from the free term matrix is
not included in H 6 •
Section VI
Coupling of Boundary and
Finite Element Methods
541
INTERFACING FINITE ELEMENT AND BOUNDARY ELEMENT

DISCRETIZATIONS
carlos A. Felippa
Applied Mechanics Laboratory
Lockheed Palo Alto Research Laboratory
Palo Alto, California 94304, USA
INTRODUCTION
This paper reviews critically various implementation
methods for coupling a mechanical finite element model
to an infinite external acoustic domain discretized
through boundary-element techniques. The associated
physical problem is that of a three-dimensional
structure submerged in an acoustic fluid, and impinged
by a pressure shock wave.
Three coupling methods for advancing the dynamic
calculations are described: field elimination,
simultaneous integration, and partitioned integration.
variants of these techniques have been tried on the
case problem over the past seven years.
The three methods are assessed from gained experience,
and their advantages and disadvantages noted. some
generalizations to more general FE/BE coupling
scenarios are then offered.
THE PROBLEM
The specific problem used as case study in this paper
is illustrated in Figure 1. A linear or nonlinear
three-dimensional structure is submerged in an
infinite acoustic fluid. A pressure shock wave
propagates through the fluid and impinges on the
structure. The structure and fluid are discretized
through finite-element (FE) and boundary-element (BE)
methods, respectively.
542
Magic Membrane 11 ,
11
Boundary -Element
Representation of
Infinite Fluid
Linear or
Nonlinear
Structure
Figure 1 Structure submerged in an acoustic fluid
Figure 2
c•
Layout of BE and FE meshes on wet surface
= BE control points; • = FE node points)
543
Before proceeding to the governing equations, two

practical considerations of relevance to this problem
should be mentioned.
First, the structural response (and most especially
the structure's survivability) is of primary concern,
whereas what happens in the fluid is of little
interest.
Second, the FE and BE meshes on the "wet surface" are

not necessarily in one-to-one correspondence, as
illustrated in the two-dimensional sketch of Figure 2.
Rather, a "fluid BE" typically overlaps several
structural elements. This ties up with the first
consideration in the sense that determination of
structural deformations and stresses demands a finer
subdivision.
Structural response equations
The governing matrix equation of motion for the

dynamic response of a discrete structure is
~s ! + ~3 ! + ~s x = f + N (l)
where x = x (t) is the structural displacement

vector, -~s ~ £s and ~s are the structural mass,
damping and stiffness matrices, respectively, i is
the external force vector, N = N(x) is a nonlinear
psudo-force vector, and a dot - denotes temporal
differentiation.
For excitation of a submerged structure by an acoustic
wave, 1 is given by
i = -g ~f <Er + Es > (2)
where ~ and Ps are nodal pressure vectors for the

wet-surface r1uid mesh pertaining to the (known)
incident wave and the (unknown) scattered wave,
respectively, ~f is the diagonal area matrix
associated with elements in the fluid mesh, and G is
the transformation matrix that relates the structural
and fluid nodal forces. Introduction of this matrix
takes care of the FE/BE "mesh-mismash" noted
previously (cf. Figure 2).
Fluid equations
The response of the fluid is modelled by the Doubly

Asymptotic Approximation (DAA) of Geers (1971,1978)
r:!r Es + pc 8r Es = pc r:!r .l!s (3)

544
where !!s is the vector of scattered-wave

fluid-particle velocity normal to the structure's wet
surface, p and c are the density and sound velocity of
the fluid, respectively, and ~f is the symmetric fluid
mass matrix for the wet-surface fluid mesh. This
matrix is produced by a boundary-element treatment of
Laplace's equation for the irrotational flow generated
in an infinite, inviscid, incompressi~ fluid by
motions of the structure's wet surface; it is fully
populated with non-zero matrix elements. When
transformed into structural coordinates, the fluid
mass matrix yields the added mass matrix, which, when
combined with the structural mass matrix, yields the
virtual mass matrix for motions of a structure
submerged in an incompressible fluid. Details of the
calculation procedure may be found in a paper by
DeRuntz and Geers (1978).
The approximate pressure-velocity relation (3) is
called "doubly asymptotic" because it approaches
exactness in both the high-frequency (early-time) and
low-frequency (late-time> limits. For high-frequency
motions, 1g8 1 >> 1p8 1, so that (3) approaches the
relation Ps = pc !!s , which is the correct limit for
short acoustic wavelengths. For low-frequency
motions, IPs I « IPs I, so that (3) approaches the
incompressible-flow re~ation ~f Ps = ~f ~s , which
is the correct limit for long acoustic wavelengths.
For excitation by an incident acoustic wave, g is
related to structural response by the kinematic
compatibility relation
(4)
where the prime superscript denotes matrix

transposition. Equation (4) expresses the constraint
that normal fluid-particle velocity match normal
structural velocity on the wet surface of the
structure. The fact that the transformation matrix
relating these velocities is Q' follows from the
invariance of virtual work with respect to either of
the wet-surface coordinate systems.
Generally, Q is a rectangular matrix whose height
greatly exceeds its width, inasmuch as the number of
structural DOF usually exceeds considerably the number
of fluid DOF, as noted previously. Typical numbers:
5000 structural DOFs and 160 fluid DOFs.
545
Interaction Equations
The introduction of Equation (2) into (1) and (4) into
(3) yields the interaction equations
~s ~ + ~s ~ + !Ss ~ + N = -~ ~f (!II + Ps>

(5)
~f .Ps + pc ~f
.Ps = pc r:!r ( 2, x -
I
~I)
The computational structure of these coupled systems

is very different. As can be expected, the FE
(structural) system is usually ~ but §Parse. The
BE (fluid) system is typically small but dense. It is
therefore of interest to design solution methods that
exploit these attributes to maximum advantage.
SOLUTION APPROACHES
Three approaches to solving the coupled FE/BE system
(4) are reviewed here. They are presented in
chronological order, i.e., in roughly the same
sequence as they were tried and evaluated over the
past seven years.
Field Elimination
The first approach tried is now known as "field
elimination", where the term "field" refers to one of
the physical components of the coupled system
(structural and fluid in our case). As noted
previously, the structural response is of primary
interest. It is therefore natural to think of
eliminating the scattered-pressure vector Ps from the
coupled equations of motion (5). If - ~ is the
identity matrix, this results in the following
third-order ODE system for the structural
displacements ~:
-1
~s ~· + [Qs + ~f ~f f:!s 1 ~ + ISs ~ +
-1
+ 8r f:lr !5s ~ = r ( ~f, ~f· .PI' f!_I ' !:!) (6)
where the RHS vector ! accomodates incident

pressure and incident fluid-particle velocity boundary
conditions, and nonlinear effects. If ~ is not the
identity, as invariable happens in realistic
three-dimensional problems, the coefficient matrices
in Equation (6) become considerably more complex,
because the generalized inverse of g enters the
elimination process.
546
The structural response ~ (t) can no~1 be determined

by numerically integrating Equation (6).
This was in fact the first approach tried to tackle

the time-integration of the coupled system; cf.
Felippa et al. (1974). Although moderately sucessful
for the first problem series (submerged shells of
revolution, linear structural behavior>, from current
perspective it can be properly characterized as a poor
strategy that eventually leads to a "computational
horror show" for more general problems. Why?
1. The order of the reduced differential system is

raised (in this example, from two to three). The
appearance of higher derivatives can be the source
of many difficulties, the worst of which is noted
next.
2. Proper treatment of initial conditions is

complicated by the increased ODE order. In our
case study, it turned out that (6) had to be
integrated once (yielding an integrodifferential
system) so as to regularize the treatment of
wavefront-induced singularities. Time integrals
of forcing terms had then to be carrying along in
the calculations -- a grievous burden.
3. Sparseness and symmetry attributes of the original
matrices is adversely affected by the elimination
process, as can be observed in (6) for the
identity g. For a general transformation matrix,
all left-hand matrices become unsymmetric and
dense.
4. The development of specialized software is
required. For example, available software for
dealing with the uncoupled problems (structural
dynamics and acoustic shocks) separately is not
likely to be of much use in solving the reduced
system (6).
Simultanous integration
In this approach Equations (5) are viewed as a single

second-order system
8d~r~
Mf Y.r
• • • ( 7)
547
This approach removes many of the objections raised

against the field elimination technique. Inasmuch as
the ODE order is not raised, initial-condition
difficulties do not arise and better use can be made
of existing software for dealing with second-order ODE
systems.
The computational burden for realistic
three-dimensional problems can be prohibitive,
however. Note that the coefficient matrix of
higher-derivative terms is singular, which means that
implicit integration is required to construct a
marching scheme in time.
But the assembly and formation of the implicit
coefficient matrix was found to pose enormous
computational demands because of presence of BE/FE
coupling terms that can extend across thousands of
equations.
For example, it was estimated that the factorization
of that matrix for a 5000-DOF problem would require 3
hours on a Cyber 175 computer. Carrying out a
nonlinear transient- response analysis of a realistic
model was then adjudged infeasible.
Partitioned integration
In the partitioned integration approach, the solution
state is advanced over each of the two subsystems: FE
structural model and BE fluid model, in a staged
fashion. Interaction terms are treated as "forcing"
actions that have to be judiciously extrapolated.
What is now called the staggered solution procedure is
a specific partitioned-integration method originally
formulated for the system (5) by Park, Felippa and
DeRuntz (1977). A version of this procedure was
implemented in a production- level computer program
described by DeRuntz et al. (1978).
success of this method led to further applications and
eventually the development of a general theory of
partitioned time integration; see Park (1980), Park
and Felippa (1980). A state-of-the-art review of
formulation aspects is provided in a recent survey by
Felippa and Park (1980) •
The staggered solution procedure was found to offer
two important advantages: enhanced software
modularity and computational efficiency.
548
The first advantage accrues from the fact that

relativ~ly few modifications to programs available for
process1ng the uncoupled systems are necessary. Given
current costs in software development, augmentation
and maintenance, this is indeed an important virtue of
this approach. For our specific problem, a BE fluid
analysis module was written, and data-coupled to
existing large-scale structural analysis codes such as
NASTRAN, SPAR and STAGS.
An obvious advantage of "plug-in" modularity is the
freedom afforded the analyst as regards the selection
of a structural analysis code that best fits the
problem at hand; for example, the nonlinear analyzer
STAGS when plasticity or finite displacements had to
be considered. Moreover, if there is a choice among
structural analyzers that can do just about the same
thing, the user can select the one he or she is most
comfortable with.
As regards computational efficiency, the cost per time
step is roughly the same as adding up those incurred
in processing the FE and BE models as isolated
entities. This is because the overhead introduced by
the flow of information (which consists primarily of
computational vectors) among the two analysis modules
becomes comparatively insignificant in large-scale
problems. It follows that the staggered integration
procedure appears as economically attractive should
time stepsize considerations be excluded from
consideration.
Unfortunately, the latter assumption was not easy to
realize in practice. The high computational
efficiency per time step is counteracted by the fact
that satisfactory numerical stability properties are
hard to achieve; in fact, the practical feasibility
of this technique hinges almost entirely on the
stability analysis. The reader is referred to the
cited sources for additional details. suffice here to
say that a specific integration algorithm of
unconditional stability was found for (7) when the BE
fluid was suitably modified through an "augmentation"
technique.
549
FINAL REMARKS
It is time now to summarize some lessons gained from
this case study of FE/BE interfacing, and to venture
various generalizations. This will be done following
a question-and-answer style.
We have used boundary-element techniques primarily for

dynamic problems; in particular, wave propagation,
impact and reentry studies. This has provided a body
of experience "complementary" to that gathered by
investigators dealing with static problems.
Moreover, the computer implementation of dynamic and
nonlinear-static analysis share many common facets.
In fact, using the dynamic relaxation concept, the
latter can be always viewed as pseudo-dynamic systems
whose steady-state solution is sought, and much of the
discussion on time-marching algorithms apply.
Our experience has been that BE techniques are

primarily useful for discretizing unboundeq
homogeneous domains governed by linear equations. For
linear problems in bounded homogeneous domains, they
are not competitive with properly written finite
element codes, and the latter are far easier to extend
to nonhomogeneous regions and nonlinear problems.
Advertised reductions in mesh preparation efforts are
largely illusory in these days of powerful
pre-processors complemented by inexpensive interactive
graphics.
But boundary-element methods come on their own for
treating unbounded linear media, notably when their
"interior" response is of little interest. The
model-description effort is greatly reduced, and the
construction of "quite boundaries" is simplified.
As "loose" as possible. This goes along with the

philosophy of maximizing softwar~ modulariiY. One
facet of this philosophy says that the analyst ought
to have the freedom of selecting FE and BE meshes
independently according to the physics of the problem
and response-resolution requirements.
In our problem, for instance, it would be plain silly
to force the BE fluid mesh to be constrained by the
presence of internal structural stiffeners.
550
The analysis of fluid-solid impact and similar

problems requiring moving and sliding interfaces
provide further ammunition on the argument for a high
degree of independence in FE/BE mesh definitions .
~lhat' s 2. most compellj.n.g reason for modularity?
Keeping complexity down and analyst's sanity up.

Nontrivial FE and BE programs tend to be fairly
complex beasts even when taken separately. And the
complexity of a monolithic marriage can easily escape
anybody's control.
There are many things that can go wrong and (true to

Murphy's law) will: modelling, numerics, machine
problems, data management, result interpretati on.
Nonlinear dynamics problems, for example, are
particularly vulnerable to many trouble sources.
Experience has shown that keeping modular interfaces
not only reduces the chances for troubles, but also
makes their resolution prompter.
Do modularity requirements impact solution strategies?
Yes. And that's a major reason for the advent and

high degree of success of partitioned analysis
procedures.
1§ success or failure of partitioned integration
proced~. con~~ upon implementati on details?
Very much so. The survey paper by Felippa and Park

(1980) highlights this theme.
ACKNOWLEDGEHENT
Preparation of this paper was supported by the
Independent Research Program of Lockheed Missiles and
Space Co., Inc.
REFERENCES
DeRuntz, J. A. and Geers, T. L. (1978) Added Mass

Computation by the Boundary Integral Method, Int.
Journal Numerical Methods in Engineering, ~. 531-550.
DeRuntz, J. A., Geers, T. L., and Felippa, c. A.

(1978) The Underwater Shock Analysis (USA) Code: A
Reference Manual, Report DNA-4524F to Defense Nuclear
Agency, Lockheed Palo Alto Research Lab., Palo Alto,
California.
551
Felippa, C. A., Geers, T. L., and DeRuntz, J. A.

(1974) Response of a Ring-Stiffened Cylindrical Shell
to a Transient Acoustic wave, Report LMSC-D403671 to
Office of Naval Research, Structural Mechanics
Laboratory, Lockheed Palo Alto Research Lab., Palo
Alto, California
Felippa, c. A. and Park, K. C. (1980), Staggered

Solution Transient Analysis Procedures for Coupled
Hechanical Systems: Formulation, Computer Methods in
Applied Hechanics and Engineering, 2,!, 61-lll.
Geers, T. L. <1971) Residual Potential and
Approximate Methods for Three-Dimensional
Fluid-Structure Interaction, Journal of the Acoustical
Society of America, ~. 1505-1510.
Geers, T. L. (1978) Doubly Asymptotic Approximations

for Transient Motions of Submerged Structures, Journal
of the Acoustical Society of America, QA, 1500-1508.
Park, K. c. ( 1980) Partitioned Transient Analysis

Procedures for Coupled-Field Problems: Stability
Analysis, ASHE Journal of Applied J.'.lechanics, 47,
370-376.
Park, K. c. and Felippa, c. A. (1980) Partitioned

Transient Analysis Procedures for Coupled-Field
Problems: Accuracy Analysis, ASME Journal of Applied
Mechanics, fl, 919-926.
Park, K. c., Felippa, C. A. and DeRuntz, J. A. (1977)

Stabilization of Staggered Solution Procedures for
Fluid-Structure Interaction Analysis, in:
T. Belytschko and T. L. Geers (eds.l, Computational
Methods for Fluid-Structure Interaction Problems, ASME
Applied Mechanics Symposia Series, AMD-Vol. 26,
94-124.
552
THE DERIVATION OF STIFFNESS MATRICES FROM INTEGRAL EQUATIONS
Friedel Hartmann
University of Dortmund, Dept. of Structural Mechanics
Introduction
Besides Finite Elements in the domain we apply today also in-
tegral equation methods on the boundary to problems in elasticity.
This boundary element method (B.E.M.), as it is called, leads to
non-symmetric and fully populated matrices and it is not easy to
convince engineers, that anything good can come out of such
strange matrices.
Immagine these same engineers when they discover that the
stiffness matrices which they derive from these linear systems
turn out to be non-symmetric •••• God knows there is not much
engineers believe in, but they certainly believe that stiffness
matrices have to be symmetric and they are therefore deeply
irritated, finding themselves in a situation where two funda-
mental beliefs clash (Mathematics and Mechanics) and either one
is equally convincing.
But naturally the situation is not that dramatic and we will
see, as you might have guessed, that also B.E.M. lead to sym-
metric stiffness matrices, though not so easily as F.E.M. To see
how this is achieved and to justify the modifications we must
necessarily bring about, we make a small detour by studying in-
stead of th~ 1 original equatio~ 1 Hu = Gt the equivalent integral
equation G Hu_l t, where G H is a self-adjoint operator. If
the equation G Hu = t is solved with Galerkin's method or by
minimizing a potential IT (u), we obtain*indeed a stiffness ma-
tr~x, which has all three*properties Pl (Rigid body movement),
P2 (Equilibrium) and P3 (Symmetry), commonly associated with
such a matrix. So there is nothing wrong with integral equation
methods.
But solving the equation G- 1Hu = t on a finite set o~ 1 boun
dary elements amounts_£o substituting for the operator G H a
modified operator PG HP, where P stands for a collocation or
an orthogonal L2 -projecti£j onto the set of nodal shape functions.
While the 1 2 -operator PG HP (i.e. if P is the L2 -projection)
preserves the properties, which guar~jtee a correct stiffness
matrix, the collocation operator PG HP is no longer self-adjoint
553
and looses also the feature that guarantees the property P2 * (Eq.)
in its stiffness matrix.
These remarks are first hints at the problems we encounter if
we do numerical computations. But prior to computations, we h~ye
to cope with the fact that an expression for the operator G H
is unknown and we therefore have to return to the original equa-
tion Hu = Gt and use the matrices PHP, PGP (the operators P~l'
PGP etc. coincide wit~ 1matrices) to approximate the matrix PG HP
with the product PGP PHP. Using in practical computations nu-
merical i~Iegration amounts furthermore ~~ approximating the pro-
duct PGP PHP with the product P'G'P' P'H'P', where the primes
denote the modified operators. So we end up with the chain:
The only opeE!tor at hand is the one last in line, i.e. the ma-
trix P'G'P' P'H'P' and neither the collocation nor the L2 -
version of it preserves the thEre properties_ 1Pl, P2, P3 (see
below), which originally in G H or in PG HP (if P is the
L2 -projection) guaranteed a correct stiffness matrix.
Being confronted with the task to obtain a correct stiffness
matrix from P'G'P' and P'H'P' we can now, on going back-
wards, analyze where we loose the single properties needed for
a corr~It stiffness matrix and modify the approximation
P'G'P' P'H'P' accordingly.
---our main concern is the symmetry of the stiffness matrix and
this is automatically achieved, if we calculate the stiffness
matrix by minimizing the pote~£ial IT(u), but we could also
ma~I the L2 -operator P'G'P' P'H'P' self-adjoint, (because
PG HP has this property) and then u~e Galerkin's method.
With*respect to the properties Pl (Rigid body movement)
and P2 (Eq.) we must either modify the matrices to ensure these
properties or hope that the defect is negligible. This should
occur rather in the L7 -matrices than in the collocation matrices.
In the following we wiil discuss all this in more detail.
554
1. Finite elements in the domain
W~dstudy in the bounded domain B with smooth boundary S the

2 b.v.p. ("Neumann problem") for the vector valued displacement
u(x)
L u 0 X E B lim T(u) on S (1)

i
where L is the differential operator of classical elastostatics
L= 1111 + --11- - grad div (2)

1 - 2\!
and
T( ) 2 11 _L +
<ln ·
~
1 - 2v
n div + 11 (n 1\ rot)
is the traction operator.

We approximate the solution u on ! set of nodal shape
functions { cp 1 , cp 2 , cp 3 , cj>N} C H (B) by a function
Pu Lu r cp (x) =
r
~T u
r
and the traction t by the function
the "vectorcomponents" ui etc. are themselves vectors, i.e. y etc.

represent "hypervectors" and consequently are all the matrices
"hypermatrices", i.e. their elements K.. etc. are themselves
(3x3) matrices). l.J
and we choose the coefficients with Galerkin's method
(see ref.[l], p. 94 and 95)
(LPu, cj> i) jE(Pu)·C·E(cj>i) dv J Pt·cJ>i ds 0 (4)

B s i = 1, 2 ...
which yields the linear system
Ku F t f (5)
=- =-
t n is the outward normal and 11 > 0, 0 ~ \) < 1/2 are the

elastic constants.
555
with
f E(¢.)·C·E(¢.) dv
]. J
B
or we choose as solution the approximation Pu
minimizes the functional
Il(u)
1
2
f E(u)·C·E(u) dv- f Pt·u ds (7)
B s
which yields again Eq.(S). The vector f in Eq. (5) represents
the so called "equivalent nodal forces":
2. Boundary elements on the surface
If t is sm2oth enou~h~ then the solution u of Eq. (1) is

regular, u € C (B) n C (B), and the Somigliana identity holds
on S
C(x)u(x) + f T(y,x)u(y)ds
y
= f
s
U(y,x)t(y)ds
y
(8)
s
or shortly
Hu G t
with t = lim 'e (u). The operator G is self-adjoint, while H

is not. i
In boundary element methods we use Eq.(8) to determine the
unknown trace u(x) approxima§Nly by a projection method, which
yields a linear equation on R
PGP t PHP u =
=-
where u and t are the vectors of (generalized) nodal values
with respect to-a projection p
u + Pu t + Pt cj>T t (10)
A formal madpulation of Eq. (9) leads then to the following equation

between the nodal displacements u.]. and the equivalent nodal
forces f.].
F PGP- 1 PHP
===
~ F t f (11)
t see ref. [2] for a definition of C(x), T(y,x) and U(y,x).

556
-1
where the (usually) non-symmetric matrix E PGP PHP appears in
the role of a stiffness matrix which irritates engineers.
In the following we want to discuss what the transition from
Eq,(9) to Eq.(ll) means and show how we can use Eq.(9) to derive
a symmetric stiffness matrix.
Suppose we study instead of the problem
H u G t (12)
the problem
t (13)
G-lH is a linear operator on c1 •Y(s) with the properties
(Pl,Ker) 0 \;j u=a+bAX (rigid body movement)
(P2, Eq.) Hu,a+bll.x)

(G
-1 ~
0 '\J ~ (equilibrium condition)
-1 -1 -2 - 1 -2
(P3,Adj .) (G H u , u ) = (u~ G- H u ) (Al 2 = A2 1 )
' '
and it is with Korn's second inequality, see ref.[l] p. 381,
-1 - -- - 2
(G H u,u) = ( T(u), u) f E(u)·C·E(u) dv > c \\u\\Hl(B)-\\u\\ 1 2(B
B
if u is the solution, L u 0, with the trace u = lim u.
Property (P3, Adj.) is a consequence of the Betti equation
2 -1
2
u · Lu ] dv
1
J [T(u 1 )•u-2 - T(u )·u ] ds (15)
s
and the fact, that for any pair of functions ~i £ c1 •Y(s),
£1 E c0 'Y (S), which stand in the relation
H ~i = G ~i (16)
i 2( ) 1(-) i . h
t ~ereiexisis aandf~?c1 t1.mion u. I: C l3 f'\ C B , Lu = 0, wlt
1
.. J.mu =u T(u 1 ) £1 ,seeref.[2],theorem8.land
i i
see Kupradze's (et alia) result on the regularity of elastic
potentials, ref.[3], theorem 5.2, p. 313 and theorem 6.2, p.315.
t The Galerkin method is identical with an orthogonal L2 -pro-

jection. Eq.(9) represents therefore both, the collocation me-
thod and the Galer kin method. For a definition of the two matri-
ces PGP, PHP see Eq.(27 a,b) (Collocation) and Eq.(28a,b)
(Galer kin)-.-
557
Furthermore, if u satisfies L u 0, then the value of IT(u)

at u is
il(u) = ~ JT(u) u ds J t u ds
1 - sf t u ds (17)
2
and the trace u, i.e. the value of u on the syrface S, mini-

mizes, as u minimizes_ ll(u) i~ 2 Eq.(7) on H (B), the func-
tional fi(u) in Eq.(l7) on H (S).
The Galerkin method on a set of functions
i 1,2 ••• N (18)
yields the same linear equations for u as the problem to mini-

mize rr (u)
li(Pu) = j f G-lH Pu ·Pu ds J Pt•Pu ds + Min (19)
namely
Ku = F t (20)
=-
with -1
Kij = (G H cjlj' cjli) (21)
The matrix K in Eq.(20) is a true stiffness matrix, i.e. it

has all the properties, commonly associated with such a matrix.
3. A modification
To steer into the direction of Eq.(ll) we modify the Galerkin

method. We do not solve Eq.(l8), but the problem
.._T t 0 (22) t
'>' '
i=l,2 ••• N
tWe delete in the following text the upper tilde, which we used
to designate the trace.
558
and the same holds for the functional IT where we solve instead
of Eq.(l9) the problem
1
IT(Pu) = 2 f PG -1HPu·Pu ds f Pt·Pu ds
s s
f ¢T t · ~T u ds + Min (23)
s
i.e. we restrict the procedures to the projections of G- 1H Pu
onto the space ~ spanned by the functions¢ ..
We use either a collocation projection (we kssume ¢i(xj) oij)
i
p = p with (u, o(x-x ))
c
T
or an orthogonal L2 - projection (i.e. (u ~u,¢.)=0)
- - l.
-1
u + with u = F w(u) and wi (u)=(u,¢i)
= -
-1
On the operator u + PG HPu is defined by
a matrix,
u ~T PG- 1HP u
+ (24)
-1 i
=- - i
with PG HP ij (o(x-x ), G-lH¢.) = T(u(¢j))(x ) if p p
J c
or
-1 N -1 -1
PG HP ..
l.J l.
k
Fik (G H\,¢j)
if P = PL2.
The introduction of the L2 -projection is camouflage, because

the thus "modified" problems, Eq.(22) and Eq.(23), have the same
solutions as the original problems, Eq.(l8) and Eq.(l9), but we
hoped to gain some consistency in the derivation by doing so.
Note, that the projection of a function that is contained in
the set ~ coincides with the function itself.
2 Keeping this_femark in mind, it follows zasily, that the
L -operator PG HP, (i.e. if P is the L -projection), has
all three properties (Pl), (P2) and (P3), if ~T contains the
traces of the six basic rigid body movements, wHile the collo-
cation operator has only the property (Pl,Ker).
The loss of (P2,Eq.) and (P3,Adj.) under the collocation pro-
jection is naturally due to the fait that the basis functions
¢. and the weight functions o(x-x ) differ, but with N
artd (consequently) ¢. + o(x-x 1 ), we might expect to regain these
properties in the limit.
559
With this modification, i.e. replacing G-~P

are free to couple the two methods with either a
collocation projection.
Method Projection
Galer kin
II - Min Collocation
w~fch yields essentially only thrze different proced~res to solve

G H u = t because (Galerkin x L ) and ( II- Min x L ) coincide,
see below. The combination (Galerkin x Collocation) does not lead
to a symmetric stiffnes2 matrix and so we are left with the com-
binations (Galerkin x L ) and (II- Min x Collocation).
The Galerkin vector u, the solution of Eq. (22), satisfies
the equation
F t (25)
while the II - Min vector u, the solution of Eq.(23), satisfies

the equation
(26)
(it is ,E = l>
Both vect2rs coincide, if PG-1 HP = M is symmetric, i.e. if p
is the L -projection and it easily follows, then that the
stiffness matrix 1 = F M has the properties
* T T
K s = P2 * \j
Pl :
*
=-
T
0
fT
s
i
f = s
- Ku = 0
=-
u, s
P3 : ~1 ~2 -2 ~1
if Ku
=-
= fi i = 1, 2
-1
because PG HP has the properties (Pl), (P2) and (P3) or, what
is equivalent, the matrix ~ = PG-lHP satisfies the Eqs.(33) -
(35), see below. Analogously the stiffness matrix
1 = t <r tl + tlT.V
-1
with ~+ ~T has the properties PI*, P2* and P3* if PG HP has
the properties (Pl) and (P2).
4. A substitution
But now the operator G-lH is unknown and we therefore use

the operators G and H from Eq.(8), H u = G t, and their
modifications with respect to P
PHP u + <t>TPHP u
..... = ...
<t>TPGP t
PGP t +
-=-
560
where PGP and PHP are the matrices

i i
PHP ij (o(x-x ), H<jl.) PGP ij (o(x-x ), G<jl.) (27a,b)
J J
N
p PHP .
iJ
PGP i. =
J
I
k
(28a, b)
to substitute for the matrix PG-~P the matrix PGP-lPHP
M (29)
This substituUon rests on the following argument: consider t

fixed and let u. be the function that satisfies PHPu = PGPt or
eq~fvalent BHP ~ = PGP ~ and SU£~ose the same u satisfies
PG HPu = Pt, or equivalently PG HP u = t, then follows Eq.(29).
But this assumption with respect to ~ hold~ 1naturally only for
the original equations, i.e. Hu = Gt ==9 G Hu = t. Note that
(Gt,t) = J E(u)·C·E(u) dv (30) t

R3
where
u = Gt J U(y,x)t(y)dsy
s
x e R
3
(31)
and therefore (Gt,t) = 0 yields u = a+ b " x and due to
0 - 0 = lim
i
Te(u) - lim Te(u) =
e
!2 t +
S'
r T*tds y - (- !2 t +
s
fT*tds )
y
t
~
also t = 0, which means, that the matrix ! PGP = [(G<jl.,<jli)] is

non-singular, which should also hold, if N is large eJough, for
the collocation matrix PGP = [(G<jl.)(x~)] due to Gt=O ~ t=O.
= J "'
5. Some corrections
-1
!~ practical computations we do not approximate PG HP with
PGP PHP, but,due,~o numerical quadratu~;rather with the product
P'G'P'-lP'H'P' =
= =
~'
=
(32)
where a prime denotes the modified operator due to numerical

qu~~ratm.e. The_~roperties (Pl), (P2) and (P3) read, i f we replace
~G HP by PGP PHP (we delete the primes)
-1
(Pl,Ker) PGP PHP s = 0 s (33)
= =-
(P2 ,Eq.) 0 u, s (34)
t See the appendix for a proof of Eq.(30).

~ The index e on Te( ) means, that the normal points into the
exterior.
561
or equivalently
(34a)
(P3 ,Adj.) (35)
where s is the vector of nodal values of a rigid body movement

a+ b Ax with respect to the functions ¢ ..
The collocation matrix ~· has none of fhe properties (Pl),
(P2) and (P3). While the property (Pl,Ker) gets lost under nu-
merical integration, the lack of properties (P2,Eq.) and (P3,Adj.)
is due to the collocatio~ 1projection ~Iself and has nothing to
do with slfbstitutin~ PGP PHP for PG HP.
The L -matrix ~· also looses (Pl,Ker) under numerical quad-
rature, while the lack of (P2,Eq.) and (P3,Adj.) is due to the sub-
stitution, but we might hope to miss these three properties only
by a narrow margin. * * *
The properties Pl , P2 and P3 are considered essential
for a stiffness matrix and depend, as we have seen, on the pro-
perties Pl, P2 and P3 of M.
Introducing appro~imati~ns R' for ~ = PG- 1HP we have,
therefore, to ensur~ that the properties P~and 23 are me~
by the substitute ~· . Consequently we replace the L -matrix ~·
by the symmetrix matrix -
(36)
and modify M" to guarantee (P2,Eq.) and (Pl,Ker).

Analogously we correct the collocation matrix M' to ensure
(Pl,Ker) and furthermore mo~fy it, to obtain property (P2,Eq.),
which is originally no property of ~. but will be needed to ob-
tain a correct stiffness matrix. Naturally there is no need to
symmetrize the collocation matrix, if we couple it with the
II - Min method. 2
With these corrections we obtain from (Galerkin x L ) and
(II -Minx Collocation) correct stiffness matrices.
6. Stiffness matrices for exterior unbounded regions

-1 -1
Let us denote for clarity by G H = G H the operator of
interior regions which we studied up fo now.
The approximation and derivation of sti~fness matrices for
exterior regions, i.e. for the operator G ~E proceeds as be-
fore and we theE!fore restrict the discussion to an artalysis of
the operator G HE.
We study the b.v.p.
~
L u = 0 limTi(u)=t (37)
e
and call u 1 solution, if u satisfies the equations above and
belongs to H (B).
We know, that e if "'t is smo2th enou¥h~ then £here exists a
unique regular solution u E C (B )n C (B ) n H (B ) ,
e e e
562
which satisfies the Somigliana identity
Ce(x)u(x) + J T.(y,x)u(y)ds
s ]. y
J U(y,x)t(y)ds (38)
s y
or shortly
In interior regions we encountered the operator H = HI' namely
J Te (y,x)u(y)dsY
S
(40)
and it follows with
ci (x)s(x) + J Te (y,x)s(y)ds y
S
0 v s = a + b 1\ x (41)
and
I - C (x) (42)
e
that
0 s (43)
or equivalently
s (44)
The operator G-l is defined by G-lG t t and due to
Gt = J U(y,x)t(y)ds y
s
(45)
it follows, with the well known jump of the traction of the

potentials of the first kind, that
G-l( ) = lim T ( ) - lim 'e ( ) (46)

i e e
Note that this definition makes sense only for those u E H112 (s),
whose continuation into B. and B are known. The rigid body
movement belongs certainly 1 to this ~lass, and so we obtain
G-lH s
E
= G-ls = lim
.
T (s) - lim T (s)
e e
=0 - 0 0 (47)
-1 l. e
i.e. the operator G HE has the property (Pl,Ker).
Exterior regular solutions satisfy furthermore the Betti
equation in unbounded reg~~ns and with theorem 8.2 in ref.[2] it
follows as above, that G HE_ 1has the property (P3,Adj.). Only
(P2,Eq.) is no property of G HE • Exterior solutions must not
satisfy the equilibrium conditionson the surface S. (This is no
longer true for regular exterior solutions in the plane. The
equilibrium condition is a necessary condition in the plane. Other-
wise the solution would not satisfy a radiation condition.)
*Consequ*ntly have exterior stiffness matrices the properties
Pl and P3 , (in 3-D), which should make the approximation of these
matrices perhaps a little bit easier.
563
7. Practical computations
In this section we have finally a short look at the numzrical

effort needed to derive a stiffness matrix with either a L - or
a collocation projection from the equation Hu = Gt.
Let us suppose the surface S to consist of M elements S ,
m.= 1, 2 ..• M and we assume furthermore, that there are N no~es
x 1 on S and that integration is done on a reference element
5 0 with a numerical integration technique with, (say), K weights
wk:
L J fm(~)Jm(~)d~
M M K
J f(x)ds = { L fm(~k)~(~k)wk = { f(~~)J~w~
m o ~ m km f
where ~ = (m-1)! + k, J = J (~ ), w = wk if~= (m-l)K + k.
~
~The points ~ , ~ = 1,2 •.. L are the lntegration points on S,
w are the corresponding w~ights and J is the corresponding
value of the Jacobian in x . The nodal points usually do not
coincide with an integration point. ~ext we introduce the diago-
nal rna trix ~(L L) with W. . = 8 i. JJ ~ and introduce N nodal
shape functionsx ~., whicEJgenerate the matrix !(N L) with
Vij ~i(xJ) and witE the matrices £(NxL) and ~(NxL)x
i -1 . i i . i
H.. = C(x )oijwij
1J
+ T(z.J ,x ) Gij= U(x ,yj) U(xJ,y)
and c
=(LxL)
and H
=(LxL)
H.. =
1J
i -1
c(:J:£ )oij wij + T(~j •.~h 'G1J
.. = U(~
i
,z.j) U(~ ,i) t
we obtain the collocation matrices
PHP = H W VT
and the L~ matrices

PHP = F- 1v w
= =====
Hw vT
(for completeness we state: F = V W VT = D VT)
which yields with £(NxL)= !E for=the collocation matrix
M' = (£~1T)-l ~,!!,Y.T =
and for the L2 -matrix M' :

M' = (,Y_,!;!, £,!;!, ,Y_T) -11!! li!! ,Y.T = (£ £ ~T) -1~ E~T
where D GDT is a symmetric matrix, while G DT is not.
t H and G are not defined on the diagonal, but this is irrele-

vant here and no major obstacle to the notation.
564
So, while the L2 -method (or Galerkin method) calls for a higher
resolution of G ~ [(L L and H ~ !(LxL) and involves two
more matrix multiplicat!o6s
-1
(Q£ .. ) Q~
the matrix to invert, Q~QT, is symmetric and as approximation

of K PGP = [(G~ .• ~i)]-probably positive definite.~
Furthermore w~ m~ght expect that the defect of ~· with re-
s~ect to (Pl,Ker) and (P2,Eq.) becomes 2 smaller, if we use the
L -projection, which should give the L -method a fair chance in
comparison with the collocation projection.
Having obtained a symmetric stiffness matrix with the means
proposed above, we now turn to the modifications we must bring
about to ensure the properties (Pl,Ker) and (P2,Eq.).
The property (P2,Eq.) seems to be of considerable influence
on the exactness of a solution. This we conclude from similar
experience with the Laplace operator ~ , where the analogue
property is (au/ an, 1) = 0 and remarks in ref.[S), which con-
firm this view.
The necessary modifications would be (relatively) simple if we
neglect possible unbalanced couples and satisfy only the equili-
brium condition with respect to translations. The following three
simple vectors are a basis of the set of all possible transla-
tions:
s(l)= [1' o, o, 1, O]T
s(2)= [0, 1, 0, o, 1, O]T
s(3)= [0, o, 1, o, o, 1, l]T
Let us denote by E: (k) the residuum of the translation s (k)
s(k) T rE'= E:(k) T (48)
The residua E:(k) would be zero vectors, if F M' was a

correct matrix, see Eq.(34a). The modified matrix !M has then
the components
3
(k)
£k)
j
si i, jI 1, 2 •.• 3N (49)
k=l
The property (Pl,Ker) should give us the least trouble. There

is a technique to calculate the Cauchy principal value of the
kernel T(y,x) with a rigid body movement. If we do this, then
the matrix fl' automatically has the property (Pl,Ker). If we
do not use this technique, the property (Pl,Ker) might not be ob-
tained exactly, but the defect should be very small. Otherwise
our numerical quadratu~needs mending.
565
8. Conclusion
Summing it all up, we state:

To derive from the equation
a stiffness matrix with the properties Pl * P2 * and P3 * pro-

ceed as follows:
If P is the 1 2-projection, calculate M', see Eq.(29), sym-
metrize it' see Eq. (36)' and modify the new matrix M" to en-
sure property (Pl,Ker), calculate the stiffness matrix .K = K M_•:
according to Eq.(24), and modify it to guarantee property (PZ,Eq.).
If P is the collocation projection calculate see Eq.(29), M',
modify it to ensure (Pl,Ker} and modify I~-' to satl-sfy _E~. (34)
and then calculate the stiffness matrix K = 1/2 (F M' + M' F).
These ma!1ices are the stiffness matri~es, if w~~olve=th~
equation G Hu = t with Galerkin's method or by minimizing a
potential in co~~ection with a_~rojection method and if we
approximate PG HP with PGP PHP.
APPENDIX
Proof of Eq.(30)
The function u(x)= J U(y,x)t(y)dsy satisfies Lu = 0 in a .• y

3
point X E R = B. u B and has a finite strain energy in the
1 e
continuum, i.e.
f
R3
E(u)·C·E(u) dv < 00 (a)
Due to Lu = 0 it follows, see ref. [1], p.95,

-i
(t ' u)
B.
J E(u)·C·E(u) dv (b)
-e
(t ' u) = f E(u)·C·E(u) dv
1
(c)
B
e
and by adding Eq.(b) to Eq.(c) we obtain
(~i + ~e
'
u) = f E(u)· C· E(u) dv (d)
R3
where -i
z-t + J T*t ds y
1
t lim Te(u) (e)
-e
t
i
lim T. (u)
1
(- 21 t + f T*t ds y ) (f)
e
Substituting for ~i and ~e the expression in Eqs.(~ and (f)

we obtain with u = Gt finally the result:
(Gt,t) = f E(u)•C•E(u) dv (g)
R3
566
References
[1] Flugge, S.(Ed.), Encyclopedia of Physics, Volume VIa/2,

Mechanics of Solids II, (Vol. Ed. Truesdell), Springer
Verlag 1972
[2] Hartmann, F. Elastische Potentiale in Gebieten mit Ecken.

Dissertation, University of Dortmund, 1980.
[3] Kupradze, Gegelia, Bashelishivili, Burchuladze. Three

dimensional problems of the mathematical theory of elas-
ticity and thermoelasticity. North-Holland Publishing
Company, Amsterdam, New York, Oxford, 1979.
[4] Brebbia, C.A. and Georgiou, P. Combination of boundary

and finite elements in elastostatics, Appl. Math. Modelling
1979, Vol.3
[5] Zienkiewicz, O.C., D.W. Kelly, and P.Bettess. Marriage

a la mode - The Best of Both Worlds (Finite Elements and
Boundary Integrals), in Energy methods in F.E. analysis,
ed. by Glowinski, Rodin and Zienkiewicz, John Wiley & Son,
1979.
[6] Mattioli, F. Numerical instabilities of the integral

approach to the interior boundary value problem for the
two-dimensional Helmholtz equation. Int. Journal for
Numerical Methods in Eng. Vol. 15, 1303 - 1313 (1980).
[7] Weyl, H., Symmetry. Princeton, Princeton University Press

1952.
ACKNOWLEDGEMENT
This work was completed while the author was visiting the
School of Engineering at the University of California,
Irvine. The financial support of the NATO Science Committee
(Grant 490/420/585/0) is gratefully acknowledged.
567
THREE DIMENSIONAL SUPER-ELEMENT BY THE BOUNDARY INTEGRAL

EQUATION METHOD FOR ELASTOSTATICS
F. Volait
CETIM - France
ABSTRACT
In this paper, we describe the formation of a stiffness

matrix using a boundary integral method for three dimensio-
nal elastostatics problems.
This technique is based upon energy minimization and leads

directly to a symmetric form. A special discretization is
adopted in order to avoid problems due to eventual discon-
tinuities of the traction vector (geometrical boundary dis-
continuities). Such a matrix enables a coupling with a
finite element discretization or an incorporation within a
standard finite element software. Numerical aspects induced
by the integration are discussed.
INTRODUCTION
For an homogeneous isotropic medium V of boundary S, a
linear elasticity problem can be formulated by using
Navier's equation :
/'o ;: u. (y) + f. (y)

~ ~
=0 for y E v (I)
and classical boundary conditions

u. (y) u. (y) for y E s (2)
~ ~ u
t. (y) t. (y) for y E st (3)
~ ~
where t. (y) a .. (y) n. (y) s = s us

~ ~J J ' u t
n. (y) is the unit outward normal vector at point y
J
568
CLASSICAL BOUNDARY ELEMENT METHOD
Numerical solution of this problem em be computed by using

classical boundary element method (see ref[!], [2/)
This solution derives from an equivalent formulation of
the first problem :
JV Uij (x - y) ll'l:tuj (y) + fj (y)] dV =0 (4)
for x E v
with boundary conditions u. (y) = u. (y) on S
l l u
t. (y)
l
In equation (4), U.. (x- y) is the elementary solution of

the Navier's equat~Jn which is nothing but the Kelvin's
tensor.
In most of numerical formulations, equation (4) is

transformed by using Betti's theorem and becomes
c .. (x) u. (x) +
lJ J
fs
):
T ••
lJ
(x - y) u. (y)
J
dS
(5)
/.s u .. (x - y) t. (y) dS +
lJ J
fv uij (x - y) fj(y)dV
which, in the case of zero body forces, is reduced to a

boundary integral equation. Then, boundary conditions are
taken into account by r~lacing the functions u. (resp.t.)
-- l l
by their known values ui (resp. ti)
Finally, we obtain an equation where unknowns are of two

kinds, displacements and tractions, which can be resolved
by analytical or numerical methods (conocation method for
example).
A SYMMETRICAL BOUNDARY INTEGRAL EQUATION
In the previous formulation, we obtained a relation bet-

ween u. and t. (Eq 5) in which we imposed boundary con-
ditions7 1
The initial problem can be now formulated in such a way that

displacements are principal unknowns. This approach, des-
cribed by Zienkiewicz et al. in Lo/, [7] is based on the
minimization of the energy functional.
IT (u)
2
I
jv a ij
e: •• dV -
lJ fv u.f.dV
l l
- Js u.t.ds l l
(6)
t
with u. u. (7)
l l
569
If we now impose as an extra condition that

~X u. + f. 0 (8)
l. l.
1s
the expression of the functional (6) becomes :
IT (u) = .!.
2
1s u. t. dS -
1 1
u. t.dS + -2I
1 1
{ u.f. dV
lv 1 1
(9)
which is to be minimized with both conditions (7) and (8).

The solution of the initial problem is then achieved by
calculating a stationary value of n , which is given by :
oIT (u) =0 (IO)
Calculating the expression of li IT we obtain

li rr (u) = ~ fs [ ui li ti + ti li ui] dV
Jv
(I I)
fs t o u.
l.
t.
l.
ds + -2I o u. f. dV
l. l.
0
with u.l. = u.l. on S

u
and with the equilibrium condition ~*u. + f.
l. l.
=0
which is equivalent to the integral equation (5). It is to
be noticed that this integral equation gives a relation
between tractions and displacements which will be used
to eliminate tractions from the expression (II)
NUMERICAL TREATMENT
A classical discretization procedure can be applied to

this problem and, assuming that f. = 0 in V, this discre-
tization is reduced to the boundarj {7}. Coordinates, dis-
placements and tractions are approximated by a set of
polynomial functions. A collocation procedure is used to
obtain a linear system from integral equation (5)
{ u} = (B] {t} (12)
In most situations, this technique gives a well conditio

ned solvable system of equations and tr.actions may be ex-
pressed as a linear function of displacements.
However, an important feature, which has been related by
M. Chaudonneret Ill, and by the author in previous papers
[4!, ~. is the possible discontinuity of the tractions
which cannot be taken into account by a classical discre-
tization.
In order to avoid introducing numerical errors, the discre-
tization procedure is modified. Nodal values of the tracdons
570
are now chosen inside the element as shown in Figure I.

In this way, discontinuities of the traction vector occur-
ring on the outline of the elements may be precisely ap-
proximated.
?
tn
Traction nodal
• '
~ . Geometric nodes and
displacement nodal values
'
values and '-----·- -?>
I;;
collocation
positions •
• •
Figure I - Discretization scheme

However, such a discretization increases the number of
tension nodal values and consequently the number of columns
of the matrix B. This leads to a non square system and
therefore a non solvable one.
To overcome this problem, the collocation procedure is
modified so that the collocation positions coincide with
the traction nodal value positions.
It is to be noticed that the displacements interpolations
can be kept continuous.
Finally, the integral equation (5) leads to the relation :
{t} { u} (I 3)
By using the same discretization in equation (11) a matrix

form is obtained :
(14)
2I < ou > [c] {t} +_!_

2 <t(o u) >[ct } ~u >
{u- {F}= o
with [ C]
and {F}
in these expressions N (resp. M) denotes the interpo-

lant functions of u (resp. t).
Applying the relation (I3) to the displacements {u} and
[o u} , the equation (14) can now be expressed in terms
of displacements only so that
1
2< ou >[(Ej + [E] T 1 {u} < oU> {F} 0
(15)
where [E] =[c][Br 1

571
Numerical solutions of the initial problem are then achie-

ved by solving the approximate problem
u} - { F } = 0 (16)
with u.]. u.]. on S

u
As has been pointed out in /6/ , this method leads to a

symmetric linear system (16) which can be assembled to a
system of finite element equations as the contribution
from a nev1 element. The compatibility is ensured if the
boundary interpolant for displacement is chosen to be
identical to the adjacent finite element interpolation on
the interface boundary.
NUMERICAL INTEGRATION
One of the most important problemsto solve when computing

the coefficients of the matrices[A] and [B] is certainly
the numerical integration accuracy. This problem has been
solved by Lachat and Watson /7/ in the three dimensional
elastostatics program E.I.T.D. %which was based on a
classical boundary integral equation formulation.
However, due to the new discretization, we had to modify

it.
Coefficients to be computed are of the kind
I (x) = j J
+I +I
K (x,ye (t; ,n) Nc (t;,n) Je (t;, n)dt; dn
-1 -1
(17)
where K(x,y) is a singular kernel depending on the dis-

tance r (x,y) ; ye ( t;, n ) represents the global coor-
dinates of a point y on the element e and Je ( t; , n )
is the Jacobian of the geometrical transformation.
In most cases, an approximate value can be achieved by a
Gauss numerical integrating with a ( 3 x 3 ) points sche-
me. However, when the distance becomes small compared with
the size of the element, this formula has to be modified
by dividing the element and by increasing the number of
integration points.
% E.I.T.D. CETIM'S Program running on a C.D.C. 7600

572
When the collocation point belongs to the element over which

the integration has to be performed, the integral becomes
improper and has to be evaluated with particular care.
For K (x,y) U.. (x,y) , the kernel contains a singula-
1J
rity of the order 1 and good accuracy is obtained by
dividing the element as shown in figure 2. Here, the
number of integration points is chosen according to the
value of e~
Collocation position
Figure 2 - Integration scheme on singular element
This integration scheme was tested on the function 1 which

is simpler than U.. (x, y) but singular. Result~ were
compared with anal~tical calculations and the error was
only of 0.03 % .
ForK (x,y) = T.. (x,y) , the kernel varies as 1 and the

1J r-2
integral has to be interpreted in the sense of the Cauchy
principal value. Numerical evaluation of this integral
is very difficult.
To overcome this problem, we tried to use the solution des-

cribed by Lachat and Watson in [7].
This consists of decreasing the order of the singularity
by using displacement interpolant vanishin@ when x = y.
The product of both functions T .. (x,y ( ~. n ) and
1J
Nc ( ~. n ) is reduced to a l singularity (Extended cal-
culation is related in (5]), rfor the bidimensional case.
However, due to the collocation position (inside the ele-
ment), we had to modify the discretization of the displace-
ments and use the same interpolation functions as the
ones used for the traction vectors. In this way, colloca-
tion positions coincide with displacements nodal value and
d1e integrating scheme previous!* described way be used.
After computing th~ matrices A and B (A corresponding
to displacements u - fig. 3) and in order to keep the
compatibility with a possible finite element zone, we have
to return to the previous interpolation. This is simply
done by expressing the inside nodal values in terms
573
of outline nodal values (see fig. 3)

= [ M] {u} '
~Outline nodal values u

• • •
• •-..:
~Inside nodal values u
::t
./
• •
Figure 3 - Position of the displacement nodal values
CONCLUSION
In this paper, we have emphasized the numerical treat-

ment of the symmetrical boundary element method applied to
three dimensional elasticity problems.
At present, a computer program is being developed at CETIM

and numerical results should be available in a short time.
However, the results obtained with the two dimensional
program [5! showed a good accuracy. But, due to the compu-
tation of the inverse matrix,computing time was considera-
bly increased with respect to a classical formulation.
ACKNOWLEDGEMENTS
This work was carried out with the support of the"Direction

des Recherches et Etudes Techniques" • Thanks are due to
Prof. O.C. Zienkiewicz, Drs. A. Chaudouet, D.W. Kelly,
G.G.W. Mustoe andG. Loubignac for their help and encoura-
gement.
574
BIBLIOGRAPHY
/1/ Brebbia, C.A., Walker, S., Boundary Element Technique

in engineering , Newnes- Butterworths,l980.
/2/ Lachat, J.C., A further development on the Boundary
Integral technique for elastostatics, Ph. D. Thesis,
University of Southampton, 1975.
/3/ Chaudonneret, M. , On the discontinuity of the stress
vector in the boundary integral equation method for
elastic analysis. Int. Symp. on recent advances in
boundary element methods, Southampton. 1978.
/4/ Mustoe, G.G.W., Volait,F.,A symmetric direct boundary
integral equation method for two dimensional elastosta-
tics. Int. Symp. on recent advances in boundary element
methods, Southampton. 1980.
/5/ Volait, F., Calcul d'un super element par une methode
d'equation integrale. These de 3eme cycle - Universite
de Technologie de Compiegne. 1980.
/6/ Zienkiewicz, O.C., Kelly, D.W. and Bettess, P.
Marriage a la mode - The best of both worlds (finite
elements and boundary integra~ - Int. Symp. on inno-
vative Numerical analysis in applied engineering Scien-
ce, Versailles France. (1977)
/7/ Lachat, J.C. and Watson, J.O. Effective numerical treat-
ment of boundary integral equations : A formulation
for three - dimensional elastostatics.
Int. J. for num. meth. in eng., Vol 10, 1976.
575
THE COUPLING OF BOUNDARY AND FINITE ELEMENT METHODS FOR

INFINITE DOMAIN PROBLEMS IN ELASTO- PLASTICITY
G. Beer and J.L. Meek
Dept. of Civil Engineering, University of Queensland, Australia
ABSTRACT
The implementation of a coupled analysis capability into an

existing Finite Element computer program is discussed. The
coupled analysis is then applied to a circular excavation in
a infinite domain where the region of plasticity is confined to
the Finite Element mesh. Further potential usage of the
coupled analysis is then discussed in relation to mine design.
INTRODUCTION
The coupling of Boundary Element and Finite Element methods

was first discussed in a general context by Zienkiewicz et al.,
(1977) although some more specific applications appeared
earlier (Chen 1974). The method was applied to a number of
field problems and problems in fluid mechanics (Kelly 1979).
Application to problems in elasto-statics have appeared more
recently (Brebbia 1979 and Mustoe 1980). None of these,
however, deal with the 'exterior' problem i.e. one involving
an infinite domain. The principle approach of the above
methods is to obtain a stiffness matrix of the region which is
bounded by Boundary Elements. The stiffness matrix is then
made symmetric by an energy approach or by minimising the
errors in the non-symmetric terms. This method will be called
the Symmetric Dire~t Boundary Element method.
At the same time as the above developments took place, a

completely different approach was presented by Ungless for
elasto-static problems and by Bettess for field problems and
problems in fluid mechanics. This approach used special Finite
Elements which extended to infinity in one or more directions.
The functions or displacements were assumed to decay in the
infinite direction exponentially or inversely proportional to
the distance (l) from a decay origin. The authors have later
R
576
1
shown that a decay of R gives good results for underground
excavation problems. In contrast with the Boundary Elements,
these Elemen~still require a volume integration which is
carried out numerically in a finite mapped region of the
infinite domain.
These two approaches will be examined herein in more

detail. Both methods have been programmed, incorporated in an
existing Finite Element program, and run on a PDP-11 mini
computer at the University of Queensland.
SYMMETRIC DIRECT BOUNDARY ELEMENT METHOD
Boundary integral equations

The derivation of Boundary Integral equations for elasto-
statics is well known (Rizzo 1967) and will not be repeated
here. This paper follows the notation by Watson. For the
exterior problem, the following integral equation is obtained
(see Fig. 1) :
c .. (x)u.(x)
~J J
+
s
I T .. (x,y)u.(y) dSy
~J J
=I
s
u .. (x,y)t.(y)
~J J
dSy (1)
where
lim I Ti.(x,y) dSy (2)

e:-+0 J
S(x,e:)
INFINITE DOMAIN
t.
x2
S(X,E)
L x1
Figure 1. Opening in an infinite domain

577
In the above equation, x is a point on the surface and y

the integration variable; u. and t. are the displacements and
tractions at the boundary. 1 The kefnels U.. (x,y) and Ti.(x,y)
for the case of plane strain are: 1J J
1 (xi- yi) (xj - y .)

.. 1n(-) + J
ui/x,y) c1 {c 2 o1J r r2 }
(3)
and
~~
(x. - y .)
c • J J
Tij (x,y) 3 c4 Ei (y) nj (y)
r
(xi-yi)J
r
+ [ c 4 oij + 2
(xi - Yi) (xj - yjj
r2
. n s (y)
(x
s
- y )
s 1 (4)
r {
where
( l+v)
c1 4TIE (1-\!)
c2 3 - 4\!
1
c3 4TI( 1-\!)
c4 1 - 2\!
In the above, x.,y. are the coordinates of points x andy at

the surface, r 1 is 1 the distance between the points and n.(y)
is the vector normal to the surface at point y pointing iffward;
o .. is the Kroneker Delta.
1J
Discretisation of boundary integral equations
Here a point collocation procedure is used to discretize the
Boundary Integral Equation described previously. This means
that equation 1 is written for n points along the Boundary.
These points are the nodal points of Boundary Elements. For
each Boundary Element, the geometry and functions are described
by
(5)
and
(6)
578
t 0 (!;) ( 7)
]_
k
In the above, N (!;) are suitable shape functions of the local
(element) coordinate !;, and the superscript refers to the kth
node of an Element.
Two forms of the shape functions will be examined later,

namely parabolic and linear. The parabolic function involves
3 element nodes and the serendipity shape function can be
written as:
N1 (I;) 1;(1;+1) t (8)
N2(!;) 1 - 1;2 (9)

1
N3 (0 - !;(!; - 1) (10)
2
The linear variation involves only 2 nodes and the correspond-
ing shape functions are:
1 (/; + 1) ( 11)
2
1 (!; - 1) (12)
2
As with Finite Elements, a global numbering system is

used for the nodes. Let d(b,c) be the global number of node
c of Element b. The discretized form of equation 1 becomes:
p n(b)
c .. (xa) u. (xa) + I I u. (xd(b,c)) T .. (xa,y(i;J)
l-J J b=l c=l J
J l-J
sb
p n(b)
a
Nc(O J(!;) d!; I I t. (xd(b ,c))
J J
uij(x ,y(i;))
b=l c=1
sb
(13)
In Equation 13, the summation is over all elements p and

nodal points n(b) of each Element and J(i;) is the Jacobian of
Equation 5.
Numerical Integration of Boundary Elements

Due to the complexity of the integrands, the integration has to
be carried out numerically. The following scheme is used:
579
For d(b,c) # a, i.e. where the point a does not lie

within an Element, a 4 point Gauss-Legendre integration for a
parabolic shape function and a 2-pointscheme for a linear
shape function are used.
When d(b,c) =a, the product U.. (xa,y(~))• Nc(~) tends to

infinity and requires special attenEfon. For i = j the
integrand becomes:
J c1 <t
sb
rc 1
21n<-;> +
(x. - y .)
1.
r2
1.
2
} Nc(O J(O d~ (14)
The integral
C ln (.!.)Nc(~) J(l;) di; (15)

r
I
sb
will be examined in more detail. Substituting r = fwhere R
is the length of Boundary Element b Equation 15 can be
rewritten as:
1
f
sb
C Nc(!;) ln (f~) J(!;) ds
1
J c Nc(t;(r)) 1n (.!.)
r
at;
()f
J(!;) dr + fc Nc (!;) ln ta) JCO di; (16)
-1
The first part of the integral above can now be evaluated

numerically using the integration formulae given by Stroud
which integrate
1
r 1 - -
OJ ln(~) f(r) dr
whereas, for the second part simple Gauss-Legendre is

sufficient. Watson has shown that the integrals T.. (xa,y(!;))
Nc(!;) and the free term c .. (xa) need not be evaluaEJd at
points d(b,c) = a. The l.Jvalue can be obtained by summing
all off-diagonal terms and changing sign. For the exterior
problem, a value of o.. has to be subtracted before changing
sign. l.J
580
Symmetric stiffness matrix of boundary element region

Using matrix notation equation 13 can be rewritten as:
[A] {a} = {B}[b] (17)
where matrices [A] and [B] contain the integrals of the kernel
functions and the vectors {a} and {b} list the displacements
and tractions at all nodes.
It should be mentioned here that at a node where 2 Elements

join the node has two different local numbers, for example,
d(1,3) = d(2,1). When constructing the matrices [A] and [B]
this causes the addition of some terms. Equation 17 can now be
solved to give the tractions as a function of the displacements:
~} [B]-l[A] {a} = [C] {a} (18)
To obtain the stiffness matrix of the Boundary Element

region, two approaches have been used which the authors
believe to be identical. The first approach is by a minimis-
ation of the potential Energy of the system and has been
suggested by Kelly and later used by Mustoe. The potential
energy of the exterior region is
df (19)
where {o} and {E} are the stress and strain vectors and {t} are
tractions applied at the boundary. The strain energy term may
be transformed into a boundary integral and the total potential
energy can be written as
TI = t J {u}T {t} df (20)
r
Upon substitution of equation 6 and 7 the discrete form of
equation 20 is:
1 {a}T [M] {b} - {a}T {F} (21)

2
Where the matrix [M] contains the integrals of shape function

products, for example:
581
r
J N 1 (~)N 2 (~) J(~) d~ etc. (22)
sl
The vector {F} contains integrals of the product of shape
functions and traction values, for example:
F1 = j N 1 (~)· t1 (~) J(~)d~ etc. (23)
sl
Upon substitution of equation 18 relationship 21 can be

rewritten as:
1
7T = 2 {a}T [M] [C] {a} (24)
Setting the variation to {a} zero gives
1
2
[ [M] [C] + ([M] [C]) T] {a} -{F} 0 (25)
where
[K] ; ( [M] [C] + ([Mj [Cj) T) (26)
is the stiffness matrix.
In the second approach, Brebbia uses virtual work principles

to obtain
[K] LM] [ C] (27)

us
which is not in general symmetric.The matrix is made symmetric

by the substitution
[K] 12 ( [K] + [K] T (28)

s us us
which is identical to the previous approach.

582
Coupling of boundary and finite element meshes

Once the stiffness matrix of the Boundary Element region has
been found, the analysis of coupled problems proceeds as in the
standard Finite Element process. All stiffness matrices are
assembled and the system of equations solved for applied forces.
Since the stiffness matrix of the Boundary Superelement is
symmetric, standard Finite Element software with symmetric
equation solvers can be used,
It is needless to say that to ensure compatability between

the Finite and Boundary Element meshes, the interpolation
functions for the displacements must be identical. Thus,
parabolic Boundary Elements should be combined with Finite
Elements which have a parabolic variation of the displacements
at the connecting boundary although Brebbia apparently obtained
good results by coupling constant Boundary Elements with
parabolic Finite Elements
Tests on accuracy - choice of shape functions

Tests were made on a circular opening in an infinite medium to
determine the accuracy obtainable with the symmetric direct
Boundary Element method for infinite domain problems. The
virgin stress field was assumed to be OH = -0.4 in horizontal
direction and a = -1.0 in vertical direction. A Poisson's
ratio of 0.2 waX used. The problem was solved by using
equation 17, that is, a conventional Boundary Element method
and using equation 25, that is, the symmetric Direct Boundary
Element method.
In the first series of runs parabolic Boundary Elements

and two different meshes were used. The displacements along
the boundary are plotted in Figure 2 together with the
theoretical values by Pender (continuous curve). It can be
seen that although the conventional Boundary Element method
gives excellent results with relatively few parabolic elements,
there is an appreciable loss in accuracy where a symmetric
stiffness matrix is used. The symmetrising causes minor
oscillations of the displacements. No reason could be found
except that the stiffness matrix [K] may be not well
conditioned. The tests were made on a PDP 11/34 mini computer
with a wordlength of 32 bits. Mustoe apparently has not
experienced any problems with interior regions using parabolic
Elements although he used a CDC 7600 for his analyses. This
computer has a wordlength of 60 bits.
583
1,6
0
0 BEM MESH 1
,. D
X SYM [K]
+ BEM MESH 2
o SYM [K]
•/o lnfl "h

LOCATION ANGLE
Figure 2. Plot of u 2 E/cr vs. ~ along the surface

of opening usiXg parabolic Boundary
Elements
When using Boundary Elements with a linear variation of

displacements and tractions, the problem disappears and
results by the conventional and symmetric Boundary Element
method are practically identical. This is shown in Figure
3 where 2 different meshes of linear Boundary Elements are
used. Obviously, a finer mesh is required to obtain results
similar to the ones obtained with parabolic Boundary Elements.
It is interesting to note here that Brebbia has also reported
no loss in accuracy between conventional Boundary Element
method and symmetric Boundary Element method for constant
Boundary Elements.
584
1.5
Ill
1-
z
w
::E
w
~ 0 BEM MEStj 1
it 1. X SYM [K]
c
1/)
+ BEM MESH 2
0 SYM [K]
"/a '"/a
LOCATION ANGLE
Figure 3. Plot of u 2 E/0 vs. S along the surface

using linear BZundary Elements
Coupled analysis of circular opening
After the previous tests, it was decided to use linear
Boundary Elements and couple them with the variable number of
node quadilateral Finite Elemens available in the existing
program. This Element allows the choice of either a linear
or parabolic variation of displacement along an element
boundary depending on the presence or omission of the midside
node.
The mesh used for the coupled analysis is shown in Figure

4 where also the displacements along the tunnel periphery are
shown. It can be seen th~t there is good agreement with the
theory. There is also good agreement with the theoretical
stresses.
585
1,5
VERTICAL
,_
z
w
:;;
......
w
u
Q,
"'0 0,5
0 Fl NITE & BOUNDARY ELEM.
HORIZONTAL
"/a l'Ya "A

LOCATION ANGLE
Figure 4. Results of Coupled analysis
INFINITE ELEMENTS
The theory of the Infinite Element has been presented elsewhere

(Beer 1980) and here only a brief summary will be given.
The infinite Element is a special Finite Element which is

of infinite extension in one direction. The geometry of the
element is described by
(29)
where "'k
N are singular shape functions of the local coordinate
~2 • The variation of the displacement within the Element is
(30)
where r is the distance from a decay origin in this case the

centre of the opening.
586
To obtain the stiffness matrix of the infinite element a

volume integration is still needed. The accuracy of the
solution now depends not only on the fineness of the mesh in
the finite direction but also on the remoteness of the Element
from the excavation surface because the variation of
displacements is based on a far field solution. Thus, these
elements can not in general be used on the free surface.
However, recent tests have shown (Beer 1981) that convergence
of the element is good and in most cases only 2 or 3 layers of
Finite Elements between the surface and the infinite Element
are needed to give good results. The main advantage of the
infinite Elements is that they are easy to program and that
they do not destroy the banded nature of the Finite Element
equations thus making the equation solving more efficient. The
accuracy and computational efficiency of the two methods is
compared in Table 1 for the circular opening.
TABLE 1
Max. Execution time

Displacement PDP 11/34
Theory 1.632
Finite Elements and
Boundary Elements 1.6007 6:05
Finite Elements and
Infinite Elements 1.6077 3:19
It can be seen that whereas the accuracy is practically the

same, the coupled approach requires significantly more
computing time.
APPLICATIONS OF COUPLED ANALYSIS
As mentioned in the Introduction, a coupled analysis could be

applied to infinite domain problems in elasto-plasticity.
Figure 5 shows such as application where a deep circular
opening in a visco plastic soil is analysed. The plasticity
parameters where such that the region of inelastic deformation
was confined to the Finite Element mesh. For this particular
example, a similar result can be obtained using infinite
Elements.
587
I
+
Figure 5. Region of plasticity for angle of friction
40° and cohesion of 0.5
Another applicatio n of the coupled analysis which occurs
in undergrou nd mining is the study of pillar behaviour . Figure
6 shows a typical cross-sec tion of a mine opening. The aim of
the study is to observe the non-linea r behaviour in the pillar.
This is essential ly a parameter study and may have to be
carried out a large number of times with different material
propertie s, failure criteria and geometry.
PILLAR
F.E. MESH
EXCAVATION
rMESH
Figure 6. Mine pillar problem

588
For such a study, it is convenient to discretize the

opening into Boundary Elements and the pillar into Finite
Elements. There only the stiffness matrix of the nodes
connecting to the Finite Element mesh is needed and the
symmetrisation process has to be modified slightly. The vector
{a} is partitioned into vector {a 1} which lists displacements
of the nodes not connected to the Finite Element mesh and a
vector {a 2 }. Similarly, the vector {b} is partitioned into
{b 1} and Lb 2 }. Equation 17 is then rewritten as:
Ja1} l ~B]1 [B]~

f{b 1}1
l
EA]1 [A]J (31)
{a2} J lb2}r
and rearranged:
[B]~
J {a 1 l }
E•Jl [All[-
r {b'}}
EA]1
l- {b2}
{a2}
(32)
Multiplying both sides with the inverse of the matrix on the

left hand side gives:
)
' {a!} J [E) '

r {b }
1 J
1 (33)
- {b } -{a }
2 l 2
Where the new matrix [E] is the inverse of the matrix on the
left hand side, multiplied by the matrix on the right hand
side in (32). The tractions on the Boundary Finite Element
interface can be computed from (33) by partitioning:
(34)
In the above {b 1} are the (known) tractions at the Boundary

Element surface and {a 2 } the displacement of the Boundary-
Finite Element interface. Equation 34 can then be substituted
into (21):
7T (35)
589
where [M] now only involves the Boundary Elements connected to

the Finite Element mesh. Minimisation of the potential energy
yields:
The stiffness matrix of the interface nodes is:
K Z1 [[M][E]22
- (36)
and the nodal force vector becomes:
{F} (37)
CONCLUSIONS
This paper has concentrated on the implementation of a coupled

Finite Element Boundary Element program. The attention was
focused on the exterior problems of a deep underground
excavation although the computer code also allows the treatment
of interior domains. Details of implementation of a Symmetric
Direct Boundary Element method were discussed and test runs
made to assess the accuracy and computing time of a coupled
analysis. The implementation of the coupled analysis capability
into an existing general purpose Finite Element code only
required minor modifications to the program and the addition of
12 Subroutines and a total of 500 Fortran statements.
In the paper, it was also shown that the coupled analysis

can be applied successfully to the excavation of a circular
tunnel with localised plasticity. It was also shown that a
completely different approach using infinite Elements gives
similar answers with greater economy. Infinite Elements,
however, can not be used on the free surface internal boundary.
Thus, potential application of the coupled Finite-Boundary
Element analysis is for parameter studies on rock pillars in
the mining industry. In this application, the excavation
surface isdiscretized into Boundary Elements and the pillar
region into Finite Elements. The Boundary problem is solved
giving the stiffness matrix of the elastic rock mass only at
the interface nodes. Once this has been done, a large number
of non-linear analyses can be carried out (changing the material
models, pillar dimensions, properties etc.) on a relatively
small system. Work is currently underway in this direction.
590
ACKNOWLEDGEMENT
The work reported herein was carried out as part of the

research project 'Computer Based Excavation Design Techniques
for Multiple Orebodies in Jointed Rock' which is sponsored by
the Mining Research section of Mount !sa Mines Ltd., Australia.
REFERENCES
Anderson, D.L. and Ungless, R.L. (1979) Infinite Finite

Elements. Int. Symposium on Innovative Numerical Analysis in
Applied Engineering Science, Versailles, France.
Beer, G. (1981) On the Finite Element Analysis of Underground

Excavations. Submitted to Int. Jnl. Anal. and Num. Meth. in
Geomechan.
Beer, G. and Meek, L.L. (1980) A Boundary Finite Element for

underground mining applications. in New Developments in
Boundary Element Methods, CML Publications, Southampton, U.K.
pp. 281-294.
Bettess, P. (1977) Infinite Elements. Int. Jnl. Num. Meth.

Engng., Vol. 11: 53-64.
Brebbia, C.A. and Georgiou, P. (1979) Combination of Boundary

and Finite Elements in elastostatics. Appl. Math. Modelling,
Vol. 3.
Chen, H.S. and Mei, C.C. (1974) Oscillations and Wave Forces in
a Man-made Harbour. lOth Naval Hydro Symp., Dept. of Civil Eng.,
M.I.T., Cambridge, U.S.A.
Kelly, D.W., Mustoe, G.G. and Zienkiewicz, O.C. (1979) Coupling

Boundary Element methods with other numericalmethods. Ch. 10 in
Developments in Boundary Element Methods-!, Applied Science
Publishers Ltd., London.
~ustoe, G.G. and Volait, F. (1980) A Symmetric Direct Integral

Equation method for two-dimensional elastostatics. Paper
presented at 2nd Int. Seminar on Boundary Element methods,
Southampton.
Pender, M.T. (1980) Elastic Solutions for a deep circular

tunnel, Geotechnique XXX, 2: 216-222.
Rizzo, F.J. (1967) An integral equation approach to boundary

value problems of classical elasto-statics. Quart. Appl.
Math., 25: 83-95.
591
Stroud, A.J. and Secrest, D. (1966) Gaussian Quadrature

Formulas, Prectice-Hall.
Watson, T.O. (1979) Advanced Implementation of the Boundary

Element method for two and three dimensional elastostatics.
Ch. 3 in Developments in Boundary Element methods-1,
Banerjee ed., Applied Science Publ.
Zienkiewicz, O.C., Kelly, D.W. andBettess, P. (1977) Marriage

a la mode - The best of both worlds (Finite Elements and
Boundary Integrals). Int. Symp. on Innovative Num. Anal. in
Appl. Eng. Science, Versailles, France.
Zienkiewicz, O.C., Kelly, D.W. and Bettess, P. (1977) The

Coupling of Finite Element and Boundary Solution Procedures,
Int. J. Num. Meth. Engng., Vol. 11: 355-376.
592
A FINITE ELEMENT - BOUNDARY INTEGRAL SCHEME TO SIMULATE ROCK-

EFFECTS ON THE LINER OF AN UNDERGROUND INTERSECTION
B. A. Dendroul and s. A. Dendrou2
1Agbabian Asociates, El Segundo, California
2camp Dresser & McKee, Annandale, Virginia
INTRODUCTION
The collapsing mechanism of an underground facility in a blocky

jointed rock is generally associated with rock wedges in
loosened conditions along preexisting surfaces of discontinui-
ties. During the movement of the rock fallout the initial
stress field is released, resulting in a new state of equilib-
rium that affects the liner of the underground opening.
A three-dimensional finite element analysis is generally

required to simulate the excavation process and the installa-
tion of the liner. The analytical description of the rock
fallout occurring after construction of the liner, however,
requires either a great computational effort, if one uses
"joint" elements, or an over-simplification if only the gravity
load of the rock wedge is applied on the liner. A viable
alternative is an iterative procedure that progressively loads
the finite element discretization of the liner through the use
of a rock fallout and that takes into consideration the con-
finement of the rock medium that results from the excavation
process.
In this study the main effort is placed in the coupling of the

above finite element and boundary integral procedures, as
applied in the study of rock fallouts affecting an underground
intersection. The results of this scheme are compared with the
results of an oversimplified approach in which the gravity
loads of the rock wedge are applied directly on the liner.
This reveals that the proposed methodology offers both
accuracy and cost efficiency.
FORMULATION OF THE FINITE ELEMENT-BOUNDARY INTEGRAL SCHEME
The possibility of rock fallout following an excavation stage

is of primary concern for the designer of the liner of an
underground facility. The conventional method consists of
593
assuming that the rock medium around the opening is in a

loosening state of stress, which allows large movements of
rock blocks along preexisting discontinuous surfaces. These
rock blocks are free to transmit all or part of their weight
to the supporting liner and affect the redistribution of the
stresses in the continuous rock medium. Clearly the rock
block represents a collapsing mechanism and the liner should be
designed with respect to this ultimate load. This important
fact is also recognized by Cording (1976) who provided design
recommendations that were based on possible configurations of
rock fallout. At the present state of knowledge, location of
the initiation of rock failure is not known, and consequently
it is not possible to simulate the propagation of failure for
a known pattern of discontinuities. Therefore, a possible
approach is a postfailure analysis in which the failure
surface is assumed known and its effect on the conventional
analytical model is estimated. Many investigators, Ghaboussi
(1973), Goodmann (1968), and Zien~ewicz (1970) have imple-
mented this concept of slip surfaces in a finite element
computer code through the use of element representations.
These "joint elements" have met with moderate success because
of the required large computational effort resulting from the
increase of the degrees of freedom that describe the sliding
friction. The proposed technique is conceptually simple and is
based on the premises that the slip surface between the rock
medium and the rock fallout is a stress free area described
by a series of adjacent nodal points. The described solution
is obtained by prescribing nodal pressures at the slip nodes in
such a manner that the friction law is satisfied. In what
follows the governing equations for a finite-element model of
a liner and a boundary-integral formulation for the rock-
fallout-liner intersection are developed. These equations are
then combined to form the proposed iterative scheme.
The finite element equilibrium equation for the liner structure

is given, in matrix form as follows
u (1)
where K~ is the stiffness matrix, u is the structural

displa;~ment vector, ~ is the vector associated with the
gravity load of the liner and ~E is the force vector associated
with the rock fallout action on the liner. The stiffness
matrix and vector ~ are easily obtained from a Finite Element
Solution. The force vector FE on the other hand is the
resultant of two vectors
~E = ~RG - ~I (2)
594
where ~RG is the load vector for a completely rigid liner, and
~I is the correction vector that accounts for the rock/liner
interaction. This vector is defined as:
T (3)
~I =! ~R X~
where ~R is the stiffness of the rock medium and ! is a trans-

formation matrix that selects the structural degrees of freedom
used to define the rock/liner interface. Substituting equation
(3) in equation (1) we obtain the expression:
(~ + !
T
JSR :v ~ = !!,s + ~RG (4)
Matrix~~ and vector ~R are now determined through the

applicat1on of the boun~ary-integral technique as follows.
The fundamental boundary-integral equation is given by Brebbia

(1978). Neglecting body forces and assuming a virtual state
(denoted by superscript *) on the domain n of the rock medium
with a boundary s, and applying the Betti-Maxwell reciprocity
principle, the following relation is obtained:
(5)
s s
where u. and t. are respectively the displacement and traction
on bounaary s. 1
Furthermore, if we assume the virtual state to be the funda-

mental solution state provided by Kelvin's problem then the
virtual displacements are given by:
< 16 7f~(l-\!)r l(3-4\!) oij + ~~i . ~~j } ej (6)
where r is the distance from the point of application of a unit

load to the point of evaluation of the virtual displacement.
G is the shear modulus of elasticity and\! is poisson's ratio
of the rock medium, o .. is Kronecker's delta and e. the unit
vector in the directiBrt of the unit load. J
Similarly, the tractions at a point on the surface with a

normal direction n are given by:
t.*
1
1
2
~ar
an
~~ J
[cl-2\!l 0ij + 3 ax. ax.
B7r(l-\!)r 1 J
ar
+ (1-2\! l <n.
J ax.1
- ni k)
ax. } ej
(7)
J
595
Through the division or rock's external boundaries into a

series of boundary elements, equation (5) may be expressed in
matrix notation as:
H u
~- --
=G t
In the present implementation the displacement and traction
(8)
shape functions are assumed to be constant over each boundary

element. Once the matrices in equation (8) are computed, the
rock stiffness matrix ~ can be obtained from the following
expression:
-1 (9)
.!SaE.=Q. ,!!"U
Attention should be given to the fact that the derived stiff-
ness matrix is not symmetric and should be symmetrized in order
to use it in expression (4).
Only a brief presentation of the boundary integral formulation

is given here since an extensive coverage of the subject is
available in the literature. Rather, in what follows the
effort is placed in the description of the iterative procedure
adopted for the simulation of the rock fallout. If n 2 repre-
sents the rock medium and n 1 represents the rock fallout
(Figure 4), then our main concern is the description of the
sliding of nl with respect to n~ during the loosening process.
The phenomenon of sliding frict~on is not well understood at
the present time but the adopted friction law is representative
of many experimental observations, Jaeger (1971).
TWo are the important observations: (a) the frictional

resistance of two surfaces in contact is proportional to the
normal force across the friction surface, and (b) the "kinetic"
friction is smaller than the "static" friction. In this
application the former case is considered. Then the friction~!
resistance prior to slippage obeys the following relationship
(10)
where fT is the tangential force acting at a particular slip

node f~ is the normal force acting at a particular slip node
and k ~s the coefficient of friction.
In a typical computational procedure, the slip nodes are first

assumed to be attached to the adjacent material surface. The
inception of slippage occurs when the tangential force exceeds
the frictional resistance given in expression (10). This is
expressed as
(11)
596
where f~ is the force existing on the rock medium, Figure 4.

After sliding begins the adhesion of the frictional surface is
lost and at this point the tangential force acting at the slip
node is given by
(12)
where a equals ± 1, so as to ensure that the frictional force

acts always in a-direction opposite to that of slippage.
Application of this frictional force implies a new equilibrium
state in which adjacent slip nodes can move only in the tan-
gential direction as shown in Figure 4. The .new equilibrium
state is identified using the boundary integral technique and
the compatibility of the displacements of the slip nodes in the
normal direction is assured by a correction displacement
vector that has to be incorporated in the subsequent computa-
tional step. In a formal way, the above methodology is
presented by the following iterative steps:
{i) Define the boundaries of the rock fallout and compute

tangential and normal forces acting on the boundaries
based on the computed stress field of the three-
dimensional finite element analysis.
(ii) Identify boundary elements where frictional resistance

is exceeded and apply corresponding tangential forces.
(iii) Compute loading vector and displacements of the liner

according to expression (4).
(iv) Compute the displacements of the slip nodes resulting

from the tangential forces and the response of the
liner.
(v) Check the relative normal displacements of the slip

nodes. If compatibility is violated impose a correc-
tion loading vector and return to step (iii).
APPLICATION OF THE PROPOSED SCHEME-DISCUSSION AND RESULTS
The case of an underground intersection and in particular the

study of the rock-fallout effects on the liner at the inter-
section are used as an application of the above described
methodology. Two are the main issues associated with the
application of the above methodology, namely (a) the evaluation
of the stress field of the rock medium at the final excavation
stage, before initiation of the loosening of the rock fallout,
and (b) the comparison of the proposed approach with other
simplified methods.
The initial stress field for the application of the finite-

element Boundary Integral method is evaluated by a three-
597
dimensional finite element BMINES code, Agbabian (1976). The

three-dimensional model of the intersection illustrated in
Figure 1 includes 490 brick elements. Thirteen excavation
stages disposed in three consecutive horizontal layers are used
to simulate the removal of the rock. The liner is modeled
with a grid of beams that are adequately calibrated to respond
in a similar way as the real liner made of shotcretc and
steel ribs.
The corresponding vertical stress fields at stages 8, 9, andl3,

and for the diagonal section shown in Figure 1, are illustrated
in Figure 2. The stress field at stage 13 is retained as the
initial stress field of the rock fallout study. Interestingly,
it can be observed that loosening conditions are seen to
develop close to the base of the rock wedge, Figure 2, stage
13. The geometric configuration of the assumed failure
surface that defines the rock wedge is given in Figure 3, a.
Investigation of the effect of the above defined rock fallout
on the liner requires the retention of a smaller domain of
analysis as shown in Figure 4. If a finite element model
is used in conjunction with the simulation of the relative
displacements of the rock wedge with respect to the rock
medium using "joint elements," an estimated 1200 degrees of
freedom would be necessary to describe accurately the physical
phenomenon. Retaining the same discretization of the
boundaries as in the finite element approach, but using the
boundary integral formulation, reduces the degrees of freedom
to 180, resulting in a more tractable analysis.
The initial stresses and displacements of the boundaries of

the domains nl and n2 are obtained by the previously described
finite element model. This requires compatibility between
domains n and n at the interfacing boundary, achieved by
using spafial 1-b elements. Sliding between domains n 1 and n 2
is modeled to be initiated at the base of the interface
surface, and is simulated by releasing the tangential fric-
tional resistance, Figure 4. The resulting displacements at
the interface between nl and n2 obtained by the boundary
integral model used assuming a rigid liner are shown in Table
1. The equality between normal and tangential stresses in
iteration 1, Table 1, is due to the 45° slope of the assumed
discontinuity surface. The displacement fields in Ql and n 2
are seen to be incompatible along the interface. Re axing
the assumption of rigidity of the liner, the vertical displace-
ment field of the liner is obtained by use of a simplified
finite element discretization of the structural components of
the liner, namely the steel ribs and the shotcrete. This
technique was originally applied by Brierley (1975). In a
second iteration, the incompatibilities of the normal
displacements at the interfacing boundary are removed resulting
in an adjusted tangential displacement field and normal force
field along the interface, Table 1. Reduction of the normal
598
force incompatibilities indicate smooth convergence of the

iterative scheme. In iteration 2, incremental tangential
stresses are computed as resulting from the adjustment of the
relative normal displacement, Table 1. To this displacement
field corresponds a new set of displacements of the liner
shown in Table 2. The displacement field of the final itera-
tion is added to the displacement field obtained from the
finite element model of the excavation process. This total
displacement field corresponds to the stresses of the steel
ribs and the stresses of the shotcrete shown in Table 4.
The above results are compared to a simplified model in which

the action of the rock fallout is simply simulated by the
action of the gravity load. This comparison is shown in
Tables 2 and 3. Interestingly, the difference between the
above two approaches is larger at the springline of the inter-
section, thesimplifiedmodelshowing smaller displacements at
that location. Thus, the proposed Finite Element Boundary
Integral method can produce a more realistic analysis. How-
ever, attention should be given to the importance of the
initiation of the loosening conditions. In the above applica-
tion, loosening was assumed to be initiated at the springline
of the liner.
In fact many causes can initiate loosening conditions, such as,

for example, the alteration of physical properties of the rock
mass due to the presence of water. Finally, concerning the
slippage conditions at the discontinuity surface, kinetic
friction and dynamic effects ought to be included in the model
if blasting operations are to be conducted at a nearby
construction site.
ACKNOWLEDGMENTS
The investigation reported above was partially supported by

NSF, grant No. DAR76-80044.
599
Agbabian Assoc. (1976) User's Guide for a Computer Program for

Analytical Modeling of Rock/Structure Interaction, U-7638-3-
4183. El Segundo, CA:AA, Jun.
Brebbia, C.A. (1978) "The Boundary Element Method for

Engineers," Pen tech Press.
Brierley, G.S. (1975) "The Performance During Construction of

the Liner for a Large, Shallow Underground Opening in the Rock."
Ph.D. Dissertation, Univ. of Illinois, Urbana.
Brown, E.T. and Hocking, G. (1976) "The Use of Three Dimen-

sional Boundary Integral Equation Method for Determining
Stresses at Tunnel Intersections," in Proc. 2nd Australian
Tunnelling Conf. Melbourne, 1976, pp. 55-64.
Cording, E.J.; Mathews, A.A.; and Peck, R.B. (1976) Design

Criteria for Permanent Structural Linings for Station Excava-
tions in Rock, Washington Metropolitan Area Transit Authority.
Unpublished draft report prepared for DeLeuw, Cather and Co.,
General Engineering Consultant, Washington Metropolitan Area
Transit Authority, Washington, D.C., Jun.
Cruse, T.A. and Rizzo, F.J., eds. (1975) Boundary Integral

Equation Method: Computational Applications in Applied
Mechanics, AMD-Vol. 11, ASME, New York.
Ghaboussi, J. ; Wilson, E. L. ; and Isenberg, J. (197 3) "Finite

Element for Rock Joints and Interfaces," J. Soil Mech. Found.
Div., ASCE Vol. 99, No. SMlO.
Goodman, R.E.; Taylor, R.L.; and Brekke, T.L. (1968) "A Model
for the Mechanics of Jointed Rock," J. Soil Mechanics and
Foundations Div., ASCE, Vol. 94, No. SM3.
Jaeger, J.C. (1971) "Friction of Rocks and Stability of Rock

Slopes," Geotechnique, Vol. 21.
Zienkiewicz, o.c. et al. (1970) "Analysis of Nonlinear

Problems in Rock Mechanics with Particular Reference to Jointed
Rock Systems," Proc: 2nd Cong. of Int. Soc. for Rock Mechn.,
Belgrade.
Table 1. ~~mputed
8
Displacements and Pressures at the Interface Between n1 and n2
Normal Direction Tangential Direction

Normal Normal Tangential 1 1 Tangential
Displacement Traction Displacement Traction
[in.] [psi] [in.] [psi]
Itera- Loca-* Relative Relative ~elative Relative
tions tions Ql Q2 Value Ql Q2 Value Ql Q2 Ql Q2
Value Value
Given Given
1 a -.0124 -.0113 .0011 50 -so .0800 .0908 .0108 -so +50
b -.0130 -.0117 .0013 50 -40 .0935 .0974 .0039 -40 +40
c -.0132 -.0107 .0025 35 -35 .0970 .0978 .0009 -35 +35
Given
2 a .ooos -.0005 3.3 -3.5 .20 ').8 -0.7
b .0006 -.0006 4. -4.3 .30 1.2 -1.
c .0012 -.0012 s. -5.2 .20 1.3 -1.5
*Locations are indicated in Figure 4.

Table 2. Computed Vertical Displacements (in inches)
F.E. Gravity Model B.I. - F.E. Model

Displ. at
Steel Final
Ribs Nodes Excavation Partial Displ. Total Displ. Partial Displ. Total Displ.
1 -.076 -.0137 -.0897 -.014 -.090

3* 2 -.07 +.0074 -.064 +.0082 -.078
3 -.082 +.0119 -.071 -.00203 -.084
4 -.096 -.0024 -.098 -.003 -.099
1 -.08 -.0182 -.0982 -.026 -.106

4 2 -.068 -.0208 -.0888 -.029 -.097
3 -.073 -.0018 -.074 -.0025 -.075
4 -.099 +.0023 -.101 -.003 -.132
1 -.085 -.0530 -.138 -.054 -.139

5 2 -.017 -.0991 -.170 -.098 -.169
3 -.073 -.0465 -.119 -.049 -.122
4 -.09 -.0104 -.100 -.015 -.105
1 -.085 -.0543 -.139 -.055 -.140

6 2 -.07 -. 0771 -.149 -.078 -.148
3 -.075 -.0325 -.107 -.035 -.110
4 -.094 -.0098 -.103 -.012 -.106
*Numbers of steel ribs are given in Figure 5.

en
~
en
0
1\)
Table 3. Compressive Stresses in Steel Ribs (in p.si)
Final Excavation F.E. Gravity Model B.I. - F.E. Model

Steel
Ribs Members Strains Stresses Strains Stresses Strains Stresses
1 1. 53 E-3 46.10 2.43 E-3 70.0 2.50 E-3 72.0

3* 2 1. 57 E-3 45.70 2.42 E-3 70.0 2.48 E-3 72.0
3 1.47 E-3 42.70 1. 51 E-3 44.0 2.0 E-3 58.0
1 1. 61 E-3 46.6 2.40 E-3 69.0 2.45 E-3 70.0

4 2 1.62 E-3 47.0 2. 39 E-3 69.0 2.50 E-3 71.0
3 1. 52 E-3 44.10 2.23 E-3 65.0 2.48 E-3 72.0
1 1.23 E-3 35.50 1.64 E-3 47.0 1.7 E-3 49.0

5 2 1. 36 E-3 39.40 2. 39 E-3 69.0 2.4 E-3 69.0
3 0.88 E-3 25.64 2.3 E-3 67.0 2.4 E-3 70.0
1 2.3 E-3 67.28 0.89 E-3 26.0 1.0 E-3 29.0

6 2 2.31 E-3 67.10 1. 25 E-3 36.0 1.28 E-3 38.0
3 0.55 E-3 16.10 1. 36 E-3 39.0 1.4 E-3 40.0
1 0.83 E-3 24.10 1.29 E-3 37.0 1.3 E-3 38.0

7 2 0.67 E-3 19.37 0. 79 E-3 23.0 0.8 E-3 23.0
3 1.01 E-3 29.50 1.08 E-3 31.0 1.42 E-3 32.0
*
Numbers of steel ribs are given in Figure 5.
Table 4. Stresses in Shotcrete at Four Gaussian Points (in psi) for Panel A Shown in Figure 5
Principal Stresses
Gauss
(J (J T (Jl (J2
Point X y xy
Final
Excavation 1 -666.0 -327. 895. -1407.0 419.0
2 -1130.0 -316. 666. -1504.0 58.0
3 -444.0 -340. 190. -1099.0 345.0
4 -471.0 -353. 283. -918.0 94.0
F.E. 1 -691.0 70.0 371 -842 222

Gravity 2 -1150.0 -10.0 33 -1151 9
Model 3 -467.0 51.0 745 -997 581
4 -486.0 -96.0 482 -881 229
B.I. - F.E. 1 -670.0 -108.0 700

Model 2 -1140.0 -60.0 350
3 -470.0 -150.0 260
4 -490.0 -408.0 200
(-) sign indicates compression.
m
0
w
604
FIGURE 1. THREE-DIMENSIONAL CONFIGURATION OF POSSIBLE ROCK

FALLOUT AT INTERSECTION WITH RESPECT TO FINITE
ELEMENT DISCRETIZATION
STAGE 8 STAGE 9 STAGE 13
- - - VOLUME OF ROCK EXCAVATED AT PART I CULAR EXCAVATION STAGE
FIGURE 2. FINITE ELEMENT ESTIMATION OF VERTICAL STRESS DISTRIBUTION IN ROCK MEDIUM en

0
AT DIAGONAL SECTION OF INTERSECTION t11
606
MATERIAL PROPERTIES
E = 500 KSI
NU = .1
1110 PSI
]0 PSI
20 PSI I NITI AL
STRESS
10 PSI
FI El D
Ill AOU
K£01UII'I
0 PSI
(a) Geometric configuration (b) Stress field applied

of rock fa l lout to 1 i ner
FIGURE 3. GEOMETRIC CONFIGURATION AND INITIAL VERTICAL

STRESS OF ASSUMED ROCK FALLOUT
BOUNDARIES STRESS FIELD
OBTAINED FROM FINITE
ELEMENT SOLUTION
f~ = TANGENTIAL FORCE
APPLIED TO ROCK
Q2
f~ =TANGENTIAL FORCE APPLIED

TO ROCK FALLOUT
FIGURE 4. DISCRETIZATION OF BOUNDRIES RETAINED FOR STUDY OF ROCK FALLOUT EFFECTS ~

.....
t 8
1 r--•S'---i! CO
I 0 I I 0
" 0 '" ,J I I I I .oc•

II(OIU~ l rI - >EDIUN ~~ rI -
~~ I I~ ~~ ~
~ r I 0
,.. II " II 3 II I CD ll' " l CD

IHTERSEC"T tDH t NTtRstcTtOH oF
ll...-J~" MIN OP[NtNC
or I MIN OPEN ING
WI TH AUESS WI TM A((£SS
TUNHEir. TUNNEl
~ ~
(a) Disp lacements at excavation stage 13 (b) Displacements due to rock fallout
FIGURE 5. CONTOURS OF COMPUTED DISPLACEMENTS AT UNDERGROUND INTERSECTION

609
THE USE OF GREEN'S FUNCTIONS IN THE NUMERICAL ANALYSIS

OF POTENTIAL, ELASTIC AND PLATE BENDING PROBLEMS
c. Katz
Technical University Munich, Germany
INTRODUCTION
Neither the Boundary Element Method (BEM) nor the Finite

Element Method (FEM) is the unique technique to solve
engineering problems. While the first method satisfies
the differential equation exactly and approximates the
boundary conditions, the latter does the reverse. This
implies in general that a solution obtained by the BEM
will not fulfil global conditions as equilibrium or
balance of fluxes in potential theory.
But this is not the only reason to link the BEM and the
FEM. An effective computer code exists for the FEM, but on
the other hand, the modelling of infinite regions is
difficult in FEM and the advantage of the BEM to dis-
cretize the boundary only follows the thinking of
practical engineers.
Linking procedures will in general try to calculate a
stiffness matrix (symmetric) using the BEM. Zienkiewicz
et al. (1977) showed a way to do this by using a vari-
ational principle and the direct integral equation
approach. Their method has two disadvantages, first it
implies the inversion of a fully populated nonsymmetric
matrix, second the independent interpolation of dis-
placements and stresses or potential and fluxes intro-
duces some errors.
The next step in a new direction was taken by Margulies
(1980). He showed that using the Green's function for a
special region a symmetric stiffness matrix can be cal-
culated. This approach however has the handicap that
this particular Green's function has to be known and the
severe restriction that the kernels of his equations are
610
strongly singular and any numerical integration is im-

possible for the diagonal coefficients.
This paper pushes forward Margulies' approach and shows
some new ways to get a stiffness matrix by the BEM.
THE METHOD OF MARGULIES
Margulies started with the fundamental equation of the

BEM which gives the values of the potential in an ob-
servation point z 0 by the boundary values ¢(z) and
~
a n
¢(z 0 ) = JG(z,z 0 ) · acp(z) ds -Ja G(z,z 0 ) • cjl(z)ds (1)

an an
1. ~I' H .. l'
In general G is the free space Green's function which is
the potential due to a unit source in z and is symmetric.
with respect to_z and z 0 • Changing G to the real Green's
function, that ~s the solution of the unit source in z
imposed to the boundary condition
G(z,z 0 ) =0 for z or z0 e r (2)
equation (1) simplifies to

cjl(z)
0
=-fa G(z,zo) cjl(z)ds ( 3)
an
... £,...
which is an explicit representation for the solution
given by the boundary values .¢ ( z ) •
Using this expression for the variational formulation of
J
the problem which is for an interior problem
TI = ~ ¢ (zo). (4)
lo€ ro
we obtain
'IT =- 1
-
2
f cjl(z 0 ) • JI a 2 G(zo,z)
· cjl(z) ds ds 0 (5)
anan
z 0e.r., zer o
and using the interpolation trial functions as usual for
¢
611
From this the symmetric stiffness matrix is deduced
a .. =- p§N. (z ) •
2
d G(zo,z) N]. (z) ds ds 0 (7)
l] r r l 0 anan
0 0
Margulies has calculated the stiffness matrix for the
cases when r is a circle or a sphere and for constant or
linear trial functions. However, the integration of the
diagonal coefficients has to be performed analytically
or by some additional constraints. In the following I
restrict myself to the potential problem. The extension
to elastic problems will be touched on shortly afterwards
and will be reported in detail elsewhere.
FIRST EXTENSION: GREEN'S FUNCTIONS FOR OTHER REGIONS
Though the Green's Function for the circle of radius a is

2 2 r
(r/a) + (a/r 0 ) - 2 (--) cos (a - a 0 )
rg
G - 4n ln ------------------------·------------~~ (8)
2
+ (r/r ) - 2(E_) cos (a- a 0 )
o r0
and the derivatives can be calculated from there, it is
much easier to use the complex representation. Each
harmonic function is the real part of an analytic funct-
ion. For the free space function the Green's function is
the real part of
(9)
The imaginary part of (9) is a harmonic function too and

is called the stream function.
To obtain the analytic Green's function for the circle
three components are needed
- a source in z 0 in the interior of the circle
- a sink in the inverse point a 2 / o z-
- a constant factor to shift the potential ln (z 0 /a)
The complex Green's function is then given by
-1 a(z - zQ)
GG = 2n ln ~~2~--~~ ( 10)
(a - zz0 )
The real part of (10) is
2
G = -1 a (Z - Zo) • ( Z - Zo) ( 11 )
4n ln
(a
2
- zz0 ) ·(a 2
- zz 0 )
and is equivalent to (8).
612
It is worth noting that z is a complex variable x + i y

while z is x - i y but is independent of z when calculat-
ing complex derivatives.
Similar the complex Green's function for the half plane
is given by
GG = - -1- ln z - Zo ( 1 2)
21T
z - z0
For other simply connected regions the complex Green's

function is given by the function for the unit circle or
half plane and the conformal mapping from the region to
the circle or half plane. Let c; represent a point in the
mapped plane, z in the original one, then the conformal
mapping is
c; = w(z) ( 1 3)
and for the complex Green's function the expression holds:

GGz (z,z 0 ) = GG(; (w(z), w(z 0 )) (14)
The proof can be found in Kantorowitsch/Krylow (1964).
If the region to be mapped is the unit square, the con-
formal mapping to the unit circle is given by Gaier (1964)
5 9 13
(; = a·z + ~ . z 5 + ~ · z 9 +~ . z 13 + .•.
10 120 15600
a = 0.92703733865 ( 1 5)
The mapping is of high quality as is shown in the follow-

ing figure
n II'~ _J
·-
_/'\
ki
__ I
\- ~ ·-I
~
II
1-·
l\i
~
\
LI I
I
-\ L-\'
I I! 1
Figure I Conformal mapping of the unit square (15)
using 4 coefficients
This function is used to calculate the Green's function
for the unit square. The following figure gives the lines
613
of equal potential for this function.
Figure II Equipotentials, Green's function for a square
To calculate the derivatives the following expressions

can be used
2 . az
ClG ClG
ax - i
ClG
ay ( 1 6)
2 . aG aG + i
aG ( 1 7)
az ax ay
when the normal and tangent of the boundary are given by
n = e i8 cos8 + i • sine ( 18)

s = i · n
the derivatives in the direction of n and s will be
( 1 9)
an az az
i aG e
-ie aG e
i8 aG ( 20)
as az az
and from equation (19) follows
a2G e
i8
e
i8o a2G + e
-i8 -i8o
e
a2G
C:lnC:ln 0 az 0 az az0 az
-i8 i8o Cl2G i8 -i8o Cl2G ( 21 )

e e + e e
az 0 az a"Za az
614
SECOND EXTENSION: NUMERICAL EVALUATION OF THE INTEGRALS
Though the kernel of the stiffness matrix is now known

for all regions which can be mapped on the unit circle
or the half plane the numerical evaluation of integral
(7) is impossible. The original Green's function is of
logarithmic order, the second derivative is of order
(1/r2) and this can not be integrated in a double in-
tegral without special tricks.
This is a problem which seems spurious because there is
no phrsical reason for it. Indeed the difficulties come
from the singularity of the unit potential on the
boundary, while the potential in nature has no jumps at
all. In the following figures the flow-field for a unit
potential and the real potential on the boundary is
shown. It is evident that in the former case the fluxes
will be infinite, while the latter case has finite
fluxes.
<P =1
Figure III Unit Potential
,_ ,
// I I
... ~
,' ,I
'
'
' ....
Figure IV Distributed Potential {real case)

615
Moreover, the nature of the trial functions is that they

are bounded in some parts of the boundary and vanish on
the rest of the boundary.
All this leads to the idea of integrating by parts. This
is first performed on equation (3) which gives the value
of cJ> at z 0 •
cj> (zo) = -jr ClG(z,zQ)

Cln
cj> ( z) ds
- JoG
-
Cln
ds•cj>(z) +f • asClcj>(z). J~~ ds ds (22)
r
The first part will vanish i f we integrate in a closed

contour or if we consider only one trial form. However,
there might be some additional values due to the multi-
valued functions but this is discussed later. So we have
the new function
G*< z' zo ) = JClG(z,zQ)

Cln
ds (23)
which I call the Green's function of the second kind.

With this function the expression for an interior value
will be
= f
r
a cJ> <z >
Cls
• G* ds - cj> ( z 1 ) 1
I"'
G* ds
If:J ClNi
as
*
G ds • cj>i + cj>o (24)
...
Similar a second partial integration is performed on the

other part of the boundary r 0 • This leads to the ex-
pression for the stiffness matrix
a ..
l.J
tf ClNi • G**<z,zb) • -
Clsi
ClNi
- ds. dsJ.
Clsj l.
( 2 5)
rr
in which
G**(z,z )
o
= JJ ~ n ()G(z,zo)ds
2
n 0
ds
o
(26)
is called the Green;s function of the third kind.

616
Both new functions remain of logarithmic order and can,

therefore, be integrated easily using any quadrature
formula. As one can see, a constant potential gives no
value to (24) than ~ 0 and not any to the stiffness matrix.
Indeed this means that the row-sums of the stiffness
matrix are exactly zero. This even holds for each stage
of the integration because the feature of the trial
functions (their sum is one in every point) yields that
their sum of the derivatives will vanish in every point.
So far this approach seems very easy. To get the Green's
functions of the second or third kind, a normal complex
integration would be very cumbersome. But there is some
more physical meaning in these functions.
To show this, let us first assume that our Green's
function of the first kind G can be r&presented as the
sum of four analytic functions as follows
( 2 7)
This is true in all cases because G is the real part of

an analytic function and is harmonic in z and z 0 • Using
(27) the two derivatives will become
<lG
e
-18 (~ + a!) + e18 (~ + .£.1) (28)
an az az az az
and
·
e-~ 8 e-~ 8 o
· a2 i + e18e18 0 ~2,."'
a
-18 18 0 ()2~ 18 -18 0 ()21/J

+ e e + e e (29)
azaz 0 azaz 0
I claim now that the functioll:; G* and G** are given by
G*= 1
~
I(z,z 0 )-i(i,i 0 ) + 1/J(z,i 0 ) - ~(i,z 0 ) I (30)
1/J(z,i 0 ) + ~(i,z
0
)I ( 31 )
617
The proof is given by the fact that
G* =JaG ds + Canst ~ aG*

as
3G
an
(32)
an
and
**-If
G- - 2 - dsds + Canst
aan G
ana
a ~
a 2 G**
asasa
ri 2 G
anan a
(33)
and applying the following rules to (30) and (31)
-ie aG* ie aG*

aG*
= - e (34)
az-
i e
as az
and
.2 32G** 2 2
l eieei8a a G** + e -ie e -ie a a G**
as a sa ~ az a a~
a
(35)
- e
ie e-iea a 2 G**
- e-ieeiea a 2 G**
azaza a~aza
Looking at (30) and (31) closer one can see that the
Green's function of the second kind is nothing else but
the imaginary part of the complex Green's function or,
if one prefers, the so called stream function. Greenis
function of the third kind can be interpreted as comple-
ment function to the one of the first kind because the
two singularities now have the same sign.
SPECIAL NUMERICAL CONSIDERATIONS
Let us start with the Green's function of the second kind.
l {
For the unit circle it will be
z-z
1_ • Im [ ln ~ = 2~ z-z ~
~ (36)
G* = __ • arctan
2rr 1-zz 1-zz
a a
when going round about the circle, this function will
jump by the value 1. So¢ in equation (24) has the
value of the potential wh~re the jump occurs. It is con-
venient to shift this jump to a node to avoid difficult
interpolations. Normally the function will jump when the
argument is - a. (This is dependent on the computer).
This will be when
618
z-z 0 z-z 0
< 0 (37)
1-zz 0 1-zz 0
Noting that z is on the unit circle and
z = o = 1;-cr (38)
one gets a quadratic equation with the following solut-
ions 1-z
0 1+z
01 = - --_- 0
(39)
1-zo 1+z0
where o1 is the right one.
The evaluation of the integral (24) starts at the chosen
node counterclockwise. When position o1 is passed the
value 1 has to be added to G*.
To calculate the fluxes in some point z equation (24) is
differentiated with respect to -z 0 which gives
0
f
zer
cp.
1
aNi
as
(;j"'i'(Z ) d s
0
(40)
This is the only formula where the derivative of the con-

formal mapping occurs. In all other relations only the
coordinates of z and z or of their images are needed.
This is caused by the 1act that the solution of the
potential problem and finding the conformal mapping are
equivalent. ·
For the calculation of the stiffness matrix only the
upper half has to be calculated. It is useful to inter-
prate the double integral as two dimensional integral and
using an isoparametric concept. So we have to integrate
in a unit square where each dimension represents one
boundary element and can use n by n Gauss points:
~'> ,!' Yy
NL
r.
* * *
1 f·t
* * *
* * *
t-+r ,.......v
~j
Figure v Integration schemes
619
If the two elements coincide, the logarithmic singularity

is on the diagonal of the square. As we calculate only
the half of the stiffness matrix we have to restrict our-
selves to one half of the square. Therefore, we can in-
tegrate in one triangle only and use special quadrature
formulas or transform our n by n Gauss points by
rr = i [r (1-r 0 ) - (1 + r 0 )]
ss = !1 [r (1-r 0 ) + (1 + r 0 l] ( 41 )
The determinant of this transformation will be

det !J! = i (1 - r 0 ) (42)
with which the integrand has to be multiplied.
If desired, the conformal mapping can be improved by
using our knowledge that the images of the boundary
points are located on the unit circle reap. the real
axis and applying the following corrections to the images
circle z.; : (43)
half plane z.;:= Re[z.;] (44)
NUMERICAL EXAMPLE
To demonstrate the new approach we consider a square

with prescribed potential at two opposite sides. The
fluxes will be constant, the potential will vary linearly.
Each side is divided in three linear elements with four
Gauss points for each element.
<P=O
<1>=300 +4 +1
+5 +2
t-6 +3
Figure VI Constant flow example

620
As we use only a trial form for the potential no corner

problem exists besides that in the conformal mapping.
First we calculate some values in interior points
applying the Green's function of the first and the second
kind (equations 3,24,40) prescribing the exact solution
of our problem at the boundary. The conformal mapping
was done by equation (15) using four coefficients. The
following table gives the calculated values in the points
marked in Figure VI. The exact values of the fluxes are
~
,x = 5.000 and ¢ ,y = 0.000.
As one can see, the values obtained by the Green's
function of the second kind are of much higher quality
than the original ones. The fluxes in the points near
the boundary are very poor when one uses the Green's
function of the first kind.
Table I Values at internal points
Point exact Values calculated with Values calculated with

no. value Green 1 s function 1st kind Green 1 s function 2nd kind
<!> <!> <!>'X <l>,y <!> <!>'X <l>,y
1 15o 15o.o 4.999 o.aoo 15o.o 4.999 o.ooo

2 15o 149.9 5.a17 o.o65 15o.o 4.995 o.ooo
3 15o 146.2 5.378 3.o36 15o.o 4.872 o.ooo
4 275 267.4 -1.o7a o.ooo 275.o 5.oo1 o.oao

5 275 267.4 -1.2o3 o.2o4 274.9 5.o11 a.o22
6 275 2Bo.o 2.5o2 3.210 274.8 4.981 o.o14
Then a stiffness matrix was calculated following equation

(25). This matrix is fully populated blt exactly symmetric
and has the eigen-value zero i.e. the row sums are zero
within the computer precision. This matrix was used to
solve the given problem. After a solution is obtained the
original matrix is multiplied with the potential to give
the nodal sources.
Q. = a ..• cp. (45)
~ ~J J
The next table gives the potential in the two nodes on

the upper resp.lower side and the nodal sources in the
nodes with prescribed potential. The values do not con-
621
verge to the exact solution but to that one which is

prescribed by the conformal mapping.
Table II Values obtained by the stiffness matrix
Nu111ber of Gauss SUIII

points/ ele11an t <1>1 <1>2 Q1 Q2 of Q
exact value 2oo.oo 1oo.oa 5.ooo 1o.ooo 3o.oo

4 199.9o 1oo.1o 4.633 1o.143 29.55
5 199.98 1oo.o2 4.714 1o.142 29.71
6 2oo.o3 99.97 4.832 1o.o64 29.8o
8 2oa.17 99.83 4.882 1o.o69 29.9a
EXTENSION TO ELASTIC PROBLEMS
The technique for elastic systems is similar but much

more difficult. First and most severe restriction is the
Green's function.
Elastic problems are prescribed by the biharmonic equat-
ion. The use of conformal mapping is therefore not easy
and a simple substitution does not give the GreenJs
function for the mapped region. Moreover, Green's funct-
ions for even very simple regions are rather complicate.
The only way I can see to obtain those functions is the
method of Muskelishvilli (1951). He used the complex
representation of a biharmonic equation called the
formula by Goursat:
MG = o {==:9 G = Re (z ~II' ( z) + X ( z) ] ( 4 6)
where 10 and X are analytic functions.

The partial integration is more complicatedthan the
previous one but facilitated by the use of equation (46).
The stretching of plates is described by a logarithmic
kernel. So the technique remains quite similar to that
of the potential problem. The kernel for bending of
plates however is of order r 2ln r and here higher de-
rivatives occur. Despite of this, it is possible to keep
in the logarithmic order integrating the rotational
degrees by part once and the deflections twice. The re-
search on these problems is still going on and will be
reported elsewhere.
622
CONCLUSION
A technique was developed in this paper to calculate

exact stiffness matrices for a given region using
Boundary Elements and the complex representation of
harmonic or biharmonic functions. A numerically stable
approach is given which makes use of a continuous
potential or displacement at the boundary. New types of
Green's functions are introduced, the so called Green's
functions of the second and third kind.
REFERENCES
Banerjee, P.K., Butterfield, R. (1979) Developments in

Boundary Element Methods I, Applied Science Publishers
Ltd., London.
Bettess, P. (1977) Infinite Elements, Int. Journ. for
num. Meth. in Engineering Vol 11, 53-64.
Brebbia, C.A. (1978) The Boundary Element Method for
Engineers. Pentech Press, Plymouth.
Gaier, D. (1964) Konstruktive Methoden der konformen
Abbildung. Springer Verlag, Berlin.
laswon, M.A., Symm, G.T. (1977) Integral Equation
Methods in Potential Theory and Elastostatics, Academic
Press, London.
Kantorowitsch /Krylow (1964) Approximative Methods of
Higher Analysis, Noordhof Ltd., Groningen.
Margulies, M. (1980) Exact treatment of the exterior
problem in the combined FEM-BEM. New Developments in
Boundary Element Methods (c. Brebbia Edit.) CML Publi-
cations.
Morse, P., Feshbach, H. (1953) Methods of Theoretical
Physics. McGraw-Hill, New York.
MUskhelishvilli, N.I. (1951) Some basic problems of
mathematical theory of elasticity. Noodrhof Ltd.,
Groningen.
MUst9e, G.G.W., Volait, F. (1980) A Symmetric Direct
Boundary Integral Equation Method for two Dimensional
Elastostatics.
Zienkiewicz, O.C., Kelly, D.W., Bettess, P. (1977) The
Coupling of the Finite Element Method and Boundary
Solution Procedures. Int. Journal Num. Meth. in En-
gineering Vol 11 pp 355-375.

(Boundary Elements 3) George T. Symm (auth.), Dr. Carlos A. Brebbia (eds.) - Boundary Element Methods_ Proceedings of the Third International Seminar, Irvine, California, July 1981-Springer-Verlag Ber.pdf

Uploaded by

Copyright:

Available Formats

(Boundary Elements 3) George T. Symm (auth.), Dr. Carlos A. Brebbia (eds.) - Boundary Element Methods_ Proceedings of the Third International Seminar, Irvine, California, July 1981-Springer-Verlag Ber.pdf

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Boundary Elements 3) George T. Symm (auth.), Dr. Carlos A. Brebbia (eds.) - Boundary Element Methods_ Proceedings of the Third International Seminar, Irvine, California, July 1981-Springer-Verlag Ber.pdf

Uploaded by

Copyright:

Available Formats

What is the document about?

What is the document about?

What methods are discussed for solving potential and fluid flow problems?

What methods are discussed for solving potential and fluid flow problems?

Boundary

Seminar sponsored by the International Society for Computational

With 232 Figures

Springer-Verlag Berlin Heidelberg GmbH 1981

ISBN 978-3-662-11272-4 ISBN 978-3-662-11270-0 (eBook)

© Springer-Verlag Berlin Heidelberg 1981

SECTION I POTENTIAL AND FLUID FLOI-1 PROBLEMS

Two Methods for Computing the Capacitance of a 3

On the Construction of the Boundary Integral 71

A Comparative Study of Several Boundary Elements in 123

Boundary Elements applied to Soil-Foundation 293

SECTION IV MATERIAL PROBLEMS 335

Numerical Analysis of Cyclic Plasticity using the 337

Some Theoretical Aspects of Boundary Integral 399

SECTION VI COUPLING OF BOUNDARY AND FINITE ELEMENT 539

Interfacing Finite Element and Boundary Element 541

The great interest that finite element methods have

The first of the above points came as a natural conse-

These techniques can be traced to pre-computer times [1 ,2]

Pian [4]. An excellent exposition of mixed methods in

It is difficult to point out precisely who was the first

Since the early 1960s a small research group started

More recently this work has been e~anded to encompass

One of the most important recent applications of boundary

The boundary element method generally offers the advant-

It may be convenient in many applications to combine

Ways of s~trizing the BE matrices have been suggested

2. REVIEW OF THE PAPERS PRESENTED

The 3rd International Conference on Boundary Element Methods

One of the most interesting set of papeiS is that dealing

ideas have been put forward.

Many papers deal with the coupling of BE and FE solutions.

Several of the papers presented here refer to potential

The paper by Curran et al. [45] discusses the solution of

Pignole [4~ dealing with linear elastic problems, dis-

Another integral formulation called by the authors the

The important field of using BEM to solve crack problems

Antes [57] extends the BEM for the solution of circular

BEM has been particularly successful to study contact

Wong and Hutchinson j}o] use BEM to determine the natural

but numerical results are not givEn as the computer code is

Part III of this book .deals with geomechanic problems.

BEM is increasingly becoming an accepted method of solu-

Several important papers are presented in the section

Herrera [69] tries to establish a systematic approach to the

The work of Futagami [7!] presents the first attempt to

Section VI of these proceedings deal with the topic of

Hartmann [77] reviews in a very clear manner the differ-

3. SUGGESTIONS FOR FURTHER RESEARCH

Several important topics related to boundary solutions or more

It is also evidentto finite element practitioners that the

Although current boundary element practice is heavily

Time dependent problems may be one of the most fruitful

I. Kantorovich,L.V. and Krylov, V.I., Approximate Methods of

2. Courant, R. and Hilbert, D. Methods of Mathematical

3. Reissner, E." A Note on Variational Principles in Elasticity