DK3139_half 1/20/05 11:13 AM Page 1
Boundary
Methods
Elements, Contours, and Nodes
© 2005 by Taylor & Francis Group, LLC
MECHANICAL ENGINEERING
A Series of Textbooks and Reference Books
Founding Editor
L. L. Faulkner
Columbus Division, Battelle Memorial Institute
and Department of Mechanical Engineering
The Ohio State University
Columbus, Ohio
1. Spring Designer’s Handbook, Harold Carlson
2. ComputerAided Graphics and Design, Daniel L. Ryan
3. Lubrication Fundamentals, J. George Wills
4. Solar Engineering for Domestic Buildings, William A. Himmelman
5. Applied Engineering Mechanics: Statics and Dynamics, G. Boothroyd
and C. Poli
6. Centrifugal Pump Clinic, Igor J. Karassik
7. ComputerAided Kinetics for Machine Design, Daniel L. Ryan
8. Plastics Products Design Handbook, Part A: Materials and Components;
Part B: Processes and Design for Processes, edited by Edward Miller
9. Turbomachinery: Basic Theory and Applications, Earl Logan, Jr.
10. Vibrations of Shells and Plates, Werner Soedel
11. Flat and Corrugated Diaphragm Design Handbook, Mario Di Giovanni
12. Practical Stress Analysis in Engineering Design, Alexander Blake
13. An Introduction to the Design and Behavior of Bolted Joints, John H.
Bickford
14. Optimal Engineering Design: Principles and Applications, James N. Siddall
15. Spring Manufacturing Handbook, Harold Carlson
16. Industrial Noise Control: Fundamentals and Applications, edited by
Lewis H. Bell
17. Gears and Their Vibration: A Basic Approach to Understanding Gear Noise,
J. Derek Smith
18. Chains for Power Transmission and Material Handling: Design and Appli
cations Handbook, American Chain Association
19. Corrosion and Corrosion Protection Handbook, edited by
Philip A. Schweitzer
20. Gear Drive Systems: Design and Application, Peter Lynwander
21. Controlling InPlant Airborne Contaminants: Systems Design
and Calculations, John D. Constance
22. CAD/CAM Systems Planning and Implementation, Charles S. Knox
23. Probabilistic Engineering Design: Principles and Applications,
James N. Siddall
DK3139_series.qxd 1/20/05 11:17 AM Page 1
© 2005 by Taylor & Francis Group, LLC
24. Traction Drives: Selection and Application, Frederick W. Heilich III
and Eugene E. Shube
25. Finite Element Methods: An Introduction, Ronald L. Huston
and Chris E. Passerello
26. Mechanical Fastening of Plastics: An Engineering Handbook,
Brayton Lincoln, Kenneth J. Gomes, and James F. Braden
27. Lubrication in Practice: Second Edition, edited by W. S. Robertson
28. Principles of Automated Drafting, Daniel L. Ryan
29. Practical Seal Design, edited by Leonard J. Martini
30. Engineering Documentation for CAD/CAM Applications, Charles S. Knox
31. Design Dimensioning with Computer Graphics Applications,
Jerome C. Lange
32. Mechanism Analysis: Simplified Graphical and Analytical Techniques,
Lyndon O. Barton
33. CAD/CAM Systems: Justification, Implementation, Productivity Measure
ment, Edward J. Preston, George W. Crawford, and Mark E. Coticchia
34. Steam Plant Calculations Manual, V. Ganapathy
35. Design Assurance for Engineers and Managers, John A. Burgess
36. Heat Transfer Fluids and Systems for Process and Energy Applications,
Jasbir Singh
37. Potential Flows: Computer Graphic Solutions, Robert H. Kirchhoff
38. ComputerAided Graphics and Design: Second Edition, Daniel L. Ryan
39. Electronically Controlled Proportional Valves: Selection and Application,
Michael J. Tonyan, edited by Tobi Goldoftas
40. Pressure Gauge Handbook, AMETEK, U.S. Gauge Division,
edited by Philip W. Harland
41. Fabric Filtration for Combustion Sources: Fundamentals and Basic
Technology, R. P. Donovan
42. Design of Mechanical Joints, Alexander Blake
43. CAD/CAM Dictionary, Edward J. Preston, George W. Crawford,
and Mark E. Coticchia
44. Machinery Adhesives for Locking, Retaining, and Sealing, Girard S. Haviland
45. Couplings and Joints: Design, Selection, and Application, Jon R. Mancuso
46. Shaft Alignment Handbook, John Piotrowski
47. BASIC Programs for Steam Plant Engineers: Boilers, Combustion, Fluid
Flow, and Heat Transfer, V. Ganapathy
48. Solving Mechanical Design Problems with Computer Graphics,
Jerome C. Lange
49. Plastics Gearing: Selection and Application, Clifford E. Adams
50. Clutches and Brakes: Design and Selection, William C. Orthwein
51. Transducers in Mechanical and Electronic Design, Harry L. Trietley
52. Metallurgical Applications of ShockWave and HighStrainRate Phenomena,
edited by Lawrence E. Murr, Karl P. Staudhammer, and Marc A. Meyers
53. Magnesium Products Design, Robert S. Busk
54. How to Integrate CAD/CAM Systems: Management and Technology,
William D. Engelke
DK3139_series.qxd 1/20/05 11:17 AM Page 2
© 2005 by Taylor & Francis Group, LLC
55. Cam Design and Manufacture: Second Edition; with cam design software
for the IBM PC and compatibles, disk included, Preben W. Jensen
56. SolidState AC Motor Controls: Selection and Application,
Sylvester Campbell
57. Fundamentals of Robotics, David D. Ardayfio
58. Belt Selection and Application for Engineers, edited by Wallace D. Erickson
59. Developing ThreeDimensional CAD Software with the IBM PC, C. Stan Wei
60. Organizing Data for CIM Applications, Charles S. Knox, with contributions
by Thomas C. Boos, Ross S. Culverhouse, and Paul F. Muchnicki
61. ComputerAided Simulation in Railway Dynamics, by Rao V. Dukkipati
and Joseph R. Amyot
62. FiberReinforced Composites: Materials, Manufacturing, and Design,
P. K. Mallick
63. Photoelectric Sensors and Controls: Selection and Application, Scott M. Juds
64. Finite Element Analysis with Personal Computers, Edward R. Champion, Jr.
and J. Michael Ensminger
65. Ultrasonics: Fundamentals, Technology, Applications: Second Edition,
Revised and Expanded, Dale Ensminger
66. Applied Finite Element Modeling: Practical Problem Solving for Engineers,
Jeffrey M. Steele
67. Measurement and Instrumentation in Engineering: Principles and Basic
Laboratory Experiments, Francis S. Tse and Ivan E. Morse
68. Centrifugal Pump Clinic: Second Edition, Revised and Expanded,
Igor J. Karassik
69. Practical Stress Analysis in Engineering Design: Second Edition,
Revised and Expanded, Alexander Blake
70. An Introduction to the Design and Behavior of Bolted Joints: Second Edition,
Revised and Expanded, John H. Bickford
71. High Vacuum Technology: A Practical Guide, Marsbed H. Hablanian
72. Pressure Sensors: Selection and Application, Duane Tandeske
73. Zinc Handbook: Properties, Processing, and Use in Design, Frank Porter
74. Thermal Fatigue of Metals, Andrzej Weronski and Tadeusz Hejwowski
75. Classical and Modern Mechanisms for Engineers and Inventors,
Preben W. Jensen
76. Handbook of Electronic Package Design, edited by Michael Pecht
77. ShockWave and HighStrainRate Phenomena in Materials, edited by
Marc A. Meyers, Lawrence E. Murr, and Karl P. Staudhammer
78. Industrial Refrigeration: Principles, Design and Applications, P. C. Koelet
79. Applied Combustion, Eugene L. Keating
80. Engine Oils and Automotive Lubrication, edited by Wilfried J. Bartz
81. Mechanism Analysis: Simplified and Graphical Techniques, Second Edition,
Revised and Expanded, Lyndon O. Barton
82. Fundamental Fluid Mechanics for the Practicing Engineer,
James W. Murdock
83. FiberReinforced Composites: Materials, Manufacturing, and Design,
Second Edition, Revised and Expanded, P. K. Mallick
84. Numerical Methods for Engineering Applications, Edward R. Champion, Jr.
DK3139_series.qxd 1/20/05 11:17 AM Page 3
© 2005 by Taylor & Francis Group, LLC
85. Turbomachinery: Basic Theory and Applications, Second Edition,
Revised and Expanded, Earl Logan, Jr.
86. Vibrations of Shells and Plates: Second Edition, Revised and Expanded,
Werner Soedel
87. Steam Plant Calculations Manual: Second Edition, Revised and Expanded,
V. Ganapathy
88. Industrial Noise Control: Fundamentals and Applications, Second Edition,
Revised and Expanded, Lewis H. Bell and Douglas H. Bell
89. Finite Elements: Their Design and Performance, Richard H. MacNeal
90. Mechanical Properties of Polymers and Composites: Second Edition,
Revised and Expanded, Lawrence E. Nielsen and Robert F. Landel
91. Mechanical Wear Prediction and Prevention, Raymond G. Bayer
92. Mechanical Power Transmission Components, edited by David W. South
and Jon R. Mancuso
93. Handbook of Turbomachinery, edited by Earl Logan, Jr.
94. Engineering Documentation Control Practices and Procedures,
Ray E. Monahan
95. Refractory Linings Thermomechanical Design and Applications,
Charles A. Schacht
96. Geometric Dimensioning and Tolerancing: Applications and Techniques
for Use in Design, Manufacturing, and Inspection, James D. Meadows
97. An Introduction to the Design and Behavior of Bolted Joints: Third Edition,
Revised and Expanded, John H. Bickford
98. Shaft Alignment Handbook: Second Edition, Revised and Expanded,
John Piotrowski
99. ComputerAided Design of PolymerMatrix Composite Structures,
edited by Suong Van Hoa
100. Friction Science and Technology, Peter J. Blau
101. Introduction to Plastics and Composites: Mechanical Properties
and Engineering Applications, Edward Miller
102. Practical Fracture Mechanics in Design, Alexander Blake
103. Pump Characteristics and Applications, Michael W. Volk
104. Optical Principles and Technology for Engineers, James E. Stewart
105. Optimizing the Shape of Mechanical Elements and Structures, A. A. Seireg
and Jorge Rodriguez
106. Kinematics and Dynamics of Machinery, Vladimír Stejskal
and Michael Valásek
107. Shaft Seals for Dynamic Applications, Les Horve
108. ReliabilityBased Mechanical Design, edited by Thomas A. Cruse
109. Mechanical Fastening, Joining, and Assembly, James A. Speck
110. Turbomachinery Fluid Dynamics and Heat Transfer, edited by Chunill Hah
111. HighVacuum Technology: A Practical Guide, Second Edition,
Revised and Expanded, Marsbed H. Hablanian
112. Geometric Dimensioning and Tolerancing: Workbook and Answerbook,
James D. Meadows
113. Handbook of Materials Selection for Engineering Applications, edited by
G. T. Murray
DK3139_series.qxd 1/20/05 11:17 AM Page 4
© 2005 by Taylor & Francis Group, LLC
114. Handbook of Thermoplastic Piping System Design, Thomas Sixsmith
and Reinhard Hanselka
115. Practical Guide to Finite Elements: A Solid Mechanics Approach,
Steven M. Lepi
116. Applied Computational Fluid Dynamics, edited by Vijay K. Garg
117. Fluid Sealing Technology, Heinz K. Muller and Bernard S. Nau
118. Friction and Lubrication in Mechanical Design, A. A. Seireg
119. Influence Functions and Matrices, Yuri A. Melnikov
120. Mechanical Analysis of Electronic Packaging Systems, Stephen A. McKeown
121. Couplings and Joints: Design, Selection, and Application, Second Edition,
Revised and Expanded, Jon R. Mancuso
122. Thermodynamics: Processes and Applications, Earl Logan, Jr.
123. Gear Noise and Vibration, J. Derek Smith
124. Practical Fluid Mechanics for Engineering Applications, John J. Bloomer
125. Handbook of Hydraulic Fluid Technology, edited by George E. Totten
126. Heat Exchanger Design Handbook, T. Kuppan
127. Designing for Product Sound Quality, Richard H. Lyon
128. Probability Applications in Mechanical Design, Franklin E. Fisher
and Joy R. Fisher
129. Nickel Alloys, edited by Ulrich Heubner
130. Rotating Machinery Vibration: Problem Analysis and Troubleshooting,
Maurice L. Adams, Jr.
131. Formulas for Dynamic Analysis, Ronald L. Huston and C. Q. Liu
132. Handbook of Machinery Dynamics, Lynn L. Faulkner and Earl Logan, Jr.
133. Rapid Prototyping Technology: Selection and Application,
Kenneth G. Cooper
134. Reciprocating Machinery Dynamics: Design and Analysis,
Abdulla S. Rangwala
135. Maintenance Excellence: Optimizing Equipment LifeCycle Decisions,
edited by John D. Campbell and Andrew K. S. Jardine
136. Practical Guide to Industrial Boiler Systems, Ralph L. Vandagriff
137. Lubrication Fundamentals: Second Edition, Revised and Expanded,
D. M. Pirro and A. A. Wessol
138. Mechanical Life Cycle Handbook: Good Environmental Design
and Manufacturing, edited by Mahendra S. Hundal
139. Micromachining of Engineering Materials, edited by Joseph McGeough
140. Control Strategies for Dynamic Systems: Design and Implementation,
John H. Lumkes, Jr.
141. Practical Guide to Pressure Vessel Manufacturing, Sunil Pullarcot
142. Nondestructive Evaluation: Theory, Techniques, and Applications,
edited by Peter J. Shull
143. Diesel Engine Engineering: Thermodynamics, Dynamics, Design,
and Control, Andrei Makartchouk
144. Handbook of Machine Tool Analysis, Ioan D. Marinescu, Constantin Ispas,
and Dan Boboc
DK3139_series.qxd 1/20/05 11:17 AM Page 5
© 2005 by Taylor & Francis Group, LLC
145. Implementing Concurrent Engineering in Small Companies,
Susan Carlson Skalak
146. Practical Guide to the Packaging of Electronics: Thermal and Mechanical
Design and Analysis, Ali Jamnia
147. Bearing Design in Machinery: Engineering Tribology and Lubrication,
Avraham Harnoy
148. Mechanical Reliability Improvement: Probability and Statistics
for Experimental Testing, R. E. Little
149. Industrial Boilers and Heat Recovery Steam Generators: Design,
Applications, and Calculations, V. Ganapathy
150. The CAD Guidebook: A Basic Manual for Understanding and Improving
ComputerAided Design, Stephen J. Schoonmaker
151. Industrial Noise Control and Acoustics, Randall F. Barron
152. Mechanical Properties of Engineered Materials, Wolé Soboyejo
153. Reliability Verification, Testing, and Analysis in Engineering Design,
Gary S. Wasserman
154. Fundamental Mechanics of Fluids: Third Edition, I. G. Currie
155. Intermediate Heat Transfer, KauFui Vincent Wong
156. HVAC Water Chillers and Cooling Towers: Fundamentals, Application,
and Operation, Herbert W. Stanford III
157. Gear Noise and Vibration: Second Edition, Revised and Expanded,
J. Derek Smith
158. Handbook of Turbomachinery: Second Edition, Revised and Expanded,
edited by Earl Logan, Jr. and Ramendra Roy
159. Piping and Pipeline Engineering: Design, Construction, Maintenance,
Integrity, and Repair, George A. Antaki
160. Turbomachinery: Design and Theory, Rama S. R. Gorla
and Aijaz Ahmed Khan
161. Target Costing: MarketDriven Product Design, M. Bradford Clifton,
Henry M. B. Bird, Robert E. Albano, and Wesley P. Townsend
162. Fluidized Bed Combustion, Simeon N. Oka
163. Theory of Dimensioning: An Introduction to Parameterizing Geometric
Models, Vijay Srinivasan
164. Handbook of Mechanical Alloy Design, edited by George E. Totten, Lin Xie,
and Kiyoshi Funatani
165. Structural Analysis of Polymeric Composite Materials, Mark E. Tuttle
166. Modeling and Simulation for Material Selection and Mechanical Design,
edited by George E. Totten, Lin Xie, and Kiyoshi Funatani
167. Handbook of Pneumatic Conveying Engineering, David Mills, Mark G. Jones,
and Vijay K. Agarwal
168. Clutches and Brakes: Design and Selection, Second Edition,
William C. Orthwein
169. Fundamentals of Fluid Film Lubrication: Second Edition,
Bernard J. Hamrock, Steven R. Schmid, and Bo O. Jacobson
170. Handbook of LeadFree Solder Technology for Microelectronic Assemblies,
edited by Karl J. Puttlitz and Kathleen A. Stalter
171. Vehicle Stability, Dean Karnopp
DK3139_series.qxd 1/20/05 11:17 AM Page 6
© 2005 by Taylor & Francis Group, LLC
172. Mechanical Wear Fundamentals and Testing: Second Edition,
Revised and Expanded, Raymond G. Bayer
173. Liquid Pipeline Hydraulics, E. Shashi Menon
174. Solid Fuels Combustion and Gasification, Marcio L. de SouzaSantos
175. Mechanical Tolerance Stackup and Analysis, Bryan R. Fischer
176. Engineering Design for Wear, Raymond G. Bayer
177. Vibrations of Shells and Plates: Third Edition, Revised and Expanded,
Werner Soedel
178. Refractories Handbook, edited by Charles A. Schacht
179. Practical Engineering Failure Analysis, Hani M. Tawancy,
Anwar UlHamid, and Nureddin M. Abbas
180. Mechanical Alloying and Milling, C. Suryanarayana
181. Mechanical Vibration: Analysis, Uncertainties, and Control,
Second Edition, Revised and Expanded, Haym Benaroya
182. Design of Automatic Machinery, Stephen J. Derby
183. Practical Fracture Mechanics in Design: Second Edition,
Revised and Expanded, Arun Shukla
184. Practical Guide to Designed Experiments, Paul D. Funkenbusch
185. Gigacycle Fatigue in Mechanical Practive, Claude Bathias
and Paul C. Paris
186. Selection of Engineering Materials and Adhesives, Lawrence W. Fisher
187. Boundary Methods: Elements, Contours, and Nodes, Subrata Mukherjee
and Yu Xie Mukherjee
DK3139_series.qxd 1/20/05 11:17 AM Page 7
© 2005 by Taylor & Francis Group, LLC
DK3139_title 1/20/05 11:12 AM Page 1
Boundary
Methods
Elements, Contours, and Nodes
Boca Raton London New York Singapore
A CRC title, part of the Taylor & Francis imprint, a member of the
Taylor & Francis Group, the academic division of T&F Informa plc.
Subrata Mukherjee
Cornell University
Ithaca, New York, U.S.A.
Yu Xie Mukherjee
Cornell University
Ithaca, New York, U.S.A.
© 2005 by Taylor & Francis Group, LLC
Published in 2005 by
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW
Boca Raton, FL 334872742
© 2005 by Taylor & Francis Group, LLC
CRC Press is an imprint of Taylor & Francis Group
No claim to original U.S. Government works
Printed in the United States of America on acidfree paper
10 9 8 7 6 5 4 3 2 1
International Standard Book Number10: 0824725999 (Hardcover)
International Standard Book Number13: 9780824725990 (Hardcover)
Library of Congress Card Number 2004063489
This book contains information obtained from authentic and highly regarded sources. Reprinted material is
quoted with permission, and sources are indicated. A wide variety of references are listed. Reasonable efforts
have been made to publish reliable data and information, but the author and the publisher cannot assume
responsibility for the validity of all materials or for the consequences of their use.
No part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic,
mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and
recording, or in any information storage or retrieval system, without written permission from the publishers.
For permission to photocopy or use material electronically from this work, please access www.copyright.com
(http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC) 222 Rosewood Drive,
Danvers, MA 01923, 9787508400. CCC is a notforprofit organization that provides licenses and registration
for a variety of users. For organizations that have been granted a photocopy license by the CCC, a separate
system of payment has been arranged.
Trademark Notice:
Product or corporate names may be trademarks or registered trademarks, and are used only
for identification and explanation without intent to infringe.
Library of Congress CataloginginPublication Data
Mukherjee, Subrata.
Boundary methods : elements, contours, and nodes / Subrata Mukherjee and Yu Mukherjee.
p. cm.  (Mechanical engineering ; 185)
ISBN 0824725999 (alk. paper)
1. Boundary element methods. I. Mukherjee, Yu. II. Title. III. Mechanical engineering
(Marcel Dekker, Inc.) ; 185.
TA347.B69M83 2005
621'.01'51535dc22 2004063489
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
Taylor & Francis Group
is the Academic Division of T&F Informa plc.
DK3139_discl Page 1 Wednesday, January 19, 2005 9:05 AM
© 2005 by Taylor & Francis Group, LLC
iii
To our boys
Anondo and Alok
and
To Yu’s teacher, Professor Zhicheng Xie of Tsinghua University,
a distinguished scholar
who has dedicated himself to China.
© 2005 by Taylor & Francis Group, LLC
v
PREFACE
The general subject area of concern to this book is computational science and
engineering, with applications in potential theory and in solid mechanics (linear
elasticity). This ﬁeld has undergone a revolution during the past several decades
along with the exponential growth of computational power and memory. Problems
that were too large for main frame computers 15 or 20 years ago can now be
routinely solved on desktop personal computers.
There are several popular computational methods for solving problems in po
tential theory and linear elasticity. The most popular, versatile and most commonly
used is the ﬁnite element method (FEM). Many hundreds of books already exist
on the subject and new books get published frequently on a regular basis. Another
popular method is the boundary element method (BEM). Compared to the FEM,
we view the BEM as a niche method, in that it is particularly well suited, from the
point of view of accuracy as well as computational eﬃciency, for linear problems.
The principal advantage of the BEM, relative to the FEM, is its dimensionality
advantage. The FEM is a domain method that requires discretization of the entire
domain of a body while the BEM, for linear problems, only requires discretization
of its bounding surface.
The process of discretization (or meshing) of a threedimensional (3D) object of
complex shape is a popular research area in computational geometry. Even though
great strides have been made in recent years, meshing, for many applications, still
remains an arduous task. During the past decade, meshfree (also called mesh
less) methods have become a popular research area in computational mechanics.
The main purpose here is to substantially simplify the task of meshing of an object.
Advantages of meshfree methods become more pronounced, for example, for prob
lems involving optimal shape design or adaptive meshing, since many remeshings
must be typically carried out for such problems. One primary focus of this book
is a marriage of these two ideas, i.e. a discussion of a boundarybased meshfree
method  the boundary node method (BNM)  which combines the dimensionality
advantage of the BEM with the ease of discretization of meshfree methods.
Following an introductory chapter, this book consists of three parts related to
the boundary element, boundary contour and boundary node methods. The ﬁrst
part is short, in order not to duplicate information on the BEM that is already
available in many books on the subject. Only some novel topics related to the
BEM are presented here. The second part is concerned with the boundary contour
method (BCM). This method is a novel variant of the BEM in that it further reduces
the dimensionality of a problem. Only onedimensional line integrals need to be
numerically computed when solving threedimensional problems in linear elasticity
by the BCM. The third part is concerned with the boundary node method (BNM).
The BNM combines the BEM with moving leastsquares (MLS) approximants,
thus producing a meshfree boundaryonly method. In addition to the solution of
3D problems, Part II of the book on the BCM presents shape sensitivity analysis,
shape optimization, and error estimation and adaptivity; while Part III on the BNM
includes error analysis and adaptivity.
© 2005 by Taylor & Francis Group, LLC
vi
This book is written in the style of a research monograph. Each topic is clearly
introduced and developed. Numerical results for selected problems appear through
out the book, as do references to related work (research publications and books).
This book should be of great interest to graduate students, researchers and
practicing engineers in the ﬁeld of computational mechanics; and to others inter
ested in the general areas of computational mathematics, science and engineering.
It should also be of value to advanced undergraduate students who are interested
in this ﬁeld.
We wish to thank a number of people and organizations who have contributed
in various ways to making this book possible. Two of Subrata’s former graduate
students, Glaucio Paulino and Mandar Chati, as well as Yu’s associate Xiaolan Shi,
have made very signiﬁcant contributions to the research that led to this book. Sin
cere thanks are expressed to Subrata’s former graduate students Govind Menon and
Ramesh Gowrishankar, to one of his present graduate students, Srinivas Telukunta,
and to Vasanth Kothnur, for their contributions to the BNM. Earlin Lutz, Anan
tharaman Nagarajan and AnhVu Phan have signiﬁcantly contributed to the early
development of the BCM; while Subrata’s justgraduated student Zhongping Bao
has made excellent contributions to the research on microelectromechanical sys
tems (MEMS) by the BEM. Sincere thanks are expressed to our dear friend Ashim
Datta for his help and encouragement throughout the writing of this book.
Much of the research presented here has been ﬁnancially supported by the Na
tional Science Foundation and Ford Motor Company, and this support is gratefully
acknowledged. Most of the ﬁgures and tables in this book have been published
before in journals. They were all originally created by the authors of this book,
together with their coauthors. These items have been printed here by permission of
the original copyright owner (i.e. the publishers of the appropriate journal), and this
permission is very much appreciated. The original source has been acknowledged
in this book at the end of the caption for each item.
Subrata and Yu Mukherjee
Ithaca, New York
October 2004
© 2005 by Taylor & Francis Group, LLC
Contents
Preface v
INTRODUCTION TO BOUNDARY METHODS xiii
I SELECTED TOPICS IN BOUNDARY ELEMENT
METHODS 1
1 BOUNDARY INTEGRAL EQUATIONS 3
1.1 Potential Theory in Three Dimensions . . . . . . . . . . . . . . . 3
1.1.1 Singular Integral Equations . . . . . . . . . . . . . . . . . 3
1.1.2 Hypersingular Integral Equations . . . . . . . . . . . . . . 5
1.2 Linear Elasticity in Three Dimensions . . . . . . . . . . . . . . . 6
1.2.1 Singular Integral Equations . . . . . . . . . . . . . . . . . 6
1.2.2 Hypersingular Integral Equations . . . . . . . . . . . . . . 8
1.3 Nearly Singular Integrals in Linear Elasticity . . . . . . . . . . . 12
1.3.1 Displacements at Internal Points Close to the Boundary . 12
1.3.2 Stresses at Internal Points Close to the Boundary . . . . . 13
1.4 Finite Parts of Hypersingular Equations . . . . . . . . . . . . . . 14
1.4.1 Finite Part of a Hypersingular Integral Collocated at an
Irregular Boundary Point . . . . . . . . . . . . . . . . . . 14
1.4.2 Gradient BIE for 3D Laplace’s Equation . . . . . . . . . 17
1.4.3 Stress BIE for 3D Elasticity . . . . . . . . . . . . . . . . 19
1.4.4 Solution Strategy for a HBIE Collocated at an Irregular
Boundary Point . . . . . . . . . . . . . . . . . . . . . . . . 20
2 ERROR ESTIMATION 23
2.1 Linear Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 Iterated HBIE and Error Estimation . . . . . . . . . . . . . . . . 25
2.2.1 Problem 1 : Displacement Boundary Conditions . . . . . 25
2.2.2 Problem 2 : Traction Boundary Conditions . . . . . . . . 28
2.2.3 Problem 3 : Mixed Boundary Conditions . . . . . . . . . 30
2.3 ElementBased Error Indicators . . . . . . . . . . . . . . . . . . . 32
2.4 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . 33
vii
© 2005 by Taylor & Francis Group, LLC
viii CONTENTS
2.4.1 Example 1: Lam´e’s Problem of a ThickWalled Cylinder
under Internal Pressure . . . . . . . . . . . . . . . . . . . 34
2.4.2 Example 2: Kirsch’s Problem of an Inﬁnite Plate with a
Circular Cutout . . . . . . . . . . . . . . . . . . . . . . . 36
3 THIN FEATURES 39
3.1 Exterior BIE for Potential Theory: MEMS . . . . . . . . . . . . 39
3.1.1 Introduction to MEMS . . . . . . . . . . . . . . . . . . . . 39
3.1.2 Electric Field BIEs in a Simply Connected Body . . . . . 41
3.1.3 BIES in Inﬁnite Region Containing Two Thin Conducting
Plates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.1.4 Singular and Nearly Singular Integrals . . . . . . . . . . . 46
3.1.5 Numerical Results . . . . . . . . . . . . . . . . . . . . . . 49
3.1.6 The Model Problem  a Parallel Plate Capacitor . . . . . 50
3.2 BIE for Elasticity: Cracks and Thin Shells . . . . . . . . . . . . 54
3.2.1 BIES in LEFM . . . . . . . . . . . . . . . . . . . . . . . . 54
3.2.2 Numerical Implementation of BIES in LEFM . . . . . . . 60
3.2.3 Some Comments on BIEs in LEFM . . . . . . . . . . . . . 61
3.2.4 BIEs for Thin Shells . . . . . . . . . . . . . . . . . . . . . 62
II THE BOUNDARY CONTOUR METHOD 65
4 LINEAR ELASTICITY 67
4.1 Surface and Boundary Contour Equations . . . . . . . . . . . . . 67
4.1.1 Basic Equations . . . . . . . . . . . . . . . . . . . . . . . 67
4.1.2 Interpolation Functions . . . . . . . . . . . . . . . . . . . 68
4.1.3 Boundary Elements . . . . . . . . . . . . . . . . . . . . . 71
4.1.4 Vector Potentials . . . . . . . . . . . . . . . . . . . . . . . 73
4.1.5 Final BCM Equations . . . . . . . . . . . . . . . . . . . . 74
4.1.6 Global Equations and Unknowns . . . . . . . . . . . . . . 76
4.1.7 Surface Displacements, Stresses, and Curvatures . . . . . 76
4.2 Hypersingular Boundary Integral Equations . . . . . . . . . . . . 78
4.2.1 Regularized Hypersingular BIE . . . . . . . . . . . . . . . 78
4.2.2 Regularized Hypersingular BCE . . . . . . . . . . . . . . 78
4.2.3 Collocation of the HBCE at an Irregular Surface Point . . 80
4.3 Internal Displacements and Stresses . . . . . . . . . . . . . . . . 82
4.3.1 Internal Displacements . . . . . . . . . . . . . . . . . . . . 82
4.3.2 Displacements at Internal Points Close to the Bounding
Surface . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.3.3 Internal Stresses . . . . . . . . . . . . . . . . . . . . . . . 83
4.3.4 Stresses at Internal Points Close to the Bounding Surface 84
4.4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.4.1 Surface Displacements from the BCM and the
HBCM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
© 2005 by Taylor & Francis Group, LLC
CONTENTS ix
4.4.2 Surface Stresses . . . . . . . . . . . . . . . . . . . . . . . . 87
4.4.3 Internal Stresses Relatively Far from the Bounding Surface 90
4.4.4 Internal Stresses Very Close to the Bounding Surface . . . 90
5 SHAPE SENSITIVITY ANALYSIS 93
5.1 Sensitivities of Boundary Variables . . . . . . . . . . . . . . . . . 93
5.1.1 Sensitivity of the BIE . . . . . . . . . . . . . . . . . . . . 93
5.1.2 The Integral I
k
. . . . . . . . . . . . . . . . . . . . . . . . 94
5.1.3 The Integral J
k
. . . . . . . . . . . . . . . . . . . . . . . . 96
5.1.4 The BCM Sensitivity Equation . . . . . . . . . . . . . . . 98
5.2 Sensitivities of Surface Stresses . . . . . . . . . . . . . . . . . . . 99
5.2.1 Method One . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.2.2 Method Two . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.2.3 Method Three . . . . . . . . . . . . . . . . . . . . . . . . 100
5.2.4 Method Four . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.3 Sensitivities of Variables at Internal Points . . . . . . . . . . . . 101
5.3.1 Sensitivities of Displacements . . . . . . . . . . . . . . . . 101
5.3.2 Sensitivities of Displacement Gradients and Stresses . . . 103
5.4 Numerical Results: Hollow Sphere . . . . . . . . . . . . . . . . . 106
5.4.1 Sensitivities on Sphere Surface . . . . . . . . . . . . . . . 107
5.4.2 Sensitivities at Internal Points . . . . . . . . . . . . . . . 108
5.5 Numerical Results: Block with a Hole . . . . . . . . . . . . . . . 110
5.5.1 Geometry and Mesh . . . . . . . . . . . . . . . . . . . . . 110
5.5.2 Internal Stresses . . . . . . . . . . . . . . . . . . . . . . . 112
5.5.3 Sensitivities of Internal Stresses . . . . . . . . . . . . . . . 112
6 SHAPE OPTIMIZATION 115
6.1 Shape Optimization Problems . . . . . . . . . . . . . . . . . . . . 115
6.2 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.2.1 Shape Optimization of a Fillet . . . . . . . . . . . . . . . 116
6.2.2 Optimal Shapes of Ellipsoidal Cavities Inside Cubes . . . 118
6.2.3 Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
7 ERROR ESTIMATION AND ADAPTIVITY 125
7.1 Hypersingular Residuals as Local Error Estimators . . . . . . . . 125
7.2 Adaptive Meshing Strategy . . . . . . . . . . . . . . . . . . . . . 126
7.3 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . 127
7.3.1 Example One  Short Clamped Cylinder under Tension . 127
7.3.2 Example Two  the Lam´e Problem for a Hollow Cylinder 130
III THE BOUNDARY NODE METHOD 133
8 SURFACE APPROXIMANTS 135
8.1 Moving Least Squares (MLS) Approximants . . . . . . . . . . . 135
8.2 Surface Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . 139
© 2005 by Taylor & Francis Group, LLC
x CONTENTS
8.3 Weight Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
8.4 Use of Cartesian Coordinates . . . . . . . . . . . . . . . . . . . . 142
8.4.1 Hermite Type Approximation . . . . . . . . . . . . . . . . 142
8.4.2 Variable Basis Approximation . . . . . . . . . . . . . . . . 143
9 POTENTIAL THEORY AND ELASTICITY 151
9.1 Potential Theory in Three Dimensions . . . . . . . . . . . . . . . 151
9.1.1 BNM: Coupling of BIE with MLS Approximants . . . . . 151
9.1.2 HBNM: Coupling of HBIE with MLS Approximants . . . 155
9.1.3 Numerical Results for Dirichlet Problems on a Sphere . . 156
9.2 Linear Elasticity in Three Dimensions . . . . . . . . . . . . . . . 165
9.2.1 BNM: Coupling of BIE with MLS Approximants . . . . . 165
9.2.2 HBNM: Coupling of HBIE with MLS Approximants . . . 167
9.2.3 Numerical Results . . . . . . . . . . . . . . . . . . . . . . 168
10 ADAPTIVITY FOR 3D POTENTIAL THEORY 175
10.1 Hypersingular and Singular Residuals . . . . . . . . . . . . . . . 175
10.1.1 The Hypersingular Residual . . . . . . . . . . . . . . . . . 175
10.1.2 The Singular Residual . . . . . . . . . . . . . . . . . . . . 176
10.2 Error Estimation and Adaptive Strategy . . . . . . . . . . . . . . 177
10.2.1 Local Residuals and Errors  Hypersingular Residual Ap
proach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
10.2.2 Local Residuals and Errors  Singular Residual Approach 178
10.2.3 Cell Reﬁnement Criterion . . . . . . . . . . . . . . . . . . 179
10.2.4 Global Error Estimation and Stopping Criterion . . . . . 179
10.3 Progressively Adaptive Solutions: Cube
Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
10.3.1 Exact Solution . . . . . . . . . . . . . . . . . . . . . . . . 181
10.3.2 Initial Cell Conﬁguration # 1 (54 Surface Cells) . . . . . 181
10.3.3 Initial Cell Conﬁguration # 2 (96 Surface Cells) . . . . . 182
10.4 OneStep Adaptive Cell Reﬁnement . . . . . . . . . . . . . . . . 188
10.4.1 Initial Cell Conﬁguration # 1 (54 Surface Cells) . . . . . 190
10.4.2 Initial Cell Conﬁguration # 2 (96 Surface Cells) . . . . . 191
11 ADAPTIVITY FOR 3D LINEAR ELASTICITY 193
11.1 Hypersingular and Singular Residuals . . . . . . . . . . . . . . . 193
11.1.1 The Hypersingular Residual . . . . . . . . . . . . . . . . . 193
11.1.2 The Singular Residual . . . . . . . . . . . . . . . . . . . . 194
11.2 Error Estimation and Adaptive Strategy . . . . . . . . . . . . . . 194
11.2.1 Local Residuals and Errors  Hypersingular Residual Ap
proach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
11.2.2 Local Residuals and Errors  Singular Residual Approach 195
11.2.3 Cell Reﬁnement Global Error Estimation and Stopping
Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
11.3 Progressively Adaptive Solutions: Pulling a Rod . . . . . . . . . 195
© 2005 by Taylor & Francis Group, LLC
CONTENTS xi
11.3.1 Initial Cell Conﬁguration . . . . . . . . . . . . . . . . . . 197
11.3.2 Adaptivity Results . . . . . . . . . . . . . . . . . . . . . . 197
11.4 OneStep Adaptive Cell Reﬁnement . . . . . . . . . . . . . . . . 198
Bibliography 203
© 2005 by Taylor & Francis Group, LLC
INTRODUCTION TO
BOUNDARY METHODS
This chapter provides a brief introduction to various topics that are of interest
in this book.
Boundary Element Method
Boundary integral equations (BIE), and the boundary element method (BEM),
based on BIEs, are mature methods for numerical analysis of a large variety of
problems in science and engineering. The standard BEM for linear problems
has the wellknown dimensionality advantage in that only the twodimensional
(2D) bounding surface of a threedimensional (3D) body needs to be meshed
when this method is used. Examples of books on the subject, published dur
ing the last 15 years, are Banerjee [4], Becker [9], Bonnet [14], Brebbia and
Dominguez [16], Chandra and Mukherjee [22], Gaul et al. [47], Hartmann [62],
Kane [68] and Par´ıs and Ca˜ nas [121]. BEM topics of interest in this book are
ﬁnite parts (FP) in Chapter 1, error estimation in Chapter 2 and thin features
(cracks and thin objects) in Chapter 3.
Hypersingular Boundary Integral Equations
Hypersingular boundary integral equations (HBIEs) are derived from a diﬀer
entiated version of the usual boundary integral equations (BIEs). HBIEs have
diverse important applications and are the subject of considerable current re
search (see, for example, Krishnasamy et al. [76], Tanaka et al. [162], Paulino
[122] and Chen and Hong [30] for recent surveys of the ﬁeld). HBIEs, for exam
ple, have been employed for the evaluation of boundary stresses (e.g. Guiggiani
[60], Wilde and Aliabadi [173], Zhao and Lan [185], Chati and Mukherjee [24]),
in wave scattering (e.g. Krishnasamy et al. [75]), in fracture mechanics (e.g.
Cruse [38], Gray et al. [54], Lutz et al. [89], Paulino [122], Gray and Paulino
[58], Mukherjee et al. [110]), to obtain symmetric Galerkin boundary element
formulations (e.g. Bonnet [14], Gray et al. [55], Gray and Paulino ([56], [57]), to
xiii
© 2005 by Taylor & Francis Group, LLC
xiv INTRODUCTION TO BOUNDARY METHODS
evaluate nearly singular integrals (Mukherjee et al. [104]), to obtain the hyper
singular boundary contour method (Phan et al. [131], Mukherjee and Mukherjee
[99]), to obtain the hypersingular boundary node method (Chati et al. [27]), and
for error analysis (Paulino et al. [123], Menon [95], Menon et al. [96], Chati et
al. [27], Paulino and Gray [125]) and adaptivity [28].
An elegant approach of regularizing singular and hypersingular integrals, us
ing simple solutions, was ﬁrst proposed by Rudolphi [143]. Several researchers
have used this idea to regularize hypersingular integrals before collocating an
HBIE at a regular boundary point. Examples are Cruse and Richardson [39],
Lutz et al. [89], Poon et al. [138], Mukherjee et al. [110] and Mukherjee [106].
The relationship between ﬁnite parts of strongly singular and hypersingular in
tegrals, and the HBIE, is discussed in [168], [101] and [102]. A lively debate (e.g.
[92], [39]), on smoothness requirements on boundary variables for collocating
an HBIE on the boundary of a body, has apparently been concluded recently
[93]. An alternative way of satisfying this smoothness requirement is the use of
the hypersingular boundary node method (HBNM).
MeshFree Methods
Meshfree (also called meshless) methods [82], that only require points rather
than elements to be speciﬁed in the physical domain, have tremendous potential
advantages over methods such as the ﬁnite element method (FEM) that require
discretization of a body into elements.
The idea of moving least squares (MLS) interpolants, for curve and surface
ﬁtting, is described in a book by Lancaster and Salkauskas [78]. Nayroles et
al. [117] proposed a coupling of MLS interpolants with Galerkin procedures in
order to solve boundary value problems. They called their method the diﬀuse
element method (DEM) and applied it to twodimensional (2D) problems in
potential theory and linear elasticity.
During the relatively short span of less than a decade, great progress has
been made in solid mechanics applications of meshfree methods. Meshfree
methods proposed to date include the elementfree Galerkin (EFG) method
[10, 11, 12, 13, 67, 174, 175, 176, 108], the reproducingkernel particle method
(RKPM) [83, 84], h−p clouds [42, 43, 120], the meshless local PetrovGalerkin
(MLPG) approach [3], the local boundary integral equation (LBIE) method
[152, 188], the meshless regular local boundary integral equation (MRLBIE)
method [189], the natural element method (NEM) [158, 160], the general
ized ﬁnite element method (GFEM) [157], the extended ﬁnite element method
(XFEM) [97, 41, 159], the method of ﬁnite spheres (MFS) [40], the ﬁnite
cloud method (FCM) [2], the boundary cloud method (BCLM) [79, 80], the
boundary point interpolation method (BPIM) [82], the boundaryonly radial
basis function method (BRBFM) [32] and the boundary node method (BNM)
[107, 72, 25, 26, 27, 28, 52].
© 2005 by Taylor & Francis Group, LLC
xv
Boundary Node Method
S. Mukherjee, together with his research collaborators, has recently pioneered
a new computational approach called the boundary node method (BNM) [26,
25, 27, 28, 72, 107]. Other examples of boundarybased meshless methods are
the boundary cloud method (BCLM) [79, 80], the boundary point interpolation
method (BPIM) [82], the boundary only radial basis function method (BRBFM)
[32] and the local BIE (LBIE) [188] approach. The LBIE, however, is not a
boundary method since it requires evaluation of integrals over certain surfaces
(called L
s
in [188]) that can be regarded as “closure surfaces” of boundary
elements.
The BNM is a combination of the MLS interpolation scheme and the stan
dard boundary integral equation (BIE) method. The method divorces the tra
ditional coupling between spatial discretization (meshing) and interpolation (as
commonly practiced in the FEM or in the BEM). Instead, a “diﬀuse” interpo
lation, based on MLS interpolants, is used to represent the unknown functions;
and surface cells, with a very ﬂexible structure (e.g. any cell can be arbitrarily
subdivided without aﬀecting its neighbors [27]) are used for integration. Thus,
the BNM retains the meshless attribute of the EFG method and the dimen
sionality advantage of the BEM. As a consequence, the BNM only requires the
speciﬁcation of points on the 2D bounding surface of a 3D body (including
crack faces in fracture mechanics problems), together with surface cells for in
tegration, thereby practically eliminating the meshing problem (see Figures i
and ii). The required cell structure is analogous to (but not the same as) a
tiling [139]. The only requirements are that the intersection of any two surface
cells is the null set and that the union of all the cells is the bounding surface of
the body. In contrast, the FEM needs volume meshing, the BEM needs surface
meshing, and the EFG needs points throughout the domain of a body.
It is important to point out another important advantage of MLS inter
polants. They can be easily designed to be suﬃciently smooth to suit a given
purpose, e.g. they can be made C
1
or higher [10] in order to collocate the HBNM
at a point on the boundary of a body.
The BNM is described in Chapters 8 and 9 of this book.
Figure i: BNM with nodes and cells
(from [28])
Figure ii: BEM with nodes and elements
(from [28])
© 2005 by Taylor & Francis Group, LLC
xvi INTRODUCTION TO BOUNDARY METHODS
Boundary Contour Method
The Method
The usual boundary element method (BEM), for threedimensional (3D) lin
ear elasticity, requires numerical evaluations of surface integrals on boundary
elements on the surface of a body (see, for example, [98]). [115] (for 2D linear
elasticity) and [116] (for 3D linear elasticity) have recently proposed a novel
approach, called the boundary contour method (BCM), that achieves a further
reduction in dimension! The BCM, for 3D linear elasticity problems, only re
quires numerical evaluation of line integrals over the closed bounding contours
of the usual (surface) boundary elements.
The central idea of the BCM is the exploitation of the divergencefree prop
erty of the usual BEM integrand and a very useful application of Stokes’ the
orem, to analytically convert surface integrals on boundary elements to line
integrals on closed contours that bound these elements. [88] ﬁrst proposed an
application of this idea for the Laplace equation and Nagarajan et al. gen
eralized this idea to linear elasticity. Numerical results for twodimensional
(2D) problems, with linear boundary elements, are presented in [115], while
results with quadratic boundary elements appear in [129]. Threedimensional
elasticity problems, with quadratic boundary elements, are the subject of [116]
and [109]. Hypersingular boundary contour formulations, for twodimensional
[131] and threedimensional [99] linear elasticity, have been proposed recently.
A symmetric Galerkin BCM for 2D linear elasticity appears in [119]. Recent
work on the BCM is available in [31, 134, 135, 136, 186].
The BCM is described in Chapter 4 of this book.
Shape Sensitivity Analysis with the BCM and the HBCM
Design sensitivity coeﬃcients (DSCs), which are deﬁned as rates of change of
physical response quantities with respect to changes in design variables, are
useful for various applications such as in judging the robustness of a given
design, in reliability analysis and in solving inverse and design optimization
problems. There are three methods for design sensitivity analysis (e.g. [63]),
namely, the ﬁnite diﬀerence approach (FDA), the adjoint structure approach
(ASA) and the direct diﬀerentiation approach (DDA). The DDA is of interest
in this work.
The goal of obtaining BCM sensitivity equations can be achieved in two
equivalent ways. In the 2D work by [130], design sensitivities are obtained
by ﬁrst converting the discretized BIEs into their boundary contour version,
and then applying the DDA (using the concept of the material derivative) to
this BCM version. This approach, while relatively straightforward in principle,
becomes extremely algebraically intensive for 3D elasticity problems. [100]
oﬀers a novel alternative derivation, using the opposite process, in which the
DDA is ﬁrst applied to the regularized BIE and then the resulting equations
© 2005 by Taylor & Francis Group, LLC
xvii
are converted to their boundary contour version. It is important to point out
that this process of converting the sensitivity BIE into a BCM form is quite
challenging. This new derivation, for sensitivities of surface variables [100], as
well as for internal variables [103], for 3D elasticity problems, is presented in
Chapter 5 of this book. The reader is referred to [133] for a corresponding
derivation for 2D elasticity
Shape Optimization with the BCM
Shape optimization refers to the optimal design of the shape of structural com
ponents and is of great importance in mechanical engineering design. A typical
gradientbased shape optimization procedure is an iterative process in which
iterative improvements are carried out over successive designs until an optimal
design is accepted. A domainbased method such as the ﬁnite element method
(FEM) typically requires discretization of the entire domain of a body many
times during this iterative process. The BEM, however, only requires surface
discretization, so that mesh generation and remeshing procedures can be carried
out much more easily for the BEM than for the FEM. Also, surface stresses are
typically obtained very accurately in the BEM. As a result, the BEM has been
a popular method for shape optimization in linear mechanics. Some examples
are references [33], [145], [178], [144], [169], [177], [161] and the book [184].
In addition to having the same meshing advantages as the usual BEM, the
BCM, as explained above, oﬀers a further reduction in dimension. Also, surface
stresses can be obtained very easily and accurately by the BCM without the
need for additional shape function diﬀerentiation as is commonly required with
the BEM. These properties make the BCM very attractive as the computational
engine for stress analysis for use in shape optimization. Shape optimization in
2D linear elasticity, with the BCM, has been presented by [132]. The corre
sponding 3D problem is presented in [150] and is discussed in Chapter 6.
Error Estimation and Adaptivity
A particular strength of the ﬁnite element method (FEM) is the welldeveloped
theory of error estimation, and its use in adaptive methods (see, for example,
Ciarlet [34], Eriksson et al. [44]). In contrast, error estimation in the boundary
element method (BEM) is a subject that has attracted attention mainly over
the past decade, and much work remains to be done. For recent surveys on
error estimation and adaptivity in the BEM, see Sloan [155], Kita and Kamiya
[70], Liapis [81] and Paulino et al. [124].
Many error estimators in the BEM are essentially heuristic and, unlike for
the FEM, theoretical work in this ﬁeld has been quite limited. Rank [140]
proposed error indicators and an adaptive algorithm for the BEM using tech
niques similar to those used in the FEM. Most notable is the work of Yu and
Wendland [171, 172, 181, 182], who have presented local error estimates based
© 2005 by Taylor & Francis Group, LLC
xviii INTRODUCTION TO BOUNDARY METHODS
on a linear errorresidual relation that is very eﬀective in the FEM. More re
cently, Carstensen et al. [18, 21, 19, 20] have presented error estimates for
the BEM analogous to the approach of Eriksson [44] for the FEM. There are
numerous stumbling blocks in the development of a satisfactory theoretical
analysis of a generic boundary value problem (BVP). First, theoretical analy
ses are easiest for Galerkin schemes, but most engineering codes, to date, use
collocationbased methods (see, for example, Banerjee [4]). Though one can
view collocation schemes as variants of PetrovGalerkin methods, and, in fact,
numerous theoretical analyses exist for collocation methods (see, for example,
references in [155]), the mathematical analysis for this class of problems is
diﬃcult. Theoretical analyses for mixed boundary conditions are limited and
involved (Wendland et al. [170]) and the presence of corners and cracks has
been a source of challenging problems for many years (Sloan [155], Costabel
and Stephan [35], Costabel et al. [36]). Of course, problems with corners and
mixed boundary conditions are the ones of most practical interest, and for such
situations one has to rely mostly on numerical experiments.
During the past few years, there has been a marked interest, among mathe
maticians in the ﬁeld, in extending analyses for the BEM with singular integrals
to hypersingular integrals ([21, 19, 156, 45]. For instance, Feistauer et al. [45]
have studied the solution of the exterior Neumann problem for the Helmholtz
equation formulated as an HBIE. Their paper contains a rigorous analysis of
hypersingular integral equations and addresses the problem of noncompatibil
ity of the residual norm, where additional hypotheses are needed to design a
practical error estimate. These authors use residuals to estimate the error,
but they do not use the BIE and the HBIE simultaneously. Finally, Goldberg
and Bowman [51] have used superconvergence of the Sloan iterate [153, 154] to
show the asymptotic equivalence of the error and the residual. They have used
Galerkin methods, an iteration scheme that uses the same integral equation for
the approximation and for the iterates, and usual residuals in their work.
Paulino [122] and Paulino et al. [123] ﬁrst proposed the idea of obtaining
a hypersingular residual by substituting the BEM solution of a problem into
the hypersingular BEM (HBEM) for the same problem; and then using this
residual as an element error estimator in the BEM. It has been proved that
([95], [96], [127]), under certain conditions, this residual is related to a measure
of the local error on a boundary element, and has been used to postulate local
error estimates on that element. This idea has been applied to the collocation
BEM ([123], [96], [127]) and to the symmetric Galerkin BEM ([125]). Recently,
residuals have been obtained in the context of the BNM [28] and used to obtain
local error estimates (at the element level) and then to drive an hadaptive
mesh reﬁnement process. An analogous approach for error estimation and h
adaptivity, in the context of the BCM, is described in [111]. Ref. [91] has a
bibliography of work on mesh generation and reﬁnement up to 1993.
Error analysis with the BEM is presented in Chapter 2, while error analysis
and adaptivity in the context of the BCM and the BNM are discussed in Chapter
7, and Chapters 10, 11, respectively, of this book.
© 2005 by Taylor & Francis Group, LLC
Part I
SELECTED TOPICS IN
BOUNDARY ELEMENT
METHODS
1
© 2005 by Taylor & Francis Group, LLC
Chapter 1
BOUNDARY INTEGRAL
EQUATIONS
Integral equations, usual as well as hypersingular, for internal and boundary
points, for potential theory in three dimensions, are ﬁrst presented in this chap
ter. This is followed by their linear elasticity counterparts. The evaluation of
ﬁnite parts (FPs) of some of these equations, when the source point is an irreg
ular boundary point (situated at a corner on a onedimensional plane curve or
at a corner or edge on a twodimensional surface), is described next.
1.1 Potential Theory in Three Dimensions
The starting point is Laplace’s equation in three dimensions (3D) governing a
potential function u(x
1
, x
2
, x
3
) ∈ B, where B is a bounded region (also called
the body):
∇
2
u(x
1
, x
2
, x
3
) ≡
∂
2
u
∂x
2
1
+
∂
2
u
∂x
2
2
+
∂
2
u
∂x
2
3
= 0 (1.1)
along with prescribed boundary conditions on the bounding surface ∂B of B.
1.1.1 Singular Integral Equations
Referring to Figure 1.1, let ξ and η be (internal) source and ﬁeld points ∈ B
and x and y be (boundary) source and ﬁeld points ∈ ∂B, respectively. (Source
and ﬁeld points are also referred to as p and q (for internal points) and as P
and Q (for boundary points), respectively, in this book).
The wellknown integral representation for (1.1), at an internal point ξ ∈ B,
is:
u(ξ) =
∂B
[G(ξ, y)τ(y) −F(ξ, y)u(y)]dS(y) (1.2)
3
© 2005 by Taylor & Francis Group, LLC
4 CHAPTER 1. BOUNDARY INTEGRAL EQUATIONS
B
x(P)
y(Q)
ξ(p)
r
(
ξ
,
y
)
η(q)
n(y)
n(x)
∂B
Figure 1.1: Notation used in integral equations (from [6])
An inﬁnitesimal surface area on ∂B is dS = dSn, where n is the unit outward
normal to ∂B at a point on it and τ = ∂u/∂n. The kernels are written in terms
of source and ﬁeld points ξ ∈ B and y ∈ ∂B. These are :
G(ξ, y) =
1
4πr(ξ, y)
(1.3)
F(ξ, y) =
∂G(ξ, y)
∂n(y)
=
(ξ
i
−y
i
)n
i
(y)
4πr
3
(ξ, y)
(1.4)
in terms of r(ξ, y), the Euclidean distance between the source and ﬁeld points
ξ and y. Unless speciﬁed otherwise, the range of indices in these and all other
equations in this chapter is 1,2,3.
An alternative form of equation (1.2) is:
u(ξ) =
∂B
[G(ξ, y)u
,k
(y) −H
k
(ξ, y)u(y)]e
k
· dS(y) (1.5)
where e
k
, k = 1, 2, 3, are the usual Cartesian unit vectors, e
k
· dS(y) =
n
k
(y)dS(y), and:
H
k
(ξ, y) =
(ξ
k
−y
k
)
4πr
3
(ξ, y)
(1.6)
The boundary integral equation (BIE) corresponding to (1.2) is obtained by
taking the limit ξ → x. A regularized form of the resulting equation is:
0 =
∂B
[G(x, y)τ(y) −F(x, y){u(y) −u(x)}]dS(y) (1.7)
© 2005 by Taylor & Francis Group, LLC
1.1. POTENTIAL THEORY IN THREE DIMENSIONS 5
with an alternate form (from (1.5)):
0 =
∂B
[G(x, y)u
,k
(y) −H
k
(x, y){u(y) −u(x)}]e
k
· dS(y) (1.8)
1.1.2 Hypersingular Integral Equations
Equation (1.2) can be diﬀerentiated at an internal source point ξ to obtain the
gradient
∂u
∂ξ
m
of the potential u. The result is:
∂u(ξ)
∂ξ
m
=
∂B
∂G(ξ, y)
∂ξ
m
τ(y) −
∂F(ξ, y)
∂ξ
m
u(y)
dS(y) (1.9)
An interesting situation arises when one takes the limit ξ → x (x can even
be an irregular point on ∂B but one must have u(y) ∈ C
1,α
at y = x) in
equation (1.9). As discussed in detail in Section 1.4.2, one obtains:
∂u(x)
∂x
m
=
∂B
=
∂G(x, y)
∂x
m
τ(y) −
∂F(x, y)
∂x
m
u(y)
dS(y) (1.10)
where the symbol
= denotes the ﬁnite part (FP) of the integral. Equation (1.10)
is best regularized before computations are carried out. The regularized version
given below is applicable even at an irregular boundary point x provided that
u(y) ∈ C
1,α
at y = x. This is:
0 =
∂B
∂G(x, y)
∂x
m
u
,p
(y) −u
,p
(x)
n
p
(y)dS(y)
−
∂B
∂F(x, y)
∂x
m
u(y) −u(x) −u
,p
(x)(y
p
−x
p
)
dS(y) (1.11)
An alternative form of (1.11), valid at a regular boundary point x, [76] is:
0 =
∂B
∂G(x, y)
∂x
m
τ(y) −τ(x)
dS(y)
− u
,k
(x)
B
∂G(x, y)
∂x
m
n
k
(y) −n
k
(x)
dS(y)
−
∂B
∂F(x, y)
∂x
m
u(y) −u(x) −u
,p
(x)(y
p
−x
p
)
dS(y) (1.12)
Carrying out the inner product of (1.12) with the source point normal n(x),
one gets:
0 =
∂B
∂G(x, y)
∂n(x)
τ(y) −τ(x)
dS(y)
© 2005 by Taylor & Francis Group, LLC
6 CHAPTER 1. BOUNDARY INTEGRAL EQUATIONS
− u
,k
(x)
B
∂G(x, y)
∂n(x)
n
k
(y) −n
k
(x)
dS(y)
−
∂B
∂F(x, y)
∂n(x)
u(y) −u(x) −u
,p
(x)(y
p
−x
p
)
dS(y) (1.13)
1.1.2.1 Potential gradient on the bounding surface
The gradient of the potential function is required in the regularized HBIEs (1.11
 1.13). For potential problems, the gradient (at a regular boundary point) can
be written as,
∇u = τn +
∂u
∂s
1
t
1
+
∂u
∂s
2
t
2
(1.14)
where τ = ∂u/∂n is the ﬂux, n is the unit normal, t
1
, t
2
are the appropriately
chosen unit vectors in two orthogonal tangential directions on the surface of the
body, and ∂u/∂s
i
, i = 1, 2 are the tangential derivatives of u (along t
1
and t
2
)
on the surface of the body.
1.2 Linear Elasticity in Three Dimensions
The starting point is the NavierCauchy equation governing the displacement
u(x
1
, x
2
, x
3
) in a homogeneous, isotropic, linear elastic solid occupying the
bounded 3D region B with boundary ∂B; in the absence of body forces:
0 = u
i,jj
+
1
1 −2ν
u
k,ki
(1.15)
along with prescribed boundary conditions that involve the displacement and
the traction τ on ∂B. The components τ
i
of the traction vector are:
τ
i
= λu
k,k
n
i
+µ(u
i,j
+u
j,i
)n
j
(1.16)
In equations (1.15) and (1.16), ν is Poisson’s ratio and λ and µ are Lam´e
constants. As is well known, µ is the shear modulus of the material and is also
called G in this book. Finally, the Young’s modulus is denoted as E.
1.2.1 Singular Integral Equations
The wellknown integral representation for (1.15), at an internal point ξ ∈ B
(Rizzo [141]) is:
u
k
(ξ) =
∂B
[U
ik
(ξ, y)τ
i
(y) −T
ik
(ξ, y)u
i
(y)] dS(y) (1.17)
where u
k
and τ
k
are the components of the displacement and traction respec
tively, and the wellknown Kelvin kernels are:
© 2005 by Taylor & Francis Group, LLC
1.2. LINEAR ELASTICITY IN THREE DIMENSIONS 7
U
ik
=
1
16π(1 −ν)Gr
[(3 −4ν)δ
ik
+r
,i
r
,k
] (1.18)
T
ik
= −
1
8π(1 −ν)r
2
{(1 −2ν)δ
ik
+ 3r
,i
r
,k
}
∂r
∂n
+ (1 −2ν)(r
,i
n
k
−r
,k
n
i
)
(1.19)
In the above, δ
ik
denotes the Kronecker delta and, as before, the normal n
is deﬁned at the (boundary) ﬁeld point y. A comma denotes a derivative with
respect to a ﬁeld point, i.e.
r
,i
=
∂r
∂y
i
=
y
i
−ξ
i
r
(1.20)
An alternative form of equation (1.17) is:
u
k
(ξ) =
∂B
[U
ik
(ξ, y)σ
ij
(y) −Σ
ijk
(ξ, y)u
i
(y)] e
j
· dS(y) (1.21)
where σ is the stress tensor, τ
i
= σ
ij
n
j
and T
ik
= Σ
ijk
n
j
. (Please note that
e
j
· dS(y) = n
j
(y)dS(y)). The explicit form of the kernel Σ is:
Σ
ijk
= E
ijmn
∂U
km
∂y
n
= −
1
8π(1 −ν)r
2
[ (1 −2ν)(r
,i
δ
jk
+r
,j
δ
ik
−r
,k
δ
ij
) + 3r
,i
r
,j
r
,k
] (1.22)
where E is the elasticity tensor (for isotropic elasticity):
E
ijmn
= λδ
ij
δ
mn
+µ[δ
im
δ
jn
+δ
in
δ
jm
] (1.23)
The boundary integral equation (BIE) corresponding to (1.17) is obtained
by taking the limit ξ → x. The result is:
u
k
(x) = lim
ξ→x
∂B
[U
ik
(ξ, y)τ
i
(y) −T
ik
(ξ, y)u
i
(y)] dS(y)
=
∂B
= [U
ik
(x, y)τ
i
(y) −T
ik
(x, y)u
i
(y)] dS(y) (1.24)
where the symbol
∂B
= denotes the ﬁnite part of the appropriate integral (see
Section 1.4).
A regularized form of equation (1.24) is:
0 =
∂B
[U
ik
(x, y)τ
i
(y) −T
ik
(x, y){u
i
(y) −u
i
(x)}]dS(y) (1.25)
© 2005 by Taylor & Francis Group, LLC
8 CHAPTER 1. BOUNDARY INTEGRAL EQUATIONS
with an alternate form (from (1.21)):
0 =
∂B
[U
ik
(x, y)σ
ij
(y) −Σ
ijk
(x, y){u
i
(y) −u
i
(x)}]e
j
· dS(y) (1.26)
1.2.2 Hypersingular Integral Equations
Equation (1.17) can be diﬀerentiated at an internal source point ξ to obtain
the displacement gradient at this point:
∂u
k
(ξ)
∂ξ
m
=
∂B
∂U
ik
∂ξ
m
(ξ, y)τ
i
(y) −
∂T
ik
∂ξ
m
(ξ, y)u
i
(y)
dS(y) (1.27)
An alternative form of equation (1.27) is:
∂u
k
(ξ)
∂ξ
m
=
∂B
∂U
ik
∂ξ
m
(ξ, y)σ
ij
(y) −
∂Σ
ijk
∂ξ
m
(ξ, y)u
i
(y)
e
j
· dS(y) (1.28)
Stress components at an internal point ξ can be obtained from either of
equations (1.27) or (1.28) by using Hooke’s law:
σ
ij
= λu
k,k
δ
ij
+µ(u
i,j
+u
j,i
) (1.29)
It is sometimes convenient, however, to write the internal stress directly.
This equation, corresponding (for example) to (1.27) is:
σ
ij
(ξ) =
∂B
[D
ijk
(ξ, y)τ
k
(y) −S
ijk
(ξ, y)u
k
(y)] dS(y) (1.30)
where the new kernels D and S are:
D
ijk
= E
ijmn
∂U
km
∂ξ
n
= λ
∂U
km
∂ξ
m
δ
ij
+µ
∂U
ki
∂ξ
j
+
∂U
kj
∂ξ
i
= −Σ
ijk
(1.31)
S
ijk
= E
ijmn
∂Σ
kpm
∂ξ
n
n
p
= λ
∂Σ
kpm
∂ξ
m
n
p
δ
ij
+µ
∂Σ
kpi
∂ξ
j
+
∂Σ
kpj
∂ξ
i
n
p
=
G
4π(1 −ν)r
3
3
∂r
∂n
[(1 −2ν)δ
ij
r
,k
+ν(δ
ik
r
,j
+δ
jk
r
,i
) −5r
,i
r
,j
r
,k
]
+
G
4π(1 −ν)r
3
[3ν(n
i
r
,j
r
,k
+n
j
r
,i
r
,k
)
+(1 −2ν)(3n
k
r
,i
r
,j
+n
j
δ
ik
+n
i
δ
jk
) −(1 −4ν)n
k
δ
ij
] (1.32)
Again, the normal n is deﬁned at the (boundary) ﬁeld point y. Also:
© 2005 by Taylor & Francis Group, LLC
1.2. LINEAR ELASTICITY IN THREE DIMENSIONS 9
∂U
ik
∂ξ
m
(ξ, y) = −U
ik,m
,
∂Σ
ijk
∂ξ
m
(ξ, y) = −Σ
ijk,m
(1.33)
It is important to note that D becomes strongly singular, and S hypersin
gular as a source point approaches a ﬁeld point (i.e. as r → 0).
For future use in Chapter 4, it is useful to rewrite (1.28) using (1.33). This
equation is:
u
k,m
(ξ) = −
∂B
[U
ik,m
(ξ, y)σ
ij
(y) −Σ
ijk,m
(ξ, y)u
i
(y)] n
j
(y)dS(y) (1.34)
Again, as one takes the limit ξ → x in any of the equations (1.27), (1.28) or
(1.30), one must take the ﬁnite part of the corresponding right hand side (see
Section 1.4.3). For example, (1.28) and (1.30) become, respectively:
∂u
k
(x)
∂x
m
= lim
ξ→x
∂B
∂U
ik
∂ξ
m
(ξ, y)σ
ij
(y) −
∂Σ
ijk
∂ξ
m
(ξ, y)u
i
(y)
n
j
(y)dS(y)
=
∂B
=
∂U
ik
∂x
m
(x, y)σ
ij
(y) −
∂Σ
ijk
∂x
m
(x, y)u
i
(y)
n
j
(y)dS(y) (1.35)
σ
ij
(x) = lim
ξ→x
∂B
[D
ijk
(ξ, y)τ
k
(y) −S
ijk
(ξ, y)u
k
(y)] dS(y)
=
∂B
= [D
ijk
(x, y)τ
k
(y) −S
ijk
(x, y)u
k
(y)] dS(y) (1.36)
Also, for future reference, one notes that the traction at a boundary point
is:
τ
i
(x) = n
j
(x) lim
ξ→x
∂B
[D
ijk
(ξ, y)τ
k
(y) −S
ijk
(ξ, y)u
k
(y)] dS(y) (1.37)
Fully regularized forms of equations (1.35) and (1.36), that only contain
weakly singular integrals, are available in the literature (see, for example, Cruse
and Richardson [39]). These equations, that can be collocated at an irregular
point x ∈ ∂B provided that the stress and displacement ﬁelds in (1.38, 1.39)
satisfy certain smoothness requirements (see Martin et al. [93] and, also, Section
1.4.4 of this chapter) are:
0 =
∂B
U
ik,m
(x, y) [σ
ij
(y) −σ
ij
(x)] n
j
(y)dS(y)
−
∂B
Σ
ijk,m
(x, y) [u
i
(y) −u
i
(x) −u
i,
(x) (y
−x
)] n
j
(y)dS(y) (1.38)
© 2005 by Taylor & Francis Group, LLC
10 CHAPTER 1. BOUNDARY INTEGRAL EQUATIONS
0 =
∂B
D
ijk
(x, y) [σ
kp
(y) −σ
kp
(x)] n
p
(y)dS(y)
−
∂B
S
ijk
(x, y) [u
k
(y) −u
k
(x) −u
k,p
(x)(y
p
−x
p
)] dS(y) (1.39)
An alternate version of (1.39) that can only be collocated at a regular point
x ∈ ∂B is:
0 =
∂B
D
ijk
(x, y)[τ
k
(y) −τ
k
(x)]dS(y)
− σ
km
(x)
∂B
D
ijk
(x, y)(n
m
(y) −n
m
(x))dS(y)
−
∂B
S
ijk
(x, y) [u
k
(y) −u
k
(x) −u
k,m
(x)(y
m
−x
m
)] dS(y) (1.40)
Finally, taking the inner product of (1.40) with the normal at the source
point gives:
0 =
∂B
D
ijk
(x, y)n
j
(x)[τ
k
(y) −τ
k
(x)]dS(y)
− σ
km
(x)
∂B
D
ijk
(x, y)n
j
(x)[n
m
(y) −n
m
(x)]dS(y)
−
∂B
S
ijk
(x, y)n
j
(x) [u
k
(y) −u
k
(x) −u
k,m
(x)(y
m
−x
m
)] dS(y) (1.41)
1.2.2.1 Displacement gradient on the bounding surface
The gradient of the displacement u is required for the regularized HBIEs (1.38
 1.41). Lutz et al. [89] have proposed a scheme for carrying this out. Details
of this procedure are available in [27] and are given below.
The (righthanded) global Cartesian coordinates, as before, are (x
1
, x
2
, x
3
).
Consider (righthanded) local Cartesian coordinates (x
1
, x
2
, x
3
) at a regular
point P on ∂B as shown in Figure 1.2. The local coordinate system is oriented
such that the x
1
and x
2
coordinates lie along the tangential unit vectors t
1
and t
2
while x
3
is measured along the outward normal unit vector n to ∂B as
deﬁned in equation (1.14).
Therefore, one has:
x
= Qx (1.42)
u
= Qu (1.43)
where u
k
, k = 1, 2, 3 are the components of the displacement vector u in the
local coordinate frame, and the orthogonal transformation matrix Q has the
components:
© 2005 by Taylor & Francis Group, LLC
1.2. LINEAR ELASTICITY IN THREE DIMENSIONS 11
x
x
x
x
x
1
2
3
3
1
2
'
'
'
x
P
Figure 1.2: Local coordinate system on the surface of a body (from [27])
Q =
t
11
t
12
t
13
t
21
t
22
t
23
n
1
n
2
n
3
(1.44)
with t
ij
the j
th
component of the i
th
unit tangent vector and (n
1
, n
2
, n
3
) the
components of the unit normal vector.
The tangential derivatives of the displacement, in local coordinates, are
u
i,k
, i = 1, 2, 3; k = 1, 2. These quantities are obtained as follows:
u
i,k
≡
∂u
i
∂s
k
= Q
ij
∂u
j
∂s
k
(1.45)
where ∂u
i
/∂s
k
are tangential derivatives of u
i
at P with s
1
= x
1
and s
2
= x
2
.
The remaining components of ∇u in local coordinates are obtained from
Hooke’s law (see [89]) as:
∂u
1
∂x
3
=
τ
1
G
−
∂u
3
∂x
1
∂u
2
∂x
3
=
τ
2
G
−
∂u
3
∂x
2
∂u
3
∂x
3
=
(1 −2ν)τ
3
2G(1 −ν)
−
ν
1 −ν
∂u
1
∂x
1
+
∂u
2
∂x
2
(1.46)
where τ
k
, k = 1, 2, 3, are the components of the traction vector in local coordi
nates.
The components of the displacement gradient tensor, in the local coordinate
system, are now known. They can be written as:
© 2005 by Taylor & Francis Group, LLC
12 CHAPTER 1. BOUNDARY INTEGRAL EQUATIONS
(∇u)
local
≡ A
=
u
1,1
u
1,2
u
1,3
u
2,1
u
2,2
u
2,3
u
3,1
u
3,2
u
3,3
(1.47)
Finally, the components of ∇u in the global coordinate frame are obtained
from those in the local coordinate frame by using the tensor transformation
rule:
(∇u)
global
≡ A = Q
T
A
Q =
u
1,1
u
1,2
u
1,3
u
2,1
u
2,2
u
2,3
u
3,1
u
3,2
u
3,3
(1.48)
The gradient of the displacement ﬁeld in global coordinates is now ready for
use in equations (1.38  1.41).
1.3 Nearly Singular Integrals in Linear Elastic
ity
It is well known that the ﬁrst step in the BEM is to solve the primary problem
on the bounding surface of a body (e.g. equation (1.25)) and obtain all the
displacements and tractions on this surface. The next steps are to obtain the
displacements and stresses at selected points inside a body, from equations
such as (1.17) and (1.30). It has been known in the BEM community for many
years, dating back to Cruse [37], that one experiences diﬃculties when trying
to numerically evaluate displacements and stresses at points inside a body that
are close to its bounding surface (the socalled nearsingular or boundary layer
problem). Various authors have addressed this issue over the last 3 decades.
This section describes a new method recently proposed by Mukherjee et al.
[104].
1.3.1 Displacements at Internal Points Close to the Bound
ary
The displacement at a point inside an elastic body can be determined from
either of the (equivalent) equations (1.17) or (1.21). A continuous version of
(1.21), from Cruse and Richardson [39] is:
u
k
(ξ) = u
k
ˆ
(x) +
∂B
[ U
ik
(ξ, y)σ
ij
(y) −Σ
ijk
(ξ, y){u
i
(y) −u
i
(ˆ x)} ] n
j
(y)dS(y)
(1.49)
where ξ ∈ B is an internal point close to ∂B and a target point ˆ x ∈ ∂B is close
to the point ξ (see Fig. 1.3). An alternative form of (1.49) is:
© 2005 by Taylor & Francis Group, LLC
1.3. NEARLY SINGULAR INTEGRALS IN LINEAR ELASTICITY 13
B
∂B
ξ
z
z
v
x
v
y
Figure 1.3: A body with source point ξ, ﬁeld point y and target point ˆ x
(from [104])
u
k
(ξ) = u
k
(ˆ x)+
∂B
[ U
ik
(ξ, y)τ
i
(y) −T
ik
(ξ, y){u
i
(y) −u
i
(ˆ x)} ] dS(y) (1.50)
Equation (1.49) (or (1.50)) is called “continuous” since it has a continuous
limit to the boundary (LTB as ξ → ˆ x ∈ ∂B) provided that u
i
(y) ∈ C
0,α
(i.e.
H¨older continuous). Taking this limit is the standard approach for obtaining
the wellknown regularized form (1.26) (or (1.25)).
In this work, however, equation (1.49) (or (1.50)) is put to a diﬀerent, and
novel use. It is ﬁrst observed that T
ik
in equation (1.50) is O(1/r
2
(ξ, y)) as
ξ → y, whereas {u
i
(y) − u
i
(ˆ x)} is O(r(ˆ x, y)) as y → ˆ x. Therefore, as y → ˆ x,
the product T
ik
(ξ, y){u
i
(y) −u
i
(ˆ x)}, which is O(r(ˆ x, y)/r
2
(ξ, y)), → 0 ! As a
result, equation (1.50) (or (1.49)) can be used to easily and accurately evaluate
the displacement components u
k
(ξ) for ξ ∈ B close to ∂B. This idea is the
main contribution of [104].
It is noted here that while it is usual to use (1.17) (or (1.21)) to evaluate
u
k
(ξ) when ξ is far from ∂B, equation (1.49) (or (1.50)) is also valid in this
case. (The target point ˆ x can be chosen as any point on ∂B when ξ is far from
∂B). Therefore, it is advisable to use the continuous equation (1.49) (or (1.50))
universally for all points ξ ∈ B. This procedure would eliminate the need to
classify, a priori, whether ξ is near to, or far from ∂B.
1.3.2 Stresses at Internal Points Close to the Boundary
The displacement gradient at a point ξ ∈ B can be obtained from equation
(1.34) or the stress at this point from (1.30). Continuous versions of (1.34) and
(1.30) can be written as [39]:
© 2005 by Taylor & Francis Group, LLC
14 CHAPTER 1. BOUNDARY INTEGRAL EQUATIONS
u
k,n
(ξ) = u
k,n
(ˆ x) −
∂B
U
ik,n
(ξ, y) [σ
ij
(y) −σ
ij
(ˆ x)] n
j
(y)dS(y)
+
∂B
Σ
ijk,n
(ξ, y) [u
i
(y) −u
i
(ˆ x) −u
i,
(ˆ x) (y
− ˆ x
)] n
j
(y)dS(y) (1.51)
σ
ij
(ξ) = σ
ij
(ˆ x) +
∂B
D
ijk
(ξ, y)[τ
k
(y) −σ
km
(ˆ x)n
m
(y)]dS(y)
−
∂B
S
ijk
(ξ, y)[u
k
(y) −u
k
(ˆ x) −u
k,
(ˆ x)(y
− ˆ x
)] dS(y) (1.52)
The integrands in equations (1.51) (or (1.52)) are O(r(ˆ x, y)/r
2
(ξ, y)) and
O(r
2
(ˆ x, y)/r
3
(ξ, y)) as y → ˆ x. Similar to the behavior of the continuous BIEs
in the previous subsection, the integrands in equations (1.51) and (1.52) → 0
as y → ˆ x. Either of these equations, therefore, is very useful for evaluating the
stresses at an internal point ξ that is close to ∂B. Of course (please see the
discussion regarding displacements in the previous section), they can also be
conveniently used to evaluate displacement gradients or stresses at any point
ξ ∈ B.
Henceforth, use of equations (1.17), (1.21), (1.30) or (1.34) will be referred
to as the standard method, while use of equations (1.49), (1.50), (1.51) or (1.52)
will be referred to as the new method.
1.4 Finite Parts of Hypersingular Equations
A discussion of ﬁnite parts (FPs) of hypersingular BIEs (see e.g. equations
(1.9 1.11)) is the subject of this section. The general theory of ﬁnite parts
is presented ﬁrst. This is followed by applications of the theory in potential
theory and in linear elasticity. Further details are available in Mukherjee [102].
1.4.1 Finite Part of a Hypersingular Integral Collocated
at an Irregular Boundary Point
1.4.1.1 Deﬁnition
Consider, for speciﬁcity, the space R
3
, and let S be a surface in R
3
. Let the
points x ∈ S and ξ / ∈ S. Also, let
ˆ
S and
¯
S ⊂
ˆ
S be two neighborhoods (in S) of
x such that x ∈
¯
S (Figure 1.4). The point x can be an irregular point on S.
Let the function K(x, y) , y ∈ S, have its only singularity at x = y of
the form 1/r
3
where r = x − y , and let φ(y) be a function that has no
singularity in S and is of class C
1,α
at y = x for some α > 0.
The ﬁnite part of the integral
© 2005 by Taylor & Francis Group, LLC
1.4. FINITE PARTS OF HYPERSINGULAR EQUATIONS 15
S
S
x
S
y
S
ξ
S
S
Figure 1.4: A surface S with regions
ˆ
S and
¯
S and points ξ, x and y (from [102])
I(x) =
S
K(x, y)φ(y)dS(y) (1.53)
is deﬁned as:
S
= K(x, y)φ(y)dS(y) =
S\
ˆ
S
K(x, y)φ(y)dS(y)
+
ˆ
S
K(x, y)[φ(y) −φ(x) −φ
,p
(x)(y
p
−x
p
)]dS(y)
+ φ(x)A(
ˆ
S) +φ
,p
(x)B
p
(
ˆ
S) (1.54)
where
ˆ
S is any arbitrary neighborhood (in S) of x and:
A(
ˆ
S) =
ˆ
S
= K(x, y)dS(y) (1.55)
B
p
(
ˆ
S) =
ˆ
S
= K(x, y)(y
p
−x
p
)dS(y) (1.56)
The above FP deﬁnition can be easily extended to any number of physical
dimensions and any order of singularity of the kernel function K(x, y). Please
refer to Toh and Mukherjee [168] for further discussion of a previous closely re
lated FP deﬁnition for the case when x is a regular point on S, and to Mukherjee
[101] for a discussion of the relationship of this FP to the CPV of an integral
when its CPV exists.
© 2005 by Taylor & Francis Group, LLC
16 CHAPTER 1. BOUNDARY INTEGRAL EQUATIONS
1.4.1.2 Evaluation of A and B
There are several equivalent ways for evaluating A and B.
Method one. Replace S by
ˆ
S and
ˆ
S by
¯
S in equation (1.54). Now, setting
φ(y) = 1 in (1.54) and using (1.55), one gets:
A(
ˆ
S) −A(
¯
S) =
ˆ
S\
¯
S
K(x, y)dS(y) (1.57)
Next, setting φ(y) = (y
p
− x
p
) (note that, in this case, φ(x) = 0 and
φ
,p
(x) = 1) in (1.54), and using (1.56), one gets:
B
p
(
ˆ
S) −B
p
(
¯
S) =
ˆ
S\
¯
S
K(x, y)(y
p
−x
p
)dS(y) (1.58)
The formulae (1.57) and (1.58) are most useful for obtaining A and B when
ˆ
S is an open surface and Stoke regularization is employed. An example is the
application of the FP deﬁnition (1.54) (for a regular collocation point) in Toh
and Mukherjee [168], to regularize a hypersingular integral that appears in the
HBIE formulation for the scattering of acoustic waves by a thin scatterer. The
resulting regularized equation is shown in [168] to be equivalent to the result of
Krishnasamy et al. [75]. Equations (1.57) and (1.58) are also used in Mukherjee
and Mukherjee [99] and in Section 3.2 of [102].
Method two. From equation (1.57):
A(
ˆ
S) −A(
¯
S) =
ˆ
S\
¯
S
K(x, y)dS(y) = lim
ξ→x
ˆ
S\
¯
S
K(ξ, y)dS(y) (1.59)
The second equality above holds since K(x, y) is regular for x ∈
¯
S and
y ∈
ˆ
S\
¯
S. Assuming that the limits:
lim
ξ→x
ˆ
S
K(ξ, y)dS(y), lim
ξ→x
¯
S
K(ξ, y)dS(y)
exist, then:
A(
ˆ
S) = lim
ξ→x
ˆ
S
K(ξ, y)dS(y) (1.60)
Similarly:
B
p
(
ˆ
S) = lim
ξ→x
ˆ
S
K(ξ, y)(y
p
−x
p
)dS(y) (1.61)
Equations (1.60) and (1.61) are most useful for evaluating A and B when
ˆ
S = ∂B, a closed surface that is the entire boundary of a body B. Examples
appear in Sections 1.4.2 and 1.4.3 of this chapter.
© 2005 by Taylor & Francis Group, LLC
1.4. FINITE PARTS OF HYPERSINGULAR EQUATIONS 17
Method three. A third way for evaluation of A and B is to use an auxiliary
surface (or “tent”) as ﬁrst proposed for fracture mechanics analysis by Lutz et
al. [89]. (see, also, Mukherjee et al. [110], Mukherjee [105] and Section 3.2.1 of
[102]. This method is useful if S is an open surface.
1.4.1.3 The FP and the LTB
There is a very simple connection between the FP, deﬁned above, and the
LTB approach employed by Gray and his coauthors. With, as before, ξ / ∈ S,
x ∈ S (x can be an irregular point on S), K(x, y) = O(x − y
−3
) as y → x
and φ(y) ∈ C
1,α
at y = x, this can be stated as:
lim
ξ→x
S
K(ξ, y)φ(y)dS(y) =
S
= K(x, y)φ(y)dS(y) (1.62)
Of course, ξ can approach x from either side of S.
Proof of equation (1.62). Consider the ﬁrst and second terms on the right
hand side of equation (1.54). Since these integrands are regular in their respec
tive domains of integration, one has:
S\
ˆ
S
K(x, y)φ(y)dS(y) = lim
ξ→x
S\
ˆ
S
K(ξ, y)φ(y)dS(y) (1.63)
and
ˆ
S
K(x, y)[φ(y) −φ(x) −φ
,p
(x) (y
p
−x
p
)]dS(y)
= lim
ξ→x
ˆ
S
K(ξ, y)[φ(y) −φ(ξ) −φ
,p
(ξ)(y
p
−ξ
p
)]dS(y) (1.64)
Use of equations (1.60, 1.61, 1.63 and 1.64) in (1.54) proves equation (1.62).
1.4.2 Gradient BIE for 3D Laplace’s Equation
This section is concerned with an application of equation (1.54) for collocation
of the HBIE (1.9), for the 3D Laplace equation, at an irregular boundary point.
A complete exclusion zone,
ˆ
S = ∂B is used here. An application of a vanishing
exclusion zone, for collocation of the HBIE for the 2D Laplace equation, at an
irregular boundary point, is presented in Mukherjee [102].
Using equations (1.4) and (1.6), equations (1.9) and (1.10) are ﬁrst written
in the slightly diﬀerent equivalent forms:
∂u(ξ)
∂ξ
i
=
∂B
[D
i
(ξ, y)τ(y) −S
i
(ξ, y)u(y)] dS(y) (1.65)
∂u(x)
∂x
i
=
∂B
= [D
i
(x, y)τ(y) −S
i
(x, y)u(y)] dS(y) (1.66)
© 2005 by Taylor & Francis Group, LLC
18 CHAPTER 1. BOUNDARY INTEGRAL EQUATIONS
where:
D
i
(x, y) = −G
,i
(x, y) , S
i
= −H
k,i
(x, y)n
k
(y) (1.67)
Use of (1.54) in (1.66), with S =
ˆ
S = ∂B, results in:
u
,i
(x) =
∂B
D
i
(x, y)
u
,p
(y) −u
,p
(x)
n
p
(y)dS(y)
−
∂B
S
i
(x, y)
u(y) −u(x) −u
,p
(x)(y
p
−x
p
)
dS(y)
− A
i
(∂B)u(x) +C
ip
(∂B)u
,p
(x) (1.68)
where, using method two in Section 1.4.1.2:
A
i
(∂B) = lim
ξ→x
∂B
S
i
(ξ, y)dS(y) (1.69)
C
ip
(∂B) = lim
ξ→x
∂B
[D
i
(ξ, y)n
p
(y) −S
i
(ξ, y)(y
p
−ξ
p
)] dS(y) (1.70)
It is noted here that the (possibly irregular) boundary point x is approached
from ξ ∈ B, i.e. from inside the body B.
The quantities A and C can be easily evaluated using the imposition of
simple solutions. Following Rudolphi [143], use of the uniform solution u(y) = c
(c is a constant) in equation (1.65) gives:
∂B
S
i
(ξ, y)dS(y) = 0 (1.71)
while use of the linear solution:
u = u(ξ) + (y
p
−ξ
p
)u
,p
(ξ)
τ(y) =
∂u
∂y
k
n
k
(y) = u
,p
(ξ)n
p
(y) (with p = 1, 2, 3) (1.72)
in equation (1.65) (together with (1.71)) gives:
∂B
[D
i
(ξ, y)n
p
(y) −S
i
(ξ, y)(y
p
−ξ
p
)] dS(y) = δ
ip
(1.73)
Therefore, (assuming continuity) A
i
(∂B) = 0, C
ip
(∂B) = δ
ip
, and (1.68)
yields a simple, fully regularized form of (1.66) as:
0 =
∂B
D
i
(x, y)[u
,p
(y) −u
,p
(x)]n
p
(y)dS(y)
−
∂B
S
i
(x, y)[u(y) −u(x) −u
,p
(x)(y
p
−x
p
)]dS(y) (1.74)
© 2005 by Taylor & Francis Group, LLC
1.4. FINITE PARTS OF HYPERSINGULAR EQUATIONS 19
which is equivalent to equation (1.11).
A few comments are in order. First, equation (1.74) is the same as Rudol
phi’s [143] equation (20) with (his) κ = 1 and (his) S
0
set equal to S and
renamed ∂B. (See, also, Kane [68], equation (17.34)). Second, this equation
can also be shown to be valid for the case ξ / ∈ B, i.e. for an outside approach
to the boundary point x . Third, as noted before, x can be an edge or corner
point on ∂B (provided, of course, that u(y) ∈ C
1,α
at y = x  Rudolphi
had only considered a regular boundary collocation point in his excellent paper
that was published in 1991). Finally, as discussed in the Section 1.4.3, equation
(1.74) is analogous to the regularized stress BIE in linear elasticity  equation
(28) in Cruse and Richardson [39] .
1.4.3 Stress BIE for 3D Elasticity
This section presents a proof of the fact that equation (1.39) is a regularized
version of (1.36), valid at an irregular point x ∈ ∂B, provided that the stress
and displacement ﬁelds in (1.39) satisfy certain smoothness requirements. These
smoothness requirements are discussed in Section 1.4.4. The approach is very
similar to that used in Section 1.4.2.
The ﬁrst step is to apply the FP equation (1.54) to regularize (1.36). With
S =
ˆ
S = ∂B, the result is:
σ
ij
(x) =
∂B
D
ijk
(x, y) [σ
kp
(y) −σ
kp
(x)] n
p
(y)dS(y)
−
∂B
S
ijk
(x, y) [u
k
(y) −u
k
(x) −u
k,p
(x)(y
p
−x
p
)] dS(y)
− A
ijk
(∂B)u
k
(x) +C
ijkp
(∂B)u
k,p
(x) (1.75)
where, using method two in Section 1.4.1.2:
A
ijk
(∂B) = lim
ξ→x
∂B
S
ijk
(ξ, y)dS(y) (1.76)
C
ijkp
(∂B) = lim
ξ→x
∂B
E
mkp
D
ijm
(ξ, y)n
(y)dS(y)
− lim
ξ→x
∂B
S
ijk
(ξ, y)(y
p
−ξ
p
)dS(y) (1.77)
with E the elasticity tensor (see (1.23)) which appears in Hooke’s law:
σ
m
= E
mkp
u
k,p
(1.78)
Simple (rigid body and linear) solutions in linear elasticity (see, for example,
Lutz et al. [89], Cruse and Richardson [39]) are now used in order to determine
© 2005 by Taylor & Francis Group, LLC
20 CHAPTER 1. BOUNDARY INTEGRAL EQUATIONS
the quantities A and C. Using the rigid body mode u
k
= c
k
(c
k
are arbitrary
constants) in (1.30), one has:
0 =
∂B
S
ijk
(ξ, y)dS(y) (1.79)
while, using the linear solution:
u
k
(y) = (y
p
−ξ
p
)u
k,p
(ξ), u
k,m
(y) = u
k,m
(ξ),
τ
k
(y) = σ
km
(y)n
m
(y) = E
kmrs
u
r,s
(ξ)n
m
(y) (1.80)
in equation (1.30) gives:
σ
ij
(ξ) = u
k,p
(ξ)
∂B
[E
mkp
D
ijm
(ξ, y)n
(y) −S
ijk
(ξ, y)(y
p
−ξ
p
)] dS(y)
(1.81)
Taking the limit ξ → x of (1.79), using continuity of the integral and com
paring with (1.76), gives A = 0. Taking the limit ξ → x of (1.81) and comparing
with (1.77), one has:
σ
ij
(x) = C
ijkp
u
k,p
(x) (1.82)
Comparing (1.82) with (1.78) yields C(∂B) = E.
Therefore, equation (1.75) reduces to the simple regularized equation (1.39).
Equation (1.39) is equation (28) of Cruse and Richardson [39] in the present
notation. As is the case in the present work, Cruse and Richardson [39] have
also proved that their equation (28) is valid at a corner point, provided that
the stress is continuous there.
It has been proved in this section that the regularized stress BIE (28) of
Cruse and Richardson [39] can also be obtained from the FP deﬁnition (1.54)
with a complete exclusion zone.
1.4.4 Solution Strategy for a HBIE Collocated at an Ir
regular Boundary Point
Hypersingular BIEs for a body B with boundary ∂B are considered here. Regu
larized HBIEs, obtained by using complete exclusion zones, e.g. equation (1.74)
for potential theory or (1.39) for linear elasticity, are recommended as starting
points.
An irregular collocation point x for 3D problems is considered next. Let
∂B
n
, (n = 1, 2, 3, ..., N) be smooth pieces of ∂B that meet at an irregular point
x ∈ ∂B. Also, as before, let a source point, with coordinates x
k
, be denoted by
P, and a ﬁeld point, with coordinates y
k
, be denoted by Q.
Martin et al. [93] state the following requirements for collocating a regular
ized HBIE, such as (1.39) at an irregular point P ∈ ∂B. These are:
© 2005 by Taylor & Francis Group, LLC
1.4. FINITE PARTS OF HYPERSINGULAR EQUATIONS 21
(i) The displacement u must satisfy the equilibrium equations in B.
(ii) (a) The stress σ must be continuous in B.
(b) The stress σ must be continuous on ∂B.
(iii) u
i
(Q
n
) −u
i
L
(Q
n
; P) = O(r
(1+α)
n
) as r
n
→ 0, for each n.
(iv) [σ
ij
(Q
n
) −σ
ij
(P)]n
j
(Q
n
) = O(r
α
n
) as r
n
→ 0, for each n.
Box 1.1 Requirements for collocation of a HBIE at an irregular point
(from [93]).
In the above, r
n
= y(Q
n
) −x(P), Q
n
∈ ∂B
n
, and α > 0. Also,
u
L
i
(Q
n
; P) = u
i
(P) +u
i,j
(P)[y
j
(Q
n
) −x
j
(P)] (1.83)
There are two important issues to consider here.
The ﬁrst is that, if there is to be any hope for collocating (1.39) at an
irregular point P, the exact solution of a boundary value problem must satisfy
conditions (iiv) in Box 1.1. Clearly, one should not attempt this collocation
if, for example, the stress is unbounded at P (this can easily happen  see an
exhaustive study on the subject in Glushkov et al. [50]), or is bounded but
discontinuous at P (e.g. at the tip of a wedge  see, for example, Zhang and
Mukherjee [183]). The discussion in the rest of this book is limited to the class
of problems, referred to as the admissible class, whose exact solutions satisfy
conditions (i  iv).
The second issue refers to smoothness requirements on the interpolation
functions for u, σ and the traction τ = n · σ in (1.39). It has proved very
diﬃcult, in practice, to ﬁnd BEM interpolation functions that satisfy, a priori,
(ii(b)(iv)) in Box 1.1, for collocation at an irregular surface point on a 3D
body [93]. It has recently been proved in Mukherjee and Mukherjee [111],
however, that interpolation functions used in the boundary contour method
(BCM  see, for example, Mukherjee et al. [109], Mukherjee and Mukherjee
[99]) satisfy these conditions a priori. Another important advantage of using
these interpolation functions is that ∇u can be directly computed from them
at an irregular boundary point [99], without the need to use the (undeﬁned)
normal and tangent vectors at this point. In principle, these BCM interpolation
functions can also be used in the BEM.
The BCM and the hypersingular BCM (HBCM) are discussed in detail
in Chapter 4 of this book. Numerical results from the hypersingular BCM,
collocated on edges and at corners, from Mukherjee and Mukherjee [111], are
available in Chapter 4.
© 2005 by Taylor & Francis Group, LLC
Chapter 2
ERROR ESTIMATION
Pointwise (i.e. that the error is evaluated at selected points) residualbased error
estimates for Dirichlet, Neumann and mixed boundary value problems (BVPs)
in linear elasticity are presented ﬁrst in this chapter. Interesting relationships
between the actual error and the hypersingular residuals are proved for the ﬁrst
two classes of problems, while heuristic error estimators are presented for mixed
BVPs. Elementbased error indicators, relying on the pointwise error measures
presented earlier, are proposed next. Numerical results for two mixed BVPs
in 2D linear elasticity complete this chapter. Further details are available in
[127].
2.1 Linear Operators
Boundary integral equations can be analyzed by viewing them as linear equa
tions in a Hilbert space. A very readable account of this topic is available in
Kress [73]. Following Sloan [155], it is assumed here that the boundary ∂B is
a C
1
continuous closed Jordan curve given by the mapping:
z : [0, 1] → ∂B, z ∈ C
1
, [z
[ = 0
where z ∈ C, the space of complex numbers. The present analysis excludes
domains with corners. It is also assumed that any integrable function v on ∂B
may be represented in a Fourier series:
v ∼
∞
k=−∞
ˆ v(k)e
2πikx
1
= a
0
+
∞
k=1
(a
k
cos(2πkx
1
) + b
k
sin(2πkx
1
) (2.1)
where i ≡
√
−1 and:
ˆ v(k) =
1
0
e
−2πikx
1
v(x
1
)dx
1
, k ∈ Z (2.2)
23
© 2005 by Taylor & Francis Group, LLC
24 CHAPTER 2. ERROR ESTIMATION
in which Z denotes the space of integers.
The following Lemma is very useful for the work presented in this chapter.
Lemma 1. If / : B
1
→ B
2
is a continuous linear operator that has a
continuous inverse, and /x = y, then there exist real positive constants C
1
and
C
2
, such that:
C
1
y
B
2
≤ x
B
1
≤ C
2
y
B
2
where  
B
i
denotes a suitable norm of the appropriate function (in the Banach
space B
i
).
Proof: The linearity and continuity of / and /
−1
imply that / and /
−1

are ﬁnite. From the CauchySchwarz inequality, one has:
y = /x ≤ /x
x = /
−1
y ≤ /
−1
y
The result now follows by choosing C
1
= 1// and C
2
= /
−1
.
Returning to the problem at hand, the following operators are deﬁned as:
(
ij
v
j
)(ξ) :=
∂B
U
ij
(ξ, y)v
j
(y)dS(y) (2.3)
(T
ij
v
j
)(ξ) :=
∂B
T
T
ij
(ξ, y)v
j
(y)dS(y) (2.4)
(T
ijk
v
k
)(ξ) :=
∂B
D
ijk
(ξ, y)v
k
(y)ds(y) (2.5)
(o
ijk
v
k
)(ξ) :=
∂B
S
ijk
(ξ, y)v
k
(y)ds(y) (2.6)
(T
(N)
ik
v
k
)(ξ, x) := n
j
(x)
∂B
D
ijk
(ξ, y)v
k
(y)ds(y) (2.7)
(o
(N)
ik
v
k
)(ξ, x) := n
j
(x)
∂B
S
ijk
(ξ, y)v
k
(y)ds(y) (2.8)
The operator 
ij
is continuous onto the boundary, whereas T
ij
and T
(N)
ij
are
not continuous (Tanaka et al. [162]) and give rise to additional bounded free
terms in the limit. The hypersingular operator o
(N)
ij
gives rise to unbounded
terms that vanish when the integral is considered, for example, in the LTB
sense. These terms depend on the smoothness of the boundary at the source
© 2005 by Taylor & Francis Group, LLC
2.2. ITERATED HBIE AND ERROR ESTIMATION 25
point x. In this work, the HBIE is collocated only at regular boundary points
(where the boundary is locally smooth) inside boundary elements.
Using the operators deﬁned above, the BIE (1.24) and HBIE (1.37) become,
respectively:
BIE : u
i
= 
ij
τ
j
−T
ij
u
j
(2.9)
HBIE : τ
i
= T
(N)
ij
τ
j
−o
(N)
ij
u
j
(2.10)
As in the case of potential theory (Menon et al. [96]), the LTB of the above
integral operators has been used to obtain the singular integral equations (2.9)
and (2.10).
Remark 1 One should note, however, that key properties of the operators, such
as continuity and invertibility, assume a certain regularity of the boundary (for
instance no corners or cusps) [155]. These assumptions, of course, are too re
strictive for the solution of practical engineering problems. Such assumptions
have, nevertheless, been made here in order to obtain some mathematical un
derstanding of the error estimation process that is described in Section 2.2. The
numerical example problems do contain corners. The HBIE (2.10), however,
has only been collocated at regular points on the boundary of a body.
2.2 Iterated HBIE and Error Estimation
The heuristic idea that is at the heart of the pointwise error estimation proce
dure described below (see also [122, 123, 96]) is simple : the amount by which
an approximate solution to the BIE fails to satisfy the HBIE is a measure of the
error in the approximation. The main result of this work is that this heuristic
idea, when stated formally, leads to a simple characterization of the error. In
essence, the method reduces to ﬁnding a second approximation to the solution
by iterating the ﬁrst approximation with the HBIE. This idea is ﬁrst illustrated
in the context of two basic cases : the interior Dirichlet and Neumann problems.
Mixed boundary conditions are considered thereafter.
2.2.1 Problem 1 : Displacement Boundary Conditions
Solve the NavierCauchy equations:
(λ + µ)∇(∇ u) + µ∇
2
u = 0 in B
subject to the boundary conditions:
u = f on ∂B
This problem is analogous to the Dirichlet problem of potential theory. Un
der suitable restrictions on the domain, it is possible to prove existence and
uniqueness of a solution to this BVP [46].
© 2005 by Taylor & Francis Group, LLC
26 CHAPTER 2. ERROR ESTIMATION
Either the displacement BIE (2.9) or the traction HBIE (2.10) may be used
to formulate a method of solution for the unknown traction on the boundary.
The displacement BIE leads to a system of singular integral equations of the
ﬁrst kind for the (unknown) traction:

ij
τ
j
= f
i
+T
ij
f
j
=: g
1
i
i = 1, 2 (2.11)
while the traction BIE gives rise to equations of the second kind (for the trac
tion):
τ
i
−T
(N)
ij
τ
j
= −o
(N)
ij
f
j
=: g
2
i
i = 1, 2 (2.12)
Recall that 
ij
is logsingular, T
ij
and T
(N)
ij
are Cauchy singular, and o
(N)
ij
is hypersingular. As in potential theory, since 
ij
has a logarithmic kernel, one
again encounters the problem of the transﬁnite diameter. For instance, one
may show that if the domain B is a circle of radius exp[(1/2)(3 −4ν)], then the
BIE (2.11) does not admit a unique solution.
2.2.1.1 Error estimate for the primary problem
Using the two BIEs (2.11) and (2.12), one can formulate an error estimation
process that is analogous to the Dirichlet problem in potential theory [96].
• Step 1: Solve the displacement BIE (2.11) for the traction τ
(1)
i
:

ij
τ
(1)
j
= (1
ij
+T
ij
)f
j
(2.13)
where 1 is the identity operator and 1
ij
f
j
= δ
ij
f
j
= f
i
, with δ
ij
the
components of the Kronecker delta.
• Step 2: Use the traction HBIE (2.12) to iterate the traction and obtain a
second approximation τ
(2)
i
:
τ
(2)
i
= T
(N)
ij
τ
(1)
j
−o
(N)
ij
f
j
(2.14)
This approximation, called the HBIE iterate, will be used for error estima
tion.
Let the error (in traction) in the primary solution and iterate be:
e
τ(1)
i
= τ
(1)
i
−τ
i
(2.15)
e
τ(2)
i
= τ
(2)
i
−τ
i
(2.16)
respectively. Deﬁne the hypersingular residual to be:
© 2005 by Taylor & Francis Group, LLC
2.2. ITERATED HBIE AND ERROR ESTIMATION 27
r
(τ)
i
= τ
(1)
i
−τ
(2)
i
(2.17)
One can now show that:
r
(τ)
i
(2.17)
= τ
(1)
i
−τ
(2)
i
(2.14)
= τ
(1)
i
−(T
(N)
ij
τ
(1)
j
−o
(N)
ij
f
j
)
(2.15)
= τ
i
−(T
(N)
ij
τ
j
−o
(N)
ij
f
j
) + e
τ(1)
i
−T
(N)
ij
e
τ(1)
j
(2.10)
= (1
ij
−T
(N)
ij
)e
τ(1)
j
so that:
r
(τ)
i
= (1
ij
−T
(N)
ij
)e
τ(1)
j
(2.18)
Theorem 1 For a suﬃciently smooth domain, and suﬃciently smooth data
and solutions (as detailed above), if the solution to the integral equations (2.11)
and (2.12) is unique, then there exist real positive constants C
1
and C
2
such
that:
C
1
r
(τ)
i
 ≤ e
τ(1)
i
 ≤ C
2
r
(τ)
i

Proof: The continuity of the operators is a manifestation of the elliptic
nature of the partial diﬀerential equation (PDE). Uniqueness of solutions to
the integral formulations implies that the operators (1
ij
−T
(N)
ij
) and 
ij
have
continuous inverses [172]. Now use Lemma 1.
2.2.1.2 Error estimate for the iterate
In a manner similar to the previous subsection, one can show that:
e
τ(2)
i
(2.16)
= τ
(2)
i
−τ
i
(2.14)
= T
(N)
ij
τ
(1)
j
−o
(N)
ij
f
j
−τ
i
(2.15)
= T
(N)
ij
(e
τ(1)
j
+ τ
j
) −o
(N)
ij
f
j
−τ
i
(2.10)
= T
(N)
ij
e
τ(1)
j
so that:
e
τ(2)
i
= T
(N)
ij
e
τ(1)
j
(2.19)
© 2005 by Taylor & Francis Group, LLC
28 CHAPTER 2. ERROR ESTIMATION
2.2.2 Problem 2 : Traction Boundary Conditions
Solve the NavierCauchy equations:
(λ + µ)∇(∇ u) + µ∇
2
u = 0 in B
subject to the boundary conditions:
t = g on ∂B
where the tractions satisfy the consistency conditions of static equilibrium:
∂B
t ds = 0
∂B
(r t) ds = 0
It is known that the solution to the above problem exists, and is unique up
to a rigid body motion (Fung [46]). The space of twodimensional rigid body
motions may be characterized as (Chen and Zhou [29]):
{ := r
0
+ ω r (2.20)
where r
0
∈ R
2
is a translation, and ω = ω k is an axial vector representing a
rotation.
The ﬁrst integral equation formulation for the problem follows from the
displacement BIE (2.9). One has an integral equation of the second kind for
the (unknown) displacement:
u
i
+T
ij
u
j
= 
ij
g
j
=: h
1
i
(2.21)
and using the traction HBIE (2.10):
−o
(N)
ij
u
j
= g
i
−T
(N)
ij
g
j
=: h
2
i
(2.22)
one obtains an integral equation of the ﬁrst kind for the displacement.
It is important to mention again that the solution of the traction prescribed
BVP is arbitrary within a rigid body motion, and, to eliminate this arbitrari
ness, one must work in a restricted function space as has been done before [96]
for Neumann problems in potential theory. An elegant practical way to solve
traction prescribed problems in linear elasticity is outlined in a recent paper by
Lutz et al. [90] where the singular matrix from the BIE is suitably regularized
at the discretized level by eliminating rigid body modes.
© 2005 by Taylor & Francis Group, LLC
2.2. ITERATED HBIE AND ERROR ESTIMATION 29
2.2.2.1 Error estimate for the primary problem
The error estimation technique is analogous to the Neumann problem investi
gated previously by Menon et al. [96]. First, construct an approximation to
the displacement ﬁeld, u
(1)
i
. Next ﬁnd τ
(2)
i
, an iterated approximation to the
traction, and use it to estimate the error in the primary solution.
• Step 1: Solve the displacement BIE (2.21) for the displacement u
(1)
i
:
(1
ij
+T
ij
)u
(1)
j
= 
ij
g
j
(2.23)
• Step 2: Use the traction HBIE (2.10) to obtain τ
(2)
i
:
τ
(2)
i
= T
(N)
ij
g
j
−o
(N)
ij
u
(1)
j
(2.24)
Deﬁne the hypersingular residual:
r
(τ)
i
= τ
(1)
i
−τ
(2)
i
= g
i
−τ
(2)
i
(2.25)
Also, the error in the displacement is deﬁned as:
e
u(1)
i
= u
(1)
i
−u
i
(2.26)
One can now show that:
r
(τ)
i
(2.25)
= g
i
−τ
(2)
i
(2.24)
= g
i
−(T
(N)
ij
g
j
−o
(N)
ij
u
(1)
j
)
(2.26)
= g
i
−T
(N)
ij
g
j
+o
(N)
ij
(u
j
+ e
u(1)
j
)
(2.10)
= o
(N)
ij
e
u(1)
j
so that:
r
(τ)
i
= S
(N)
ij
e
u(1)
j
(2.27)
Theorem 2 The hypersingular traction residual bounds the error in the dis
placement globally:
C
1
r
(τ)
i
 ≤ e
u(1)
i
 ≤ C
2
r
(τ)
i

Proof: The proof is quite analogous to that of Theorem 1. It follows from
using equation (2.27).
© 2005 by Taylor & Francis Group, LLC
30 CHAPTER 2. ERROR ESTIMATION
2.2.2.2 The displacement residual
In the traction boundary condition problem, the unknown is the displacement
but equation (2.27) relates the traction residual to the error in the displacement.
It is proved below, however, that the traction residual is also equal to a suitably
deﬁned displacement residual for this problem.
The HBIE (2.10), with u
i
added to both sides of it, and upon rearrangement,
becomes:
u
i
= (1
ij
−o
(N)
ij
)u
j
−(1
ij
−T
(N)
ij
)g
j
(2.28)
Let u
(1)
i
be the solution of the BIE (2.9). Iterate (2.28) with this solution
and deﬁne:
u
(2)
i
= (1
ij
−o
(N)
ij
)u
(1)
j
−(1
ij
−T
(N)
ij
)g
j
(2.29)
Deﬁne the displacement residual:
r
(u)
i
≡ u
(1)
i
−u
(2)
i
(2.30)
One can now show that:
r
(u)
i
≡ u
(1)
i
−u
(2)
i
(2.29)
= u
(1)
i
−(1
ij
−o
(N)
ij
)u
(1)
j
+ (1
ij
−T
(N)
ij
)g
j
= o
(N)
ij
u
(1)
j
+ g
i
−T
(N)
ij
g
j
(2.10)
= o
(N)
ij
u
(1)
j
−o
(N)
ij
u
j
(2.26)
= o
N
ij
e
u(1)
j
(2.27)
= r
(τ)
i
(2.31)
so that:
r
(u)
i
= r
(τ)
i
(2.32)
and r
(τ)
i
can be replaced by r
(u)
i
in Theorem 2 !
Remark 2 An analogous result in potential theory appears in Menon et al.
[96]
2.2.3 Problem 3 : Mixed Boundary Conditions
The general boundary value problem in linear elasticity is:
Solve the NavierCauchy equations:
© 2005 by Taylor & Francis Group, LLC
2.2. ITERATED HBIE AND ERROR ESTIMATION 31
(λ + µ)∇(∇ u) + µ∇
2
u = 0 in B
subject to the boundary conditions:
Au +Bt = f on ∂B
where the matrices A and B and the vector f are prescribed quantities.
This class of problems is the most commonly encountered one in linear elas
ticity. Indeed, the numerical examples presented in Section 2.4 of this chapter
all have mixed boundary conditions imposed upon them. In this case, however,
a heuristic approach to error estimation is adopted here.
2.2.3.1 Traction residual
One computes the traction components τ
(1)
j
on ∂B by solving the primary BIE
(2.9) and then obtains the HBIE iterate τ
(2)
j
from the HBIE (2.10). As before,
the traction residual is deﬁned as:
r
(τ)
i
= τ
(1)
i
−τ
(2)
i
(2.33)
The corresponding pointwise error measure is as follows. At a ﬁxed bound
ary point, if the traction is speciﬁed in one direction, and the displacement in
the other, then the error in the boundary data is the error in displacement in
the ﬁrst direction, and the error in traction in the second direction. This issue
is discussed further in Section 2.3 of this chapter.
2.2.3.2 Stress residual
The stress residual is another important quantity in this work. The primary
BIE (2.9) is solved ﬁrst. This yields the boundary tractions and displacements
τ
(1)
j
and u
(1)
j
. The boundary stresses σ
(1)
ij
are next obtained from the boundary
values of the tractions and the tangential derivatives of the displacements, to
gether with Hooke’s law. This is a wellknown procedure in the BIE literature
(see, for example, Mukherjee [98] or Sladek and Sladek [151]).
Next, the iterated boundary stress is obtained from the HBIE (1.36) as
follows:
σ
(2)
ij
= T
ijk
τ
(1)
k
−o
ijk
u
(1)
k
(2.34)
where the required operators are deﬁned in equations (2.5) and (2.6) and the
LTB of the above operators are used in equation (2.34). Also, equation (2.34)
is collocated only at regular boundary points (where the boundary is locally
smooth) inside boundary elements.
© 2005 by Taylor & Francis Group, LLC
32 CHAPTER 2. ERROR ESTIMATION
One now gets the error in stress, for the BIE and the HBIE iterate, respec
tively, as:
e
s(1)
ij
= σ
(1)
ij
−σ
ij
(2.35)
e
s(2)
ij
= σ
(2)
ij
−σ
ij
(2.36)
and the stress residual is deﬁned as:
r
(s)
ij
= σ
(1)
ij
−σ
(2)
ij
(2.37)
Remark 3 The stress residual, deﬁned above in equation (2.37), can also be
used for problems with displacement or traction boundary conditions, which are
special cases of problems with mixed boundary conditions.
2.3 ElementBased Error Indicators
The main objective of error estimation is the development of suitable element
error indicators, which are denoted by η
i
. These indicators should satisfy the
following criteria:
C
1
η
i
≤ e
A(∂B
i
)
≤ C
2
η
i
(2.38)
D
1
N
i=1
η
2
i
≤ e
2
A
≤ D
2
N
i=1
η
2
i
(2.39)
where A is a suitable norm, A(∂B
i
) denotes the restriction of this norm to the
ith element, and C
1
, C
2
, D
1
and D
2
are appropriate constants. It is often
diﬃcult to prove these properties analytically, and one usually takes recourse
to numerical experiments. As in the potential theory case [96], this method
leads to two natural error indicators. The ﬁrst is based on the traction residual
deﬁned in equation (2.17); the second is based on the stress residual deﬁned in
equation (2.37).
These pointwise error measures may be used to deﬁne element error indica
tors. The following are proposed : the ﬁrst based on the traction residual and
the second on the stress residual:
η
(τ)
j
:= r
(τ)
i

L
2
(∂B
j
)
(2.40)
η
(s)
j
:= r
(s)
k

L
2
(∂B
j
)
(2.41)
Note that the subscript j refers to the jth element  the error indicator is a
scalar, not a vector. The L
2
norm is used for convenience and other norms can
be used, if desired.
© 2005 by Taylor & Francis Group, LLC
2.4. NUMERICAL EXAMPLES 33
The error estimates, as deﬁned above, do not depend directly on the bound
ary conditions on an element. The traction residual has been shown to be
related to the pointwise error in the boundary unknowns for Dirichlet and Neu
mann problems in elasticity (Section 2.2, Theorems 1 and 2). Note that, even
though the traction residual uses the diﬀerence in the primary and iterated trac
tions, but no explicit information about the displacement, it has been proved
in Section 2.2 that the traction residual is equal to the displacement residual
for a traction prescribed boundary value problem.
In general, at a local level, on a particular element, the actual error will
depend on the boundary conditions. In mixed boundary value problems, for
instance, the traction may be prescribed in the x
1
direction and the displace
ment in the x
2
direction at a boundary point. The errors are, therefore, in
the displacement in the x
1
direction, and in the traction in the x
2
direction.
Ideally, the traction residualbased element error indicator will capture the L
2
norm of these errors on an element, even for mixed boundary value problems.
The stress residual is also used as a measure of the error in stress on the
boundary. At any boundary point in a 2D problem, at most two components
of the stress are known from the prescribed boundary conditions. Thus, there is
always some error in a computed stress tensor at a boundary point. Numerical
experiments presented below (Section 2.4) suggest that this error is eﬀectively
tracked by the stress residualbased error indicator.
2.4 Numerical Examples
Two basic problems from the theory of planar elasticity are considered in this
section. The numerical implementation consists of two modules: a standard
code for twodimensional elastostatics, and a set of routines that calculate the
hypersingular residual for error estimation. For the ﬁrst part (i.e. the BIE), a
code due to Becker [9] is employed. This code uses collocation with quadratic
isoparametric elements. Numerical integration is done using Gaussian quadra
ture, except on elements that contain the collocation point. Singular integration
is avoided using the rigidbody mode, i.e., diagonal terms are evaluated by sum
ming the oﬀdiagonal terms. For the second part of the code (i.e. the HBIE),
collocation is carried out at points on the boundary where it is locally smooth,
and which are inside boundary elements, in order to determine the components
of stress using the traction HBIE. The numerical method used for evaluation
of the necessary hypersingular integrals here is due to Guiggiani [60]. In the
following examples, the stress tensor at a boundary point is evaluated by using
the HBIE (1.36) at three boundary points inside a boundary element, and then
a quadratic polynomial is employed to approximate the stress components over
each element.
© 2005 by Taylor & Francis Group, LLC
34 CHAPTER 2. ERROR ESTIMATION
2.4.1 Example 1: Lam´e’s Problemof a ThickWalled Cylin
der under Internal Pressure
Consider an inﬁnitely long hollow cylinder subjected to an internal pressure
p = 1. The inner radius r
i
= 3, and the outer radius r
o
= 6. Material properties
are also chosen of O(1), namely Young’s modulus, E = 1.0, and Poisson’s ratio
ν = 0.3. Consistent units are assumed throughout this paper.
Symmetry is employed and the problem is formulated as a mixed boundary
value problem on a quarter of the cylinder. The mesh used to solve this problem
is shown in Figure 2.1(a). Notice that the mesh is not biased a priori in the
sense that the element density is not increased on parts of the boundary where
the error is expected to be high.
*
: BIE collocation nodes
o : HBIE collocation nodes
(a)
0 5 10 15 20 25 30 35 40
1
0.5
0
0.5
1
1.5
2
Collocation point
o:stress from HBIE
+:stress from BIE
:exact solution
σ
1
1
__
(b)
Figure 2.1: (a) Discretized domain for the Lam´e problem with 12 elements. (b)
Comparison of analytical and numerical solutions for σ
11
(from [127])
Figure 2.1(b) presents a comparison of the computed and analytical solu
tions for the stress component σ
11
at the collocation points used for the HBIE.
Note that the continuous line for the exact solution is just used as a matter
of convenience and does not have any meaning except at discrete points, since
the xaxis is the collocation point number. The results for σ
12
and σ
22
display
similar accuracy and are not shown here.
More importantly, a pointwise comparison between the absolute value of the
error and pointwise measurements of the hypersingular residual is considered
next. Unlike the numerical examples in potential theory [96], the hypersingular
residual and error are often of opposite signs in this elasticity example. Also,
it is seen that the residual is not an upper bound as it underestimates the
error at some points. Since the stress residual is a symmetric tensor with three
independent components, a comparison between the pointwise error and stress
residual in each direction is carried out here. The error in stress, for the BIE
and the HBIE iterate, and the stress residual, are deﬁned in (2.35  2.37).
© 2005 by Taylor & Francis Group, LLC
2.4. NUMERICAL EXAMPLES 35
0 5 10 15 20 25 30 35 40
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
Collocation point
V
a
r
i
o
u
s
q
u
a
n
t
i
t
i
e
s
r
e
l
a
t
e
d
t
o
σ
o: error in the BIE solution
+: error in the HBIE iterate
: hypersingular residual
__
1
1
(a)
0 5 10 15 20 25 30 35 40
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
Collocation point
V
a
r
i
o
u
s
q
u
a
n
t
i
t
i
e
s
r
e
l
a
t
e
d
t
o
σ
1
2
o: error in the BIE solution
+: error in the HBIE iterate
: hypersingular residual
__
(b)
Figure 2.2: Absolute values of error in the BIE solution, error in the HBIE
iterate, and the hypersingular residual, for (a) σ
11
and (b) σ
12
. All variables
are unscaled (from [127])
Now consider a comparison of absolute values of errors, and the residual
in σ
11
in Figure 2.2(a), and in σ
12
in Figure 2.2(b). It is seen that the stress
residual provides good pointwise tracking of the error on a relatively coarse
mesh.
Of most practical importance (e.g. in adaptivity) is the performance of ele
ment error indicators. In particular, the performance of the two indicators η
(τ)
j
and η
(s)
j
deﬁned in equations (2.40) and (2.41), respectively, is studied here.
The ﬁrst is a traction residualbased error indicator, and the second uses the
stress residual.
The element error indicator based on the traction residual (η
(τ)
j
from equa
tion (2.40)) is compared with the elementbased L
2
norm of the error in the
unspeciﬁed boundary data in Figure 2.3(a). On the other hand, the element
error indicator based on the stress residual (η
(s)
j
from equation (2.41)) is com
pared to the elementbased L
2
norm of the error in stress, on all the boundary
elements, in Figure 2.3(b). The stress residuals are seen to capture the error
trends quite eﬀectively.
Remark 4 The comparison between the traction residual and error in the dis
placement is diﬃcult unless one uses normalized values. A simple way to do
this is to use nondimensional quantities to begin with. This is the approach
followed in this work.
© 2005 by Taylor & Francis Group, LLC
36 CHAPTER 2. ERROR ESTIMATION
0 2 4 6 8 10 12
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
Element number
element error indicator
L norm of error
2
E
r
r
o
r
i
n
d
i
c
a
t
o
r
a
n
d
e
r
r
o
r
(a)
0 2 4 6 8 10 12
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
Element number
L norm of error
element error indicator
2
E
r
r
o
r
i
n
d
i
c
a
t
o
r
a
n
d
e
r
r
o
r
(b)
Figure 2.3: Comparisons of element error indicators with actual errors, for (a)
traction residualbased error with error in the unspeciﬁed boundary data, and
(b) stress residualbased error with error in boundary stresses (from [127])
2.4.2 Example 2: Kirsch’s Problem of an Inﬁnite Plate
with a Circular Cutout
Consider an inﬁnite plate with a circular cutout of radius 1, subject to a traction
τ
1
= 1 at inﬁnity. Material properties are the same as in the previous example.
The displacement and stress ﬁelds for this problem may be found in Tim
oshenko and Goodier [167]. In order to simulate this problem with a ﬁnite
geometry, the boundaries of a ﬁnite square domain are subjected to tractions
computed from the exact solution of an inﬁnite plate subjected to traction at
inﬁnity. Using symmetry, only a quarter of the plate is used in the computer
model. The mesh is shown in Figure 2.4.
Pointwise comparisons between the errors (in the BIE solution and in the
HBIE iterate), and the residual, in some components of the stress, are presented
ﬁrst. Figures 2.5(a) and 2.5(b) show the absolute values of these errors in two
stress components, together with the corresponding stress residuals. Of course,
the error in BIE stress component is zero at points where that particular stress
component is prescribed as a boundary condition.
The more important comparison is between the computed element error
indicators and the L
2
norms of the error on each element (Figure 2.6). Figure
2.6(a) uses errors in the unspeciﬁed boundary data while Figure 2.6(b) uses
errors in stress components. It is seen that the tractionbased error indicator
underestimates the error on some elements. The stressbased error indicator
performs better and accurately captures the error in stress on most of the
elements.
In conclusion, the error estimation method presented here has the advantage
of capturing errors in the stress ﬁeld. On the other hand, the usual residual
© 2005 by Taylor & Francis Group, LLC
2.4. NUMERICAL EXAMPLES 37
HBIE collocation nodes
*
:
BIE collocation nodes
o:
Figure 2.4: Discretized domain for the problem of a plate with a cutout with
10 elements. Plate side is 4 units and cutout radius is 1 unit of length (from
[127])
0 5 10 15 20 25 30
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Collocation point
σ
: error in the BIE solution
: error in the HBIE iterate
: hypersingular residual
__
V
a
r
i
o
u
s
q
u
a
n
t
i
t
i
e
s
r
e
l
a
t
e
d
t
o
o
+
1
2
(a)
0 5 10 15 20 25 30
0
0.1
0.2
0.3
0.4
0.5
0.6
Collocation point
σ
: error in the BIE solution
: error in the HBIE iterate
: hypersingular residual
+
__
V
a
r
i
o
u
s
q
u
a
n
t
i
t
i
e
s
r
e
l
a
t
e
d
t
o
o
2
2
(b)
Figure 2.5: Absolute values of error in the BIE solution, error in the HBIE
iterate, and the hypersingular residual for (a) σ
12
and (b) σ
22
. All variables are
unscaled (from [127])
© 2005 by Taylor & Francis Group, LLC
38 CHAPTER 2. ERROR ESTIMATION
1 2 3 4 5 6 7 8 9 10
0
0.05
0.1
0.15
0.2
0.25
0.3
Element number
o: element error indicator
+: L norm of error
2
E
r
r
o
r
i
n
d
i
c
a
t
o
r
a
n
d
e
r
r
o
r
(a)
1 2 3 4 5 6 7 8 9 10
0
0.1
0.2
0.3
0.4
0.5
0.6
Element number
E
r
r
o
r
i
n
d
i
c
a
t
o
r
a
n
d
e
r
r
o
r
o: element error indicator
+: L norm of error
2
(b)
Figure 2.6: Comparisons of element error indicators with actual errors, for (a)
traction residualbased error with error in the unspeciﬁed boundary data, and
(b) stress residualbased error with error in boundary stresses (from [127])
techniques may only be used to compute the error in displacements. On phys
ical grounds, a measure of the error in stress is preferable to a measure of
displacement error. It may also be viewed as a stronger measure of convergence
 i.e. the approximate displacement ﬁeld, and its gradient have converged to
the actual solution in this case.
© 2005 by Taylor & Francis Group, LLC
Chapter 3
THIN FEATURES
Many boundary value problems that are solved by the BEM involve thin fea
tures. Common examples are cracks and thin plates and shells. This chapter
ﬁrst presents BEM formulations for 3D potential theory in a region exterior to
thin plates, for applications in microelectromechanical systems (MEMS). This
is followed by a discussion of crack problems (linear elastic fracture mechanics
 LEFM) in 3D linear elasticity.
3.1 Exterior BIE for Potential Theory: MEMS
Exterior BIEs for potential theory, suitable for applications in MEMS, are pre
sented in this section. Numerical results for a simple system with two thin
conducting plates follow. Further details are available in [6].
3.1.1 Introduction to MEMS
The ﬁeld of microelectromechanical systems (MEMS) is a very broad one
that includes ﬁxed or moving microstructures; encompassing microelectro
mechanical, microﬂuidic, microoptoelectromechanical and microthermal
mechanical devices and systems. MEMS usually consists of released microstruc
tures that are suspended and anchored, or captured by a hubcap structure and
set into motion by mechanical, electrical, thermal, acoustical or photonic energy
source(s).
Typical MEMS structures consist of arrays of thin beams with crosssections
in the order of microns (µm) and lengths in the order of ten to hundreds of
microns. Sometimes, MEMS structural elements are plates. An example is a
small rectangular silicon plate with sides in the order of mm and thickness of
the order of microns, that deforms when subjected to electric ﬁelds. Owing
to its small size, signiﬁcant forces and/or deformations can be obtained with
the application of low voltages (≈ 10 volts). Examples of devices that utilize
vibrations of such plates are synthetic microjets ([142, 8]  for mixing, cooling
39
© 2005 by Taylor & Francis Group, LLC
40 CHAPTER 3. THIN FEATURES
of electronic components, micropropulsion and ﬂow control), microspeakers [71]
etc.
Numerical simulation of electrically actuated MEMS devices have been car
ried out for around a decade or so by using the BEM to model the exterior
electric ﬁeld and the FEM (see, e.g. [179, 190, 66]) to model deformation of
the structure. The commercial software package MEMCAD [147], for example,
uses the commercial FEM software package ABAQUS for mechanical analysis,
together with a BEM code FastCap [112] for the electric ﬁeld analysis. Other
examples of such work are [49, 148, 1]; as well as [147, 149] for dynamic analysis
of MEMS.
The focus of this section is the BEM analysis of the electric ﬁeld exterior to
very thin conducting plates. A convenient way to model such a problem is to
assume plates with vanishing thickness and solve for the sum of the charges on
the upper and lower surfaces of each plate [61]. The standard BIE with a weakly
singular kernel is used here and this approach works well for determining, for
example, the capacitance of a parallel plate capacitor. For MEMS calculations,
however, one must obtain the charge densities separately on the upper and lower
surfaces of a plate since the traction at a surface point on a plate depends on the
square of the charge density at that point. The gradient BIE is employed in the
present work to obtain these charge densities separately. Careful regularization
of the gradient equation, to take care of singular and nearly singular integrals
that arise, is the principal contribution of the present work. The work of Liu
[87], on thin shells, is of great value to the research reported here. Gray [53] and
Nishimura and his coworkers [118, 180] have considered the 3D Laplace equa
tion in a region exterior to a narrow slit or crack. Gray addresses applications
in electroplating problems. The research described in Refs. [118, 180] is quite
diﬀerent from the problem under consideration here. Their primary interest
is in the crack opening displacement with zero normal displacement derivative
on the crack faces, while the separate charge densities (that are proportional
to the normal derivative of the potential) are of interest in this chapter. The
formulation given in the present work is a BEM scheme that is particularly well
suited for MEMS analysis of very thin plates  for h/L ≤ .001  in terms of the
length L (of a side of a square plate) and its thickness h. A similar approach has
also been developed for MEMS and nanoelectromechanical systems (NEMS)
with very thin beams [7], but this work is not presented in this book.
As a byproduct of the development of the thin plate BEM, an enhanced
BEM, suitable for MEMS analysis of moderately thick plates, has also been
developed in this work. It is shown that accurate evaluation of weakly singular
and nearly weakly singular integrals plays a key role here.
This section is organized as follows. The usual and gradient BIEs for po
tential theory, in an inﬁnite region exterior to a structure composed of thin
conducting plates, are ﬁrst presented and regularized. Singular and nearly sin
gular integrals, both weak and strong, are discussed next. Numerical results are
presented and discussed for a model problem (a parallel plate capacitor) from
three methods  the usual BEM, the enhanced BEM and the thin plate BEM.
© 2005 by Taylor & Francis Group, LLC
3.1. EXTERIOR BIE FOR POTENTIAL THEORY: MEMS 41
3.1.2 Electric Field BIEs in a Simply Connected Body
First consider the solution of Laplace’s equation in a threedimensional (3D)
simply connected body.
3.1.2.1 Usual BIE  indirect formulation
Referring to Figure 1.1, for a source point ξ ∈ B (with bounding surface ∂B),
one has the usual indirect BIE:
φ(ξ) =
∂B
ν(y)
4πr(ξ, y)
ds(y) (3.1)
where φ is the potential, r(ξ, y) = y −ξ, r = r, is the dielectric constant of
the medium and ν is the (unknown) surface density function on ∂B.
3.1.2.2 Gradient BIE  indirect formulation
Taking the gradient of φ at the source point ξ results in:
∇
ξ
φ(ξ) =
∂B
ν(y)
4π
∇
ξ
1
r(ξ, y)
ds(y) =
∂B
ν(y)r(ξ, y)
4πr
3
(ξ, y)
ds(y) (3.2)
Alternatively, one can write (3.2) as:
∂φ
∂ξ
k
(ξ) =
∂B
ν(y)(y
k
−ξ
k
)
4πr
3
(ξ, y)
ds(y) (3.3)
Note that, in general, the function ν(y) is not the charge density. It be
comes equal to the charge density when B is the inﬁnite region exterior to the
conductors. This is discussed in Section 3.1.3.
3.1.3 BIES in Inﬁnite Region Containing Two Thin Con
ducting Plates
Now consider the situation shown in Figure 3.1. Of interest is the solution of
the following Dirichlet problem for Laplace’s equation:
∇
2
φ(x) = 0, x ∈ B, φ(x) prescribed for x ∈ ∂B (3.4)
where B is now the region exterior to the two plates. The unit normal n to ∂B
is deﬁned to point away from B (i.e. into a plate).
© 2005 by Taylor & Francis Group, LLC
42 CHAPTER 3. THIN FEATURES
+
1
+ +
+
1
x
σ
1
V
n plate 1


1
x
σ
σ
+
2
1
V 
plate 2
+
σ

2
ground plane
0 = V
ξ
^




x
~
+
g
n
h
s
s
s
s
s
+
^
s
2




Figure 3.1: Parallel plate capacitor with two plates (from [6])
3.1.3.1 Regular BIE  source point approaching a plate surface S
+
1
Let
ˆ
S
1
+
⊂ S
+
1
be a small neighborhood of x
+
. As ξ → x
+
∈
ˆ
S
1
+
⊂ S
+
1
(see
Figure 3.1), one has:
φ(x
+
) =
S
+
1
−
ˆ
S
1
+
β(y)
4πr(x
+
, y)
dS(y) +
ˆ
S
1
+
β(y)
4πr(x
+
, y)
dS(y)
+
S
+
2
β(y)
4πr(x
+
, y)
dS(y) (3.5)
Here β(y) = σ(y
+
) + σ(y
−
), where σ is now the charge density at a point
on a plate surface. The second integral above is weakly singular, while the rest
are usually regular. It should be noted, however, that the last integral above
becomes nearly weakly singular when both h and g are small.
A similar equation can be written for x
+
∈ S
+
2
. For the case shown in
Figure 3.1, however, this is not necessary since β(y) is equal and opposite on
the two plates. Therefore, for this case, equation (3.5) is suﬃcient to solve for
β on both the plates !
3.1.3.2 Gradient BIE  source point approaching a plate surface S
+
1
The governing equation. It is ﬁrst noted that for x
+
∈ S
+
k
∪S
−
k
, k = 1, 2:
σ(x) =
∂φ
∂n
(x) = n(x) · [∇
ξ
φ(ξ)]
ξ=x
(3.6)
© 2005 by Taylor & Francis Group, LLC
3.1. EXTERIOR BIE FOR POTENTIAL THEORY: MEMS 43
Consider the limit ξ → x
+
∈
ˆ
S
1
+
⊂ S
+
1
. It is important to realize that
this limit is meaningless for a point x on the edge of a plate, since the charge
density is singular on its edges. One has:
σ(x
+
) =
S
+
1
−
ˆ
S
1
+
β(y)r(x
+
, y) · n(x
+
)
4πr
3
(x
+
, y)
dS(y)
+
ˆ
S
1
+
r(x
+
, y) · [β(y)n(x
+
) −β(x)n(y)]
4πr
3
(x
+
, y)
dS(y)
+
β(x)
4π
Ω(
ˆ
S
1
+
, x
+
) +
S
+
2
β(y)r(x
+
, y) · n(x)
4πr
3
(x
+
, y)
dS(y) (3.7)
x
y
r
(
x
,
y
)
+
+
+
+
θ
φ
n
ψ
Z axis
1
s
ξ
L
Figure 3.2: Line integral for evaluation of solid angle (from [6])
In the above, the solid angle subtended by the surface element
ˆ
S
1
+
at the
point x
+
is (see [87, 105] and Figure 3.2):
Ω(
ˆ
S
1
+
, x
+
) =
ˆ
S
1
+
=
r(x
+
, y) · n(y)
r
3
(x
+
, y)
dS(y) =
2π
0
[1 −cos(ψ(θ))]dθ (3.8)
where the symbol
= denotes the Finite Part (FP) of the integral in the sense of
Mukherjee [101, 102].
Equations (3.7) and (3.8) give the ﬁnal equation:
1
2
[σ(x
+
) −σ(x
−
)] =
S
+
1
−
ˆ
S
1
+
β(y)r(x
+
, y) · n(x
+
)
4πr
3
(x
+
, y)
dS(y)
© 2005 by Taylor & Francis Group, LLC
44 CHAPTER 3. THIN FEATURES
+
ˆ
S
1
+
r(x
+
, y) · [β(y)n(x
+
) −β(x)n(y)]
4πr
3
(x
+
, y)
dS(y)
−
β(x)
4π
2π
0
cos(ψ(θ))dθ +
S
+
2
β(y)r(x
+
, y) · n(x
+
)
4πr
3
(x
+
, y)
dS(y) (3.9)
Here (see Figure 3.2), a local coordinate system (x, y, z) is set up with the
origin at x
+
such that the positive z axis intersects the surface
ˆ
S
1
+
. Now, ψ
is the angle between the positive z axis and r(x
+
, y) with y ∈
ˆ
L
+
, and θ the
angle between the positive x axis and the projection of r(x
+
, y) in the xy plane.
In the above, the second integral on the righthand side is weakly singular,
while the rest are usually regular. The last integral above, however, becomes
nearly strongly singular if both the thickness h and the gap g are small. Once β
is known on both plates, (3.9) can be used, as a postprocessing step, to obtain
σ
+
and σ
−
on both plates.
Equation (3.9) for ﬂat plates. First, it should be emphasized that equa
tions (3.5) and (3.9) are valid for thin curved shells as well as for ﬂat plates.
For a pair of symmetric ﬂat plates (see Figure 3.1), the ﬁrst, second and third
integrals on the righthand side of (3.9) vanish, and one is left with the simple
equation:
1
2
[σ(x
+
) −σ(x
−
)] =
S
+
2
β(y)r(x
+
, y) · n(x
+
)
4πr
3
(x
+
, y)
dS(y) (3.10)
Equation (3.10) implies that σ(x
+
) = σ(x
−
) if one has only one plate. This,
of course, is true. The existence of the second plate in Figure 3.1 is the reason
for (in general) σ(x
+
) = σ(x
−
).
3.1.3.3 Two plates very close together
For cases in which the gap g between the thin plates in Figure 3.1 is also of
the order of the (small) plate thickness, the last integral on the righthand side
of equation (3.5) must be treated as nearly weakly singular. In this case, this
integral should be written as:
S
+
2
β(y)
4πr(x
+
, y)
dS(y) =
S
+
2
−
ˆ
S
2
+
β(y)
4πr(x
+
, y)
dS(y)
+
ˆ
S
2
+
β(y) −β(˜ x)
4πr(x
+
, y)
dS(y) +
β(˜ x)
4π
ˆ
S
2
+
1
r(x
+
, y)
dS(y) (3.11)
where ˜ x
+
∈
ˆ
S
2
+
, ˜ x
−
∈
ˆ
S
2
−
and β(˜ x) = σ(˜ x
+
) + σ(˜ x
−
). The ﬁrst and second
integrals on the righthand side of (3.11) are regular. (The second integral is
O(˜ r/r) where ˜ r = y − ˜ x
+
. As ˜ r → 0, r → g + h, so that this integrand
© 2005 by Taylor & Francis Group, LLC
3.1. EXTERIOR BIE FOR POTENTIAL THEORY: MEMS 45
→0.) The last integral is nearly singular. A procedure for accurate evaluation
of nearly singular integrals is presented in Section 3.1.4.
Also, the last integral on the righthand side of (3.9) now becomes nearly
strongly singular. This integral, called J, can be evaluated as follows. One can
write:
J =
S
+
2
−
ˆ
S
2
+
β(y)r(x
+
, y) · n(x
+
)
4πr
3
(x
+
, y)
dS(y)
+
ˆ
S
2
+
r(x
+
, y) · [β(y)n(x
+
) −β(˜ x)n(y)]
4πr
3
(x
+
, y)
dS(y)
+
β(˜ x)
4π
Ω(
ˆ
S
2
+
, x
+
) (3.12)
where (see Figure 3.2):
Ω(
ˆ
S
2
+
, x
+
) =
ˆ
S
2
+
r(x
+
, y) · n(y)
r
3
(x
+
, y)
dS(y) =
2π
0
[1 −cos(ψ(θ))]dθ (3.13)
It is noted that, in this case, the point x
+
is slightly above
ˆ
S
2
+
and that
the second term in (3.13) denotes a “nearly FP” integral.
The idea of regularizing (3.12) with β(˜ x) has been inspired by earlier work
on evaluation of nearly singular integrals [104].
x
x
~
+
+
+
x
+
n( )
x
+
n( )
~
s
1
+
s
2
V=0
y
.
.
.
ground plane
Figure 3.3: Symmetric deformation of two plates (from [6])
Let the integrals on the righthand side of (3.12) be called J
1
, J
2
, J
3
. Each
of the three integrals is regular. The fact that the second integral J
2
is regular
can be proved as follows.
© 2005 by Taylor & Francis Group, LLC
46 CHAPTER 3. THIN FEATURES
It is assumed that the two plates always remain symmetric with respect to
the ground plane, even after deformation (see Figure 3.3). Their equations,
therefore, are of the form x
3
= ±f(x
1
, x
2
). One now has:
n(x
+
) ∝ −k +if
,1
+jf
,2
, n(˜ x
+
) ∝ −k −if
,1
−jf
,2
(3.14)
As y → ˜ x
+
,
[β(y)n(x
+
) −β(˜ x)n(y)] ∝
k[β(˜ x) −β(y)] + [ia +jb][β(˜ x) +β(y)] (3.15)
where i, j, k are Cartesian unit vectors and a and b are some numbers.
As y → ˜ x
+
, r(x
+
, y) ∝ −k, so that the integrand of J
2
is O(˜ r/r
2
) where
˜ r = y −˜ x
+
. In this limit, ˜ r →0, r →g +h, so that the integrand of J
2
→0.
3.1.3.4 Consistency check
It is interesting to examine the forms of equations (3.5) and (3.9) when col
located at x
−
on S
−
1
(see Figure 3.1). Equation (3.5) yields φ(x
+
) = φ(x
−
).
Referring to equation (3.7), its lefthand side becomes σ(x
−
), and the signs of
the ﬁrst, second and fourth terms on its righthand side change (since n(x
−
) =
−n(x
+
) and n(y) for y ∈ S
+
1
equals −n(y) for y ∈ S
−
1
). The solid angle
expression (see (3.8)) now becomes:
Ω(
ˆ
S
1
−
, x
−
) =
2π
0
[1 + cos(ψ(θ))]dθ (3.16)
Using these facts, it is easy to show that, as expected, equation (3.9) remains
unchanged when collocated at x
−
!
3.1.4 Singular and Nearly Singular Integrals
Certain BEM integrals require special care for thin plates and when thin plates
come close together. The usual BEM must deal with weakly and nearly weakly
singular integrals in such cases, while the thin plate BEM must deal with both
nearly weakly singular (the last integral on the righthand side of (3.5)) and
nearly strongly singular (the last integral on the righthand side of (3.9)) in
tegrals; as well as weakly and strongly singular integrals. The weakly singular
case involving this kernel has been addressed before by many authors (see, e.g.
[114, 26]). The nearly (also called quasi) weakly singular case, along with other
nearly singular integrals of various orders, can be eﬀectively evaluated by em
ploying a cubic polynomial transformation due to Telles [165] and Telles and
Oliveira [166]. Several other authors have also considered similar problems (e.g.
[64, 104])  many of these references are available in [165, 114, 166, 104] and
are not repeated here in the interest of brevity. A new simple approach for
evaluation of nearly weakly singular integrals is presented below.
© 2005 by Taylor & Francis Group, LLC
3.1. EXTERIOR BIE FOR POTENTIAL THEORY: MEMS 47
h α α h
ˆ
ˆ
ˆ
ˆ
(a) (b)
r
r r
r
x
x
x
x
y y
Figure 3.4: Nearly weakly singular integrals (from [6])
Proposed new method for evaluating nearly weakly singular integrals.
A typical situation where one encounters weakly and nearly weakly singular
integrals in the usual BEM is shown in Figure 3.4. Consider a source point x on
the top face of a plate and its image point ˆ x on the bottom face in Figure 3.4(a).
Two kinds of singular (O(1/r)) integrals arise  a weakly singular integral on the
boundary element ∆ on the top face of the plate that contains x, and, since h is
small, a nearly weakly singular integral on the boundary element
ˆ
∆ (the image
of ∆) on the bottom face of the plate that contains ˆ x. The weakly singular
integral is evaluated by employing the methods outlined in [114, 26]. A nearly
weakly singular integral (see above and also the last term on the righthand
side of (3.11)) has the form:
I(x) =
ˆ
∆
σ(y)ds(y)
4πr(x, y)
(3.17)
The integrand above is multiplied by ˆ r/ˆ r with the result:
I(x) =
ˆ
∆
[σ(y)(ˆ r/r)]ds(y)
4πˆ r(ˆ x, y)
(3.18)
Since ˆ r/r is O(1) and →0 as y → ˆ x (i.e. as ˆ r →0), the integrand in (3.18)
is weakly singular, of O(1/ˆ r) as ˆ r → 0. Therefore, the integral (3.18) can be
evaluated by employing the methods described in [114, 26].
The source point x may also lie on a side face of the plate (in the 3D BEM)
as shown in Figure 3.4(b). The same idea (3.18) can be applied in this case as
well.
Performance of new method for nearly weakly singular integrals.
The performance of the new method is compared with that of standard Gauss
integration. Figure 3.5(a) shows the source point and region of integration (a
triangle). The triangle is purposely chosen to be fairly elongated. Numerical
results (for σ = 1 and = 1) appear in Figure 3.5(b). It is seen that for
h < 1/100, standard Gauss integration, even with 19 Gauss points, cannot re
duce the error below around 4%. The new method is seen to take care of these
nearly weakly singular integrals very well, even for very small values of h.
© 2005 by Taylor & Francis Group, LLC
48 CHAPTER 3. THIN FEATURES
3
x
h
2
x
1
x
(0,6,0)
(2,0,0)
.
.
(0,0,h) x
y
.
n(y)
(a)
10 10 10 10 10 10 1
0
1
2
3
4
5
6
7
8
.
: 7 Gauss points
: 19 Gauss points
: New method
Out of plane distance h
P
e
r
c
e
n
t
a
g
e
e
r
r
o
r
_ _ _ _ _ _
6 5 4 3 2 1
(b)
Figure 3.5: Nearly weakly singular integral over a triangle. (a) Schematic (b)
Errors in numerical results from various methods (from [6])
10
6
10
5
10
4
10
3
10
2
10
1
1
0.2
0
0. 2
0. 4
0. 6
0. 8
1
1. 2
1. 4
1. 6
1. 8
N
e
a
r
l
y
s
t
r
o
n
g
l
y
s
i
n
g
u
l
a
r
i
n
t
e
g
r
a
l
Out of plane distance h
_
_ _ _ _ _ _
Result from Maple
19 Gauss points
π/2
Figure 3.6: Nearly strongly singular integral (from [6])
© 2005 by Taylor & Francis Group, LLC
3.1. EXTERIOR BIE FOR POTENTIAL THEORY: MEMS 49
Nearly strongly singular integrals. When two plates get close together,
one encounters nearly strongly singular integrals of the form (see (3.13)):
I =
∆
r(x, y) · n(y)
r
3
(x, y)
ds(y) =
∆
h
r
3
(x, y)
ds(y) (3.19)
where the second expression above is the special version for the situation de
picted in Figure 3.5(a). It is absolutely critical that this integral be evaluated
carefully. The pathetic result of applying standard Gauss integration (even with
19 Gauss points) appears in Figure 3.6. The correct results in this ﬁgure are
obtained from the symbolic computational code Maple [65]. A Stoke transfor
mation to convert this nearly strongly singular surface integral to a line integral
[87] is recommended in this work (see (3.13)).
3.1.5 Numerical Results
3.1.5.1 Three BEM Models
Three BEM models, the usual, the enhanced and the thin plate, are of concern
to this work. These are brieﬂy described below.
Usual BEM. The usual BEM is the standard version with weakly singular
integrals evaluated by the method outlined in [26]. For the sake of simplicity
and consistency between diﬀerent BEM models, only uniform BEM meshes,
composed of T6 triangles, are used on the square faces of objects in this work.
It is well known that the charge density is singular on the edges of a plate.
Therefore, nonconforming boundary elements should be used and collocation
points should not be placed at plate edges. This procedure, however, makes
the BEM code somewhat cumbersome and expensive. For simplicity, regular
T6 elements are used on square plate faces, in all three types of BEM models,
with collocation points placed everywhere including on plate edges. Further, in
the interest of standardization, the usual BEM only uses T6 elements in this
work.
The usual BEM code is ﬁrst veriﬁed by solving for the charge distribution on
the surface of a unit cube, subjected to unit surface potential. The capacitance
(the total charge on the cube surface divided by the voltage) (with = 1) from
the BEM is 8.28 while that from FastCap [113] is 8.3. The BEM mesh for this
problem has an 8 × 8 array of squares on each face, each divided into two T6
triangles, for a total of 768 boundary elements.
Enhanced BEM. The enhanced BEM is designed to solve problems with
moderately thick plates accurately and eﬃciently. Like the usual BEM, weakly
singular integrals are again evaluated by the methods outlined in [26]. The
enhanced BEM has two additional features:
• Nearly weakly singular integrals are evaluated by the method outlined in
Section 3.1.4 above.
© 2005 by Taylor & Francis Group, LLC
50 CHAPTER 3. THIN FEATURES
• Detailed modeling of side faces for small values of h/L is not desirable
for two reasons. One reason is that this would either lead to triangular
boundary elements (on these side faces) with very large aspect ratios,
or, alternatively, to a prohibitively large number of boundary elements.
The second reason is a matter of eﬃciency. In the enhanced BEM, T3
elements are used on the side faces. This assures linear interpolation
of the charge density across a side face with no new unknowns being
introduced on these faces (and, therefore, no additional nearly weakly
singular integrals). Obviously, this idea is good for moderate values of
h/L. It is not a good idea for cubelike conductors. Also, the usual as well
as the enhanced BEM breaks down for very small values of h/L. This
issue is discussed further in Section 3.1.6 below.
Thin plate BEM. This is the model that has been presented in detail in
Section 3.1.3 of this chapter. In this case, nearly singular integrals only arise
when plates are very close together (i.e. for small values of g/L). Again, weakly
singular and nearly weakly singular integrals (when they arise) are evaluated
by the method outlined in [26], and in Section 3.1.4, respectively. Strongly
singular integrals and nearly strongly singular integrals (when they arise) are
evaluated by Stoke regularization (see equations (3.8) and (3.13)). The value
of g/L = 0.2 in many examples below. The corresponding integrals are treated
as regular in these cases.
Collocation points. It is important to point out that, in general, the usual
BEM requires collocation points on the entire bounding surface of a plate, the
enhanced BEM on the upper and lower surfaces, and the thin plate BEM only
on the upper surface. Of course, symmetry can be used, whenever possible, to
further reduce the number of collocation points.
3.1.6 The Model Problem  a Parallel Plate Capacitor
Numerical results are obtained from three methods  the usual, enhanced and
thin plate BEM. The model problem is a parallel plate capacitor with two
square plates (Figure 3.1). The length L of the side of each plate is unity. The
gap between the plates is g. The dielectric constant of the external medium, ,
is unity. The voltages on the upper and lower plates are V
1
= 1 and V
2
= −1,
respectively.
3.1.6.1 Capacitance evaluation from the thin plate BEM model
Results from the thin plate BEM model (with h/L = 10
−6
) are compared to
those from Harrington [61] in Figure 3.7. The capacitance C = Q/2V , where
Q is the total charge on the top plate and V is the potential on it. A = L
2
is
the area of each plate. Calculation of the capacitance, of course, only requires
equation (3.5). Harrington’s approach is essentially the same as the present one.
© 2005 by Taylor & Francis Group, LLC
3.1. EXTERIOR BIE FOR POTENTIAL THEORY: MEMS 51
The diﬀerences are that the weakly singular integral in (3.5) is now evaluated
from the method outlined in [26], and that [61] uses constant while the present
work uses quadratic (T6) elements. The present results are believed to be more
accurate than the older ones.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.5
1
1.5
2
2.5
3
3.5
Thin plate BEM
Harrington (1993)
g/L
Cg
Aε
__
Figure 3.7: Normalized capacitance from the thin plate model compared with
numerical results from [61]. Ref. [61] uses 36 constant square elements while
the present work uses 16 × 16 × 2 T6 elements on the upper surface of each
plate (from [6])
3.1.6.2 Comparison of various methods
The performance of the three BEM models  the usual, the enhanced and the
thin plate  are compared for the parallel plate capacitor problem (Figure 3.1)
in Table 3.1. Diﬀerent values of h/L are considered here while g/L is kept ﬁxed
at 0.2. Each method has diﬀerent regions where it excels  the enhanced BEM
being best suited for intermediate thicknesses and the thin plate BEM being
best for very thin plates.
Total charge. Table 3.1 is only concerned with the total charge on the plates
 therefore only with equation (3.5) of the thin plate BEM. It is ﬁrst observed
that there is a signiﬁcant discrepancy between the usual and enhanced BEM
results for h/L = 0.1. This is primarily due to the fact that the enhanced
BEM computes nearly weakly singular integrals by the method proposed in
Section 3.1.4 while the usual BEM employs standard Gaussian quadrature for
evaluating these integrals. With the given mesh, the enhanced BEM delivers
© 2005 by Taylor & Francis Group, LLC
52 CHAPTER 3. THIN FEATURES
Cg/A Computational eﬀort
h/L Usual Enhanced Thin plate Usual Enhanced Thin plate
BEM BEM BEM BEM BEM BEM
1 2.3975 1053.9
0.1 3.3542 2.6631 1.2351 95.7 99.7 4.2
0.05 1.7405 1.3879 100.4 4.2
0.01 1.6899 1.5611 101.2 4.2
0.005 1.6652 1.5874 100.1 4.2
0.001 1.6221 1.6094 100.1 4.2
10
−6
1.6200 4.2
Table 3.1: Summary of various cases for the parallel plate capacitor problem.
Mesh : All cases  on each (unit) square (top and bottom) plate face 8×8×2 T6
triangles. Usual BEM : for cube  each side face 8 × 8 × 2 T6 triangles; for
h/L = 0.1, each side face 8 × 2 T6 triangles. Enhanced BEM : each side face
16 ×2 T3 triangles. g/L = 0.2. Computational eﬀort in cpu seconds (from [6])
Cg/A = 2.6631. The same approach, with the same mesh, but with Gaussian
quadrature used to evaluate nearly weakly singular integrals, yields a value of
1.9785  a diﬀerence of over 25% ! The usual BEM model with 16 T6 triangles
on each side face has an additional 16 midside nodes on each side face that
generate additional nearly weakly singular integrals (see Figure 3.4(b)). When
all these nearly weakly singular integrals are evaluated with standard Gaussian
quadrature, the usual BEM yields Cg/A = 3.3542. (In fact, many of the
values of charge on the midside nodes on the side faces come out wrong  they
are negative !) The idea here is not to deliberately downgrade the usual BEM.
Rather, it is to emphasize that signiﬁcant errors can arise from the evaluation
of nearly weakly singular integrals by standard Gaussian quadrature. One tries,
of course, to compensate for the use of Gaussian quadrature for the evaluation
of nearly weakly singular integrals by using a ﬁne mesh, or a large number of
Gauss points, or both. As seen from Figure 3.5(b), however, just using a large
number of Gauss points may not help very much.
The enhanced and the thin plate BEM show best agreement for h/L = 0.001.
It is noted that even though the thin plate equations have no explicit dependence
on the plate thickness h, the ﬁnal results from it do depend on h because the
distance from S
+
1
to S
+
2
is g + h (see Figure 3.1) and the gap g is kept ﬁxed
here.
Separate charges. Perhaps even more interesting is Table 3.2, which com
pares how the separate charges σ
+
and σ
−
(on the top plate) are calculated by
the enhanced BEM and the thin plate BEM. Of course, equation (3.10) now
comes into play in the thin plate BEM. As is well known, accurate determi
nation of σ
+
and σ
−
is critical for MEMS applications since the traction on a
© 2005 by Taylor & Francis Group, LLC
3.1. EXTERIOR BIE FOR POTENTIAL THEORY: MEMS 53
Enhanced BEM Thin plate BEM
h/L = 0.01 h/L = 0.001 h/L = 0.01 h/L = 0.001
σ
+
at plate center 1.9972 4.5300 1.4539 1.4500
σ
−
at plate center 9.8761 12.150 9.4399 9.8600
s
+
1
σ
+
ds 0.3183 0.7904 0.4303 0.4348
s
−
1
σ
−
ds 1.2622 0.8182 1.1308 1.1746
Table 3.2: Comparison of enhanced BEM and thin plate BEM for the parallel
plate capacitor problem. Mesh : Both cases  on each (unit) square (top and
bottom) plate face 8 × 8 × 2 T6 triangles. Enhanced BEM : each side face
16 ×2 T3 triangles. g/L = 0.2 (from [6])
plate, at a surface point, depends on the square of the charge density at that
point. Agreement between the two methods is tolerable for h/L = 0.01. The
enhanced BEM, however, breaks down when determining the charges separately
for h/L = 0.001 (even though it successfully calculates the capacitance for this
value of h/L). The reason for this failure is that the usual (and enhanced) BIEs,
collocated at x
+
and x
−
in Figure 3.1, become nearly identical as h/L → 0.
This is analogous to the wellknown failure of the usual elasticity BIE for frac
ture mechanics problems (see, e.g. [87, 105]), and, of course, gets worse as h/L
gets smaller. The thin plate BEM performs well for h/L = 0.001 (and for a
much lower value of h/L, see Table 3.3). This is because equation (3.5) is now
only used for obtaining the sum of the charges on each plate, and (3.10) is used
for obtaining the separate charges. The bottom line is that equation (3.10) is
crucial for obtaining the charges separately.
mesh on unit square σ
+
at plate center σ
−
at plate center Cg/A
8 ×8 ×2 1.455 9.913 1.615
12 ×12 ×2 1.465 9.908 1.593
16 ×16 ×2 1.470 9.907 1.583
Table 3.3: Convergence of the thin plate BEM model for the parallel plate
capacitor problem for various meshes. h/L = 10
−6
, g/L = 0.2 (from [6])
The convergence characteristics for the thin plate BEM are investigated
for a super thin plate (h/L = 10
−6
) in Table 3.3. It is found that the mesh
dependence of the results is minimal.
Computational eﬃciency. Finally, Table 3.1 demonstrates huge savings in
computational eﬀort for the thin plate BEM  around a factor of 25 compared
to the enhanced BEM.
© 2005 by Taylor & Francis Group, LLC
54 CHAPTER 3. THIN FEATURES
∂B
0
∂B
0
Crack
x
x
x
S
S
S
S
+
+
+



^
^
^
ξ
Figure 3.8: Geometry of the crack problem (from [105])
3.2 BIE for Elasticity: Cracks and Thin Shells
Regularized BIEs for linear elasticity, suitable for applications in linear elastic
fracture mechanics (LEFM), are presented in this section. This is followed by a
discussion of BIEs suitable for the analysis of thin elastic shells. Further details
are available in [105].
3.2.1 BIES in LEFM
This section is concerned with regularization of BIEs in linear elastic fracture
mechanics. The starting point is FP integrals that involve traction sums and
crack opening displacements when the collocation point approaches a crack
surface in the limit to the boundary (LTB) sense. The appropriate FP integrals
are regularized by an additionsubtraction procedure followed by applications
of Stokes’ theorem to convert singular or hypersingular integrals on surfaces
into regular line integrals on the bounding contours of these surfaces.
The geometry of the problem appears in Figure 3.8. The total surface of a
cracked body B is ∂B = ∂B
0
∪S
+
∪S
−
, where S
+
and S
−
are the (coincident)
upper and lower surfaces of a crack in the body. These surfaces are geometrically
identical and have opposite normals at any pair of twin points x
+
and x
−
. A
singular element containing x
+
is
ˆ
S
+
and similarly one containing x
−
is
ˆ
S
−
.
It is well known (e.g. Cruse [38]), that both the displacement and stress
BIEs must be used appropriately in order to solve this problem by the BEM.
A discussion of this, and related issues, follows.
© 2005 by Taylor & Francis Group, LLC
3.2. BIE FOR ELASTICITY: CRACKS AND THIN SHELLS 55
3.2.1.1 Source point approaching the outer surface
Let an internal point ξ approach a point x on the outer surface ∂B
0
of the body
(see Figure 3.8).
Regularized displacement BIE. The usual rigid body mode is used to
regularize the integral on ∂B
0
in the displacement BIE (1.24). With ∂B =
∂B
0
∪ S
+
∪ S
−
, a regularized form of equation (1.24) is:
0 =
∂B
0
[U
ik
(x, y)τ
i
(y) −T
ik
(x, y)(u
i
(y) −u
i
(x)]dS(y)
+
S
+
[U
ik
(x, y)q
i
(y) −T
ik
(x, y)v
i
(y)dS(y) (3.20)
where the sum of the tractions across a crack, q, and the crack opening dis
placement, v, are deﬁned as:
q
i
(y) = τ
i
(y
+
) +τ
i
(y
−
) (3.21)
v
i
(y) = u
i
(y
+
) −u
i
(y
−
) (3.22)
for twin points y
+
and y
−
across a crack.
Regularized stress BIE Use of both rigid body and linear displacement
modes (see Lutz et al. [89] and Mukherjee et al. [99]) leads to a regularized
form of the stress BIE (1.36). This is:
0 =
∂B
0
¸
D
ijk
(x, y)
σ
kh
(y) −σ
kh
(x)
n
h
(y)
−S
ijk
(x, y)
u
k
(y) −u
k
(x) −u
k,
(x)(y
−x
)
¸
dS(y)
+
S
+
[D
ijk
(x, y)q
k
(y) −S
ijk
(x, y)v
k
(y)] dS(y) (3.23)
Either of the equations (3.20) or (3.23) can be collocated at x ∈ ∂B
0
when
solving a LEFM problem. The simpler (3.20) is preferable and is recommended.
3.2.1.2 Source point approaching a crack surface
Let an internal point ξ approach a point x
+
on the crack surface S
+
(see Figure
3.8).
© 2005 by Taylor & Francis Group, LLC
56 CHAPTER 3. THIN FEATURES
The displacement BIE. An FP representation of the displacement BIE
(1.24) with x →x
+
∈ S
+
is:
u
k
(x
+
) =
∂B
0
U
ik
(x
+
, y)τ
i
(y) −T
ik
(x
+
, y)u
i
(y)
dS(y)
+
S
+
=
U
ik
(x
+
, y) q
i
(y) −T
ik
(x
+
, y)v
i
(y)
dS(y) (3.24)
With x
+
∈
ˆ
S
+
, equation (3.24) can be written as:
u
k
(x
+
) =
∂B
0
U
ik
(x
+
, y) τ
i
(y) −T
ik
(x
+
, y)u
i
(y)
dS(y)
+
S
+
−
ˆ
S
+
U
ik
(x
+
, y)q
i
(y) −T
ik
(x
+
, y)v
i
(y)
dS(y)
+
ˆ
S
+
U
ik
(x
+
, y)q
i
(y) dS(y)
−
ˆ
S
+
T
ik
(x
+
, y)[v
i
(y) −v
i
(x)]dS(y)
−v
i
(x)
ˆ
S
+
= T
ik
(x
+
, y)dS(y) (3.25)
It is ﬁrst noted that (see below (1.21)) Σ
ijk
(x, y)n
j
(y) = T
ik
(x, y). Follow
ing [116, 99] (see, also, [87] and Figure 3.2):
ˆ
S
+
= T
ik
(x
+
, y)dS(y) = −g
ik
(x
+
) −
Ω(
ˆ
S
+
, x
+
)
4π
δ
ik
(3.26)
where:
g
ik
(x
+
) =
1 −2ν
8π(1 −ν)
ik
ˆ
L
+
1
r(x
+
, y)
dz
+
km
8π(1 −ν)
ˆ
L
+
r
,i
(x
+
, y)r
,
(x
+
, y)
r(x
+
, y)
dz
m
(3.27)
and the solid angle Ω(
ˆ
S
+
, x
+
), subtended by
ˆ
S
+
at x
+
, is given by equation
(3.8).
In equations (3.26  3.27), z = y − x and , i ≡ ∂/∂y
i
. Also, is the
alternating tensor, δ is the Kronecker delta and the line integrals in (3.27) are
evaluated in a clockwise sense when viewed from above (see Figure 3.2).
Using equations (3.26  3.27) and (3.8) in (3.25), a regularized form of (3.25)
is obtained as:
© 2005 by Taylor & Francis Group, LLC
3.2. BIE FOR ELASTICITY: CRACKS AND THIN SHELLS 57
1
2
(u
k
(x
+
) +u
k
(x
−
)) =
∂B
0
U
ik
(x
+
, y)τ
i
(y) −T
ik
(x
+
, y)u
i
(y)
dS(y)
+
S
+
−
ˆ
S
+
U
ik
(x
+
, y)q
i
(y) −T
ik
(x
+
, y)v
i
(y)
dS(y)
+
ˆ
S
+
U
ik
(x
+
, y)q
i
(y)dS(y)
−
ˆ
S
+
T
ik
(x
+
, y)[v
i
(y) −v
i
(x)]dS(y)
+g
ik
(x
+
)v
i
(x) −
v
k
(x)
4π
2π
0
cos(ψ(θ))dθ (3.28)
The stress BIE. An FP representation of the stress BIE with ξ →x
+
∈ S
+
is:
σ
ij
(x
+
) =
∂B
0
D
ijk
(x
+
, y)τ
k
(y) −S
ijk
(x
+
, y)u
k
(y)
dS(y)
+
S
+
=
D
ijk
(x
+
, y)q
k
(y) −S
ijk
(x
+
, y)v
k
(y)
dS(y) (3.29)
With x
+
∈
ˆ
S
+
, equation (3.29) can be written as:
σ
ij
(x
+
) =
∂B
0
D
ijk
(x
+
, y)τ
k
(y) −S
ijk
(x
+
, y)u
k
(y)
dS(y)
+
S
+
−
ˆ
S
+
D
ijk
(x
+
, y)q
k
(y) −S
ijk
(x
+
, y)v
k
(y)
dS(y)
+
ˆ
S
+
D
ijk
(x
+
, y)[s
k
(y) −s
k
(x)]n
(y)dS(y)
−
ˆ
S
+
S
ijk
(x
+
, y)[v
k
(y) −v
k
(x) −d
kn
(x)(y
n
−x
+
n
)]dS(y)
−v
k
(x)
ˆ
S
+
= S
ijk
(x
+
, y)dS(y)
+d
mn
(x)
ˆ
S
+
=
E
kmn
D
ijk
(x
+
, y)n
(y) −S
ijm
(x
+
, y)(y
n
−x
+
n
)
dS(y)
(3.30)
where the newly deﬁned quantities are:
s
ij
(x) = σ
ij
(x
+
) −σ
ij
(x
−
) (3.31)
d
ij
(x) = u
i,j
(x
+
) −u
i,j
(x
−
) (3.32)
© 2005 by Taylor & Francis Group, LLC
58 CHAPTER 3. THIN FEATURES
Of course, s is related to d by Hooke’s law in the same manner as σ is
related to ∇u, i.e.:
s
ij
= E
ijmn
d
mn
(3.33)
The last two terms on the righthand side of equation (3.30) involve FP
integrals. These are now regularized by applying Stokes’ theorem.
Using the deﬁnition of S
ijk
from (1.32), one gets:
ˆ
S
+
= S
ijk
(x
+
, y)dS(y) = −
ˆ
S
+
= E
ijmn
Σ
km,n
(x
+
, y)n
(y)dS(y) (3.34)
The hypersingular integral on the righthand side of (3.34) can be trans
formed into a line integral. A careful derivation is provided in Mukherjee and
Mukherjee [99], Appendix D. Applying Stokes’ theorem, one has:
ˆ
S
+
= Σ
km,n
n
dS =
ˆ
L
+
Σ
km
nt
dz
t
(3.35)
Finally, the desired result is:
ˆ
S
+
= S
ijk
(x
+
, y)dS(y) = −E
ijmn
ˆ
L
+
Σ
km
(x
+
, y)
nt
dz
t
(3.36)
where, as before, the line integral in (3.36) is evaluated in a clockwise sense
when viewed from above (see Figure 3.2). This result also appears in Liu et al.
[86].
The ﬁnal task is to convert the last term in the righthand side of equation
(3.30) into line integrals. In order to succeed in this task, the entire term must
be converted together. Otherwise, some surface integrals will survive, as in Liu
et al. [86].
First deﬁne:
J
ijmn
=
ˆ
S
+
=
E
kmn
D
ijk
(x
+
, y)n
(y) −S
ijm
(x
+
, y)(y
n
−x
+
n
)
dS(y) (3.37)
Using equation (3.34) and the deﬁnition of the kernel D in terms of U from
equation (1.31), one can write:
J
ijmn
= E
ijpq
I
mnpq
(3.38)
where:
I
mnpq
= −
ˆ
S
+
=
E
kmn
U
pk,q
(x
+
, y) −Σ
mp,q
(x
+
, y)(y
n
−x
+
n
)
n
(y)dS(y)
(3.39)
Integrating by parts:
© 2005 by Taylor & Francis Group, LLC
3.2. BIE FOR ELASTICITY: CRACKS AND THIN SHELLS 59
I
mnpq
= −
ˆ
S
+
=
E
kmn
U
pk
(x
+
, y) −Σ
mp
(x
+
, y)(y
n
−x
+
n
)
,q
n
(y)dS(y)
−
ˆ
S
+
= Σ
mp
δ
nq
n
(y)dS(y) ≡ I
(1)
mnpq
+ I
(2)
mnpq
(3.40)
Once again, the hypersingular integrals in equation (3.40) need careful treat
ment. The ideas presented in Mukherjee and Mukherjee [99], Appendix D, need
to be employed here. (An alternative proof, using Lutz et al.’s ([89]) “tent” in
tegrals, is available in [105]).
Deﬁne
F
mnp
(x
+
, y) = E
kmn
U
pk
(x
+
, y) −Σ
mp
(x
+
, y)(y
n
−x
+
n
) (3.41)
Using Stokes’ theorem in the form:
S
[F
mnp,q
n
−F
mnp,
n
q
] ds =
L
F
mnp
tq
dz
t
(3.42)
(where S =
ˆ
S
+
−
˜
S is a punctured surface with x
+
∈
˜
S and L =
ˆ
L
+
−
˜
L the
bounding contour of S), and the fact that F
mnp,
= 0 for x
+
/ ∈ S, one ﬁnally
gets:
I
(1)
mnpq
(x
+
) = −h
mnpq
(x
+
) (3.43)
where:
h
mnpq
(x
+
) =
ˆ
L
+
E
kmn
U
pk
(x
+
, y) −Σ
mp
(x
+
, y)(y
n
−x
+
n
)
qt
dz
t
(3.44)
Also, using equation (3.26) in (3.40):
I
(2)
mnpq
(x
+
) = −
ˆ
S
+
T
mp
(x
+
, y)δ
nq
ds(y)
= g
mp
(x
+
)δ
nq
+
Ω(
ˆ
S
+
, x
+
)
4π
δ
mp
δ
nq
(3.45)
Finally, use of equations (3.33, 3.36 3.38, 3.40, 3.43 and 3.45) in (3.30)
yields a fully regularized version of (3.30). This is:
(1/2)[σ
ij
(x
+
) +σ
ij
(x
−
)] =
∂B
0
D
ijk
(x
+
, y)τ
k
(y) −S
ijk
(x
+
, y)u
k
(y)
dS(y)
© 2005 by Taylor & Francis Group, LLC
60 CHAPTER 3. THIN FEATURES
+
S
+
−
ˆ
S
+
D
ijk
(x
+
, y)q
k
(y) −S
ijk
(x
+
, y)v
k
(y)
dS(y)
+
ˆ
S
+
D
ijk
(x
+
, y)[s
k
(y) −s
k
(x)]n
(y)dS(y)
−
ˆ
S
+
S
ijk
(x
+
, y)[v
k
(y) −v
k
(x) −d
kn
(x)(y
n
−x
+
n
)]dS(y)
+ E
ijmn
v
k
(x)
ˆ
L
+
Σ
km
(x
+
, y)
nt
dz
t
−E
ijpq
d
mn
(x) h
mnpq
(x
+
)
+ E
ijpn
d
mn
(x)g
mp
(x
+
) −
s
ij
(x)
4π
2π
0
cos(ψ(θ))dθ (3.46)
3.2.2 Numerical Implementation of BIES in LEFM
Of interest here are the regularized equations (3.20), (3.23), (3.28) and (3.46).
The ﬁrst step in the implementation is to choose shape functions for u and τ
on ∂B
0
and q and v on s
+
. A dot product of equation (3.46) must be taken
with n(x
+
).
Next, it is necessary to obtain u
k,
(x) (x ∈ ∂B
0
) in (3.23) and d
k
(x) (x ∈
s
+
) in (3.46). The former is obtained as suggested by Lutz et al. [89]. For 3D
problems, local coordinates α
k
(k = 1, 2, 3), are chosen at x ∈ ∂B
0
such that
the α
3
axis is normal and the α
1
and α
2
axes are tangential to ∂B
0
at x .
Now, tangential diﬀerentiation of the displacement shape functions provides
the quantities u
k,δ
, k = 1, 2, 3; δ = 1, 2. The remaining displacement gradients
at x are obtained from the formulae:
u
1,3
=
τ
1
G
−u
3,1
u
2,3
=
τ
2
G
−u
3,2
u
3,3
=
(1−2ν)τ
3
2(1−ν)G
−
ν
1−ν
(u
1,1
+ u
2,2
)
(3.47)
The quantities d
k
(x) with (x ∈ s
+
) require application of the above idea
on both surfaces of a crack (Figure 3.9). First, in the α
k
coordinate frame, one
has:
d
kγ
= v
k,γ
, k = 1, 2, 3 : γ = 1, 2 (3.48)
The remaining components of d are obtained as follows:
On S
+
, in the α
k
frame : u
1,3
(x
+
) = τ
+
1
/G−(u
+
3
)
,1
On S
−
, in the β
k
frame : u
1,3
(x
−
) = τ
−
1
/G−(u
−
3
)
,1
On S
−
, in the α
k
frame : −u
1,3
(x
−
) = τ
−
1
/G + (u
−
3
)
,1
(3.49)
Adding the ﬁrst and last of equations (3.49), one has:
© 2005 by Taylor & Francis Group, LLC
3.2. BIE FOR ELASTICITY: CRACKS AND THIN SHELLS 61
S
+
S

α
1
β
1
α
2
β
2
α
3
β
3
Figure 3.9: Local coordinates on crack surfaces (from [105])
At x
+
in the α
k
frame : d
13
=
q
1
G
−d
31
(3.50)
Similarly:
d
23
=
q
2
G
−d
32
(3.51)
d
33
=
(1 −2ν)q
3
2(1 −ν)G
−
ν
1 −ν
(d
11
+ d
22
) (3.52)
It should be noted that in equations (3.49), u
+
k
≡ u
k
(x
+
), τ
+
k
≡ τ
k
(x
+
),
and similarly at x
−
.
Finally, s is related to d by Hooke’s law (3.33).
It is interesting to point out that if the surface (tangential) gradient of the
displacement ﬁeld (rather than its total gradient) is used to regularize the rel
evant HBIE, as is done, for example, in [85]  [86], then the procedure outlined
in this section is no longer required. The surface gradient regularization proce
dure, however, leads to some extra terms in a regularized HBIE such as equation
(3.23) in this chapter.
3.2.3 Some Comments on BIEs in LEFM
As has been mentioned before, equation (3.20) is the most convenient one for
collocation on ∂B
0
. On S
+
, both the regularized displacement BIE (3.28) and
© 2005 by Taylor & Francis Group, LLC
62 CHAPTER 3. THIN FEATURES
the stress BIE (3.46) are, in some sense, “defective.” Equation (3.28) contains
u
+
and u
−
on the crack surfaces, but only q (not the individual tractions) on
S
+
and on S
−
. Thus, it can be used to solve for q given u
+
and u
−
, but not
for the important practical case of solving for the crack surface displacements
given the tractions. “Overcollocation” on crack surfaces does not help since
q alone does not specify a problem. For example, both a tractionfree and a
pressurized crack have q = 0 but very diﬀerent solutions. This failure of the
displacement BIE for fracture mechanics problems, is, of course, well known
(e.g. Cruse [38]) and has led to the search for a remedy in the form of a stress
BIE, as early as at least 1975 ([17], [59]).
When one takes a dot product of the stress BIE (3.46) with n
+
, it contains
the tractions τ
+
, τ
−
, and the crack opening displacement (COD) v. This
equation would fail if one wished to calculate the individual tractions on the
crack surfaces given the COD. Fortunately, however, it is perfectly suited for
the practical case of solving for v given τ
+
and τ
−
. (As mentioned in the
previous subsection, the components of the tensor d can be easily calculated
from the tractions and the tangential derivatives of the COD (3.48, 3.50 3.52),
and s is related to d by Hooke’s law (3.33)).
In summary, therefore, it is most convenient to collocate equation (3.20) on
∂B
0
and (3.46) (i.e. its dot product with n
+
) on S
+
. (Lutz et al. [89] used
u and τ on crack surfaces as primary variables and collocated the regular BIE
on one crack face and the HBIE on the other). The individual crack surface
displacements u
+
and u
−
can then be calculated, if desired, at a postprocessing
step, from (3.28) on S
+
.
3.2.4 BIEs for Thin Shells
This section is concerned with BIEs and HBIEs for thin shells, especially in the
limit as the thickness of a shell → 0. The question of what constitutes a thin
plate or shell, in the limit as the thickness → 0, is an interesting one that has
been discussed extensively in the literature, from both a mathematical as well
as an engineering perspective.
3.2.4.1 Mathematical formulation
Regularized BIEs and HBIEs for thin shells are presented next. The geometry
of the problem is shown in Figure 3.10. Again, the “tent” is chosen such that
the unit outward normal to the body at x
+
is also the outward normal to the
“tent” at that point.
The situation here is quite analogous to the crack problem with the outer
surface ∂B
0
in Figure 3.8 replaced by the “edge surface” S
E
of the shell. It is
important to retain the integral over S
E
because even when the shell thickness
h → 0 and S
E
degenerates from a surface to a curve, this integral can remain
nonzero with ﬁnite applied tractions per unit length along this curve.
It has been proved in [105] that, in this case:
© 2005 by Taylor & Francis Group, LLC
3.2. BIE FOR ELASTICITY: CRACKS AND THIN SHELLS 63
x
x
S
S
S
S
S
c
S
E
S
E
+
+
+
n
+
+



^
^
tent
^
Thin Shell
ξ
Figure 3.10: Geometry of the thin shell problem (from [105])
v
i
(x) = u
i
(x
+
) −u
i
(x
−
) = 0 (3.53)
s
ij
(x) = σ
ij
(x
+
) −σ
ij
(x
−
) = 0 (3.54)
Equations (3.53  3.54) simply state that, starting with a continuous dis
placement and stress ﬁeld in a shell of ﬁnite thickness, the displacements and
stresses on the upper and lower surfaces of the shell must approach each other
as the shell thickness h →0.
Unfortunately, however, equation (3.54) implies, (in the limit h → 0), a
strong restriction on the applied tractions on S
+
and S
−
, namely that:
q
k
(x) = τ
k
(x
+
) +τ
k
(x
−
) = 0 (3.55)
Strictly speaking, (3.55) appears to limit the class of thin shell problems that
can be solved by this BIE approach, or, for that matter, by any other approach,
since equations (3.53  3.55) are always true in the thin shell limit h → 0.
Nevertheless, it is well known, of course, that ﬁnite element shell formulations,
for example, work well in practice.
There are interesting situations that satisfy (3.53 3.54) approximately. One
of these cases appears in Liu [87] and is discussed in the next subsection.
3.2.4.2 Some comments on thin shell problems
First, it is useful to examine equations (1113) in Liu [87]. These equations are
valid for h → 0 in the absence of the edge integrals on S
E
. Subtracting (12)
from (11) immediately gives ∆u = 0 which implies that (since the matrix B
+
in [87] is nonsingular) Σt = 0, i.e. equations (3.53) and (3.55) of the present
chapter. Therefore, equation (13) of Liu [87] (which is a restatement of his
© 2005 by Taylor & Francis Group, LLC
64 CHAPTER 3. THIN FEATURES
equations (1112) in terms of displacements and tractions on the upper and
lower surfaces of the shell) only admits the solutions v = q = 0  no other
solutions are possible in the thin shell limit !
One is, of course, primarily interested in solving problems for shells of small
but ﬁnite thickness, i.e. application of equations (45) of Liu [87] (including, in
general, the integral on S
E
) to thin shells. (Note that (13) of [87] is the limiting
form of (45) of [87] as h →0.) Diﬃculties might arise when applying equations
(45) of [87] to shells of small but ﬁnite thickness h if the applied tractions do
not satisfy (3.55), since one would then have a contradiction as h →0. Liu [87],
however, has demonstrated the usefulness of his equations (45) for the case of
a thin spherical shell under internal pressure with radius to thickness ratios of
up to about 200. One reason for his success is that in his spherically symmetric
numerical example the only nonzero displacement component, the radial one, is
nearly uniform as a function of the radius, so that (3.53) is satisﬁed in the thin
shell limit. Also, (3.54) is satisﬁed by the tangential stress components (in the
thin shell limit) while the radial stress component, that violates (3.54), is small
compared to the tangential ones. The shearing stress components are zero in
a spherical coordinate system. Finally, the only nonzero applied traction, the
radial one, violates (3.55) but its magnitude is small compared to the tangential
stress components.
In view of the above discussion, the authors of this book feel that the eﬀect
of the asymptotic requirement (3.55) on the numerical performance of the stan
dard BIE, collocated at twin points on opposite surfaces of a shell, for general
thin shell problems, remains an open question that needs further investigation.
© 2005 by Taylor & Francis Group, LLC
Part II
THE BOUNDARY
CONTOUR METHOD
65
© 2005 by Taylor & Francis Group, LLC
Chapter 4
LINEAR ELASTICITY
Derivations of the BCM and HBCM, for 3D linear elasticity, together with
representative numerical results for selected problems, are presented in this
chapter.
4.1 Surface and Boundary Contour Equations
4.1.1 Basic Equations
A regularized form of the standard boundary integral equation (Rizzo [141]),
for 3D linear elasticity (see equation (1.26) in Chapter 1), is:
0 =
∂B
[U
ik
(x, y)σ
ij
(y) −Σ
ijk
(x, y)¦u
i
(y) −u
i
(x)¦] e
j
dS(y)
≡
∂B
F
k
dS(y) (4.1)
Here, as before, ∂B is the bounding surface of a body B (B is an open set)
with inﬁnitesimal surface area dS = dSn, where n is the unit outward normal
to ∂B at a point on it. The stress tensor is σ, the displacement vector is u
and e
j
(j = 1, 2, 3) are global Cartesian unit vectors. The BEM Kelvin kernels
are written in terms of (boundary) source and ﬁeld points. These are given in
Chapter 1 as equations (1.22) and (1.18), respectively.
The ﬁrst task is to show that the integrand vector F
k
in equation (4.1) is
divergence free (except at the point of singularity x= y). Writing in component
form:
F
k
= F
jk
e
j
= (σ
ij
U
ik
−Σ
ijk
u
i
)e
j
+ Σ
ijk
u
i
(x)e
j
(4.2)
Taking the divergence of the above at a ﬁeld point y, one gets:
67
© 2005 by Taylor & Francis Group, LLC
68 CHAPTER 4. LINEAR ELASTICITY
∇
y
F
k
= F
jk,j
= (σ
ij
E
ijk
−E
ijk
ij
)
+ (σ
ij,j
U
ik
−Σ
ijk,j
u
i
) + Σ
ijk,j
u
i
(x) (4.3)
where the Kelvin strain tensor E and the strain ﬁeld e are:
E
ijk
= (1/2)(U
ik,j
+U
jk,i
),
ij
= (1/2)(u
i,j
+u
j,i
) (4.4)
Let (u, σ) correspond to a body force free electrostatic state with the same
elastic constants as the Kelvin solution. The stress and strain tensors, σ
ij
and
ij
, respectively, are related to each other through the usual Hooke’s law. The
corresponding Kelvin stress and strain tensors Σ
ijk
and E
ijk
, respectively, are
related by Hooke’s law in exactly the same manner (see, e.g. Mukherjee [98]).
As a consequence, the ﬁrst expression on the righthand side (RHS) of equation
(4.3) vanishes. Also, equilibrium in the absence of body forces demands that
σ
ij
be divergence free. The corresponding Kelvin stress tensor Σ
ijk
is also
divergence free, except at the point of singularity. Therefore, the second and
third expressions on the RHS of (4.3) also vanish everywhere, except at the
point of singularity. Thus, F
k
in equation (4.3) is divergence free.
The above property demonstrates the existence of vector potential functions
V
k
such that:
F
k
= ∇V
k
(4.5)
As a consequence of equation (4.5), the surface integral in equation (4.1)
over any open surface patch S ∈ ∂B, can be converted to a contour integral
around the bounding curve C of S, by applying Stokes’ theorem, i.e.:
S
F
k
dS =
C
V
k
dr (4.6)
Stokes’ theorem is valid under very general conditions. The closed curve
C and the open surface S need not be ﬂat  they only need to be piecewise
smooth.
4.1.2 Interpolation Functions
Since the vectors F
k
contain the unknown ﬁelds u and σ, interpolation functions
must be chosen for these variables, and potential functions derived for each lin
early independent interpolation function, in order to determine the vectors V
k
.
Also, since the kernels in equation (4.1) are functions only of z
k
= y
k
−x
k
(and
not of the source and ﬁeld coordinates separately), these interpolation functions
must also be written in the coordinates z
k
in order to determine the potential
vectors V
k
. Finally, these interpolation functions are global in nature and are
chosen to satisfy, a priori, the NavierCauchy equations of equilibrium. (Such
functions are called Treﬀtz functions). The weights, in linear combinations
© 2005 by Taylor & Francis Group, LLC
4.1. SURFACE AND BOUNDARY CONTOUR EQUATIONS 69
¯ u
αi
1 2 3 4 5 6 7
i = 1 1 0 0 y
1
0 0 y
2
i = 2 0 1 0 0 y
1
0 0
i = 3 0 0 1 0 0 y
1
0
¯ u
αi
8 9 10 11 12 13 14
i = 1 0 0 y
3
0 0 y
2
1
y
2
2
i = 2 y
2
0 0 y
3
0 k
1
y
1
y
2
k
2
y
1
y
2
i = 3 0 y
2
0 0 y
3
0 0
¯ u
αi
15 16 17 18 19 20 21
i = 1 k
2
y
1
y
2
k
1
y
1
y
2
k
1
y
1
y
3
k
2
y
1
y
3
y
2
3
y
2
1
0
i = 2 y
2
1
y
2
2
0 0 0 0 y
2
2
i = 3 0 0 y
2
3
y
2
1
k
2
y
1
y
3
k
1
y
1
y
3
k
1
y
2
y
3
¯ u
αi
22 23 24 25 26 27
i = 1 0 0 0 y
2
y
3
0 0
i = 2 y
2
3
k
2
y
2
y
3
k
1
y
2
y
3
0 y
1
y
3
0
i = 3 k
2
y
2
y
3
y
2
2
y
2
3
0 0 y
1
y
2
Table 4.1: Treﬀtz functions for interpolating displacements. α = 1, 2, 3 are
constant, α = 4, 5, ..., 12 are linear and α = 13, 14, ..., 27 are quadratic. k
1
=
−4(1 −ν), k
2
= −2(1 −2ν), k
3
= k
1
−4
of these interpolation functions, however, are deﬁned piecewise on boundary
elements.
Quadratic interpolation functions are used in this work. With :
z
k
= y
k
−x
k
(4.7)
one has, on a boundary element :
u
i
=
27
¸
α=1
β
α
u
αi
(y
1
, y
2
, y
3
) =
27
¸
α=1
ˆ
β
α
(x
1
, x
2
, x
3
)u
αi
(z
1
, z
2
, z
3
) (4.8)
σ
ij
=
27
¸
α=1
β
α
σ
αij
(y
1
, y
2
, y
3
) =
27
¸
α=1
ˆ
β
α
(x
1
, x
2
, x
3
)σ
αij
(z
1
, z
2
, z
3
) (4.9)
where u
αi
, σ
αij
(with i = 1, 2, 3 and α = 1, 2, .., 27) are the interpolation func
tions and β
α
are the weights in the linear combinations of the shape functions.
Each boundary element has, associated with it, 27 constants β
α
that will be re
lated to physical variables on that element. This set β diﬀers from one element
to the next.
The displacement interpolation functions for α = 1, 2, 3 are constants, those
for α = 4, ..., 12 are of ﬁrst degree and those for α = 13, ..., 27 are of second
© 2005 by Taylor & Francis Group, LLC
70 CHAPTER 4. LINEAR ELASTICITY
degree. There are a total of 27 linearly independent (vector) interpolation
functions u
α
. The interpolation functions for the stresses are obtained from
those for the displacements through the use of Hooke’s law. The interpolation
functions u
αi
and σ
αij
are given in Tables 4.1 and 4.2, respectively.
¯ σ
αij/G
1 2 3 4 5 6 7
i, j = 1, 1 0 0 0 k
1
/k
2
0 0 0
i, j = 1, 2 0 0 0 0 1 0 1
i, j = 1, 3 0 0 0 0 0 1 0
i, j = 2, 2 0 0 0 λ/G 0 0 0
i, j = 2, 3 0 0 0 0 0 0 0
i, j = 3, 3 0 0 0 λ/G 0 0 0
¯ σ
αij/G
8 9 10 11 12 13 14
i, j = 1, 1 λ/G 0 0 0 λ/G −k
1
y
1
−4νy
1
i, j = 1, 2 0 0 0 0 0 k
1
y
2
4νy
2
i, j = 1, 3 0 0 1 0 0 0 0
i, j = 2, 2 k
1
/k
2
0 0 0 λ/G k
3
y
1
k
1
y
1
i, j = 2, 3 0 1 0 1 0 0 0
i, j = 3, 3 λ/G 0 0 0 k
1
/k
2
−4νy
1
−4νy
1
¯ σ
αij/G
15 16 17 18 19 20 21
i, j = 1, 1 k
1
y
2
k
3
y
2
k
3
y
3
k
1
y
3
−4νy
1
−k
1
y
1
−4νy
2
i, j = 1, 2 4νy
1
k
1
y
1
0 0 0 0 0
i, j = 1, 3 0 0 k
1
y
1
4νy
1
4νy
3
k
1
y
3
0
i, j = 2, 2 −4νy
2
−k
1
y
2
−4νy
3
−4νy
3
−4νy
1
−4νy
1
−k
1
y
2
i, j = 2, 3 0 0 0 0 0 0 k
1
y
3
i, j = 3, 3 −4νy
2
−4νy
2
−k
1
y
3
−4νy
3
k
1
y
1
k
3
y
1
k
3
y
2
¯ σ
αij/G
22 23 24 25 26 27
i, j = 1, 1 −4νy
2
−4νy
3
−4νy
3
0 0 0
i, j = 1, 2 0 0 0 y
3
y
3
0
i, j = 1, 3 0 0 0 y
2
0 y
2
i, j = 2, 2 −4νy
2
k
1
y
3
k
3
y
3
0 0 0
i, j = 2, 3 4νy
3
4νy
2
k
1
y
2
0 y
1
y
1
i, j = 3, 3 k
1
y
2
−4νy
3
−k
1
y
3
0 0 0
Table 4.2: Treﬀtz functions for interpolating stresses. α = 1, 2, 3 are zero,
α = 4, 5, ..., 12 are constant and α = 13, 14, ..., 27 are linear. k
1
= −4(1 −
ν), k
2
= −2(1 −2ν), k
3
= k
1
−4
It is easy to show that the coordinate transformation (4.7) results in the
constants
ˆ
β
j
being related to the constants β
α
as follows :
ˆ
β
i
=
27
¸
α=1
S
iα
(x
1
, x
2
, x
3
)β
α
, i = 1, 2, 3 (4.10)
© 2005 by Taylor & Francis Group, LLC
4.1. SURFACE AND BOUNDARY CONTOUR EQUATIONS 71
ˆ
β
k
=
27
¸
α=1
R
nα
(x
1
, x
2
, x
3
)β
α
, k = 4, 5, ...12, n = k −3 (4.11)
ˆ
β
α
= β
α
, α = 13, 14, .., 27 (4.12)
where
S
iα
= u
αi
(x
1
, x
2
, x
3
) , i = 1, 2, 3, α = 1, 2, ..., 27
R
kα
=
∂u
α
(y
1
, y
2
, y
3
)
∂y
j
(x
1
,x
2
,x
3
)
, k = 1, 2, .., 9, α = 1, 2, ..., 27
with j = 1 +
k−1
3
 and = k − 3j + 3. Here, the symbol n, called the
ﬂoor of n, denotes the largest integer less than or equal to n.
It is useful to note that the matrices S and R are functions of only the
source point coordinates (x
1
, x
2
, x
3
).
4.1.3 Boundary Elements
The BCM is a perfectly general approach that can be used to solve any well
posed problem in linear elasticity. A departure from the usual BEM, however,
is that a set of primary physical variables a
k
are ﬁrst chosen at appropriate
points on a boundary element. Some of these would be speciﬁed as boundary
conditions and the rest would be unknown in a given problem. The ﬁrst step in
the BCM solution procedure is to determine the unspeciﬁed primary physical
variables in terms of those that are prescribed as boundary conditions. Once
all the primary physical variables are known, the rest of the physical variables
(the secondary ones) are obtained at a simple postprocessing step. Also, unlike
in the standard BEM, it is particularly easy to obtain surface variables, such as
stresses and curvatures, in the BCM. Surface stresses are discussed in Section
4.1.7.
The number of primary physical variables on a boundary element must
match the number of artiﬁcial variables β
k
associated with it, in order that the
transformation matrix T relating the vectors a and β on element m is square.
This relationship is expressed as:
m
a =
m
T
m
β (4.13)
Of course, the matrix T must be invertible. The issue of invertibility is
discussed in Nagarajan et al. [116].
The CIM9 boundary element, shown in Figure 4.1(a), is used in the present
work. The displacement u is the primary physical variable at the three corner
nodes C
i
and the three midside nodes M
i
, while tractions are primary variables
at the internal nodes I
i
. Thus, there are a total of 27 primary physical variables.
© 2005 by Taylor & Francis Group, LLC
72 CHAPTER 4. LINEAR ELASTICITY
*
*
*
C
1
C
2
C
3
M
1
M
2
M
3
I
1
I
2
I
3
(a)
(1,0)
(1,0)
1
2
3
4
5
6
η
ξ
(0, 3)
(b)
Figure 4.1: (a) The CIM9 boundary element. (b) Intrinsic coordinates (from
[109])
The BCM equations are collocated at the six peripheral nodes as well as at
the centroid of the element. In a typical discretization procedure, some of
the peripheral nodes may lie on corners or edges, while the internal nodes are
always located at regular points where the boundary ∂B is locally smooth. It is
of obvious advantage to have to deal only with displacement components, that
are always continuous, on edges and corners, while having traction components
only at regular boundary points. This approach eliminates the wellknown
problems associated with modeling of corners and edges in the usual BEM.
The boundary elements, in general, are curved (quadratic) with their shapes
deﬁned by the six points C
k
, M
k
, k = 1, 2, 3. (see Figure 4.1). The relative
coordinates z
i
(see equation (4.7)) of a point on one of the sides of a triangle
are written as:
z
i
= N
k
(ξ, η)z
k
i
, i = 1, 2, 3; k = 1, 2, ..., 6 (4.14)
where z
k
i
are the relative coordinates of the peripheral nodes 1, 2, .., 6 on the
CIM9 element (see Figure 4.1), and ξ and η are intrinsic coordinates. The shape
functions are:
N
k
= (2L
k
−1)L
k
, k = 1, 2, 3; no sum over k
N
4
= 4L
1
L
2
, N
5
= 4L
2
L
3
, N
6
= 4L
1
L
3
(4.15)
with:
© 2005 by Taylor & Francis Group, LLC
4.1. SURFACE AND BOUNDARY CONTOUR EQUATIONS 73
L
1
= η/
√
3
L
2
= (1/2)(1 −ξ) −η/(2
√
3)
L
3
= (1/2)(1 +ξ) −η/(2
√
3) (4.16)
The reference triangle in intrinsic coordinates is shown in Figure 4.1(b).
The unit outward normal to a boundary element, at a point on it, is given
by:
n =
∂r
∂ξ
∂r
∂η
/
∂r
∂ξ
∂r
∂η
(4.17)
where r = z
i
e
i
. It is important to point out that the elements of the transfor
mation matrix T in equation (4.13) contain the components of the normal n at
the points I
1
, I
2
, I
3
of the CIM9 element shown in Figure 4.1(a).
4.1.4 Vector Potentials
The homogeneous nature of the Kelvin kernels is exploited in deriving the po
tential functions. From equations (1.22) and (1.18), it is clear that both Σ
ijk
and U
ik
are ratios of homogeneous polynomials and are, therefore, homoge
neous. Here, Σ is of degree −2 and u is of degree −1. If an interpolation
function is of degree n, then the resulting force vector F
αk
(which is F
k
corre
sponding to the αth interpolation function) is of degree n − 2. In this work,
interpolation functions with n = 0, 1, 2 are used (see Tables 4.1 and 4.2).
4.1.4.1 The nonsingular case : n = 0
The homogeneous nature of the Kelvin kernels greatly facilitates the use of an
inversion integral to calculate the inverse curl of a given vector ﬁeld of zero
divergence (see, e.g. [69]). Thus:
F
αk
= ∇V
αk
⇒ V
αk
(z
1
, z
2
, z
3
) =
¸
1
0
tF
αk
(tz
1
, tz
2
, tz
3
)dt
r
(4.18)
where r = z
i
e
i
.
Because of the homogeneous nature of F
αk
:
F
αk
(tz
1
, tz
2
, tz
3
) = t
n−2
F
αk
(z
1
, z
2
, z
3
) (4.19)
Therefore, for the nonsingular case n = 0 equations (4.18, 4.19) yield:
V
αk
(z
1
, z
2
, z
3
) = (1/n)F
αk
(z
1
, z
2
, z
3
) r (4.20)
© 2005 by Taylor & Francis Group, LLC
74 CHAPTER 4. LINEAR ELASTICITY
4.1.4.2 The singular case : n=0
The singular case (n = 0) corresponds to constant displacement interpolation
functions ¯ u
αi
= δ
αi
, α = 1, 2, 3; i = 1, 2, 3, (where δ is the Kronecker delta 
see Table 4.1). Referring to equation (4.1), one must now deal with the term
Σ
αjk
given in equation (1.22). The expression for Σ
ijk
e
j
can be partitioned
into three terms, each of which are divergence free [116]. One writes:
−Σ
αjk
e
j
=
1
8π(1 −ν)r
2
¸
3r
,α
r
,j
r
,k
−r
,j
δ
αk
e
j
+
1 −2ν
8π(1 −ν)r
2
¸
r
,α
δ
jk
−r
,k
δ
αj
e
j
+
r
,j
δ
αk
4πr
2
e
j
(4.21)
Each of the above three terms on the righthand side of equation (4.21) are
divergence free and can be written as the curl of a potential function. Corre
sponding to the ﬁrst two terms, one has [116]:
1
r
2
¸
3r
,α
r
,j
r
,k
−r
,j
δ
αk
e
j
= ∇
kmj
r
,α
r
,m
r
e
j
(4.22)
1
r
2
¸
r
,α
δ
jk
−r
,k
δ
αj
e
j
= ∇
αkm
r
e
m
(4.23)
where is the alternating tensor.
It is well known that the solid angle Ω subtended at the source point x by
the open surface S has the expression:
Ω =
S
r
,j
r
2
e
j
dS =
S
r.dS
r
3
(4.24)
Therefore, the surface integral of the third term on the righthand side of
(4.21) over S equals (Ω/4π)δ
αk
. While it is possible to convert the surface inte
gral in (4.24) to a line integral [116], use of this line integral has proved to lack
robustness in general numerical computations involving 3D bodies of complex
shape. Therefore, the solid angle Ω is computed as a surface integral according
to equation (4.24). This is the only surface integral that is ever computed in the
BCM. Fortunately, the solid angle is a purely geometrical quantity that can be
computed easily as a surface integral in a robust fashion. It is also noted here
that algebraic expressions exist for the solid angle for the special case when S
is a plane.
4.1.5 Final BCM Equations
Use of the interpolating functions for displacement and stress from Tables 4.1
and 4.2 respectively, together with equations (4.7  4.13) and (4.20  4.24) trans
forms the regularized BIE (4.1) into a regularized boundary contour equation
(BCE) that can be collocated (as in the usual BEM) at any boundary (surface)
© 2005 by Taylor & Francis Group, LLC
4.1. SURFACE AND BOUNDARY CONTOUR EQUATIONS 75
point  including those on edges and corners, as long as the displacement is
continuous there. This equation is:
0 =
1
2
M
¸
m=1
27
¸
α=13
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jnt
z
n
dz
t
¸
m
T
−1
m
a
α
+
M
¸
m=1
12
¸
α=4
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jnt
z
n
dz
t
¸
R
m
T
−1
m
a
α−3
+
M
¸
m=1
m/ ∈S
3
¸
α=1
¸
L
m
D
αjk
dz
j
¸
S
m
T
−1
m
a −
P
T
−1
P
a
¸
α
(4.25)
with
L
m
D
αjk
dz
j
= −
S
m
Σ
αjk
e
j
dS
=
1
8π(1 −ν)
L
m
kij
r
,α
r
,i
r
dz
j
+
1 −2ν
8π(1 −ν)
L
m
αkj
1
r
dz
j
+
Ω
4π
δ
αk
(4.26)
Here L
m
is the bounding contour of the surface element S
m
. Also,
m
T
and
m
a
are the transformation matrix and primary physical variable vectors on element
m,
P
T
and
P
a are the same quantities evaluated on any element that belongs to
the set o of elements that contain the source point x, and
ijk
is the usual
alternating symbol.
This method of integrating Σ
ijk
has been presented before, using spherical
coordinates, by Ghosh and Mukherjee [48]. Also, Manti˘c [94] has independently
derived the result in equation (4.26) with the use of the tangential diﬀerential
operator:
D
ij
(f(y)) ≡ n
i
(y)f
,j
(y) −n
j
(y)f
,i
(y) (4.27)
The equations are assembled by making use of the fact that displacements
are continuous across elements. The ﬁnal result is:
Ka = 0 (4.28)
which is written as:
Ax = By (4.29)
where x contains the unknown and y the known (from the boundary conditions)
values of the primary physical variables on the surface of the body. Once these
© 2005 by Taylor & Francis Group, LLC
76 CHAPTER 4. LINEAR ELASTICITY
equations are solved, the vector a is completely known. Now, at a postprocess
ing step,
m
β
α
can be easily obtained on each boundary element from equation
(4.13).
4.1.6 Global Equations and Unknowns
The global system equation (4.29) generally leads to a rectangular matrix A.
The system of linear equations is usually overdetermined but always consistent.
A count of equations and unknowns is given below.
For any general polyhedron, Euler’s theorem states that:
F +V = E + 2 (4.30)
where F is the number of faces (here the number of elements), V is the number
of vertices (here the number of corner nodes) and E is the number of edges
(here 1.5F for triangular elements). Thus, one has:
Number of corner nodes =
1
2
Number of elements + 2 (4.31)
Also, a CIM9 element has 1.5 midside nodes.
One is now in a position to count the number of (vector) equations and
unknowns in the global system (4.29). In a CIM9 element, the BCM equations
are enforced at all the peripheral nodes and also at the centroid of the element.
Thus, for M boundary elements, one has M equations at the centroids, 1.5M
equations at the midside nodes and 0.5M + 2 at the corner nodes, for a total
of 3M + 2 equations. For a Dirichlet problem, in which all the displacements
are prescribed on ∂B, there are a total of 3M vector unknowns  the tractions
at the nodes interior to the elements. One therefore has two extra vector (six
extra scalar) equations. This is the worstcase scenario in the sense that for
mixed boundary value problems the number of equations remain the same while
the number of unknowns decreases. For example, for a Neumann problem in
which all the tractions are prescribed on ∂B, one only has 2M + 2 (vector)
unknowns (displacements at the peripheral nodes of the elements). Of course,
a Neumann problem is illposed since rigid body displacements of the body are
not constrained.
In summary, use of the CIM9 element results in overdetermined, consistent
linear systems for wellposed problems in linear elasticity.
4.1.7 Surface Displacements, Stresses, and Curvatures
A very useful consequence of using global shape functions is that, once the
standard BCM is solved, it is very easy to obtain the displacements, stresses and
curvatures at a regular oﬀcontour boundary point (ROCBP) on the bounding
surface of a body. Here, a point at an edge or corner is called an irregular
point while at a regular point the boundary is locally smooth. Also, a regular
boundary point can lie on or away from a boundary contour. The former is
© 2005 by Taylor & Francis Group, LLC
4.1. SURFACE AND BOUNDARY CONTOUR EQUATIONS 77
called a regular contour point (RCP), the latter a regular oﬀcontour boundary
point. A point inside a body is called an internal point.
First, one obtains
m
β
α
from equation (4.13), then uses equations (4.10 and
4.11) to get
ˆ
β
m
α
, α = 1, 2, .., 12. The curvatures, which are piecewise constant
on each boundary element, are obtained by direct diﬀerentiation of equation
(4.8). Finally, one has the following results.
4.1.7.1 Surface displacements
[u
i
(x)] =
ˆ
β
1
ˆ
β
2
ˆ
β
3
¸
¸
x
(4.32)
4.1.7.2 Surface stresses
[u
i,j
(x)] =
ˆ
β
4
ˆ
β
7
ˆ
β
10
ˆ
β
5
ˆ
β
8
ˆ
β
11
ˆ
β
6
ˆ
β
9
ˆ
β
12
¸
¸
x
(4.33)
4.1.7.3 Surface curvatures
¸
∂
2
u
1
∂x
i
∂x
j
=
2(β
13
+β
20
) k
2
β
15
+k
1
β
16
k
1
β
17
+k
2
β
18
2β
14
β
25
symmetric 2β
19
¸
¸
¸
¸
(4.34)
¸
∂
2
u
2
∂x
i
∂x
j
=
2β
15
k
1
β
13
+k
2
β
14
β
26
2(β
16
+β
21
) k
2
β
23
+k
1
β
24
symmetric 2β
22
¸
¸
¸
¸
(4.35)
¸
∂
2
u
3
∂x
i
∂x
j
=
2β
18
β
27
k
2
β
19
+k
1
β
20
2β
23
k
1
β
21
+k
2
β
22
symmetric 2(β
17
+β
24
)
¸
¸
¸
¸
(4.36)
with k
1
= −4(1 −ν) and k
2
= −2(1 −2ν)
Equation (4.33) can be used to ﬁnd the displacement gradients at x. Hooke’s
law would then give the stress σ
ij
(x). An alternative approach is to use equation
(4.9) together with all the β
α
on an element.
It should be noted that the simple approach, described above, cannot be
used to ﬁnd internal stresses since the constants
m
β
α
are only meaningful on
the boundary of a body. Therefore, an internal point representation of the
diﬀerentiated BCE, for the internal displacement gradients u
i,j
(p), is necessary
© 2005 by Taylor & Francis Group, LLC
78 CHAPTER 4. LINEAR ELASTICITY
(see Section 4.3). It is also of interest to examine the limiting process of a
diﬀerentiated BCE as an internal point ξ (also denoted as p) approaches a
boundary point x (also denoted as P). This issue is of great current interest
in the BEM community in the context of the standard and hypersingular BIE
(HBIE  see, for example, [92], [39], [93], [111]). Further, the hypersingular BCE
(HBCE) must be understood if one wishes to collocate the HBCE as the primary
integral equation, as may be necessary, for example, in applications such as
fracture mechanics, symmetric Galerkin formulations or adaptive analysis. This
topic is the subject of Section 4.2 of this chapter.
4.2 Hypersingular Boundary Integral Equations
4.2.1 Regularized Hypersingular BIE
A hypersingular boundary integral equation (HBIE) can be obtained by diﬀer
entiating the standard BIE at an internal point, with respect to the coordinates
of this internal source point. A regularized version of this equation, containing,
at most, weakly singular integrals (see equation (1.38) in Chapter 1) is:
0 =
∂B
U
ik,n
(x, y) [σ
ij
(y) −σ
ij
(x)] n
j
(y)dS(y)
−
∂B
Σ
ijk,n
(x, y) [u
i
(y) −u
i
(x) −u
i,
(x) (y
−x
)] n
j
(y)dS(y)
x, y ∈ ∂B (4.37)
Martin et al. [93]  Appendix II2, p. 905, (see, also, [111]) have proved that
(4.37) can be collocated even at an edge or corner point x on the surface of
a 3D body, provided that the displacement and stress ﬁelds in (4.37) satisfy
certain smoothness requirements. These smoothness requirements are discussed
in Section 1.4.4 in Chapter 1. Please note that in (4.37) above, , n ≡ ∂y
n
, not
the normal derivative.
4.2.2 Regularized Hypersingular BCE
The regularized HBIE (4.37) can be converted to a regularized hypersingular
boundary contour equation (HBCE). Details are available in [99] and are given
below.
The ﬁrst task at hand is to transform equation (4.37) into a boundary con
tour form. The integrands in equation (4.37), without n
j
(y)dS(y), are ﬁrst
evaluated at an internal ﬁeld point q very near Q (i.e. on a surface ∂
ˆ
B inside
the body, very near and parallel to ∂B), the derivatives are transferred from
the kernels to the quantities inside the square brackets by the product rule,
and then the limit q → Q is taken again. This is possible since the integrals in
equation (4.37) are regular. The result is:
© 2005 by Taylor & Francis Group, LLC
4.2. HYPERSINGULAR BOUNDARY INTEGRAL EQUATIONS 79
0 =
∂B
U
ik
(x, y) [σ
ij
(y) −σ
ij
(x)] −Σ
ijk
(x, y)
u
i
(y) −u
(L)
i
,n
n
j
(y)dS(y)
−
∂B
[U
ik
(x, y)σ
ij,n
(y) −Σ
ijk
(x, y) [u
i,n
(y) −u
i,n
(x)]] n
j
(y)dS(y)
x, y ∈ ∂B (4.38)
where
u
(L)
i
= u
i
(x) +u
i,
(x)(y
−x
)
Note also that:
u
(L)
i,n
= u
i,n
(x)
Therefore, the linear displacement ﬁeld u
(L)
i
gives the stress ﬁeld σ
ij
(x), so
that the stress ﬁeld σ
ij
(y) − σ
ij
(x) is obtained from the displacement ﬁeld
u
i
(y) −u
(L)
i
.
Using the identity [116]
v
,n
= ∇(v e
n
) (4.39)
(which is valid if the vector ﬁeld v is divergencefree) and Stokes’ theorem, the
ﬁrst integral in equation (4.38), over S
m
, is converted to the contour integral:
I
1
=
L
m
U
ik
(x, y) [σ
ij
(y) −σ
ij
(x)] −Σ
ijk
(x, y)
u
i
(y) −u
(L)
i
jnt
dz
t
(4.40)
An explicit form of equation (4.40) is derived in Appendix A of this chapter.
Next, it is noted from equations (4.32 and 4.10) that:
u
i,n
(x) =
¸
S
,N
P
β
i
(4.41)
where , N ≡ ∂/∂x
n
.
Further, as proved in Appendix B of this chapter:
u
i,n
(y) =
¸
S
,N
m
β
i
+
12
¸
α=4
¸
R
,N
m
β
α−3
u
αi
(z
1
, z
2
, z
3
) (4.42)
Now, the second integral (called I
2
) in equation (4.38) is written as:
I
2
= −
∂B
[U
ik
(x, y)σ
ij,n
(y) −Σ
ijk
(x, y)u
i,n
(y)] n
j
(y)dS(y)
−u
i,n
(x)
∂B
Σ
ijk
(x, y)n
j
(y)dS(y) (4.43)
© 2005 by Taylor & Francis Group, LLC
80 CHAPTER 4. LINEAR ELASTICITY
The next steps involve writing ∂B ≡ ∪S
m
, separating the constant and
linear parts of u
i,n
in the ﬁrst integrand above, and using equation (4.42). This
sets the stage for converting I
2
into two contour integrals. Details are given in
Appendix C of this chapter.
The ﬁnal result is a contour integral form (HBCE) of the regularized HBIE
(4.37). This equation is valid at any point on the boundary ∂B as long as the
stress is continuous there. This includes edge, corner and regular points, that
lie on or oﬀ contours.
The regularized hypersingular boundary contour equation (HBCE) is:
0 = −
M
¸
m=1
27
¸
α=13
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jnt
dz
t
¸
m
T
−1
m
a
α
+
M
¸
m=1
12
¸
α=4
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jst
z
s
dz
t
¸
R
,N
m
T
−1
m
a
α−3
−
M
¸
m=1
m/ ∈S
12
¸
α=4
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jnt
dz
t
¸
R
m
T
−1
m
a −
P
T
−1
P
a
¸
α−3
+
M
¸
m=1
m/ ∈S
3
¸
α=1
¸
L
m
D
αjk
dz
j
¸
S
,N
m
T
−1
m
a −
P
T
−1
P
a
¸
α
+
M
¸
m=1
m/ ∈S
3
¸
α=1
¸
L
m
Σ
αjk
jnt
dz
t
¸
S
m
T
−1
m
a −
P
T
−1
P
a
¸
α
(4.44)
where, as before, o is the set of boundary elements that contains the source
point x . The derivatives R
,N
and S
,N
in (4.44) are taken with respect to the
source point coordinates x
n
. In equation (4.44), the integrands in the ﬁrst two
terms are regular (O(1)). The third and fourth (potentially strongly singular,
O(1/r)) as well as the ﬁfth (potentially hypersingular, O(1/r
2
)) need to be
evaluated only on nonsingular elements.
4.2.3 Collocation of the HBCE at an Irregular Surface
Point
Martin et al. [93] have stated the requirements for collocating a regularized
HBIE at an irregular point on ∂B. This matter is discussed in Section 1.4.4 in
Chapter 1 of this book. It has been proved in Mukherjee and Mukherjee [111]
that the BCE interpolation functions given in equations (4.8) and (4.9) satisfy,
a priori, all these smoothness requirements for collocation of the HBCE (4.44)
at an irregular surface point. This proof is repeated below.
It is ﬁrst noted that the interpolation functions in (4.8  4.9) have both a
global (they are initially deﬁned as functions of y) as well as a local (the weights
© 2005 by Taylor & Francis Group, LLC
4.2. HYPERSINGULAR BOUNDARY INTEGRAL EQUATIONS 81
β
k
are only deﬁned piecewise on the boundary elements) character. Consider
a singular boundary element containing the source and ﬁeld points P and Q,
with P an irregular point on ∂B. Let this element be any one of the smooth
pieces of ∂B that meet at P (see Section 1.4.4 in Chapter 1). From equations
(4.8  4.9), it is easy to show that :
u
i
(Q) −u
L
i
(Q; P) = u
i
(Q) −u
i
(P) −u
i,j
(P)[y
j
(Q) −x
j
(P)]
=
27
¸
α=13
β
α
¯ u
αi
(z) = O(r
2
) (4.45)
σ
ij
(Q) −σ
ij
(P) =
27
¸
α=13
β
α
¯ σ
αij
(z) = O(r) (4.46)
where r = [y(Q) − x(P)[ = [z[. The last equalities in the above equations are
true in view of the fact that the shape functions ¯ u
αi
and ¯ σ
αij
are quadratic and
linear, respectively, in z
k
, for α = 13, 14, ..., 27 (see Tables 4.1 and 4.2). Note
that these weights β
α
belong to the element containing P and Q and are unique
on that element (see below equation (4.9)).
As an aside, it is interesting to connect with Toh and Mukherjee ([168]
p.2304) where, for the same problem, the requirement [∇u(Q) − ∇u(P)[ =
O(r
α
) is prescribed as r → 0. It is easy to show that (Mukherjee and Mukherjee
[99]), for the BCM shape functions on a singular element:
u
i,j
(Q) −u
i,j
(P) =
12
¸
α=4
[R
,J
(x) β]
α−3
¯ u
αi
(z) = O(r) (4.47)
since ¯ u
αi
(z) is linear in z for α = 4, 5, ..., 12.
In view of equations (4.45  4.46), conditions (iiiiv) in Box 1.1 (in Section
1.4.4 in Chapter 1) are satisﬁed a priori by the BCM shape functions deﬁned
by equations (4.8  4.9). Satisfaction of condition (ii(b)) on ∂B follows from
equation (4.46). The conditions inside B ((i) and (ii(a))), of course, have noth
ing to do with BEM shape functions that are only deﬁned on the bounding
surface ∂B, but rather with the boundary element method itself. The BCM
is derived from the BEM and it satisﬁes these internal point conditions in the
same way as does the BEM. (As a bonus, the BCM shape functions satisfy the
NavierCauchy equilibrium equations of linear elasticity a priori (see Section
4.1.2)), although weights are not deﬁned at points p ∈ B). Please note that the
above arguments have been made for the “worstcase scenario” of the colloca
tion point P ∈ ∂B being an irregular point. Of course, these arguments also
go through for regular points (on or oﬀ contour) on ∂B.
In view of the above, all the conditions (iiv) in Box 1.1 are satisﬁed a priori
by the BCM, and there is no need to consider “relaxed smoothness requirements”
in this method. It is worth repeating again that it is extremely diﬃcult to
ﬁnd, in general, BEM shape functions (for 3D elasticity problems) that satisfy
© 2005 by Taylor & Francis Group, LLC
82 CHAPTER 4. LINEAR ELASTICITY
conditions (ii(b)iv) a priori. The primary reason for this is that BEM shape
functions are deﬁned only on the bounding surface of a body, while the BCM
ones are deﬁned in B (although the weights are deﬁned only on ∂B).
4.3 Internal Displacements and Stresses
At this stage, it is a simple matter to derive equations for displacements and
stresses at a point inside a body. The general equations are derived in [103]
and equations for internal points close to the bounding surface are derived in
[104]. They are given below.
4.3.1 Internal Displacements
One has to compare the regularized BIE (4.1), the regularized BCE (4.25) and
the usual BIE at an internal point p = ξ (equation (1.21) in Chapter 1). The
result is:
u
k
(ξ) =
1
2
M
¸
m=1
27
¸
α=13
¸
L
m
(σ
αij
(z)U
ik
(z) −u
αi
(z)Σ
ijk
(z))
jst
z
s
dz
t
¸
m
T
−1
m
a
α
+
M
¸
m=1
12
¸
α=4
¸
L
m
(σ
αij
(z)U
ik
(z) −u
αi
(z)Σ
ijk
(z))
jst
z
s
dz
t
¸
R(ξ)
m
T
−1
m
a
α−3
+
M
¸
m=1
3
¸
α=1
¸
L
m
D
αjk
(z)dz
j
¸
S(ξ)
m
T
−1
m
a
α
(4.48)
where z = y −ξ.
4.3.2 Displacements at Internal Points Close to the Bound
ing Surface
This section is concerned with evaluation of displacements at internal points
that are close to the bounding surface of a body. Section 4.3.4 concerns internal
stresses also at points close to the bounding surface. Please refer to Section 1.3
of this book for a similar discussion related to the BEM. (See, also, [104]).
The ﬁrst step is to choose the target point ˆ x at or close to the centroid
of a boundary element (see Figure 1.3). Since all other terms in equation
(4.48), except the solid angle, are evaluated as contour integrals, these terms
are already regularized. There are at least two ways of regularizing the solid
angle term in equation (4.48). The ﬁrst is to evaluate the solid angle Ω (see
equation (4.24)) as a line integral by employing equation (16) in Liu [87]. The
© 2005 by Taylor & Francis Group, LLC
4.3. INTERNAL DISPLACEMENTS AND STRESSES 83
second is to use a boundary contour version of equation (1.49) and still evaluate
Ω as a surface integral. The latter approach is adopted in this work.
The boundary contour version of equation (1.49) can be obtained easily.
This equation is:
u
k
(ξ) = u
k
(ˆ x)
+
1
2
M
¸
m=1
27
¸
α=13
¸
L
m
(σ
αij
(z)U
ik
(z) −u
αi
(z)Σ
ijk
(z))
jst
z
s
dz
t
¸
m
T
−1
m
a
α
+
M
¸
m=1
12
¸
α=4
¸
L
m
(σ
αij
(z)U
ik
(z) −u
αi
(z)Σ
ijk
(z))
jst
z
s
dz
t
¸
R(ξ)
m
T
−1
m
a
α−3
+
M
¸
m=1
3
¸
α=1
¸
L
m
D
αjk
(z)dz
j
¸
S(ξ)
m
T
−1
m
a −S(ˆ x)β
ˆ
P
α
(4.49)
where:
u
k
(ˆ x) =
ˆ
β
ˆ
P
k
=
27
¸
α=1
S
kα
(ˆ x)β
ˆ
P
α
(4.50)
Note that the point
ˆ
P has coordinates ˆ x.
It is important to note that, on a singular element (i.e. when integration is
being carried out on an element that contains the point ˆ x), one has:
m
T
−1
m
a=
m
β= β
ˆ
P
(4.51)
In this case, the numerator of the last integrand in equation (4.49) is O(r(ξ, ˆ x))
while the denominator in the solid angle term is O(r
2
(ξ, ˆ x)) as y → ˆ x, so that
equation (4.49) is “nearly weakly singular” as y → ˆ x. It is useful to remember
that the integral of D
αjk
in equation (4.49) contains the solid angle term which
is evaluated as a surface integral.
4.3.3 Internal Stresses
This time, one has to compare the regularized HBIE (4.37), the regularized
HBCE (4.44) and the usual integral expression for the displacement gradient
at an internal point p = ξ (equation (1.34) in Chapter 1). The result is:
u
k,n
(ξ) = −
M
¸
m=1
27
¸
α=13
¸
L
m
(σ
αij
(z)U
ik
(z) −u
αi
(z)Σ
ijk
(z))
jnt
dz
t
¸
m
T
−1
m
a
α
+
M
¸
m=1
12
¸
α=4
¸
L
m
(σ
αij
(z)U
ik
(z) −u
αi
(z)Σ
ijk
(z))
jst
z
s
dz
t
¸
R
,n
(ξ)
m
T
−1
m
a
α−3
© 2005 by Taylor & Francis Group, LLC
84 CHAPTER 4. LINEAR ELASTICITY
−
M
¸
m=1
12
¸
α=4
¸
L
m
(σ
αij
(z)U
ik
(z) −u
αi
(z)Σ
ijk
(z))
jnt
dz
t
¸
R(ξ)
m
T
−1
m
a
α−3
+
M
¸
m=1
3
¸
α=1
¸
L
m
D
αjk
(z)dz
j
¸
S
,n
(ξ)
m
T
−1
m
a
α
+
M
¸
m=1
3
¸
α=1
¸
L
m
Σ
αjk
(z)
jnt
dz
t
¸
S(ξ)
m
T
−1
m
a
α
(4.52)
Hooke’s law is now used to obtain the internal stress from the internal displace
ment gradient.
Curvatures at an internal point are given in equations (4.34  4.36).
4.3.4 Stresses at Internal Points Close to the Bounding
Surface
As before for the case of displacement evaluation at an internal point close to
the boundary of a body (see start of section 4.3.2), one has two choices with
respect to the strategy for evaluation of the solid angle. Again, for the sake of
uniformity, a boundary contour version of equation (1.51) is used here, together
with evaluation of the solid angle as a surface integral.
A boundary contour version of equation (1.51) is obtained in a manner that
is quite analogous to the approach discussed in Section 4.2.2 (see, also, [99]).
The ﬁrst step is to use the product rule to transform equation (1.51) to the
form:
u
k,n
(ξ) = u
k,n
(ˆ x)
−
∂B
U
ik
(ξ, y) [σ
ij
(y) −σ
ij
(ˆ x)] −Σ
ijk
(ξ, y)
u
i
(y) −u
(L)
i
,n
n
j
(y)dS(y)
+
∂B
[U
ik
(ξ, y)σ
ij,n
(y) −Σ
ijk
(ξ, y) [u
i,n
(y) −u
i,n
(ˆ x)]] n
j
(y)dS(y) (4.53)
where (see Figure 1.3):
u
(L)
i
= u
i
(ˆ x) −u
i,
(ˆ x)ˆ z
(4.54)
with:
ˆ z
= y
− ˆ x
(4.55)
The BCM version of equation (1.51) is:
u
k,n
(ξ) = u
k,n
(ˆ x)
© 2005 by Taylor & Francis Group, LLC
4.4. NUMERICAL RESULTS 85
−
M
¸
m=1
27
¸
α=13
¸
L
m
(σ
αij
(ˆ z)U
ik
(z) −u
αi
(ˆ z)Σ
ijk
(z))
jnt
dz
t
¸
m
T
−1
m
a
α
+
M
¸
m=1
12
¸
α=4
¸
L
m
(σ
αij
(z)U
ik
(z) −u
αi
(z)Σ
ijk
(z))
jst
z
s
dz
t
¸
R
,n
(ξ)
m
T
−1
m
a
α−3
−
M
¸
m=1
m/ ∈S
12
¸
α=4
¸
L
m
(σ
αij
(ˆ z)U
ik
(z) −u
αi
(ˆ z)Σ
ijk
(z))
jnt
dz
t
×
¸
R(ˆ x)
m
T
−1
m
a −β
ˆ
P
α−3
+
M
¸
m=1
3
¸
α=1
¸
L
m
D
αjk
(z)dz
j
¸
S
,n
(ξ)
m
T
−1
m
a −S
,n
(ˆ x)β
ˆ
P
α
+
M
¸
m=1
m/ ∈S
3
¸
α=1
¸
L
m
Σ
αjk
(z)
jnt
dz
t
¸
S(ˆ x)
m
T
−1
m
a −β
ˆ
P
α
(4.56)
where ˆ z = y − ˆ x (see Figure 1.3).
It should be noted that the ﬁrst, third and ﬁfth terms, with summations
and integrals, on the righthand side of equation (4.56), arise from the ﬁrst
integral in equation (4.53); while the second and fourth arise from the second
integral in equation (4.53). Again, as in the case of equation (4.49), the last but
one term on the righthand side of equation (4.56) is “nearly weakly singular”
(O(1/r(ξ, ˆ x)) as y → ˆ x).
4.4 Numerical Results
Numerical results from the BCM, for selected 3D examples, are available in
[116], [109] and, from the HBCM, in [99]. Typical results, for a thick hollow
sphere under internal pressure, are given below. For the results in Sections 4.4.1
 4.4.3, the inner and outer radii of the sphere are 1 and 2 units, respectively,
the shear modulus G = 1, the Poisson’s ratio ν = 0.3, and the internal pressure
is 1. For the results in Section 4.4.4, the inner and outer radii of the sphere
are 1 and 4 units, respectively, the Young’s modulus E = 1, the Poisson’s ratio
ν = 0.25, and the internal pressure is 1. A generic surface mesh on a oneeighth
sphere is shown in Figure 4.2. Two levels of discretization  medium and ﬁne,
are used in this work. Mesh statistics are shown in Table 4.3.
4.4.1 Surface Displacements from the BCM and the
HBCM
First, please note that (4.44) has two free indices, k and n, so that it repre
sents nine equations. These equations arise from u
k,n
. Diﬀerent strategies are
© 2005 by Taylor & Francis Group, LLC
86 CHAPTER 4. LINEAR ELASTICITY
X
1
X
2
X
3
Figure 4.2: A typical mesh on the surface of a oneeighth sphere (from [100])
mesh number of elements
on each ﬂat plane on each curved surface total
coarse 12 9 54
medium 36 36 180
ﬁne 64 64 320
Table 4.3: Mesh statistics on a oneeighth sphere (from [100])
possible for collocating (4.44) at a boundary point. The ﬁrst is to use all nine
equations. The second is to use six corresponding to
kn
= (1/2)(u
k,n
+ u
n,k
).
The sixequation strategy amounts to replacing E
kn
, the righthand side of
(4.44), by (1/2)(E
kn
+E
nk
). Both the nineequation and sixequation strategies
lead to overdetermined systems, but are convenient for collocating at irregular
boundary points since the source point normal is not involved in these cases.
A third, the threeequation strategy, suitable for collocation at regular points,
corresponds to the traction components τ
n
. In this case, the righthand side
(E
kn
) of (4.44) is replaced by [λE
mm
δ
kn
+ µ(E
kn
+ E
nk
)]n
k
(P), where λ and
µ are Lam´e constants, δ
ij
are components of the Kronecker delta and Hooke’s
law is used. The threeequation strategy involving the traction components is
not convenient for collocating the HBCE at a point on an edge or a corner of a
body where the normal to the body surface has a jump discontinuity. In view
of the assumed continuity of the stress tensor at such a point, this situation
leads to a jump in traction at that point, unless the stress tensor is zero there.
One would, therefore, need to use multiple source points, each belonging to
a smooth surface meeting at that irregular point, and collocate separately at
these points. Since the primary purpose here is to demonstrate collocation of
(4.37) at irregular boundary points, only the nineequation and sixequation
strategies are used below.
It should be mentioned here that, for the HBCM in 2D elasticity, a strategy
© 2005 by Taylor & Francis Group, LLC
4.4. NUMERICAL RESULTS 87
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
0.1
0.15
0.2
0.25
0.3
0.35
Radial coordinate
R
a
d
i
a
l
d
i
s
p
l
a
c
e
m
e
n
t
Figure 4.3: Hollow sphere under internal pressure. Radial displacement as a
function of radius along the x
1
axis. Numerical solutions are obtained from the
medium mesh. Exact solution: —, BCM solution: ◦ ◦ ◦◦, sixequation HBCM
solution: ∗ ∗ ∗∗, nineequation HBCM solution: ++++ (from [111])
corresponding to the ﬁrst one above has been successfully employed by [131]
and a strategy corresponding to the third one above has been implemented by
[187].
The overdetermined system of linear algebraic equations, resulting from the
nineequation and sixequation strategies mentioned above, have been solved
by using a subroutine based on QR decomposition of the system matrix. This
subroutine has been obtained from the IMSL software package.
It is seen from Figure 4.2 that many of the collocation points lie on edges
and six of them lie on corners of the surface of the oneeighth sphere. These
results, displayed in Figure 4.3, show a comparison of the BCM (from equation
(4.25)) and HBCM (from equation (4.44)) results with the exact solution of the
problem [167]. The ﬁrst and last points along the axis lie on corners, the rest
lie along an edge. The agreement between the exact, BCM and nineequation
HBCM solution is seen to be excellent.
4.4.2 Surface Stresses
Stresses on the inner (R = a) and on the outer (R = b) surface of the sphere,
obtained from equation (4.33) and Hooke’s law, are shown in Figure 4.4. The
nodes are chosen at the centroids of the boundary elements. The agreement
between the numerical and analytical solutions is seen to be excellent.
© 2005 by Taylor & Francis Group, LLC
88 CHAPTER 4. LINEAR ELASTICITY
0 5 10 15 20 25 30 35 40
2
1.5
1
0.5
0
0.5
1
1.5
2
Nodes on the inner surface R=a
S
t
r
e
s
s
c
o
m
p
o
n
e
n
t
s
(a)
0 5 10 15 20 25 30 35 40
0.5
0.4
0.3
0.2
0.1
0
0.1
0.2
0.3
0.4
0.5
Nodes on the outer surface R=b
S
t
r
e
s
s
c
o
m
p
o
n
e
n
t
s
(b)
Figure 4.4: Hollow sphere under internal pressure. Stresses (a) on the inner
surface R = a and (b) on the outer surface R = b. Exact solutions —. Numer
ical solutions from the medium mesh: σ
θθ
= σ
φφ
∗ ∗ ∗ ∗, σ
RR
◦ ◦ ◦ ◦ (from
[100])
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
1.2
1
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
Radial coordinate
S
t
r
e
s
s
c
o
m
p
o
n
e
n
t
s
Figure 4.5: Hollow sphere under internal pressure. Stresses as functions of
radius along the line x
1
= x
2
= x
3
. Exact solutions —. Numerical solutions
from the ﬁne mesh: σ
θθ
= σ
φφ
∗ ∗ ∗ ∗, σ
RR
◦ ◦ ◦ ◦ (from [103])
© 2005 by Taylor & Francis Group, LLC
4.4. NUMERICAL RESULTS 89
1 1.02 1.04 1.06 1.08 1.1 1.12
1
0.95
0.9
0.85
0.8
0.75
0.7
0.65
Radial coordinate
: Exact solution
: BCM new
: BCM standard
R
a
d
i
a
l
s
t
r
e
s
s
(a)
1 1.02 1.04 1.06 1.08 1.1 1.12
0.35
0.4
0.45
0.5
0.55
Radial coordinate
: Exact solution
: BCM new
: BCM standard
C
i
r
c
u
m
f
e
r
e
n
t
i
a
l
s
t
r
e
s
s
(b)
Figure 4.6: Hollow sphere under internal pressure. Radial and circumferential
stresses (σ
rr
and σ
θθ
) as functions of radial distance from the center of the
sphere. The new and standard BCM solutions from the ﬁne mesh, together
with exact solutions, for points very close to the inner surface of the sphere
(from [104])
1 1.5 2 2.5 3 3.5 4
1
0.5
0
0.5
BCM solution for σ
θθ
BCM solution for σ
rr
Exact σ
θθ
Exact σ
rr
S
t
r
e
s
s
c
o
m
p
o
n
e
n
t
s
Radial coordinate
Figure 4.7: Hollow sphere under internal pressure. Radial and circumferential
stresses as functions of the radial distance from the center of the sphere. Results
from the new BCM (ﬁne mesh) together with the exact solution (from [104])
© 2005 by Taylor & Francis Group, LLC
90 CHAPTER 4. LINEAR ELASTICITY
4.4.3 Internal Stresses Relatively Far from the Bounding
Surface
Internal stresses along the line x
1
= x
2
= x
3
, obtained from equation (4.52) and
Hooke’s law, appear in Figure 4.5. Excellent agreement is observed between the
numerical and analytical solutions.
4.4.4 Internal Stresses Very Close to the Bounding Sur
face
Numerical results for stress components, from the standard (equation (4.52))
and the new (equation (4.56)) BCM, are shown in Figures 4.6 (a) and (b),
respectively. The results from the standard BCM exhibit large errors whereas
results from the new BCM faithfully track the exact solutions in both cases.
Finally, Figure 4.7 gives the global picture for stresses throughout the sphere.
The new BCM performs beautifully, even at points that are very close to the
surfaces of the hollow sphere.
Appendix A
An Explicit Form of Equation (4.40)
Using the equations :
u
i
(y) =
27
¸
α=1
ˆ
β
m
α
u
αi
(z
1
, z
2
, z
3
)
u
i
(x) =
3
¸
α=1
ˆ
β
P
α
u
αi
(z
1
, z
2
, z
3
)
u
i,
(x)[y
−x
] =
12
¸
α=4
ˆ
β
P
α
u
αi
(z
1
, z
2
, z
3
)
(the last equation above can be proved from equation (4.33)), the integral in
equation (4.40) can be written as:
I
1
=
L
m
U
ik
(x, y)
¸
12
¸
α=4
(
ˆ
β
m
α
−
ˆ
β
P
α
)σ
αij
+
27
¸
α=13
ˆ
β
m
α
σ
αij
¸
jnt
dz
t
−
L
m
Σ
ijk
(x, y)
¸
12
¸
α=1
(
ˆ
β
m
α
−
ˆ
β
P
α
)u
αi
+
27
¸
α=13
ˆ
β
m
α
u
αi
¸
jnt
dz
t
where the fact that σ
αij
= 0 for α = 1, 2, 3 has been used.
© 2005 by Taylor & Francis Group, LLC
4.4. NUMERICAL RESULTS 91
Appendix B
Proof of Equation (4.42)
u
i
(y) =
27
¸
α=1
ˆ
β
m
α
u
αi
(z
1
, z
2
, z
3
)
u
i,n
(y) =
27
¸
α=4
ˆ
β
m
α
u
αi,n
(z
1
, z
2
, z
3
)
since u
αi
, α = 1, 2, 3, are constant.
Now,
u
i,n
(y)
constant
≡
12
¸
α=4
ˆ
β
m
α
u
αi,n
=
¸
S
,N
m
β
i
where the last equality is obtained by observing equation (4.41).
Let
u
i,n
(y)
linear
≡
27
¸
α=13
ˆ
β
m
α
u
αi,n
(z
1
, z
2
, z
3
) =
27
¸
α=13
m
β
α
u
αi,n
(z
1
, z
2
, z
3
)
Now, with n = 1,
27
¸
α=13
m
β
α
u
α,1
= 2
m
β
13
+
m
β
20
u
4
+ 2
m
β
15
u
5
+ 2
m
β
18
u
6
+, ......, +
k
2
m
β
19
+k
1
m
β
20
u
12
=
12
¸
α=4
¸
R
,1
m
β
α−3
u
α
(z
1
, z
2
, z
3
)
Similar expressions can be obtained for n = 2,3.
In the above, a vector displacement shape function is deﬁned as:
u
α
= [u
α1
, u
α2
, u
α3
]
T
(where T denotes the transpose of the vector), and the constants k
1
and k
2
are
deﬁned in Table 4.1
© 2005 by Taylor & Francis Group, LLC
92 CHAPTER 4. LINEAR ELASTICITY
Appendix C
Conversion of Equation (4.43)
m
I 2
= −
12
¸
α=4
¸
S
m
[U
ik
(x, y)σ
αij
−Σ
ijk
(x, y)u
αi
] n
j
dS
¸
R
,N
m
β
α−3
+
3
¸
α=1
¸
S
m
Σ
αjk
(x, y)n
j
dS
¸
S
,N
m
β
α
−
3
¸
α=1
u
α,n
(x)
S
m
Σ
αjk
(x, y)n
j
dS
Applying Kaplan [69]’s formula (see [116], [109]), to the ﬁrst term above (its
integrand is O(1/r)), one gets the contour integral:
m
I 2
1
= −
12
¸
α=4
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jst
z
s
dz
t
¸
R
,N
m
β
α−3
The remaining terms cancel on a singular element (see equation (4.41)),
while, on a nonsingular element, one gets (see equation (4.26)):
m
I 2
2
= −
3
¸
α=1
¸
L
m
D
αjk
dz
j
¸
S
,N
m
β −
P
β
α
© 2005 by Taylor & Francis Group, LLC
Chapter 5
SHAPE SENSITIVITY
ANALYSIS
Shape sensitivity analysis of surface and internal displacements and stresses,
obtained from the BCM for 3D linear elasticity, is the subject of this chapter.
Further details are available in [100, 103].
5.1 Sensitivities of Boundary Variables
The starting point in this chapter is the standard BIE (1.26) collocated at an
internal point ξ . (This regularized BIE is valid both at an internal point ξ ∈ B
as well as at a boundary point x ∈ ∂B.) The sensitivity (total or material
derivative) of this equation is taken with respect to a design variable b. The
resulting sensitivity BIE is split into three parts. The ﬁrst part vanishes and
the surface integrals in the second and third parts are systematically converted
into contour integrals by using Stokes’ theorem.
5.1.1 Sensitivity of the BIE
Equation (4.1), at an internal point ξ ∈ ∂B is rewritten as:
0 =
∂B
[U
ik
(ξ, y)σ
ij
(y, b) −Σ
ijk
(ξ, y)[u
i
(y, b) −u
i
(ξ, b)]] n
j
(y)dS(y)
≡
∂B
F
jk
(ξ, y, b)n
j
(y)dS(y) (5.1)
As mentioned above, b is a shape design variable and the spatial coordinates of
the source and ﬁeld points depend on b , i.e. ξ(b), y(b).
From equation (5.1), one has:
93
© 2005 by Taylor & Francis Group, LLC
94 CHAPTER 5. SHAPE SENSITIVITY ANALYSIS
F
jk
(ξ, y, b) = U
ik
(ξ, y)σ
ij
(y, b) −Σ
ijk
(ξ, y)[u
i
(y, b) −u
i
(ξ, b)] (5.2)
Now the (total) sensitivity of a function f(ξ(b), y(b), b), in a materials derivative
sense, is deﬁned as :
∗
f≡
df
db
=
∂f
∂ξ
k
∗
ξ
k
+
∂f
∂y
k
∗
y
k
+
∂f
∂b
(5.3)
while its partial sensitivity is deﬁned as :
f≡
∂f
∂b
(5.4)
It should be noted that, for u
i
(y, b):
u
i
=
∗
u
i
−u
i,k
∗
y
k
,
σ
ij
=
∗
σ
ij
−σ
ij,k
∗
y
k
(5.5)
where , k ≡
∂
∂y
k
; and similarly for u
i
(ξ, b).
Taking the sensitivity (total derivative) of equation (5.1) with respect to b,
one gets:
0 =
∗
ξ
r
∂B
∂F
jk
(ξ, y, b)
∂ξ
r
n
j
(y)dS(y)
+
∂B
∂F
jk
(ξ, y, b)
∂y
r
∗
y
r
n
j
(y)dS(y) +
∂B
F
jk
(ξ, y, b)
d
db
[n
j
(y)dS(y)]
+
∂B
U
ik
(x, y)
σ
ij
(y, b) −Σ
ijk
(ξ, y)(
u
i
(y, b)−
u
i
(ξ, b))
n
j
(y)dS(y)
(5.6)
It should be noted that the last integrand above is
∂F
jk
(ξ,y,b)
∂b
.
The ﬁrst integral on the righthand side of equation (5.6) is zero because
the integral in equation (5.1) vanishes for all values of ξ ∈ B. Let the second
and third integrals together be called I
k
and the last integral J
k
. Thus:
I
k
+J
k
= 0 (5.7)
Each of these surface integrals will be converted to line integrals in the next
two sections.
5.1.2 The Integral I
k
The surface integral I
k
is converted to a sum of contour integrals in this section.
As mentioned before, a point ξ, inside the body, is considered ﬁrst. From
Bonnet and Xiao [15] (see also Petryk and Mr´oz [128]):
∗
n
j
= −
∗
y
r,j
n
r
+
∗
y
r,m
n
r
n
m
n
j
(5.8)
© 2005 by Taylor & Francis Group, LLC
5.1. SENSITIVITIES OF BOUNDARY VARIABLES 95
∗
dS
dS
=
∗
y
r,r
−
∗
y
r,m
n
r
n
m
(5.9)
so that:
d
db
[n
j
dS] =
∗
y
r,r
n
j
−
∗
y
r,j
n
r
dS (5.10)
Using equation (5.10) in (5.6), one has:
I
k
=
∂B
F
jk,r
∗
y
r
n
j
+F
jk
∗
y
r,r
n
j
−F
jk
∗
y
r,j
n
r
dS (5.11)
Since F
jk,j
= 0 (except at a point of singularity  here ξ is an internal point),
the above expression can be written as:
I
k
=
∂B
¸
F
jk
∗
y
r
,r
n
j
−
F
jk
∗
y
r
,j
n
r
dS (5.12)
Let ∂B = ∪S
m
and let L
m
be the bounding contour of the boundary element
S
m
. Using Stokes’ theorem in the form (see Toh and Mukherjee [168]):
S
m
(F
,r
n
j
−F
,j
n
r
) dS(y) =
L
m
qjr
Fdy
q
(5.13)
with F ≡ F
jk
∗
y
r
, one gets:
I
k
=
M
¸
m=1
L
m
qjr
F
jk
∗
y
r
dy
q
(5.14)
It is useful to state here that formula (5.14) is a general one, in the sense
that, a surface integral I (over a closed surface ∂B) of any divergencefree vector
function F(y) , of the form:
I =
∂B
F
j
(y) n
j
(y) dS(y) (5.15)
has the sensitivity expression :
∗
I
=
M
¸
m=1
L
m
jnt
F
j
(y)
∗
y
n
dy
t
(5.16)
Of course, from Gauss’ theorem, I = 0. Therefore,
∗
I
= 0.
One can show that:
M
¸
m=1
L
m
qjr
F
jk
dy
q
=
∂B
F
jk,r
n
j
dS = 0 ⇒
M
¸
m=1
∗
ξ
r
L
m
qjr
F
jk
dy
q
= 0
(5.17)
© 2005 by Taylor & Francis Group, LLC
96 CHAPTER 5. SHAPE SENSITIVITY ANALYSIS
Since
∗
y
r
=
∗
ξ
r
+
∗
z
r
, one can replace
∗
y
r
with
∗
z
r
in equation (5.14). Also, dy
q
= dz
q
since dx
q
= 0 at a ﬁxed source point.
Substituting equation (5.2) into (5.14) (with y replaced by z) , one gets:
I
k
(ξ) =
M
¸
m=1
L
m
[U
ik
(ξ, y)σ
ij
(y, b) −Σ
ijk
(ξ, y)(u
i
(y, b) −u
i
(ξ, b))]
jnt
∗
z
n
dz
t
(5.18)
Since
∗
z
n
is O(r) as r = y − x → 0, the above expression is completely
regular. Therefore, it remains valid at a boundary point x ∈ ∂B.
5.1.3 The Integral J
k
The surface integral J
k
is converted to a sum of contour integrals in this section.
5.1.3.1 Shape functions for partial sensitivities  a simple example
Series expressions for the partial sensitivities of displacements and stresses, in
terms of global BCM shape functions, are derived next.
It is useful to start with a very simple example. Let:
f(y) = a
0
+a
1
y +a
2
y
2
(5.19)
With the change of variables:
y = x +z (5.20)
one can write:
f(x, z) = ˆ a
0
(x) + ˆ a
1
(x)z + ˆ a
2
(x)z
2
(5.21)
where:
ˆ a
0
= a
0
+a
1
x +a
2
x
2
= f(x)
ˆ a
1
= a
1
+ 2a
2
x =
d
dx
f(x)
ˆ a
2
= a
2
(5.22)
It is easy to show that taking sensitivities of the above equations results in:
∗
f (y) =
∗
a
0
+
∗
a
1
y+
∗
a
2
y
2
+a
1
∗
y
+2a
2
y
∗
y
(5.23)
f (y) =
∗
f (y) −f
,y
∗
y
=
∗
a
0
+
∗
a
1
y+
∗
a
2
y
2
(5.24)
© 2005 by Taylor & Francis Group, LLC
5.1. SENSITIVITIES OF BOUNDARY VARIABLES 97
As expected, the partial sensitivity of f(x) is its sensitivity at a ﬁxed point
in space.
Now, with the change of variables (5.20), one has:
f (x, z) =
a
0
(x)+
a
1
(x)z+
a
2
(x)z
2
(5.25)
where
a
k
, k = 0, 1, 2, are related to
∗
a
k
in the same manner as ˆ a
k
are related
to a
k
in equation (5.22).
5.1.3.2 BCM interpolation functions
The displacement and stress interpolation functions for the BCM are considered
next.
One starts with equation (4.8) for the displacements. Following the proce
dure outlined above for the simple example, it is easy to show that:
u
i
=
27
¸
α=1
∗
β
α
u
αi
(y
1
, y
2
, y
3
) =
27
¸
α=1
β
α
(x
1
, x
2
, x
3
)u
αi
(z
1
, z
2
, z
3
) (5.26)
where,
β
α
, α = 1, 2, ...27, are related to
∗
β
α
in the same manner as
ˆ
β
α
are related
to β
α
(see equations 4.10, 4.11 and 4.12), i.e.:
β
i
=
27
¸
α=1
S
iα
(x
1
, x
2
, x
3
)
∗
β
α
, i = 1, 2, 3 (5.27)
β
k
=
27
¸
α=1
R
nα
(x
1
, x
2
, x
3
)
∗
β
α
, k = 4, 5, ...12, n = k −3 (5.28)
β
α
=
∗
β
α
, α = 13, 14, .., 27 (5.29)
Of course, the partial sensitivities for the stresses can now be expressed as:
σ
ij
=
27
¸
α=1
∗
β
α
σ
αij
(y
1
, y
2
, y
3
) =
27
¸
α=1
β
α
(x
1
, x
2
, x
3
)σ
αij
(z
1
, z
2
, z
3
) (5.30)
5.1.3.3 The ﬁnal form of J
k
The conversion procedure is entirely analogous to the derivation of the primary
BCM equation presented in Chapter 4. Series expansions (5.26) and (5.30) are
ﬁrst substituted into the last integral on the righthand side of the sensitivity
BIE (5.6). The potential functions are the same as before. Finally, the surface
integral J
k
is converted into a sum of contour integrals. The result is:
© 2005 by Taylor & Francis Group, LLC
98 CHAPTER 5. SHAPE SENSITIVITY ANALYSIS
J
k
(x) =
1
2
M
¸
m=1
27
¸
α=13
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jnt
z
n
dz
t
β
m
α
+
M
¸
m=1
12
¸
α=4
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jnt
z
n
dz
t
β
m
α
+
M
¸
m=1
m/ ∈S
3
¸
α=1
¸
L
m
D
αjk
dz
j
¸
β
m
α
−
β
P
α
(5.31)
It is interesting to comment on the physical meaning of the quantities
β
P
α
for the case when the surface source point P is a regular oﬀcontour boundary
point (ROCBP). Comparing equations (4.10  4.12) with (5.27  5.29), it is clear
that
ˆ
β
k
=
β
k
, k = 1, 2, 3, .., 27. Now, the quantities
β
P
α
can be easily interpreted
in terms of the partial sensitivities of displacements and their derivatives from
equations (4.32  4.36). For example,
β
P
k
=
u
k
(P), k = 1, 2, 3, etc.
5.1.4 The BCM Sensitivity Equation
An explicit form of the BCM sensitivity equation is now derived.
On any boundary element:
∗
β= T
−1
∗
a +(T
−1
)
∗
a (5.32)
in which it is convenient to evaluate the sensitivity of T
−1
from the formula:
(T
−1
)
∗
= −T
−1
∗
T
T
−1
(5.33)
Expressions (5.27  5.29) for
β
α
are substituted into (5.31), and (5.32) is
used to write
∗
β in explicit form. Next, an explicit expression for I
k
is obtained
by substituting the series expressions (4.8) for u
i
and (4.9) for σ
ij
into (5.18).
Finally, the explicit expression for I
k
(obtained from (5.18)) and the explicit
expression for J
k
(obtained from (5.31)) are input into the BIE sensitivity equa
tion (5.6) (see also 5.7) evaluated at a general boundary point x ∈ ∂B. The
result is:
0 =
M
¸
m=1
27
¸
α=13
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jnt
∗
z
n
dz
t
¸
m
T
−1
m
a
α
+
M
¸
m=1
12
¸
α=4
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jnt
∗
z
n
dz
t
¸
R
m
T
−1
m
a
α−3
© 2005 by Taylor & Francis Group, LLC
5.2. SENSITIVITIES OF SURFACE STRESSES 99
−
M
¸
m=1
m/ ∈S
3
¸
α=1
¸
L
m
Σ
αjk
jnt
∗
z
n
dz
t
¸
S
m
T
−1
m
a −
P
T
−1
P
a
¸
α
+
1
2
M
¸
m=1
27
¸
α=13
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jnt
z
n
dz
t
×
¸
m
T
−1
∗
a
m
+
m
T
−1
∗
m
a
α
+
M
¸
m=1
12
¸
α=4
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jnt
z
n
dz
t
×
¸
R
m
T
−1
∗
a
m
+
m
T
−1
∗
m
a
α−3
+
M
¸
m=1
m/ ∈S
3
¸
α=1
¸
L
m
D
αjk
dz
j
×
¸
S
m
T
−1
∗
a
m
−
P
T
−1
∗
a
P
+
m
T
−1
∗
m
a −
P
T
−1
∗
P
a
¸
α
(5.34)
Comparison of the above sensitivity equation (5.34) with the standard BCE
(4.25) reveals that the integrals in its last three terms are identical to those in
the standard BCE. Therefore, its discretized form can be written with the same
coeﬃcient matrix A as for the standard BCM, i.e. :
K
∗
a= h (5.35)
where the righthand side vector h = −
∗
K
a can be computed from equation
(5.34) by using the boundary values of the primary variables a that are known
at this stage.
Finally, the usual switching of columns leads to:
A
∗
x= B
∗
y
+h (5.36)
where
∗
x contains the unknown and
∗
y
the known values of boundary sensitivities.
In many applications,
∗
y
= 0
5.2 Sensitivities of Surface Stresses
The ﬁrst step is to use equation (5.32) to ﬁnd
∗
β on each element.
There are at least four ways to ﬁnd stress sensitivities on the surface of a
body.
© 2005 by Taylor & Francis Group, LLC
100 CHAPTER 5. SHAPE SENSITIVITY ANALYSIS
5.2.1 Method One
Equation (4.9) is diﬀerentiated to give:
∗
σ
ij
=
27
¸
α=1
∗
β
α
σ
αij
(y
1
, y
2
, y
3
) +
27
¸
α=1
β
α
∗
σ
αij
(
∗
y
1
,
∗
y
2
,
∗
y
3
) (5.37)
Note that
∗
σ
αij
(
∗
y
1
,
∗
y
2
,
∗
y
3
) involves sensitivities of the ﬁeld point coordinates
(y
1
, y
2
, y
3
).
5.2.2 Method Two
Sensitivities of displacement gradients v
ij
≡ u
i,j
are computed by diﬀerentiating
equation (4.33). The result is:
∗
v
ij
(x)
=
∗
ˆ
β
4
∗
ˆ
β
7
∗
ˆ
β
10
∗
ˆ
β
5
∗
ˆ
β
8
∗
ˆ
β
11
∗
ˆ
β
6
∗
ˆ
β
9
∗
ˆ
β
12
¸
¸
¸
¸
¸
¸
¸
x
(5.38)
where, by diﬀerentiating equation (4.11), one has:
∗
ˆ
β
k
=
27
¸
α=1
R
nα
(x
1
, x
2
, x
3
)
∗
β
α
+
27
¸
α=1
∗
Rnα
(
∗
x
1
,
∗
x
2
,
∗
x
3
)β
α
k = 4, 5, ...12, n = k −3 (5.39)
Note that
∗
Rnα
(
∗
x
1
,
∗
x
2
,
∗
x
3
) involves sensitivities of the source point coordinates
(x
1
, x
2
, x
3
)
Finally, Hooke’s law is used to determine the stress sensitivities from the
sensitivities of the displacement gradients.
5.2.3 Method Three
One writes:
∗
v
ij
=
v
ij
+v
ij,k
∗
x
k
(5.40)
Now:
v
ij,k
= u
i,jk
(5.41)
Also, from equation (4.33):
© 2005 by Taylor & Francis Group, LLC
5.3. SENSITIVITIES OF VARIABLES AT INTERNAL POINTS 101
v
ij
(x)
=
ˆ
β
4
ˆ
β
7
ˆ
β
10
ˆ
β
5
ˆ
β
8
ˆ
β
11
ˆ
β
6
ˆ
β
9
ˆ
β
12
¸
¸
¸
¸
¸
¸
¸
x
(5.42)
with, from equations (4.11) and (5.28):
ˆ
β
k
=
β
k
=
27
¸
α=1
R
nα
(x
1
, x
2
, x
3
)
∗
β
α
k = 4, 5, ..., 12, n = k −3 (5.43)
The curvature expressions needed in equation (5.40) are available in terms
of β
α
, α = 13, 14, ...27, from equations (4.34  4.36).
5.2.4 Method Four
The starting point is, again, equation (5.40). The term v
ij,k
on the righthand
side of equation (5.40) is treated in the same fashion as in Section 5.2.3 . The
other term,
v
ij
, is treated diﬀerently. It is ﬁrst observed that the operators , k
and commute and that
u
i
(P) =
β
P
i
for i = 1, 2, 3. Therefore, one has:
v
ij
= (u
i,j
)
= (
u
i
)
,j
= (
β
P
i
)
,j
(5.44)
It follows from equation (5.27) that:
v
ij
= (
β
P
i
)
,j
=
27
¸
α=1
S
iα,j
(x
1
, x
2
, x
3
)
∗
β
α
, i, j = 1, 2, 3 (5.45)
5.3 Sensitivities of Variables at Internal Points
Boundary contour integral equations of the sensitivities of internal displace
ments and stresses are derived in this section. Further details are available
[103].
5.3.1 Sensitivities of Displacements
The starting point is equation (1.21), the displacement BIE at an internal point
ξ. This equation is written as:
© 2005 by Taylor & Francis Group, LLC
102 CHAPTER 5. SHAPE SENSITIVITY ANALYSIS
u
k
(ξ, b) =
∂B
[U
ik
(ξ, y)σ
ij
(y, b) −Σ
ijk
(ξ, y)u
i
(y, b)] n
j
(y)dS(y) (5.46)
Deﬁne:
G
jk
(ξ, y, b) = U
ik
(ξ, y)σ
ij
(y, b) −Σ
ijk
(ξ, y)u
i
(y, b) (5.47)
Taking the sensitivity of equation (5.46), one gets:
∗
u
k
(ξ, b) =
∗
ξ
r
∂B
∂G
jk
(ξ, y, b)
∂ξ
r
n
j
(y)dS(y)
+
∂B
∂G
jk
(ξ, y, b)
∂y
r
∗
y
r
n
j
(y)dS(y)
+
∂B
G
jk
(ξ, y, b)
d
db
[n
j
(y)dS(y)]
+
∂B
U
ik
(ξ, y)
σ
ij
(y, b) −Σ
ijk
(ξ, y)
u
i
(y, b)
n
j
(y)dS(y)
(5.48)
Let the ﬁrst term on the righthand side of equation (5.48) be called J
1k
,
the second and third integrals together be called J
2k
and the last integral be
J
3k
. Therefore, one has:
∗
u
k
(ξ, b) = J
1k
+J
2k
+J
3k
(5.49)
It is obvious that:
J
1k
= u
k,r
(ξ)
∗
ξ
r
(5.50)
Using exactly the same procedure described in Section 5.1.2, one gets:
J
2k
=
M
¸
m=1
L
m
jnt
G
jk
∗
z
n
dz
t
(5.51)
Finally,
J
3k
=
u
k
(5.52)
Substituting equations (5.50, 5.51, 5.52 and 5.47) into (5.49), one obtains
the expression:
∗
u
k
(ξ, b) = u
k,p
(ξ)
∗
ξ
p
+
M
¸
m=1
L
m
[U
ik
(ξ, y)σ
ij
(y, b) −Σ
ijk
(ξ, y)u
i
(y, b)]
jnt
∗
z
n
dz
t
+
u
k
(5.53)
© 2005 by Taylor & Francis Group, LLC
5.3. SENSITIVITIES OF VARIABLES AT INTERNAL POINTS 103
An explicit form of equation (5.53) is obtained by writing the displacements
and stresses in terms of their interpolation functions from equations (4.8 and
4.9). Please refer to Section 5.1.3 for details of the treatment of the partial
sensitivity
u
k
. Finally, the boundary contour integral form of the displacement
sensitivity equation is:
∗
u
k
(ξ, b) = u
k,r
(ξ)
∗
ξ
r
+
M
¸
m=1
27
¸
α=13
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jnt
∗
z
n
dz
t
¸
m
T
−1
m
a
α
+
M
¸
m=1
12
¸
α=4
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jnt
∗
z
n
dz
t
¸
R
m
T
−1
m
a
α−3
−
M
¸
m=1
3
¸
α=1
¸
L
m
Σ
αjk
jnt
∗
z
n
dz
t
¸
S
m
T
−1
m
a
α
+
1
2
M
¸
m=1
27
¸
α=13
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jnt
z
n
dz
t
×
¸
m
T
−1
∗
a
m
+
m
T
−1
∗
m
a
α
+
M
¸
m=1
12
¸
α=4
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jnt
z
n
dz
t
×
¸
R
m
T
−1
∗
a
m
+
m
T
−1
∗
m
a
α−3
+
M
¸
m=1
3
¸
α=1
¸
L
m
D
αjk
dz
j
¸
S
m
T
−1
∗
a
m
+
m
T
−1
∗
m
a
α
(5.54)
5.3.2 Sensitivities of Displacement Gradients and Stresses
This time, the starting point is the displacement gradient BIE (1.28), which is
written as:
v
kr
(ξ, b) ≡ u
k,r
(ξ, b)
= −
∂B
[U
ik,r
(ξ, y)σ
ij
(y, b) −Σ
ijk,r
(ξ, y)u
i
(y, b)] n
j
(y)dS(y)
(5.55)
Now, one deﬁnes:
© 2005 by Taylor & Francis Group, LLC
104 CHAPTER 5. SHAPE SENSITIVITY ANALYSIS
H
jkr
(ξ, y) = U
ik,r
(ξ, y)σ
ij
(y, b) −Σ
ijk,r
(ξ, y)u
i
(y, b) (5.56)
One has H
jkr,j
= 0 at an internal point since:
H
jkr
(ξ, y) = −
∂G
jk
(ξ, y)
∂ξ
r
(5.57)
and G
jk,j
= 0.
The exact same reasoning as in the previous section 5.3.1 leads to an equa
tion for the sensitivities of displacement gradients at an internal point. This
is:
∗
v
kr
(ξ, b) = u
k,rp
(ξ)
∗
ξ
p
−
M
¸
m=1
L
m
[U
ik,r
(ξ, y)σ
ij
(y, b) −Σ
ijk,r
(ξ, y)u
i
(y, b)]
jnt
∗
z
n
dz
t
−
∂B
U
ik,r
(ξ, y)
σ
ij
(y, b) −Σ
ijk,r
(ξ, y)
u
i
(y, b)
n
j
(y)dS(y)
(5.58)
It is very interesting to verify (5.58) from another point of view. Diﬀerenti
ating equation (5.53) with respect to ξ
r
, one gets:
(
∗
u
k
)
,r
(ξ) = u
k,rp
(ξ)
∗
ξ
p
+u
k,p
(ξ)(
∗
ξ
p
)
,r
−
M
¸
m=1
L
m
[U
ik,r
(ξ, y)σ
ij
(y, b) −Σ
ijk,r
(ξ, y)u
i
(y, b)]
jnt
∗
z
n
dz
t
−
∂B
U
ik,r
(ξ, y)
σ
ij
(y, b) −Σ
ijk,r
(ξ, y)
u
i
(y, b)
n
j
(y)dS(y)
(5.59)
Normally, one would expect another term on the righthand side of the above
equation, namely:
M
¸
m=1
L
m
[U
ik
(ξ, y)σ
ij
(y, b) −Σ
ijk
(ξ, y)u
i
(y, b)]
jnt
(
∗
z
n
)
,r
dz
t
(5.60)
However, (see Section 5.1.2)
∗
z
n
=
∗
y
n
−
∗
ξ
n
, (
∗
z
n
)
,r
= −(
∗
ξ
n
)
,r
and:
M
¸
m=1
(
∗
ξ
n
)
,r
L
m
G
jk
(ξ, y, b)
jnt
dz
t
= 0 (5.61)
© 2005 by Taylor & Francis Group, LLC
5.3. SENSITIVITIES OF VARIABLES AT INTERNAL POINTS 105
since G
jk
is divergence free (see (5.17)). Therefore, the expression in (5.60)
vanishes.
Now, using the wellknown formula (which is valid for any suﬃciently smooth
function φ  see, for example, Haug et al. [63]):
(φ
,r
)
∗
= (
∗
φ)
,r
−φ
,p
(
∗
x
p
)
,r
(5.62)
with φ = u
k
, it is easy to show that equations (5.58) and (5.59) are consistent.
Finally, one writes the displacements and stresses in terms of their inter
polation functions in order to obtain an explicit form of equation (5.58). It
should be noted that the second term in the righthand side of equation (5.58)
is analogous to the integral on the righthand side of equation (5.53) (with G
replaced by H), while the last term in equation (5.58),
v
kr
, is analogous to an
expression for the displacement gradient at an internal point. The displacement
gradient BCE (4.52) is very useful for obtaining an explicit expression for this
integral. The ﬁnal result is:
∗
v
kr
(ξ, b) = u
k,rp
(ξ)
∗
ξ
p
−
M
¸
m=1
27
¸
α=13
¸
L
m
(σ
αij
U
ik,r
−u
αi
Σ
ijk,r
)
jnt
∗
z
n
dz
t
¸
m
T
−1
m
a
α
−
M
¸
m=1
12
¸
α=4
¸
L
m
(σ
αij
U
ik,r
−u
αi
Σ
ijk,r
)
jnt
∗
z
n
dz
t
×
¸
R
m
T
−1
m
a
α−3
+
M
¸
m=1
3
¸
α=1
¸
L
m
Σ
αjk,r
jnt
∗
z
n
dz
t
¸
S
m
T
−1
m
a
α
−
M
¸
m=1
27
¸
α=13
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jrt
dz
t
×
¸
m
T
−1
∗
a
m
+
m
T
−1
∗
m
a
α
+
M
¸
m=1
12
¸
α=4
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jnt
z
n
dz
t
×
¸
R
,r
m
T
−1
∗
a
m
+
m
T
−1
∗
m
a
α−3
−
M
¸
m=1
12
¸
α=4
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jrt
dz
t
×
© 2005 by Taylor & Francis Group, LLC
106 CHAPTER 5. SHAPE SENSITIVITY ANALYSIS
¸
R
m
T
−1
∗
a
m
+
m
T
−1
∗
m
a
α−3
+
M
¸
m=1
3
¸
α=1
¸
L
m
D
αjk
dz
j
¸
S
,r
m
T
−1
∗
a
m
+
m
T
−1
∗
m
a
α
+
M
¸
m=1
3
¸
α=1
¸
L
m
Σ
αjk
jrt
dz
t
¸
S
m
T
−1
∗
a
m
+
m
T
−1
∗
m
a
α
(5.63)
The curvatures u
k,rp
in the above equation can be obtained from (see [103]):
u
k,rp
(ξ) =
M
¸
m=1
27
¸
α=13
¸
L
m
(σ
αij
U
ik,r
−u
αi
Σ
ijk,r
)
jpt
dz
t
¸
m
T
−1
m
a
α
−
M
¸
m=1
12
¸
α=4
¸
L
m
(σ
αij
U
ik
−u
αi
Σ
ijk
)
jrt
dz
t
¸
R
,p
m
T
−1
m
a
α−3
+
M
¸
m=1
12
¸
α=4
¸
L
m
(σ
αij
U
ik,r
−u
αi
Σ
ijk,r
)
jpt
dz
t
¸
R
m
T
−1
m
a
α−3
+
M
¸
m=1
3
¸
α=1
¸
L
m
Σ
αjk
jrt
dz
t
¸
S
,p
m
T
−1
m
a
α
−
M
¸
m=1
3
¸
α=1
¸
L
m
Σ
αjk,r
jpt
dz
t
¸
S
m
T
−1
m
a
α
+
M
¸
m=1
3
¸
α=1
¸
L
m
D
αjk
dz
j
¸
S
,rp
m
T
−1
m
a
α
(5.64)
Finally, stress sensitivities can be easily obtained from
∗
v
kr
by using Hooke’s
law.
5.4 Numerical Results: Hollow Sphere
Sensitivity results are presented in this section for a thick hollow sphere under
internal pressure (see Section 4.4). As before, the inner and outer radii of the
sphere are 1 and 2 units, respectively, the shear modulus G = 1, the Poisson’s
ratio ν = 0.3 and the internal pressure is 1 unit. The design variable is the
inner radius a of the sphere.
A generic surface mesh on a oneeighth sphere is given in Figure 4.2. Three
levels of discretization, coarse, medium and ﬁne, are used in this work. Mesh
statistics are shown in Table 4.3.
© 2005 by Taylor & Francis Group, LLC
5.4. NUMERICAL RESULTS: HOLLOW SPHERE 107
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Radial coordinate
S
e
n
s
i
t
i
v
i
t
y
o
f
r
a
d
i
a
l
d
i
s
p
l
a
c
e
m
e
n
t
(a)
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Radial coordinate
S
e
n
s
i
t
i
v
i
t
y
o
f
r
a
d
i
a
l
d
i
s
p
l
a
c
e
m
e
n
t
(b)
Figure 5.1: Sensitivity of radial displacement along (a) the x
3
axis and (b) along
the line x
1
= 0, x
2
= x
3
. Exact solution —. Numerical solutions: coarse mesh
◦ ◦ ◦◦, medium mesh ∗ ∗ ∗∗, ﬁne mesh ++++ (from [100])
5.4.1 Sensitivities on Sphere Surface
5.4.1.1 Displacement sensitivities
Displacement sensitivities (from equation (5.34)), along various lines on the
sphere surface, for diﬀerent discretizations, appear in Figure 5.1. (A typical
displacement proﬁle for this example appears in Figure 4.3.) A linear design
velocity proﬁle:
∗
R
=
b −R
b −a
(5.65)
is used here. It is seen that the numerical results for the coarse mesh exhibit
large errors. The reason for this is under investigation. However, they do
appear to converge to the exact solution with increasing mesh density. Please
see Chandra and Mukherjee [22] for a discussion of analytical solutions for
design sensitivities for various examples.
5.4.1.2 Stress sensitivities
Stress sensitivities on the surface of the sphere are calculated from equations
(5.38  5.39). (Stress proﬁles for this example appear in Figure 4.4). Numerical
solutions from the medium mesh, for stress sensitivities on the outer surface
R = b, agree well with the exact solution (Figure 5.2).
The situation, however, is somewhat tricky on the inner surface R = a.
Here, one has:
© 2005 by Taylor & Francis Group, LLC
108 CHAPTER 5. SHAPE SENSITIVITY ANALYSIS
0 5 10 15 20 25 30 35 40
1
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
Nodes on the outer surface R=b
S
t
r
e
s
s
s
e
n
s
i
t
i
v
i
t
i
e
s
Figure 5.2: Stress sensitivities on the outer surface R = b. Exact solutions —.
Numerical solutions from the medium mesh:
∗
σ
RR
+ + + +,
∗
σ
θθ
◦◦◦◦,
∗
σ
φφ
∗∗∗∗
(from [100])
∗
σ
θθ
=
σ
θθ
+
∂σ
θθ
∂R
∗
R
(5.66)
(and similarly for the other components of stress). The exact solution for
σ
θθ
=
σ
φφ
is 120/49, which is positive, while the convected term in equation
(5.66) is −12/7, which is negative. As can be seen from Figures (5.35.4), this
situation makes it quite diﬃcult to calculate
∗
σ
θθ
and
∗
σ
φφ
accurately, especially
when accuracy is measured in terms of percentage errors, because one must
calculate the diﬀerence between two numbers that are reasonably close.
Table 5.1 shows percentage root mean square errors in
∗
σ
θθ
and
∗
σ
φφ
, respec
tively. These are deﬁned as:
=
100
f
exact
1
n
n
¸
i=1
(f
exact
−f
i numer
)
2
(5.67)
It is seen that while the errors on the outer surface are very low, those on the
inner surface, even with the ﬁne mesh, are quite high. Perhaps further work on
this problem, including the development of a diﬀerent algorithm for calculation
of surface stress sensitivities, needs to be carried out.
5.4.2 Sensitivities at Internal Points
5.4.2.1 Sensitivity of radial displacement
Figure 5.5(a) shows the sensitivity of radial displacement along the line x
1
=
x
2
= x
3
. These results are obtained from equation (5.54) which is used after ﬁrst
© 2005 by Taylor & Francis Group, LLC
5.4. NUMERICAL RESULTS: HOLLOW SPHERE 109
medium mesh ﬁne mesh
∗
σ
θθ
1.99 1.62
R = b ∗
σ
φφ
1.77 2.01
∗
σ
θθ
25.45 15.15
R = a ∗
σ
φφ
26.49 14.88
Table 5.1: Percentage root mean square errors in sensitivities of stress compo
nents (from [100])
0 5 10 15 20 25 30 35 40
3
2
1
0
1
2
3
4
Nodes on the inner surface R=a
S
t
r
e
s
s
s
e
n
s
i
t
i
v
i
t
i
e
s
(a)
0 10 20 30 40 50 60 70
3
2
1
0
1
2
3
4
Nodes on the inner surface R=a
S
t
r
e
s
s
s
e
n
s
i
t
i
v
i
t
i
e
s
(b)
Figure 5.3: Stress sensitivities on the inner surface R = a. Exact solutions:
—. Numerical solutions from (a) the medium mesh and (b) the ﬁne mesh:
σ
θθ
◦ ◦ ◦◦,
∂σ
θθ
∂R
∗
R
+ + + +,
∗
σ
θθ
∗ ∗ ∗∗ (from [100])
© 2005 by Taylor & Francis Group, LLC
110 CHAPTER 5. SHAPE SENSITIVITY ANALYSIS
0 5 10 15 20 25 30 35 40
3
2
1
0
1
2
3
4
Nodes on the inner surface R=a
S
t
r
e
s
s
s
e
n
s
i
t
i
v
i
t
i
e
s
(a)
0 10
20 30 40 50 60 70
3
2
1
0
1
2
3
4
Nodes on the inner surface R=a
S
t
r
e
s
s
s
e
n
s
i
t
i
v
i
t
i
e
s
(b)
Figure 5.4: Stress sensitivities on the inner surface R = a. Exact solutions:
—. Numerical solutions from (a) the medium mesh and (b) the ﬁne mesh:
σ
φφ
◦ ◦ ◦◦,
∂σ
φφ
∂R
∗
R
+ + + +,
∗
σ
φφ
∗ ∗ ∗∗ (from [100])
solving the sensitivity boundary value problem (5.34). Again, the linear velocity
proﬁle (5.65) is used here. It is seen from Figure 5.5(a) that the agreement
between the numerical and analytical solutions is very good.
5.4.2.2 Sensitivities of stresses
Sensitivities of the stress components σ
θθ
and σ
RR
, along the line x
1
= x
2
= x
3
,
are presented in Figure 5.5(b). Again, very good agreement is observed between
the numerical and analytical solutions.
5.5 Numerical Results: Block with a Hole
5.5.1 Geometry and Mesh
This example is concerned with a rectangular block with a cylindrical hole of
circular crosssection, loaded in uniform remote tension. The BCM model is
fully threedimensional but the imposed boundary conditions are chosen such
that a state of plane strain prevails in the block and the numerical results
obtained from the BCM can be compared with Kirsch’s analytical solution for
the corresponding 2D plane strain problem. Of course, the analytical solution
is only available for an inﬁnitesimal hole in a slab and this is referred to as the
“exact” solution in this section of this chapter.
The side of a square face of the full block is 20 units, the hole diameter is 2
units and the thickness (in the x
3
direction) is 6 units. Oneeighth of the block
is modeled in order to take advantage of symmetry. The mesh on the front and
© 2005 by Taylor & Francis Group, LLC
5.5. NUMERICAL RESULTS: BLOCK WITH A HOLE 111
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
0
0.1
0.2
0.3
0.4
0.5
0.6
S
e
n
s
i
t
i
v
i
t
y
o
f
r
a
d
i
a
l
d
i
s
p
l
a
c
e
m
e
n
t
Radial coordinate
(a)
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
0.5
0
0.5
1
Radial coordinate
S
t
r
e
s
s
s
e
n
s
i
t
i
v
i
t
i
e
s
(b)
Figure 5.5: (a) Sensitivity of radial displacement along the line x
1
= x
2
= x
3
.
Exact solution —. Numerical solution from the ﬁne mesh: ◦◦◦◦. (b) Sensitivity
of stresses along the line x
1
= x
2
= x
3
. Exact solutions —. Numerical solutions
from the ﬁne mesh:
∗
σ
θθ
∗ ∗ ∗∗
∗
σ
RR
◦ ◦ ◦◦ (from [103])
0 1 2 3 4 5 6 7 8 9 10
0
1
2
3
4
5
6
7
8
9
10
X
X
2
1
(a)
a
R
φ
L
L
X
X
2
1
(b)
Figure 5.6: (a) Mesh on a quarter of the front and back faces of a rectangular
block with a circular cylindrical hole. (b) Design velocities for this example
(from [103])
© 2005 by Taylor & Francis Group, LLC
112 CHAPTER 5. SHAPE SENSITIVITY ANALYSIS
back faces (x
3
= 3 and x
3
= 0, respectively) of the oneeighth block is shown in
Figure 5.6(a). Four layers of triangles (in the thickness direction) constitute the
mesh on the remaining surfaces of the oneeighth block. The complete block
is loaded by uniform remote tensions in the x
1
and x
2
directions while u
3
= 0
on the faces normal to the x
3
axis in order to simulate plane strain conditions.
Of course, boundary conditions on the symmetry planes are applied in the
computer model of the oneeighth block in the usual way.
The design variable in this example is the hole radius a. The chosen design
velocity distribution in the slab is linear along any radial direction in any square
section normal to the x
3
axis and is independent of the x
3
coordinate. In other
words, referring to Figure 5.6(b), one has:
∗
R
=
L/ cos φ −R
L/ cos φ −a
for φ < π/4 (5.68)
and
∗
R
=
L/ sinφ −R
L/ sinφ −a
for φ ≥ π/4 (5.69)
5.5.2 Internal Stresses
Comparisons of numerical and analytical solutions for internal stresses in the
block are presented in Figures 5.7(a) and (b). Two cases are considered: stresses
along the line x
1
= x
2
, x
3
= 1.5 for uniaxial remote loading in the x
1
direc
tion, and along the line x
1
=
√
3x
2
, x
3
= 1.5 for equal biaxial loading. The
agreement between the analytical and numerical results is again seen to be very
good.
5.5.3 Sensitivities of Internal Stresses
This last example is a diﬃcult one. It is concerned with stress sensitivities for
the situation depicted in Figure 5.7(a).
The numerical and analytical results for this problem are shown in Figure
5.8. The internal point numerical results, which are obtained from independent
calculations, show acceptable agreement with the analytical solutions for the
stress sensitivities, even though the analytical solutions vary rapidly near the
hole.
© 2005 by Taylor & Francis Group, LLC
5.5. NUMERICAL RESULTS: BLOCK WITH A HOLE 113
0 1 2 3 4 5 6 7 8 9 10
0.4
0.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
X
=
X
S
t
r
e
s
s
c
o
m
p
o
n
e
n
t
s
1 2
(a)
0 1 2 3 4 5 6 7 8 9 10
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
1.3
1.4
1.5
X = 3 X
S
t
r
e
s
s
c
o
m
p
o
n
e
n
t
s
2
1
(b)
Figure 5.7: Stresses (a) along the line x
1
= x
2
, x
3
= 1.5 for uniaxial loading in
the x
1
direction and (b) along the line x
1
=
√
3x
2
, x
3
= 1.5 for equal biaxial
loading. Exact solutions —. Numerical solutions: σ
11
◦ ◦ ◦ ◦, σ
22
∗ ∗ ∗ ∗ (from
[103])
S
t
r
e
s
s
s
e
n
s
i
t
i
v
i
t
i
e
s
X X
0 1 2 3 4 5 6 7 8 9 10
0.4
0.3
0.2
0.1
0
0.1
0.2
0.3
1 2
=
Figure 5.8: Stress sensitivities along the line x
1
= x
2
, x
3
= 1.5 for uniaxial
loading in the x
1
direction. Exact solutions:
∗
σ
11
— ,
∗
σ
22
− − −. Numerical
solutions:
∗
σ
11
◦ ◦ ◦◦,
∗
σ
22
∗ ∗ ∗∗ (from [103])
© 2005 by Taylor & Francis Group, LLC
Chapter 6
SHAPE OPTIMIZATION
Shape optimization of 3D elasticity problems, with the BCM, is the subject of
this chapter. Further details are available in Shi and Mukherjee [150].
6.1 Shape Optimization Problems
An optimal shape design problem can be stated as a minimization problem
under certain constraints. Its general form can be stated as follows:
Minimize f(b) (6.1)
Subject to g
i
(b) ≥ 0, i = 1, ..., N
g
(6.2)
h
j
(b) = 0, j = 1, ..., N
h
(6.3)
b
()
k
≤ b
k
≤ b
(u)
k
, k = 1, ..., N (6.4)
in which b = b
1
, b
2
, ..., b
N
T
are the design variables, f(b) is the objective
function, and g
i
(b) and h
j
(b) are the inequality and equality constraints, re
spectively. Equation (6.4) gives side constraints that are used to limit the search
region of an optimization problem. Here, the parameters b
()
k
and b
(u)
k
denote
the lower and upper bounds, respectively, of the design variable b
k
.
The most common mathematical programming approaches, used in gradient
based optimization algorithms, are the successive linear programming (SLP)
and successive quadratic programming (SQP) methods. In the SQP method,
the optimization problem is approximated by expanding the objective function
in a second order Taylor series about the current values of the design variables,
while the constraints are expanded in a ﬁrst order Taylor series.
115
© 2005 by Taylor & Francis Group, LLC
116 CHAPTER 6. SHAPE OPTIMIZATION
The subroutine DNCONF from the IMSL library is coupled with a 3D BCM
code for elastic stress analysis in order to carry out the shape optimization ex
amples that are described in this chapter. This subroutine, that uses the SQP
method, is based on the FORTRAN subroutine NLPQL developed by Schit
tkowski [146]. The required gradients of the objective functions, constraints
etc. are calculated internally by the optimization subroutine by the ﬁnite dif
ference method. Of course, these sensitivities could also have been obtained by
the direct diﬀerentiation approach described in Chapter 5.
6.2 Numerical Results
Two illustrative shape optimization examples, in 3D linear elasticity, are solved
using the BCM coupled with the IMSL optimization subroutine mentioned
above. The ﬁrst example is that of optimizing the shape of a ﬁllet in a tension
bar whose volume is selected as the optimization function. An optimal shape
is sought that minimizes the volume (and therefore the weight when the bar
material has constant density) without causing yielding anywhere in the bar.
This is the axisymmetric version of the planar problem described in Phan et al.
[132]  Section 5.3. Of course, the full 3D BCM code is used here.
The second problem is concerned with a cube with an ellipsoidal cavity
loaded in remote triaxial tension. An interesting result in 2D [5, 132], for an
inﬁnite elastic plate with an elliptical cutout, loaded in remote biaxial tension, is
as follows. Let the semimajor and semiminor axes of the elliptical hole be a
1
and
a
2
, respectively, and the coordinate axes be centered at the center of the ellipse
with the x
1
and x
2
directions along the major and minor axes, respectively.
Also, let S
1
and S
2
be the remote tensile loadings in the coordinate directions.
Now, if a
2
/a
1
= S
2
/S
1
, then the tangential stress around the elliptical cutout
is uniform ! This problem has been solved in 2D by the BCM in [132] and an
elasticplastic version of this 2D problem has been solved in [169] (see, also,
[22]) with the BEM. The 3D elasticity problem is described in this chapter and
some interesting results are obtained.
6.2.1 Shape Optimization of a Fillet
The initial crosssection of the axisymmetric bar is shown in Figure 6.1. The
Young’s modulus, Poisson’s ratio and allowable von Mises stress are taken as
E = 10
7
psi, ν = 0.3 and ˆ σ
(V M)
= 120 psi, respectively. The bar is loaded by
a uniform axial tensile traction of 100 psi . The design surface is the surface
of revolution (initially a cone) obtained by revolving the curve ED about the
symmetry axis AB. The end circles of this surface of revolution are kept ﬁxed.
The variable boundary ED is modeled as a cubic spline (using the IMSL sub
routine DCSDEC) which is deﬁned by the ﬁxed end points E and D and by the
variable points C
k
, k = 1, 2, 3. The points E, C
k
, k = 1, 2, 3 and D have equally
spaced projections on the axis of symmetry while the radii r
k
, k = 1, 2, 3 are
the design variables.
© 2005 by Taylor & Francis Group, LLC
6.2. NUMERICAL RESULTS 117
9
9
4.5
4.5
τ
3
=100psi
20
z
r
A
B
C
C
1
C
2
C
3
D
E F
G
u
3
=0
τ
1
=0
τ
2
=0
r r r
1 2 3
Figure 6.1: Modeling of a bar with a ﬁllet (from [150])
Quadratic CIM9 elements (see Figure 4.1) are used in both the numerical
examples described in this chapter. A total of 160 elements are used to discretize
the surface of the bar. These are distributed as follows. There are 16 elements
on each of the end circles (on the planes z = 0 and z = 20, respectively) of the
bar. The surface of revolution described by the edge FE has 32 elements, the
one described by DC has 32 elements, and the design surface has 64 elements,
respectively.
The objective function is the volume of the axisymmetric object bounded
by the plane z = 9, the cylindrical surface r = 4.5 and the design surface. This
is:
φ(r
1
, r
2
, r
3
) =
z
D
z
G
π
r
2
(r
1
, r
2
, r
3
, z) −(4.5)
2
dz (6.5)
The axisymmetric shape of the body is maintained during the optimization
process and the design nodes C
k
, k = 1, 2, 3 are constrained to lie within the
triangle EGD. In addition, one has:
σ
(V M)
i
/ˆ σ
(V M)
≤ 1.0, 1 ≤ i ≤ n
s
(6.6)
where σ
(V M)
i
are the values of the von Mises stress at the centroids of certain
elements on the curved surface of the bar. Here, since the physical problem
is axisymmetric, the elements chosen are the ones whose centroids lie along
FEDC, so that n
s
= 16.
The usual deﬁnition of the von Mises stress is used. This is:
[σ
(V M)
]
2
= (3/2)s
ij
s
ij
(6.7)
in terms of s
ij
, the deviatoric components of the stress σ
ij
.
The ﬁnal converged solution, which is obtained after just 1 iteration, is
shown in Figure 6.2(a) while Figure 6.2(b) shows the objective function as a
© 2005 by Taylor & Francis Group, LLC
118 CHAPTER 6. SHAPE OPTIMIZATION
9 10 11 12 13 14 15 16
4
4.5
5
5.5
6
6.5
7
7.5
8
8.5
9
r
z
E
C
C
C
D
1
2
3
(a)
1 2 3 4 5
40
60
80
100
120
140
160
180
Iteration number
φ
/
π
0
(b)
Figure 6.2: (a) Optimal shape of the ﬁllet (b) Objective function φ as a function
of iteration number (from [150])
function of iteration number. The initial value of the objective function is 551.35
with one stress constraint being violated. The ﬁnal (converged) value is 150.39
when two stress constraints are active. Thus, the ﬁnal volume of the design
portion of the bar is about 27.28 % of its initial value. In the corresponding
2D problem [132], the design area reduces to 49.85% of its initial value.
6.2.2 Optimal Shapes of Ellipsoidal Cavities Inside Cubes
A cube with a centrally located ellipsoidal cavity, loaded in remote triaxial
tension, is shown in Figure 6.3. Of course, the faces of the cube, which are not
shown in this ﬁgure, are suitably restrained. Consistent units are used. This
time the shear modulus is taken as G = 10
5
and the Poisson’s ratio ν = 0.3.
The cube is of size 30×30×30. The cavity surface is the design surface with the
ellipsoid semi axes a
1
and a
2
as the design variables (a
3
= 1). The cube surface
is ﬁxed. Three cases of remote loadings (cases (1), (2) and (3)) are considered
: S
1
: S
2
: S
3
= 1 : 1 : 1; 2 : 2 : 1 and 2 : 1.5 : 1, respectively, with S
3
= 10
5
in
all cases.
The mesh consists of 8 identical CIM9 elements on each surface of the cube
and 72 CIM9 elements on the surface of the cavity. The cavity mesh evolves
with the changing shape of the cavity during the optimization process.
As stated at the start of this section, the objective here is to ﬁnd, if possible,
the shape of an ellipsoidal cavity that would have uniform stress (i.e. some
suitable measure of stress) on its surface for a given remote loading. As a test
case, the ﬁrst case of remote loading, i.e. S
1
: S
2
: S
3
= 1 : 1 : 1, is applied.
Starting with an ellipsoidal cavity with axis ratios a
1
: a
2
: a
3
= 3 : 2 : 1,
the cavity shape is sought that would make all the stress components on it
uniform. This, of course, is a spherically symmetric problem, and, as expected,
© 2005 by Taylor & Francis Group, LLC
6.2. NUMERICAL RESULTS 119
S
3
S
1
S
2
x
1
x
2
x
3
a
1
a
2
a
3
Figure 6.3: A cube with an ellipsoidal cavity under remote loading (from [150])
the optimal shape of the cavity is found to be a sphere.
The more interesting cases, of course, are loading cases (2) and (3). This
time, two diﬀerent scalar measures of stress on the cavity surface are considered.
The ﬁrst is σ
kk
, the trace of the stress tensor, and the second is the von Mises
stress. Uniform σ
kk
on a surface means that σ
nn
+σ
ss
+σ
tt
(where n is normal
and s and t are any two orthogonal directions, tangential to the surface at a
point on it) is also uniform on the surface. Since the cavity surface is unloaded,
this implies that σ
ss
+ σ
tt
is uniform for all points on the cavity surface. The
criterion σ
kk
is uniform on a cavity surface is treated as a generalization of the
2D case in which the tangential stress on a cutout surface was made uniform
[132].
The corresponding objective functions to be minimized are deﬁned as:
φ
1
(a
1
, a
2
) =
1
n
s
n
s
i=1
σ
(i)
kk
−σ
kk
2
(6.8)
φ
2
(a
1
, a
2
) =
1
n
s
n
s
i=1
σ
(V M)
i
2
−(σ
(V M)
)
2
2
(6.9)
In the above equations, n
s
is the number of elements on the design surface,
a superscript (i) denotes evaluation of the appropriate quantity at the centroid
of the i
th
element on the design surface, and an overbar denotes the mean value
of the appropriate quantity over the design surface.
© 2005 by Taylor & Francis Group, LLC
120 CHAPTER 6. SHAPE OPTIMIZATION
Load case Loads Ellipse semiaxes Objective function
Uniform σ
kk
on cavity surface
(2) 2, 2, 1 Start 2, 2, 1 u.b. 3,3,1 1.044
Optimal 2.654, 2.654, 1 0.0027
(3) 2, 1.5, 1 Start 2, 1.5, 1 u.b. 7,7,1 0.11057
Optimal 3.8022, 1.5608, 1 0.002286
Uniform σ
(V M)
on cavity surface
(2) 2, 2, 1 Start 2, 2, 1 u.b. 9,9,1 3.4086
Optimal 3.4105, 3.4105, 1 0.0489
Table 6.1: Results for cube with ellipsoidal cavity under remote loading. Upper
bound is denoted as u.b. Lower bounds for ellipse semiaxes are 1,1,1. The loads
are normalized with S
3
= 1 (from [150])
6.2.2.1 Uniform trace of the stress tensor over the cavity surface
The load cases (2) and (3) are tried for the ﬁrst objective function φ
1
. In each
case, inspired by the 2D case, the starting geometry of the ellipsoid is taken to
be compatible with the loads. The results appear in Table 6.1 in which the side
constraints on the design variables, as well as the initial and optimal values of
the design variables and the objective function, are shown.
Details are shown in Figures 6.4  6.5. Figure 6.4(a) shows the stress σ
kk
,
and its mean value, initially and at the end of the optimization process. It is
seen that the initial distribution of σ
kk
(dots) is quite widespread while the
optimal distribution (asterisks) is concentrated within a band around its mean
value σ
kk
. The corresponding objective function, as a function of iteration
number, is shown in Figure 6.4(b). It can be seen from Table 6.1 that, while
the optimal values of a
1
and a
2
are equal (i.e. the crosssection of the optimal
shaped cavity in the x
3
= 0 plane is a circle), the loading and geometry are not
compatible as in the 2D case (i.e. the optimal a
1
: a
2
: a
3
does not match the
load ratios).
The more diﬃcult case is load case (3). The stress σ
kk
and the objective
function for this case appear in Figures 6.5(a) and 6.5(b), respectively. Again, it
can be seen from Figure 6.5(a) that the initial distribution of σ
kk
(dots) is quite
widespread while the optimal distribution (asterisks) is concentrated within a
band around its mean value σ
kk
. It is seen from Figure 6.5(b) that in this case
the behavior of the objective function is initially oscillatory. It does, however,
eventually settle down to about 2.07% of its initial value. Examination of Table
6.1 again shows that the optimal cavity geometry is not compatible with the
loading in this case.
© 2005 by Taylor & Francis Group, LLC
6.2. NUMERICAL RESULTS 121
0 10 20 30 40 50 60 70 80
4.4
4.6
4.8
5
5.2
5.4
5.6
5.8
σ
k
k
Nodes at element centroids
(a)
1 2 3 4 5 0
0
0.2
0.4
0.6
0.8
1
1.2
1.4
Iteration number
φ
1
(b)
Figure 6.4: (a) The stresses σ
kk
and σ
kk
for load case (2). Initial : σ
kk
...,
σ
kk
−−− Optimal : σ
kk
∗ ∗∗, σ
kk
— (b) Objective function φ
1
as a function of
iteration number for load case (2) (from [150])
0 10 20 30 40 50 60 70 80
3.8
4
4.2
4.4
4.6
4.8
5
Nodes at element centroids
σ
k
k
(a)
0 5 10 15 20 25
0
0.2
0.4
0.6
0.8
1
1.2
1.4
Iteration number
φ
1
(b)
Figure 6.5: (a) The stresses σ
kk
and σ
kk
for load case (3). Initial : σ
kk
...,
σ
kk
−−− Optimal : σ
kk
∗ ∗∗, σ
kk
— (b) Objective function φ
1
as a function of
iteration number for load case (3) (from [150])
© 2005 by Taylor & Francis Group, LLC
122 CHAPTER 6. SHAPE OPTIMIZATION
0 10 20 30 40 50 60 70 80
3.5
4
4.5
5
5.5
6
6.5
7
7.5
(
σ
V
M
)
2
Nodes at element centroids
(
)
(a)
1 2 3 4 5 6 0
0
0.5
1
1.5
2
2.5
3
3.5
4
φ
2
Iteration number
(b)
Figure 6.6: (a) The stresses (σ
(V M)
)
2
and (σ
(V M)
)
2
for load case (2). Initial :
(σ
(V M)
)
2
..., (σ
(V M)
)
2
−−− Optimal : (σ
(V M)
)
2
∗ ∗∗, (σ
(V M)
)
2
— (b) Objective
function φ
2
as a function of iteration number for load case (2) (from [150])
6.2.2.2 Uniform von Mises stress over the cavity surface
The ﬁnal example considers the load case (2) with the second objective function
φ
2
. The results are summarized in Table 6.1 and details are shown in Figures
6.6(a) and 6.6(b). This time the converged value of the objective function is
1.43% of its initial value. Figure 6.6(a) shows a signiﬁcant diﬀerence between
the initial and ﬁnal values of the mean value of (σ
(V M)
)
2
. Also, the optimal
distribution of (σ
(V M)
)
2
(asterisks) lies in a wider band around its mean value
compared to the situations in Figures 6.4(a) and 6.5(a) . As expected, the
optimal cavity shape has a circle in the x
3
= 0 plane.
The case of φ
2
with the load case (3) (2:1.5:1) did not converge. It is
conjectured that a more general cavity shape, rather than an ellipsoidal one,
needs to be considered for this diﬃcult problem.
6.2.3 Remarks
The optimal shape of a ﬁllet in an axially loaded bar is obtained quickly within
just one iteration ! The size of the design region reduces to nearly a quarter of its
initial value in this 3D problem, compared to about half in the corresponding
2D problem.
The next example considered is the optimal shape of an ellipsoidal cavity
in a cube loaded by remote tensions. The sought after cavity shape is the
one that makes some measure of the stress uniform on it. It is known that
© 2005 by Taylor & Francis Group, LLC
6.2. NUMERICAL RESULTS 123
in the corresponding 2D problem, for an elliptical cutout in an inﬁnite plate,
the tangential stress on the cutout surface is uniform when the loading and
geometry are compatible. (The precise meaning of compatibility is explained
in the body of this chapter). It is found that such is not the case in the 3D
problem. Also, optimization of this class of problems is often more diﬃcult
than for the previous bar with a ﬁllet example. This is true if either no two
of the remote tractions are equal, and/or if one tries to make the von Mises
stress uniform on the cutout surface. Sometimes (Figure 6.5(b)) one sees initial
oscillations of the objective function as a function of iteration number.
© 2005 by Taylor & Francis Group, LLC
Chapter 7
ERROR ESTIMATION
AND ADAPTIVITY
The subject of this chapter is error analysis and adaptivity with the BCM. The
idea of using hypersingular residuals, to obtain local error estimates for the
BIE, was ﬁrst proposed by Paulino [122] and Paulino et al. [123]. This idea has
been applied to the collocation BEM (Paulino et al. [123], Menon et al. [96]
and Paulino et al. [127]); and has been discussed in detail in Chapter 2 of this
book. The main idea, applied to the BCM, has appeared in Mukherjee and
Mukherjee [111], and is presented in this chapter.
7.1 Hypersingular Residuals as Local Error Es
timators
The usual BCM equation (4.25) is solved ﬁrst for the boundary variables (trac
tions and displacements) a. Next, this value of a is input into the righthand
side of equation (4.44) in order to obtain the hypersingular residuals v
kn
in
the displacement gradients u
k,n
. Next, the stress residuals are obtained from
Hooke’s law:
s
kn
= λv
mm
δ
kn
+ µ(v
kn
+ v
nk
) (7.1)
Finally, a scalar measure r of the residual, evaluated at the centroid of a
triangular surface element, is postulated based on the idea of energy. This is:
r = s
kn
v
kn
(7.2)
It has been proved in [96] and [127] for the BIE that, under certain favorable
conditions, real positive constants c
1
and c
2
exist such that:
c
1
r ≤ ≤ c
2
r (7.3)
125
© 2005 by Taylor & Francis Group, LLC
126 CHAPTER 7. ERROR ESTIMATION AND ADAPTIVITY
Generate mesh
Solve BCM
equations
Obtain error
estimators
Need to
subdivide
elements?
STOP
no
yes
Calculate HBCM
residuals
Figure 7.1: Flow chart for adaptive meshing (from [111])
where r is some scalar measure of a hypersingular residual and is a scalar
measure of the exact local error. Thus, a hypersingular residual is expected to
provide a good estimate of the local error on a boundary element. It should be
mentioned here that the deﬁnitions of the residuals used in [96] and [127] are
analogous to, but diﬀerent in detail from, the ones proposed in this chapter.
In the rest of this chapter, e = r, where r, deﬁned in equation (7.2) (and
evaluated at an element centroid), is the hypersingular residual, and e is the
local element error estimator that is used to drive an hadaptive procedure with
the BCM.
7.2 Adaptive Meshing Strategy
The ﬂow chart for adaptive meshing is shown in Figure 7.1.
The remeshing strategy is based on the values of the error estimator e at
each element centroid. This strategy is shown in Figure 7.2 in which ¯ e is the
average value of the error estimator e over all the boundary elements.
A possible criterion for stopping cell reﬁnement can be:
¯ e ≤ e
global
(7.4)
where e
global
has a preset value that depends on the level of overall desired
accuracy.
© 2005 by Taylor & Francis Group, LLC
7.3. NUMERICAL RESULTS 127
One element is
split into
1 < γ < 2 2 < γ < 3 3 < γ
2 elements 3 elements 4 elements
γ = e / e
_
_ _ _
Figure 7.2: Remeshing strategy (from [111])
7.3 Numerical Results
7.3.1 Example One  Short Clamped Cylinder under Ten
sion
This ﬁrst example is concerned with a short cylinder which is clamped at the
bottom and subjected to unit tensile traction on the top surface (Figure 7.3(a)).
The radius and length of the cylinder are each 2 units, the shear modulus of the
cylinder material is 1.0 and the Poisson’s ratio is 0.3 (in consistent units). The
initial mesh on the top (loaded) and bottom (clamped) faces of the cylinder are
identical and are shown in Figure 7.3 (b) while the initial uniform mesh on its
curved surface is shown in Figure 7.4 (b).
It is known ([37], [137]) that, for this problem, the normal stress component
σ
33
varies slowly over much of the clamped face, but exhibits sharp gradients
near its boundary. This stress component becomes singular on the boundary
of the clamped face. The behavior of the shearing stress component σ
zr
(here
r, θ, z ≡ 3 are the usual polar coordinates) on the clamped face is qualitatively
similar to that of σ
33
. The stresses are uniform on the loaded face.
It is seen from Figures 7.3 and 7.4 that this behavior is captured well by
the adaptive scheme. Element error estimators are obtained from equations
(4.44), (7.1) and (7.2) after ﬁrst averaging the traction results from (4.25) within
each element and then using these averaged traction values. Figure 7.3 (b)
shows that these element error estimators (denoted by vertical bars at element
centroids) are largest on the elements near the boundary of the clamped face.
As a consequence (Figures 7.3 (b), (c), (d)) the region near the boundary of the
clamped face is reﬁned most while the mesh on the loaded face of the cylinder
is left unaltered. Also, Figures 7.4 (c), (d) show that some mesh reﬁnement
© 2005 by Taylor & Francis Group, LLC
128 CHAPTER 7. ERROR ESTIMATION AND ADAPTIVITY
4
2
p = 1
(a)
2
1
0
1
2
2
1
0
1
2
0
0.5
1
1.5
2
2.5
X
2
X
3
X
1
(b)
2
1.5
1
0.5
0
0.5
1
1.5
2
2
1
0
1
2
0
0.5
1
1.5
2
X
2
X
3
X
1
(c)
2
1.5
1
0.5
0
0.5
1
1.5
2
2
1
0
1
2
0
0.5
1
1.5
2
X
2
X
3
X
1
(d)
Figure 7.3: Adaptive meshing of the top and bottom faces of a clamped cylinder
under tension: (a) geometry and loading (b) initial mesh with element error
estimators (c) mesh at the end of the ﬁrst adaptive step (d) mesh at the end
of the second adaptive step (from [111])
© 2005 by Taylor & Francis Group, LLC
7.3. NUMERICAL RESULTS 129
4
2
p = 1
(a)
2
1.5
1
0.5
0
0.5
1
1.5
2
2
1
0
1
2
0
0.5
1
1.5
2
X
2
X
3
X
1
(b)
2
1.5
1
0.5
0
0.5
1
1.5
2
2
1
0
1
2
0
0.5
1
1.5
2
X
2
X
3
X
1
(c)
2
1.5
1
0.5
0
0.5
1
1.5
2
2
1
0
1
2
0
0.5
1
1.5
2
X
2
X
3
X
1
(d)
Figure 7.4: Adaptive meshing of the curved surface of a clamped cylinder under
tension: (a) geometry and loading (b) initial uniform mesh (c) mesh at the
end of the ﬁrst adaptive step (d) mesh at the end of the second adaptive step
(from [111])
© 2005 by Taylor & Francis Group, LLC
130 CHAPTER 7. ERROR ESTIMATION AND ADAPTIVITY
mesh # of elements # of nodes ¯ e
initial 144 290 0.0086799
after 1st adaptive step 192 386 0.0048994
after 2nd adaptive step 246 494 0.0042723
Table 7.1: Mesh statistics and ¯ e for the clamped cylinder under tension (from
[111])
takes place on the bottom layer of the curved surface of the cylinder, which is
nearest to the clamped face, while the rest of the mesh on it remains unaltered.
Finally, the mesh statistics, together with ¯ e, the average value of the error
estimator e over the entire surface of the cylinder, appear in Table 7.1. As
expected, ¯ e is seen to decrease with mesh reﬁnement.
7.3.2 Example Two  the Lam´e Problem for a Hollow
Cylinder
This example is concerned with a thick hollow cylinder, in plane strain, sub
jected to external radial tensile loading. The inner and outer radii of the hollow
cylinder are 1 and 3 units, respectively. The shear modulus of the material is
1.0, the Poisson’s ratio is 0.3 and the external radial traction is 3 (in consistent
units). A quarter of the cylinder is modeled and the initial mesh on the quarter
cylinder is shown in Figure 7.5. The bars in Figure 7.5 are the error estimators
evaluated at the centroids of the boundary elements. As expected ([124], also,
please see the discussion in the following paragraph), the error estimators are
largest on the surface of the hole and on the elements on the upper and lower
surfaces (EBAF and HCDG) of the cylinder that lie near the hole. (The visible
elements are shown in Figure 7.5 and the hidden ones are not).
The next (and ﬁnal) mesh, obtained from the adaptive strategy outlined in
Section 7.2 above, is shown in Figure 7.6. It is seen that mesh reﬁnement is
carried out vigorously on the upper and lower surfaces EBAF and HCDG of
the cylinder (the hidden elements are not shown in Figure 7.6), as well as on
the surface FADG of the hole, while the symmetry planes ABCD and EFGH,
on which the stresses are independent of the x
3
coordinate, are only slightly
reﬁned in order to maintain mesh compatibility. Of course, reﬁnement of the
surfaces EBAF and HCDG is expected in view of the presence of radial stress
gradients on these surfaces. The situation on the curved surface FADG is
particularly interesting. In this axisymmetric problem, the tangential gradients
of the stress ﬁelds in the θ direction are, of course, always zero. It is important
to note, however, that the radial stress gradients are large at points on the hole
surface, and this fact leads to large error estimators and signiﬁcant reﬁnement
of the boundary elements on the surface FADG. The corresponding 2D case is
discussed, in some detail, in Paulino et al. [124].
It is important to check the behavior of the actual errors, when the exact
© 2005 by Taylor & Francis Group, LLC
7.3. NUMERICAL RESULTS 131
0
0.5
1
1.5
2
2.5
3
0
0.5
1
1.5
2
2.5
3
0
0.2
0.4
0.6
0.8
1
1.2
A
X
X
X
1
2
3
B
C
D
E
F
G
H
Figure 7.5: Lam´e problem  initial mesh on quarter cylinder together with
element error estimators (from [111])
0
0.5
1
1.5
2
2.5
3
0
0.5
1
1.5
2
2.5
3
0
0.2
0.4
0.6
0.8
1
1
2
3
X
X
X
A
B
C
D
E
F
G
H
Figure 7.6: Lam´e problem  ﬁnal mesh on quarter cylinder (from [111])
© 2005 by Taylor & Francis Group, LLC
132 CHAPTER 7. ERROR ESTIMATION AND ADAPTIVITY
solution is available, in adaptive meshing problems such as this example. The
tangential stress σ
θθ
, as a function of the radial distance r from the center
of the cylinder, is shown in Figure 7.7. The solid line in Figure 7.7 is the
exact solution (from, e.g. [167]) while the numerical results, from the initial
and the ﬁnal mesh, are designated by open circles and plus signs, respectively.
The numerical results for the tangential stress are obtained from the calculated
tractions at the traction nodes I
i
(see Figure 4.1) on the boundary elements on
the symmetry face ABCD in Figures 7.5 and 7.6. The inaccurate results from
the initial coarse mesh is a consequence of the chosen mesh, not the method
itself. This can be seen, for example, by observing the BCM results for the Lam´e
problem for a hollow sphere under internal pressure in Figure 4.3 (open circles),
obtained from a reasonably ﬁne mesh; as well as by examining other numerical
results from the BCM, in, for example, Chapter 4 (see, also, Mukherjee et al.
[109]).
1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3
3
3.5
4
4.5
5
5.5
6
6.5
7
σ
θ
θ
r
Figure 7.7: Lam´e problem for a hollow cylinder. Tangential stress σ
θθ
as a
function of radial distance r. Exact solution: —, BCM solution from initial
mesh: ◦ ◦ ◦◦, BCM solution from ﬁnal mesh: ++++ (from [111])
The L
2
error in a numerical solution in Figure 7.7 is deﬁned as:
=
100
¯ σ
θθ
n
i=1
(
i
)
2
n
(7.5)
where the pointwise error
i
= (σ
θθ
)
(i)
numerical
− (σ
θθ
)
(i)
exact
at node i, n is the
number of nodes and ¯ σ
θθ
is the average value of the exact solution for σ
θθ
(here
4.5). The resulting values of the L
2
errors are 9.83% and 3.83% for the initial
and ﬁnal mesh, respectively. The adaptive meshing procedure is seen to reduce
the error signiﬁcantly in one step.
© 2005 by Taylor & Francis Group, LLC
Part III
THE BOUNDARY NODE
METHOD
133
© 2005 by Taylor & Francis Group, LLC
Chapter 8
SURFACE
APPROXIMANTS
A moving least squares (MLS) approximation scheme, using curvilinear coor
dinates on the 1D bounding surface of a 2D body, or on the 2D bounding
surface of a 3D body, is suitable for the BNM. The 2D problem, which uses
the curvilinear coordinate s on the boundary of a body, is discussed in detail
in Mukherjee and Mukherjee [107] (potential theory) and in Kothnur et al. [72]
(elasticity) (see, also, [108, 52, 77]). The 3D problem requires the curvilinear
surface coordinate s with components (s
1
, s
2
). (Chati and Mukherjee [26], po
tential theory; Chati et al. [25], elasticity). This procedure is described below.
A brief discussion, of ongoing work on the BNM with Cartesian coordinates
[79, 80, 163, 164], is presented as well.
8.1 Moving Least Squares (MLS) Approximants
It is assumed that, for 3D problems, the bounding surface ∂B of a solid body
is the union of piecewise smooth segments called panels. On each panel, one
deﬁnes surface curvilinear coordinates (s
1
, s
2
). For problems in potential theory,
let u be the unknown potential function and τ ≡ ∂u/∂n (where n is a unit
outward normal to ∂B at a point on it). For 3D linear elasticity, let u denote a
component of the displacement vector u and τ be a component of the traction
vector τ on ∂B. One deﬁnes :
u(s) =
m
¸
i=1
p
i
(s −s
E
)a
i
= p
T
(s −s
E
)a
τ(s) =
m
¸
i=1
p
i
(s −s
E
)b
i
= p
T
(s −s
E
)b (8.1)
135
© 2005 by Taylor & Francis Group, LLC
136 CHAPTER 8. SURFACE APPROXIMANTS
The monomials p
i
(see below) are evaluated in local coordinates (s
1
−s
E
1
, s
2
−
s
E
2
) where (s
E
1
, s
E
2
) are the global coordinates of an evaluation point E. It is
important to state here that a
i
and b
i
are not constants. Their functional
dependencies are determined later. (The name “moving least squares” arises
from the fact that the quantities a
i
and b
i
are not constants). The integer m is
the number of monomials in the basis used for u and τ. Quadratic interpolants,
for example, are of the form:
p
T
(˜ s
1
, ˜ s
2
) = [1, ˜ s
1
, ˜ s
2
, ˜ s
2
1
, ˜ s
2
2
, ˜ s
1
˜ s
2
], m = 6, ˜ s
i
= s
i
−s
E
i
; i = 1, 2 (8.2)
The coeﬃcients a
i
and b
i
are obtained by minimizing the weighted discrete
L
2
norms:
R
u
=
n
¸
I=1
w
I
(d)
p
T
(s
I
−s
E
)a − ˆ u
I
2
R
τ
=
n
¸
I=1
w
I
(d)
p
T
(s
I
−s
E
)b − ˆ τ
I
2
(8.3)
1
2
3
E
4
5
BNM
EBNM
(a)
1
0.5
0
0.5
1
1
0.5
0
0.5
1
0
0.2
0.4
0.6
0.8
1
s
1
s
2
w
0 0.5 1
0
0.5
1
d
w
_
_
_
_
(b)
Figure 8.1: Domain of dependence and range of inﬂuence. (a) The nodes 1,
2 and 3 lie within the domain of dependence of the evaluation point E. The
ranges of inﬂuence of nodes 1, 2, 3, 4 and 5 are shown as gray regions. In the
standard BNM, the range of inﬂuence of a node near an edge, e.g. node 4, is
truncated at the edges of a panel. In the EBNM, the range of inﬂuence can
reach over to neighboring panels and contain edges and/or corners  see, e.g.
node 5 (b) Gaussian weight function deﬁned on the range of inﬂuence of a node
(from [163])
where the summation is carried out over the n boundary nodes for which the
weight function w
I
(d) = 0 (Weight functions are deﬁned in Section 8.3). The
© 2005 by Taylor & Francis Group, LLC
8.1. MOVING LEAST SQUARES (MLS) APPROXIMANTS 137
quantity d = g(s, s
I
) is the length of the geodesic on ∂B between s and s
I
.
These n nodes are said to be within the domain of dependence of a point s
(evaluation point E in Figure 8.1(a)). Also, (s
I
1
− s
E
1
, s
I
2
− s
E
2
) are the local
surface coordinates of the boundary nodes with respect to the evaluation point
s
E
= (s
E
1
, s
E
2
) and ˆ u
I
and ˆ τ
I
are the approximations to the nodal values u
I
and
τ
I
. These equations above can be rewritten in compact form as:
R
u
= [P(s
I
−s
E
)a − ˆ u]
T
W(s, s
I
)[P(s
I
−s
E
)a − ˆ u] (8.4)
R
τ
= [P(s
I
−s
E
)b − ˆ τ]
T
W(s, s
I
)[P(s
I
−s
E
)b − ˆ τ] (8.5)
where ˆ u
T
= (ˆ u
1
, ˆ u
2
, · · · , ˆ u
n
), ˆ τ
T
= (ˆ τ
1
, ˆ τ
2
, · · · , ˆ τ
n
), P(s
I
) is an n×m matrix
whose k
th
row is:
[1, p
2
(s
(k)
1
, s
(k)
2
), ...., p
m
(s
(k)
1
, s
(k)
2
)]
and W(s, s
I
) is an n ×n diagonal matrix with w
kk
= w
k
(d) (no sum over k).
The stationarity of R
u
and R
τ
, with respect to a and b, respectively, leads
to the equations:
a(s) = A
−1
(s)B(s)ˆ u , b(s) = A
−1
(s)B(s)ˆ τ (8.6)
where
A(s) = P
T
(s
I
−s
E
)W(s, s
I
)P(s
I
−s
E
)
B(s) = P
T
(s
I
−s
E
)W(s, s
I
) (8.7)
It is noted from above that the coeﬃcients a
i
and b
i
turn out to be functions
of s. Substituting equations (8.6) into equations (8.1), leads to:
u(s) =
n
¸
I=1
Φ
I
(s)ˆ u
I
, τ(s) =
n
¸
I=1
Φ
I
(s)ˆ τ
I
(8.8)
where the approximating functions Φ
I
are:
Φ
I
(s) =
m
¸
j=1
p
j
(s −s
E
)(A
−1
B)
jI
(s) (8.9)
An alternative form of (8.9) is:
Φ(s) = p
T
(s −s
E
)(A
−1
B)(s) (8.10)
where Φ(s) is 1 ×n.
© 2005 by Taylor & Francis Group, LLC
138 CHAPTER 8. SURFACE APPROXIMANTS
As mentioned previously, ˆ u and ˆ τ are approximations to their real values u
and τ. Matrix versions of (8.8) can be written as:
[H]{ˆ u} = {u} , [H]{ˆ τ} = {τ} (8.11)
Equations (8.11) relate the nodal approximations of u and τ to their nodal
values.
Remarks
• Remark 1: Invertibility of A
The matrix A is an m× m matrix, composed of the matrices P and W
(equation 8.7). It needs to be invertible for the construction of the shape
functions. It is also desirable that A be well conditioned. From a well
known fact in linear algebra about ranks of products of matrices, it is
necessary that the rank of matrix P be m. However, if n < m i.e. the
number of nodes n in the domain of dependence of an evaluation point
is less than the order of the polynomial basis m, then matrix A would
be rank deﬁcient and would become noninvertible. So, it is essential to
choose the parameter which controls the range of inﬂuence of a node,
namely
ˆ
d, such that n ≥ m. However, even if the condition n ≥ m is
satisﬁed, but the n nodes in the domain of dependence of the evaluation
point E lie on a straight line on the surface, then the matrix A becomes
singular. Also, it has been observed that choosing n ∼ m may lead to an
unacceptably large condition number of the matrix A.
• Remark 2: Matrix H
As noted above, the matrix Hrelates the actual nodal values to their nodal
approximations. It is observed through numerical experiments that the
matrix H has m eigenvalues equal to unity. The associated m eigenvectors
are described by the monomials used in the bases for constructing the
approximation. Thus, when looking for solutions that cannot be spanned
by the monomials used in the bases, the matrix H plays a signiﬁcant role
in the success of the method.
• Remark 3: Boundary conditions
The H matrix plays a crucial role in the satisfaction of essential boundary
conditions in the EFG method [108] and in the satisfaction of all boundary
conditions in the BNM [107].
• Remark 4: Deﬁnition of a panel
Curvilinear coordinates (s
1
, s
2
) are used to measure distances over curved
surfaces. However, real life objects consist of piecewise smooth surfaces,
referred to as panels in this work, and deﬁning curvilinear coordinates
across edges and corners is a formidable task. In this work, collocation
nodes are placed inside panels, and, in order to circumvent the problem
© 2005 by Taylor & Francis Group, LLC
8.2. SURFACE DERIVATIVES 139
of “reaching over edges,” it was decided that the range of inﬂuence (ROI)
of each node would be truncated at an edge or corner (Figure 8.1(a)). It
will be seen through numerical experiments that restricting the range of
inﬂuence of a node to the panel to which it belongs still yields acceptable
results. An alternative to truncation of ROIs at edges is to use Cartesian
coordinates (see Figure 8.1(a) and Section 8.4).
• Remark 5: The nature of s
1
, s
2
The coordinates (s
1
, s
2
) are the curvilinear coordinates measured along
the bounding surface ∂B. These coordinates are local and not global. In
other words, these are constructed with the origin at the evaluation point
E i.e. these curvilinear coordinates will always be (0,0) at the evaluation
point E. This simpliﬁes the computation of the shape functions to some
extent. Since, (s
1
= 0, s
2
= 0), one has p
1
= 1 and p
i
= 0 for i = 2, · · · , m.
This further implies that the shape function is just the ﬁrst row of the
matrix C.
8.2 Surface Derivatives
Surface derivatives of the potential (or displacement) ﬁeld u are required for
the HBIE. These are computed as follows. With
C = A
−1
B
equations (8.8) and (8.9) give:
u(s) =
n
¸
I=1
m
¸
j=1
p
j
(s −s
E
)C
jI
(s)ˆ u
I
(8.12)
and the tangential derivatives of u can be written as:
∂u(s)
∂s
k
=
n
¸
I=1
m
¸
j=1
¸
∂p
j
∂s
k
(s −s
E
)C
jI
(s) + p
j
(s −s
E
)
∂C
jI
(s)
∂s
k
ˆ u
I
k = 1, 2 (8.13)
The derivatives of the monomials p
j
can be easily computed. These are:
∂p
T
∂s
1
(s
1
−s
E
1
, s
2
−s
E
2
) = [0, 1, 0, 2(s
1
−s
E
1
), 0, (s
2
−s
E
2
)] (8.14)
∂p
T
∂s
2
(s
1
−s
E
1
, s
2
−s
E
2
) = [0, 0, 1, 0, 2(s
2
−s
E
2
), (s
1
−s
E
1
)] (8.15)
After some simple algebra (Chati [23]), the derivatives of the matrix C with
respect to s
k
take the form:
© 2005 by Taylor & Francis Group, LLC
140 CHAPTER 8. SURFACE APPROXIMANTS
∂C(s)
∂s
k
= −A
−1
(s)
∂B(s)
∂s
k
P(s
I
−s
E
)A
−1
(s)B(s) +A
−1
(s)
∂B(s)
∂s
k
k = 1, 2 (8.16)
with
∂B(s)
∂s
k
= P
T
(s
I
−s
E
)
∂W(s, s
I
)
∂s
k
(8.17)
In deriving equation (8.16), the following identity has been used:
∂A
−1
(s)
∂s
k
= −A
−1
(s)
∂A(s)
∂s
k
A
−1
(s) , k = 1, 2 (8.18)
Tangential derivatives of the weight functions (described in Section 8.3) are
easily computed (Chati [23]). The ﬁnal form of the tangential derivatives of the
potential (or displacement) u, at an evaluation point E, takes the form:
∂u
∂s
k
(s
E
) =
n
¸
I=1
m
¸
j=1
¸
∂p
j
∂s
k
(0, 0)C
jI
(s
E
)
ˆ u
I
+
n
¸
I=1
m
¸
j=1
¸
p
j
(0, 0)
¸
A
−1
(s
E
)
∂B
∂s
k
(s
E
)
I −P(s
I
−s
E
)A
−1
(s
E
)B(s
E
)
jI
¸
ˆ u
I
(8.19)
with k = 1, 2. In the above equation, I is the identity matrix.
One also needs the spatial gradient of the function u in order to solve the
HBIE. For problems in potential theory, this is easily obtained from its tangen
tial and normal derivatives, i.e. ∂u/∂s
k
and ∂u/∂n (see (1.14)). For elasticity
problems, however, one must also use Hooke’s law at a point on the surface ∂B.
Details of this procedure are given in Chati et al. [27] and in Chapter 1 of this
book.
Equation (8.19) can be rewritten in compact form as:
∂u
∂s
k
(s
E
) =
n
¸
I=1
Ψ
(k)
I
(s
E
)ˆ u
I
; k = 1, 2 (8.20)
where the approximating functions Ψ
(k)
I
are:
Ψ
(k)
I
(s
E
) =
m
¸
j=1
¸
∂p
j
∂s
k
(0, 0)C
jI
(s
E
)
© 2005 by Taylor & Francis Group, LLC
8.3. WEIGHT FUNCTIONS 141
+
m
¸
j=1
¸
p
j
(0, 0)
¸
A
−1
(s
E
)
∂B
∂s
k
(s
E
)
I −P(s
I
−s
E
)A
−1
(s
E
)B(s
E
)
jI
¸
(8.21)
8.3 Weight Functions
The basic idea behind the choice of a weight function is that its value should
decrease with distance from a node and that it should have compact support so
that the region of inﬂuence of a node is of ﬁnite extent (Figure 8.1(b)). Possible
choices of weight functions [26] are:
• Gaussian (referred to as  WFA) :
w
I
(d) =
e
−(d/d
I
)
2
for d ≤ d
I
0 for d > d
I
(8.22)
• Exponential (referred to as  WFB) :
w
I
(d) =
e
−(d/c)
2
−e
(d
I
/c)
2
1−e
(d
I
/c)
2
for d ≤ d
I
0 for d > d
I
(8.23)
• Cubic Spline (referred to as  WFC) :
w
I
(d) =
2/3 −4(
ˆ
d)
2
+ 4(
ˆ
d)
3
for
ˆ
d ≤ 1/2
4/3 −4(
ˆ
d) + 4(
ˆ
d)
2
−(4/3)(
ˆ
d)
3
for 1/2 <
ˆ
d ≤ 1
0 for d > 1
(8.24)
• Quartic Spline (referred to as  WFD) :
w
I
(d) =
1 −6(
ˆ
d)
2
+ 8(
ˆ
d)
3
−3(
ˆ
d)
4
for
ˆ
d ≤ 1
0 for
ˆ
d > 1
(8.25)
where
ˆ
d = d/d
I
and c is a constant.
Here d = g(s, s
I
) is the minimum distance, measured on the surface ∂B,
(i.e. the geodesic) between a point s and the collocation node I. In the research
performed to date, the region of inﬂuence of a node has been truncated at
the edge of a panel (Figure 8.1(a)) so that geodesics, and their derivatives (for
use in equation (8.17)), need only be computed on piecewise smooth surfaces.
Finally, the quantities d
I
determine the extent of the region of inﬂuence (the
© 2005 by Taylor & Francis Group, LLC
142 CHAPTER 8. SURFACE APPROXIMANTS
compact support) of node I. They can be made globally uniform, or can be
adjusted such that approximately the same number of nodes get included in
the region of inﬂuence of any given node I or in the domain of dependence of
a given evaluation point E. Such ideas have been successfully implemented in
Chati and Mukherjee [26] and Chati et al. [25].
8.4 Use of Cartesian Coordinates
One of the drawbacks of using curvilinear surface coordinates as described above
in Section 8.1 is the need to truncate the range of inﬂuence of a node at an
edge or corner (see Remark 4 in Section 8.1). Another is the need to compute
geodesics on general surfaces. The more straightforward approach, namely the
use of Cartesian coordinates, suﬀers from the disadvantage that the matrix A
deﬁned in equation (8.7) becomes singular if all the nodes in the domain of
dependence of an evaluation point lie on a plane (see [117, 79]). Li and Aluru
have suggested, in two recent papers [79, 80], ways to use Cartesian coordinates
in a modiﬁed version of the BNM which they call the boundary cloud method.
(The acronym BCLM is used for the boundary cloud method in this book). Nice
results for 2D problems in potential theory are given in [79, 80]. This idea is
discussed below for 3D problems in potential theory with linear approximants.
Extension to 3D elasticity is relatively straightforward.
8.4.1 Hermite Type Approximation
For the 3D Laplace equation (see [79] for the 2D case), one writes:
u(x) = p
T
(x)a, τ(x) =
∂p
T
∂n
(x)a (8.26)
Collocation is not allowed at a point on an edge or a corner. For a linear
approximation:
p
T
(x) = [1, x
1
, x
2
, x
3
]
∂p
T
∂n
(x) =
¸
0,
∂x
1
∂n
,
∂x
2
∂n
,
∂x
3
∂n
= [0, n
1
, n
2
, n
3
] (8.27)
where n is the unit outward normal at a boundary point.
The coeﬃcients a
i
are obtained by minimizing the weighted discrete L
2
norm:
J =
n
¸
I=1
w
I
(x
t
, x
I
)
p
T
(x
I
)a − ˆ u
I
2
+
n
¸
I=1
w
I
(x
t
, x
I
)
¸
∂p
T
∂n
(x
I
)a − ˆ τ
I
2
(8.28)
In [79], where a ﬁxed least squares approach is adopted, x
t
are the coor
dinates of a ﬁxed point inside a cloud (chosen to be the center of a cloud),
© 2005 by Taylor & Francis Group, LLC
8.4. USE OF CARTESIAN COORDINATES 143
and x
I
, as before, are the coordinates of node I. This time, the Euclidean dis
tance between x
t
and x
I
is used in the weight function. The weight functions
are piecewise constant, i.e. they are constant within each cloud but vary from
cloud to cloud.
The stationarity of J with respect to a leads to the equations:
u(x) =
n
¸
I=1
M
I
(x)ˆ u
I
+
n
¸
I=1
N
I
(x)ˆ τ
I
(8.29)
τ(x) =
n
¸
I=1
S
I
(x)ˆ u
I
+
n
¸
I=1
T
I
(x)ˆ τ
I
(8.30)
where:
M(x) = p
T
(x)(A
−1
B), N = p
T
(x)(A
−1
D) (8.31)
S(x) =
∂p
T
∂n
(x)(A
−1
B), T(x) =
∂p
T
∂n
(x)(A
−1
D) (8.32)
with:
A(x) = P
T
(x
I
)W(x
t
, x
I
)P(x
I
) +
∂P
T
∂n
(x
I
)W(x
t
, x
I
)
∂P
∂n
(x
I
) (8.33)
B(x) = P
T
(x
I
)W(x
t
, x
I
) (8.34)
D(x) =
∂P
T
∂n
(x
I
)W(x
t
, x
I
) (8.35)
It is proved in [79] that the matrix A in (8.33) is nonsingular.
8.4.2 Variable Basis Approximation
Li and Aluru [80] present a variable basis approach for solving the 2D Laplace’s
equation with the boundary cloud method (BCLM). This is an elegant approach
in which reduced bases are appropriately employed in order to avoid singularity
of the matrix A = P
T
WP. A disadvantage of this approach, as well as that
of the standard BNM, is that discontinuities in the normal derivative of the
potential function, across edges and corners, are not addressed properly  nor
mal derivatives are modeled with continuous approximants, even across edges
and corners. Very recently, Telukunta and Mukherjee [163, 164] have combined
© 2005 by Taylor & Francis Group, LLC
144 CHAPTER 8. SURFACE APPROXIMANTS
the advantages of the variable basis approach [80], together with allowing dis
continuities in τ =
∂u
∂n
, in a new approach called the extended boundary node
method (EBNM). The EBNM, for 2D and 3D potential theory, is described
next.
It is important to mention here that the ﬁxed least squares approach has
been adopted by Li and Aluru in [79, 80] while the variable basis approach
adopted in [163, 164] uses moving least squares for both the BCLM and the
EBNM. The title “BCM” used in ﬁgures depicting numerical results in [163]
refer to a variable basis BCLM with moving least squares. These results have
been obtained from a fresh implementation of the equations presented in [80],
adapted to the moving least squares formulation. The quantity d in the ar
gument of a weight function w
I
in Section 8.3 is, in general, still the geodesic
between x and x
I
in the EBNM. As mentioned before, d is taken as the Eu
clidean distance between x
t
(a ﬁxed point inside a cloud) and x
I
in the work
of Li and Aluru [79, 80] that adopts a ﬁxed least squares approach.
The ﬁrst step is to distinguish between singular and nonsingular clouds
(DODs and ROIs are sometimes called clouds in this work  the term is taken
from the work of Li and Aluru [79, 80]). A straight cloud (for 2D problems)
is one in which the nodes lie on a straight line. Similarly, a ﬂat cloud (for
3D problems) is one in which the nodes lie on a plane. A curved cloud is a
smooth curve in 2D and a smooth surface in 3D problems. Finally, a broken
cloud contains at least one corner in 2D and at least one edge or corner in 3D
problems. The approximants for u are identical in the variable basis BCLM [80]
and in the EBNM [163]. They are, however, diﬀerent for τ  the approximants
for τ are continuous (same as those for u) in [80] but allow for jumps in τ
across corners and edges in the EBNM [163]. The speciﬁc approximants for the
EBNM are given below.
The starting point is to write the approximations (8.1) in terms of Cartesian
coordinates x:
u(x) = p
T
(x)a , τ(x) = q
T
(x)b (8.36)
A basis (i.e. the functions in p or q in (8.36)) for a cloud must satisfy two
competing requirements  it must be broad enough to include all cases, yet it
must be narrow enough such that the matrices A = P
T
WP and C = Q
T
WQ
(see (8.46  8.47)) are nonsingular.
It is noted here that the smoothness of the ﬁnal approximating functions
for u and τ (see equation (8.44)), at a regular point on the bounding surface of
a body, depends on the choice of the weight function w
I
[10].
8.4.2.1 Twodimensional problems
Bases for u and τ. The following bases are used for the various cases listed
below.
Straight cloud: x
1
= c
1
Basis [1, x
2
] for u and τ
© 2005 by Taylor & Francis Group, LLC
8.4. USE OF CARTESIAN COORDINATES 145
All other straight clouds: Basis [1, x
1
] for u and τ
All curved clouds: Basis [1, x
1
, x
2
] for u and τ
Broken cloud: Basis
[1, x
1
, x
2
] for u
[n
1
, n
1
x
2
, n
2
, n
2
x
1
] for τ with both segments
straight
[n
1
, n
1
x
1
, n
1
x
2
, n
2
, n
2
x
1
, n
2
x
2
] for τ with at least
one segment curved
Explanation for choice of basis for u on a broken cloud. It is assumed
that u(x
1
, x
2
) ∈ C
∞
in B. Let a corner C on ∂B have coordinates (x
10
, x
20
)
which, for simplicity, is written as (x
0
, y
0
). A Taylor series expansion for u
about (x
0
, y
0
) is of the form:
u(x, y) = u(x
0
, y
0
) + u
x
(x
0
, y
0
)(x −x
0
) + u
y
(x
0
, y
0
)(y −y
0
) + h.o.t. (8.37)
The linear approximation of u about C is of the form:
l(x, y) = a
0
+ a
1
x + a
2
y (8.38)
which justiﬁes the chosen basis.
Explanation for choice of basis for τ on a broken cloud. A general
curved cloud segment has the equation f(x
1
, x
2
) = 0. Assuming ∇u ∈ C
∞
in
B, one has:
τ = u
,1
(x
1
, x
2
)n
1
+ u
,2
(x
1
, x
2
)n
2
(8.39)
One can, therefore, use the basis n
1
[1, x
1
, x
2
] ∪ n
2
[1, x
1
, x
2
] on a general
curved cloud segment.
Special cases of straight segments of a broken cloud are as follows.
Straight line: n
1
x
1
= c
1
Basis n
1
[1, x
2
]
Straight line: n
2
x
2
= c
2
Basis n
2
[1, x
1
]
Straight line: n
1
x
1
+n
2
x
2
= c
3
Basis n
1
[1, x
1
or x
2
] ∪ n
2
[1, x
1
or x
2
].
It is clear that the recommended reduced basis for τ covers all these cases.
Invertibility of C = Q
T
WQ for τ on a broken cloud.
© 2005 by Taylor & Francis Group, LLC
146 CHAPTER 8. SURFACE APPROXIMANTS
2
4
1
3
L
L
1
2
= y axis
= x axis
O
(a)
1
2
3
4
1
o
L = y axis
L
2
(b)
Figure 8.2: Boundary segments of 2D regions with corners (from [163])
Straight segments with reduced basis. Let (x
1
, x
2
) → (x, y) and [n
1
, n
2
] →
[p, q]. Referring to Figure 8.2(a) (the worst case scenario), it is easy to show
that:
x
1
= x
2
= 0, y
3
= y
4
= 0
q
1
= q
2
= 0, p
1
= p
2
= −1
p
3
= p
4
= 0, q
3
= q
4
= −1 (8.40)
With the reduced basis [p, py, q, qx], one has:
[C] =
¸
L
1
w
i
¸
L
1
w
i
y
i
0 0
¸
L
1
w
i
y
i
¸
L
1
w
i
y
2
i
0 0
0 0
¸
L
2
w
i
¸
L
2
w
i
x
i
0 0
¸
L
2
w
i
x
i
¸
L
2
w
i
x
2
i
¸
¸
¸
¸
¸
¸
¸
(8.41)
This matrix is nonsingular provided that each cloud segment contains at
least two nodes. Clouds must be chosen to fulﬁll this requirement.
Straight and curved segments with full basis. Figure 8.2(b) shows a
broken cloud with one straight and one curved segment. (Without loss of gen
erality, L
1
is chosen to be part of the y axis). For this case, using the full basis
[p, px, py, q, qx, qy], one gets the matrix [C] that is displayed on the next page.
This matrix is nonsingular if the segment L
2
is curved. If, however, L
2
is
straight, one has q
i
= βp
i
for all nodes on L
2
, with β a constant. In this case,
row 5 = β ×row 2 and the matrix becomes singular!
© 2005 by Taylor & Francis Group, LLC
8
.
4
.
U
S
E
O
F
C
A
R
T
E
S
I
A
N
C
O
O
R
D
I
N
A
T
E
S
1
4
7
Σ Σ Σ
Σ
Σ
Σ
Σ
Σ
Σ
Σ Σ
Σ
Σ
Σ
Σ
Σ
Σ
Σ Σ
Σ
Σ
Σ
Σ
Σ
L
1
i
+
L
2
p
2
i
w
i
L
2
p
2
i
w
i
x
i
L
1
w
i
y
i
+
L
2
p
2
i
w
i
y
i
L
2
p
i
q
i
w
i
L
2
p
i
q
i
w
i
x
i
L
2
p
i
q
i
w
i
y
i
L
2
p
2
i
w
i
x
2
i L
2
p
2
i
w
i
x
i
y
i
L
2
p
i
q
i
w
i
x
i
L
2
p
i
q
i
w
i
x
2
i L
2
p
i
q
i
w
i
x
i
y
i
L
1
w
i
y
2
i
+
L
2
p
2
w
i
y
2
i L
2
p
i
q
i
w
i
y
i
L
2
p
i
q
i
w
i
x
i
y
i
L
2
p
i
q
i
w
i
y
2
i
symmetric
L
2
q
2
i
w
i
L
2
q
2
i
w
i
x
i
L
2
q
2
i
w
i
y
i
L
2
q
2
i
w
i
x
2
i L
2
q
2
i
w
i
x
i
y
i
L
2
q
2
i
w
i
y
2
i
The matrix [C]
i
w
© 2005 by Taylor & Francis Group, LLC
148 CHAPTER 8. SURFACE APPROXIMANTS
8.4.2.2 Threedimensional problems
First, a word of caution. In order for the matrices A and C (see (8.46  8.47))
to be invertible, a ﬂat cloud (or a ﬂat segment of a broken cloud) must always
contain three or more points, and all the points on it must not lie on a straight
line [164].
Bases for u and τ. The following bases are used for the various cases listed
below.
Flat cloud:
x
1
= c
1
or n
1
x
1
+ n
2
x
2
= c
4
Basis [1, x
2
, x
3
] for u and τ
x
2
= c
2
or n
2
x
2
+ n
3
x
3
= c
5
Basis [1, x
3
, x
1
] for u and τ
All other ﬂat clouds: Basis [1, x
1
, x
2
] for u and τ
All curved clouds: Basis [1, x
1
, x
2
, x
3
] for u and τ
Broken cloud: Basis
[1, x
1
, x
2
, x
3
] for u
see below for τ
Explanation for choice of basis for u on a broken cloud. The arguments
given for the 2D case can be easily extended to the 3D case as well.
Choice of basis for τ on a broken cloud. A general curved cloud segment
has the equation f(x
1
, x
2
, x
3
) = 0. Assuming ∇u ∈ C
∞
in B, one has:
τ = u
,1
(x
1
, x
2
, x
3
)n
1
+ u
,2
(x
1
, x
2
, x
3
)n
2
+ u
,3
(x
1
, x
2
, x
3
)n
3
(8.42)
One can, therefore, use the basis n
1
[1, x
1
, x
2
, x
3
] ∪ n
2
[1, x
1
, x
2
, x
3
] ∪
n
3
[1, x
1
, x
2
, x
3
] on a general curved cloud segment.
Special cases of ﬂat segments in a broken cloud
Plane P
1
: n
1
x
1
= c
1
Basis n
1
[1, x
2
, x
3
]
Plane P
2
: n
2
x
2
= c
2
Basis n
2
[1, x
3
, x
1
]
Plane P
3
: n
3
x
3
= c
3
Basis n
3
[1, x
1
, x
2
]
Plane P
4
: n
1
x
1
+n
2
x
2
= c
4
Basis n
1
[1, x
1
or x
2
, x
3
] ∪ n
2
[1, x
1
or x
2
, x
3
]
Plane P
5
: n
2
x
2
+n
3
x
3
= c
5
Basis n
2
[1, x
1
, x
2
or x
3
] ∪ n
3
[1, x
1
, x
2
or x
3
]
Plane P
6
: n
3
x
3
+n
1
x
1
= c
6
Basis n
3
[1, x
1
or x
3
, x
2
] ∪ n
1
[1, x
1
or x
3
, x
2
]
© 2005 by Taylor & Francis Group, LLC
8.4. USE OF CARTESIAN COORDINATES 149
Plane P
7
: n
1
x
1
+n
2
x
2
+n
3
x
3
= c
7
Basis n
1
[1, two of (x
1
, x
2
, x
3
)] ∪
n
2
[1, two of (x
1
, x
2
, x
3
)] ∪ n
3
[1, two of (x
1
, x
2
, x
3
)]
On a broken cloud, one must use a basis that is a union of the bases for its
segments.
For a union of two ﬂat segments, one has, for example:
P
1
∪ P
2
: Basis [n
1
, n
1
x
2
, n
1
x
3
, n
2
, n
2
x
3
, n
2
x
1
]
P
1
∪ P
4
: Basis [n
1
, n
1
x
2
, n
1
x
3
, n
2
, n
2
x
3
, n
2
x
1
]
P
1
∪ P
5
: Basis [n
1
, n
1
x
2
, n
1
x
3
, n
2
, n
2
x
3
, n
2
x
1
, n
3
, n
3
x
1
, n
3
x
2
]
P
1
∪ P
7
: Basis [n
1
, n
1
x
2
, n
1
x
3
, n
2
, n
2
x
3
, n
2
x
1
, n
3
, n
3
x
1
, n
3
x
2
]
For a union of three ﬂat segments, one has, for example:
P
1
∪ P
2
∪ P
3
: Basis [n
1
, n
1
x
2
, n
1
x
3
, n
2
, n
2
x
3
, n
2
x
1
, n
3
, n
3
x
1
, n
3
x
2
]
Special cases of curved segments in a broken cloud
S
1
: f
1
(x
1
, x
2
) = 0 Basis [n
1
, n
1
x
1
, n
1
x
2
, n
1
x
3
, n
2
, n
2
x
1
, n
2
x
2
, n
2
x
3
]
S
2
: f
2
(x
2
, x
3
) = 0 Basis [n
2
, n
2
x
1
, n
2
x
2
, n
2
x
3
, n
3
, n
3
x
1
, n
3
x
2
, n
3
x
3
]
S
3
: f
3
(x
3
, x
1
) = 0 Basis [n
1
, n
1
x
1
, n
1
x
2
, n
1
x
3
, n
3
, n
3
x
1
, n
3
x
2
, n
3
x
3
]
S
4
: f
4
(x
1
, x
2
, x
3
) = 0
Basis [n
1
, n
1
x
1
, n
1
x
2
, n
1
x
3
, n
2
, n
2
x
1
, n
2
x
2
, n
2
x
3
, n
3
, n
3
x
1
, n
3
x
2
, n
3
x
3
]
In the above, f
k
, k = 1, 2, 3, 4, are nonlinear functions of their arguments.
On a broken cloud, one must use a basis that is a union of the bases for its
segments.
For example:
P
1
∪ S
1
: Basis [n
1
, n
1
x
1
, n
1
x
2
, n
1
x
3
, n
2
, n
2
x
1
, n
2
x
2
, n
2
x
3
]
P
1
∪S
2
: Basis [n
1
, n
1
x
2
, n
1
x
3
, n
2
, n
2
x
1
, n
2
x
2
, n
2
x
3
, n
3
, n
3
x
1
, n
3
x
2
, n
3
x
3
]
Note that the case P
1
∪ S
1
is a 3D version of the 2D example in Figure
8.2(b).
© 2005 by Taylor & Francis Group, LLC
150 CHAPTER 8. SURFACE APPROXIMANTS
Invertibility of C = Q
T
WQ for τ on a broken cloud with ﬂat segments.
The matrix C becomes singular if and only if all the points on any ﬂat cloud
segment are colinear. A proof of this fact is available in [164].
8.4.2.3 Determination of approximating functions for u and τ
Next, analogous to (8.3), let:
S
u
=
n
¸
I=1
w
I
(d)
p
T
(x
I
)a − ˆ u
I
2
, S
τ
=
n
¸
I=1
w
I
(d)
q
T
(x
I
)b − ˆ τ
I
2
(8.43)
Following the same steps as in Section 8.1, one ﬁnally gets:
u(x) =
n
¸
I=1
α
I
(x)ˆ u
I
, τ(x) =
n
¸
I=1
β
I
(x)ˆ τ
I
(8.44)
where:
α(x) = p
T
(x)(A
−1
B)(x), β(x) = q
T
(x)(C
−1
D)(x) (8.45)
with:
A(x) = P
T
(x
I
)W(x, x
I
)P(x
I
), B(x) = P
T
(x
I
)W(x, x
I
) (8.46)
C(x) = Q
T
(x
I
)W(x, x
I
)Q(x
I
), D(x) = Q
T
(x
I
)W(x, x
I
) (8.47)
In the above, P(x
I
) is an n ×m matrix whose k
th
row is:
[p
1
(x
k
), p
2
(x
k
), ...., p
m
(x
k
)]
similarly for Q(x
I
); and W(x, x
I
) is an n×n diagonal matrix with w
kk
= w
k
(d)
(no sum over k).
Numerical results from the EBNM, for 2D and 3D problems in potential
theory [163, 164], are most encouraging.
It is important to mention here that the rest of this part of this book presents
the boundary node method with curvilinear surface coordinates (s
1
, s
2
). Use of
Cartesian coordinates in the 3D BNM is an interesting subject of continuing
research.
© 2005 by Taylor & Francis Group, LLC
Chapter 9
POTENTIAL THEORY
AND ELASTICITY
This chapter presents a procedure for coupling of 3D singular and hypersin
gular integral equations (BIE and HBIE  see Chapter 1), with moving least
squares (MLS) interpolants (see Chapter 8), to obtain the singular and hy
persingular boundary node method (BNM and HBNM) equations. Numerical
results for selected examples follow. Potential theory is presented ﬁrst; then
linear elasticity.
9.1 Potential Theory in Three Dimensions
9.1.1 BNM: Coupling of BIE with MLS Approximants
9.1.1.1 Coupled equations
The bounding surface ∂B of a body B is partitioned into N
c
cells ∂B
k
and
MLS approximations for the functions u and τ (8.8) are used in the usual BIE
for 3D potential theory (1.7). The result is:
0 =
N
c
¸
k=1
∂B
k
¸
G(x, y)
n
y
¸
I=1
Φ
I
(y)ˆ τ
I
−F(x, y)
n
y
¸
I=1
Φ
I
(y)ˆ u
I
−
n
x
¸
I=1
Φ
I
(x)ˆ u
I
¸¸
dS(y)
(9.1)
where Φ
I
(x) and Φ
I
(y) are the contributions from the I
th
node to the colloca
tion point x and ﬁeld point y respectively. Also, n
y
nodes lie in the domain of
dependence of the ﬁeld point y and n
x
nodes lie in the domain of dependence
of the source point x. When x and y belong to the same cell, the cell is treated
as a singular cell and the special techniques described in the next subsection
151
© 2005 by Taylor & Francis Group, LLC
152 CHAPTER 9. POTENTIAL THEORY AND ELASTICITY
ξ
1
ξ
2
.
.
.
.
. .
.
.
. .
.
P
P
s
t
(a) (b)
(c) (d)
x
y
z
P
I
J
.
.
.
θ=π/2
θ
ρ
ρ=1
η
η
1
1
1
1
(e)
2
1
Figure 9.1: Mapping scheme for weakly singular integrals (from [26])
are used to carry out the integrations. Otherwise, regular Gaussian integration
is used.
9.1.1.2 Weakly singular integration scheme
As mentioned above, regular Gaussian integration can be used as long as the
source and ﬁeld points are on diﬀerent cells. However, the kernels become
singular as the source and collocation points approach each other, i.e. as y → x.
The kernel F(x, y) is strongly singular (O(1/r
2
)) while the kernel G(x, y) is
weakly singular (O(1/r)) as r → 0 (here r is the Euclidean distance between x
and y). The strongly singular integral in (9.1) is regularized in the usual way
by multiplying the kernel F(x, y) by the O(r) function u(y) − u(x). Various
methods have been proposed in the literature to handle weakly singular integrals
arising in the BEM. The method suggested by Nagarajan and Mukherjee [114]
has been used to carry out the weakly singular integration in the BNM. The
details follow.
Consider evaluating the integral with G(x, y) in (9.1) over a (in general
curved) surface cell as shown in Figure 9.1(a). This integral can be represented
© 2005 by Taylor & Francis Group, LLC
9.1. POTENTIAL THEORY IN THREE DIMENSIONS 153
as:
I =
∂B
O(1/r) dS
Q
(9.2)
The cell shown contains the source point x, so that the ﬁeld point y could
coincide with the collocation point x. In this example, Quadratic (T6) trian
gles are used to describe the geometry of the bounding surface. The method
described here, however, can be easily extended for various other kinds of ge
ometric interpolations. First, the cell is mapped into the parent space (Figure
9.1(b)) using the wellknown shape functions for T6 triangles. This involves a
Jacobian and the integral I takes the form:
I =
t=1
t=0
s=1−t
s=0
O(1/r)J
1
ds dt (9.3)
Now, in the parent space the triangle is divided into six pieces. Each indi
vidual triangle is mapped into the parametric (η
1
− η
2
) space (Figure 9.1(c))
using the mapping for linear (T3) triangles. The integral I can now be written
as:
I =
6
¸
i=1
η
2
=1
η
2
=0
η
1
=1−η
2
η
1
=0
O(1/r)J
1
J
(i)
2
dη
1
dη
2
(9.4)
where J
(i)
2
is the Jacobian for each triangle. Now, consider the mapping [114]:
η
1
= ρ cos
2
θ , η
2
= ρ sin
2
θ (9.5)
which maps the ﬂat triangle from the η
1
−η
2
coordinate system into a rectangle
in the ρ −θ space (Figure 9.1(d)). The integral I now takes the form:
I =
6
¸
i=1
θ=π/2
θ=0
ρ=1
ρ=0
O(1/r)J
1
J
(i)
2
ρ sinθ dρ dθ (9.6)
As ρ is a measure of the distance between the source point and the ﬁeld
point, the integral I is now regularized. In other words, the ρ in the numerator
cancels the O(1/r) singularity. Now, to evaluate the integral I in equation (9.6),
regular Gaussian integration can be used. The ﬁnal mapping involves the use
of quadratic (Q4) shape functions to map the rectangle from the ρ − θ space
into the standard square in ξ
1
−ξ
2
space (Figure 9.1(e)). The ﬁnal form of the
integral I is:
I =
6
¸
i=1
ξ
2
=1
ξ
2
=−1
ξ
1
=1
ξ
1
=−1
O(1/r)J
1
J
(i)
2
J
3
ρ sinθ dξ
1
dξ
2
(9.7)
where J
3
is the Jacobian of the ﬁnal transformation. Finally, regular Gaussian
integration can be used to evaluate the above integral I.
© 2005 by Taylor & Francis Group, LLC
154 CHAPTER 9. POTENTIAL THEORY AND ELASTICITY
9.1.1.3 Discretized equations and boundary conditions
The discretized assembled form of equation (9.1) becomes:
[K
1
]{ˆ u} + [K
2
]{ˆ τ} = {0}, N
B
equations (9.8)
where {ˆ u} and {ˆ τ} contain the approximations to the nodal values of u and τ
at the N
B
boundary nodes.
Satisfaction of boundary conditions is carried out next. A special procedure
is needed here since, due to the lack of the delta function property of MLS
approximants, (9.1) contains approximations to the nodal values of the primary
variables rather than the variables themselves [107, 108]. It is assumed that in
a general mixed boundary value problem, either u or τ is prescribed at each
boundary node. Let the vector {¯ y} contain the prescribed boundary conditions
and {x} contain the rest. Each of these vectors is of length N
B
. Also let {ˆ y}
and {ˆ x} be their corresponding approximations. Finally, let {ˆ z} = ({ˆ u}∪{ˆ τ}).
Equation (9.8) is now written as:
[M]{ˆ z} = {0}, N
B
equations (9.9)
Referring to equations (8.11), the nodes with prescribed quantities are con
sidered ﬁrst. This gives rise to a system of equations of the form:
[H
1
]{ˆ z} = {¯ y}, N
B
equations (9.10)
Equations (9.9) and (9.10) are now solved together for 2N
B
unknowns ˆ z
k
.
Finally, (as a postprocessing step) consideration of the rest of the boundary
nodes (those without prescribed boundary conditions) results in the equations:
{x} = [H
2
]{ˆ z}, N
B
equations (9.11)
which yield the required boundary values x
k
.
9.1.1.4 Potential and potential gradient at internal points
The potential at a point ξ inside the body B is obtained from equation (1.2)
while the potential gradient is obtained from (1.9); together with (8.8) in each
case. It is well known that direct use of these equations can yield poor results for
the potential and terrible results for the potential gradient at internal points
that are close to the boundary of the body. This phenomenon, sometimes
referred to as the boundary layer eﬀect, is a consequence of the fact that the
kernels G(ξ, y) and F(ξ, y) become nearly singular and nearly hypersingular,
respectively, as ξ → y. A possible remedy for the BEM for elasticity problems
is discussed in Section 1.3, and for the BCM in Section 4.3 of this book (see,
also, [104]). An analogous procedure is used here for problems in potential
theory. The explicit equations for potential theory are available, for example,
in Kane [68] (equations (17.25) and (17.34)). It should be noted here that
the above mentioned equations in [68] were used to regularize the BIE and
© 2005 by Taylor & Francis Group, LLC
9.1. POTENTIAL THEORY IN THREE DIMENSIONS 155
HBIE, respectively, while use of these equations to obtain the primary variable
and its derivative, at internal points close to the boundary of a body, is a new
application of these equations. This approach, for the BNM for linear elasticity,
is discussed later in Section 9.2.1.3 of this chapter.
9.1.2 HBNM: Coupling of HBIE with MLS Approximants
9.1.2.1 Coupled equations
This time, MLS approximations for the functions τ and ∂u/∂s (8.8, 8.20) are
used in the HBIEs for 3D potential theory (1.12, 1.13). The resulting HBNM
equations are:
0 =
N
c
¸
i=1
∂B
i
∂G(x, y)
∂x
m
¸
n
y
¸
I=1
Φ
I
(y)ˆ τ
I
−
n
x
¸
I=1
Φ
I
(x)ˆ τ
I
¸
dS(y)
− u
,k
(x)
∂B
i
∂G(x, y)
∂x
m
n
k
(y) −n
k
(x)
dS(y)
−
∂B
i
∂F(x, y)
∂x
m
¸
n
y
¸
I=1
Φ
I
(y)ˆ u
I
−
n
x
¸
I=1
Φ
I
(x)ˆ u
I
−u
,k
(x)(y
k
−x
k
)
¸
dS(y)
(9.12)
0 =
N
c
¸
i=1
∂B
i
∂G(x, y)
∂n(x)
¸
n
y
¸
I=1
Φ
I
(y)ˆ τ
I
−
n
x
¸
I=1
Φ
I
(x)ˆ τ
I
¸
dS(y)
− u
,k
(x)
∂B
i
∂G(x, y)
∂n(x)
n
k
(y) −n(x)
dS(y)
−
∂B
i
∂F(x, y)
∂n(x)
¸
n
y
¸
I=1
Φ
I
(y)ˆ u
I
−
n
x
¸
I=1
Φ
I
(x)ˆ u
I
−u
,k
(x)(y
k
−x
k
)
¸
dS(y)
(9.13)
respectively, where Φ
I
(x) and Φ
I
(y) are the contributions from the I
th
node
to the collocation point (x) and ﬁeld point (y), respectively, with n
x
and n
y
nodes in their respective domains of dependence.
The discretized form of the potential gradient equation (1.14) at a source
point x is:
∇u(x) = n(x)
n
x
¸
I=1
Φ
I
(x)ˆ τ
I
+t
1
(x)
n
x
¸
I=1
Ψ
(1)
I
(x)ˆ u
I
+t
2
(x)
n
x
¸
I=1
Ψ
(2)
I
(x)ˆ u
I
(9.14)
The gradient of u from equation (9.14), in global coordinates, is used in
equations (9.12) and (9.13).
© 2005 by Taylor & Francis Group, LLC
156 CHAPTER 9. POTENTIAL THEORY AND ELASTICITY
9.1.2.2 Discretized equations
Equation (9.13) is used to solve boundary value problems in potential theory. As
described above in Section 9.1.1.2, nonsingular integrals in (9.13) are evaluated
by Gaussian quadrature while weakly singular integrals are evaluated by the
procedure outlined in Section 9.1.1.2. The discretized version of equation (9.13)
has the same form as (9.8) and boundary conditions are imposed in the manner
described above in Section 9.1.1.3.
9.1.3 Numerical Results for Dirichlet Problems on a Sphere
The BNM [26] and the HBNM [27] are used to solve Dirichlet problems on
a sphere. (Results for example problems on cubes are available in Chati and
Mukherjee [26]). The exact solutions presented below have been used to eval
uate the performance of the various parameters of the BNM and the HBNM.
Dirichlet problems are posed with these solutions imposed (in turn) on the sur
face of a solid sphere, and the normal derivatives of the potential are computed
on the sphere surface. Potential gradients at internal points are also computed
in some cases. The complete sphere is modeled in all cases.
Exact solutions.
• Linear solution
u = x
1
+ x
2
+ x
3
(9.15)
• Quadratic solutionone
u = x
1
x
2
+ x
2
x
3
+ x
3
x
1
(9.16)
• Quadratic solutiontwo
u = −2(x
1
)
2
+ (x
2
)
2
+ (x
3
)
2
(9.17)
• Cubic solution
u = x
3
1
+ x
3
2
+ x
3
3
−3x
2
1
x
2
−3x
2
2
x
3
−3x
2
3
x
1
(9.18)
• Trigonometric solution
u =
2r
2
R
2
cos
2
φ −
2r
2
3R
2
−
1
3
(9.19)
where R is the radius of the sphere, φ is the angle measured from the x
3
axis and r
2
= x
2
1
+ x
2
2
+ x
2
3
.
© 2005 by Taylor & Francis Group, LLC
9.1. POTENTIAL THEORY IN THREE DIMENSIONS 157
Error measures. Several error measures have been employed in order to
assess the accuracy of the numerical solutions. These are given below.
• Local
ε
local
=
φ
(num)
i
−φ
(exact)
i
φ
(exact)
i
×100 % (9.20)
• Global One (nodebased global percentage L
2
error  used for BNM re
sults)
ε
global
=
1
φ
i

max
1
N
B
N
B
¸
i=1
(φ
(num)
i
−φ
(exact)
i
)
2
% (9.21)
• Global Two (integrated L
2
error  used for HBNM results)
(φ) =
A
(φ
(num)
−φ
(exact)
)
2
dA
A
(φ
(exact)
)
2
dA
×100 % (9.22)
In the above, φ is a generic function, φ
i
is its nodal value at node i and
φ
(num)
is its numerical value.
Position of collocation nodes. One node per cell has been used for all the
numerical examples presented in this section. It has been shown (from numeri
cal experiments) in [26, 27] that, as expected, placement of the collocation node
at the centroid of the triangle in the parent space yields excellent results. This
nodal placement has been used in all the examples presented below.
9.1.3.1 Results from the BNM
A variety of problems have been solved on a sphere. The usual curvilinear co
ordinates θ and φ are used. As mentioned above, Dirichlet boundary conditions
corresponding to the exact solution have been imposed on the surface of the
sphere. Numerical results have been obtained using linear (T3) triangles and
quadratic (T6) triangles for interpolating the geometry. To carry out the inte
gration, each of these triangles are mapped into a unit triangle in the parent
space (see Figure 9.2). The results have been obtained for four diﬀerent meshes.
The crude mesh contains 72 and the ﬁne mesh 288 cells (over the surface of the
sphere). Each of the meshes has two versions  using linear (T3) and quadratic
(T6) triangles, respectively. These results are taken from [26].
© 2005 by Taylor & Francis Group, LLC
158 CHAPTER 9. POTENTIAL THEORY AND ELASTICITY
.
.
.
.
.
.
.
. Linear (T3) triangle
Parent space
s
1.0
1/3
c
t
Quadratic (T6) triangle
x
1.0
1/3
x
z
y
y
z
.
.
Figure 9.2: Mapping of linear (T3) and quadratic (T6) triangles onto the parent
space (from [26])
Choice of compact support of weight function. The parameter d
I
(the
compact support of a weight function) needs to be chosen such that a ‘reason
able’ number of nodes lie in the domain of dependence of an evaluation point E.
In Figure 9.3, the global one error (global percentage L
2
error from (9.21)) in
τ has been recorded for increasing d
I
. It is observed that choosing the smallest
possible d
I
yields the lowest L
2
error. For the optimum choice of d
I
, there
are about 12 −14 nodes in the range of inﬂuence of each node, which is about
2m−3m for a quadratic polynomial basis m = 6.
Choice of weight function and polynomial basis. Various weight func
tions, proposed in the literature, are given in equations (8.22 8.25). Tables 9.1,
9.2 and 9.3 present a convergence study that has been carried out for imposed
linear, quadraticone and cubic solutions, respectively (see (9.15  9.18)), for
various weight functions proposed in the literature. It can be seen that a crude
mesh with 72 quadratic (T6) cells already yields acceptable results. Also, the
Gaussian weight function seems to have an edge over the other weight functions
used. The results for the linear and quadraticone solutions have been obtained
using a polynomial basis m = 6, while, for the cubic solution, polynomial bases
m = 6 and m = 10 have been used. It can be seen from Table 9.3 that using
a higher order basis has only a marginal eﬀect on the solution, except for the
case with the ﬁne mesh with T6 triangles. In fact, various researchers feel that
it is best to use lower order polynomial bases in meshfree methods.
© 2005 by Taylor & Francis Group, LLC
9.1. POTENTIAL THEORY IN THREE DIMENSIONS 159
1.5 2 2.5 3 3.5 4 4.5 5
1
1.5
2
2.5
3
3.5
4
4.5
L
2
e
r
r
o
r
i
n
d
I
τ
Figure 9.3: Eﬀect of changing d
I
on the global L
2
error in τ for a prescribed
linear solution (from [26])
Weight Crude Mesh Fine Mesh Crude Mesh Fine Mesh
Functions (T3) (T3) (T6) (T6)
WFA 4.237 2.083 1.067 0.497
WFB 4.237 2.086 1.069 0.539
WFC 5.029 2.104 1.031 0.584
WFD 4.057 2.253 1.246 0.757
Table 9.1: L
2
error in τ for the linear Dirichlet problem (from [26])
Weight Crude Mesh Fine Mesh Crude Mesh Fine Mesh
Functions (T3) (T3) (T6) (T6)
WFA 4.022 1.788 1.696 0.886
WFB 4.022 1.789 1.697 1.011
WFC 5.314 1.785 1.722 1.199
WFD 5.391 1.893 2.103 1.942
Table 9.2: L
2
error in τ for the quadraticone Dirichlet problem (from [26])
© 2005 by Taylor & Francis Group, LLC
160 CHAPTER 9. POTENTIAL THEORY AND ELASTICITY
Weight Crude Mesh (T3) Fine Mesh (T3)
Functions m = 6 m = 10 m = 6 m = 10
WFA 5.641 5.944 2.284 2.319
WFB 5.641 5.944 2.285 2.318
WFC 6.007 5.525 2.298 2.301
WFD 7.896 5.674 2.345 2.318
Weight Crude Mesh (T6) Fine Mesh (T6)
Functions m = 6 m = 10 m = 6 m = 10
WFA 2.018 2.962 0.909 0.485
WFB 2.019 2.963 1.002 0.485
WFC 2.254 3.003 1.120 0.574
WFD 2.456 3.570 1.733 0.796
Table 9.3: L
2
error in τ for the cubic Dirichlet problem (from [26])
Coordinates τ
num
τ
exact
Local Error
(2.0,0.0,0.0) 0.0 0.0 
(1.5,0.866,1.0) 3.650 3.665 0.413
(0.5,0.866,1.732) 2.852 2.799 1.897
(1.0,1.0,1.414) 3.847 3.828 0.475
(1.732,1.0,0.0) 1.689 1.732 2.484
(1.414,1.414,0.0) 1.998 2.000 0.118
(0.707,0.707,1.732) 3.053 2.949 3.513
(1.225,1.225,1.0) 3.912 3.949 0.954
Table 9.4: Local error in τ at boundary points for the quadraticone Dirichlet
problem (from [26])
Results at boundary points other than at nodes. Upon solving a Dirich
let boundary value problem using the BNM, the value of τ is known at the N
B
boundary nodes. However, to compute τ at any boundary location other than
the collocation nodes, the MLS approximants, equation (8.8), need to be used.
Table 9.4 presents the results for τ at a few points on the boundary that are not
the collocation nodes. The results are obtained upon imposing a quadratic1
solution on a crude mesh (72 T6 cells) with the Gaussian weight function. It
can be seen that the accuracy of the numerical solution is well within acceptable
limits.
Results at internal points. Figures 9.4 (a) and (b) show variation in the
potential and its directional derivative at points inside the sphere. The Dirichlet
© 2005 by Taylor & Francis Group, LLC
9.1. POTENTIAL THEORY IN THREE DIMENSIONS 161
boundary value problem is solved upon imposing the cubic solution on a crude
mesh (72 T6 cells) with the Gaussian weight function and a quadratic basis
(m = 6). The gradient is dotted with the diagonal (x
1
= x
2
= x
3
) in order to
get the directional derivative along this line. Values of u and ∇u, at internal
points that are close to the surface of the body, are obtained with the usual
BNM (without modiﬁcation) and by a new application of equations (17.25) and
(17.34) in Kane [68] (with modiﬁcation). It is clear that this modiﬁcation is
essential for success of this method at computing derivatives of the potential
function at internal points that are close to the boundary of a body. This issue,
for the BEM for elasticity problems, is discussed in detail in Section 1.3 of this
book, and for the BNM for elasticity in Section 9.2.1.3 later in this chapter.
0 0.5 1 1.5 2
8
7
6
5
4
3
2
1
0
BNM solution without modification
BNM solution with modification
Exact solution
Radial coordinate
u
(a)
0 0.5 1 1.5 2
14
12
10
8
6
4
2
0
Radial coordinate
∂
u
/
∂
BNM solution without modification
BNM solution with modification
Exact solution
s
(b)
Figure 9.4: Variation of the potential u and tangential derivative of u, along
the line x
1
= x
2
= x
3
, with a prescribed cubic solution (from [26])
Comparison of BEM and BNM. The results obtained by the BNM have
also been compared with those from the conventional BEM for a Dirichlet prob
lem on a sphere with the exact trigonometric solution as given in equation
(9.19). Table 9.5 compares the wallclock times from a serial BEM code with a
serial and a parallel version of the BNM code. The parallel version is an early
one in which only the assembly phase of the BNM matrices, not the solution
phase, has been parallelized. The parallel BNM code is run on 4, 16, and 32
processors using the message passing interface (MPI) standard on the IBM SP2
(R6000 architecture, 120 MHz P2SC Processor). Certainly, the serial BNM is
considerably slower than the serial BEM and this is because the approxima
tion functions in the BNM need to be generated at each point unlike in the
BEM. (Similar observations have been made by other researchers regarding the
performance of the EFG compared to the FEM). One possible remedy is an ac
celerated BNM (Kulkarni et al. [77]). Also, the BNM is very easy to parallelize,
© 2005 by Taylor & Francis Group, LLC
162 CHAPTER 9. POTENTIAL THEORY AND ELASTICITY
Meshes Serial Serial Parallel BNM
BEM BNM 4 Procs 16 Procs 32 Procs
72 T6 cells 3.3 secs 29.5 secs 10.0 secs 2.4 secs 1.3 secs
128 T6 cells 9.3 secs 106.3 secs 35.5 secs 7.6 secs 4.4 secs
288 T6 cells 47.5 secs 690.7 secs 249.0 secs 53.0 secs 27.3 secs
Table 9.5: Comparison of wallclock times for the BNM and BEM for a Dirichlet
problem on a sphere (from [26])
and, as shown in Table 9.5, parallel versions drastically reduce wallclock times.
Figure 9.5 shows a comparison of the L
2
error for the BEM and the BNM as a
function of the number of boundary nodes. The L
2
error in τ is deﬁned here
as:
e(τ) =
¸
N
B
i=1
(τ
(num)
i
−τ
(exact)
i
)
2
¸
N
B
i=1
(τ
(exact)
i
)
2
(9.23)
It can be clearly seen that the two methods yield comparable results and
have almost identical rates of convergence.
10
1
10
2
10
3
10
3
10
2
10
1
10
0
BNM solution
BEM solution
B
N
e
(
τ
)
Figure 9.5: Convergence of the BNM and the BEM with a prescribed trigono
metric solution (from [26])
9.1.3.2 Results from the HBNM
This section follows the layout of Section 9.1.3.1 above. Once again, Dirichlet
problems have been solved with various forms of the potential function (from
© 2005 by Taylor & Francis Group, LLC
9.1. POTENTIAL THEORY IN THREE DIMENSIONS 163
Collocation point : P (x
1P
, x
2P
, x
3P
) ≡ (R, θ
P
, φ
P
)
Field point : Q (x
1Q
, x
2Q
, x
3Q
) ≡ (R, θ
Q
, φ
Q
)
Curvilinear coordinates between P and Q :
˜ s
1
= R(φ
Q
−φ
P
) ; ˜ s
2
= R(θ
Q
−θ
P
)
Exact geodesic Approximate geodesic
Ψ : Angle between P & Q d =
(˜ s
2
1
+ ˜ s
2
2
)
Ψ = arccos(¯r
P
· ¯r
Q
/R
2
)
d = RΨ
Table 9.6: Exact versus approximate geodesics on the surface of a sphere (from
[27])
equations (9.15  9.19)) prescribed on the surface of a sphere. The Gaussian
weight function (8.22) is used in all cases. Also, for all the numerical examples
presented in this section, 72 T6 triangular cells have been used on the surface
of a sphere, with one node per cell placed at the centroid of each triangle in the
parent space (Figure 9.2). Overall, this arrangement yields the best numerical
results. These results are taken from [27].
Geodesics. In order to construct the interpolating functions using MLS ap
proximants, it is necessary to compute the geodesic on the bounding surface of a
body. Computation of geodesics can get quite cumbersome on a general curved
surface described by splines. For the sphere, the exact geodesic between two
points x (collocation node) and y (ﬁeld point) is the length of the arc between
these points on the great circle containing them. However, a very simple ap
proximation to the geodesic would be to use the “Euclidean” distance between
points x and y. Table 9.6 summarizes the procedure for computing the exact
and approximate geodesic on a sphere, while Table 9.7 presents a comparison
in the (global two) L
2
errors for linear, quadratic, cubic and trigonometric solu
tions imposed on the sphere. For this speciﬁc example, with idealized geometry
and boundary conditions, it can be seen that the errors remain reasonably small
even if the approximate geodesic is used to replace the exact one. Thus, it is
expected that the computation of geodesics on complicated shapes will not be
a hindrance towards using the present methodology.
Range of inﬂuence of nodes. Another important feature of the MLS ap
proximants is the range of inﬂuence associated with each node. The parameter
which controls the socalled “compact support” associated with each node is
d
I
. In this work the parameter d
I
is chosen to be nonhomogeneous in the sense
that each evaluation point has an identical number of nodes in its domain of
dependence. Now, for a given polynomial basis (e.g. Linear/Quadratic/Cubic),
© 2005 by Taylor & Francis Group, LLC
164 CHAPTER 9. POTENTIAL THEORY AND ELASTICITY
Exact solution Exact geodesic Approximate geodesic
Linear 0.0717 % 0.109 %
Quadratic 0.313 % 0.715 %
Cubic 1.665 % 4.059 %
Trigonometric 0.172 % 0.279 %
Table 9.7: Global two L
2
error in τ for Dirichlet problems on a sphere with the
exact and an approximate computation of geodesics (from [27])
the number of nodes n in the domain of dependence of each evaluation point
becomes the parameter of interest.
The parameters d
I
are chosen as follows. For a given n, let S be the set
of nodes in the domain of dependence of a particular evaluation point E. The
values of d
I
for all nodes in S are set equal to d
max
, where d
max
is the distance
from E, along the geodesic, of the node in S which is farthest from E.
Figures 9.6 (a) and (b) show the eﬀect of varying the number of nodes n for
a linear basis (m = 3) and a quadratic basis (m = 6), respectively, for the cubic
and trigonometric solutions. It is observed that the lowest value of L
2
errors is
obtained for n ∈ (2m, 3m). This fact has also been observed for the BNM.
4 6 8 10 12 14
0
5
10
15
20
25
Number of nodes (n)
L
2
e
r
r
o
r
i
n
∂
u
/
∂
n
m = 3
Cubic function
Trigonometric function
(a) m=3
10 12 14 16 18
0
1
2
3
4
5
6
Number of nodes (n)
L
2
e
r
r
o
r
i
n
∂
u
/
∂
n
m = 6
Cubic function
Trigonometric function
(b) m=6
Figure 9.6: Global two L
2
error in τ for varying number of points in the domain
of dependence of an evaluation point E: (a) linear polynomial basis (m = 3);
(b) quadratic polynomial basis (m = 6) (72 T6 cells with one node per cell)
(from [27])
Results at internal points. Figures 9.7 (a) and (b) show variation in the
potential and its x
1
derivative, respectively, for points along the x
1
axis inside
the sphere. The Dirichlet boundary value problem is solved upon imposing the
© 2005 by Taylor & Francis Group, LLC
9.2. LINEAR ELASTICITY IN THREE DIMENSIONS 165
0 0.5 1 1.5 2
10
9
8
7
6
5
4
3
x
u
HBNM solution
Exact solution
1
_
_
_
_
_
_
_
_
x
1
0
(a)
0 0.5 1 1.5 2
7
6
5
4
3
2
1
0
x
∂
u
/
∂
x
HBNM solution
Exact solution
1
1
_
_
_
_
_
_
_
x
1
0
(b)
Figure 9.7: Variation of (a) potential and (b) ∂u/∂x
1
along the x
1
axis for a
sphere with the prescribed trigonometric solution (from [27])
trigonometric solution on a cell conﬁguration consisting of 72 T6 cells with one
node per cell and a quadratic basis (m = 6). It is seen from these ﬁgures that
the HBNM solutions match the exact solutions within plotting accuracy for
both u and ∂u/∂x
1
. Once again, the reader is reminded that a technique for
dealing with nearly singular integrals in linear elasticity is described in detail
in Section 1.3 (for the BEM), and in Section 9.2.1.3 (for the BNM) later in this
chapter. An analogous method for potential theory has been employed in order
to get accurate results at internal points that are close to the surface of the
sphere in Figure 9.7.
Comparison of BEM and HBNM. The results obtained by the HBNM
have also been compared with those from the conventional BEM for a Dirichlet
problem on a sphere with the trigonometric solution (equation (9.19)). Figure
9.8 presents a comparison in the L
2
error in τ (as deﬁned in (9.23)) for the
HBNM and BEM, as functions of the (global) number of nodes. This ﬁgure
shows that the two methods yield comparable results and have similar rates of
convergence. The HBNM solution, however, is more accurate than the BEM
solution for this example.
9.2 Linear Elasticity in Three Dimensions
9.2.1 BNM: Coupling of BIE with MLS Approximants
9.2.1.1 Coupled equations
As before for the case of potential theory (Section 9.1.1.1), the bounding surface
∂B of a body B is partitioned into N
c
cells ∂B
k
and MLS approximations for
© 2005 by Taylor & Francis Group, LLC
166 CHAPTER 9. POTENTIAL THEORY AND ELASTICITY
10
1
10
2
10
3
10
5
10
4
10
3
10
2
10
1
BEM solution
HBNM solution
B
N
e
(
τ
)





Figure 9.8: Comparison of e(τ) (L
2
error in ∂u/∂n) for the HBNM and the
conventional BEM for a Dirichlet problem on a sphere (from [27])
the functions u
i
and τ
i
(8.8) are used in the usual BIE for 3D linear elasticity
(1.25). The result is:
0 =
N
c
¸
m=1
∂B
m
¸
U
ik
(x, y)
n
y
¸
I=1
Φ
I
(y)ˆ τ
iI
− T
ik
(x, y)
n
y
¸
I=1
Φ
I
(y)ˆ u
iI
−
n
x
¸
I=1
Φ
I
(x)ˆ u
iI
¸
dS(y) (9.24)
where Φ
I
(x) and Φ
I
(y) are the contributions from the I
th
node to the collo
cation point x and ﬁeld point y respectively. Also, as before, n
y
nodes lie in
the domain of dependence of the ﬁeld point y and n
x
nodes lie in the domain
of dependence of the source point x. When x and y belong to the same cell,
the cell is treated as a singular cell and the special techniques described in Sec
tion 9.1.1.2 are used to carry out the integrations. Otherwise, regular Gaussian
integration is used.
9.2.1.2 Discretized equations and boundary conditions
This discussion closely follows that on potential theory in Section 9.1.1.3. The
discretized form of (9.24) has the same form as (9.8), but this time with 3N
B
equations for the 6N
B
quantities ˆ u
i
, ˆ τ
i
, i = 1, 2, 3, at N
B
boundary nodes. The
other 6N
B
quantities of interest are u
i
, τ
i
, i = 1, 2, 3, at N
B
boundary nodes;
3N
B
of which are prescribed by the boundary conditions. The remaining 6N
B
equations, that relate the nodal approximations of u
i
and τ
i
to their nodal
© 2005 by Taylor & Francis Group, LLC
9.2. LINEAR ELASTICITY IN THREE DIMENSIONS 167
values, have the same form as (8.11). The complete system of equations is
solved in a manner that is very similar to that described in Section 9.1.1.3 for
potential theory.
9.2.1.3 Displacements and stresses at internal points
The displacement at a point ξ ∈ B is obtained from (1.17) or (1.21), the dis
placement gradient from (1.27) or (1.28), and the stress (directly) from (1.30).
In each case, of course, (8.8) must be used.
As mentioned several times before (see Sections 1.3, 4.3), evaluation of dis
placements and stresses at internal points, that are very close to the bounding
surface of a body, requires special care. For this purpose, the BNM versions of
continuous equations such as (1.50) and (1.52) are needed. These are [104]:
u
k
(ξ) = u
k
(ˆ x)
+
N
c
¸
m=1
∂B
m
¸
U
ik
(ξ, y)
n
y
¸
I=1
Φ
I
(y)ˆ τ
iI
−T
ik
(ξ, y)
n
y
¸
I=1
Φ
I
(y)ˆ u
iI
−u
i
(ˆ x)
¸¸
dS(y)
(9.25)
σ
ij
(ξ) = σ
ij
(ˆ x) +
∂B
D
ijk
(ξ, y)
¸
n
y
¸
I=1
Φ
I
(y)ˆ τ
kI
−σ
km
(ˆ x)n
m
(y)
¸
dS(y)
−
∂B
S
ijk
(ξ, y)
¸
n
y
¸
I=1
Φ
I
(y)ˆ u
kI
−u
k
(ˆ x) −u
k,
(ˆ x)(y
− ˆ x
)
¸
dS(y)
(9.26)
Equation (9.26)) requires the displacement gradients and stresses at the
target point ˆ x (in addition to the usual displacements and tractions at ﬁeld
points y). Displacement gradients on the surface of the body are obtained
as part of the BNM solution of the original boundary value problem. This
procedure, adapted from Lutz et al. [89] for the BEM, is described in detail in
Section 1.2.2.1 of this book (see, also, [27]).
9.2.2 HBNM: Coupling of HBIE with MLS Approximants
9.2.2.1 Coupled equations
The starting point here are the HBIEs (1.40) and (1.41). The HBNM equations
are:
0 =
N
c
¸
l=1
∂B
l
D
ijk
(x, y)
¸
n
y
¸
I=1
Φ
I
(y)ˆ τ
kI
−
n
x
¸
I=1
Φ
I
(x)ˆ τ
kI
¸
dS(y)
© 2005 by Taylor & Francis Group, LLC
168 CHAPTER 9. POTENTIAL THEORY AND ELASTICITY
− σ
km
(x)
∂B
l
D
ijk
(x, y)(n
m
(y) −n
m
(x)) dS(y)
−
∂B
l
S
ijk
(x, y)
¸
n
y
¸
I=1
Φ
I
(y)ˆ u
kI
−
n
x
¸
I=1
Φ
I
(x)ˆ u
kI
−u
k,m
(x)(y
m
−x
m
)
¸
dS(y)
(9.27)
and
0 =
N
c
¸
l=1
∂B
l
D
ijk
(x, y)n
j
(x)
¸
n
y
¸
I=1
Φ
I
(y)ˆ τ
kI
−
n
x
¸
I=1
Φ
I
(x)ˆ τ
kI
¸
dS(y)
− σ
km
(x)
∂B
l
D
ijk
(x, y)n
j
(x)(n
m
(y) −n
m
(x)) dS(y)
−
∂B
l
S
ijk
(x, y)n
j
(x)
¸
n
y
¸
I=1
Φ
I
(y)ˆ u
kI
−
n
x
¸
I=1
Φ
I
(x)ˆ u
kI
−u
k,m
(x)(y
m
−x
m
)
¸
dS(y)
(9.28)
Note that, as expected, taking the limit ˆ x ∈ ∂B → x ∈ ∂B in the contin
uous equation (9.26) yields a corresponding regularized HBNM. The resulting
equation is an alternate version of (9.27) (see (1.39) and (1.40)).
9.2.2.2 Discretization
Equations (9.27) and (9.28) are the HBNM equations for linear elasticity.
The procedure followed for discretization of (9.28) is quite analogous to the
BNM case described before in Section 9.2.1.2. These equations are fully regular
ized and only contain nonsingular or weakly singular integrands. Nonsingular
integrals are evaluated using the usual Gauss quadrature over surface cells,
while the weakly singular integrals are evaluated using the procedure outlined
in Section 9.1.1.2). Displacement gradients at a boundary point x are obtained
by applying the procedure outlined in Section 1.2.2.1 and the stress components
at this point are then obtained from Hooke’s law. The discretized version of
equation (9.28) has the generic form shown in (9.8).
9.2.3 Numerical Results
Numerical results from the standard and the new BNM, as well as from the
HBNM, are presented in this section. The standard BNM results are obtained
from (the BNM versions of) equations (1.17) and (1.30) while the new BNM
results are obtained from equations (9.25) and (9.26). Three illustrative exam
ples are considered in this section : a hollow sphere under internal pressure,
a bimaterial sphere and a cube with a spherical cavity under tension (the 3D
Kirsch problem). These numerical results are taken from [25], [104] and [27].
© 2005 by Taylor & Francis Group, LLC
9.2. LINEAR ELASTICITY IN THREE DIMENSIONS 169
1 1.5 2 2.5 3 3.5 4
0
0.1
0.2
0.3
0.4
0.5
0.6
R
a
d
i
a
l
d
i
s
p
l
a
c
e
m
e
n
t
BNM standard method
BNM new method
Exact solution
Radial coordinate
(a)
1 1.5 2 2.5 3 3.5 4
0
0.1
0.2
0.3
0.4
0.5
0.6
HBNM solution
Exact solution
R
a
d
i
a
l
d
i
s
p
l
a
c
e
m
e
n
t
Radial coordinate
(b)
Figure 9.9: Radial displacement as a function of the radial coordinate in a
hollow sphere: (a) from the standard and the new BNM (b) from the HBNM,
together with the exact solution (from [104] and [27])
9.2.3.1 Hollow sphere under internal pressure
The inner and outer radii of the sphere are 1 and 4 units, respectively, the
internal pressure is 1, E = 1 and ν = 0.25. 72 quadratic T6 triangular cells
(with one node at the centroid of each cell) are used on each surface of the
hollow sphere. The entire surface of the sphere is modeled here, with tractions
prescribed over both the inner and outer surfaces of the sphere. The resulting
singular matrices are regularized using the procedure described in [25] (see, also,
[90]).
Internal displacements. The radial displacement, as a function of the radial
coordinate, from the standard as well as the new BNM, are shown in Figure 9.9
(a) while corresponding results from the HBNM appear in Figure 9.9 (b). The
exact solution from [167] is also included in these ﬁgures. It is observed that the
results from the standard BNM and HBNM are excellent as long as an internal
point is not very close to one of the surfaces of the sphere; while those from the
new BNM are very accurate everywhere, including at internal points that are
close to the bounding surfaces of the sphere.
Internal stresses. The radial and circumferential stresses from the new BNM
and from the HBNM, together with the exact solutions from [167], appear in
Figures 9.10 (a) and (b), respectively. The agreement between the exact and
numerical solutions are observed to be excellent.
Finally, the stress solutions from the standard BNM are shown in Figure
9.11, together with details near the inner surface of the sphere in Table 9.8. It
is observed that the stress solutions from the standard BNM are meaningless at
© 2005 by Taylor & Francis Group, LLC
170 CHAPTER 9. POTENTIAL THEORY AND ELASTICITY
1 1.5 2 2.5 3 3.5 4
1
0.5
0
0.5
BNM solution for σ
θ θ
BNM solution for σ
rr
Exact σ
θ θ
Exact σ
rr
Radial coordinate
S
t
r
e
s
s
c
o
m
p
o
n
e
n
t
s
(a)
1 1.5 2 2.5 3 3.5 4
1
0.5
0
0.5
HBNM solution for σ
θ θ
HBNM solution for σ
rr
Exact σ
θ θ
Exact σ
rr
Radial coordinate
S
t
r
e
s
s
c
o
m
p
o
n
e
n
t
s
_
_
(b)
Figure 9.10: Radial and circumferential stresses as functions of the radial coor
dinate in a hollow sphere: (a) from the new BNM (b) from the HBNM, together
with the exact solutions (from [104] and [27])
1 1.5 2 2.5 3 3.5 4
2
1
0
1
2
3
4
5
BNM solution for σ
θ θ
(standard method)
BNM solution for σ
rr
(standard method)
BNM solution for σ
θ θ
(new method)
BNM solution for σ
rr
(new method)
Exact σ
θ θ
Exact σ
rr
Radial coordinate
S
t
r
e
s
s
c
o
m
p
o
n
e
n
t
s
Figure 9.11: Radial and circumferential stresses as functions of the radial coor
dinate in a hollow sphere, from the new and standard BNM, together with exact
solutions (from [104])
© 2005 by Taylor & Francis Group, LLC
9.2. LINEAR ELASTICITY IN THREE DIMENSIONS 171
σ
rr
σ
θθ
r standard new exact standard new exact
1.01 4.99057 0.98020 0.97012 3.68714 0.51168 0.50887
1.03 5.43126 0.92753 0.91379 1.08215 0.49365 0.48071
1.05 3.76287 0.85789 0.86168 0.35304 0.46201 0.45465
1.07 1.39290 0.79589 0.81338 0.36737 0.42834 0.43050
1.09 0.18719 0.74915 0.76857 0.01546 0.40084 0.40809
1.11 0.90056 0.71229 0.72693 0.24717 0.37928 0.38727
Table 9.8: Radial and circumferential stresses as functions of the radial coor
dinate in a hollow sphere, from the new and standard BNM, together with the
exact solutions, at points very close to the inner surface (from [104])
internal points very near the inner surface of the sphere, and an algorithm for
improving these results, such as that presented in Section 9.2.1.3, is absolutely
essential in this case.
9.2.3.2 Bimaterial sphere
The BNM has been extended to solve problems involving material discontinu
ities. Figure 9.12 (a) shows a schematic of two perfectly bonded spheres of
two diﬀerent materials. Numerical results for this model have been obtained by
prescribing displacements on the outer boundary. One could also prescribe trac
tions over the entire outer surface and then appropriately modify the scheme
presented in [25] for solving traction prescribed problems. This is planned for
the future.
Upon prescribing a radial displacement u
0
on the outer surface of the sphere,
the exact solution for the radial displacement and stresses in each material is
given below. For material 1:
u
(1)
r
= A
1
r
1 −2ν
1
E
1
(9.29)
σ
(1)
rr
= σ
(1)
tt
= A
1
(9.30)
where the constant A
1
is deﬁned below. For material 2:
u
(2)
r
=
r
E
2
¸
A
2
(1 −2ν
2
) −
B
2
2r
3
(1 + ν
2
)
(9.31)
σ
(2)
rr
= A
2
+
B
2
r
3
; σ
(2)
tt
= A
2
−
B
2
2r
3
(9.32)
The constants are:
A
1
= A
2
+
B
2
R
3
1
(9.33)
© 2005 by Taylor & Francis Group, LLC
172 CHAPTER 9. POTENTIAL THEORY AND ELASTICITY
x
x
x
E
E
R
R
ν
ν
1
1
1
1
2
2 2
2
3
(a)
0 0.5 1 1.5 2 2.5 3 3.5 4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
BNM solution
Exact solution
Radial coordinate
R
a
d
i
a
l
d
i
s
p
l
a
c
e
m
e
n
t
(b)
Figure 9.12: Bimaterial sphere subjected to uniform external radial displace
ment : (a) conﬁguration (b) radial displacement along the x
1
axis from the new
BNM together with the exact solution (from [25])
α =
E
1
(1 −2ν
2
)
E
2
(1 −2ν
1
)
; β =
E
1
(1 + ν
2
)
2E
2
(1 −2ν
1
)
(9.34)
A
2
=
u
0
E
2
R
2
R
3
1
1 + β
C(1 + α)
; B
2
= −
u
0
E
2
R
2
C
(9.35)
C =
(1 −2ν
2
)(1 + β)
R
3
1
(1 −α)
+
1 + ν
2
2R
3
2
(9.36)
The material and geometric parameters chosen for the two materials are :
• for material 1, E
1
= 1.0, ν
1
= 0.28, R
1
= 1.0
• for material 2, E
2
= 2.0, ν
2
= 0.33, R
2
= 4.0.
A constant radial displacement is prescribed on the outer boundary of mate
rial 2 (u
0
= 1.0). Figure 9.12 (b) shows a comparison of the numerical solution
and the exact solution for the radial displacement within the two materials. Fig
ures 9.13 (a) and (b) show the radial and circumferential stress, respectively,
along the line x
1
axis. It can be seen that the jump in the circumferential stress
at the bimaterial interface is very well captured by the new BNM.
9.2.3.3 3D Kirsch problem
The 3D Kirsch’s problem consists of a cube with a small spherical cavity sub
jected to far ﬁeld uniform tension (Figure 9.14 (a)). The material and geomet
ric parameters are : E = 1.0, ν = 0.25, cutout radius a = 1.0, side of cube
2b = 20.0. Again, the loading is applied without restraining any rigid body
© 2005 by Taylor & Francis Group, LLC
9.2. LINEAR ELASTICITY IN THREE DIMENSIONS 173
0 0.5 1 1.5 2 2.5 3 3.5 4
0.9
1
1.1
1.2
1.3
1.4
1.5
BNM solution
Exact solution
Radial coordinate
R
a
d
i
a
l
s
t
r
e
s
s
(a)
0 0.5 1 1.5 2 2.5 3 3.5 4
0.9
1
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
BNM solution
Exact solution
Radial coordinate
C
i
r
c
u
m
f
e
r
e
n
t
i
a
l
s
t
r
e
s
s
(b)
Figure 9.13: Bimaterial sphere subjected to uniform external radial displace
ment : (a) radial stress σ
rr
(b) circumferential stress σ
θθ
along the x
1
axis from
the new BNM, together with the exact solutions (from [25])
modes and the scheme presented in [25] is used to obtain meaningful numerical
results.
The exact solution for the normal stress σ
33
, for points in the plane x
3
= 0,
is given as [167]:
σ
33
= σ
0
¸
1 +
4 −5ν
2(7 −5ν)
a
r
3
+
9
2(7 −5ν)
a
r
5
(9.37)
Figure 9.14 (b) shows a comparison between the new BNM, HBNM and
the exact solution for the (normalized) normal stress σ
33
/σ
0
along the x
1
axis.
Again, it can be clearly seen that the BNM and HBNM solutions are in excellent
agreement with the analytical solution. The cell structure consists of 96 Q4 cells
modeling the cube and 72 T6 cells modeling the spherical cavity, again with
one node per cell. It should be noted that the algorithm presented in Section
9.2.1.3 is essential for obtaining accurate values of stresses near the surface of
the cavity.
© 2005 by Taylor & Francis Group, LLC
174 CHAPTER 9. POTENTIAL THEORY AND ELASTICITY
X
X
X
1
2
3
(a)
0 1 2 3 4 5
0.8
1
1.2
1.4
1.6
1.8
2
2.2
x
σ
/
σ
0
BNM solution
HBNM solution
Exact solution
3
3
1
(b)
Figure 9.14: 3D Kirsch problem : (a) conﬁguration (b) σ
33
/σ
0
along the x
1
axis from the new BNM and the HBNM, together with the exact solution (from
[25, 27])
© 2005 by Taylor & Francis Group, LLC
Chapter 10
ADAPTIVITY FOR 3D
POTENTIAL THEORY
The subject of this chapter is error analysis and adaptivity with the BNM, as
applied to problems in 3D potential theory. The idea of using hypersingular
residuals, to obtain local error estimates for the BIE, was ﬁrst proposed by
Paulino [122] and Paulino et al. [123]. This idea has been applied to the collo
cation BEM (Paulino et al. [123], Menon et al. [96] and Paulino et al. [127]);
and has been discussed in detail in Chapter 2 of this book. This idea, applied
to the BCM, has appeared in Mukherjee and Mukherjee [111], and is presented
in Chapter 7 of this book. This idea has also been applied to the BNM, for
problems in 3D potential theory and linear elasticity [28]. Applications in
3D potential theory is the subject of this chapter, while applications in 3D
elasticity are discussed in the next chapter  the last one in this book.
10.1 Hypersingular and Singular Residuals
10.1.1 The Hypersingular Residual
Let the BNM (equation (9.1)) for potential theory be written in operator form
as:
L
BNM
(u, τ) = 0 (10.1)
and its numerical solution be (u
∗
, τ
∗
). Also, the HBNM (equation (9.12)) is
written in operator form as:
L
HBNM
(u, τ) = 0 (10.2)
The hypersingular residual in the potential gradient u
,j
is deﬁned as,
r
j
≡ residual(u
,j
) = L
HBNM
(u
∗
, τ
∗
) (10.3)
175
© 2005 by Taylor & Francis Group, LLC
176 CHAPTER 10. ADAPTIVITY FOR 3D POTENTIAL THEORY
Solution Solution
BIE
BIE HBIE
HBIE
Error estimate Error estimate
BVP BVP
(a) (b)
Residual Residual
Figure 10.1: Interchange of BIE and HBIE (a) hypersingular residual (b) sin
gular residual (from [28])
and is calculated from equation (9.12).
This idea is illustrated in Figure 10.1(a)
It has been proved in [96] and [127] for the BIE that, under certain favorable
conditions, real positive constants c
1
and c
2
exist such that:
c
1
r ≤ ≤ c
2
r (10.4)
where r is some scalar measure of a hypersingular residual and is a scalar
measure of the exact local error. Thus, a hypersingular residual is expected to
provide a good estimate of the local error on a boundary element. It should be
mentioned here that the deﬁnitions of the residuals used in [96] and [127] are
analogous to, but diﬀerent in detail from, the ones proposed in this chapter.
10.1.2 The Singular Residual
The argument for using the residuals as error estimates is symmetric (see
Paulino [122], Paulino et al. [123]). Therefore, one can reverse the above pro
cedure to deﬁne singular residuals by ﬁrst solving the HBIE and then iterating
with the BIE.
In this case, for potential theory, one gets from equation (9.13):
L
HBNM
(u
o
, τ
o
) = 0 (10.5)
and from equation (9.1):
© 2005 by Taylor & Francis Group, LLC
10.2. ERROR ESTIMATION AND ADAPTIVE STRATEGY 177
Initial cell
configuration
BNM
simulation
HBNM
residuals
Local error
estimates
Global error
estimates
Redefinition of cells,
nodes and regions
of influence
Final
configuration
Convergence
No
Yes
Start
Iterative Cell Design Cycle
Figure 10.2: Typical selfadaptive iterative BNM algorithm (h−version) accord
ing to the scheme of Figure 10.1(a). The BNM equation used for solving the
BVP is (9.1), and the HBNM equation used for residual computation is (9.12)
(from ([28])
r ≡ residual(u) = L
BNM
(u
o
, τ
o
) (10.6)
This idea is illustrated in Figure 10.1(b).
The above formulation for singular and hypersingular residuals is a gener
alization of the earlier work by Menon et al. [96] in the sense that Dirichlet,
Neumann and mixed problems require separate prescriptions in [96], while the
current work presents a uniﬁed residual formulation.
10.2 Error Estimation and Adaptive Strategy
There are similarities between adaptive techniques (e.g. hversion) for mesh
based methods (see Paulino et al. [126, 124]) and meshless methods. How
ever, the latter set of methods provides substantially more ﬂexibility in the
(re)discretization process than the former ones.
The h−version iterative selfadaptive procedure employed in this work is
presented in the ﬂowchart  Figure 10.2. The goal is to eﬃciently develop a
ﬁnal cell conﬁguration which leads to a reliable numerical solution, in as simple
a manner as possible.
© 2005 by Taylor & Francis Group, LLC
178 CHAPTER 10. ADAPTIVITY FOR 3D POTENTIAL THEORY
10.2.1 Local Residuals and Errors  Hypersingular Resid
ual Approach
From equation (10.3):
r
j
= residual(u
,j
) (10.7)
A scalar residual measure is deﬁned as:
r = r
j
r
j
(10.8)
The exact local error in the gradient, u
,j
, is deﬁned as:
j
= u
(exact)
,j
−u
(numerical)
,j
(10.9)
and the corresponding scalar measure is deﬁned as:
=
j
j
(10.10)
Equations (10.8) and (10.10) are used to calculate the hypersingular residual
and exact error, respectively, in the gradient u
,j
, at each node, for problems in
potential theory.
10.2.2 Local Residuals and Errors  Singular Residual Ap
proach
The singular residual is deﬁned in an analogous fashion. From equation (10.6):
r = residual(u) (10.11)
and the exact local error in u is deﬁned as,
e = u
(exact)
−u
(numerical)
 (10.12)
Here, r and are themselves scalar measures of the residual and exact er
ror, respectively. Equations (10.6) and (10.12) are used to obtain the singular
residual and exact error, respectively, in the potential u, at each node, for prob
lems in potential theory. These equations are presented here for the sake of
completeness.
The local error measure (equation (10.12)) is also used for ∂u/∂n at points
on the surface of a cube (see examples of Section 10.3). This quantity is deﬁned
as:
e
∂u
∂n
=
∂u
∂n
(exact)
−
∂u
∂n
(numerical)
(10.13)
This error measure is only used in Figure 10.8.
The scalar residual measures, deﬁned above, evaluated at nodes, are used
as error estimators. In all the adaptivity examples presented in this chapter,
© 2005 by Taylor & Francis Group, LLC
10.2. ERROR ESTIMATION AND ADAPTIVE STRATEGY 179
one node is used for each cell and is placed at its centroid. The scalar residual
measure at this centroidal node is used as an error estimator for that cell. A
comparison of the residual r and exact error demonstrates the eﬀectiveness of
residuals as error estimates.
10.2.3 Cell Reﬁnement Criterion
A simple criterion for cell reﬁnement consists of subdividing the cells for which
the error indicator is larger than a certain reference value. In this work, the
reference quantity is taken as the average value of the error indicator (here the
average residual) given by:
¯ r =
1
N
n
N
n
i=1
r
(i)
(10.14)
where N
n
is the total number of nodes. If the inequality:
r > γ ¯ r (10.15)
is satisﬁed, then the cell is subdivided into four pieces (see Figure 10.3). The
parameter γ in equation (10.15) is a weighting coeﬃcient that controls the “cell
reﬁnement velocity.” The standard procedure consists of using γ = 1.0. If
γ > 1.0, then the number of cells to be reﬁned is less than with γ = 1.0.
According to Figure 10.2, the numerical solution of the next iterative step
is expected to be more accurate than that of the current step; however, the
increase on the total number of cells is comparatively small when γ > 1.0.
If γ < 1.0, then the number of cells to be reﬁned is larger than that with
γ = 1.0. The advantage in this case is that the reﬁnement rate increases, how
ever, the computational eﬃciency may decrease owing to likely generation of an
excessive number of cells. An alternative procedure, for a ONEstep reﬁnement,
is presented in Section 10.4 of this chapter.
10.2.4 Global Error Estimation and Stopping Criterion
Global L
2
error. A global L
2
error, on a panel, or over the whole boundary
∂B, is deﬁned as
¯ (φ) =
A
(φ
(exact)
−φ
(numerical)
)
2
dA
A
(φ
(exact)
)
2
dA
100% (10.16)
where φ is a variable of interest and A is the area of a panel or of the whole
surface ∂B. These global errors are used in many of the tables that are presented
later in this paper.
An indication of overall convergence may be obtained by evaluating either
¯ r (equation (10.14)) or ¯ from equation (10.16). Of course, equation (10.16) is
only useful for test examples in which the exact solution is known.
© 2005 by Taylor & Francis Group, LLC
180 CHAPTER 10. ADAPTIVITY FOR 3D POTENTIAL THEORY
.
.
. .
.
.
.
.
Subdivided cell
Subdivided cell
.
.
Quadrilateral cell
Triangular cell
Figure 10.3: Cell reﬁnement for quadrilateral and triangular cells with one node
per cell (from ([28])
Stopping criterion. For generic problems where the exact solution is not
available (e.g. most engineering problems), cell reﬁnement (see Figure 10.3)
can be stopped when:
¯ r ≤ r
global
(10.17)
where r
global
has a preset value, which depends on the overall level of accu
racy desired. The goal of the adaptive procedure is to obtain welldistributed
(i.e. near optimal) cell conﬁgurations. Ideally, as the iterative cell reﬁnement
progresses, the error estimates should decrease both locally and globally.
10.3 Progressively Adaptive Solutions: Cube
Problem
The adaptive process illustrated by Figure 10.2 is applied to the representative
example of a Dirichlet problem on a cube. Laplace’s equation is solved using
the BNM, and the (hypersingular) residuals are obtained using the HBNM,
according to Figure 10.1(a).
This example, together with an elasticity example discussed in Chapter 11,
permits assessment of various parameters of the adaptive strategy for meshless
methods based on BIE techniques. Several aspects are investigated such as the
quality of the adaptive solution obtained for scalar (potential theory) and vector
ﬁeld (elasticity theory) problems, performance of the method on problems with
either pure or mixed boundary conditions, evaluation of the quality of error
estimates obtained by means of hypersingular or singular residuals, sensitivity
© 2005 by Taylor & Francis Group, LLC
10.3. PROGRESSIVELY ADAPTIVE SOLUTIONS: CUBE PROBLEM 181
of the “ﬁnal” solution with respect to the starting cell conﬁguration (initial
condition of the selfadaptive problem), and convergence properties.
10.3.1 Exact Solution
The following exact solution, which satisﬁes the 3D Laplace’s equation, is used
in this example :
u = sinh
πx
1
2
sin
πx
2
2
√
2
sin
πx
3
2
√
2
(10.18)
Note that the solution (10.18) is symmetric with respect to y and z but that
its dependence on x is diﬀerent from its dependence on y or z. The appropriate
value of u is prescribed on ∂B (Dirichlet problem) and ∂u/∂n is computed on
∂B. Because the exact solution cannot be represented in terms of polynomials,
this is a proper test of the meshless method and the adaptivity procedure. A
quadratic basis is used for the construction of the MLS interpolating functions,
i.e. m = 6 (see equation (8.2)). The idea behind the adaptive procedure is to
start with a rather crude cell conﬁguration and carry out cell reﬁnement in the
region where the residual is large according to a certain criterion. Hence, the
adaptivity results in this section have been obtained starting with two diﬀerent
relatively coarse initial cell conﬁgurations. This comparative procedure tests
the sensitivity of the adaptive scheme with respect to the initial conditions.
10.3.2 Initial Cell Conﬁguration # 1 (54 Surface Cells)
Figure 10.4(a) shows a discretization consisting of 54 rectangular cells with
one (centroidal) node per cell. The boundary value problem is solved using
the BNM (equation (9.1)). Then the results are used in the HBNM (equation
(9.12)) to obtain the hypersingular residual. Figure 10.5 shows a comparison
between the hypersingular residual (from equations (10.3) and (10.8)) and the
exact local error in u
,j
(from equations (10.9) and (10.10)) computed for the
initial conﬁguration # 1 (Figure 10.4(a)) at each node on the surface. It can
be clearly seen that the hypersingular residual tracks the exact error perfectly.
Cell reﬁnement is carried out using γ = 1.0 in equation (10.15), and the
resulting reﬁned cell conﬁguration consisting of 126 cells is shown in Figure
10.4(b). It can be seen from Figure 10.4(b) that the cell reﬁnement occurs only
at the corners where the exact error is the largest. This is an indication that the
procedure for error estimation and adaptivity is moving in the right direction.
Now, the boundary value problem is solved again using the BNM. Table 10.1
summarizes the various output parameters of the adaptivity procedure. It can
be seen from Table 10.1 that excellent numerical results are obtained in a single
step of the adaptivity process and hence the adaptive procedure is not continued
further.
© 2005 by Taylor & Francis Group, LLC
182 CHAPTER 10. ADAPTIVITY FOR 3D POTENTIAL THEORY
1
0.5
0
0.5
1
1
0.5
0
0.5
1
1
0.5
0
0.5
1
x
y
z
_
_
_
_
_
_
(a)
1
0.5
0
0.5
1
1
0.5
0
0.5
1
1
0.5
0
0.5
1
x
y
z
_
_
_
_
_
_
(b)
Figure 10.4: Cell conﬁgurations on the surface of a cube. (a) Initial conﬁgu
ration # 1: 54 surface cells. (b) First adapted step : 126 cells, obtained with
γ = 1 (from ([28])
Output parameters Initial Final
Number of cells 54 126
x = ±1 1.4209 % 0.0238 %
y = ±1 7.6911 % 0.2773 %
z = ±1 7.6911 % 0.2578 %
All faces 2.1450 % 0.0519 %
Average residual (¯ r) 0.2366E01 0.7605E02
Maximum residual r
max
0.5197E01 0.3068E01
Table 10.1: ¯ (∂u/∂n) (see (10.16)) and residuals ¯ r, r
max
for the initial cell
conﬁguration (Figure 10.4(a)) and the conﬁguration obtained at the end of the
ﬁrst step of the adaptivity process using γ = 1.0 (Figure 10.4(b)) (from [28])
10.3.3 Initial Cell Conﬁguration # 2 (96 Surface Cells)
The initial conﬁguration # 2 is for the same physical cube with 16 uniform cells
on each face (Figure 10.7(a)) with, as always, one node at the centroid of each
cell. As before, the boundary value problem is solved using the BNM (equa
tion (9.1)), and the results obtained are used in the HBNM (equation (9.12)).
Figure 10.6 shows a comparison between the hypersingular residual (from equa
tions (10.3) and (10.8)) and the exact local error in u
,j
(from equations (10.9)
and (10.10)), computed for the initial conﬁguration # 2 (Figure 10.7(a)). It
can be clearly seen that the hypersingular residual tracks the exact error very
accurately. In fact, the results for the ﬁner cell conﬁguration #2 are very sim
ilar to those shown in Figure 10.5 for the coarser initial cell conﬁguration #
1.
© 2005 by Taylor & Francis Group, LLC
10.3. PROGRESSIVELY ADAPTIVE SOLUTIONS: CUBE PROBLEM 183
0 10 20 30 40 50
0
0.2
0.4
0.6
0.8
1
1.2
Node number
H
y
p
e
r
s
i
n
g
u
l
a
r
r
e
s
i
d
u
a
l
a
n
d
e
x
a
c
t
e
r
r
o
r
Hypersingular residual
Exact error
Figure 10.5: Comparison of hypersingular residual and exact local error
in u
,j
for the initial conﬁguration # 1 (54 cells, one node per cell). These
quantities have been normalized by their respective maximum values, where
r
max
= 0.5197 ×10
−1
and
max
= 0.2051 (from ([28])
0 20 40 60 80
0
0.2
0.4
0.6
0.8
1
1.2
Node number
H
y
p
e
r
s
i
n
g
u
l
a
r
r
e
s
i
d
u
a
l
a
n
d
e
x
a
c
t
e
r
r
o
r
Hypersingular residual
Exact error
Figure 10.6: Comparison between hypersingular residual and exact local error
in u
,j
for the initial conﬁguration # 2 (96 cells, one node per cell). These
quantities have been normalized by their respective maximum values, where
r
max
= 0.1829 ×10
−1
and
max
= 0.6223 ×10
−1
(from ([28])
© 2005 by Taylor & Francis Group, LLC
184 CHAPTER 10. ADAPTIVITY FOR 3D POTENTIAL THEORY
1
0.5
0
0.5
1
1
0.5
0
0.5
1
1
0.5
0
0.5
1
x
y
z
(a)
0.5
0
0.5
1
1
0.5
0
0.5
1
1
0.5
0
0.5
1
x
y
z
_
_
_
_
_
1
_
_
(b)
1
0.5
0
0.5
1
1
0.5
0
0.5
1
1
0.5
0
0.5
1
x
y
z
_
_
_
_
_
_
(c)
1
0.5
0
0.5
1
1
0.5
0
0.5
1
1
0.5
0
0.5
1
x
y
_
_
_
_ _
_
(d)
Figure 10.7: Cell conﬁgurations on the surface of a cube (a) initial conﬁguration
# 2 (96 surface cells), (b) ﬁrst adapted step (168 cells), (c) second step (456
cells), (d) third step (1164 cells) (from ([28])
© 2005 by Taylor & Francis Group, LLC
10.3. PROGRESSIVELY ADAPTIVE SOLUTIONS: CUBE PROBLEM 185
Adaptivity results. In order to obtain a better understanding of the adap
tivity procedure, the local error e in ∂u/∂n (from equation (10.13)) is calculated
on each of the faces of the cube. The iterative cell design cycle of Figure 10.2
is repeated three times using γ = 0.5 in equation (10.15) and starting from
the initial conﬁguration # 2 given in Figure 10.7(a). The resulting reﬁned
cell conﬁgurations are shown in Figures 10.7(b), (c) and (d), respectively. It is
noted that the cell reﬁnement should begin at the corners of the cube where
the error in ∂u/∂n is the largest.
Figure 10.8 shows contour plots of the exact local error e in ∂u/∂n on the
y = −1 face of the cube. The underlying cell structure on the face is also
shown in the color plots. The resolution of these and subsequent contour plots
is much ﬁner than the corresponding cell discretization because the error is
actually evaluated at a large number of points on the boundary (panels) of the
body. These results conﬁrm the observation made at the end of the previous
paragraph regarding regions of large errors which demand a ﬁner discretization.
Thus, reﬁnement occurs close to the edges and corners where the error in ∂u/∂n
is largest.
Other relevant comments are in order. For the ﬁrst step of the adaptive pro
cedure (see Figure 10.7(b)), selected results are shown in Figure 10.9, which pro
vides a comparison between the hypersingular residual (from equations (10.3)
and (10.8)) and the exact local error (from equations (10.9) and (10.10)). The
results are shown on the x = −1 and z = 1 faces as a representative sample
of the results over the 168 nodes. It can be seen from Figure 10.9 that the
hypersingular residual tracks the exact error reasonably well.
Figure 10.8(b) shows a contour plot for the exact local error e in ∂u/∂n on
the y = −1 face of the cube for adapted cell conﬁguration of Figure 10.7(b).
Note that, due to the reﬁnement procedure, the error in ∂u/∂n has reduced
substantially, especially at the corners (cf Figures 10.8(a) and (b)).
Figures 10.8(c) and (d) show the exact local error e in ∂u/∂n on the y = −1
face of the cube for the adapted cell conﬁgurations consisting of 456 cells and
1164 cells, respectively (see Figures 10.9 (c) and (d)). Comparing the contour
plots of Figures 10.8(a)(d), one can readily verify that the error in ∂u/∂n
decreases substantially during the adaptive process. It is interesting to note
that the absolute value of the exact solution (equation (10.18)) has the same
functional dependence on the y = −1 and z = 1 faces and diﬀerent on the
x = −1 face of the cube. Step 1 (Figure 10.7(b)) is not sensitive to this
fact, however, Steps 2 (Figure 10.7(c)) and 3 (Figure 10.7(d)) of the adaptive
procedure are. This is a tribute to the quality of residuals as error estimates.
Table 10.2 summarizes the results of the adaptive process for the cube prob
lem starting with the initial cell conﬁguration consisting of 96 cells (Figure
10.7(a)). Note that, on the faces x = ±1, ¯ (∂u/∂n) (from equation (10.16))
increases from the initial conﬁguration to Step 1, and from Step 1 to Step 2.
However, ¯ (∂u/∂n) ﬁnally decreases from Step 2 to Step 3 and reaches its lowest
value at this step, which has the sophisticated cell pattern of Figure 10.7(d). On
the faces y = ±1 and z = ±1, ¯ (∂u/∂n) monotonically decreases as the number
© 2005 by Taylor & Francis Group, LLC
186 CHAPTER 10. ADAPTIVITY FOR 3D POTENTIAL THEORY
0.5 0 0.5
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
x
z
0.02
0.04
0.06
0.08
0.1
0.12
_
_
_
_
_
(a)
0.5 0 0.5
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
x
z
0.01
0.02
0.03
0.04
0.05
0.06
0.07
_
_
_
_
_
(b)
0.5 0 0.5
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
x
z
0.01
0.02
0.03
0.04
0.05
0.06
0.07
_
_
_
_
_
(c)
0.5 0 0.5
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
x
z
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
_
_
_
_
_
(d)
Figure 10.8: Error in ∂u/∂n (e(∂u/∂n)) on the face y = −1 of the cube (a)
initial conﬁguration # 2 (96 surface cells), (b) ﬁrst adapted step (168 cells), (c)
second step (456 cells), (d) third step (1164 cells) (from ([28])
© 2005 by Taylor & Francis Group, LLC
10.3. PROGRESSIVELY ADAPTIVE SOLUTIONS: CUBE PROBLEM 187
0 10 20 30 40 50
0
0.2
0.4
0.6
0.8
1
1.2
Node number
H
y
p
e
r
s
i
n
g
u
l
a
r
r
e
s
i
d
u
a
l
a
n
d
e
x
a
c
t
e
r
r
o
r
Hypersingular residual
Exact error
Figure 10.9: Comparison of hypersingular residual and exact local error in u
,j
on the faces x = −1 and z = 1 of the cube of Figure 10.7(b) (ﬁrst step of the
adaptive procedure). The quantities have been normalized by their respective
maximum values, where r
max
= 0.1567 × 10
−1
and
max
= 0.3645 × 10
−1
(from ([28])
© 2005 by Taylor & Francis Group, LLC
188 CHAPTER 10. ADAPTIVITY FOR 3D POTENTIAL THEORY
Output parameters Initial Step 1 Step 2 Step 3
Number of cells 96 168 456 1164
x = ±1 0.0759 % 0.1062 % 0.1135 % 0.0438 %
y = ±1 1.0654 % 0.2785 % 0.2089 % 0.0551 %
z = ±1 1.0696 % 0.2781 % 0.2091 % 0.0551 %
All faces 0.1899 % 0.1269 % 0.1247 % 0.0451 %
Average residual ¯ r 0.4963E02 0.3661E02 0.5643E03 0.1811E03
Max. residual r
max
0.1829E01 0.1567E01 0.3579E02 0.2537E02
Table 10.2: L
2
error in ∂u/∂n (¯ (∂u/∂n)) and residuals ¯ r, r
max
for the various
steps of the adaptivity process starting with the initial cell conﬁguration con
sisting of 96 cells with one node per cell (Figure 10.7(a)). Here γ = 0.5 is used
for the cell reﬁnement of the cube (from [28])
of adaptive cycles increases. Moreover, as expected, the global ¯ (∂u/∂n) for “all
faces,” as well as the average and maximum residuals, decrease as the adaptive
process progresses.
10.4 OneStep Adaptive Cell Reﬁnement
The previous section has dealt with an iterative adaptive technique for cell
reﬁnement (h−version). Here the interest is on developing a simple ONEstep
algorithm for cell reﬁnement in the meshless BNM setting. The ﬂowchart of
Figure 10.10 illustrates this idea which is based on the concept of reﬁnement
level (RL) employed by Krishnamoorthy and Umesh [74].
Reﬁnement strategy. Figures 10.11(a) and (b) show that diﬀerent degrees
of reﬁnement are carried out for diﬀerent values of the reﬁnement level. From
these ﬁgures, the expression relating the ﬁnal cell size h
f
to the reﬁnement level
RL is:
h
f
=
h
i
2
RL
(10.19)
where h
i
denotes the initial cell size. Assuming that the rate of convergence of
the error is O(h
p
), where h is a characteristic cell size in the area covered by the
cells, which are of order p, and setting the error estimate equal to η = r/(γ¯ r)
(see equations (10.14) and (10.15)), one obtains:
h
f
=
h
i
η
1/p
(10.20)
From equations (10.19) and (10.20), the RL is given by:
© 2005 by Taylor & Francis Group, LLC
10.4. ONESTEP ADAPTIVE CELL REFINEMENT 189
Redefinition of
cells, nodes, and
regions of influence
Start
Configuration
Configuration
Initial
HBNM
residuals
Error
estimates
BNM
simulation
Final
Multilevel cell refinement
Figure 10.10: ONEstep adaptive BEM algorithm based on multilevel cell re
ﬁnement (from ([28])
= 0
# of cells = 4
= 1
= 3
# of cells = 4
# of cells = 4 = 4
# of cells = 4
0 1
3
RL RL
RL RL = 2
= 16
2
= 1
= 64
(a)
= 0
# of cells = 4
= 1
= 3
# of cells = 4
# of cells = 4 = 4
# of cells = 4
3
RL RL
RL RL = 2
2
= 16 = 64
1 0
= 1
(b)
Figure 10.11: Reﬁnement level RL using (a) rectangular and (b) triangular
cells. The bold lines illustrate the idea of cell structure embedding (from ([28])
© 2005 by Taylor & Francis Group, LLC
190 CHAPTER 10. ADAPTIVITY FOR 3D POTENTIAL THEORY
RL =
log η
p log 2
for η ≥ 1
0 for η < 1
(10.21)
where p is order of the interpolating function. For the interpolation procedure
used in this work, p = m. The second condition in equation (10.21) is enforced
because cell structure coarsening is not considered in this work.
This idea of ONEstep reﬁnement is applied below to the cube problem of
Section 10.3. Errors are again estimated by means of hypersingular residuals.
10.4.1 Initial Cell Conﬁguration # 1 (54 Surface Cells)
The ONEstep multilevel strategy is implemented on the cube of Figure 10.12(a),
which consists of 54 cells with one node per cell. The boundary value problem
is solved using the BNM (equation (9.1)) by imposing the exact solution in
equation (10.18) as Dirichlet boundary conditions. The hypersingular residual
(from equations (9.12 and (10.8)) is obtained and then the multilevel reﬁne
ment procedure is carried out using γ = 0.15. The cell structure obtained in
ONEstep is shown in Figure 10.12(b), which consists of 438 cells with one
node per cell. Table 10.3 shows a comparison of the results from the ONEstep
multilevel reﬁnement scheme starting with the conﬁguration of Figure 10.12(a)
and ending with the conﬁguration of Figure 10.12(b). This table shows that
¯ (∂u/∂n) and the residual consistently decrease with reﬁnement.
1
0.5
0
0.5
1
1
0.5
0
0.5
1
1
0.5
0
0.5
1
x
y
z
_
_
_
_ _
_
(a)
1
0.5
0
0.5
1
1
0.5
0
0.5
1
1
0.5
0
0.5
1
x
y
z
_
_
_
_
_
_
(b)
Figure 10.12: ONEstep multilevel cell reﬁnement for the cube problem. (a)
Initial conﬁguration # 1 with 54 cells. (b) Adapted conﬁguration with 438 cells
using γ = 0.15 (from ([28])
© 2005 by Taylor & Francis Group, LLC
10.4. ONESTEP ADAPTIVE CELL REFINEMENT 191
Output parameters Initial Final
Number of cells 54 438
x = ±1 1.4209 % 0.0411 %
y = ±1 7.6911 % 0.0339 %
z = ±1 7.6911 % 0.0343 %
All faces 2.1450 % 0.0403 %
Average residual ¯ r 0.2366E01 0.4615E03
Maximum residual r
max
0.5197E01 0.2618E01
Table 10.3: ¯ (∂u/∂n) and residuals ¯ r, r
max
for the initial conﬁguration (Fig
ure 10.12(a)), and the ﬁnal conﬁguration (Figure 10.12(b)) obtained by the
multilevel reﬁnement strategy with γ = 0.15 (from [28])
10.4.2 Initial Cell Conﬁguration # 2 (96 Surface Cells)
The BNM (equation (9.1)) is used to solve the boundary value problem using
equation (10.18) as the exact solution, and the hypersingular residual is ob
tained by means of equations (9.12) and (10.8). Multilevel reﬁnement is carried
out using γ = 0.15. The cell structure obtained in ONEstep is shown in Figure
10.13(b), which consists of 1764 cells with one node per cell.
1
0.5
0
0.5
1
1
0.5
0
0.5
1
1
0.5
0
0.5
1
x
y
z
(a)
1
0.5
0
0.5
1
1
0.5
0
0.5
1
1
0.5
0
0.5
1
x
y
z
(b)
Figure 10.13: ONEstep multilevel cell reﬁnement for the cube problem (a)
initial conﬁguration # 2 with 96 cells (b) adapted conﬁguration with 1764 cells
using γ = 0.15 (from ([28])
Table 10.4 summarizes the results for ¯ (∂u/∂n) and the residual for the mul
tilevel reﬁnement strategy starting from the initial conﬁguration # 2 of 96 cells
(see Figure 10.13(a)). Qualitatively, these results are analogous to those of Ta
ble 10.2 for the progressive adaptive reﬁnement, which includes the peculiarity
© 2005 by Taylor & Francis Group, LLC
192 CHAPTER 10. ADAPTIVITY FOR 3D POTENTIAL THEORY
Output parameters Initial Final
Number of cells 96 1764
x = ±1 0.0759 % 0.1034 %
y = ±1 1.0654 % 0.2400 %
z = ±1 1.0696 % 0.2400 %
All faces 0.1899 % 0.1169 %
Average residual (¯ r) 0.4963E02 0.3370E03
Maximum residual (r
max
) 0.1829E01 0.1535E01
Table 10.4: ¯ (∂u/∂n) and residual for the initial conﬁguration (Figure 10.13(a)),
and the ﬁnal conﬁguration (Figure 10.13(b)) obtained by the multilevel reﬁne
ment strategy with γ = 0.15 (from [28])
observed on the x = ±1 faces of the cube. Moreover, the remarks concerning
Table 10.2 also hold for explaining the results of Table 10.4. Therefore, for
further explanations, the reader is referred to Section 10.3.
© 2005 by Taylor & Francis Group, LLC
Chapter 11
ADAPTIVITY FOR 3D
LINEAR ELASTICITY
The subject of this chapter is error analysis and adaptivity with the BNM, as
applied to 3D linear elasticity. Please see the introduction to Chapter 10.
11.1 Hypersingular and Singular Residuals
11.1.1 The Hypersingular Residual
Let the BNM (equation (9.24)) for elasticity be written in operator form as:
L
BNM
(u
k
, τ
k
) = 0 ; k = 1, 2, 3 (11.1)
with the numerical solution (u
∗
k
, τ
∗
k
). Also, the HBNM (equation (9.27)) is
written in operator form as:
L
HBNM
(u
k
, τ
k
) = 0 ; k = 1, 2, 3 (11.2)
This time, the stress residual is deﬁned from the stress HBNM (equation
(9.27)) as,
r
ij
≡ residual(σ
ij
) = L
HBNM
(u
∗
k
, τ
∗
k
) ; k = 1, 2, 3 (11.3)
This idea is illustrated in Figure 10.1(a).
It has been proved in [96] and [127] for the BIE that, under certain favorable
conditions, real positive constants c
1
and c
2
exist such that:
c
1
r ≤ ≤ c
2
r (11.4)
where r is some scalar measure of a hypersingular residual and is a scalar
measure of the exact local error. Thus, a hypersingular residual is expected to
provide a good estimate of the local error on a boundary element. It should be
193
© 2005 by Taylor & Francis Group, LLC
194 CHAPTER 11. ADAPTIVITY FOR 3D LINEAR ELASTICITY
mentioned here that the deﬁnitions of the residuals used in [96] and [127] are
analogous to, but diﬀerent in detail from, the ones proposed in this chapter.
11.1.2 The Singular Residual
From equation (9.28):
L
HBNM
(u
o
k
, τ
o
k
) = 0 ; k = 1, 2, 3 (11.5)
and from equation (9.24)
r
i
≡ residual(u
i
) = L
BNM
(u
o
k
, τ
o
k
) ; k = 1, 2, 3 (11.6)
This idea is illustrated in Figure 10.1(b).
The above formulation for singular and hypersingular residuals is a gener
alization of the earlier work by Menon et al. [96] in the sense that Dirichlet,
Neumann and mixed problems require separate prescriptions in [96], while the
current work presents a uniﬁed residual formulation.
11.2 Error Estimation and Adaptive Strategy
The h−version iterative selfadaptive procedure employed in this work is pre
sented in the ﬂowchart  Figure 10.2. The goal is to eﬃciently develop a ﬁnal
cell conﬁguration which leads to a reliable numerical solution, in as simple a
manner as possible.
11.2.1 Local Residuals and Errors  Hypersingular Resid
ual Approach
From equation (11.3) :
r
ij
= residual(σ
ij
) (11.7)
A scalar residual measure is deﬁned as:
r = r
ij
r
ij
(11.8)
The exact local error in stress is deﬁned as:
ij
= σ
(exact)
ij
−σ
(numerical)
ij
(11.9)
and the corresponding scalar measure is deﬁned as:
=
ij
ij
(11.10)
Equations (11.8) and (11.10) are used to compute the hypersingular residual
and exact error, respectively, in the stress σ
ij
, at each node, for problems in
linear elasticity.
© 2005 by Taylor & Francis Group, LLC
11.3. PROGRESSIVELY ADAPTIVE SOLUTIONS: PULLING A ROD 195
11.2.2 Local Residuals and Errors  Singular Residual Ap
proach
From equation (11.6):
r
i
= residual(u
i
) (11.11)
so that a scalar residual measure is:
r = r
i
r
i
(11.12)
The exact local error in u
i
is deﬁned as:
i
= u
(exact)
i
−u
(numerical)
i
(11.13)
with a corresponding scalar measure:
=
i
i
(11.14)
Equations (11.12) and (11.14) are used to obtain the singular residual and
exact error, respectively, in the displacement u
i
, at each node, for elasticity
problems.
The scalar residual measures, deﬁned above, evaluated at nodes, are used
as error estimators. In all the adaptivity examples presented in this chapter,
one node is used for each cell and is placed at its centroid. The scalar residual
measure at this centroidal node is used as an error estimator for that cell. A
comparison of the residual r and exact error demonstrates the eﬀectiveness of
residuals as error estimates.
11.2.3 Cell Reﬁnement Global Error Estimation and Stop
ping Criterion
The algorithms used here are the same as those described in Sections 10.2.3
and 10.2.4 in Chapter 10.
11.3 Progressively Adaptive Solutions: Pulling
a Rod
The problem under consideration here is the stretching of a cylindrical elastic
rod clamped at one end (see Figure 11.1(a)). This time, the roles of the BNM
and HBNM are reversed, i.e. the HBNM (9.28) is used to solve the boundary
value problem while the singular residuals are obtained from the standard BNM
(9.24) (see Figure 10.1(b)). This same problem has been addressed with the
BCM in Section 7.3.1 of Chapter 7.
© 2005 by Taylor & Francis Group, LLC
196 CHAPTER 11. ADAPTIVITY FOR 3D LINEAR ELASTICITY
σ
0
x
z
L
D
(a)
2 1 0 1 2
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
x
y
_ _
_
_
_
_
_
(b)
3 2 1 0 1 2 3
0
0.5
1
1.5
2
θ
z
_ _ _
(radians)
(c)
Figure 11.1: Stretching of a short clamped cylindrical rod by a uniform tensile
load (a) physical situation : L = 2.0, D = 4.0, ν = 0.25, E = 1.0, σ
0
= 1.0 (b)
and (c) initial cell conﬁguration with 144 cells (one node per cell)  (b) clamped
and loaded faces (c) curved surface of the cylindrical rod mapped onto the (z, θ)
plane (from [28])
© 2005 by Taylor & Francis Group, LLC
11.3. PROGRESSIVELY ADAPTIVE SOLUTIONS: PULLING A ROD 197
20 40 60 80 100 120 140
0
0.2
0.4
0.6
0.8
1
Node number
S
i
n
g
u
l
a
r
r
e
s
i
d
u
a
l
Traction ( σ
0
)
Clamped
Traction free
Figure 11.2: Singular residual for the initial conﬁguration of 144 cells on the
cylindrical rod of Figure 11.1 (b) and (c). The residual has been normalized
with respect to its maximum value, r
max
= 0.2419 ×10
−2
(from [28])
11.3.1 Initial Cell Conﬁguration
The geometric and material parameters chosen are: E = 1.0, ν = 0.25, σ
0
= 1.0,
L = 2.0, and D = 4.0. Figures 11.1(b) and (c) show the initial cell conﬁguration
on the clamped and loaded faces and on the curved surface of the rod. The
boundary value problem is solved by prescribing tractions on the top face of the
cylinder, with the bottom surface completely clamped, and the curved surface
traction free. Upon obtaining the solution to the boundary value problem,
the singular residual is obtained at each node. Figure 11.2 shows the singular
residual (from equations (9.28), (11.6), and (11.12)) obtained for the initial cell
conﬁguration (144 cells). It can be seen that the residual is considerably higher
on the clamped face and on the curved surface near the clamped face, than on
the loaded face. This is to be expected considering the physical nature of the
problem at hand which has a singularity on the bounding circle of the clamped
face [37, 137].
11.3.2 Adaptivity Results
The adaptive strategy is carried out according to the ﬂow chart of Figure 10.2.
However, the boxes for the “BNM simulation” and “HBNM residuals” are re
placed by “HBNM simulation” and “BNM residuals,” according to the scheme
of Figure 10.1(b). Since the singular residual is higher on the clamped and
curved faces, most of the subdivision of cells occurs on those faces. The curved
© 2005 by Taylor & Francis Group, LLC
198 CHAPTER 11. ADAPTIVITY FOR 3D LINEAR ELASTICITY
surface of the cylinder near the clamped surface is reﬁned due to the singularity
at the edge of the clamped face. However, the top face (the loaded face) is NOT
reﬁned at all and so the cell structure on that face remains as shown in Figure
11.1(b).
Three steps of adaptivity are pursued using γ = 1.25 in equation (10.15) and
starting from the initial conﬁguration of Figures 11.1(b) and (c). The resulting
reﬁned cell conﬁgurations are shown in Figures 11.3  11.5. As expected, Figures
11.3(a), 11.4(a) and 11.5(a) show that the loaded face is not reﬁned at all and
remains as in the initial conﬁguration (Figure 11.1(b)). On the clamped face, a
comparison of Figures 11.1(b), 11.3(b), 11.4(b), and 11.5(b) indicates that cell
reﬁnement only takes place near the edge of the face, which is the region where
gradients in stresses are largest.
Figures 11.1(c), 11.3(c), 11.4(c), and 11.5(c) show the progressive reﬁnement
on the curved surface of the cylinder. One can observe that reﬁnement primarily
occurs along the curved surface near the clamped edge of the cylindrical rod.
Note that, for the initial conﬁguration (Figure 11.1 (b) and (c)), the number
of subdivisions along the edge of the top and bottom faces is the same as the
number of subdivisions along the edge of the curved surface (12 subdivisions).
However, when adaptivity is carried out, a signiﬁcant mismatch in the number
of subdivisions is created at every adaptive step. This does not present any
problem for the meshless method, and such freedom in modeling is expected to
be especially advantageous in analyzing problems with complicated geometry.
11.4 OneStep Adaptive Cell Reﬁnement
The ONEstep procedure outlined in Section 10.4 is demonstrated here for the
same problem as above (Section 11.3). The initial conﬁguration is the same
as that of Figure 11.1(b) and (c). The present study also employs singular
residuals (equation (9.24)) for error estimation (see Section 11.3) rather than
hypersingular residuals (equation (9.27)). This is similar to the study carried
out for the same problem by means of iterative adaptive cell reﬁnement. The
results for the multilevel reﬁnement for the original problem of Figure 11.1(a)
and (b) are given in Figure 11.6. Comparing these results with the ones of
Section 11.3 (Figures 11.311.5), one veriﬁes that the overall trends are quite
similar in both situations. However, two main diﬀerences are noticeable. First,
the loaded face is reﬁned (a little) here (see Figures 11.1(b) and 11.6(a)), while
it is not reﬁned at all in the adapted conﬁgurations shown in Figures 11.3(a),
11.4(a) and 11.5(a). Second, as expected, the multilevel cell reﬁnement does
not allow the smooth cell gradation which occurs in progressively adapted cell
conﬁgurations (see Figures 11.5 and 11.6). Nevertheless, such gradation, which
is essential for meshbased methods, is not required at all in the present meshless
methods (BNM and HBNM).
© 2005 by Taylor & Francis Group, LLC
11.4. ONESTEP ADAPTIVE CELL REFINEMENT 199
2 1 0 1 2
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
x
y
Loaded face
_ _
_
_
_
_
_
(a)
2 1 0 1 2
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
x
y
Clamped face
_ _
_
_
_
_
_
(b)
3 2 1 0 1 2 3
0
0.5
1
1.5
2
θ
z
_ _ _
(radians)
(c)
Figure 11.3: Short clamped cylindrical rod Step 1 : Adapted conﬁguration
consisting of 228 cells obtained with γ = 1.25; (a) loaded face (b) clamped face
(c) curved surface of the rod (from [28])
© 2005 by Taylor & Francis Group, LLC
200 CHAPTER 11. ADAPTIVITY FOR 3D LINEAR ELASTICITY
2 1 0 1 2
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
x
y
Loaded face
_ _
_
_
_
_
_
(a)
2 1 0 1 2
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
x
y
Clamped face
_ _
_
_
_
_
_
(b)
3 2 1 0 1 2 3
0
0.5
1
1.5
2
θ
z
_ _
_
(radians)
(c)
Figure 11.4: Short clamped cylindrical rod Step 2 : Adapted conﬁguration
consisting of 324 cells obtained with γ = 1.25; (a) loaded face (b) clamped face
(c) curved surface of the rod (from [28])
© 2005 by Taylor & Francis Group, LLC
11.4. ONESTEP ADAPTIVE CELL REFINEMENT 201
2 1 0 1 2
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
x
y
Loaded face
_
_
_
_
_
_
_
(a)
2 1 0 1 2
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
x
y
Clamped face
_ _
_
_
_
_
_
(b)
3 2 1 0 1 2 3
0
0.5
1
1.5
2
θ
z
_ _
_
(radians)
(c)
Figure 11.5: Short clamped cylindrical rod Step 3 : Adapted conﬁguration
consisting of 576 cells obtained with γ = 1.25; (a) loaded face (b) clamped face
(c) curved surface of rod (from [28])
© 2005 by Taylor & Francis Group, LLC
202 CHAPTER 11. ADAPTIVITY FOR 3D LINEAR ELASTICITY
2 1 0 1 2
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
x
y
Loaded face
_
_
_
_
_
_
_
(a)
2 1 0 1 2
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
x
y
Clamped face
_ _
_
_
_
_
_
(b)
3 2 1 0 1 2 3
0
0.5
1
1.5
2
θ
z
_ _ _
(radians)
(c)
Figure 11.6: ONEstep multilevel cell reﬁnement for short clamped cylinder:
(a) loaded face (b) clamped face (c) curved surface of cylinder. The initial cell
conﬁguration is shown in Figure 11.1(b) and (c) (from [28])
© 2005 by Taylor & Francis Group, LLC
Bibliography
[1] Aluru, N.R. and White, J. (1997). An eﬃcient numerical technique for
electromechanical simulation of complicated microelectromechanical struc
tures. Sensors and Actuators A 58:111.
[2] Aluru, N.R. and Li, G. (2001). Finite cloud method: a true meshless tech
nique based on a ﬁxed reproducing kernel approximation. International
Journal for Numerical Methods in Engineering 50:23732410.
[3] Atluri, S.N. and Shen, S. (2002). The Meshless Local PetrovGalerkin
(MLPG) Method. Tech Science Press, Encino, CA.
[4] Banerjee, P.K. (1994). The Boundary Element Methods in Engineering.
McGraw Hill Europe, Maidenhead, Berkshire, UK.
[5] Banichuk, N.V. (1983). Problems and Methods of Optimal Structural De
sign. Plenum, New York.
[6] Bao, Z. and Mukherjee, S. (2004). Electrostatic BEM for MEMS with
thin conducting plates and shells. Engineering Analysis with Boundary
Elements. 28:14271435.
[7] Bao, Z. and Mukherjee, S. (2005). Electrostatic BEM for MEMS with thin
beams. Communications in Numerical Methods in Engineering. In press.
[8] Bao, Z., Mukherjee, S., Roman, M. and Aubry, N. (2004). Nonlinear vi
brations of beams strings plates and membranes without initial tension.
ASME Journal of Applied Mechanics 71:551559.
[9] Becker, A.A. (1992). The Boundary Element Method in Engineering. Mc
Graw Hill International, Singapore.
[10] Belytschko, T., Lu, Y.Y. and Gu, L. (1994). Elementfree Galerkin meth
ods. International Journal for Numerical Methods in Engineering 37:229
256.
[11] Belytschko, T., Krongauz, Y., Organ, D., Fleming, M. and Krysl, P. (1996).
Meshless methods: An overview and recent developments. Computer Meth
ods in Applied Mechanics and Engineering 139:347.
203
© 2005 by Taylor & Francis Group, LLC
204 BIBLIOGRAPHY
[12] Bobaru, F. and Mukherjee, S. (2001). Shape sensitivity analysis and shape
optimization in planar elasticity using the elementfree Galerkin method.
Computer Methods in Applied Mechanics and Engineering 190:43194337.
[13] Bobaru, F. and Mukherjee, S. (2002). Meshless approach to shape opti
mization of linear thermoelastic solids. International Journal for Numerical
Methods in Engineering 53:765796.
[14] Bonnet, M. (1995). Boundary Integral Equation Methods for Solids and
Fluids. Wiley, Chichester, UK.
[15] Bonnet, M. and Xiao, H. (1995). Computation of energy release rate using
material diﬀerentiation of elastic BIE for 3D elastic fracture. Engineering
Analysis with Boundary Elements 15:137149.
[16] Brebbia, C.A. and Dominguez, J. (1992). Boundary Elements: An In
troductory Course. 2nd edition. Computational Mechanics Publications,
Southampton, UK, and McGraw Hill, New York.
[17] Bui, H.D. (1975). Application des potentiels ´elastiques `a l’´etude des ﬁssures
planes de forme arbitrare en milieu tridimensionnel. Comptes Rendus de
l’Acad´emie des Sciences Paris S´erie A 280:11571160.
[18] Carstensen, C. (1995). Adaptive boundary element methods and adaptive
ﬁnite element methods and boundary element coupling. Boundary Value
Problems and Integral Equations in Nonsmooth Domains. M. Costabel, M.
Dauge and S. Nicaise eds., 4758. Marcel Dekker.
[19] Carstensen, C. and Stephan, E.P. (1995). A posteriori error estimates for
boundary element methods. Mathematics of Computation 64:483500.
[20] Carstensen, C., Estep, D. and Stephan, E.P. (1995). hadaptive boundary
element schemes. Computational Mechanics 15:372383.
[21] Carstensen, C. (1996). Eﬃciency of a posteriori BEM error estimates
for ﬁrst kind integral equations on quasiuniform meshes. Mathematics of
Computation 65:6984.
[22] Chandra, A. and Mukherjee, S. (1997). Boundary Element Methods in
Manufacturing. Oxford University Press, NY.
[23] Chati, M.K. (1999). Meshless Standard and Hypersingular Boundary Node
Method  Applications in ThreeDimensional Potential Theory and Linear
Elasticity. Ph.D. Dissertation, Cornell University, Ithaca, NY.
[24] Chati, M.K. and Mukherjee, S. (1999). Evaluation of gradients on the
boundary using fully regularized hypersingular boundary integral equa
tions. Acta Mechanica 135:4155.
© 2005 by Taylor & Francis Group, LLC
BIBLIOGRAPHY 205
[25] Chati, M.K., Mukherjee, S. and Mukherjee, Y.X. (1999). The boundary
node method for threedimensional linear elasticity. International Journal
for Numerical Methods in Engineering 46:11631184.
[26] Chati, M.K. and Mukherjee, S. (2000). The boundary node method for
threedimensional problems in potential theory. International Journal for
Numerical Methods in Engineering 47:15231547.
[27] Chati, M.K., Mukherjee, S. and Paulino, G.H. (2001). The meshless hyper
singular boundary node method for threedimensional potential theory and
linear elasticity problems. Engineering Analysis with Boundary Elements
25:639653.
[28] Chati, M.K., Mukherjee, S. and Paulino, G.H. (2001). The meshless stan
dard and hypersingular boundary node methods  applications to error esti
mation and adaptivity in threedimensional problems. International Jour
nal for Numerical Methods in Engineering 50:22332269.
[29] Chen, G. and Zhou, J. (1992). Boundary Element Methods. Academic
Press.
[30] Chen, J.T. and Hong, H.K. (1999). Review of dual boundary element meth
ods with emphasis on hypersingular integrals and divergent series. ASME
Applied Mechanics Reviews 52:1733.
[31] Chen, S.Y., Tang, W.X., Zhou, S.J., Leng, X. and Zheng, W. (1999). Evalu
ation of J integrals and stress intensity factors in a 2D quadratic boundary
contour method. Communications in Numerical Methods in Engineering
15:91100.
[32] Chen, W. and Tanaka, M. (2002). A meshless, integrationfree, boundary
only RBF technique. Computers and Mathematics with Applications
43:379391.
[33] Choi, J.H. and Kwak, B.M. (1988). Boundary integral equation method
for shape optimization of elastic structures. International Journal for Nu
merical Methods in Engineering 26:15791595.
[34] Ciarlet, P.G. (1991). Basic error estimates for elliptic problems. Finite
Element Methods (Part 1)  Vol. 2 of Handbook of Numerical Analysis. P.
G. Ciarlet and J. L. Lions eds., Elsevier Science Publishers.
[35] Costabel, M. and Stephan, E. (1985). Boundary integral equations for
mixed boundary value problems in polygonal domains and Galerkin ap
proximation. Mathematical Models and Methods in Mechanics Vol. 15. W.
Fiszdon and K. Wilmanski eds., Banach Center Publications.
[36] Costabel, M., Dauge, M. and Nicaise, S. eds. (1995). Boundary Value Prob
lems and Integral Equations in Nonsmooth Domains. Number 167 in Lec
ture Notes in Pure and Applied Mathematics. Marcel Dekker Publications.
© 2005 by Taylor & Francis Group, LLC
206 BIBLIOGRAPHY
[37] Cruse, T.A. (1969). Numerical solutions in threedimensional elastostatics.
International Journal of Solids and Structures 5:12591274.
[38] Cruse, T.A. (1988). Boundary Element Analysis in Computational Fracture
Mechanics. Kluwer, Dordrecht, The Netherlands.
[39] Cruse, T.A. and Richardson, J.D. (1996). Nonsingular Somigliana stress
identities in elasticity. International Journal for Numerical Methods in En
gineering 39:32733304.
[40] De, S. and Bathe, K.J. (2000). The method of ﬁnite spheres. Computational
Mechanics 25:329345.
[41] Dolbow, J., M¨ oes, N. and Belytschko, T. (2000). Modeling fracture in
MindlinReissner plates with the extended ﬁnite element method. Interna
tional Journal of Solids and Structures 37:71617183.
[42] Duarte, C.A.M. and Oden, J.T. (1996). Hp clouds  an hp meshless
method. Numerical Methods in Partial Diﬀerential Equations 12:673705.
[43] Duarte, C.A.M. and Oden, J.T. (1996). An hp adaptive method using
clouds. Computer Methods in Applied Mechanics and Engineering 139:237
262.
[44] Eriksson, K., Estep, D., Hansbo, P. and Johnson, C. (1995). Introduc
tion to adaptive methods for partial diﬀerential equations. Acta Numerica
Chapter 3. A. Iserles ed., 105158, Cambridge University Press, UK.
[45] Feistaur, M., Hsiao, G.C. and Kleinman, R.E. (1996). Asymptotic and a
posteriori error estimates for boundary element solutions of hypersingular
integral equations. SIAM Journal of Numerical Analysis 32:666685.
[46] Fung, YC. (1965). Foundations of Solid Mechanics. Prentice Hall, Engle
wood Cliﬀs, NJ.
[47] Gaul, L., K¨ ogl, M. and Wagner, M. (2003). Boundary Element Methods
for Engineers and Scientists. Springer Verlag, Heidelberg, Germany.
[48] Ghosh, N. and Mukherjee, S. (1987). A new boundary element method
formulation for three dimensional problems in linear elasticity. Acta Me
chanica 67:107119.
[49] Gilbert, J.R., Legtenberg, R. and Senturia, S.D. (1995). 3D coupled elec
tromechanics for MEMS  applications of CoSolve EM. Proceedings of IEEE
MEMS pp.122127.
[50] Glushkov, E., Glushkova, N. and Lapina, O. (1999). 3D elastic stress
analysis at polyhedral corner points. International Journal of Solids and
Structures 36:11051128.
© 2005 by Taylor & Francis Group, LLC
BIBLIOGRAPHY 207
[51] Goldberg, M.A. and Bowman, H. (1998). Superconvergence and the use of
the residual as an error estimator in the BEM  I. Theoretical Development.
Boundary Element Communications 8:230238.
[52] Gowrishankar, R. and Mukherjee, S. (2002). A “pure” boundary node
method in potential theory. Communications in Numerical Methods in En
gineering 18:411427.
[53] Gray, L.J. (1989). Boundary element method for regions with thin internal
cavities. Engineering Analysis with Boundary Elements 6:180184.
[54] Gray, L.J., Martha, L.F. and Ingraﬀea, A.R. (1990). Hypersingular in
tegrals in boundary element fracture analysis. International Journal for
Numerical Methods in Engineering 29:11351158.
[55] Gray, L.J., Balakrishna, C. and Kane, J.H. (1995). Symmetric Galerkin
fracture analysis. Engineering Analysis with Boundary Elements 15:103
109.
[56] Gray, L.J. and Paulino, G.H. (1997). Symmetric Galerkin boundary inte
gral formulation for interface and multizone problems. International Jour
nal for Numerical Methods in Engineering 40:30853101.
[57] Gray, L.J. and Paulino, G.H. (1997). Symmetric Galerkin boundary inte
gral fracture analysis for plane orthotropic elasticity. Computational Me
chanics 20:2633.
[58] Gray, L.J. and Paulino, G.H. (1998). Crack tip interpolation revisited.
SIAM Journal of Applied Mathematics 58:428455.
[59] Guidera, J.T. and Lardner, R.W. (1975). Penny shaped cracks. Journal of
Elasticity 5:5973.
[60] Guiggiani, M. (1994). Hypersingular formulation for boundary stress eval
uation. Engineering Analysis with Boundary Elements 13:169179.
[61] Harrington, R.F. (1993). Field Computation by Moment Methods. IEEE
Press, Piscataway, NJ.
[62] Hartmann, F. (1989). Introduction to Boundary Elements : Theory and
Applications. Springer Verlag, Berlin, New York.
[63] Haug, E.J., Choi, K.K. and Komkov, V. (1986). Design Sensitivity Analysis
of Structural Systems. Academic Press, New York.
[64] Hayami, K. and Matsumoto, H. (1994). A numerical quadrature for nearly
singular boundary element integrals. Engineering Analysis with Boundary
Elements 13:143154.
[65] Heck, A. (1993). Introduction to Maple. Springer Verlag, New York.
© 2005 by Taylor & Francis Group, LLC
208 BIBLIOGRAPHY
[66] Hughes, T.J.R. (2000). The Finite Element Method  Linear Static and
Dynamic Finite Element Analysis. Dover, Mineola, NY.
[67] Kaljevic, I. and Saigal, S. (1997). An improved element free Galerkin
formulation. International Journal for Numerical Methods in Engineering
40:29532974.
[68] Kane, J.H. (1994). Boundary Element Analysis in Engineering Continuum
Mechanics. Prentice Hall, Englewood Cliﬀs, NJ.
[69] Kaplan, W. (1984). Advanced Calculus 3rd. ed. AddisonWesley, Reading,
MA.
[70] Kita, E. and Kamiya, N. (1994). Recent studies on adaptive boundary
element methods. Advances in Engineering Software 19:2132.
[71] Ko, S.C., Kim, Y.C., Lee, S.S., Choi, S.H. and Kim, S.R. (2003). Micro
machined piezoelectric membrane acoustic device. Sensors and Actuators
A 103:130134.
[72] Kothnur, V., Mukherjee, S. and Mukherjee, Y.X. (1999). Twodimensional
linear elasticity by the boundary node method. International Journal of
Solids and Structures 36:11291147.
[73] Kress, R. (1989). Linear Integral Equations. Springer Verlag.
[74] Krishnamoorthy, C.S. and Umesh, K.R. (1993). Adaptive mesh reﬁnement
for twodimensional ﬁnite element stress analysis. Computers and Struc
tures 48:121133.
[75] Krishnasamy, G., Schmerr, L.W., Rudolphi, T.J. and Rizzo, F.J. (1990).
Hypersingular boundary integral equations : some applications in acoustic
and elastic wave scattering. ASME Journal of Applied Mechanics 57:404
414.
[76] Krishnasamy, G., Rizzo, F.J. and Rudolphi, T.J. (1992). Hypersingular
boundary integral equations: their occurrence, interpretation, regulariza
tion and computation. Developments in Boundary Element Methods7.
P.K. Banerjee and S. Kobayashi eds., Elsevier Applied Science, London,
pp. 207252.
[77] Kulkarni, S.S., Telukunta, S. and Mukherjee, S. (2003). Application of an
accelerated boundarybased meshfree method to twodimensional prob
lems in potential theory. Computational Mechanics 32:240249.
[78] Lancaster, P. and Salkauskas, K. (1990). Curve and Surface Fitting  An
Introduction. Academic Press, London.
© 2005 by Taylor & Francis Group, LLC
BIBLIOGRAPHY 209
[79] Li, G. and Aluru, N.R. (2002). Boundary cloud method: a combined scat
tered point/boundary integral approach for boundary only analysis. Com
puter Methods in Applied Mechanics and Engineering 191:23372370.
[80] Li, G. and Aluru, N.R. (2003). A boundary cloud method with a cloud
bycloud polynomial basis. Engineering Analysis with Boundary Elements
27:5771.
[81] Liapis, S. (1995). A review of error estimation and adaptivity in the
boundary element method. Engineering Analysis with Boundary Elements
14:315323.
[82] Liu, G.R. (2002). Mesh Free Methods  Moving beyond the Finite Element
Method. CRC Press, Boca Raton, FL.
[83] Liu, W.K., Chen, Y., Uras, R.A. and Chang, C.T. (1996). Generalized
multiple scale reproducing kernel particle methods. Computer Methods in
Applied Mechanics and Engineering 139:91157.
[84] Liu, W.K., Chen, Y., Jun, S., Belytschko, T., Pan, C., Uras, R.A. and
Chang, C.T. (1996). Overview and applications of the reproducing kernel
particle methods. Archives of Computational Methods in Engineering 3,
380.
[85] Liu, Y.J. and Rizzo, F.J. (1993). Hypersingular boundary integral equa
tions for radiation and scattering of elastic waves in three dimensions.
Computer Methods in Applied Mechanics and Engineering 107:131144.
[86] Liu, Y.J., Zhang, D. and Rizzo, F.J. (1993). Nearly singular and hyper
singular integrals in the boundary element method. Boundary Elements
XV. C.A. Brebbia and J.J. Rencis eds., 453468, Computational Mechan
ics Publications, Southampton and Elsevier, Barking, Essex, UK.
[87] Liu, Y.J. (1998). Analysis of shelllike structures by the boundary element
method based on 3D elasticity : formulation and veriﬁcation. International
Journal for Numerical Methods in Engineering 41:541558.
[88] Lutz, E.D. (1991). Numerical Methods for Hypersingular and NearSingular
Boundary Integrals in Fracture Mechanics. Ph.D. Dissertation, Cornell
University, Ithaca, NY.
[89] Lutz, E.D., Ingraﬀea, A.R. and Gray, L.J. (1992). Use of ‘simple solutions’
for boundary integral methods in elasticity and fracture analysis. Interna
tional Journal for Numerical Methods in Engineering 35:17371751.
[90] Lutz, E., Ye, W. and Mukherjee, S. (1998). Elimination of rigid body
modes from discretized boundary integral equations. International Journal
of Solids and Structures 35:44274436.
© 2005 by Taylor & Francis Group, LLC
210 BIBLIOGRAPHY
[91] Mackerele, J. (1993). Mesh generation and reﬁnement for FEM and BEM
 a bibliography (19901993). Finite Elements in Analysis and Design
15:177188.
[92] Martin, P.A. and Rizzo, F.J. (1996). Hypersingular integrals : how smooth
must the density be ? International Journal for Numerical methods in
Engineering 39:687704.
[93] Martin, P.A., Rizzo, F.J. and Cruse, T.A. (1998). Smoothnessrelaxation
strategies for singular and hypersingular integral equations. International
Journal for Numerical Methods in Engineering 42:885906.
[94] Manti˘c, V. (1993). A new formula for the Cmatrix in the Somigliana
identity. Journal of Elasticity 33:191201.
[95] Menon, G. (1996). Hypersingular Error Estimates in Boundary Element
Methods. M.S. Thesis, Cornell University, Ithaca, NY.
[96] Menon, G., Paulino, G.H. and Mukherjee, S. (1999). Analysis of hypersin
gular residual error estimates in boundary element methods for potential
problems. Computational Methods in Applied Mechanics and Engineering
173:449473.
[97] Mo¨es, N., Dolbow, J. and Belytschko, T. (1999). A ﬁnite element method
for crack growth without remeshing. International Journal for Numerical
Methods in Engineering 46:131150.
[98] Mukherjee, S. (1982). Boundary Element Methods in Creep and Fracture.
Elsevier, London.
[99] Mukherjee, S. and Mukherjee, Y. X. (1998). The hypersingular boundary
contour method for threedimensional linear elasticity. ASME Journal of
Applied Mechanics 65:300309.
[100] Mukherjee, S., Shi, X. and Mukherjee, Y.X. (1999). Surface variables and
their sensitivities in threedimensional linear elasticity by the boundary
contour method. Computer Methods in Applied Mechanics and Engineering
173:387402.
[101] Mukherjee, S. (2000). CPV and HFP integrals and their applications in
the boundary element method. International Journal of Solids and Struc
tures 37:66236634.
[102] Mukherjee, S. (2000). Finite parts of singular and hypersingular integrals
with irregular boundary source points. Engineering Analysis with Bound
ary Elements 24:767776.
[103] Mukherjee, S., Shi, X. and Mukherjee, Y.X. (2000). Internal variables
and their sensitivities in threedimensional linear elasticity by the boundary
© 2005 by Taylor & Francis Group, LLC
BIBLIOGRAPHY 211
contour method. Computer Methods in Applied Mechanics and Engineering
187:289306.
[104] Mukherjee, S., Chati, M.K. and Shi, X. (2000). Evaluation of nearly sin
gular integrals in boundary element contour and node methods for three
dimensional linear elasticity. International Journal of Solids and Structures
37:76337654.
[105] Mukherjee, S. (2001). On boundary integral equations for cracked and for
thin bodies. Mathematics and Mechanics of Solids 6:4764.
[106] Mukherjee, S. (2002). Regularization of hypersingular boundary integral
equations : a new approach for axisymmetric elasticity. Engineering Anal
ysis with Boundary Elements 26:839844.
[107] Mukherjee, Y.X. and Mukherjee, S. (1997). The boundary node method
for potential problems. International Journal for Numerical Methods in
Engineering 40:797815.
[108] Mukherjee, Y.X. and Mukherjee, S. (1997). On boundary conditions in
the elementfree Galerkin method. Computational Mechanics 19:264270.
[109] Mukherjee, Y.X., Mukherjee, S., Shi, X. and Nagarajan, A. (1997). The
boundary contour method for threedimensional linear elasticity with a
new quadratic boundary element. Engineering Analysis with Boundary El
ements 20:3544.
[110] Mukherjee, Y.X., Shah, K. and Mukherjee, S. (1999). Thermoelastic frac
ture mechanics with regularized hypersingular boundary integral equa
tions. Engineering Analysis with Boundary Elements 23:8996.
[111] Mukherjee, Y.X. and Mukherjee, S. (2001). Error analysis and adaptiv
ity in threedimensional linear elasticity by the usual and hypersingular
boundary contour method. International Journal of Solids and Structures
38:161178.
[112] Nabors, K. and White, J. (1991). FastCap: a multipole accelerated 3D
capacitance extraction program. IEEE Transactions on ComputerAided
Design of Integrated Circuits and Systems 10:14471459.
[113] Nabors, K., Kim, S., White, J. and Senturia, S. (1992). FastCap User’s
Guide. MIT, Cambridge, MA.
[114] Nagarajan, A. and Mukherjee, S. (1993). A mapping method for numeri
cal evaluation of twodimensional integrals with 1/r singularity. Computa
tional Mechanics 12:1926.
[115] Nagarajan, A., Lutz, E.D. and Mukherjee, S. (1994). A novel boundary
element method for linear elasticity with no numerical integration for 2D
© 2005 by Taylor & Francis Group, LLC
212 BIBLIOGRAPHY
and line integrals for 3D problems. ASME Journal of Applied Mechanics
61:264269.
[116] Nagarajan, A., Mukherjee, S. and Lutz, E.D. (1996). The boundary con
tour method for threedimensional linear elasticity. ASME Journal of Ap
plied Mechanics 63:278286.
[117] Nayroles, B., Touzot, G. and Villon, P. (1992). Generalizing the ﬁnite ele
ment method: diﬀuse approximation and diﬀuse elements. Computational
Mechanics 10:307318.
[118] Nishimura, N., Yoshida, K. and Kobayashi, S. (1999). A fast multipole
boundary integral equation method for crack problems in 3D. Engineering
Analysis with Boundary Elements 23:97105.
[119] Novati, G. and Springhetti, R. (1999). A Galerkin boundary contour
method for twodimensional linear elasticity. Computational Mechanics
23:5362.
[120] Oden, J.T., Duarte, C.A.M. and Zienkiewicz, O.C. (1998). A new cloud
based hp ﬁnite element method. Computer Methods in Applied Mechanics
and Engineering 153:117126.
[121] Par´ıs, F. and Ca˜ nas, J. (1997). Boundary Element Method : Fundamentals
and Applications. Oxford University Press, UK.
[122] Paulino, G.H. (1995). Novel Formulations of the Boundary Element
Method for Fracture Mechanics and Error Estimation. Ph.D. Dissertation,
Cornell University, Ithaca, NY.
[123] Paulino, G.H., Gray, L.J. and Zarkian, V. (1996). Hypersingular residuals
 a new approach for error estimation in the boundary element method.
International Journal for Numerical Methods in Engineering 39:20052029.
[124] Paulino, G.H., Shi, F., Mukherjee, S. and Ramesh, P. (1997). Nodal sen
sitivities as error estimates in computational mechanics. Acta Mechanica
121:191213.
[125] Paulino, G.H. and Gray, L.J. (1999). Galerkin residuals for adaptive
symmetricGalerkin boundary element methods. ASCE Journal of Engi
neering Mechanics 125:575585.
[126] Paulino, G.H., Menezes, I.F.M., Cavalcante Neto, J.B. and Martha, L.F.
(1999). A methodology for adaptive ﬁnite element analysis: towards an
integrated computational environment. Computational Mechanics 23:361
388.
[127] Paulino, G.H., Menon, G. and Mukherjee S. (2001). Error estimation
using hypersingular integrals in boundary element methods for linear elas
ticity. Engineering Analysis with Boundary Elements 25:523534.
© 2005 by Taylor & Francis Group, LLC
BIBLIOGRAPHY 213
[128] Petryk, H. and Mr´ oz, Z. (1986). Time derivatives of integrals and func
tionals deﬁned on varying volume and surface elements. Archives of Me
chanics 56:697724.
[129] Phan, A.V., Mukherjee, S. and Mayer, J.R.R. (1997). The boundary con
tour method for twodimensional linear elasticity with quadratic boundary
elements. Computational Mechanics 20:310319.
[130] Phan, A.V., Mukherjee, S. and Mayer, J.R.R. (1998). A boundary con
tour formulation for design sensitivity analysis in twodimensional linear
elasticity. International Journal of Solids and Structures 35:19811999.
[131] Phan, A.V., Mukherjee, S. and Mayer, J.R.R. (1998). The hypersingu
lar boundary contour method for twodimensional linear elasticity. Acta
Mechanica 130:209225.
[132] Phan, A.V., Mukherjee, S. and Mayer, J.R.R. (1998). Stresses, stress
sensitivities and shape optimization in twodimensional linear elasticity
by the boundary contour method. International Journal for Numerical
Methods in Engineering 42:13911407.
[133] Phan, A.V. and Mukherjee, S. (1999). On design sensitivity analysis in
linear elasticity by the boundary contour method. Engineering Analysis
with Boundary Elements 23:195199.
[134] Phan, A.V. and Phan, T.N. (1999). Structural shape optimization sys
tem using the twodimensional boundary contour method. Archive of Ap
plied Mechanics 69:481489.
[135] Phan, A.V. and Liu, Y.J. (2001). Boundary contour analysis of thin ﬁlms
and layered coatings. Sixth U.S. National Congress of Applied Mechanics,
Detroit, MI.
[136] Phan, A.V., Gray, L.J., Kaplan, T. and Phan, T.N. (2002). The bound
ary contour method for twodimensional Stokes ﬂow and incompressible
elastic materials. Computational Mechanics 28:425433.
[137] Pickett, G. (1944). Application of the Fourier method to the solution of
certain boundary value problems in the theory of elasticity. Transactions
of the ASME (Journal of Applied Mechanics) 66:A176A182.
[138] Poon, H., Mukherjee, S. and Ahmad, M.F. (1998). Use of simple solutions
in regularizing hypersingular boundary integral equations in elastoplastic
ity. ASME Journal of Applied Mechanics 65:3945.
[139] Radin, C. (1994). The pinwheel tilings of the plane. Annals of Mathemat
ics 139:661702.
© 2005 by Taylor & Francis Group, LLC
214 BIBLIOGRAPHY
[140] Rank, E. (1989). Adaptive h, p, and hp versions for boundary integral
element methods. International Journal for Numerical Methods in Engi
neering 28:13351349.
[141] Rizzo, F.J. (1967). An integral equation approach to boundary value prob
lems of classical elastostatics. Quarterly of Applied Mathematics 25:8395.
[142] Roman, M. and Aubry, N. (2003). Design and fabrication of electrically
actuated synthetic microjets. ASME Paper No. IMECE200341579. Amer
ican Society of Mechanical Engineers, New York.
[143] Rudolphi, T.J. (1991). The use of simple solutions in the regularization
of hypersingular boundary integral equations. Mathematical and Computer
Modeling 15:269278.
[144] Saigal S. and Kane, J.H. (1990). Boundaryelement shape optimization
system for aircraft structural components. AIAA Journal 28:12031204.
[145] Sandgren, E. and Wu, S.J. (1988). Shape optimization using the boundary
element method with substructuring. International Journal for Numerical
Methods in Engineering 26:19131924.
[146] Schittkowski, K. (1986). NLPQL: A FORTRAN subroutine solving con
strained nonlinear programming problems. Annals of Operations Research
5:485500.
[147] Senturia, S.D., Harris, R.M., Johnson, B.P., Kim, S., Nabors, K., Shul
man, M.A. and White, J.K. (1992). A computeraided design system for
microelectromechanical systems (MEMCAD). Journal of Microelectrome
chanical Systems 1:313.
[148] Shi, F., Ramesh, P. and Mukherjee, S. (1995). Simulation methods for
microelectromechanical structures (MEMS) with application to a mi
crotweezer. Computers and Structures 56:769783.
[149] Shi, F., Ramesh, P. and Mukherjee, S. (1996). Dynamic analysis of micro
electromechanical systems. International Journal for Numerical Methods
in Engineering 39:41194139.
[150] Shi, X. and Mukherjee, S. (1999). Shape optimization in three
dimensional linear elasticity by the boundary contour method. Engineering
Analysis with Boundary Elements 23:627637.
[151] Sladek, J. and Sladek, V. (1986). Computation of stresses by BEM in 2D
elastostatics. Acta Technica CSAV 31:523531.
[152] Sladek, J., Sladek, V. and Atluri, S.N. (2000). Local boundary integral
equation (LBIE) method for solving problems of elasticity with nonhomo
geneous material properties. Computational Mechanics 24:456462.
© 2005 by Taylor & Francis Group, LLC
BIBLIOGRAPHY 215
[153] Sloan, I.H. (1976). Improvement by iteration for compact operator equa
tions. Mathematics of Computation 30:758764.
[154] Sloan, I.H. (1990). Superconvergence. Numerical Solution of Integral
Equations Chapter 2. M. A. Goldberg ed., 3570 Plenum Press, New York.
[155] Sloan, I.H. (1992). Error analysis in boundary integral methods. Acta
Numerica Chapter 7. A. Iserles ed., 287339 Cambridge University Press,
UK.
[156] Stephan, E.P. (1996). The hp boundary element method for solving 2
and 3 dimensional problems. Computer Methods in Applied Mechanics
and Engineering 133:183208.
[157] Strouboulis, T., Babuˇska, I. and Copps, K. (2000). The design and analy
sis of the generalized ﬁnite element method. Computer Methods in Applied
Mechanics and Engineering 181:4369.
[158] Sukumar, N., Moran, B. and Belytschko, T. (1998). The natural ele
ment method. International Journal for Numerical Methods in Engineering
43:839887.
[159] Sukumar, N., M¨ oes, N., Moran, B. and Belytschko, T. (2000). Extended
ﬁnite element method for threedimensional crack modeling. International
Journal for Numerical Methods in Engineering 48:15491570.
[160] Sukumar, N., Moran, B., Semenov, A.Y. and Belytschko, T. (2001).
Natural neighbour Galerkin methods. International Journal for Numer
ical Methods in Engineering 50:127.
[161] Tafreshi, A. and Fenner, R.T. (1995). Generalpurpose computer program
for shape optimization of engineering structures using the boundary ele
ment method. Computers and Structures 56:713720.
[162] Tanaka, M., Sladek, V. and Sladek, J. (1994). Regularization techniques
applied to boundary element methods. ASME Applied Mechanics Reviews
47:457499.
[163] Telukunta, S. and Mukherjee, S. (2004). An extended boundary node
method for modeling normal derivative discontinuities in potential theory
across corners and edges. Engineering Analysis with Boundary Elements
28:10991110.
[164] Telukunta, S. and Mukherjee, S. (2004). Extended boundary node method
for threedimensional problems in potential theory. Computers and Struc
tures. Submitted.
[165] Telles, J.C.F. (1987). A selfadaptive coordinate transformation for eﬃ
cient numerical evaluation of general boundary element integrals. Interna
tional Journal for Numerical Methods in Engineering 24:959973.
© 2005 by Taylor & Francis Group, LLC
216 BIBLIOGRAPHY
[166] Telles, J.C.F. and Oliveria, R.F. (1994). Third degree polynomial trans
formation for boundary element integrals : Further improvements. Engi
neering Analysis with Boundary Elements 13:135141.
[167] Timoshenko, S.P. and Goodier, J.N. (1970). Theory of Elasticity 3rd. ed.
McGraw Hill, New York.
[168] Toh, K.C. and Mukherjee, S. (1994). Hypersingular and ﬁnite part in
tegrals in the boundary element method. International Journal of Solids
and Structures 31:22992312.
[169] Wei, X., Chandra, A., Leu, L.J. and Mukherjee, S. (1994). Shape opti
mization in elasticity and elastoviscoplasticity by the boundary element
method. International Journal of Solids and Structures 31:533550.
[170] Wendland, W.L., Stephan, E.P. and Hsiao, G.C. (1979). On the inte
gral equation method for the plane mixed boundary value problem for the
Laplacian. Mathematical Methods in the Applied Sciences 1:265321.
[171] Wendland, W.L. and Yu, DH. (1988). Adaptive boundary element meth
ods for strongly elliptic integral equations. Numerische Mathematik 53:539
558.
[172] Wendland, W.L. and Yu, DH. (1992). A posteriori local error estimates
of boundary element methods with some pseudodiﬀerential equations on
closed curves. Journal of Computational Mathematics 10:273289.
[173] Wilde, A.J. and Aliabadi, M.H. (1998). Direct evaluation of boundary
stresses in the 3D BEM of elastostatics. Communications in Numerical
Methods in Engineering 14:505517.
[174] Xu, Y. and Saigal, S. (1998). An elementfree Galerkin formulation for
stable crack growth in elastic solids. Computer Methods in Applied Me
chanics and Engineering 154:331343.
[175] Xu, Y. and Saigal, S. (1998). Elementfree Galerkin study of steady quasi
static crack growth in plane strain tension in elasticplastic materials. Com
putational Mechanics 21:276282.
[176] Xu, Y. and Saigal, S. (1999). An elementfree Galerkin analysis of steady
dynamic growth of a mode I crack in elasticplastic materials. International
Journal of Solids and Structures 36:10451079.
[177] Yamazaki, K., Sakamoto, J. and Kitano, M. (1994). Threedimensional
shape optimization using the boundary element method. AIAA Journal
32:12951301.
[178] Yang, R.J. (1990). Component shape optimization using BEM. Comput
ers and Structures 37:561568.
© 2005 by Taylor & Francis Group, LLC
BIBLIOGRAPHY 217
[179] Yang, T.Y. (1986). Finite Element Structural Analysis. Prentice Hall,
Englewood Cliﬀs, NJ.
[180] Yoshida, K., Nishimura, N. and Kobayashi, S. (2001). Application of a
new fast multipole boundary integral equation method to crack problems
in 3D. Engineering Analysis with Boundary Elements 25:239247.
[181] Yu, D.H. (1987). A posteriori error estimates and adaptive approaches
for some boundary element methods. Mathematical and Computational As
pects Vol. 1 of Boundary Elements IX. C. Brebbia, W. L. Wendland and G.
Kuhn eds., 241256. Computational Mechanics Publications and Springer
Verlag.
[182] Yu, D.H. (1988). Self adaptive boundary element methods. Zeitschrift
f¨ ur Angewandte Mathematik und Mechanik 68:T435T437.
[183] Zhang, Q. and Mukherjee, S. (1991). Design sensitivity coeﬃcients for
linear elastic bodies with zones and corners by the derivative boundary
element method. International Journal of Solids and Structures 27:983
998.
[184] Zhao, Z. (1991). Shape Design Sensitivity Analysis and Optimization Us
ing the Boundary Element Method. Springer, New York.
[185] Zhao, Z.Y. and Lan, S.R. (1999). Boundary stress calculation  a compar
ison study. Computers and Structures 71:7785.
[186] Zhou, S.J., Sun, S.X. and Cao, Z.Y. (1998). The dual boundary con
tour method for twodimensional crack problems. International Journal of
Fracture 92:201212.
[187] Zhou, S.J., Cao, Z.Y. and Sun, S.X. (1999). The traction boundary
contour method for linear elasticity. International Journal for Numerical
Methods in Engineering 46:18831895.
[188] Zhu T., Zhang, JD. and Atluri, S.N. (1998). A local boundary integral
equation (LBIE) method in computational mechanics, and a meshless dis
cretization approach. Computational Mechanics 21:223235.
[189] Zhu, T. (1999). A new meshless regular local boundary integral equa
tion (MRLBIE) method. International Journal for Numerical Methods in
Engineering 46:12371252.
[190] Zienkiewicz, O.C. and Taylor, R.L. (1994). The Finite Element Method.
Vols. 1 and 2. McGraw Hill, Maidenhead, Berkshire, UK.
© 2005 by Taylor & Francis Group, LLC
MECHANICAL ENGINEERING A Series of Textbooks and Reference Books Founding Editor L. L. Faulkner Columbus Division, Battelle Memorial Institute and Department of Mechanical Engineering The Ohio State University Columbus, Ohio
1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23.
Spring Designer’s Handbook, Harold Carlson ComputerAided Graphics and Design, Daniel L. Ryan Lubrication Fundamentals, J. George Wills Solar Engineering for Domestic Buildings, William A. Himmelman Applied Engineering Mechanics: Statics and Dynamics, G. Boothroyd and C. Poli Centrifugal Pump Clinic, Igor J. Karassik ComputerAided Kinetics for Machine Design, Daniel L. Ryan Plastics Products Design Handbook, Part A: Materials and Components; Part B: Processes and Design for Processes, edited by Edward Miller Turbomachinery: Basic Theory and Applications, Earl Logan, Jr. Vibrations of Shells and Plates, Werner Soedel Flat and Corrugated Diaphragm Design Handbook, Mario Di Giovanni Practical Stress Analysis in Engineering Design, Alexander Blake An Introduction to the Design and Behavior of Bolted Joints, John H. Bickford Optimal Engineering Design: Principles and Applications, James N. Siddall Spring Manufacturing Handbook, Harold Carlson Industrial Noise Control: Fundamentals and Applications, edited by Lewis H. Bell Gears and Their Vibration: A Basic Approach to Understanding Gear Noise, J. Derek Smith Chains for Power Transmission and Material Handling: Design and Applications Handbook, American Chain Association Corrosion and Corrosion Protection Handbook, edited by Philip A. Schweitzer Gear Drive Systems: Design and Application, Peter Lynwander Controlling InPlant Airborne Contaminants: Systems Design and Calculations, John D. Constance CAD/CAM Systems Planning and Implementation, Charles S. Knox Probabilistic Engineering Design: Principles and Applications, James N. Siddall
© 2005 by Taylor & Francis Group, LLC
24. 25. 26. 27. 28. 29. 30. 31. 32. 33. 34. 35. 36. 37. 38. 39. 40. 41. 42. 43. 44. 45. 46. 47. 48. 49. 50. 51. 52. 53. 54.
Traction Drives: Selection and Application, Frederick W. Heilich III and Eugene E. Shube Finite Element Methods: An Introduction, Ronald L. Huston and Chris E. Passerello Mechanical Fastening of Plastics: An Engineering Handbook, Brayton Lincoln, Kenneth J. Gomes, and James F. Braden Lubrication in Practice: Second Edition, edited by W. S. Robertson Principles of Automated Drafting, Daniel L. Ryan Practical Seal Design, edited by Leonard J. Martini Engineering Documentation for CAD/CAM Applications, Charles S. Knox Design Dimensioning with Computer Graphics Applications, Jerome C. Lange Mechanism Analysis: Simplified Graphical and Analytical Techniques, Lyndon O. Barton CAD/CAM Systems: Justification, Implementation, Productivity Measurement, Edward J. Preston, George W. Crawford, and Mark E. Coticchia Steam Plant Calculations Manual, V. Ganapathy Design Assurance for Engineers and Managers, John A. Burgess Heat Transfer Fluids and Systems for Process and Energy Applications, Jasbir Singh Potential Flows: Computer Graphic Solutions, Robert H. Kirchhoff ComputerAided Graphics and Design: Second Edition, Daniel L. Ryan Electronically Controlled Proportional Valves: Selection and Application, Michael J. Tonyan, edited by Tobi Goldoftas Pressure Gauge Handbook, AMETEK, U.S. Gauge Division, edited by Philip W. Harland Fabric Filtration for Combustion Sources: Fundamentals and Basic Technology, R. P. Donovan Design of Mechanical Joints, Alexander Blake CAD/CAM Dictionary, Edward J. Preston, George W. Crawford, and Mark E. Coticchia Machinery Adhesives for Locking, Retaining, and Sealing, Girard S. Haviland Couplings and Joints: Design, Selection, and Application, Jon R. Mancuso Shaft Alignment Handbook, John Piotrowski BASIC Programs for Steam Plant Engineers: Boilers, Combustion, Fluid Flow, and Heat Transfer, V. Ganapathy Solving Mechanical Design Problems with Computer Graphics, Jerome C. Lange Plastics Gearing: Selection and Application, Clifford E. Adams Clutches and Brakes: Design and Selection, William C. Orthwein Transducers in Mechanical and Electronic Design, Harry L. Trietley Metallurgical Applications of ShockWave and HighStrainRate Phenomena, edited by Lawrence E. Murr, Karl P. Staudhammer, and Marc A. Meyers Magnesium Products Design, Robert S. Busk How to Integrate CAD/CAM Systems: Management and Technology, William D. Engelke
© 2005 by Taylor & Francis Group, LLC
55. 56. 57. 58. 59. 60. 61. 62. 63. 64. 65. 66. 67. 68. 69. 70. 71. 72. 73. 74. 75. 76. 77. 78. 79. 80. 81. 82. 83. 84.
Cam Design and Manufacture: Second Edition; with cam design software for the IBM PC and compatibles, disk included, Preben W. Jensen SolidState AC Motor Controls: Selection and Application, Sylvester Campbell Fundamentals of Robotics, David D. Ardayfio Belt Selection and Application for Engineers, edited by Wallace D. Erickson Developing ThreeDimensional CAD Software with the IBM PC, C. Stan Wei Organizing Data for CIM Applications, Charles S. Knox, with contributions by Thomas C. Boos, Ross S. Culverhouse, and Paul F. Muchnicki ComputerAided Simulation in Railway Dynamics, by Rao V. Dukkipati and Joseph R. Amyot FiberReinforced Composites: Materials, Manufacturing, and Design, P. K. Mallick Photoelectric Sensors and Controls: Selection and Application, Scott M. Juds Finite Element Analysis with Personal Computers, Edward R. Champion, Jr. and J. Michael Ensminger Ultrasonics: Fundamentals, Technology, Applications: Second Edition, Revised and Expanded, Dale Ensminger Applied Finite Element Modeling: Practical Problem Solving for Engineers, Jeffrey M. Steele Measurement and Instrumentation in Engineering: Principles and Basic Laboratory Experiments, Francis S. Tse and Ivan E. Morse Centrifugal Pump Clinic: Second Edition, Revised and Expanded, Igor J. Karassik Practical Stress Analysis in Engineering Design: Second Edition, Revised and Expanded, Alexander Blake An Introduction to the Design and Behavior of Bolted Joints: Second Edition, Revised and Expanded, John H. Bickford High Vacuum Technology: A Practical Guide, Marsbed H. Hablanian Pressure Sensors: Selection and Application, Duane Tandeske Zinc Handbook: Properties, Processing, and Use in Design, Frank Porter Thermal Fatigue of Metals, Andrzej Weronski and Tadeusz Hejwowski Classical and Modern Mechanisms for Engineers and Inventors, Preben W. Jensen Handbook of Electronic Package Design, edited by Michael Pecht ShockWave and HighStrainRate Phenomena in Materials, edited by Marc A. Meyers, Lawrence E. Murr, and Karl P. Staudhammer Industrial Refrigeration: Principles, Design and Applications, P. C. Koelet Applied Combustion, Eugene L. Keating Engine Oils and Automotive Lubrication, edited by Wilfried J. Bartz Mechanism Analysis: Simplified and Graphical Techniques, Second Edition, Revised and Expanded, Lyndon O. Barton Fundamental Fluid Mechanics for the Practicing Engineer, James W. Murdock FiberReinforced Composites: Materials, Manufacturing, and Design, Second Edition, Revised and Expanded, P. K. Mallick Numerical Methods for Engineering Applications, Edward R. Champion, Jr.
© 2005 by Taylor & Francis Group, LLC
Revised and Expanded. 92. 113. 102.85. 98. 104. Blau Introduction to Plastics and Composites: Mechanical Properties and Engineering Applications. V. Ganapathy Industrial Noise Control: Fundamentals and Applications. Nielsen and Robert F. 97. 112. 96. Bell and Douglas H. Joining. edited by G. LLC . 91. 100. edited by Suong Van Hoa Friction Science and Technology. Revised and Expanded. Manufacturing. Monahan Refractory Linings Thermomechanical Design and Applications. Bell Finite Elements: Their Design and Performance. Revised and Expanded. Murray © 2005 by Taylor & Francis Group. Marsbed H. Revised and Expanded. Meadows Handbook of Materials Selection for Engineering Applications. Revised and Expanded. edited by Chunill Hah HighVacuum Technology: A Practical Guide. Second Edition. John Piotrowski ComputerAided Design of PolymerMatrix Composite Structures. 95. 107. Second Edition. T. Werner Soedel Steam Plant Calculations Manual: Second Edition. 90. edited by David W. Bayer Mechanical Power Transmission Components. Vladimír Stejskal and Michael Valásek Shaft Seals for Dynamic Applications. 106. and Assembly. James A. 110. Stewart Optimizing the Shape of Mechanical Elements and Structures. Lewis H. 108. edited by Earl Logan. 105. 94. Jr. MacNeal Mechanical Properties of Polymers and Composites: Second Edition. Volk Optical Principles and Technology for Engineers. Alexander Blake Pump Characteristics and Applications. 87. John H. Engineering Documentation Control Practices and Procedures. edited by Thomas A. A. A. and Inspection. Second Edition. Charles A. Mancuso Handbook of Turbomachinery. 111. Les Horve ReliabilityBased Mechanical Design. 86. Richard H. South and Jon R. Lawrence E. Edward Miller Practical Fracture Mechanics in Design. Peter J. Speck Turbomachinery Fluid Dynamics and Heat Transfer. Landel Mechanical Wear Prediction and Prevention. James E. Schacht Geometric Dimensioning and Tolerancing: Applications and Techniques for Use in Design. Vibrations of Shells and Plates: Second Edition. Cruse Mechanical Fastening. Raymond G. Earl Logan. Ray E. Turbomachinery: Basic Theory and Applications. 88. 93. Revised and Expanded. Jr. Bickford Shaft Alignment Handbook: Second Edition. James D. 99. Meadows An Introduction to the Design and Behavior of Bolted Joints: Third Edition. 103. 109. 101. James D. Michael W. Revised and Expanded. 89. Seireg and Jorge Rodriguez Kinematics and Dynamics of Machinery. Hablanian Geometric Dimensioning and Tolerancing: Workbook and Answerbook. Revised and Expanded.
Franklin E. Marinescu. Ralph L. edited by John D. Steven M. Jr. Heinz K. Bloomer Handbook of Hydraulic Fluid Technology. Second Edition. Melnikov Mechanical Analysis of Electronic Packaging Systems. Seireg Influence Functions and Matrices. Lepi Applied Computational Fluid Dynamics. 138. 128. Handbook of Thermoplastic Piping System Design. Yuri A. Gear Noise and Vibration. Campbell and Andrew K. A. 123. Derek Smith Practical Fluid Mechanics for Engineering Applications. 140. Jr. Shull Diesel Engine Engineering: Thermodynamics. Practical Guide to Pressure Vessel Manufacturing.114. John H. 134. edited by Peter J. Garg Fluid Sealing Technology. Dynamics. Maurice L. Abdulla S. 141. 143. Design. Formulas for Dynamic Analysis. edited by Vijay K. 116. Adams. Totten Heat Exchanger Design Handbook. 121. Revised and Expanded. 139. A. Kuppan Designing for Product Sound Quality. S. Lumkes. 120. A. 135. 144. Jr. Wessol Mechanical Life Cycle Handbook: Good Environmental Design and Manufacturing. 137. 122. Andrei Makartchouk Handbook of Machine Tool Analysis. Fisher and Joy R. Revised and Expanded. Huston and C. Faulkner and Earl Logan. Richard H. 131. McKeown Couplings and Joints: Design. 117. Liu Handbook of Machinery Dynamics. Jr. Thomas Sixsmith and Reinhard Hanselka Practical Guide to Finite Elements: A Solid Mechanics Approach. Selection. Vandagriff Lubrication Fundamentals: Second Edition. and Dan Boboc © 2005 by Taylor & Francis Group. Jon R. 124. Q. 119. 115. Jardine Practical Guide to Industrial Boiler Systems. Ioan D. Kenneth G. John J. Techniques. and Application. and Applications. 142. Earl Logan. Nau Friction and Lubrication in Mechanical Design. D. and Control. Fisher Nickel Alloys. Muller and Bernard S. 127. T. 126. 136. Rapid Prototyping Technology: Selection and Application. 132. Mancuso Thermodynamics: Processes and Applications. 133. edited by Mahendra S. Pirro and A. edited by Ulrich Heubner Rotating Machinery Vibration: Problem Analysis and Troubleshooting. Hundal Micromachining of Engineering Materials. edited by George E. Cooper Reciprocating Machinery Dynamics: Design and Analysis. J. Constantin Ispas. 118. Rangwala Maintenance Excellence: Optimizing Equipment LifeCycle Decisions. 129. Lyon Probability Applications in Mechanical Design. 125. Stephen A. LLC . Ronald L. 130. Lynn L. edited by Joseph McGeough Control Strategies for Dynamic Systems: Design and Implementation. M. Sunil Pullarcot Nondestructive Evaluation: Theory.
Second Edition. Applications. Derek Smith Handbook of Turbomachinery: Second Edition. and Kiyoshi Funatani Handbook of Pneumatic Conveying Engineering. Agarwal Clutches and Brakes: Design and Selection. Bradford Clifton. Stanford III Gear Noise and Vibration: Second Edition. Antaki Turbomachinery: Design and Theory. Schoonmaker Industrial Noise Control and Acoustics. M. R. David Mills. edited by Karl J. Henry M. Wolé Soboyejo Reliability Verification. Susan Carlson Skalak Practical Guide to the Packaging of Electronics: Thermal and Mechanical Design and Analysis. 155. Integrity. Testing. 152. Revised and Expanded. and Analysis in Engineering Design. LLC . 171. 157. 167. Gary S. Orthwein Fundamentals of Fluid Film Lubrication: Second Edition. and Calculations. and Operation. Schmid. 165. and Repair. 169. Mark G. Avraham Harnoy Mechanical Reliability Improvement: Probability and Statistics for Experimental Testing. Bernard J. and Wesley P. Construction. R. Revised and Expanded. Maintenance. Bird. 170. edited by George E. 156. Randall F. Little Industrial Boilers and Heat Recovery Steam Generators: Design. 151. Currie Intermediate Heat Transfer. 166. Jacobson Handbook of LeadFree Solder Technology for Microelectronic Assemblies. B. George A. Rama S. Steven R. Jones. 161. edited by George E. and Ramendra Roy Piping and Pipeline Engineering: Design. I. 158. 159. Tuttle Modeling and Simulation for Material Selection and Mechanical Design. 150. and Bo O. edited by Earl Logan. Albano. Implementing Concurrent Engineering in Small Companies. Townsend Fluidized Bed Combustion. Ganapathy The CAD Guidebook: A Basic Manual for Understanding and Improving ComputerAided Design. Totten. 162. J. V. 147. 154. Application. Stalter Vehicle Stability. Barron Mechanical Properties of Engineered Materials. Vijay Srinivasan Handbook of Mechanical Alloy Design. Dean Karnopp © 2005 by Taylor & Francis Group. Robert E. Simeon N. KauFui Vincent Wong HVAC Water Chillers and Cooling Towers: Fundamentals.145. Wasserman Fundamental Mechanics of Fluids: Third Edition. 163. William C. 153. and Kiyoshi Funatani Structural Analysis of Polymeric Composite Materials. Jr. Lin Xie. 148. Herbert W. E. Oka Theory of Dimensioning: An Introduction to Parameterizing Geometric Models. Stephen J. G. Ali Jamnia Bearing Design in Machinery: Engineering Tribology and Lubrication. Totten. Gorla and Aijaz Ahmed Khan Target Costing: MarketDriven Product Design. 146. 160. 164. Mark E. Hamrock. and Vijay K. 149. 168. Puttlitz and Kathleen A. Lin Xie.
174. 179. Revised and Expanded. Lawrence W. Raymond G. Second Edition. 187. Revised and Expanded. 181. Paul D. Contours. and Control. 184. Revised and Expanded. 185. 173. Bayer Liquid Pipeline Hydraulics. Subrata Mukherjee and Yu Xie Mukherjee © 2005 by Taylor & Francis Group. Shashi Menon Solid Fuels Combustion and Gasification. Marcio L.172. Raymond G. and Nodes. Paris Selection of Engineering Materials and Adhesives. Bayer Vibrations of Shells and Plates: Third Edition. LLC . 175. Mechanical Wear Fundamentals and Testing: Second Edition. 178. E. Haym Benaroya Design of Automatic Machinery. Abbas Mechanical Alloying and Milling. 182. Fisher Boundary Methods: Elements. de SouzaSantos Mechanical Tolerance Stackup and Analysis. Schacht Practical Engineering Failure Analysis. Derby Practical Fracture Mechanics in Design: Second Edition. Fischer Engineering Design for Wear. Revised and Expanded. Funkenbusch Gigacycle Fatigue in Mechanical Practive. Suryanarayana Mechanical Vibration: Analysis. Bryan R. 180. 176. Arun Shukla Practical Guide to Designed Experiments. Claude Bathias and Paul C. 186. and Nureddin M. Stephen J. C. 183. Uncertainties. edited by Charles A. 177. Tawancy. Werner Soedel Refractories Handbook. Hani M. Anwar UlHamid.
part of the Taylor & Francis imprint. LLC . Yu Xie Mukherjee Cornell University Ithaca. New York.A.Boundary Methods Elements. U.S.S. the academic division of T&F Informa plc. U. New York. a member of the Taylor & Francis Group. Contours. © 2005 by Taylor & Francis Group.A. Boca Raton London New York Singapore A CRC title. and Nodes Subrata Mukherjee Cornell University Ithaca.
without written permission from the publishers. Reprinted material is quoted with permission. Title. Boundary methods : elements. but the author and the publisher cannot assume responsibility for the validity of all materials or for the consequences of their use. Trademark Notice: Product or corporate names may be trademarks or registered trademarks.copyright. or in any information storage or retrieval system.Published in 2005 by CRC Press Taylor & Francis Group 6000 Broken Sound Parkway NW Boca Raton. LLC CRC Press is an imprint of Taylor & Francis Group No claim to original U.taylorandfrancis. Boundary element methods. II.S. TA347. No part of this book may be reprinted. paper) 1. Subrata. A wide variety of references are listed. For permission to photocopy or use material electronically from this work. Inc. Inc. MA 01923.com © 2005 by Taylor & Francis Group. please access www. Danvers. mechanical. including photocopying.com Taylor & Francis Group is the Academic Division of T&F Informa plc. and are used only for identification and explanation without intent to infringe.B69M83 2005 621'. Mechanical engineering (Marcel Dekker.com (http://www.copyright.) . cm.01'51535dc22 2004063489 Visit the Taylor & Francis Web site at http://www. Yu. and sources are indicated. . (CCC) 222 Rosewood Drive. contours. Library of Congress CataloginginPublication Data Mukherjee.com/) or contact the Copyright Clearance Center. and recording. For organizations that have been granted a photocopy license by the CCC. 9787508400. a separate system of payment has been arranged. reproduced. Mukherjee. p. or other means.(Mechanical engineering . transmitted. 185) ISBN 0824725999 (alk. and nodes / Subrata Mukherjee and Yu Mukherjee. FL 334872742 © 2005 by Taylor & Francis Group. I.crcpress. Reasonable efforts have been made to publish reliable data and information. 185. or utilized in any form by any electronic. Government works Printed in the United States of America on acidfree paper 10 9 8 7 6 5 4 3 2 1 International Standard Book Number10: 0824725999 (Hardcover) International Standard Book Number13: 9780824725990 (Hardcover) Library of Congress Card Number 2004063489 This book contains information obtained from authentic and highly regarded sources. and the CRC Press Web site at http://www. LLC . III. now known or hereafter invented. CCC is a notforprofit organization that provides licenses and registration for a variety of users. microfilming.
iii To our boys Anondo and Alok and To Yu’s teacher. Professor Zhicheng Xie of Tsinghua University. © 2005 by Taylor & Francis Group. a distinguished scholar who has dedicated himself to China. LLC .
is its dimensionality advantage. In addition to the solution of 3D problems. for example. for problems involving optimal shape design or adaptive meshing. The third part is concerned with the boundary node method (BNM). There are several popular computational methods for solving problems in potential theory and linear elasticity. meshfree (also called meshless) methods have become a popular research area in computational mechanics. this book consists of three parts related to the boundary element. Many hundreds of books already exist on the subject and new books get published frequently on a regular basis. Another popular method is the boundary element method (BEM). from the point of view of accuracy as well as computational eﬃciency. while Part III on the BNM includes error analysis and adaptivity. The most popular. © 2005 by Taylor & Francis Group. still remains an arduous task. One primary focus of this book is a marriage of these two ideas.e. Only onedimensional line integrals need to be numerically computed when solving threedimensional problems in linear elasticity by the BCM. Even though great strides have been made in recent years. for many applications. for linear problems. Following an introductory chapter. Advantages of meshfree methods become more pronounced. During the past decade. relative to the FEM. This method is a novel variant of the BEM in that it further reduces the dimensionality of a problem. Only some novel topics related to the BEM are presented here.which combines the dimensionality advantage of the BEM with the ease of discretization of meshfree methods. versatile and most commonly used is the ﬁnite element method (FEM). The BNM combines the BEM with moving leastsquares (MLS) approximants. shape optimization. The main purpose here is to substantially simplify the task of meshing of an object. for linear problems. Problems that were too large for main frame computers 15 or 20 years ago can now be routinely solved on desktop personal computers. Compared to the FEM. and error estimation and adaptivity. in that it is particularly well suited. The FEM is a domain method that requires discretization of the entire domain of a body while the BEM. we view the BEM as a niche method. The process of discretization (or meshing) of a threedimensional (3D) object of complex shape is a popular research area in computational geometry. in order not to duplicate information on the BEM that is already available in many books on the subject. The principal advantage of the BEM. This ﬁeld has undergone a revolution during the past several decades along with the exponential growth of computational power and memory. meshing.v PREFACE The general subject area of concern to this book is computational science and engineering. LLC . The ﬁrst part is short. The second part is concerned with the boundary contour method (BCM). Part II of the book on the BCM presents shape sensitivity analysis.the boundary node method (BNM) . with applications in potential theory and in solid mechanics (linear elasticity). a discussion of a boundarybased meshfree method . only requires discretization of its bounding surface. boundary contour and boundary node methods. since many remeshings must be typically carried out for such problems. i. thus producing a meshfree boundaryonly method.
Two of Subrata’s former graduate students. Each topic is clearly introduced and developed. New York October 2004 © 2005 by Taylor & Francis Group. as well as Yu’s associate Xiaolan Shi. Srinivas Telukunta. It should also be of value to advanced undergraduate students who are interested in this ﬁeld. Sincere thanks are expressed to our dear friend Ashim Datta for his help and encouragement throughout the writing of this book. Sincere thanks are expressed to Subrata’s former graduate students Govind Menon and Ramesh Gowrishankar. Much of the research presented here has been ﬁnancially supported by the National Science Foundation and Ford Motor Company. Numerical results for selected problems appear throughout the book. while Subrata’s justgraduated student Zhongping Bao has made excellent contributions to the research on microelectromechanical systems (MEMS) by the BEM. The original source has been acknowledged in this book at the end of the caption for each item. This book should be of great interest to graduate students. have made very signiﬁcant contributions to the research that led to this book. Anantharaman Nagarajan and AnhVu Phan have signiﬁcantly contributed to the early development of the BCM.e. as do references to related work (research publications and books). science and engineering. Earlin Lutz. for their contributions to the BNM. and to others interested in the general areas of computational mathematics. Subrata and Yu Mukherjee Ithaca. researchers and practicing engineers in the ﬁeld of computational mechanics. Most of the ﬁgures and tables in this book have been published before in journals. and to Vasanth Kothnur. LLC .vi This book is written in the style of a research monograph. and this permission is very much appreciated. These items have been printed here by permission of the original copyright owner (i. to one of his present graduate students. together with their coauthors. We wish to thank a number of people and organizations who have contributed in various ways to making this book possible. the publishers of the appropriate journal). They were all originally created by the authors of this book. Glaucio Paulino and Mandar Chati. and this support is gratefully acknowledged.
. . . 2 ERROR ESTIMATION 2. .2 Linear Elasticity in Three Dimensions . . . . .4 Numerical Examples .2. . . . . . . . . . . . . . . . . . . . 23 23 25 25 28 30 32 33 . . . . . . . . . . . LLC . 1. .3 Problem 3 : Mixed Boundary Conditions . . . .3.4. . . . . .2 Stresses at Internal Points Close to the Boundary . 1. . . . . . . . . . 1. 1. . .2. . .3 Stress BIE for 3D Elasticity . . . .1. . . . . 1. . . . . . .2 Hypersingular Integral Equations . . . . . . .2. . .1 Problem 1 : Displacement Boundary Conditions 2. . 19 . . . . . . . . . . . . . . . .3 Nearly Singular Integrals in Linear Elasticity .2 Iterated HBIE and Error Estimation . . . .3 ElementBased Error Indicators . . . . . .2 Problem 2 : Traction Boundary Conditions . . . . 2. . . . 1. 2. . .4 Solution Strategy for a HBIE Collocated at an Irregular Boundary Point . . . . .4. . . . 17 .2. . 1. . . . . 1. . 1. . . 2. .3.1 Linear Operators . . . .1 Singular Integral Equations . . . . . . . . . . . . . . .1 Finite Part of a Hypersingular Integral Collocated at an Irregular Boundary Point . . . . . . . . .1. . . . . . . . . . . . . . 1. . . . . . . . . . . . . . . . .4.Contents Preface INTRODUCTION TO BOUNDARY METHODS v xiii I SELECTED TOPICS IN BOUNDARY ELEMENT METHODS 1 1 BOUNDARY INTEGRAL EQUATIONS 1. . .4. .1 Singular Integral Equations . . . . . . . . 20 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2. . . . . . . 3 3 3 5 6 6 8 12 12 13 14 . . . 1. . . . . .2. .1 Potential Theory in Three Dimensions . . . . . . . . . .4 Finite Parts of Hypersingular Equations . . . . .2 Hypersingular Integral Equations . 14 . vii © 2005 by Taylor & Francis Group. . . . . . . . .2 Gradient BIE for 3D Laplace’s Equation . . . . . . . . . . .1 Displacements at Internal Points Close to the Boundary 1. . . . . . 1. . . 2.
4. . . . . . . . . . and Curvatures . . . . .1. . . . . . . II 4 THE BOUNDARY CONTOUR METHOD LINEAR ELASTICITY 4. . . . . . .1. . . . . . . . .2 BIE for Elasticity: Cracks and Thin Shells . . . . . . . . . . . . . . . . . . 3. . . . . . . . .1 BIES in LEFM .3. . .1 Exterior BIE for Potential Theory: MEMS . . . . . . . . . . . .6 The Model Problem . . . . . . . 4. . . . . 3. . . . .5 Numerical Results . . . . . . . . . .1. . . . . . . . . . . . . . . . . . . . . . . . . . 34 Example 2: Kirsch’s Problem of an Inﬁnite Plate with a Circular Cutout . . . . . .2. . . . .1. . . . . . . . .2. . . . . . . . . . . . . .1 Regularized Hypersingular BIE . . . .4.4. . . 3.1. . . . . . .1. . . . .2. . . . . . . . . . 65 67 67 67 68 71 73 74 76 76 78 78 78 80 82 82 82 83 84 85 85 © 2005 by Taylor & Francis Group.viii 2. . . . .2. . . . . . . .2 Hypersingular Boundary Integral Equations . . . . . . . .3. . . . . . . . . . . . . .3 Boundary Elements . . . 4. . . . . . . . .1 Basic Equations . .4. . . . . . . 3. . . . . . . . . . . .4 Stresses at Internal Points Close to the Bounding Surface 4. 41 46 49 50 54 54 60 61 62 THIN FEATURES 3. .4 Vector Potentials . . . . . .1. . . . . . . . . . 39 . .3 Internal Stresses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 39 . . . . . 3. . 39 . . . . . . . . . . .2 Electric Field BIEs in a Simply Connected Body . . . 3. . . . . . . . . . . . . . . 4. . . . .a Parallel Plate Capacitor . . . . . . . . 4. . . . . . . 4. . . 3. . . . 41 . . . . . .2 3 CONTENTS Example 1: Lam´’s Problem of a ThickWalled Cylinder e under Internal Pressure .1 Introduction to MEMS . . . . .1. . . . . .2 Numerical Implementation of BIES in LEFM . . . .2 Interpolation Functions .1. . . . 4. . . . . . . . . . .3 Internal Displacements and Stresses .4 Numerical Results . . . . .2. .1 Surface and Boundary Contour Equations . . .2. . . . . . .1. 3. . . . Stresses. .1. . . .1.3 Some Comments on BIEs in LEFM . . . 4. . . .3 Collocation of the HBCE at an Irregular Surface Point . . .3 BIES in Inﬁnite Region Containing Two Thin Conducting Plates . 4. . . . . .1 2. . 4. . . . . . . . 3. . LLC . . . . . . . . . . 4. . . . . . . . . . . .7 Surface Displacements. . . . . . . . . . . . . . . .1 Surface Displacements from the BCM and the HBCM .5 Final BCM Equations . . . 3. . . . . . . . . . . . .4 BIEs for Thin Shells .3. . . . . . . 4. 4. . . 4. . . .1. . . . . . .2. . . . . . . .6 Global Equations and Unknowns . . . . . . . . .4 Singular and Nearly Singular Integrals . .2 Displacements at Internal Points Close to the Bounding Surface . . 4. . . . . 3. 4. .2 Regularized Hypersingular BCE . . . . .3.1 Internal Displacements . . 4. . . . . . . . .
. . . . .3. . . .4.3 Sensitivities of Internal Stresses . . . . . . . . . . . . . . . . . . . . . . . . . .1. . . . . .3. . . . . . . . . 5. 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 The BCM Sensitivity Equation . . . 125 7. 5. . . . . . 5. . . . . . . .2 Optimal Shapes of Ellipsoidal Cavities 6.5. . .2 Surface Derivatives . . . . . . . . . . . 7 ERROR ESTIMATION AND ADAPTIVITY 125 7. . . . . . . . . .1 Hypersingular Residuals as Local Error Estimators . . . . .3 The Integral Jk . . . . . . . .4 Method Four .1 Sensitivity of the BIE . . . . . . . 127 7. LLC . . . . . . . . . 5. . . . . . . .2 Numerical Results . .1 Geometry and Mesh . . . . . . . . . . . . . . .1 Sensitivities of Displacements . . .2. . . . . . . . . . . 5. . . . . . . . . . .4 ix Surface Stresses . . . . . . . . . . . . . . . . .1 Sensitivities on Sphere Surface . 6. . 87 Internal Stresses Relatively Far from the Bounding Surface 90 Internal Stresses Very Close to the Bounding Surface .1 Shape Optimization of a Fillet . . . . . . . . . . .1. .4. . . . . . . 126 7. . . .4. . . . . . 5. . .4. . . . . . . .4 Numerical Results: Hollow Sphere . . . . . . . . . . 139 © 2005 by Taylor & Francis Group. .3 Sensitivities of Variables at Internal Points .2 Adaptive Meshing Strategy .2 Internal Stresses . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4. . . . . . . .1 Example One . . . . . 5. . . 5. .5 Numerical Results: Block with a Hole . 5. . . . . . 6. . 5. .1 Method One . . . . . .3 Numerical Results . . . . . . .1 Sensitivities of Boundary Variables . .1 Shape Optimization Problems . . .2. . . . . . Inside Cubes .2. . . . 5. . . . . . . .2. . . 135 8. . 127 7. . .Short Clamped Cylinder under Tension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Example Two . . . . . . . .3.2 4. . . . . .3 Remarks . . . . 5. .2. . . . . . . . . . . . . .the Lam´ Problem for a Hollow Cylinder 130 e III THE BOUNDARY NODE METHOD 133 8 SURFACE APPROXIMANTS 135 8. .5. . . . 5. . . . . . . . . . . .1. . . . . . . . . . . . .5. . . . 90 . . 5. .3 Method Three . . . . . 5. . . . . . . . . . . . . . . . . .1 Moving Least Squares (MLS) Approximants . 6 SHAPE OPTIMIZATION 6. . . . . . . .1. .2 Sensitivities of Displacement Gradients and Stresses 5. . . . . . . . . 6. . . . . . . . . 93 93 93 94 96 98 99 100 100 100 101 101 101 103 106 107 108 110 110 112 112 115 115 116 116 118 122 5 SHAPE SENSITIVITY ANALYSIS 5. . . . . . . . .2 Method Two . . . . . .2. .2 The Integral Ik . . . . . . . . . .2. . . . . . . . . . .3 4. . . . . . . .CONTENTS 4. . . . . . . . . . 5. . . . 5. . . . . . . . . . .2 Sensitivities at Internal Points . . . . .3. . . . . .2 Sensitivities of Surface Stresses .
2 Local Residuals and Errors . . 9. . . .4. . .x 8. . . . . . . 188 10.4. . . . . . . .3. . . . . . . . . . . . . . . . . . 10 ADAPTIVITY FOR 3D POTENTIAL THEORY 175 10. .3 8. . . . . . . . . . . . 195 11. . . . . . . . . . . . . . . . . .Singular Residual Approach 195 11. . 176 10. . . . . . . .2 Initial Cell Conﬁguration # 1 (54 Surface Cells) . . 194 11. . 181 10. . . .2. . .2.2 Local Residuals and Errors . 9. . . .3 Numerical Results . .2 The Singular Residual . .3. .2. . 9. . . . . . . . . . . 193 11. . . 195 © 2005 by Taylor & Francis Group.1.1 Local Residuals and Errors . . . . . . . . . . . . 191 11 ADAPTIVITY FOR 3D LINEAR ELASTICITY 193 11. . . . . . . . . . . . . . . .3 Progressively Adaptive Solutions: Cube Problem . . . . . . . 9. . . . . . 193 11. . . . . . . .2. .1. . . . . . . . . . . . . . . . . . . 175 10. . . .4 Weight Functions . . . . .1 Hypersingular and Singular Residuals . . . . . .1. . .3 Numerical Results for Dirichlet Problems on a Sphere 9. . . . . . . . . .1 Exact Solution . . . .2 HBNM: Coupling of HBIE with MLS Approximants . 9.1. . . CONTENTS . . .Hypersingular Residual Approach . . . . . . . 177 10.2 Error Estimation and Adaptive Strategy . .2 Error Estimation and Adaptive Strategy . . . .2 Initial Cell Conﬁguration # 2 (96 Surface Cells) . . 194 11. . . . . . . LLC . .1 The Hypersingular Residual . . . .2 Linear Elasticity in Three Dimensions . . . . . . . . . . . . . . . . . .2. . . .2. . . . . . . . . . . . . . . . . . . . . . . 175 10. . . . .2 The Singular Residual . . . . . 178 10. . 9. . . . . .3. . . . . . . . . . . . . . . .4 OneStep Adaptive Cell Reﬁnement . . 182 10. . . . . . . . . . . 179 10. . . . . . 141 142 142 143 151 151 151 155 156 165 165 167 168 9 POTENTIAL THEORY AND ELASTICITY 9. . . . . .1 Initial Cell Conﬁguration # 1 (54 Surface Cells) . . . . . . . . . .1 BNM: Coupling of BIE with MLS Approximants . . . . . . . . . .3 Cell Reﬁnement Global Error Estimation and Stopping Criterion . . . . . . . . . .1. . . . . . . . . 180 10.1 BNM: Coupling of BIE with MLS Approximants .2 Variable Basis Approximation . . . . . . . . . . . . .1 Hermite Type Approximation 8. . .4. . . . .2. . . . . . . . . . . . . . . . . 194 11. . . . . .1. . .Hypersingular Residual Approach . . . . .1. . . . . . . .4. . .1 Hypersingular and Singular Residuals . . . 181 10.2 HBNM: Coupling of HBIE with MLS Approximants . . . . . . . . . . . . . . . . . .3 Cell Reﬁnement Criterion .Singular Residual Approach 178 10. . . .1 Potential Theory in Three Dimensions . . . 190 10. . .3 Progressively Adaptive Solutions: Pulling a Rod . .1 Local Residuals and Errors . . . . .2. . .2. . . .2. . . . . . . . . . . . . .3 Initial Cell Conﬁguration # 2 (96 Surface Cells) . . 179 10. . . . . .1 The Hypersingular Residual . . . . . Use of Cartesian Coordinates . . . . . .4 Global Error Estimation and Stopping Criterion . 8.
. . . . . . .4 OneStep Adaptive Cell Reﬁnement . . .1 Initial Cell Conﬁguration . 198 Bibliography 203 © 2005 by Taylor & Francis Group. . . . . . . . .3. . .2 Adaptivity Results . . .3. . 197 11. . . . . . . . .CONTENTS xi 11. . . . . 197 11. . . . . . . . LLC . . . . . . . . . . . . .
in fracture mechanics (e. Chandra and Mukherjee [22]. [54]. Brebbia and Dominguez [16]. Wilde and Aliabadi [173]. Krishnasamy et al. to xiii © 2005 by Taylor & Francis Group. [75]). [47]. Mukherjee et al. Gray and Paulino [58]. Boundary Element Method Boundary integral equations (BIE). [89]. error estimation in Chapter 2 and thin features (cracks and thin objects) in Chapter 3. Krishnasamy et al. [55]. for example. [110]). to obtain symmetric Galerkin boundary element formulations (e. [57]). for example. HBIEs. Examples of books on the subject.g. have been employed for the evaluation of boundary stresses (e. Hartmann [62]. Cruse [38].INTRODUCTION TO BOUNDARY METHODS This chapter provides a brief introduction to various topics that are of interest in this book. Gray and Paulino ([56]. BEM topics of interest in this book are ıs n ﬁnite parts (FP) in Chapter 1. Bonnet [14]. based on BIEs. Zhao and Lan [185]. Hypersingular Boundary Integral Equations Hypersingular boundary integral equations (HBIEs) are derived from a diﬀerentiated version of the usual boundary integral equations (BIEs). Paulino [122]. HBIEs have diverse important applications and are the subject of considerable current research (see. Tanaka et al. are mature methods for numerical analysis of a large variety of problems in science and engineering. Chati and Mukherjee [24]). The standard BEM for linear problems has the wellknown dimensionality advantage in that only the twodimensional (2D) bounding surface of a threedimensional (3D) body needs to be meshed when this method is used.g. and the boundary element method (BEM). Kane [68] and Par´ and Ca˜as [121]. Gray et al.g. Bonnet [14]. [162]. published during the last 15 years. in wave scattering (e. Guiggiani [60]. Gaul et al. LLC . Lutz et al. Becker [9]. [76]. Gray et al. Paulino [122] and Chen and Hong [30] for recent surveys of the ﬁeld). are Banerjee [4].g.
the extended ﬁnite element method (XFEM) [97. have tremendous potential advantages over methods such as the ﬁnite element method (FEM) that require discretization of a body into elements. h − p clouds [42. The relationship between ﬁnite parts of strongly singular and hypersingular integrals. [138]. Meshfree methods proposed to date include the elementfree Galerkin (EFG) method [10. 11. [92]. 80]. 28. using simple solutions. Mukherjee et al. the ﬁnite cloud method (FCM) [2]. Poon et al. that only require points rather than elements to be speciﬁed in the physical domain. is described in a book by Lancaster and Salkauskas [78]. is discussed in [168]. 108]. [117] proposed a coupling of MLS interpolants with Galerkin procedures in order to solve boundary value problems. the boundary point interpolation method (BPIM) [82]. 13.xiv INTRODUCTION TO BOUNDARY METHODS evaluate nearly singular integrals (Mukherjee et al.g. [104]). 67. the local boundary integral equation (LBIE) method [152. Lutz et al. 43. the method of ﬁnite spheres (MFS) [40]. 26. 160]. the meshless local PetrovGalerkin (MLPG) approach [3]. 159]. was ﬁrst proposed by Rudolphi [143]. LLC . An alternative way of satisfying this smoothness requirement is the use of the hypersingular boundary node method (HBNM). [96]. 41. 72. An elegant approach of regularizing singular and hypersingular integrals. 174. Chati et al. Mukherjee and Mukherjee [99]). the boundary cloud method (BCLM) [79. on smoothness requirements on boundary variables for collocating an HBIE on the boundary of a body. 27. 188]. 176. the meshless regular local boundary integral equation (MRLBIE) method [189]. 12. for curve and surface ﬁtting. and for error analysis (Paulino et al. Examples are Cruse and Richardson [39]. [89]. Paulino and Gray [125]) and adaptivity [28]. MeshFree Methods Meshfree (also called meshless) methods [82]. 25. to obtain the hypersingular boundary contour method (Phan et al. © 2005 by Taylor & Francis Group. [27]. A lively debate (e. [110] and Mukherjee [106]. Nayroles et al. 52]. to obtain the hypersingular boundary node method (Chati et al. [123]. Menon et al. the boundaryonly radial basis function method (BRBFM) [32] and the boundary node method (BNM) [107. 84]. [39]). great progress has been made in solid mechanics applications of meshfree methods. and the HBIE. 175. They called their method the diﬀuse element method (DEM) and applied it to twodimensional (2D) problems in potential theory and linear elasticity. [131]. the generalized ﬁnite element method (GFEM) [157]. [101] and [102]. 120]. the natural element method (NEM) [158. Several researchers have used this idea to regularize hypersingular integrals before collocating an HBIE at a regular boundary point. has apparently been concluded recently [93]. [27]). The idea of moving least squares (MLS) interpolants. Menon [95]. During the relatively short span of less than a decade. the reproducingkernel particle method (RKPM) [83.
and the EFG needs points throughout the domain of a body. e.g. The method divorces the traditional coupling between spatial discretization (meshing) and interpolation (as commonly practiced in the FEM or in the BEM). The only requirements are that the intersection of any two surface cells is the null set and that the union of all the cells is the bounding surface of the body. the FEM needs volume meshing. They can be easily designed to be suﬃciently smooth to suit a given purpose. they can be made C 1 or higher [10] in order to collocate the HBNM at a point on the boundary of a body. however. Figure i: BNM with nodes and cells Figure ii: BEM with nodes and elements (from [28]) (from [28]) © 2005 by Taylor & Francis Group. Mukherjee. LLC . The BNM is a combination of the MLS interpolation scheme and the standard boundary integral equation (BIE) method. Other examples of boundarybased meshless methods are the boundary cloud method (BCLM) [79. 25. 107]. is not a boundary method since it requires evaluation of integrals over certain surfaces (called Ls in [188]) that can be regarded as “closure surfaces” of boundary elements. Instead. 28. the BEM needs surface meshing. together with his research collaborators. 80]. the BNM only requires the speciﬁcation of points on the 2D bounding surface of a 3D body (including crack faces in fracture mechanics problems). 72. the boundary only radial basis function method (BRBFM) [32] and the local BIE (LBIE) [188] approach. any cell can be arbitrarily subdivided without aﬀecting its neighbors [27]) are used for integration.g. is used to represent the unknown functions. the BNM retains the meshless attribute of the EFG method and the dimensionality advantage of the BEM. Thus. has recently pioneered a new computational approach called the boundary node method (BNM) [26. a “diﬀuse” interpolation. The BNM is described in Chapters 8 and 9 of this book. In contrast. together with surface cells for integration.xv Boundary Node Method S. with a very ﬂexible structure (e. based on MLS interpolants. The LBIE. and surface cells. It is important to point out another important advantage of MLS interpolants. As a consequence. thereby practically eliminating the meshing problem (see Figures i and ii). The required cell structure is analogous to (but not the same as) a tiling [139]. 27. the boundary point interpolation method (BPIM) [82].
xvi
INTRODUCTION TO BOUNDARY METHODS
Boundary Contour Method
The Method
The usual boundary element method (BEM), for threedimensional (3D) linear elasticity, requires numerical evaluations of surface integrals on boundary elements on the surface of a body (see, for example, [98]). [115] (for 2D linear elasticity) and [116] (for 3D linear elasticity) have recently proposed a novel approach, called the boundary contour method (BCM), that achieves a further reduction in dimension! The BCM, for 3D linear elasticity problems, only requires numerical evaluation of line integrals over the closed bounding contours of the usual (surface) boundary elements. The central idea of the BCM is the exploitation of the divergencefree property of the usual BEM integrand and a very useful application of Stokes’ theorem, to analytically convert surface integrals on boundary elements to line integrals on closed contours that bound these elements. [88] ﬁrst proposed an application of this idea for the Laplace equation and Nagarajan et al. generalized this idea to linear elasticity. Numerical results for twodimensional (2D) problems, with linear boundary elements, are presented in [115], while results with quadratic boundary elements appear in [129]. Threedimensional elasticity problems, with quadratic boundary elements, are the subject of [116] and [109]. Hypersingular boundary contour formulations, for twodimensional [131] and threedimensional [99] linear elasticity, have been proposed recently. A symmetric Galerkin BCM for 2D linear elasticity appears in [119]. Recent work on the BCM is available in [31, 134, 135, 136, 186]. The BCM is described in Chapter 4 of this book.
Shape Sensitivity Analysis with the BCM and the HBCM
Design sensitivity coeﬃcients (DSCs), which are deﬁned as rates of change of physical response quantities with respect to changes in design variables, are useful for various applications such as in judging the robustness of a given design, in reliability analysis and in solving inverse and design optimization problems. There are three methods for design sensitivity analysis (e.g. [63]), namely, the ﬁnite diﬀerence approach (FDA), the adjoint structure approach (ASA) and the direct diﬀerentiation approach (DDA). The DDA is of interest in this work. The goal of obtaining BCM sensitivity equations can be achieved in two equivalent ways. In the 2D work by [130], design sensitivities are obtained by ﬁrst converting the discretized BIEs into their boundary contour version, and then applying the DDA (using the concept of the material derivative) to this BCM version. This approach, while relatively straightforward in principle, becomes extremely algebraically intensive for 3D elasticity problems. [100] oﬀers a novel alternative derivation, using the opposite process, in which the DDA is ﬁrst applied to the regularized BIE and then the resulting equations
© 2005 by Taylor & Francis Group, LLC
xvii are converted to their boundary contour version. It is important to point out that this process of converting the sensitivity BIE into a BCM form is quite challenging. This new derivation, for sensitivities of surface variables [100], as well as for internal variables [103], for 3D elasticity problems, is presented in Chapter 5 of this book. The reader is referred to [133] for a corresponding derivation for 2D elasticity
Shape Optimization with the BCM
Shape optimization refers to the optimal design of the shape of structural components and is of great importance in mechanical engineering design. A typical gradientbased shape optimization procedure is an iterative process in which iterative improvements are carried out over successive designs until an optimal design is accepted. A domainbased method such as the ﬁnite element method (FEM) typically requires discretization of the entire domain of a body many times during this iterative process. The BEM, however, only requires surface discretization, so that mesh generation and remeshing procedures can be carried out much more easily for the BEM than for the FEM. Also, surface stresses are typically obtained very accurately in the BEM. As a result, the BEM has been a popular method for shape optimization in linear mechanics. Some examples are references [33], [145], [178], [144], [169], [177], [161] and the book [184]. In addition to having the same meshing advantages as the usual BEM, the BCM, as explained above, oﬀers a further reduction in dimension. Also, surface stresses can be obtained very easily and accurately by the BCM without the need for additional shape function diﬀerentiation as is commonly required with the BEM. These properties make the BCM very attractive as the computational engine for stress analysis for use in shape optimization. Shape optimization in 2D linear elasticity, with the BCM, has been presented by [132]. The corresponding 3D problem is presented in [150] and is discussed in Chapter 6.
Error Estimation and Adaptivity
A particular strength of the ﬁnite element method (FEM) is the welldeveloped theory of error estimation, and its use in adaptive methods (see, for example, Ciarlet [34], Eriksson et al. [44]). In contrast, error estimation in the boundary element method (BEM) is a subject that has attracted attention mainly over the past decade, and much work remains to be done. For recent surveys on error estimation and adaptivity in the BEM, see Sloan [155], Kita and Kamiya [70], Liapis [81] and Paulino et al. [124]. Many error estimators in the BEM are essentially heuristic and, unlike for the FEM, theoretical work in this ﬁeld has been quite limited. Rank [140] proposed error indicators and an adaptive algorithm for the BEM using techniques similar to those used in the FEM. Most notable is the work of Yu and Wendland [171, 172, 181, 182], who have presented local error estimates based
© 2005 by Taylor & Francis Group, LLC
xviii
INTRODUCTION TO BOUNDARY METHODS
on a linear errorresidual relation that is very eﬀective in the FEM. More recently, Carstensen et al. [18, 21, 19, 20] have presented error estimates for the BEM analogous to the approach of Eriksson [44] for the FEM. There are numerous stumbling blocks in the development of a satisfactory theoretical analysis of a generic boundary value problem (BVP). First, theoretical analyses are easiest for Galerkin schemes, but most engineering codes, to date, use collocationbased methods (see, for example, Banerjee [4]). Though one can view collocation schemes as variants of PetrovGalerkin methods, and, in fact, numerous theoretical analyses exist for collocation methods (see, for example, references in [155]), the mathematical analysis for this class of problems is diﬃcult. Theoretical analyses for mixed boundary conditions are limited and involved (Wendland et al. [170]) and the presence of corners and cracks has been a source of challenging problems for many years (Sloan [155], Costabel and Stephan [35], Costabel et al. [36]). Of course, problems with corners and mixed boundary conditions are the ones of most practical interest, and for such situations one has to rely mostly on numerical experiments. During the past few years, there has been a marked interest, among mathematicians in the ﬁeld, in extending analyses for the BEM with singular integrals to hypersingular integrals ([21, 19, 156, 45]. For instance, Feistauer et al. [45] have studied the solution of the exterior Neumann problem for the Helmholtz equation formulated as an HBIE. Their paper contains a rigorous analysis of hypersingular integral equations and addresses the problem of noncompatibility of the residual norm, where additional hypotheses are needed to design a practical error estimate. These authors use residuals to estimate the error, but they do not use the BIE and the HBIE simultaneously. Finally, Goldberg and Bowman [51] have used superconvergence of the Sloan iterate [153, 154] to show the asymptotic equivalence of the error and the residual. They have used Galerkin methods, an iteration scheme that uses the same integral equation for the approximation and for the iterates, and usual residuals in their work. Paulino [122] and Paulino et al. [123] ﬁrst proposed the idea of obtaining a hypersingular residual by substituting the BEM solution of a problem into the hypersingular BEM (HBEM) for the same problem; and then using this residual as an element error estimator in the BEM. It has been proved that ([95], [96], [127]), under certain conditions, this residual is related to a measure of the local error on a boundary element, and has been used to postulate local error estimates on that element. This idea has been applied to the collocation BEM ([123], [96], [127]) and to the symmetric Galerkin BEM ([125]). Recently, residuals have been obtained in the context of the BNM [28] and used to obtain local error estimates (at the element level) and then to drive an hadaptive mesh reﬁnement process. An analogous approach for error estimation and hadaptivity, in the context of the BCM, is described in [111]. Ref. [91] has a bibliography of work on mesh generation and reﬁnement up to 1993. Error analysis with the BEM is presented in Chapter 2, while error analysis and adaptivity in the context of the BCM and the BNM are discussed in Chapter 7, and Chapters 10, 11, respectively, of this book.
© 2005 by Taylor & Francis Group, LLC
Part I
SELECTED TOPICS IN BOUNDARY ELEMENT METHODS
1
© 2005 by Taylor & Francis Group, LLC
1. respectively. y)τ (y) − F (ξ. respectively. 1.1. LLC . usual as well as hypersingular. The wellknown integral representation for (1. are ﬁrst presented in this chapter. The evaluation of ﬁnite parts (FPs) of some of these equations. This is followed by their linear elasticity counterparts. (Source and ﬁeld points are also referred to as p and q (for internal points) and as P and Q (for boundary points). for potential theory in three dimensions.1 Potential Theory in Three Dimensions The starting point is Laplace’s equation in three dimensions (3D) governing a potential function u(x1 . in this book). x2 . is: u(ξ) = ∂B [G(ξ. x3 ) ∈ B.1). x3 ) ≡ ∂2u ∂2u ∂2u + + =0 ∂x2 ∂x2 ∂x2 1 2 3 (1. let ξ and η be (internal) source and ﬁeld points ∈ B and x and y be (boundary) source and ﬁeld points ∈ ∂B. when the source point is an irregular boundary point (situated at a corner on a onedimensional plane curve or at a corner or edge on a twodimensional surface). where B is a bounded region (also called the body): ∇2 u(x1 . y)u(y)]dS(y) 3 (1. is described next.2) © 2005 by Taylor & Francis Group.1 Singular Integral Equations Referring to Figure 1. at an internal point ξ ∈ B. 1. x2 .1) along with prescribed boundary conditions on the bounding surface ∂B of B. for internal and boundary points.Chapter 1 BOUNDARY INTEGRAL EQUATIONS Integral equations.
A regularized form of the resulting equation is: 0= ∂B [G(x. An alternative form of equation (1. y) (1. ek · dS(y) = nk (y)dS(y). the Euclidean distance between the source and ﬁeld points ξ and y. are the usual Cartesian unit vectors.4 CHAPTER 1.2) is: u(ξ) = ∂B [G(ξ.4) ∂G(ξ. y) (1. y). 3. y) (ξi − yi )ni (y) = ∂n(y) 4πr3 (ξ. y)u(y)]ek · dS(y) (1.k (y) − Hk (ξ.1: Notation used in integral equations (from [6]) An inﬁnitesimal surface area on ∂B is dS = dSn.2) is obtained by taking the limit ξ → x.5) where ek .2. and: Hk (ξ. These are : G(ξ. where n is the unit outward normal to ∂B at a point on it and τ = ∂u/∂n. the range of indices in these and all other equations in this chapter is 1. 2. The kernels are written in terms of source and ﬁeld points ξ ∈ B and y ∈ ∂B. k = 1. Unless speciﬁed otherwise. BOUNDARY INTEGRAL EQUATIONS n(y) y(Q) r(ξ . y) = (ξk − yk ) 4πr3 (ξ. LLC . y) = 1 4πr(ξ. y)τ (y) − F (x. y){u(y) − u(x)}]dS(y) (1.y ) B ξ(p) η(q) x(P) ∂B n(x) Figure 1.3) (1. y) in terms of r(ξ.3.7) © 2005 by Taylor & Francis Group. y) = F (ξ.6) The boundary integral equation (BIE) corresponding to (1. y)u.
5)): 0= ∂B 5 [G(x.p (x)(yp − xp ) dS(y) ∂xm 0 = ∂B − ∂B (1.k (y) − Hk (x.10) where the symbol = denotes the ﬁnite part (FP) of the integral.p (y) − u. The result is: ∂u(ξ) = ∂ξm ∂F (ξ.p (x)(yp − xp ) dS(y) ∂xm ∂B = 0 (1. The regularized version given below is applicable even at an irregular boundary point x provided that u(y) ∈ C 1. Equation (1. y) ∂G(ξ. one obtains: ∂u(x) ∂F (x.2) can be diﬀerentiated at an internal source point ξ to obtain the ∂u gradient ∂ξm of the potential u. LLC . y) τ (y) − τ (x) dS(y) ∂n(x) 0 = ∂B © 2005 by Taylor & Francis Group.8) 1. y) − u(y) − u(x) − u.p (x) np (y)dS(y) ∂xm ∂F (x. y) u.1.12) Carrying out the inner product of (1. y)u.10) is best regularized before computations are carried out.α at y = x) in equation (1.12) with the source point normal n(x). y) ∂G(x. POTENTIAL THEORY IN THREE DIMENSIONS with an alternate form (from (1.11). valid at a regular boundary point x.k (x) nk (y) − nk (x) dS(y) ∂xm B ∂F (x. one gets: ∂G(x. y) τ (y) − τ (x) dS(y) ∂xm ∂B ∂G(x. y) == τ (y) − u(y) dS(y) ∂xm ∂xm ∂xm ∂B (1. y){u(y) − u(x)}]ek · dS(y) (1.9) ∂B An interesting situation arises when one takes the limit ξ → x (x can even be an irregular point on ∂B but one must have u(y) ∈ C 1.2. y) − u.4.1. y) τ (y) − u(y) dS(y) ∂ξm ∂ξm (1. y) u(y) − u(x) − u.1.11) An alternative form of (1.α at y = x. [76] is: ∂G(x. As discussed in detail in Section 1. This is: ∂G(x.2 Hypersingular Integral Equations Equation (1.9).
16) In equations (1. and the wellknown Kelvin kernels are: © 2005 by Taylor & Francis Group. y)ui (y)] dS(y) (1. at an internal point ξ ∈ B (Rizzo [141]) is: uk (ξ) = ∂B [Uik (ξ.k ni + µ(ui. x3 ) in a homogeneous.k (x) − ∂B ∂G(x. the gradient (at a regular boundary point) can be written as. x2 .6 CHAPTER 1. 1. ν is Poisson’s ratio and λ and µ are Lam´ e constants. y) u(y) − u(x) − u. As is well known.1. y) nk (y) − nk (x) dS(y) B ∂n(x) ∂F (x.p (x)(yp − xp ) dS(y) ∂n(x) (1.11 .14) where τ = ∂u/∂n is the ﬂux.1 Potential gradient on the bounding surface The gradient of the potential function is required in the regularized HBIEs (1. the Young’s modulus is denoted as E. For potential problems.15) and (1.17) where uk and τk are the components of the displacement and traction respectively.13). y)τi (y) − Tik (ξ.15). in the absence of body forces: 0 = ui. linear elastic solid occupying the bounded 3D region B with boundary ∂B. BOUNDARY INTEGRAL EQUATIONS − u.2. t1 .i )nj (1. and ∂u/∂si .1.16).2 Linear Elasticity in Three Dimensions The starting point is the NavierCauchy equation governing the displacement u(x1 . t2 are the appropriately chosen unit vectors in two orthogonal tangential directions on the surface of the body. Finally. 1.j + uj. 2 are the tangential derivatives of u (along t1 and t2 ) on the surface of the body.2. ∇u = τ n + ∂u ∂u t1 + t2 ∂s1 ∂s2 (1. isotropic. n is the unit normal.jj + 1 uk.13) 1.1 Singular Integral Equations The wellknown integral representation for (1. i = 1. µ is the shear modulus of the material and is also called G in this book. LLC . The components τi of the traction vector are: τi = λuk.15) along with prescribed boundary conditions that involve the displacement and the traction τ on ∂B.ki 1 − 2ν (1.
i = ∂r yi − ξi = ∂yi r (1.i r.k ni ) {(1 − 2ν)δik + 3r.24) is: 0= ∂B [Uik (x.i r.2. the normal n is deﬁned at the (boundary) ﬁeld point y. LINEAR ELASTICITY IN THREE DIMENSIONS 7 Uik = 1 [(3 − 4ν)δik + r. δik denotes the Kronecker delta and. τi = σij nj and Tik = Σijk nj .i nk − r.19) In the above. The explicit form of the kernel Σ is: ∂Ukm ∂yn Σijk =− = Eijmn 1 [ (1 − 2ν)(r.21) where σ is the stress tensor. as before.24) where the symbol = ∂B denotes the ﬁnite part of the appropriate integral (see Section 1.i δjk + r.23) The boundary integral equation (BIE) corresponding to (1.k δij ) + 3r. The result is: [Uik (ξ.4). y)τi (y) − Tik (x.18) 1 ∂r + (1 − 2ν)(r.25) © 2005 by Taylor & Francis Group. i.17) is: [Uik (ξ.i r.k } 8π(1 − ν)r2 ∂n (1. y)ui (y)] dS(y) ∂B (1.20) An alternative form of equation (1. LLC .k ] 16π(1 − ν)Gr (1. y)τi (y) − Tik (ξ. Tik = − r. y)ui (y)] ej · dS(y) ∂B uk (ξ) = (1.e. y)τi (y) − Tik (x. y)ui (y)] dS(y) ∂B uk (x) = ξ→x lim = = [Uik (x.22) 8π(1 − ν)r2 where E is the elasticity tensor (for isotropic elasticity): Eijmn = λδij δmn + µ[δim δjn + δin δjm ] (1.j δik − r. y){ui (y) − ui (x)}]dS(y) (1. A regularized form of equation (1.1.k ] (1. y)σij (y) − Σijk (ξ.17) is obtained by taking the limit ξ → x. A comma denotes a derivative with respect to a ﬁeld point. (Please note that ej · dS(y) = nj (y)dS(y)).j r.
26) 1.30) where the new kernels D and S are: ∂Ukm ∂Ukm =λ δij + µ ∂ξn ∂ξm ∂Uki ∂Ukj + ∂ξj ∂ξi Dijk = Eijmn = −Σijk (1.2 Hypersingular Integral Equations Equation (1.27) or (1. This equation.27) is: σij (ξ) = ∂B [Dijk (ξ.j + uj. BOUNDARY INTEGRAL EQUATIONS with an alternate form (from (1.31) Sijk ∂Σkpm ∂Σkpm ∂Σkpj ∂Σkpi np np = λ np δij + µ + ∂ξn ∂ξm ∂ξj ∂ξi ∂r G 3 = [(1 − 2ν)δij r.i r. corresponding (for example) to (1. however.32) = Eijmn Again.29) It is sometimes convenient. to write the internal stress directly.i ) (1.k ) 4π(1 − ν)r3 +(1 − 2ν)(3nk r.i r.k + nj r.j r. y){ui (y) − ui (x)}]ej · dS(y) ∂B 0= (1.k + ν(δik r.k ] 3 4π(1 − ν)r ∂n G + [3ν(ni r. y)σij (y) − (ξ. y)uk (y)] dS(y) (1.17) can be diﬀerentiated at an internal source point ξ to obtain the displacement gradient at this point: ∂uk (ξ) = ∂ξm ∂Tik ∂Uik (ξ.8 CHAPTER 1. y)σij (y) − Σijk (x.2.j + nj δik + ni δjk ) − (1 − 4ν)nk δij ] (1.27) ∂B An alternative form of equation (1.21)): [Uik (x. Also: © 2005 by Taylor & Francis Group. LLC . y)ui (y) ej · dS(y) ∂ξm ∂ξm (1.i ) − 5r.27) is: ∂uk (ξ) = ∂ξm ∂Σijk ∂Uik (ξ.i r.28) ∂B Stress components at an internal point ξ can be obtained from either of equations (1.28) by using Hooke’s law: σij = λuk.j r. y)ui (y) dS(y) ∂ξm ∂ξm (1. the normal n is deﬁned at the (boundary) ﬁeld point y. y)τi (y) − (ξ.j + δjk r.k δij + µ(ui. y)τk (y) − Sijk (ξ.
respectively: ∂uk (x) ∂xm ∂Σijk ∂Uik (ξ. also. (1.4.m (x. one notes that the traction at a boundary point is: [Dijk (ξ.30) become.34) Again.m ∂ξm (1. y) [σij (y) − σij (x)] nj (y)dS(y) ∂B 0 − = Σijk. y)σij (y) − (ξ. that only contain weakly singular integrals.e. for future reference.m (ξ.4 of this chapter) are: Uik. as r → 0). 1.37) Fully regularized forms of equations (1.35) and (1. This equation is: uk. y)ui (y) nj (y)dS(y) ∂ξm ∂ξm ∂B ∂Σijk ∂Uik (x. For future use in Chapter 4.m (x. For example. for example.m . y)uk (y)] dS(y) ∂B = = [Dijk (x. These equations.3).m (ξ.39) satisfy certain smoothness requirements (see Martin et al. y)uk (y)] dS(y) ∂B (1. y)uk (y)] dS(y) ∂B τi (x) = nj (x) lim ξ→x (1. y) = −Uik. LINEAR ELASTICITY IN THREE DIMENSIONS 9 ∂Uik (ξ. (1.38) ∂B © 2005 by Taylor & Francis Group. that can be collocated at an irregular point x ∈ ∂B provided that the stress and displacement ﬁelds in (1. as one takes the limit ξ → x in any of the equations (1. [93] and.28) using (1. y)σij (y) − Σijk.33).28) and (1.m (ξ) = − ∂B [Uik. y)ui (y)] nj (y)dS(y) (1. one must take the ﬁnite part of the corresponding right hand side (see Section 1.35) ∂xm ∂xm = ξ→x lim = = ∂B σij (x) = ξ→x lim [Dijk (ξ.36) Also. y)τk (y) − Sijk (ξ. are available in the literature (see.1. LLC .4.36). y)τk (y) − Sijk (ξ.2.27). it is useful to rewrite (1. Section 1. y)σij (y) − (x. y)ui (y) nj (y)dS(y) (1. y) [ui (y) − ui (x) − ui. y) = −Σijk.30).33) It is important to note that D becomes strongly singular. ∂ξm ∂Σijk (ξ.28) or (1. Cruse and Richardson [39]).38. y)τk (y) − Sijk (x. and S hypersingular as a source point approaches a ﬁeld point (i. (x) (y − x )] nj (y)dS(y) (1.
m (x)(ym − xm )] dS(y) (1.43) where uk . y)nj (x)[τk (y) − τk (x)]dS(y) Dijk (x.1 The gradient of the displacement u is required for the regularized HBIEs (1.2. LLC . one has: x = Qx u = Qu (1.39) that can only be collocated at a regular point x ∈ ∂B is: 0 = ∂B Dijk (x. are (x1 . y)nj (x) [uk (y) − uk (x) − uk.42) (1. x2 .41). 3 are the components of the displacement vector u in the local coordinate frame. BOUNDARY INTEGRAL EQUATIONS 0 = ∂B Dijk (x.p (x)(yp − xp )] dS(y) (1. k = 1.39) ∂B − An alternate version of (1. [89] have proposed a scheme for carrying this out. The local coordinate system is oriented such that the x1 and x2 coordinates lie along the tangential unit vectors t1 and t2 while x3 is measured along the outward normal unit vector n to ∂B as deﬁned in equation (1. y)[τk (y) − τk (x)]dS(y) Dijk (x. x2 .1. Lutz et al.2.38 . x3 ). and the orthogonal transformation matrix Q has the components: © 2005 by Taylor & Francis Group. 2.2. y) [σkp (y) − σkp (x)] np (y)dS(y) Sijk (x. taking the inner product of (1. y)(nm (y) − nm (x))dS(y) ∂B − σkm (x) − ∂B Sijk (x.14).40) Finally. Consider (righthanded) local Cartesian coordinates (x1 . y) [uk (y) − uk (x) − uk. The (righthanded) global Cartesian coordinates. Details of this procedure are available in [27] and are given below.10 CHAPTER 1. Therefore.m (x)(ym − xm )] dS(y) (1. y) [uk (y) − uk (x) − uk. x3 ) at a regular point P on ∂B as shown in Figure 1.41) Displacement gradient on the bounding surface 1.40) with the normal at the source point gives: 0 = ∂B Dijk (x. as before. y)nj (x)[nm (y) − nm (x)]dS(y) ∂B − σkm (x) − ∂B Sijk (x.
are the components of the traction vector in local coordinates. The remaining components of ∇u in local coordinates are obtained from Hooke’s law (see [89]) as: ∂u1 ∂x3 ∂u2 ∂x3 ∂u3 ∂x3 τ1 ∂u3 − G ∂x1 ∂u3 τ2 − G ∂x2 ν (1 − 2ν)τ3 ∂u2 ∂u1 − + 2G(1 − ν) 1 − ν ∂x1 ∂x2 = = = (1. 2. LINEAR ELASTICITY IN THREE DIMENSIONS 11 x '3 P ' x1 x '2 x2 x1 x3 Figure 1. 3. in the local coordinate system. 3. k = 1. 2. n3 ) the components of the unit normal vector. in local coordinates. They can be written as: © 2005 by Taylor & Francis Group. are ui. n2 . These quantities are obtained as follows: ui.k . The components of the displacement gradient tensor. k = 1.46) where τk . i = 1.44) with tij the j th component of the ith unit tangent vector and (n1 .2: Local coordinate system on the surface of a body (from [27]) t13 t23 n3 t11 Q = t21 n1 t12 t22 n2 (1.2.1.k ≡ ∂ui ∂uj = Qij ∂sk ∂sk (1. are now known. LLC . 2. The tangential derivatives of the displacement.45) where ∂ui /∂sk are tangential derivatives of ui at P with s1 = x1 and s2 = x2 .
1 ≡ A = QT A Q = u2.3. from Cruse and Richardson [39] is: ˆ uk (ξ) = uk (x) + ∂B [ Uik (ξ.1 u3.3 (∇u)global (1.25)) and obtain all the displacements and tractions on this surface.1 u1.30). [104]. dating back to Cruse [37]. y){ui (y) − ui (ˆ )} ] nj (y)dS(y) x (1.3 (1.48) The gradient of the displacement ﬁeld in global coordinates is now ready for use in equations (1. Various authors have addressed this issue over the last 3 decades.2 u3.1 ≡ A = u2.17) and (1.3). equation (1. from equations such as (1. 1.2 u1. The next steps are to obtain the displacements and stresses at selected points inside a body. A continuous version of (1.2 u1.3 u3.38 .17) or (1.1 u1.41).1.2 u2.12 CHAPTER 1. LLC . This section describes a new method recently proposed by Mukherjee et al. that one experiences diﬃculties when trying to numerically evaluate displacements and stresses at points inside a body that are close to its bounding surface (the socalled nearsingular or boundary layer problem).21).1 Displacements at Internal Points Close to the Boundary The displacement at a point inside an elastic body can be determined from either of the (equivalent) equations (1.21). the components of ∇u in the global coordinate frame are obtained from those in the local coordinate frame by using the tensor transformation rule: u1. BOUNDARY INTEGRAL EQUATIONS (∇u)local u1.3 Nearly Singular Integrals in Linear Elasticity It is well known that the ﬁrst step in the BEM is to solve the primary problem on the bounding surface of a body (e.3 u2. 1. An alternative form of (1.47) Finally.49) is: © 2005 by Taylor & Francis Group. y)σij (y) − Σijk (ξ.g.49) ˆ where ξ ∈ B is an internal point close to ∂B and a target point x ∈ ∂B is close to the point ξ (see Fig. It has been known in the BEM community for many years.2 u2.3 u2.2 u3.3 u3. 1.1 u3.
it is advisable to use the continuous equation (1. equation (1. or far from ∂B. LLC . Taking this limit is the standard approach for obtaining o the wellknown regularized form (1. y){ui (y) − ui (ˆ )}. whether ξ is near to. as y → x.49) (or (1.49) (or (1. 1.α (i. x x x x the product Tik (ξ. (The target point x can be chosen as any point on ∂B when ξ is far from ∂B). y)τi (y) − Tik (ξ.30). and novel use. which is O(r(ˆ .49) (or (1.50) (or (1. y)) as ˆ ˆ ξ → y. Continuous versions of (1.49) (or (1. → 0 ! As a result. y){ui (y) − ui (ˆ )} ] dS(y) (1.3. whereas {ui (y) − ui (ˆ )} is O(r(ˆ . Therefore.50)) universally for all points ξ ∈ B.34) and (1.50) x Equation (1.3. In this work.34) or the stress at this point from (1. equation (1.30) can be written as [39]: © 2005 by Taylor & Francis Group. y)). This idea is the main contribution of [104]. y)) as y → x. It is ﬁrst observed that Tik in equation (1. NEARLY SINGULAR INTEGRALS IN LINEAR ELASTICITY 13 B y z ξ v v z x ∂B ˆ Figure 1. H¨lder continuous). This procedure would eliminate the need to classify.21)) to evaluate uk (ξ) when ξ is far from ∂B.2 Stresses at Internal Points Close to the Boundary The displacement gradient at a point ξ ∈ B can be obtained from equation (1. equation (1. y)/r2 (ξ.50) is O(1/r2 (ξ.25)).26) (or (1.50)) is called “continuous” since it has a continuous ˆ limit to the boundary (LTB as ξ → x ∈ ∂B) provided that ui (y) ∈ C 0.50)) is put to a diﬀerent.e. a priori. It is noted here that while it is usual to use (1.17) (or (1. however.3: A body with source point ξ. ﬁeld point y and target point x (from [104]) uk (ξ) = uk (ˆ )+ x ∂B [ Uik (ξ.1.50)) is also valid in this ˆ case. Therefore.49)) can be used to easily and accurately evaluate the displacement components uk (ξ) for ξ ∈ B close to ∂B.
y)) and x ˆ O(r2 (ˆ . (ˆ )(y − x )] dS(y) x x ˆ ∂B The integrands in equations (1. Henceforth. have its only singularity at x = y of the form 1/r3 where r = x − y . Let the ˆ ¯ ˆ points x ∈ S and ξ ∈ S. equations (1.4 Finite Parts of Hypersingular Equations A discussion of ﬁnite parts (FPs) of hypersingular BIEs (see e. Let the function K(x.11)) is the subject of this section. for speciﬁcity. y)[uk (y) − uk (ˆ ) − uk.51) and (1. Similar to the behavior of the continuous BIEs x in the previous subsection. LLC .49).52) → 0 ˆ as y → x.α at y = x for some α > 0. (1.1. (ˆ ) (y − x )] nj (y)dS(y) x x ˆ σij (ξ) − = σij (ˆ ) + x ∂B Dijk (ξ.n (ξ) + ∂B = uk. (1.n (ξ. The ﬁnite part of the integral © 2005 by Taylor & Francis Group. y)/r3 (ξ.n (ˆ ) − x ∂B Uik.51) Σijk. while use of equations (1. (1. is very useful for evaluating the stresses at an internal point ξ that is close to ∂B. Either of these equations.51) (or (1. y) [σij (y) − σij (ˆ )] nj (y)dS(y) x (1. the space R3 .1 Finite Part of a Hypersingular Integral Collocated at an Irregular Boundary Point Deﬁnition Consider. y) .52) Sijk (ξ. use of equations (1. they can also be conveniently used to evaluate displacement gradients or stresses at any point ξ ∈ B. Further details are available in Mukherjee [102].9 1.50).17).n (ξ. (1. and let φ(y) be a function that has no singularity in S and is of class C 1. y) [ui (y) − ui (ˆ ) − ui. y)[τk (y) − σkm (ˆ )nm (y)]dS(y) x (1.21). 1. BOUNDARY INTEGRAL EQUATIONS uk. the integrands in equations (1.14 CHAPTER 1. The point x can be an irregular point on S.g.4. y)/r2 (ξ.1 1. let S and S ⊂ S be two neighborhoods (in S) of / ¯ x such that x ∈ S (Figure 1. 1. therefore. This is followed by applications of the theory in potential theory and in linear elasticity.51) or (1. The general theory of ﬁnite parts is presented ﬁrst.4. Also.30) or (1.52)) are O(r(ˆ .52) will be referred to as the new method. y)) as y → x. Of course (please see the discussion regarding displacements in the previous section).4). y ∈ S.34) will be referred to as the standard method. and let S be a surface in R3 .
and to Mukherjee [101] for a discussion of the relationship of this FP to the CPV of an integral when its CPV exists. y)(yp − xp )dS(y) ˆ S (1.4: A surface S with regions S and S and points ξ.53) is deﬁned as: = K(x. y)φ(y)dS(y) S (1. LLC . y)dS(y) ˆ S (1.56) The above FP deﬁnition can be easily extended to any number of physical dimensions and any order of singularity of the kernel function K(x.1.54) + ˆ S ˆ ˆ + φ(x)A(S) + φ. x and y (from [102]) I(x) = S K(x. Please refer to Toh and Mukherjee [168] for further discussion of a previous closely related FP deﬁnition for the case when x is a regular point on S. © 2005 by Taylor & Francis Group. y)φ(y)dS(y) = S ˆ S\S K(x.4. y). FINITE PARTS OF HYPERSINGULAR EQUATIONS 15 ξ S S x S y S S ˆ ¯ Figure 1.55) ˆ Bp (S) = = K(x. y)[φ(y) − φ(x) − φ.p (x)Bp (S) ˆ where S is any arbitrary neighborhood (in S) of x and: ˆ A(S) = = K(x.p (x)(yp − xp )]dS(y) (1. y)φ(y)dS(y) K(x.
p (x) = 1) in (1. Equations (1. LLC . in this case.54).61) K(ξ. From equation (1.57) Next.59) ¯ The second equality above holds since K(x. a closed surface that is the entire boundary of a body B. setting φ(y) = (yp − xp ) (note that.2 CHAPTER 1. and using (1.16 1. y)(yp − xp )dS(y) (1. Method two.58) are most useful for obtaining A and B when ˆ S is an open surface and Stoke regularization is employed.61) are most useful for evaluating A and B when ˆ S = ∂B.3 of this chapter.2 and 1.57) and (1. y)dS(y).57) and (1. y) is regular for x ∈ S and ˆ S.54).1. y)dS(y) ¯ S exist. y)dS(y) ˆ ¯ S\S (1.4. © 2005 by Taylor & Francis Group. Assuming that the limits: ¯ y ∈ S\ ξ→x lim K(ξ. [75]. Now.58) ˆ ¯ S\S The formulae (1.60) ξ→x ˆ S Equations (1. setting φ(y) = 1 in (1. φ(x) = 0 and φ. An example is the application of the FP deﬁnition (1.4. y)dS(y) = lim ˆ ¯ S\S ξ→x K(ξ. Examples appear in Sections 1. one gets: ˆ ¯ A(S) − A(S) = K(x.60) and (1. The resulting regularized equation is shown in [168] to be equivalent to the result of Krishnasamy et al.58) are also used in Mukherjee and Mukherjee [99] and in Section 3.57): ˆ ¯ A(S) − A(S) = K(x. one gets: ˆ ¯ Bp (S) − Bp (S) = K(x. ˆ ˆ ¯ Method one. y)(yp − xp )dS(y) (1. y)dS(y) ˆ ¯ S\S (1.4.2 of [102]. y)dS(y) ˆ S ξ→x (1.56). Replace S by S and S by S in equation (1.55). ˆ S ξ→x lim K(ξ.54) and using (1. BOUNDARY INTEGRAL EQUATIONS Evaluation of A and B There are several equivalent ways for evaluating A and B. to regularize a hypersingular integral that appears in the HBIE formulation for the scattering of acoustic waves by a thin scatterer. then: ˆ A(S) = lim Similarly: ˆ Bp (S) = lim K(ξ.54) (for a regular collocation point) in Toh and Mukherjee [168].
62). y)φ(y)dS(y) S S (1. A third way for evaluation of A and B is to use an auxiliary surface (or “tent”) as ﬁrst proposed for fracture mechanics analysis by Lutz et al.62).66) ∂u(x) = = [Di (x. S = ∂B is used here.p (x) (yp − xp )]dS(y) K(ξ.2 Gradient BIE for 3D Laplace’s Equation This section is concerned with an application of equation (1. y) = O(x − y−3 ) as y → x and φ(y) ∈ C 1. 1.4. An application of a vanishing exclusion zone. Consider the ﬁrst and second terms on the righthand side of equation (1. / x ∈ S (x can be an irregular point on S).54).3 The FP and the LTB There is a very simple connection between the FP. y)φ(y)dS(y) = = K(x. y)[φ(y) − φ(x) − φ. ˆ A complete exclusion zone. y)τ (y) − Si (ξ. 1. 1.65) (1. [110]. ξ ∈ S. 1. This method is useful if S is an open surface. LLC . y)[φ(y) − φ(ξ) − φ.62) Of course.2.64) in (1. K(x.63) and K(x. this can be stated as: ξ→x lim K(ξ. one has: K(x.4. Since these integrands are regular in their respective domains of integration. y)τ (y) − Si (x.9) and (1. Proof of equation (1. for collocation of the HBIE for the 2D Laplace equation. for the 3D Laplace equation. and the LTB approach employed by Gray and his coauthors. [89]. deﬁned above. at an irregular boundary point.1.4. as before. (see.10) are ﬁrst written in the slightly diﬀerent equivalent forms: ∂u(ξ) = ∂ξi [Di (ξ. equations (1. y)u(y)] dS(y) ∂xi ∂B © 2005 by Taylor & Francis Group. ξ can approach x from either side of S.α at y = x. Mukherjee et al.60.64) ˆ S = lim ξ→x ˆ S Use of equations (1. Mukherjee [105] and Section 3.p (ξ)(yp − ξp )]dS(y) (1. at an irregular boundary point. Using equations (1. With.9).54) proves equation (1.1.4) and (1. also.54) for collocation of the HBIE (1. y)φ(y)dS(y) ˆ S\S (1. is presented in Mukherjee [102]. FINITE PARTS OF HYPERSINGULAR EQUATIONS 17 Method three. y)φ(y)dS(y) = lim ˆ S\S ξ→x K(ξ.1 of [102].6). y)u(y)] dS(y) ∂B (1.63 and 1.61.
results in: u.i (x.71)) gives: [Di (ξ.69) Cip (∂B) = lim ξ→x [Di (ξ.1. y) = −G. y)[u(y) − u(x) − u.70) It is noted here that the (possibly irregular) boundary point x is approached from ξ ∈ B. (assuming continuity) Ai (∂B) = 0. y) u.66). y)nk (y) ˆ Use of (1.68) ξ→x (1. y)(yp − ξp )] dS(y) = δip ∂B (with p = 1.65) gives: Si (ξ.67) Di (x. use of the uniform solution u(y) = c (c is a constant) in equation (1.p (x)(yp − xp )]dS(y) ∂B − (1.p (x)(yp − xp ) dS(y) ∂B − − Ai (∂B)u(x) + Cip (∂B)u.p (y) − u.e. Cip (∂B) = δip .2: Ai (∂B) = lim Si (ξ. The quantities A and C can be easily evaluated using the imposition of simple solutions.73) Therefore. using method two in Section 1. Si = −Hk.p (y) − u. y)np (y) − Si (ξ. y) .p (ξ) ∂u τ (y) = nk (y) = u.18 where: CHAPTER 1. y) u(y) − u(x) − u.66) as: 0 = ∂B Di (x. y)(yp − ξp )] dS(y) ∂B (1. with S = S = ∂B.74) © 2005 by Taylor & Francis Group. y)dS(y) ∂B (1.72) (1.p (x)]np (y)dS(y) Si (x. y)[u.68) yields a simple. i.65) (together with (1.4. BOUNDARY INTEGRAL EQUATIONS Di (x. from inside the body B.71) while use of the linear solution: = u(ξ) + (yp − ξp )u. y)np (y) − Si (ξ. and (1.i (x.p (x) where. fully regularized form of (1.p (ξ)np (y) ∂yk u in equation (1. 3) (1.54) in (1. LLC . Following Rudolphi [143].p (x) np (y)dS(y) Si (x. 2.i (x) = ∂B (1. y)dS(y) = 0 ∂B (1.
77) − lim ξ→x Sijk (ξ. y) [uk (y) − uk (x) − uk.4.74) is analogous to the regularized stress BIE in linear elasticity . i.3.76) Cijkp (∂B) = ξ→x lim Em ∂B kp Dijm (ξ. 1.2. for an outside approach / to the boundary point x . First.75) ξ→x (1. The approach is very similar to that used in Section 1. FINITE PARTS OF HYPERSINGULAR EQUATIONS 19 which is equivalent to equation (1.36). Lutz et al. Cruse and Richardson [39]) are now used in order to determine © 2005 by Taylor & Francis Group. for example. as noted before.p (1. A few comments are in order. using method two in Section 1. equation (17.23)) which appears in Hooke’s law: σ m = Em kp uk. Second. y) [σkp (y) − σkp (x)] np (y)dS(y) Sijk (x. that u(y) ∈ C 1.Rudolphi had only considered a regular boundary collocation point in his excellent paper that was published in 1991). of course.4.3 Stress BIE for 3D Elasticity This section presents a proof of the fact that equation (1. as discussed in the Section 1. valid at an irregular point x ∈ ∂B. (See.equation (28) in Cruse and Richardson [39] .4.39) is a regularized version of (1. y)(yp − ξp )dS(y) ∂B with E the elasticity tensor (see (1.e. provided that the stress and displacement ﬁelds in (1.74) is the same as Rudolphi’s [143] equation (20) with (his) κ = 1 and (his) S0 set equal to S and renamed ∂B.4. Finally.4. also.39) satisfy certain smoothness requirements. Third.54) to regularize (1. y)n (y)dS(y) (1. [89]. y)dS(y) ∂B (1. the result is: σij (x) = ∂B Dijk (x. The ﬁrst step is to apply the FP equation (1. These smoothness requirements are discussed in Section 1. equation (1.36). x can be an edge or corner point on ∂B (provided.4.p (x) where.34)).p (x)(yp − xp )] dS(y) ∂B − − Aijk (∂B)uk (x) + Cijkp (∂B)uk. this equation can also be shown to be valid for the case ξ ∈ B.4. Kane [68].11).78) Simple (rigid body and linear) solutions in linear elasticity (see.2: Aijk (∂B) = lim Sijk (ξ. equation (1.1. With ˆ S = S = ∂B.1.α at y = x . LLC .
such as (1. one has: 0= ∂B Sijk (ξ. 3. using continuity of the integral and comparing with (1. Let ∂Bn . y)(yp − ξp )] dS(y) σij (ξ) = uk. gives A = 0. be denoted by Q. are recommended as starting points.m (ξ).78) yields C(∂B) = E. BOUNDARY INTEGRAL EQUATIONS the quantities A and C. let a source point.m (y) = uk.79) while. 1. An irregular collocation point x for 3D problems is considered next.74) for potential theory or (1. (n = 1. using the linear solution: uk (y) τk (y) = (yp − ξp )uk. [93] state the following requirements for collocating a regularized HBIE.77).. equation (1. y)dS(y) (1.80) = σkm (y)nm (y) = Ekmrs ur.81) Taking the limit ξ → x of (1.79). LLC . .82) Comparing (1.20 CHAPTER 1. Therefore.30) gives: (y) − Sijk (ξ..75) reduces to the simple regularized equation (1.p (ξ).p (x) (1. with coordinates yk . Regularized HBIEs. Using the rigid body mode uk = ck (ck are arbitrary constants) in (1.82) with (1.81) and comparing with (1. Equation (1.54) with a complete exclusion zone.p (ξ) ∂B [Em kp Dijm (ξ. provided that the stress is continuous there. These are: © 2005 by Taylor & Francis Group. obtained by using complete exclusion zones.39) for linear elasticity.4. one has: σij (x) = Cijkp uk.s (ξ)nm (y) in equation (1.39) at an irregular point P ∈ ∂B.76). As is the case in the present work.g. with coordinates xk .39). It has been proved in this section that the regularized stress BIE (28) of Cruse and Richardson [39] can also be obtained from the FP deﬁnition (1.30).4 Solution Strategy for a HBIE Collocated at an Irregular Boundary Point Hypersingular BIEs for a body B with boundary ∂B are considered here. and a ﬁeld point. N ) be smooth pieces of ∂B that meet at an irregular point x ∈ ∂B. Cruse and Richardson [39] have also proved that their equation (28) is valid at a corner point. e. Martin et al. y)n (1. be denoted by P .. (1. equation (1. 2.39) is equation (28) of Cruse and Richardson [39] in the present notation. Also. as before. uk. Taking the limit ξ → x of (1.
from Mukherjee and Mukherjee [111]. referred to as the admissible class. for each n. in practice.39). σ and the traction τ = n · σ in (1.39) at an irregular point P . (ii) (a) The stress σ must be continuous in B. Another important advantage of using these interpolation functions is that ∇u can be directly computed from them at an irregular boundary point [99]. It has proved very diﬃcult. whose exact solutions satisfy conditions (i . [50]).see. Box 1. or is bounded but discontinuous at P (e. [109]. α (iv) [σij (Qn ) − σij (P )]nj (Qn ) = O(rn ) as rn → 0.j (P )[yj (Qn ) − xj (P )] i 21 (1. The discussion in the rest of this book is limited to the class of problems. for example. Mukherjee et al. P ) = ui (P ) + ui. a priori. one should not attempt this collocation if.see. Numerical results from the hypersingular BCM. In principle. without the need to use the (undeﬁned) normal and tangent vectors at this point. (b) The stress σ must be continuous on ∂B. for collocation at an irregular surface point on a 3D body [93]. the exact solution of a boundary value problem must satisfy conditions (iiv) in Box 1. P ) = O(rn ) as rn → 0. that interpolation functions used in the boundary contour method (BCM . The BCM and the hypersingular BCM (HBCM) are discussed in detail in Chapter 4 of this book. for example. The ﬁrst is that.1 Requirements for collocation of a HBIE at an irregular point (from [93]). Also. (1+α) (iii) ui (Qn ) − ui L (Qn . It has recently been proved in Mukherjee and Mukherjee [111]. © 2005 by Taylor & Francis Group. to ﬁnd BEM interpolation functions that satisfy.1. The second issue refers to smoothness requirements on the interpolation functions for u. however. for each n.4. LLC . for example. Mukherjee and Mukherjee [99]) satisfy these conditions a priori. Zhang and Mukherjee [183]). uL (Qn .83) There are two important issues to consider here. FINITE PARTS OF HYPERSINGULAR EQUATIONS (i) The displacement u must satisfy the equilibrium equations in B.1. Qn ∈ ∂Bn .1. at the tip of a wedge . if there is to be any hope for collocating (1. Clearly. these BCM interpolation functions can also be used in the BEM.g. collocated on edges and at corners.see an exhaustive study on the subject in Glushkov et al. and α > 0. rn = y(Qn ) − x(P ). are available in Chapter 4.iv). (ii(b)(iv)) in Box 1. In the above. the stress is unbounded at P (this can easily happen .
e.1 Linear Operators Boundary integral equations can be analyzed by viewing them as linear equations in a Hilbert space. that the error is evaluated at selected points) residualbased error estimates for Dirichlet. relying on the pointwise error measures presented earlier.Chapter 2 ERROR ESTIMATION Pointwise (i. 2. z  = 0 where z ∈ C. Neumann and mixed boundary value problems (BVPs) in linear elasticity are presented ﬁrst in this chapter. Interesting relationships between the actual error and the hypersingular residuals are proved for the ﬁrst two classes of problems. The present analysis excludes domains with corners. Elementbased error indicators. while heuristic error estimators are presented for mixed BVPs. Further details are available in [127].2) © 2005 by Taylor & Francis Group. the space of complex numbers. z ∈ C 1 . 23 k∈Z (2. LLC . Following Sloan [155]. A very readable account of this topic is available in Kress [73]. it is assumed here that the boundary ∂B is a C 1 continuous closed Jordan curve given by the mapping: z : [0. Numerical results for two mixed BVPs in 2D linear elasticity complete this chapter. It is also assumed that any integrable function v on ∂B may be represented in a Fourier series: ∞ ∞ v∼ k=−∞ v (k)e2πikx1 = a0 + ˆ k=1 (ak cos(2πkx1 ) + bk sin(2πkx1 ) (2.1) where i ≡ √ −1 and: 1 v (k) = ˆ 0 e−2πikx1 v(x1 )dx1 . are proposed next. 1] → ∂B.
in the LTB sense. y)vk (y)ds(y) ∂B (Sik vk )(ξ. y)vj (y)dS(y) T Tij (ξ. These terms depend on the smoothness of the boundary at the source © 2005 by Taylor & Francis Group. ERROR ESTIMATION in which Z denotes the space of integers.8) (N ) (Tij vj )(ξ) := (Dijk vk )(ξ) := ∂B Dijk (ξ. y)vj (y)dS(y) ∂B (2. one has: y = Ax ≤ A x x = A−1 y ≤ A−1 y The result now follows by choosing C1 = 1/ A and C2 = A−1 .6) (2. The following Lemma is very useful for the work presented in this chapter.3) (2. From the CauchySchwarz inequality. x) := nj (x) (N ) Sijk (ξ.7) (2.4) (2. Proof: The linearity and continuity of A and A−1 imply that A and A−1 are ﬁnite.24 CHAPTER 2. for example. whereas Tij and Dij are not continuous (Tanaka et al. Returning to the problem at hand. If A : B1 → B2 is a continuous linear operator that has a continuous inverse. y)vk (y)ds(y) Sijk (ξ. x) := nj (x) (N ) Dijk (ξ. Lemma 1. y)vk (y)ds(y) ∂B (Sijk vk )(ξ) := (Dik vk )(ξ. The hypersingular operator Sij gives rise to unbounded terms that vanish when the integral is considered. [162]) and give rise to additional bounded free (N ) terms in the limit. then there exist real positive constants C1 and C2 .5) (2. LLC . and Ax = y. such that: C1 y B2 ≤ x B1 ≤ C2 y B2 where · Bi denotes a suitable norm of the appropriate function (in the Banach space Bi ). y)vk (y)ds(y) ∂B The operator Uij is continuous onto the boundary. the following operators are deﬁned as: (Uij vj )(ξ) := ∂B Uij (ξ.
10). assume a certain regularity of the boundary (for instance no corners or cusps) [155].24) and HBIE (1. © 2005 by Taylor & Francis Group. The main result of this work is that this heuristic idea. the method reduces to ﬁnding a second approximation to the solution by iterating the ﬁrst approximation with the HBIE. This idea is ﬁrst illustrated in the context of two basic cases : the interior Dirichlet and Neumann problems. that key properties of the operators. In this work. the HBIE is collocated only at regular boundary points (where the boundary is locally smooth) inside boundary elements. Remark 1 One should note.2. when stated formally. nevertheless. Using the operators deﬁned above. respectively: BIE : HBIE : ui = Uij τj − Tij uj τi = Dij τj − Sij uj (N ) (N ) (2. leads to a simple characterization of the error. 123. In essence. [96]). it is possible to prove existence and uniqueness of a solution to this BVP [46]. however.2. such as continuity and invertibility. ITERATED HBIE AND ERROR ESTIMATION 25 point x.9) and (2. Mixed boundary conditions are considered thereafter. are too restrictive for the solution of practical engineering problems. however. of course. 2. 2.9) (2. been made here in order to obtain some mathematical understanding of the error estimation process that is described in Section 2. The numerical example problems do contain corners.10) As in the case of potential theory (Menon et al.2.1 Problem 1 : Displacement Boundary Conditions (λ + µ)∇(∇ · u) + µ∇2 u = 0 in B Solve the NavierCauchy equations: subject to the boundary conditions: u = f on ∂B This problem is analogous to the Dirichlet problem of potential theory. Under suitable restrictions on the domain. the BIE (1. The HBIE (2.37) become.10).2. has only been collocated at regular points on the boundary of a body.2 Iterated HBIE and Error Estimation The heuristic idea that is at the heart of the pointwise error estimation procedure described below (see also [122. the LTB of the above integral operators has been used to obtain the singular integral equations (2. Such assumptions have. These assumptions. LLC . 96]) is simple : the amount by which an approximate solution to the BIE fails to satisfy the HBIE is a measure of the error in the approximation.
2. one may show that if the domain B is a circle of radius exp[(1/2)(3 − 4ν)]. 2 (2.9) or the traction HBIE (2.16) τ (2) ei (2) τi respectively. Deﬁne the hypersingular residual to be: © 2005 by Taylor & Francis Group. with δij the • Step 2: Use the traction HBIE (2. 2. • Step 1: Solve the displacement BIE (2. and Sij is hypersingular. will be used for error estimation.12) to iterate the traction and obtain a (2) second approximation τi : τi (2) = Dij τj (N ) (1) − Sij fj (N ) (2. then the BIE (2. For instance. called the HBIE iterate.15) (2. LLC . As in potential theory.11) and (2.11) while the traction BIE gives rise to equations of the second kind (for the traction): 2 τi − Dij τj = −Sij fj =: gi (N ) (N ) (N ) i = 1.26 CHAPTER 2.12). one again encounters the problem of the transﬁnite diameter. Iij fj = δij fj = fi . ERROR ESTIMATION Either the displacement BIE (2.11) for the traction τi : Uij τj (1) (1) = (Iij + Tij )fj (2.11) does not admit a unique solution.1.1 Error estimate for the primary problem Using the two BIEs (2. Let the error (in traction) in the primary solution and iterate be: ei τ (1) = τi = (1) − τi − τi (2. one can formulate an error estimation process that is analogous to the Dirichlet problem in potential theory [96]. since Uij has a logarithmic kernel.13) where I is the identity operator and components of the Kronecker delta.10) may be used to formulate a method of solution for the unknown traction on the boundary. 2 (2. The displacement BIE leads to a system of singular integral equations of the ﬁrst kind for the (unknown) traction: 1 Uij τj = fi + Tij fj =: gi i = 1.14) This approximation.12) (N ) Recall that Uij is logsingular. Tij and Dij are Cauchy singular.
15) (2. Uniqueness of solutions to (N ) the integral formulations implies that the operators (Iij − Dij ) and Uij have continuous inverses [172].17) (2.14) (2) ei = τi = − τi − Sij fj − τi + τj ) − Sij fj − τi (N ) (N ) (N ) (1) Dij τj (N ) (2. Now use Lemma 1.19) © 2005 by Taylor & Francis Group.15) (2.10) = Dij (ej = Dij ej τ (1) (N ) τ (1) so that: ei τ (2) = Dij ej (N ) τ (1) (2.12) is unique. then there exist real positive constants C1 and C2 such that: C1 ri (τ ) ≤ ei τ (1) ≤ C2 ri (τ ) Proof: The continuity of the operators is a manifestation of the elliptic nature of the partial diﬀerential equation (PDE). and suﬃciently smooth data and solutions (as detailed above). ITERATED HBIE AND ERROR ESTIMATION 27 ri One can now show that: (τ ) (2.16) (2.18) Theorem 1 For a suﬃciently smooth domain. if the solution to the integral equations (2.2 Error estimate for the iterate In a manner similar to the previous subsection.11) and (2.1.17) ri = τi = τi − τi (2) (N ) (1) − (Dij τj (N ) − Sij fj ) (N ) τ (1) (N ) = τi − (Dij τj − Sij fj ) + ei = (Iij − Dij )ej (N ) τ (1) − Dij ej (N ) τ (1) so that: ri (τ ) = (Iij − Dij )ej (N ) τ (1) (2.2.2.14) (2.10) (1) (1) (τ ) = τi (1) − τi (2) (2.2. LLC . one can show that: τ (2) (2. 2.
9). ERROR ESTIMATION 2. © 2005 by Taylor & Francis Group.2. and ω = ω k is an axial vector representing a rotation. An elegant practical way to solve traction prescribed problems in linear elasticity is outlined in a recent paper by Lutz et al.10): −Sij uj = gi − Dij gj =: h2 i (N ) (N ) (2.20) where r0 ∈ R2 is a translation.22) one obtains an integral equation of the ﬁrst kind for the displacement.28 CHAPTER 2. to eliminate this arbitrariness. and. [90] where the singular matrix from the BIE is suitably regularized at the discretized level by eliminating rigid body modes. The ﬁrst integral equation formulation for the problem follows from the displacement BIE (2. LLC . It is important to mention again that the solution of the traction prescribed BVP is arbitrary within a rigid body motion.21) (2.2 Problem 2 : Traction Boundary Conditions Solve the NavierCauchy equations: (λ + µ)∇(∇ · u) + µ∇2 u = 0 in B subject to the boundary conditions: t = g on ∂B where the tractions satisfy the consistency conditions of static equilibrium: t ds = 0 ∂B (r × t) ds = 0 ∂B It is known that the solution to the above problem exists. one must work in a restricted function space as has been done before [96] for Neumann problems in potential theory. and is unique up to a rigid body motion (Fung [46]). The space of twodimensional rigid body motions may be characterized as (Chen and Zhou [29]): R := r0 + ω × r (2. One has an integral equation of the second kind for the (unknown) displacement: ui + Tij uj = Uij gj =: h1 i and using the traction HBIE (2.
© 2005 by Taylor & Francis Group. construct an approximation to (1) (2) the displacement ﬁeld. ITERATED HBIE AND ERROR ESTIMATION 2.26) ri = gi − τ i = gi − (Dij gj − Sij uj ) = gi − Dij gj + Sij (uj + ej = Sij ej (N ) u(1) (N ) (N ) u(1) ) so that: ri (τ ) = Sij ej (N ) u(1) (2. Next ﬁnd τi .10) to obtain τi : τi (2) = Dij gj − Sij uj (N ) (N ) (1) (2.23) • Step 2: Use the traction HBIE (2. the error in the displacement is deﬁned as: ei One can now show that: (τ ) (2. [96]. and use it to estimate the error in the primary solution. First.27). ui .24) Deﬁne the hypersingular residual: ri (τ ) = τi (1) − τi (2) = gi − τ i (2) (2. an iterated approximation to the traction.25) (2. LLC .25) Also. • Step 1: Solve the displacement BIE (2.2.2.1 Error estimate for the primary problem 29 The error estimation technique is analogous to the Neumann problem investigated previously by Menon et al.26) (2.2.24) (2. It follows from using equation (2.2.10) (2) (N ) (N ) (1) u(1) = ui (1) − ui (2.21) for the displacement ui : (Iij + Tij )uj (1) (1) = Uij gj (2) (2.27) Theorem 2 The hypersingular traction residual bounds the error in the displacement globally: C1 ri (τ ) ≤ ei u(1) ≤ C2 ri (τ ) Proof: The proof is quite analogous to that of Theorem 1.
10). Iterate (2. [96] 2. and upon rearrangement. LLC . ERROR ESTIMATION The displacement residual In the traction boundary condition problem.2 CHAPTER 2.10) (2. becomes: ui = (Iij − Sij )uj − (Iij − Dij )gj (1) (N ) (N ) (2.2.28) with this solution and deﬁne: ui (2) = (Iij − Sij )uj − (Iij − Dij )gj (N ) (1) (N ) (2.28) Let ui be the solution of the BIE (2. The HBIE (2.29) Deﬁne the displacement residual: ri One can now show that: ri (u) (u) ≡ ui (1) − ui (2) (2. however. the unknown is the displacement but equation (2.27) relates the traction residual to the error in the displacement. It is proved below.26) (2.2. with ui added to both sides of it.31) so that: ri and ri (τ ) (u) = ri (τ ) (2. that the traction residual is also equal to a suitably deﬁned displacement residual for this problem.29) (1) − ui (1) (2) (N ) (1) (N ) = ui − (Iij − Sij )uj + (Iij − Dij )gj (N ) (N ) = Sij uj + gi − Dij gj (2.30) ≡ ui (2.30 2.9).27) (N ) (1) = Sij uj − Sij uj N = Sij ej u(1) (N ) (1) = ri (τ ) (2.2.3 Problem 3 : Mixed Boundary Conditions The general boundary value problem in linear elasticity is: Solve the NavierCauchy equations: © 2005 by Taylor & Francis Group.32) can be replaced by ri (u) in Theorem 2 ! Remark 2 An analogous result in potential theory appears in Menon et al.
The primary BIE (2. The boundary stresses σij are next obtained from the boundary values of the tractions and the tangential derivatives of the displacements. ITERATED HBIE AND ERROR ESTIMATION 31 (λ + µ)∇(∇ · u) + µ∇2 u = 0 in B subject to the boundary conditions: Au + Bt = f on ∂B where the matrices A and B and the vector f are prescribed quantities.9) is solved ﬁrst.34) where the required operators are deﬁned in equations (2.1 Traction residual (1) One computes the traction components τj (2.3 of this chapter.34) is collocated only at regular boundary points (where the boundary is locally smooth) inside boundary elements.10).2.33) The corresponding pointwise error measure is as follows.9) and then obtains the HBIE iterate the traction residual is deﬁned as: ri (τ ) on ∂B by solving the primary BIE from the HBIE (2. the numerical examples presented in Section 2.34). however. together with Hooke’s law. Mukherjee [98] or Sladek and Sladek [151]). Also. In this case. This class of problems is the most commonly encountered one in linear elasticity. Next.2 Stress residual The stress residual is another important quantity in this work.3.6) and the LTB of the above operators are used in equation (2. if the traction is speciﬁed in one direction.4 of this chapter all have mixed boundary conditions imposed upon them. As before. then the error in the boundary data is the error in displacement in the ﬁrst direction. equation (2. At a ﬁxed boundary point.2. 2. LLC . a heuristic approach to error estimation is adopted here. This yields the boundary tractions and displacements (1) (1) (1) τj and uj .5) and (2.2. 2. This issue is discussed further in Section 2. This is a wellknown procedure in the BIE literature (see. (2) (2) τj = τi (1) − τi (2. for example. and the error in traction in the second direction. the iterated boundary stress is obtained from the HBIE (1. and the displacement in the other. © 2005 by Taylor & Francis Group. Indeed.2.36) as follows: σij = Dijk τk − Sijk uk (2) (1) (1) (2.3.
32 CHAPTER 2. The following are proposed : the ﬁrst based on the traction residual and the second on the stress residual: ηj (τ ) := ri (τ ) L2 (∂Bj ) (2. respectively. C2 . deﬁned above in equation (2. These indicators should satisfy the following criteria: C1 ηi N ≤ e A(∂Bi ) 2 A ≤ C2 ηi N (2. These pointwise error measures may be used to deﬁne element error indicators. as: eij eij s(1) = σij − σij = σij − σij (2) (1) (2. It is often diﬃcult to prove these properties analytically. and one usually takes recourse to numerical experiments. this method leads to two natural error indicators.38) 2 ηi D1 i=1 2 ηi ≤ e ≤ D2 i=1 (2.3 ElementBased Error Indicators The main objective of error estimation is the development of suitable element error indicators. not a vector.41) ηj (s) := rk (s) L2 (∂Bj ) Note that the subscript j refers to the jth element .17). if desired. D1 and D2 are appropriate constants.40) (2. which are denoted by ηi . the second is based on the stress residual deﬁned in equation (2. which are special cases of problems with mixed boundary conditions.35) (2.36) s(2) and the stress residual is deﬁned as: rij = σij − σij (s) (1) (2) (2. can also be used for problems with displacement or traction boundary conditions. As in the potential theory case [96].37). The ﬁrst is based on the traction residual deﬁned in equation (2.39) where A is a suitable norm.37). LLC . for the BIE and the HBIE iterate. © 2005 by Taylor & Francis Group. A(∂Bi ) denotes the restriction of this norm to the ith element. The L2 norm is used for convenience and other norms can be used.37) Remark 3 The stress residual. and C1 . ERROR ESTIMATION One now gets the error in stress. 2.the error indicator is a scalar.
the traction residualbased element error indicator will capture the L2 norm of these errors on an element. Singular integration is avoided using the rigidbody mode. the traction may be prescribed in the x1 direction and the displacement in the x2 direction at a boundary point. NUMERICAL EXAMPLES 33 The error estimates. therefore.e. in order to determine the components of stress using the traction HBIE. in the displacement in the x1 direction. Note that. the HBIE). except on elements that contain the collocation point. the BIE). The numerical implementation consists of two modules: a standard code for twodimensional elastostatics.4 Numerical Examples Two basic problems from the theory of planar elasticity are considered in this section. on a particular element. This code uses collocation with quadratic isoparametric elements. Theorems 1 and 2). at most two components of the stress are known from the prescribed boundary conditions. i. and a set of routines that calculate the hypersingular residual for error estimation.2. Thus. collocation is carried out at points on the boundary where it is locally smooth. In general. The stress residual is also used as a measure of the error in stress on the boundary.2 that the traction residual is equal to the displacement residual for a traction prescribed boundary value problem. diagonal terms are evaluated by summing the oﬀdiagonal terms. and in the traction in the x2 direction. © 2005 by Taylor & Francis Group. and then a quadratic polynomial is employed to approximate the stress components over each element. The numerical method used for evaluation of the necessary hypersingular integrals here is due to Guiggiani [60]. At any boundary point in a 2D problem. the actual error will depend on the boundary conditions..36) at three boundary points inside a boundary element. the stress tensor at a boundary point is evaluated by using the HBIE (1.e. as deﬁned above. a code due to Becker [9] is employed. In the following examples. do not depend directly on the boundary conditions on an element. For the second part of the code (i.4) suggest that this error is eﬀectively tracked by the stress residualbased error indicator. but no explicit information about the displacement. even for mixed boundary value problems. 2. there is always some error in a computed stress tensor at a boundary point. Numerical integration is done using Gaussian quadrature. The traction residual has been shown to be related to the pointwise error in the boundary unknowns for Dirichlet and Neumann problems in elasticity (Section 2. it has been proved in Section 2. even though the traction residual uses the diﬀerence in the primary and iterated tractions.e. for instance.2. In mixed boundary value problems. LLC . at a local level. and which are inside boundary elements. Ideally. The errors are. Numerical experiments presented below (Section 2. For the ﬁrst part (i.4.
1 Example 1: Lam´’s Problem of a ThickWalled Cyline der under Internal Pressure Consider an inﬁnitely long hollow cylinder subjected to an internal pressure p = 1. are deﬁned in (2. Unlike the numerical examples in potential theory [96].0.35 .3. Symmetry is employed and the problem is formulated as a mixed boundary value problem on a quarter of the cylinder. Note that the continuous line for the exact solution is just used as a matter of convenience and does not have any meaning except at discrete points. ERROR ESTIMATION 2. E = 1. (b) e Comparison of analytical and numerical solutions for σ11 (from [127]) Figure 2.5 0 0.1(a). for the BIE and the HBIE iterate. The mesh used to solve this problem is shown in Figure 2.5 1 0 5 10 15 20 25 30 35 40 Collocation point (a) (b) Figure 2. 2 1. a comparison between the pointwise error and stress residual in each direction is carried out here. and the outer radius ro = 6. LLC . Since the stress residual is a symmetric tensor with three independent components.1: (a) Discretized domain for the Lam´ problem with 12 elements. it is seen that the residual is not an upper bound as it underestimates the error at some points. Consistent units are assumed throughout this paper.1(b) presents a comparison of the computed and analytical solutions for the stress component σ11 at the collocation points used for the HBIE. Material properties are also chosen of O(1). The error in stress. More importantly. the hypersingular residual and error are often of opposite signs in this elasticity example. since the xaxis is the collocation point number.2. The inner radius ri = 3. and the stress residual. The results for σ12 and σ22 display similar accuracy and are not shown here.34 CHAPTER 2. and Poisson’s ratio ν = 0.5 __ 1 o:stress from HBIE +:stress from BIE :exact solution σ 11 * : BIE collocation nodes o : HBIE collocation nodes 0. © 2005 by Taylor & Francis Group. Also. a pointwise comparison between the absolute value of the error and pointwise measurements of the hypersingular residual is considered next. Notice that the mesh is not biased a priori in the sense that the element density is not increased on parts of the boundary where the error is expected to be high.4.37). namely Young’s modulus.
2: Absolute values of error in the BIE solution. A simple way to do this is to use nondimensional quantities to begin with. and in σ12 in Figure 2. On the other hand.05 0.04 0.41)) is compared to the elementbased L2 norm of the error in stress.3(a). the performance of the two indicators ηj and ηj deﬁned in equations (2.1 0. (τ ) (s) Remark 4 The comparison between the traction residual and error in the displacement is diﬃcult unless one uses normalized values. error in the HBIE iterate.g.03 0. LLC . This is the approach followed in this work. The ﬁrst is a traction residualbased error indicator. All variables are unscaled (from [127]) Now consider a comparison of absolute values of errors.06 0.1 0. The stress residuals are seen to capture the error trends quite eﬀectively.09 0.40) and (2. In particular.07 0. in adaptivity) is the performance of ele(τ ) ment error indicators. and the hypersingular residual. for (a) σ11 and (b) σ12 .05 0. on all the boundary elements.04 0.08 0.01 0 0 5 10 15 20 25 30 35 40 5 10 15 20 25 30 35 40 Collocation point Collocation point (a) (b) Figure 2.03 0.08 35 Various quantities related to σ 11 0. respectively. in Figure 2. © 2005 by Taylor & Francis Group. Of most practical importance (e.40)) is compared with the elementbased L2 norm of the error in the unspeciﬁed boundary data in Figure 2.01 0 0 Various quantities related to σ12 o: error in the BIE solution +: error in the HBIE iterate __ : hypersingular residual __ : o: error in the BIE solution +: error in the HBIE iterate hypersingular residual 0.3(b). and the second uses the stress residual.07 0. the element (s) error indicator based on the stress residual (ηj from equation (2.06 0.2.2(b).09 0. The element error indicator based on the traction residual (ηj from equation (2.41).2(a).4. NUMERICAL EXAMPLES 0. and the residual in σ11 in Figure 2.02 0. is studied here. It is seen that the stress residual provides good pointwise tracking of the error on a relatively coarse mesh.02 0.
Of course.6(b) uses errors in stress components. only a quarter of the plate is used in the computer model. The stressbased error indicator performs better and accurately captures the error in stress on most of the elements. and the residual.035 CHAPTER 2.6).03 0.06 Error indicator and error element error indicator element error indicator 0. Figure 2. together with the corresponding stress residuals.4.03 0. the error in BIE stress component is zero at points where that particular stress component is prescribed as a boundary condition. the error estimation method presented here has the advantage of capturing errors in the stress ﬁeld.04 0.005 0 0 Error indicator and error 0.01 L norm of error 0 0 2 4 6 8 10 12 2 Element number Element number (a) (b) Figure 2. in some components of the stress. subject to a traction τ1 = 1 at inﬁnity. The more important comparison is between the computed element error indicators and the L2 norms of the error on each element (Figure 2. ERROR ESTIMATION 0.07 0.05 0.5(a) and 2.025 0. The displacement and stress ﬁelds for this problem may be found in Timoshenko and Goodier [167]. and (b) stress residualbased error with error in boundary stresses (from [127]) 2. Figures 2.01 0.04 0. the usual residual © 2005 by Taylor & Francis Group.36 0. Pointwise comparisons between the errors (in the BIE solution and in the HBIE iterate).2 Example 2: Kirsch’s Problem of an Inﬁnite Plate with a Circular Cutout Consider an inﬁnite plate with a circular cutout of radius 1. On the other hand.015 0. for (a) traction residualbased error with error in the unspeciﬁed boundary data.02 0. are presented ﬁrst. Using symmetry.5(b) show the absolute values of these errors in two stress components. It is seen that the tractionbased error indicator underestimates the error on some elements. In conclusion.3: Comparisons of element error indicators with actual errors. the boundaries of a ﬁnite square domain are subjected to tractions computed from the exact solution of an inﬁnite plate subjected to traction at inﬁnity. The mesh is shown in Figure 2.02 L2 norm of error 2 4 6 8 10 12 0.6(a) uses errors in the unspeciﬁed boundary data while Figure 2.045 0. In order to simulate this problem with a ﬁnite geometry.4. Material properties are the same as in the previous example. LLC .
5 __ +: error in the BIE solution o : error in the HBIE iterate : hypersingular residual 0.4: Discretized domain for the problem of a plate with a cutout with 10 elements. NUMERICAL EXAMPLES 37 *: BIE o: HBIE collocation nodes collocation nodes Figure 2.6 0. error in the HBIE iterate.1 0 0 5 10 15 20 25 30 0 0 5 10 15 20 25 30 Collocation point Collocation point (a) (b) Figure 2.2 0.2 0.1 0.35 0.4.2.15 0.3 0.5: Absolute values of error in the BIE solution.3 Various quantities related to σ 12 0. All variables are unscaled (from [127]) © 2005 by Taylor & Francis Group. Plate side is 4 units and cutout radius is 1 unit of length (from [127]) 0. and the hypersingular residual for (a) σ12 and (b) σ22 .05 0.4 0. LLC .25 Various quantities related to σ22 +: error in the BIE solution o: error in the HBIE iterate __ : hypersingular residual 0.
6: Comparisons of element error indicators with actual errors. for (a) traction residualbased error with error in the unspeciﬁed boundary data.15 0.i. © 2005 by Taylor & Francis Group. and (b) stress residualbased error with error in boundary stresses (from [127]) techniques may only be used to compute the error in displacements.4 Error indicator and error 0.3 CHAPTER 2. On physical grounds. ERROR ESTIMATION 0.25 o: element error indicator +: L2 norm of error 0.2 0. and its gradient have converged to the actual solution in this case.2 0.e.6 0. a measure of the error in stress is preferable to a measure of displacement error.5 o: element error indicator 2 +: L norm of error Error indicator and error 0. It may also be viewed as a stronger measure of convergence .1 0. LLC .38 0.05 0. the approximate displacement ﬁeld.1 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 Element number Element number (a) (b) Figure 2.3 0.
cooling 39 © 2005 by Taylor & Francis Group. Sometimes. encompassing microelectromechanical. or captured by a hubcap structure and set into motion by mechanical. for applications in microelectromechanical systems (MEMS). electrical. Further details are available in [6]. microﬂuidic.1 Introduction to MEMS The ﬁeld of microelectromechanical systems (MEMS) is a very broad one that includes ﬁxed or moving microstructures. thermal.LEFM) in 3D linear elasticity.for mixing. Owing to its small size. This is followed by a discussion of crack problems (linear elastic fracture mechanics .Chapter 3 THIN FEATURES Many boundary value problems that are solved by the BEM involve thin features. Common examples are cracks and thin plates and shells. that deforms when subjected to electric ﬁelds.1. MEMS structural elements are plates. 3. are presented in this section. This chapter ﬁrst presents BEM formulations for 3D potential theory in a region exterior to thin plates. signiﬁcant forces and/or deformations can be obtained with the application of low voltages (≈ 10 volts). Typical MEMS structures consist of arrays of thin beams with crosssections in the order of microns (µm) and lengths in the order of ten to hundreds of microns. acoustical or photonic energy source(s). MEMS usually consists of released microstructures that are suspended and anchored. Numerical results for a simple system with two thin conducting plates follow. suitable for applications in MEMS.1 Exterior BIE for Potential Theory: MEMS Exterior BIEs for potential theory. 8] . 3. An example is a small rectangular silicon plate with sides in the order of mm and thickness of the order of microns. LLC . Examples of devices that utilize vibrations of such plates are synthetic microjets ([142. microoptoelectromechanical and microthermalmechanical devices and systems.
together with a BEM code FastCap [112] for the electric ﬁeld analysis. suitable for MEMS analysis of moderately thick plates. has also been developed in this work. 148. [118. Other examples of such work are [49. to take care of singular and nearly singular integrals that arise. the capacitance of a parallel plate capacitor.for h/L ≤ .the usual BEM. are ﬁrst presented and regularized. 149] for dynamic analysis of MEMS. A similar approach has also been developed for MEMS and nanoelectromechanical systems (NEMS) with very thin beams [7]. however. LLC . 180] is quite diﬀerent from the problem under consideration here.in terms of the length L (of a side of a square plate) and its thickness h. Numerical simulation of electrically actuated MEMS devices have been carried out for around a decade or so by using the BEM to model the exterior electric ﬁeld and the FEM (see. e. is of great value to the research reported here. The work of Liu [87]. an enhanced BEM. as well as [147.001 .40 CHAPTER 3. Careful regularization of the gradient equation. The formulation given in the present work is a BEM scheme that is particularly well suited for MEMS analysis of very thin plates . Gray addresses applications in electroplating problems. The standard BIE with a weakly singular kernel is used here and this approach works well for determining. Numerical results are presented and discussed for a model problem (a parallel plate capacitor) from three methods . 1]. Gray [53] and Nishimura and his coworkers [118. © 2005 by Taylor & Francis Group. for example. microspeakers [71] etc. micropropulsion and ﬂow control). but this work is not presented in this book. on thin shells. in an inﬁnite region exterior to a structure composed of thin conducting plates. 66]) to model deformation of the structure. both weak and strong. It is shown that accurate evaluation of weakly singular and nearly weakly singular integrals plays a key role here. The usual and gradient BIEs for potential theory. The gradient BIE is employed in the present work to obtain these charge densities separately. The focus of this section is the BEM analysis of the electric ﬁeld exterior to very thin conducting plates. The commercial software package MEMCAD [147]. [179. A convenient way to model such a problem is to assume plates with vanishing thickness and solve for the sum of the charges on the upper and lower surfaces of each plate [61]. This section is organized as follows. 180] have considered the 3D Laplace equation in a region exterior to a narrow slit or crack. Singular and nearly singular integrals. while the separate charge densities (that are proportional to the normal derivative of the potential) are of interest in this chapter. the enhanced BEM and the thin plate BEM. uses the commercial FEM software package ABAQUS for mechanical analysis. Their primary interest is in the crack opening displacement with zero normal displacement derivative on the crack faces.g. The research described in Refs. one must obtain the charge densities separately on the upper and lower surfaces of a plate since the traction at a surface point on a plate depends on the square of the charge density at that point. THIN FEATURES of electronic components. 190. for example. is the principal contribution of the present work. As a byproduct of the development of the thin plate BEM. are discussed next. For MEMS calculations.
2 Gradient BIE .1.2. one has the usual indirect BIE: φ(ξ) = ∂B ν(y) ds(y) 4πr(ξ. It becomes equal to the charge density when B is the inﬁnite region exterior to the conductors.3. LLC . φ(x) prescribed for x ∈ ∂B (3.1) where φ is the potential. Of interest is the solution of the following Dirichlet problem for Laplace’s equation: ∇2 φ(x) = 0. for a source point ξ ∈ B (with bounding surface ∂B).2) as: ∂φ (ξ) = ∂ξk ν(y)(yk − ξk ) ds(y) 4πr3 (ξ. y) (3.3.1. This is discussed in Section 3.2) Alternatively. The unit normal n to ∂B is deﬁned to point away from B (i.e. one can write (3.1 Usual BIE . r = r.3 BIES in Inﬁnite Region Containing Two Thin Conducting Plates Now consider the situation shown in Figure 3. 3. the function ν(y) is not the charge density.1.1.1. y) ds(y) 4πr3 (ξ. r(ξ.1.2. 3.2 Electric Field BIEs in a Simply Connected Body First consider the solution of Laplace’s equation in a threedimensional (3D) simply connected body. © 2005 by Taylor & Francis Group.3) ∂B Note that.indirect formulation Taking the gradient of φ at the source point ξ results in: ∇ξ φ(ξ) = ∂B ν(y) ∇ξ 4π 1 r(ξ.1.4) where B is now the region exterior to the two plates. EXTERIOR BIE FOR POTENTIAL THEORY: MEMS 41 3. into a plate).indirect formulation Referring to Figure 1. y) ds(y) = ∂B ν(y)r(ξ. in general. 3. y) = y − ξ.1. x ∈ B. y) (3. is the dielectric constant of the medium and ν is the (unknown) surface density function on ∂B. y) (3.
σ~ x+ σ + s+ 2 n s2 plate 2 Figure 3.5) is suﬃcient to solve for β on both the plates ! 3. k = 1. where σ is now the charge density at a point on a plate surface. y) ˆ S1 + β(y) dS(y) 4πr(x+ .6) © 2005 by Taylor & Francis Group. for this case. As ξ → x+ ∈ S1 ⊂ S1 (see Figure 3.1. Therefore.2 + Gradient BIE . 2: The governing equation.5) + + S2 β(y) dS(y) 4πr(x+ . this is not necessary since β(y) is equal and opposite on the two plates.source point approaching a plate surface S1 + + ˆ+ ˆ+ Let S1 ⊂ S1 be a small neighborhood of x+ . one has: φ(x+ ) = + ˆ+ S1 −S1 β(y) dS(y) + 4πr(x+ . while the rest are usually regular. however.1).3.1. The second integral above is weakly singular.V1 . For the case shown in Figure 3.3. + A similar equation can be written for x+ ∈ S2 . however. It should be noted.1. σ(x) = ∂φ (x) = n(x) · [∇ξ φ(ξ)]ξ=x ∂n (3. y) Here β(y) = σ(y+ ) + σ(y− ). THIN FEATURES ξ V1 ^+ s 1 h g x σ+ σ + s+ 1 n s1^+ s 2 plate 1 ground plane V =0 x .1: Parallel plate capacitor with two plates (from [6]) 3.1 + Regular BIE . that the last integral above becomes nearly weakly singular when both h and g are small. LLC . y) (3.42 CHAPTER 3.source point approaching a plate surface S1 + − It is ﬁrst noted that for x+ ∈ Sk ∪ Sk . equation (3.
One has: σ(x+ ) = + ˆ S1 −S1 + β(y)r(x+ . x+ ) = = + [1 − cos(ψ(θ))]dθ 0 (3. y) · [β(y)n(x+ ) − β(x)n(y)] dS(y) 4πr3 (x+ . y) (3. y) · n(x+ ) dS(y) 4πr3 (x+ . y) = + ˆ S1 −S1 + © 2005 by Taylor & Francis Group. y) 2π ˆ Ω(S1 . EXTERIOR BIE FOR POTENTIAL THEORY: MEMS 43 + ˆ+ Consider the limit ξ → x+ ∈ S1 ⊂ S1 . the solid angle subtended by the surface element S1 at the + point x is (see [87.1. y) n Z axis θ L+ s1+ y Figure 3. y) · n(x) dS(y) 4πr3 (x+ . x+ ) + 4π β(y)r(x+ . since the charge density is singular on its edges. 105] and Figure 3. Equations (3. It is important to realize that this limit is meaningless for a point x on the edge of a plate. y) · n(x+ ) dS(y) 4πr3 (x+ . LLC . y) · n(y) dS(y) = ˆ+ 3 + S1 r (x .8) give the ﬁnal equation: 1 [σ(x+ ) − σ(x− )] 2 β(y)r(x+ . y) + + ˆ+ S1 r(x+ .2: Line integral for evaluation of solid angle (from [6]) ˆ+ In the above.8) where the symbol = denotes the Finite Part (FP) of the integral in the sense of Mukherjee [101.3.7) and (3. y) + S2 β(x) ˆ+ Ω(S1 . 102].7) ξ x+ φ ψ r(x + .2): r(x+ .
In the above. y. of course. LLC . the second integral on the righthand side is weakly singular. The existence of the second plate in Figure 3. 3.1 is the reason for (in general) σ(x+ ) = σ(x− ). (3.9) vanish. ψ ˆ is the angle between the positive z axis and r(x+ . and one is left with the simple equation: 1 2 [σ(x+ ) − σ(x− )] = + S2 β(y)r(x+ .3 Two plates very close together For cases in which the gap g between the thin plates in Figure 3. r → g + h.9) are valid for thin curved shells as well as for ﬂat plates. this integral should be written as: β(y) dS(y) = 4πr(x+ . The last integral above. y) 1 ˆ S2 + β(y) − β(˜ ) x β(˜ ) x dS(y) + + . y) · n(x+ ) dS(y) (3. x− ∈ S2 and β(˜ ) = σ(˜ + ) + σ(˜ − ). y) 4πr(x 4π r(x+ . y) · n(x+ ) dS(y) 4πr3 (x+ . is true. The ﬁrst and second x x x integrals on the righthand side of (3. y) 2π − β(x) 4π cos(ψ(θ))dθ + 0 + S2 β(y)r(x+ . a local coordinate system (x. z) is set up with the ˆ+ origin at x+ such that the positive z axis intersects the surface S1 .1. y) in the xy plane. As r → 0.2). y) Here (see Figure 3. First. Now. For a pair of symmetric ﬂat plates (see Figure 3.1). Once β is known on both plates. y) dS(y) (3.5) must be treated as nearly weakly singular. y) (3.11) ˆ+ ˜ ˆ− ˜ where x+ ∈ S2 . (The second integral is ˜ O(˜/r) where r = y − x+ . so that this integrand r ˜ ˜ © 2005 by Taylor & Francis Group.1 is also of the order of the (small) plate thickness. however.3. the last integral on the righthand side of equation (3. as a postprocessing step.10) implies that σ(x+ ) = σ(x− ) if one has only one plate. becomes nearly strongly singular if both the thickness h and the gap g are small.10) Equation (3. it should be emphasized that equations (3. This. y) with y ∈ L+ . to obtain σ + and σ − on both plates. THIN FEATURES + r(x+ . while the rest are usually regular.9) can be used. In this case. the ﬁrst.9) for ﬂat plates. and θ the + angle between the positive x axis and the projection of r(x . y) · [β(y)n(x+ ) − β(x)n(y)] dS(y) 4πr3 (x+ .11) are regular. second and third integrals on the righthand side of (3. Equation (3. y) + ˆ S2 + + S2 + ˆ+ S2 −S2 β(y) dS(y) 4πr(x+ .9) 4πr3 (x+ .5) and (3.44 + ˆ S1 CHAPTER 3.
can be evaluated as follows. A procedure for accurate evaluation of nearly singular integrals is presented in Section 3.) The last integral is nearly singular. ~+ © 2005 by Taylor & Francis Group. J3 . x+ ) 4π (3.13) denotes a “nearly FP” integral. y) J = + ˆ+ S2 −S2 + r(x+ .12) be called J1 .3. ~+ s2 + n(x ) Figure 3.3: Symmetric deformation of two plates (from [6]) Let the integrals on the righthand side of (3. y) 2π [1 − cos(ψ(θ))]dθ 0 (3.9) now becomes nearly strongly singular.12) where (see Figure 3. .12) with β(˜ ) has been inspired by earlier work x on evaluation of nearly singular integrals [104]. One can write: β(y)r(x+ . The fact that the second integral J2 is regular can be proved as follows.1. in this case. y) + 4πr ˆ S2 β(˜ ) x ˆ+ + Ω(S2 . y) · n(x+ ) dS(y) 4πr3 (x+ .1. This integral. the last integral on the righthand side of (3. the point x+ is slightly above S2 and that the second term in (3.13) ˆ+ It is noted that. Also.4. x . y) · [β(y)n(x+ ) − β(˜)n(y)] x dS(y) 3 (x+ . Each of the three integrals is regular. The idea of regularizing (3. n(x ) + x + s1 ground plane V=0 + y . J2 . EXTERIOR BIE FOR POTENTIAL THEORY: MEMS 45 → 0. called J.2): + ˆ Ω(S2 . x+ ) = ˆ+ S2 r(x+ . LLC . y) · n(y) dS(y) = r3 (x+ .
k are Cartesian unit vectors and a and b are some numbers. A new simple approach for evaluation of nearly weakly singular integrals is presented below. In this limit.4 Singular and Nearly Singular Integrals Certain BEM integrals require special care for thin plates and when thin plates come close together. Their equations.many of these references are available in [165. as expected. as well as weakly and strongly singular integrals. 26]).g. its lefthand side becomes σ(x− ). second and fourth terms on its righthand side change (since n(x− ) = + − −n(x+ ) and n(y) for y ∈ S1 equals −n(y) for y ∈ S1 ). can be eﬀectively evaluated by employing a cubic polynomial transformation due to Telles [165] and Telles and Oliveira [166]. equation (3.15) n(˜+ ) ∝ −k − if. Referring to equation (3. Equation (3. even after deformation (see Figure 3.9) when col− located at x− on S1 (see Figure 3.7).5)) and nearly strongly singular (the last integral on the righthand side of (3. 166. ˜ As y → x+ .1. it is easy to show that. along with other nearly singular integrals of various orders. r → 0. © 2005 by Taylor & Francis Group.2 . 104] and are not repeated here in the interest of brevity. [64. are of the form x3 = ±f (x1 . y) ∝ −k.4 Consistency check It is interesting to examine the forms of equations (3. The usual BEM must deal with weakly and nearly weakly singular integrals in such cases.1 − jf. One now has: n(x+ ) ∝ −k + if. Several other authors have also considered similar problems (e.16) Using these facts. LLC . e. THIN FEATURES It is assumed that the two plates always remain symmetric with respect to the ground plane. ˜ As y → x+ . r(x+ .46 CHAPTER 3. 114.14) where i.9) remains unchanged when collocated at x− ! 3.1 + jf.5) yields φ(x+ ) = φ(x− ). The weakly singular case involving this kernel has been addressed before by many authors (see.3.1).5) and (3.2 x (3. ˜ ˜ 3. therefore. so that the integrand of J2 → 0.g. 104]) . while the thin plate BEM must deal with both nearly weakly singular (the last integral on the righthand side of (3. The nearly (also called quasi) weakly singular case.8)) now becomes: ˆ Ω(S1 . j. so that the integrand of J2 is O(˜/r2 ) where r ˜ r = y − x+ .3).1. [114. The solid angle expression (see (3. r → g + h. x2 ). and the signs of the ﬁrst.9)) integrals. [β(y)n(x+ ) − β(˜)n(y)] x ∝ k[β(˜ ) − β(y)] + [ia + jb][β(˜ ) + β(y)] x x (3. x− ) = 0 − 2π [1 + cos(ψ(θ))]dθ (3.
The performance of the new method is compared with that of standard Gauss integration. The weakly singular integral is evaluated by employing the methods outlined in [114. Figure 3.a weakly singular integral on the boundary element ∆ on the top face of the plate that contains x. Numerical results (for σ = 1 and = 1) appear in Figure 3. EXTERIOR BIE FOR POTENTIAL THEORY: MEMS x x h 47 α ˆ x r ˆ r y α ˆ x ˆ r r y h (a) (b) Figure 3. since h is ˆ small.17) The integrand above is multiplied by r/ˆ with the result: ˆ r I(x) = ˆ ∆ [σ(y)(ˆ/r)]ds(y) r 4π r(ˆ . Therefore. The triangle is purposely chosen to be fairly elongated.18) ˆ ˆ is weakly singular.5(b).4(a). 26].18) ˆ Since r/r is O(1) and → 0 as y → x (i. of O(1/ˆ) as r → 0. and.4: Nearly weakly singular integrals (from [6]) Proposed new method for evaluating nearly weakly singular integrals.4(b).18) can be r ˆ evaluated by employing the methods described in [114.1. The new method is seen to take care of these nearly weakly singular integrals very well. as r → 0). LLC . Two kinds of singular (O(1/r)) integrals arise . the integral (3. Performance of new method for nearly weakly singular integrals. even with 19 Gauss points.3.e. Consider a source point x on ˆ the top face of a plate and its image point x on the bottom face in Figure 3.18) can be applied in this case as well. the integrand in (3. The same idea (3.11)) has the form: I(x) = ˆ ∆ σ(y)ds(y) 4π r(x. standard Gauss integration. a nearly weakly singular integral on the boundary element ∆ (the image ˆ of ∆) on the bottom face of the plate that contains x. The source point x may also lie on a side face of the plate (in the 3D BEM) as shown in Figure 3.5(a) shows the source point and region of integration (a triangle). A typical situation where one encounters weakly and nearly weakly singular integrals in the usual BEM is shown in Figure 3. It is seen that for h < 1/100. © 2005 by Taylor & Francis Group.4. A nearly weakly singular integral (see above and also the last term on the righthand side of (3. cannot reduce the error below around 4%. y) (3. 26]. y) ˆx (3. even for very small values of h.
2 π/2 Result from Maple 19 Gauss points 0 _ 0. 6 0. 4 1. x2 x1 10 10 10 10 10 1 Out of plane distance h (b) Figure 3.5: Nearly weakly singular integral over a triangle. 8 1.0) .0) (0.2 10 _6 10 _5 10 10 10 Out of plane distance h _4 _3 _2 _ 10 1 1 Figure 3. 4 0. (a) Schematic (b) Errors in numerical results from various methods (from [6]) 1.y n(y) (a) (0.0. LLC .6. x3 x h (2. 6 Nearly strongly singular integral 1.6: Nearly strongly singular integral (from [6]) © 2005 by Taylor & Francis Group. THIN FEATURES 8 .h) Percentage error 7 6 5 4 3 2 1 0 _ 10 6 _5 _4 _3 _2 _1 : 7 Gauss points : 19 Gauss points : New method .48 CHAPTER 3. . 8 0. 2 1 0.0.
The usual BEM is the standard version with weakly singular integrals evaluated by the method outlined in [26]. The enhanced BEM is designed to solve problems with moderately thick plates accurately and eﬃciently. This procedure. LLC .13)). composed of T 6 triangles. the enhanced and the thin plate. regular T 6 elements are used on square plate faces. only uniform BEM meshes. the usual BEM only uses T 6 elements in this work. These are brieﬂy described below. © 2005 by Taylor & Francis Group. in all three types of BEM models. for a total of 768 boundary elements. The pathetic result of applying standard Gauss integration (even with 19 Gauss points) appears in Figure 3. Usual BEM. subjected to unit surface potential. EXTERIOR BIE FOR POTENTIAL THEORY: MEMS 49 Nearly strongly singular integrals. The usual BEM code is ﬁrst veriﬁed by solving for the charge distribution on the surface of a unit cube. in the interest of standardization.28 while that from FastCap [113] is 8.1 Numerical Results Three BEM Models Three BEM models. It is well known that the charge density is singular on the edges of a plate.1. are of concern to this work.5(a). The capacitance (the total charge on the cube surface divided by the voltage) (with = 1) from the BEM is 8.13)): r(x.6. one encounters nearly strongly singular integrals of the form (see (3. y) ∆ ∆ where the second expression above is the special version for the situation depicted in Figure 3.4 above. The BEM mesh for this problem has an 8 × 8 array of squares on each face. Like the usual BEM. makes the BEM code somewhat cumbersome and expensive. I= 3. are used on the square faces of objects in this work. A Stoke transformation to convert this nearly strongly singular surface integral to a line integral [87] is recommended in this work (see (3. The correct results in this ﬁgure are obtained from the symbolic computational code Maple [65]. y) · n(y) h ds(y) = ds(y) (3.1.5 3. however. each divided into two T 6 triangles.3. For the sake of simplicity and consistency between diﬀerent BEM models. When two plates get close together. The enhanced BEM has two additional features: • Nearly weakly singular integrals are evaluated by the method outlined in Section 3. with collocation points placed everywhere including on plate edges. Therefore. the usual. weakly singular integrals are again evaluated by the methods outlined in [26].3. For simplicity.19) r3 (x. Enhanced BEM. nonconforming boundary elements should be used and collocation points should not be placed at plate edges.1.1. It is absolutely critical that this integral be evaluated carefully. y) r3 (x. Further.5.
THIN FEATURES • Detailed modeling of side faces for small values of h/L is not desirable for two reasons. Collocation points. The second reason is a matter of eﬃciency.6. and the thin plate BEM only on the upper surface. 3. Of course. symmetry can be used.the usual.1 Capacitance evaluation from the thin plate BEM model Results from the thin plate BEM model (with h/L = 10−6 ) are compared to those from Harrington [61] in Figure 3.8) and (3. One reason is that this would either lead to triangular boundary elements (on these side faces) with very large aspect ratios.6 below. Harrington’s approach is essentially the same as the present one. The gap between the plates is g. It is not a good idea for cubelike conductors. the usual BEM requires collocation points on the entire bounding surface of a plate.e. This issue is discussed further in Section 3. Obviously. respectively. The model problem is a parallel plate capacitor with two square plates (Figure 3.1). The dielectric constant of the external medium. the enhanced BEM on the upper and lower surfaces. whenever possible. © 2005 by Taylor & Francis Group. Thin plate BEM. and in Section 3.1. where Q is the total charge on the top plate and V is the potential on it. It is important to point out that.a Parallel Plate Capacitor Numerical results are obtained from three methods . 3. this idea is good for moderate values of h/L. enhanced and thin plate BEM. to a prohibitively large number of boundary elements. The voltages on the upper and lower plates are V1 = 1 and V2 = −1. is unity. This is the model that has been presented in detail in Section 3.2 in many examples below. The length L of the side of each plate is unity.1.1. In the enhanced BEM. In this case. or. therefore.3 of this chapter. Strongly singular integrals and nearly strongly singular integrals (when they arise) are evaluated by Stoke regularization (see equations (3. T 3 elements are used on the side faces.5).50 CHAPTER 3. Also. The value of g/L = 0. the usual as well as the enhanced BEM breaks down for very small values of h/L. The corresponding integrals are treated as regular in these cases. nearly singular integrals only arise when plates are very close together (i. The capacitance C = Q/2V .1. A = L2 is the area of each plate. This assures linear interpolation of the charge density across a side face with no new unknowns being introduced on these faces (and.7. alternatively.6 The Model Problem . no additional nearly weakly singular integrals). only requires equation (3. of course. .4.13)). respectively. Calculation of the capacitance. to further reduce the number of collocation points. for small values of g/L). in general. weakly singular and nearly weakly singular integrals (when they arise) are evaluated by the method outlined in [26]. Again. LLC .1.
4 while the usual BEM employs standard Gaussian quadrature for evaluating these integrals.1 is only concerned with the total charge on the plates . and that [61] uses constant while the present work uses quadratic (T 6) elements.2 Comparison of various methods The performance of the three BEM models . With the given mesh.1) in Table 3.5) is now evaluated from the method outlined in [26].1.5 1 0.9 1 Thin plate BEM Harrington (1993) Figure 3.5 0 0 0.therefore only with equation (3.7: Normalized capacitance from the thin plate model compared with numerical results from [61]. The present results are believed to be more accurate than the older ones.2. the enhanced and the thin plate . 3. LLC .5 3 2.1.1. EXTERIOR BIE FOR POTENTIAL THEORY: MEMS 51 The diﬀerences are that the weakly singular integral in (3.7 0. Total charge.1 0.8 0. Ref.5) of the thin plate BEM.2 0.1.1. the enhanced BEM delivers © 2005 by Taylor & Francis Group. Diﬀerent values of h/L are considered here while g/L is kept ﬁxed at 0. This is primarily due to the fact that the enhanced BEM computes nearly weakly singular integrals by the method proposed in Section 3. It is ﬁrst observed that there is a signiﬁcant discrepancy between the usual and enhanced BEM results for h/L = 0.4 0.are compared for the parallel plate capacitor problem (Figure 3.3 0.the usual.the enhanced BEM being best suited for intermediate thicknesses and the thin plate BEM being best for very thin plates.6.5 g/L 0. Each method has diﬀerent regions where it excels .5 2 Cg __ Aε 1. [61] uses 36 constant square elements while the present work uses 16 × 16 × 2 T 6 elements on the upper surface of each plate (from [6]) 3.6 0.3. Table 3.
1 0. yields a value of 1. each side face 8 × 2 T 6 triangles.2 100. however. As seen from Figure 3. to compensate for the use of Gaussian quadrature for the evaluation of nearly weakly singular integrals by using a ﬁne mesh.7 4.3879 1.7 99.6631 1. which compares how the separate charges σ + and σ − (on the top plate) are calculated by the enhanced BEM and the thin plate BEM.3542 Thin plate BEM 1.1. the ﬁnal results from it do depend on h because the + + distance from S1 to S2 is g + h (see Figure 3. LLC . It is noted that even though the thin plate equations have no explicit dependence on the plate thickness h.10) now comes into play in the thin plate BEM. for h/L = 0.4(b)). Separate charges. or both.6652 1. Usual BEM : for cube . Computational eﬀort in cpu seconds (from [6]) Cg/A = 2.5874 1. THIN FEATURES h/L 1 0. g/L = 0.2 101.7405 1.01 0. When all these nearly weakly singular integrals are evaluated with standard Gaussian quadrature.5611 1. but with Gaussian quadrature used to evaluate nearly weakly singular integrals.3542.6094 1.2 4.they are negative !) The idea here is not to deliberately downgrade the usual BEM.2351 1.2 Table 3.52 Cg/A Enhanced BEM 2.9 95. Of course.6221 CHAPTER 3.6631. Enhanced BEM : each side face 16 × 2 T 3 triangles.each side face 8 × 8 × 2 T 6 triangles. the usual BEM yields Cg/A = 3.001 10−6 Usual BEM 2.9785 .2 100. many of the values of charge on the midside nodes on the side faces come out wrong .on each (unit) square (top and bottom) plate face 8×8×2 T 6 triangles. or a large number of Gauss points.2 100.001. Mesh : All cases . Rather.2 4. accurate determination of σ + and σ − is critical for MEMS applications since the traction on a © 2005 by Taylor & Francis Group.5(b). of course. The same approach. it is to emphasize that signiﬁcant errors can arise from the evaluation of nearly weakly singular integrals by standard Gaussian quadrature. equation (3. just using a large number of Gauss points may not help very much. with the same mesh.2.1: Summary of various cases for the parallel plate capacitor problem. As is well known.1 4. (In fact.05 0. The enhanced and the thin plate BEM show best agreement for h/L = 0.6899 1. Perhaps even more interesting is Table 3.4 4.2.6200 Computational eﬀort Usual Enhanced Thin plate BEM BEM BEM 1053. One tries.1 4.3975 3.005 0.1) and the gap g is kept ﬁxed here.a diﬀerence of over 25% ! The usual BEM model with 16 T 6 triangles on each side face has an additional 16 midside nodes on each side face that generate additional nearly weakly singular integrals (see Figure 3.
01.1 demonstrates huge savings in computational eﬀort for the thin plate BEM .615 1.2622 4.10) is used for obtaining the separate charges.4399 0. see Table 3. of course.4348 1.4303 1.465 1.1308 1. at a surface point.913 9.01 h/L = 0.01 h/L = 0. become nearly identical as h/L → 0.583 Table 3. The bottom line is that equation (3. Finally. collocated at x+ and x− in Figure 3.4500 9.1.2 (from [6]) plate.908 9.001 (and for a much lower value of h/L.8600 0.4539 9.7904 0.1.2: Comparison of enhanced BEM and thin plate BEM for the parallel plate capacitor problem.on each (unit) square (top and bottom) plate face 8 × 8 × 2 T 6 triangles. e. mesh on unit square 8×8×2 12 × 12 × 2 16 × 16 × 2 σ + at plate center 1.around a factor of 25 compared to the enhanced BEM. Agreement between the two methods is tolerable for h/L = 0.455 1. This is because equation (3. and. Computational eﬃciency.3).8761 0. however.8182 1.g.470 σ − at plate center 9. [87. and (3. g/L = 0. breaks down when determining the charges separately for h/L = 0.3.3: Convergence of the thin plate BEM model for the parallel plate capacitor problem for various meshes. Enhanced BEM : each side face 16 × 2 T 3 triangles.001 σ at plate center σ − at plate center σ + ds s+ 1 σ − ds s− 1 + 1. The enhanced BEM.593 1.3183 1. The reason for this failure is that the usual (and enhanced) BIEs. gets worse as h/L gets smaller.150 0.10) is crucial for obtaining the charges separately. 105]). LLC .9972 9.5300 12. g/L = 0.001 (even though it successfully calculates the capacitance for this value of h/L).1746 Table 3. It is found that the mesh dependence of the results is minimal.3. h/L = 10−6 . Table 3. This is analogous to the wellknown failure of the usual elasticity BIE for fracture mechanics problems (see.907 Cg/A 1. EXTERIOR BIE FOR POTENTIAL THEORY: MEMS Enhanced BEM h/L = 0. depends on the square of the charge density at that point.001 53 Thin plate BEM h/L = 0. © 2005 by Taylor & Francis Group.5) is now only used for obtaining the sum of the charges on each plate.2 (from [6]) The convergence characteristics for the thin plate BEM are investigated for a super thin plate (h/L = 10−6 ) in Table 3. Mesh : Both cases . The thin plate BEM performs well for h/L = 0.
It is well known (e. LLC . follows. THIN FEATURES ∂B0 ^ ∂B0 ξ x+ xS ^ S+ + x S^S Crack Figure 3.8: Geometry of the crack problem (from [105]) 3. A discussion of this.2 BIE for Elasticity: Cracks and Thin Shells Regularized BIEs for linear elasticity. The total surface of a cracked body B is ∂B = ∂B0 ∪ S + ∪ S − . where S + and S − are the (coincident) upper and lower surfaces of a crack in the body.g. These surfaces are geometrically identical and have opposite normals at any pair of twin points x+ and x− .2. The appropriate FP integrals are regularized by an additionsubtraction procedure followed by applications of Stokes’ theorem to convert singular or hypersingular integrals on surfaces into regular line integrals on the bounding contours of these surfaces. Further details are available in [105]. are presented in this section.8. and related issues. Cruse [38]). A ˆ ˆ singular element containing x+ is S + and similarly one containing x− is S − . suitable for applications in linear elastic fracture mechanics (LEFM).54 CHAPTER 3. The starting point is FP integrals that involve traction sums and crack opening displacements when the collocation point approaches a crack surface in the limit to the boundary (LTB) sense. The geometry of the problem appears in Figure 3. This is followed by a discussion of BIEs suitable for the analysis of thin elastic shells. 3. © 2005 by Taylor & Francis Group. that both the displacement and stress BIEs must be used appropriately in order to solve this problem by the BEM.1 BIES in LEFM This section is concerned with regularization of BIEs in linear elastic fracture mechanics.
1 Let an internal point ξ approach a point x on the outer surface ∂B0 of the body (see Figure 3. y)vi (y)dS(y) (3. v.23) can be collocated at x ∈ ∂B0 when solving a LEFM problem.21) (3. y)qk (y) − Sijk (x.8).23) [Dijk (x. Regularized displacement BIE. The simpler (3.8).1.36). a regularized form of equation (1. LLC .2 Source point approaching a crack surface Let an internal point ξ approach a point x+ on the crack surface S + (see Figure 3. [89] and Mukherjee et al. are deﬁned as: qi (y) = τi (y+ ) + τi (y− ) vi (y) = ui (y+ ) − ui (y− ) for twin points y+ and y− across a crack. y)(ui (y) − ui (x)]dS(y) + S+ [Uik (x. y)vk (y)] dS(y) Either of the equations (3.3.20) where the sum of the tractions across a crack.2. © 2005 by Taylor & Francis Group.2. The usual rigid body mode is used to regularize the integral on ∂B0 in the displacement BIE (1.20) is preferable and is recommended. Regularized stress BIE Use of both rigid body and linear displacement modes (see Lutz et al. (x)(y − x ) + S+ dS(y) (3.24). y) uk (y) − uk (x) − uk.24) is: 0 = ∂B0 [Uik (x. BIE FOR ELASTICITY: CRACKS AND THIN SHELLS Source point approaching the outer surface 55 3.1. [99]) leads to a regularized form of the stress BIE (1. y)τi (y) − Tik (x. q. 3.2.20) or (3.22) 0 = ∂B0 Dijk (x. This is: (3. and the crack opening displacement. y)qi (y) − Tik (x. With ∂B = ∂B0 ∪ S + ∪ S − . y) σkh (y) − σkh (x) nh (y) − Sijk (x.
y)dS(y) = −gik (x+ ) − ˆ S+ ˆ Ω(S + . LLC . y) ˆ ˆ and the solid angle Ω(S + . Following [116. y)r.24) can be written as: uk (x+ ) = ∂B0 Uik (x+ .26 .27) are evaluated in a clockwise sense when viewed from above (see Figure 3.3. [87] and Figure 3. y)vi (y) dS(y) + ˆ S+ Uik (x+ . y)ui (y) dS(y) + ˆ S + −S + Uik (x+ .21)) Σijk (x.i (x+ . y) dzm r(x+ . y)qi (y) − Tik (x+ . y)qi (y) dS(y) Tik (x+ . is the alternating tensor. y)nj (y) = Tik (x. equation (3.25). y) dz (3. y) qi (y) − Tik (x+ .2). z = y − x and . y)τi (y) − Tik (x+ . δ is the Kronecker delta and the line integrals in (3. also. y)[vi (y) − vi (x)]dS(y) (3.3. x+ ). x+ ) δik 4π (3. 99] (see. y) τi (y) − Tik (x+ .8).2): = Tik (x+ .27) 8π(1 − ν) ˆ L+ r.25) − ˆ S+ −vi (x) = Tik (x+ . y)vi (y) dS(y) (3. Also.27) and (3. y)dS(y) ˆ S+ It is ﬁrst noted that (see below (1. i ≡ ∂/∂yi .24) ˆ With x+ ∈ S + .25) is obtained as: © 2005 by Taylor & Francis Group. An FP representation of the displacement BIE (1. Using equations (3.26 .56 CHAPTER 3. y)ui (y) dS(y) += S+ Uik (x+ .26) where: 1 − 2ν 8π(1 − ν) + k m gik (x+ ) = 1 ik ˆ L+ r(x+ . subtended by S + at x+ . THIN FEATURES The displacement BIE.24) with x → x+ ∈ S + is: uk (x+ ) = ∂B0 Uik (x+ . is given by equation (3. (x+ .8) in (3.27). y). a regularized form of (3. In equations (3.
y)τk (y) − Sijk (x+ .2. y)qi (y)dS(y) Tik (x+ .j (x+ ) − ui. y)τk (y) − Sijk (x+ . y)uk (y) dS(y) Dijk (x+ . y)[sk (y) − sk (x)]n (y)dS(y) Sijk (x+ . y)[vi (y) − vi (x)]dS(y) vk (x) 4π 2π − ˆ S+ +gik (x+ )vi (x) − cos(ψ(θ))dθ 0 (3. y)τi (y) − Tik (x+ .3. BIE FOR ELASTICITY: CRACKS AND THIN SHELLS 57 1 (uk (x+ ) + uk (x− )) 2 = ∂B0 Uik (x+ . y)qk (y) − Sijk (x+ . y)vi (y) dS(y) + ˆ S+ Uik (x+ . y)uk (y) dS(y) + ˆ S + −S + Dijk (x+ . y)dS(y) +dmn (x) = Ek ˆ S+ + mn Dijk (x .31) (3. y)[vk (y) − vk (x) − dkn (x)(yn − x+ )]dS(y) n ˆ S+ − ˆ S+ −vk (x) = Sijk (x+ .30) where the newly deﬁned quantities are: sij (x) = σij (x+ ) − σij (x− ) dij (x) = ui. LLC (3. equation (3.32) .j (x− ) © 2005 by Taylor & Francis Group. y)(yn − x+ ) dS(y) n (3. y)qi (y) − Tik (x+ . y)n (y) − Sijm (x+ .29) can be written as: σij (x+ ) = ∂B0 Dijk (x+ . y)vk (y) dS(y) (3.29) += S+ ˆ With x+ ∈ S + . y)ui (y) dS(y) + ˆ S + −S + Uik (x+ . y)qk (y) − Sijk (x+ . y)vk (y) dS(y) + ˆ S+ Dijk (x+ . An FP representation of the stress BIE with ξ → x+ ∈ S + is: σij (x+ ) = ∂B0 Dijk (x+ .28) The stress BIE.
the line integral in (3. y) (3. the entire term must be converted together.36) is evaluated in a clockwise sense when viewed from above (see Figure 3. [86].e. one can write: Jijmn = Eijpq Imnpq where: Imnpq = − = Ek ˆ S+ + mn Upk. as in Liu et al. y) nt dzt (3.33) The last two terms on the righthand side of equation (3. y)(yn − x+ ) dS(y) n (3. y)dS(y) = −Eijmn ˆ S+ ˆ L+ Σk + m (x .31). as before. y)dS(y) = − = Eijmn Σk ˆ S+ ˆ S+ + m. In order to succeed in this task.34) The hypersingular integral on the righthand side of (3.32).37) Using equation (3.q (x . y)n (y)dS(y) (3. one gets: = Sijk (x+ .36) where. some surface integrals will survive. The ﬁnal task is to convert the last term in the righthand side of equation (3.30) involve FP integrals. y)n Jijmn = = Ek ˆ S+ (y) − Sijm (x+ .34) and the deﬁnition of the kernel D in terms of U from equation (1.30) into line integrals. Otherwise.n (x . s is related to d by Hooke’s law in the same manner as σ is related to ∇u.34) can be transformed into a line integral. one has: = Σk ˆ S+ m. First deﬁne: + mn Dijk (x . the desired result is: = Sijk (x+ . Using the deﬁnition of Sijk from (1. A careful derivation is provided in Mukherjee and Mukherjee [99].58 CHAPTER 3. This result also appears in Liu et al.q (x .35) Finally. i. THIN FEATURES Of course.38) − Σm + p.2). Applying Stokes’ theorem. Appendix D. LLC . y)(yn − x+ ) n (y)dS(y) n (3.: sij = Eijmn dmn (3.39) Integrating by parts: © 2005 by Taylor & Francis Group. [86].n n dS = ˆ L+ Σk m nt dzt (3. These are now regularized by applying Stokes’ theorem.
use of equations (3.44) Also. using Lutz et al. BIE FOR ELASTICITY: CRACKS AND THIN SHELLS 59 Imnpq = − = Ek ˆ S+ + mn Upk (x .45) in (3. 3.40) need careful treatment. y)uk (y) dS(y) ∂B0 © 2005 by Taylor & Francis Group.30) yields a fully regularized version of (3.40) (1) (2) − = Σm p δnq n (y)dS(y) ≡ Imnpq + Imnpq ˆ S+ Once again. 3.45) = gmp (x+ )δnq + Finally. y)δnq ds(y) ˆ Ω(S + . y) − Σm p (x+ . using equation (3.42) ˆ ˜ ˜ ˆ ˜ (where S = S + − S is a punctured surface with x+ ∈ S and L = L+ − L the bounding contour of S). Deﬁne F + mnp (x . y) qt dzt (3. nq ] ds = L F mnp t q dzt (3.38.40.36. is available in [105]).40): = − (2) Imnpq (x+ ) ˆ S+ Tmp (x+ . This is: (1/2)[σij (x+ ) + σij (x− )] = Dijk (x+ . the hypersingular integrals in equation (3. y) = Ek + mn Upk (x . and the fact that F mnp.43) where: − Σm p (x+ .q n (y)dS(y) (3. The ideas presented in Mukherjee and Mukherjee [99].2.26) in (3.’s ([89]) “tent” integrals. LLC . y)τk (y) − Sijk (x+ . (An alternative proof.30).3. y)(yn − x+ ) n . y)(yn − x+ ) n (3. = 0 for x+ ∈ S. x+ ) δmp δnq 4π (3. need to be employed here. one ﬁnally / gets: (1) Imnpq (x+ ) = −hmnpq (x+ ) (3. 3.q n −F mnp. Appendix D.43 and 3. y) − Σm p (x+ .3. y)(yn − x+ ) n hmnpq (x+ ) = ˆ L+ Ek + mn Upk (x .41) Using Stokes’ theorem in the form: [F S mnp.33.
23).3 = u2.2 − ν 1−ν (u1. The ﬁrst step in the implementation is to choose shape functions for u and τ on ∂B0 and q and v on s+ . in the αk frame : Adding the ﬁrst and last of equations (3. (3. 3).48) The remaining components of d are obtained as follows: + u1. (3. 2.47) (1−2ν)τ3 2(1−ν)G The quantities dk (x) with (x ∈ s+ ) require application of the above idea on both surfaces of a crack (Figure 3. y)vk (y) dS(y) + ˆ S+ Dijk (x+ . The remaining displacement gradients at x are obtained from the formulae: u1.3 = u3. δ = 1. in the αk coordinate frame.3 (x− ) = τ1 /G − (u− ). THIN FEATURES Dijk (x+ .3 (x+ ) = τ1 /G − (u+ ). y)[vk (y) − vk (x) − dkn (x)(yn − x+ )]dS(y) n ˆ L+ − ˆ S+ + Eijmn vk (x) Σk + m (x .49). A dot product of equation (3.2. are chosen at x ∈ ∂B0 such that the α3 axis is normal and the α1 and α2 axes are tangential to ∂B0 at x . Next.28) and (3.3 = τ1 G τ2 G − u3.23) and dk (x) (x ∈ s+ ) in (3.20). in the αk frame : −u1.2 ) (3.9). The former is obtained as suggested by Lutz et al. For 3D problems. tangential diﬀerentiation of the displacement shape functions provides the quantities uk.γ . one has: dkγ = vk. 2 (3.1 − u3.1 3 − On S − .46). LLC . (x) (x ∈ ∂B0 ) in (3.1 3 On S + .1 + u2. y)qk (y) − Sijk (x+ .46).46) must be taken with n(x+ ).2 Numerical Implementation of BIES in LEFM Of interest here are the regularized equations (3.60 + ˆ S + −S + CHAPTER 3.49) © 2005 by Taylor & Francis Group. 2. First. 2. y)[sk (y) − sk (x)]n (y)dS(y) Sijk (x+ . y) nt dzt 2π − Eijpq dmn (x) hmnpq (x+ ) cos(ψ(θ))dθ (3. in the βk frame : u1. local coordinates αk (k = 1. [89]. 3. k = 1. it is necessary to obtain uk. 2.3 (x ) = τ1 /G + (u− ).46) + Eijpn dmn (x)gmp (x+ ) − sij (x) 4π 0 3.1 3 − − − On S . one has: (3.δ . k = 1. Now. 3 : γ = 1.
however. leads to some extra terms in a regularized HBIE such as equation (3. BIE FOR ELASTICITY: CRACKS AND THIN SHELLS 61 α1 β1 β2 β3 α3 α2 S S +  Figure 3. in [85] . On S + . for example. then the procedure outlined in this section is no longer required.52) (1 − 2ν)q3 ν − (d11 + d22 ) 2(1 − ν)G 1−ν + It should be noted that in equations (3. LLC .51) (3. both the regularized displacement BIE (3.2. The surface gradient regularization procedure. k − and similarly at x .9: Local coordinates on crack surfaces (from [105]) At x+ in the αk frame : d13 = Similarly: d23 = d33 = q2 − d32 G q1 − d31 G (3. τk ≡ τk (x+ ).50) (3.23) in this chapter. It is interesting to point out that if the surface (tangential) gradient of the displacement ﬁeld (rather than its total gradient) is used to regularize the relevant HBIE.49).[86]. u+ ≡ uk (x+ ). as is done.3 Some Comments on BIEs in LEFM As has been mentioned before.3. Finally. s is related to d by Hooke’s law (3. equation (3. 3.20) is the most convenient one for collocation on ∂B0 .28) and © 2005 by Taylor & Francis Group.33).2.
Thus.3. in this case: © 2005 by Taylor & Francis Group. τ − . and the crack opening displacement (COD) v. it is perfectly suited for the practical case of solving for v given τ + and τ − . as early as at least 1975 ([17]. its dot product with n+ ) on S + . For example. is. however. it can be used to solve for q given u+ and u− .8 replaced by the “edge surface” SE of the shell. “Overcollocation” on crack surfaces does not help since q alone does not specify a problem.1 Mathematical formulation Regularized BIEs and HBIEs for thin shells are presented next. The geometry of the problem is shown in Figure 3.46) are.46) (i.g. “defective.28) contains u+ and u− on the crack surfaces. both a tractionfree and a pressurized crack have q = 0 but very diﬀerent solutions. it is most convenient to collocate equation (3. 3.20) on ∂B0 and (3. This failure of the displacement BIE for fracture mechanics problems.4 BIEs for Thin Shells This section is concerned with BIEs and HBIEs for thin shells. especially in the limit as the thickness of a shell → 0. the “tent” is chosen such that the unit outward normal to the body at x+ is also the outward normal to the “tent” at that point. 3. (Lutz et al. Again. from both a mathematical as well as an engineering perspective. In summary. THIN FEATURES the stress BIE (3. but only q (not the individual tractions) on S + and on S − . [59]). it contains the tractions τ + .10. therefore.28) on S + . When one takes a dot product of the stress BIE (3.62 CHAPTER 3. at a postprocessing step. The individual crack surface displacements u+ and u− can then be calculated.46) with n+ . the components of the tensor d can be easily calculated from the tractions and the tangential derivatives of the COD (3. [89] used u and τ on crack surfaces as primary variables and collocated the regular BIE on one crack face and the HBIE on the other).” Equation (3.e. this integral can remain nonzero with ﬁnite applied tractions per unit length along this curve. of course.2. in the limit as the thickness → 0. The question of what constitutes a thin plate or shell.48. 3.33)). in some sense. The situation here is quite analogous to the crack problem with the outer surface ∂B0 in Figure 3. Cruse [38]) and has led to the search for a remedy in the form of a stress BIE.52). It has been proved in [105] that. LLC .2. Fortunately.50. if desired.4. This equation would fail if one wished to calculate the individual tractions on the crack surfaces given the COD. and s is related to d by Hooke’s law (3. from (3. It is important to retain the integral over SE because even when the shell thickness h → 0 and SE degenerates from a surface to a curve. is an interesting one that has been discussed extensively in the literature. (As mentioned in the previous subsection. but not for the important practical case of solving for the crack surface displacements given the tractions. well known (e.
3. LLC . the displacements and stresses on the upper and lower surfaces of the shell must approach each other as the shell thickness h → 0. of course. Nevertheless.55) appears to limit the class of thin shell problems that can be solved by this BIE approach.2. equations (3. or.10: Geometry of the thin shell problem (from [105]) vi (x) = ui (x+ ) − ui (x− ) = 0 sij (x) = σij (x+ ) − σij (x− ) = 0 (3.53 3. a strong restriction on the applied tractions on S + and S − . equation (3. Subtracting (12) from (11) immediately gives ∆u = 0 which implies that (since the matrix B + in [87] is nonsingular) Σt = 0. work well in practice.53) and (3. equation (13) of Liu [87] (which is a restatement of his © 2005 by Taylor & Francis Group. (in the limit h → 0).53) (3.55) are always true in the thin shell limit h → 0. 3.54) simply state that. BIE FOR ELASTICITY: CRACKS AND THIN SHELLS 63 ξ n+ SE Thin Shell ^+ tent S c x+ xS + ^ S+ SE S^S Figure 3. i.2 Some comments on thin shell problems First.3. for that matter. for example. starting with a continuous displacement and stress ﬁeld in a shell of ﬁnite thickness. it is well known.54) approximately. (3. These equations are valid for h → 0 in the absence of the edge integrals on SE . namely that: qk (x) = τk (x+ ) + τk (x− ) = 0 (3. however.54) implies. by any other approach.2. Therefore.54) Equations (3.53 . Unfortunately.53 .55) of the present chapter. One of these cases appears in Liu [87] and is discussed in the next subsection. There are interesting situations that satisfy (3. that ﬁnite element shell formulations.e.55) Strictly speaking.4.3. since equations (3. it is useful to examine equations (1113) in Liu [87].
the radial one.53) is satisﬁed in the thin shell limit. of course.no other solutions are possible in the thin shell limit ! One is. Also. that violates (3. In view of the above discussion.54) is satisﬁed by the tangential stress components (in the thin shell limit) while the radial stress component. Finally. (Note that (13) of [87] is the limiting form of (45) of [87] as h → 0. is small compared to the tangential ones.64 CHAPTER 3. LLC . violates (3. the only nonzero applied traction.54). the integral on SE ) to thin shells. for general thin shell problems.55) but its magnitude is small compared to the tangential stress components.e. One reason for his success is that in his spherically symmetric numerical example the only nonzero displacement component. collocated at twin points on opposite surfaces of a shell. The shearing stress components are zero in a spherical coordinate system. application of equations (45) of Liu [87] (including. Liu [87]. THIN FEATURES equations (1112) in terms of displacements and tractions on the upper and lower surfaces of the shell) only admits the solutions v = q = 0 . remains an open question that needs further investigation. however.55). so that (3. (3. © 2005 by Taylor & Francis Group. the authors of this book feel that the eﬀect of the asymptotic requirement (3.55) on the numerical performance of the standard BIE. in general. i. the radial one.) Diﬃculties might arise when applying equations (45) of [87] to shells of small but ﬁnite thickness h if the applied tractions do not satisfy (3. is nearly uniform as a function of the radius. primarily interested in solving problems for shells of small but ﬁnite thickness. since one would then have a contradiction as h → 0. has demonstrated the usefulness of his equations (45) for the case of a thin spherical shell under internal pressure with radius to thickness ratios of up to about 200.
Part II THE BOUNDARY CONTOUR METHOD 65 © 2005 by Taylor & Francis Group. LLC .
1 Surface and Boundary Contour Equations Basic Equations A regularized form of the standard boundary integral equation (Rizzo [141]).1. 2. for 3D linear elasticity (see equation (1. where n is the unit outward normal to ∂B at a point on it.Chapter 4 LINEAR ELASTICITY Derivations of the BCM and HBCM. 3) are global Cartesian unit vectors.1) Here. for 3D linear elasticity.2) . one gets: 67 © 2005 by Taylor & Francis Group. y){ui (y) − ui (x)}] ej · dS(y) Fk · dS(y) ∂B ≡ (4. LLC (4. together with representative numerical results for selected problems.1 4. The ﬁrst task is to show that the integrand vector Fk in equation (4. These are given in Chapter 1 as equations (1. is: 0 = ∂B [Uik (x. respectively. are presented in this chapter.22) and (1.1) is divergence free (except at the point of singularity x= y). The BEM Kelvin kernels are written in terms of (boundary) source and ﬁeld points. The stress tensor is σ. ∂B is the bounding surface of a body B (B is an open set) with inﬁnitesimal surface area dS = dSn. 4. as before. Writing in component form: Fk = Fjk ej = (σij Uik − Σijk ui )ej + Σijk ui (x)ej Taking the divergence of the above at a ﬁeld point y. the displacement vector is u and ej (j = 1. y)σij (y) − Σijk (x.18).26) in Chapter 1).
LLC .68 CHAPTER 4. the surface integral in equation (4. respectively. since the kernels in equation (4. except at the point of singularity. the ﬁrst expression on the righthand side (RHS) of equation (4. are related by Hooke’s law in exactly the same manner (see. LINEAR ELASTICITY ∇y · Fk = Fjk. interpolation functions must be chosen for these variables.j ui ) + Σijk. equilibrium in the absence of body forces demands that σij be divergence free. these interpolation functions must also be written in the coordinates zk in order to determine the potential vectors Vk . The above property demonstrates the existence of vector potential functions Vk such that: Fk = ∇ × Vk (4. a priori.3) where the Kelvin strain tensor E and the strain ﬁeld e are: Eijk = (1/2)(Uik.j ui (x) (4. The corresponding Kelvin stress and strain tensors Σijk and Eijk . Mukherjee [98]). The corresponding Kelvin stress tensor Σijk is also divergence free. these interpolation functions are global in nature and are chosen to satisfy.5) As a consequence of equation (4. in order to determine the vectors Vk . σij and ij . Fk in equation (4.4) Let (u. ij = (1/2)(ui. the NavierCauchy equations of equilibrium. σ) correspond to a body force free electrostatic state with the same elastic constants as the Kelvin solution.j + Ujk.j Uik − Σijk.5). i. Finally.3) also vanish everywhere.g. except at the point of singularity.3) vanishes. Therefore.1. As a consequence. The closed curve C and the open surface S need not be ﬂat .e. The weights.j + uj.i ) (4. and potential functions derived for each linearly independent interpolation function.1) are functions only of zk = yk − xk (and not of the source and ﬁeld coordinates separately). the second and third expressions on the RHS of (4. (Such functions are called Treﬀtz functions).1) over any open surface patch S ∈ ∂B. by applying Stokes’ theorem. can be converted to a contour integral around the bounding curve C of S.they only need to be piecewise smooth. respectively. are related to each other through the usual Hooke’s law. The stress and strain tensors. e. in linear combinations © 2005 by Taylor & Francis Group. 4. Also.i ).6) Stokes’ theorem is valid under very general conditions. Also.3) is divergence free.: Fk · dS = S C Vk · dr (4.2 Interpolation Functions Since the vectors Fk contain the unknown ﬁelds u and σ. Thus.j = (σij Eijk − Eijk ij ) + (σij.
3 are constant.. α = 1.. 5. 2.4. are deﬁned piecewise on boundary elements.7) ui = α=1 βα uαi (y1 . .. x3 )uαi (z1 .9) where uαi . x2 . σ αij (with i = 1. . Quadratic interpolation functions are used in this work. 3 are constants.1: Treﬀtz functions for interpolating displacements. k1 = −4(1 − ν).. 27 are of second © 2005 by Taylor & Francis Group.. 27 are quadratic. y2 . 27 constants βα that will be related to physical variables on that element. 14. SURFACE AND BOUNDARY CONTOUR EQUATIONS uαi ¯ i=1 i=2 i=3 uαi ¯ i=1 i=2 i=3 uαi ¯ i=1 i=2 i=3 uαi ¯ i=1 i=2 i=3 1 1 0 0 8 0 y2 0 15 k2 y1 y2 2 y1 0 22 0 2 y3 k2 y2 y3 2 0 1 0 9 0 0 y2 16 k1 y1 y2 2 y2 0 23 0 k2 y2 y3 2 y2 3 0 0 1 10 y3 0 0 17 k1 y1 y3 0 2 y3 24 0 k1 y2 y3 2 y3 4 y1 0 0 11 0 y3 0 18 k2 y1 y3 0 2 y1 25 y 2 y3 0 0 5 0 y1 0 12 0 0 y3 19 2 y3 0 k2 y1 y3 26 0 y1 y3 0 6 0 0 y1 13 2 y1 k1 y1 y2 0 20 2 y1 0 k1 y1 y3 27 0 0 y1 y2 7 y2 0 0 14 2 y2 k2 y1 y2 0 21 0 2 y2 k1 y2 y3 69 Table 4. 2.. . those for α = 4. 12 are of ﬁrst degree and those for α = 13... k2 = −2(1 − 2ν). This set β diﬀers from one element to the next.. z2 . x2 . The displacement interpolation functions for α = 1. however. y3 ) = α=1 ˆ βα (x1 . z2 . 27) are the interpolation functions and βα are the weights in the linear combinations of the shape functions.. .. 12 are linear and α = 13. . LLC . 2... With : zk = yk − xk one has. z3 ) (4. on a boundary element : 27 27 (4. associated with it. 2. α = 4.. Each boundary element has. y2 .1. 3 and α = 1. x3 )σ αij (z1 . k3 = k1 − 4 of these interpolation functions. y3 ) = α=1 ˆ βα (x1 .8) 27 27 σij = α=1 βα σ αij (y1 . z3 ) (4.
α = 4.. 2 i. 3 i. 2 i. j = 2. k1 = −4(1 − ν). j = 2. . 2 i. x3 )βα . j = 1.. j = 1. 2 i. LLC .. The interpolation functions for the stresses are obtained from those for the displacements through the use of Hooke’s law. j = 3.. 3 1 0 0 0 0 0 0 8 λ/G 0 0 k1 /k2 0 λ/G 15 k1 y2 4νy1 0 −4νy2 0 −4νy2 22 −4νy2 0 0 −4νy2 4νy3 k1 y2 2 0 0 0 0 0 0 9 0 0 0 0 1 0 16 k3 y2 k1 y1 0 −k1 y2 0 −4νy2 23 −4νy3 0 0 k1 y3 4νy2 −4νy3 3 0 0 0 0 0 0 10 0 0 1 0 0 0 17 k3 y3 0 k1 y1 −4νy3 0 −k1 y3 24 −4νy3 0 0 k3 y3 k1 y2 −k1 y3 4 k1 /k2 0 0 λ/G 0 λ/G 11 0 0 0 0 1 0 18 k1 y3 0 4νy1 −4νy3 0 −4νy3 25 0 y3 y2 0 0 0 5 0 1 0 0 0 0 12 λ/G 0 0 λ/G 0 k1 /k2 19 −4νy1 0 4νy3 −4νy1 0 k1 y1 26 0 y3 0 0 y1 0 6 0 0 1 0 0 0 13 −k1 y1 k1 y2 0 k3 y1 0 −4νy1 20 −k1 y1 0 k1 y3 −4νy1 0 k3 y1 27 0 0 y2 0 y1 0 7 0 1 0 0 0 0 14 −4νy1 4νy2 0 k1 y1 0 −4νy1 21 −4νy2 0 0 −k1 y2 k1 y3 k3 y2 Table 4. j = 1. 1 i. 3 i. 3 σαij/G ¯ i. j = 3. . j = 1. 2 i. 3 i.7) results in the ˆ constants βj being related to the constants βα as follows : 27 ˆ βi = α=1 Siα (x1 . 3 i. j = 1. 3 σαij/G ¯ i. 1 i. i = 1. 2 i. 3 i.70 CHAPTER 4. j = 3. 2. LINEAR ELASTICITY degree.1 and 4. j = 2. j = 1. j = 2. j = 2.. j = 3. x2 . j = 2. 2. σαij/G ¯ i. 3 i.2: Treﬀtz functions for interpolating stresses. j = 1. k3 = k1 − 4 It is easy to show that the coordinate transformation (4. 14. k2 = −2(1 − 2ν).2. 3 i.. j = 1. respectively. 2 i. j = 1. j = 1. 3 are zero. 3 i. 3 σαij/G ¯ i. j = 1. 3 (4. j = 2. 5. 2 i.10) © 2005 by Taylor & Francis Group. 1 i. 12 are constant and α = 13. j = 1. The interpolation functions uαi and σ αij are given in Tables 4. α = 1. 27 are linear. There are a total of 27 linearly independent (vector) interpolation functions uα . 1 i. j = 2.
= ∂uα (y1 .12) where Siα Rkα = uαi (x1 .. . in order that the transformation matrix T relating the vectors a and β on element m is square. The CIM9 boundary element. Thus.1(a). however. 2.. y3 ) ∂yj i = 1. Also.. denotes the largest integer less than or equal to n.1. .. n = k − 3 ˆ βα = βα . Some of these would be speciﬁed as boundary conditions and the rest would be unknown in a given problem.13) Of course. x2 . the matrix T must be invertible. This relationship is expressed as: m a = Tβ mm (4. SURFACE AND BOUNDARY CONTOUR EQUATIONS 71 27 ˆ βk = α=1 Rnα (x1 . unlike in the standard BEM. 2. such as stresses and curvatures.. 2. 2. x2 . Once all the primary physical variables are known. 27 (4.x2 .11) (4. [116].. 14. 9. 4. Surface stresses are discussed in Section 4.. The issue of invertibility is discussed in Nagarajan et al..7. A departure from the usual BEM. α = 1. The ﬁrst step in the BCM solution procedure is to determine the unspeciﬁed primary physical variables in terms of those that are prescribed as boundary conditions. x3 ). called the 3 ﬂoor of n. α = 13. shown in Figure 4. The displacement u is the primary physical variable at the three corner nodes Ci and the three midside nodes Mi .3 Boundary Elements The BCM is a perfectly general approach that can be used to solve any wellposed problem in linear elasticity. . (x1 . it is particularly easy to obtain surface variables.1..4. It is useful to note that the matrices S and R are functions of only the source point coordinates (x1 .. . y2 . k = 1. while tractions are primary variables at the internal nodes Ii . x2 . 3. is that a set of primary physical variables ak are ﬁrst chosen at appropriate points on a boundary element. the rest of the physical variables (the secondary ones) are obtained at a simple postprocessing step. the symbol n . is used in the present work.. The number of primary physical variables on a boundary element must match the number of artiﬁcial variables βk associated with it. k = 4.12. © 2005 by Taylor & Francis Group. there are a total of 27 primary physical variables. 5. x3 ) . x3 )βα .x3 ) with j = 1 + k−1 and = k − 3j + 3. LLC . 27 α = 1. Here. . 27 .1. in the BCM..
. some of the peripheral nodes may lie on corners or edges.14) k where zi are the relative coordinates of the peripheral nodes 1. 3. This approach eliminates the wellknown problems associated with modeling of corners and edges in the usual BEM.15) © 2005 by Taylor & Francis Group. 6 (4.0) 4 6 * I2 3 5 (1. 3. N5 = 4L2 L3 . on edges and corners. The boundary elements. in general. The relative coordinates zi (see equation (4. 2. LLC .. N6 = 4L1 L3 (4. no sum over k = 4L1 L2 . 3. 2.. k = 1.1). The shape functions are: Nk N4 with: = (2Lk − 1)Lk . 3) 1 C1 M1 C2 * I1 I3 * M2 (a) M3 C3 2 (1. are curved (quadratic) with their shapes deﬁned by the six points Ck . k = 1.. i = 1. LINEAR ELASTICITY η (0. 2. while having traction components only at regular boundary points. In a typical discretization procedure. (see Figure 4.1).7)) of a point on one of the sides of a triangle are written as: k zi = Nk (ξ.. 2.. 6 on the CIM9 element (see Figure 4.1: (a) The CIM9 boundary element. . Mk . It is of obvious advantage to have to deal only with displacement components. (b) Intrinsic coordinates (from [109]) The BCM equations are collocated at the six peripheral nodes as well as at the centroid of the element.0) ξ (b) Figure 4. and ξ and η are intrinsic coordinates. k = 1. 2. η)zi . while the internal nodes are always located at regular points where the boundary ∂B is locally smooth. that are always continuous.72 CHAPTER 4.
13) contain the components of the normal n at the points I1 .1(b). z3 ) = (1/n)Fαk (z1 .4.17) where r = zi ei .20) © 2005 by Taylor & Francis Group.1.g. homogeneous. In this work. tz2 . tz2 . 4. at a point on it. Thus: 1 Fαk = ∇ × Vαk ⇒ Vαk (z1 . 4.18. The unit outward normal to a boundary element.16) The reference triangle in intrinsic coordinates is shown in Figure 4. z3 ) = 0 tFαk (tz1 . [69]). e. tz3 ) = tn−2 Fαk (z1 . for the nonsingular case n = 0 equations (4.18) where r = zi ei . is given by: n= ∂r ∂r × ∂ξ ∂η / ∂r ∂r × ∂ξ ∂η (4.2). 1. Σ is of degree −2 and u is of degree −1. 2 are used (see Tables 4. I2 .22) and (1.1. From equations (1. it is clear that both Σijk and Uik are ratios of homogeneous polynomials and are. LLC . z2 .19) Therefore. tz3 )dt × r (4.1 and 4. z2 . z2 . 4.1(a).1 The nonsingular case : n = 0 The homogeneous nature of the Kelvin kernels greatly facilitates the use of an inversion integral to calculate the inverse curl of a given vector ﬁeld of zero divergence (see. z3 ) (4. then the resulting force vector Fαk (which is Fk corresponding to the αth interpolation function) is of degree n − 2. It is important to point out that the elements of the transformation matrix T in equation (4.4 Vector Potentials The homogeneous nature of the Kelvin kernels is exploited in deriving the potential functions. Here.18). If an interpolation function is of degree n. z3 ) × r (4. therefore. z2 .1.19) yield: Vαk (z1 .4. SURFACE AND BOUNDARY CONTOUR EQUATIONS 73 L1 L2 L3 √ = η/ 3 √ = (1/2)(1 − ξ) − η/(2 3) √ = (1/2)(1 + ξ) − η/(2 3) (4. I3 of the CIM9 element shown in Figure 4. Because of the homogeneous nature of Fαk : Fαk (tz1 . interpolation functions with n = 0.
2 respectively.13) and (4.k − r. Corresponding to the ﬁrst two terms. The expression for Σijk ej can be partitioned into three terms. It is well known that the solid angle Ω subtended at the source point x by the open surface S has the expression: Ω= S r.5 Final BCM Equations Use of the interpolating functions for displacement and stress from Tables 4. one must now deal with the term Σαjk given in equation (1.k − r. 3.24) Therefore.j δαk ej = ∇ × r2 1 r.1) into a regularized boundary contour equation (BCE) that can be collocated (as in the usual BEM) at any boundary (surface) © 2005 by Taylor & Francis Group. This is the only surface integral that is ever computed in the BCM.1.j δαk r. Referring to equation (4. use of this line integral has proved to lack robustness in general numerical computations involving 3D bodies of complex shape. Therefore.20 .α δjk − r. LLC . 3.1 and 4.1. While it is possible to convert the surface integral in (4. It is also noted here that algebraic expressions exist for the solid angle for the special case when S is a plane.k δαj ej = ∇ × r2 kmj r.74 4.k δαj ej + + ej 2 8π(1 − ν)r 4πr2 −Σαjk ej = (4.4.7 . The singular case : n=0 LINEAR ELASTICITY The singular case (n = 0) corresponds to constant displacement interpolation functions uαi = δαi .24) transforms the regularized BIE (4. (where δ is the Kronecker delta ¯ see Table 4.4.j δαk ej 8π(1 − ν)r2 1 − 2ν r.4. 4.dS r3 (4.21) over S equals (Ω/4π)δαk .α r.α r.23) αkm r where is the alternating tensor. the surface integral of the third term on the righthand side of (4.22). 2. Fortunately.m ej r em (4.22) (4. α = 1.21) are divergence free and can be written as the curl of a potential function.α r.1). i = 1. each of which are divergence free [116]. together with equations (4. the solid angle Ω is computed as a surface integral according to equation (4. 2. the solid angle is a purely geometrical quantity that can be computed easily as a surface integral in a robust fashion. One writes: 1 3r.24) to a line integral [116]. one has [116]: 1 3r.21) Each of the above three terms on the righthand side of equation (4.α δjk − r.j ej · dS = r2 S r.2 CHAPTER 4.j r.1).24).j r.
using spherical coordinates. and ijk is the usual alternating symbol. as long as the displacement is continuous there. The ﬁnal result is: Ka = 0 which is written as: Ax = By (4.α r.25) with = − Sm Dαjk dzj Lm Σαjk ej · dS = 1 r.j (y) − nj (y)f. by Ghosh and Mukherjee [48]. Once these © 2005 by Taylor & Francis Group.4. Manti˘ [94] has independently c derived the result in equation (4.28) where x contains the unknown and y the known (from the boundary conditions) values of the primary physical variables on the surface of the body.26) m Here Lm is the bounding contour of the surface element Sm .1. SURFACE AND BOUNDARY CONTOUR EQUATIONS 75 point .27) P P The equations are assembled by making use of the fact that displacements are continuous across elements. Also. Also. T and a are the same quantities evaluated on any element that belongs to the set S of elements that contain the source point x.i (y) (4. LLC .29) (4. T and a are the transformation matrix and primary physical variable vectors on element m. This equation is: 1 2 m=1 α=13 M 12 M 27 m 0 = (σ αij Uik − uαi Σijk ) Lm jnt zn dzt T −1 a m α + m=1 α=4 M 3 Lm (σ αij Uik − uαi Σijk ) m jnt zn dzt P RT m −1 m a α−3 + m∈S / m=1 Dαjk dzj α=1 Lm S T −1 a − T −1 a α m P (4.26) with the use of the tangential diﬀerential operator: Dij (f (y)) ≡ ni (y)f. This method of integrating Σijk has been presented before.i dzj kij 8π(1 − ν) Lm r 1 − 2ν Ω 1 + δαk αkj dzj + 8π(1 − ν) Lm r 4π m (4.including those on edges and corners.
one has: 1 Number of elements + 2 (4.5M equations at the midside nodes and 0. For example. stresses and curvatures at a regular oﬀcontour boundary point (ROCBP) on the bounding surface of a body.1. use of the CIM9 element results in overdetermined.7 Surface Displacements.5F for triangular elements). Of course. 1. One therefore has two extra vector (six extra scalar) equations. This is the worstcase scenario in the sense that for mixed boundary value problems the number of equations remain the same while the number of unknowns decreases. the BCM equations are enforced at all the peripheral nodes and also at the centroid of the element. at a postprocessm ing step. in which all the displacements are prescribed on ∂B. Thus. for a Neumann problem in which all the tractions are prescribed on ∂B.29). Thus. One is now in a position to count the number of (vector) equations and unknowns in the global system (4.76 CHAPTER 4. In summary. for M boundary elements. Now.29) generally leads to a rectangular matrix A. β α can be easily obtained on each boundary element from equation (4. once the standard BCM is solved. one only has 2M + 2 (vector) unknowns (displacements at the peripheral nodes of the elements).5M + 2 at the corner nodes. Also.31) 2 Also.6 Global Equations and Unknowns The global system equation (4.the tractions at the nodes interior to the elements. a regular boundary point can lie on or away from a boundary contour. consistent linear systems for wellposed problems in linear elasticity. the vector a is completely known. V is the number of vertices (here the number of corner nodes) and E is the number of edges (here 1. a CIM9 element has 1. Euler’s theorem states that: F +V =E+2 (4. Number of corner nodes = 4. Here. The system of linear equations is usually overdetermined but always consistent.5 midside nodes. The former is © 2005 by Taylor & Francis Group. For a Dirichlet problem.30) where F is the number of faces (here the number of elements). there are a total of 3M vector unknowns .13). In a CIM9 element. For any general polyhedron.1. LINEAR ELASTICITY equations are solved. it is very easy to obtain the displacements. 4. A count of equations and unknowns is given below. one has M equations at the centroids. for a total of 3M + 2 equations. a Neumann problem is illposed since rigid body displacements of the body are not constrained. a point at an edge or corner is called an irregular point while at a regular point the boundary is locally smooth. and Curvatures A very useful consequence of using global shape functions is that. LLC . Stresses.
3 Surface curvatures 2(β13 + β20 ) ∂ 2 u1 = ∂xi ∂xj symmetric ∂ 2 u2 ∂xi ∂xj = 2β15 k2 β15 + k1 β16 2β14 k1 β17 + k2 β18 β25 2β19 (4. Hooke’s law would then give the stress σij (x). LLC . is necessary © 2005 by Taylor & Francis Group. 12.10 and ˆm 4. 4.1. for the internal displacement gradients ui.32) x ˆ β7 ˆ β8 ˆ β9 ˆ β10 ˆ β11 ˆ β12 (4.13). cannot be m used to ﬁnd internal stresses since the constants βα are only meaningful on the boundary of a body. described above..7.. . α = 1. then uses equations (4. one has the following results. an internal point representation of the diﬀerentiated BCE. one obtains βα from equation (4. the latter a regular oﬀcontour boundary point. are obtained by direct diﬀerentiation of equation (4. Therefore.34) k1 β13 + k2 β14 2(β16 + β21 ) symmetric ∂ 2 u3 ∂xi ∂xj = 2β18 β27 2β23 symmetric with k1 = −4(1 − ν) and k2 = −2(1 − 2ν) k2 β23 + k1 β24 2β22 k2 β19 + k1 β20 k1 β21 + k2 β22 2(β17 + β24 ) β26 (4.1 Surface displacements ˆ β1 ˆ [ui (x)] = β2 ˆ β3 4. m First.8).1.1. which are piecewise constant on each boundary element. The curvatures.9) together with all the βα on an element. An alternative approach is to use equation (4. SURFACE AND BOUNDARY CONTOUR EQUATIONS 77 called a regular contour point (RCP).j (p).7.7.36) Equation (4.33) can be used to ﬁnd the displacement gradients at x.33) x 4.2 Surface stresses ˆ β4 ˆ [ui.11) to get βα . 2.j (x)] = β5 ˆ β6 (4. Finally.4.1. It should be noted that the simple approach.35) (4. A point inside a body is called an internal point.
78
CHAPTER 4.
LINEAR ELASTICITY
(see Section 4.3). It is also of interest to examine the limiting process of a diﬀerentiated BCE as an internal point ξ (also denoted as p) approaches a boundary point x (also denoted as P ). This issue is of great current interest in the BEM community in the context of the standard and hypersingular BIE (HBIE  see, for example, [92], [39], [93], [111]). Further, the hypersingular BCE (HBCE) must be understood if one wishes to collocate the HBCE as the primary integral equation, as may be necessary, for example, in applications such as fracture mechanics, symmetric Galerkin formulations or adaptive analysis. This topic is the subject of Section 4.2 of this chapter.
4.2
4.2.1
Hypersingular Boundary Integral Equations
Regularized Hypersingular BIE
A hypersingular boundary integral equation (HBIE) can be obtained by diﬀerentiating the standard BIE at an internal point, with respect to the coordinates of this internal source point. A regularized version of this equation, containing, at most, weakly singular integrals (see equation (1.38) in Chapter 1) is: Uik,n (x, y) [σij (y) − σij (x)] nj (y)dS(y)
∂B
0
= −
Σijk,n (x, y) [ui (y) − ui (x) − ui, (x) (y − x )] nj (y)dS(y)
∂B
x, y ∈ ∂B
(4.37)
Martin et al. [93]  Appendix II2, p. 905, (see, also, [111]) have proved that (4.37) can be collocated even at an edge or corner point x on the surface of a 3D body, provided that the displacement and stress ﬁelds in (4.37) satisfy certain smoothness requirements. These smoothness requirements are discussed in Section 1.4.4 in Chapter 1. Please note that in (4.37) above, , n ≡ ∂yn , not the normal derivative.
4.2.2
Regularized Hypersingular BCE
The regularized HBIE (4.37) can be converted to a regularized hypersingular boundary contour equation (HBCE). Details are available in [99] and are given below. The ﬁrst task at hand is to transform equation (4.37) into a boundary contour form. The integrands in equation (4.37), without nj (y)dS(y), are ﬁrst ˆ evaluated at an internal ﬁeld point q very near Q (i.e. on a surface ∂ B inside the body, very near and parallel to ∂B), the derivatives are transferred from the kernels to the quantities inside the square brackets by the product rule, and then the limit q → Q is taken again. This is possible since the integrals in equation (4.37) are regular. The result is:
© 2005 by Taylor & Francis Group, LLC
4.2. HYPERSINGULAR BOUNDARY INTEGRAL EQUATIONS
79
0=
∂B
Uik (x, y) [σij (y) − σij (x)] − Σijk (x, y) ui (y) − ui
(L) ,n
nj (y)dS(y)
−
∂B
[Uik (x, y)σij,n (y) − Σijk (x, y) [ui,n (y) − ui,n (x)]] nj (y)dS(y) x, y ∈ ∂B (4.38)
where ui Note also that:
(L)
= ui (x) + ui, (x)(y − x ) ui,n = ui,n (x)
(L)
Therefore, the linear displacement ﬁeld ui gives the stress ﬁeld σij (x), so that the stress ﬁeld σij (y) − σij (x) is obtained from the displacement ﬁeld (L) ui (y) − ui . Using the identity [116] v,n = ∇ × (v × en ) (4.39)
(L)
(which is valid if the vector ﬁeld v is divergencefree) and Stokes’ theorem, the ﬁrst integral in equation (4.38), over Sm , is converted to the contour integral: I1 =
Lm
Uik (x, y) [σij (y) − σij (x)] − Σijk (x, y) ui (y) − ui
(L)
jnt dzt
(4.40) An explicit form of equation (4.40) is derived in Appendix A of this chapter. Next, it is noted from equations (4.32 and 4.10) that:
P
ui,n (x) = S,N β
i
(4.41)
where , N ≡ ∂/∂xn . Further, as proved in Appendix B of this chapter:
m 12 m
ui,n (y) = S,N β
i
+
α=4
R,N β
α−3
uαi (z1 , z2 , z3 )
(4.42)
Now, the second integral (called I2 ) in equation (4.38) is written as: I2 = −
∂B
[Uik (x, y)σij,n (y) − Σijk (x, y)ui,n (y)] nj (y)dS(y) Σijk (x, y)nj (y)dS(y)
∂B
−ui,n (x)
(4.43)
© 2005 by Taylor & Francis Group, LLC
80
CHAPTER 4.
LINEAR ELASTICITY
The next steps involve writing ∂B ≡ ∪Sm , separating the constant and linear parts of ui,n in the ﬁrst integrand above, and using equation (4.42). This sets the stage for converting I2 into two contour integrals. Details are given in Appendix C of this chapter. The ﬁnal result is a contour integral form (HBCE) of the regularized HBIE (4.37). This equation is valid at any point on the boundary ∂B as long as the stress is continuous there. This includes edge, corner and regular points, that lie on or oﬀ contours. The regularized hypersingular boundary contour equation (HBCE) is:
M 27 m
0
= −
m=1 α=13 M 12 Lm
(σ αij Uik − uαi Σijk ) (σ αij Uik − uαi Σijk )
m=1 α=4 M 12 Lm
jnt dzt
T −1 a
m α
+ −
m∈S / M
m=1
jst zs dzt
R,N T −1 a
m m
m
m α−3
(σ αij Uik − uαi Σijk )
α=4 3 Lm m
jnt dzt
R T −1 a − T −1 a
α−3 P
P
P
+
m∈S / M
m=1
Dαjk dzj
α=1 3 Lm
S,N
T −1 a − T −1 a
α m P
m
P
+
m∈S /
m=1
Σαjk
α=1 Lm
jnt dzt
S
T −1 a − T −1 a
α
m
P
(4.44)
where, as before, S is the set of boundary elements that contains the source point x . The derivatives R,N and S,N in (4.44) are taken with respect to the source point coordinates xn . In equation (4.44), the integrands in the ﬁrst two terms are regular (O(1)). The third and fourth (potentially strongly singular, O(1/r)) as well as the ﬁfth (potentially hypersingular, O(1/r2 )) need to be evaluated only on nonsingular elements.
4.2.3
Collocation of the HBCE at an Irregular Surface Point
Martin et al. [93] have stated the requirements for collocating a regularized HBIE at an irregular point on ∂B. This matter is discussed in Section 1.4.4 in Chapter 1 of this book. It has been proved in Mukherjee and Mukherjee [111] that the BCE interpolation functions given in equations (4.8) and (4.9) satisfy, a priori, all these smoothness requirements for collocation of the HBCE (4.44) at an irregular surface point. This proof is repeated below. It is ﬁrst noted that the interpolation functions in (4.8  4.9) have both a global (they are initially deﬁned as functions of y) as well as a local (the weights
© 2005 by Taylor & Francis Group, LLC
4.2. HYPERSINGULAR BOUNDARY INTEGRAL EQUATIONS
81
βk are only deﬁned piecewise on the boundary elements) character. Consider a singular boundary element containing the source and ﬁeld points P and Q, with P an irregular point on ∂B. Let this element be any one of the smooth pieces of ∂B that meet at P (see Section 1.4.4 in Chapter 1). From equations (4.8  4.9), it is easy to show that : ui (Q) − uL (Q; P ) i = ui (Q) − ui (P ) − ui,j (P )[yj (Q) − xj (P )]
27
=
α=13
βα uαi (z) = O(r2 ) ¯
27
(4.45)
σij (Q) − σij (P ) =
α=13
βα σαij (z) = O(r) ¯
(4.46)
where r = y(Q) − x(P ) = z. The last equalities in the above equations are true in view of the fact that the shape functions uαi and σαij are quadratic and ¯ ¯ linear, respectively, in zk , for α = 13, 14, ..., 27 (see Tables 4.1 and 4.2). Note that these weights βα belong to the element containing P and Q and are unique on that element (see below equation (4.9)). As an aside, it is interesting to connect with Toh and Mukherjee ([168]p.2304) where, for the same problem, the requirement ∇u(Q) − ∇u(P ) = O(rα ) is prescribed as r → 0. It is easy to show that (Mukherjee and Mukherjee [99]), for the BCM shape functions on a singular element:
12
ui,j (Q) − ui,j (P ) =
α=4
[R,J (x) β]α−3 uαi (z) = O(r) ¯
(4.47)
since uαi (z) is linear in z for α = 4, 5, ..., 12. ¯ In view of equations (4.45  4.46), conditions (iiiiv) in Box 1.1 (in Section 1.4.4 in Chapter 1) are satisﬁed a priori by the BCM shape functions deﬁned by equations (4.8  4.9). Satisfaction of condition (ii(b)) on ∂B follows from equation (4.46). The conditions inside B ((i) and (ii(a))), of course, have nothing to do with BEM shape functions that are only deﬁned on the bounding surface ∂B, but rather with the boundary element method itself. The BCM is derived from the BEM and it satisﬁes these internal point conditions in the same way as does the BEM. (As a bonus, the BCM shape functions satisfy the NavierCauchy equilibrium equations of linear elasticity a priori (see Section 4.1.2)), although weights are not deﬁned at points p ∈ B). Please note that the above arguments have been made for the “worstcase scenario” of the collocation point P ∈ ∂B being an irregular point. Of course, these arguments also go through for regular points (on or oﬀ contour) on ∂B. In view of the above, all the conditions (iiv) in Box 1.1 are satisﬁed a priori by the BCM, and there is no need to consider “relaxed smoothness requirements” in this method. It is worth repeating again that it is extremely diﬃcult to ﬁnd, in general, BEM shape functions (for 3D elasticity problems) that satisfy
© 2005 by Taylor & Francis Group, LLC
The primary reason for this is that BEM shape functions are deﬁned only on the bounding surface of a body. The © 2005 by Taylor & Francis Group.48) where z = y − ξ. The result is: uk (ξ) = 1 2 m=1 α=13 M 12 M 27 (σ αij (z)Uik (z) − uαi (z)Σijk (z)) Lm jst zs dzt T −1 a m m α + m=1 α=4 M 3 Lm (σ αij (z)Uik (z) − uαi (z)Σijk (z)) Dαjk (z)dzj m=1 α=1 Lm jst zs dzt R(ξ) T m −1 m a α−3 + S(ξ) T −1 a m m α (4.3). (See. They are given below. LLC .2 Displacements at Internal Points Close to the Bounding Surface This section is concerned with evaluation of displacements at internal points that are close to the bounding surface of a body. also.3.3.48). 4.3.24)) as a line integral by employing equation (16) in Liu [87]. 4. There are at least two ways of regularizing the solid angle term in equation (4. these terms are already regularized. except the solid angle. 4.3 of this book for a similar discussion related to the BEM. the regularized BCE (4. are evaluated as contour integrals. [104]). Since all other terms in equation (4. The general equations are derived in [103] and equations for internal points close to the bounding surface are derived in [104]. Please refer to Section 1.1 Internal Displacements One has to compare the regularized BIE (4.25) and the usual BIE at an internal point p = ξ (equation (1.82 CHAPTER 4. LINEAR ELASTICITY conditions (ii(b)iv) a priori. The ﬁrst is to evaluate the solid angle Ω (see equation (4. it is a simple matter to derive equations for displacements and stresses at a point inside a body.1).3 Internal Displacements and Stresses At this stage.4 concerns internal stresses also at points close to the bounding surface.21) in Chapter 1).48). while the BCM ones are deﬁned in B (although the weights are deﬁned only on ∂B). Section 4. ˆ The ﬁrst step is to choose the target point x at or close to the centroid of a boundary element (see Figure 1.
44) and the usual integral expression for the displacement gradient at an internal point p = ξ (equation (1.4. INTERNAL DISPLACEMENTS AND STRESSES 83 second is to use a boundary contour version of equation (1.e. The latter approach is adopted in this work. the numerator of the last integrand in equation (4. one has: T −1 a= β= β P m m m ˆ (4. the regularized HBCE (4. It is important to note that.n (ξ) = − m=1 α=13 M 12 Lm (σ αij (z)Uik (z) − uαi (z)Σijk (z)) (σ αij (z)Uik (z) − uαi (z)Σijk (z)) jst zs dzt jnt dzt T −1 a m m m α + m=1 α=4 Lm R.49) and still evaluate Ω as a surface integral.3 Internal Stresses This time.3. so that ˆ equation (4.50) ˆ ˆ Note that the point P has coordinates x. when integration is ˆ being carried out on an element that contains the point x). one has to compare the regularized HBIE (4.49) contains the solid angle term which is evaluated as a surface integral.n (ξ) T −1 a α−3 © 2005 by Taylor & Francis Group.49) is “nearly weakly singular” as y → x. The result is: M 27 m uk.37). It is useful to remember that the integral of Dαjk in equation (4.51) ˆ In this case.49) can be obtained easily.49) is O(r(ξ. on a singular element (i. LLC . x)) as y → x.3.49) where: ˆ ˆP uk (ˆ ) = βk = x 27 Skα (ˆ )βα x P α=1 ˆ (4. This equation is: uk (ξ) = uk (ˆ ) x + 1 2 m=1 α=13 M 12 M 27 (σ αij (z)Uik (z) − uαi (z)Σijk (z)) Lm jst zs dzt T −1 a m m m α + m=1 α=4 M 3 Lm (σ αij (z)Uik (z) − uαi (z)Σijk (z)) m jst zs dzt R(ξ) T −1 a m α−3 + m=1 α=1 Lm Dαjk (z)dzj S(ξ) T −1 a −S(ˆ )β P x m ˆ α (4. The boundary contour version of equation (1. 4.34) in Chapter 1). x)) ˆ ˆ while the denominator in the solid angle term is O(r2 (ξ.
A boundary contour version of equation (1.n (ˆ )]] nj (y)dS(y) x where (see Figure 1. y) ui (y) − ui x (L) . a boundary contour version of equation (1.4 Stresses at Internal Points Close to the Bounding Surface As before for the case of displacement evaluation at an internal point close to the boundary of a body (see start of section 4. LLC .n nj (y)dS(y) (4.2.51) is obtained in a manner that is quite analogous to the approach discussed in Section 4.2 (see. also.53) + ∂B [Uik (ξ.n (ˆ ) x − ∂B Uik (ξ.84 M 12 CHAPTER 4.51) is used here.2). (σ αij (z)Uik (z) − uαi (z)Σijk (z)) m=1 α=4 M 3 Lm LINEAR ELASTICITY R(ξ) T −1 a m m α−3 − + m=1 α=1 M 3 Lm jnt dzt Dαjk (z)dzj Σαjk (z) m=1 α=1 Lm S. (ˆ )ˆ x xz (4.54) uk.52) Hooke’s law is now used to obtain the internal stress from the internal displacement gradient.55) (L) = ui (ˆ ) − ui. The ﬁrst step is to use the product rule to transform equation (1.n (ξ) = uk.4.n (y) − ui. y) [σij (y) − σij (ˆ )] − Σijk (ξ. Curvatures at an internal point are given in equations (4. Again. for the sake of uniformity. y) [ui.34 . one has two choices with respect to the strategy for evaluation of the solid angle. [99]).3.3): ui with: z =y −x ˆ ˆ The BCM version of equation (1.n (ξ) T m −1 m a α jnt dzt + S(ξ) T −1 a m m α (4.n (ξ) = uk.3.36). together with evaluation of the solid angle as a surface integral. y)σij.51) is: (4.51) to the form: uk. 4.n (ˆ ) x © 2005 by Taylor & Francis Group.n (y) − Σijk (ξ.
the inner and outer radii of the sphere are 1 and 2 units.1 . These equations arise from uk. on the righthand side of equation (4. A generic surface mesh on a oneeighth sphere is shown in Figure 4. It should be noted that the ﬁrst.56) is “nearly weakly singular” ˆ ˆ (O(1/r(ξ. 4. Mesh statistics are shown in Table 4. k and n. while the second and fourth arise from the second integral in equation (4. are used in this work. Two levels of discretization . respectively. the shear modulus G = 1.n (ξ) T −1 a −S. the last but one term on the righthand side of equation (4.4. with summations and integrals.2.44) has two free indices.53).3. and the internal pressure is 1.n .4.25. for a thick hollow sphere under internal pressure.1 Surface Displacements from the BCM and the HBCM First.n (ˆ )β P x S(ˆ ) T x m −1 m m m ˆ α + α=1 Lm jnt dzt a −β P ˆ α (4. For the results in Section 4.4.4.3). For the results in Sections 4. and the internal pressure is 1. x)) as y → x). respectively. as in the case of equation (4.4.n (ξ) T −1 a × m m α−3 (σ αij (ˆ)Uik (z) − uαi (ˆ)Σijk (z)) z z α=4 Lm jnt dzt R(ˆ ) T −1 a −β P x M 3 m m ˆ α−3 + m=1 α=1 M 3 Lm Dαjk (z)dzj Σαjk (z) m∈S / m=1 S. the Young’s modulus E = 1. [109] and.3. so that it represents nine equations. the Poisson’s ratio ν = 0.56). 4. Typical results.49). are available in [116]. from the HBCM. for selected 3D examples. third and ﬁfth terms. arise from the ﬁrst integral in equation (4.56) ˆ ˆ where z = y − x (see Figure 1.4. Diﬀerent strategies are © 2005 by Taylor & Francis Group. NUMERICAL RESULTS M 27 85 T −1 a m m α jst zs dzt − m=1 α=13 M 12 Lm (σ αij (ˆ)Uik (z) − uαi (ˆ)Σijk (z)) z z (σ αij (z)Uik (z) − uαi (z)Σijk (z)) m=1 α=4 M 12 Lm jnt dzt + − m∈S / m=1 R.4. Again. the inner and outer radii of the sphere are 1 and 4 units.3. LLC . are given below. please note that (4. in [99].4 Numerical Results Numerical results from the BCM.medium and ﬁne.53). the Poisson’s ratio ν = 0.4.
the threeequation strategy.44) is replaced by [λEmm δkn + µ(Ekn + Enk )]nk (P ). for the HBCM in 2D elasticity. In this case.3: Mesh statistics on a oneeighth sphere (from [100]) possible for collocating (4.44). corresponds to the traction components τn .86 X2 CHAPTER 4. and collocate separately at these points. It should be mentioned here that. the righthand side of (4.n + un. only the nineequation and sixequation strategies are used below.2: A typical mesh on the surface of a oneeighth sphere (from [100]) mesh coarse medium ﬁne number of elements on each ﬂat plane on each curved surface 12 9 36 36 64 64 total 54 180 320 Table 4. A third. need to use multiple source points.44) at a boundary point. the righthand side (Ekn ) of (4. a strategy © 2005 by Taylor & Francis Group. but are convenient for collocating at irregular boundary points since the source point normal is not involved in these cases. therefore. δij are components of the Kronecker delta and Hooke’s e law is used. The second is to use six corresponding to kn = (1/2)(uk.37) at irregular boundary points. Since the primary purpose here is to demonstrate collocation of (4. unless the stress tensor is zero there. where λ and µ are Lam´ constants. LINEAR ELASTICITY X3 X1 Figure 4. each belonging to a smooth surface meeting at that irregular point. In view of the assumed continuity of the stress tensor at such a point. by (1/2)(Ekn + Enk ). The ﬁrst is to use all nine equations. suitable for collocation at regular points.k ). LLC . this situation leads to a jump in traction at that point. One would. The sixequation strategy amounts to replacing Ekn . The threeequation strategy involving the traction components is not convenient for collocating the HBCE at a point on an edge or a corner of a body where the normal to the body surface has a jump discontinuity. Both the nineequation and sixequation strategies lead to overdetermined systems.
show a comparison of the BCM (from equation (4.3 Radial displacement 0. The agreement between the numerical and analytical solutions is seen to be excellent. Exact solution: —.3 1.4. NUMERICAL RESULTS 0. resulting from the nineequation and sixequation strategies mentioned above.33) and Hooke’s law. The overdetermined system of linear algebraic equations. nineequation HBCM solution: ++++ (from [111]) corresponding to the ﬁrst one above has been successfully employed by [131] and a strategy corresponding to the third one above has been implemented by [187]. The nodes are chosen at the centroids of the boundary elements. © 2005 by Taylor & Francis Group. The ﬁrst and last points along the axis lie on corners.25)) and HBCM (from equation (4. Numerical solutions are obtained from the medium mesh. have been solved by using a subroutine based on QR decomposition of the system matrix. The agreement between the exact. LLC .8 1. It is seen from Figure 4. These results.3: Hollow sphere under internal pressure.7 1.2 that many of the collocation points lie on edges and six of them lie on corners of the surface of the oneeighth sphere.3.4. displayed in Figure 4. This subroutine has been obtained from the IMSL software package.44)) results with the exact solution of the problem [167]. the rest lie along an edge. BCM and nineequation HBCM solution is seen to be excellent.1 1. BCM solution: ◦ ◦ ◦◦.4. 4.6 1.35 87 0. sixequation HBCM solution: ∗ ∗ ∗∗.4.4 1.9 2 Radial coordinate Figure 4. Radial displacement as a function of radius along the x1 axis.1 1 1.2 0.5 1.25 0. are shown in Figure 4. obtained from equation (4.2 1.2 Surface Stresses Stresses on the inner (R = a) and on the outer (R = b) surface of the sphere.15 0.
1 0 0.3 0.4 1.6 0. Stresses as functions of radius along the line x1 = x2 = x3 . Exact solutions —.5: Hollow sphere under internal pressure.2 0. σRR ◦ ◦ ◦ ◦ (from [100]) 0.1 0.8 1.6 0.2 1.4 0.2 0 0.4 Stress components 0.5 1.4 0.8 0.5 1 0. Numerical solutions from the ﬁne mesh: σθθ = σφφ ∗ ∗ ∗ ∗.5 2 0 5 10 15 20 25 30 35 40 Stress components Nodes on the inner surface R=a Nodes on the outer surface R=b (a) (b) Figure 4.1 1.8 1 1.3 0.5 0.2 1 1. σRR ◦ ◦ ◦ ◦ (from [103]) © 2005 by Taylor & Francis Group. Exact solutions —.5 0 5 10 15 20 25 30 35 40 Stress components 1.5 0 0.4: Hollow sphere under internal pressure.88 CHAPTER 4.6 1. Numerical solutions from the medium mesh: σθθ = σφφ ∗ ∗ ∗ ∗.3 1.2 0. LLC .5 1 1.7 1.2 0. Stresses (a) on the inner surface R = a and (b) on the outer surface R = b.4 0. LINEAR ELASTICITY 2 0.9 2 Radial coordinate Figure 4.
5 2 2.35 1 1.08 1.75 0.4.5 3 3. Radial and circumferential stresses (σrr and σθθ ) as functions of radial distance from the center of the sphere. Radial and circumferential stresses as functions of the radial distance from the center of the sphere.08 1. together with exact solutions.7 0.95 1 0.12 0.12 Radial coordinate Radial coordinate (a) (b) Figure 4.1 1.4.4 1 1.6: Hollow sphere under internal pressure.5 Stress components 0 0.06 1.02 1. LLC .9 0.5 BCM solution for σ θθ BCM solution for σrr Exact σ θθ Exact σ rr 1 1 1.65 0.45 0.1 1.5 4 Radial coordinate Figure 4.06 1. NUMERICAL RESULTS 89 0. for points very close to the inner surface of the sphere (from [104]) 0. The new and standard BCM solutions from the ﬁne mesh.02 1.04 1.85 0. Results from the new BCM (ﬁne mesh) together with the exact solution (from [104]) © 2005 by Taylor & Francis Group.8 0.7: Hollow sphere under internal pressure.5 : Exact solution : BCM new : BCM standard Circumferential stress Radial stress 0.04 1.55 : Exact solution : BCM new : BCM standard 0.
Finally. Figure 4.3 Internal Stresses Relatively Far from the Bounding Surface Internal stresses along the line x1 = x2 = x3 .7 gives the global picture for stresses throughout the sphere.90 CHAPTER 4.56)) BCM. respectively. 4.6 (a) and (b). even at points that are very close to the surfaces of the hollow sphere. the integral in equation (4.33)). obtained from equation (4.4. z2 .4. are shown in Figures 4. 2. The results from the standard BCM exhibit large errors whereas results from the new BCM faithfully track the exact solutions in both cases. z2 . LLC . Appendix A An Explicit Form of Equation (4. (x)[y − x ] = α=4 ˆP βα uαi (z1 .52)) and the new (equation (4. y) α=1 ˆm ˆP (βα − βα )uαi + α=13 ˆm βα uαi jnt dzt where the fact that σ αij = 0 for α = 1. 3 has been used. z3 ) 12 ui. z3 ) (the last equation above can be proved from equation (4.40) can be written as: 12 27 I1 = Lm Uik (x. The new BCM performs beautifully.4 Internal Stresses Very Close to the Bounding Surface Numerical results for stress components. © 2005 by Taylor & Francis Group. z2 . appear in Figure 4. from the standard (equation (4. LINEAR ELASTICITY 4. z3 ) ui (x) = α=1 ˆP βα uαi (z1 .5. y) α=4 ˆm ˆP (βα − βα )σ αij + α=13 12 ˆm βα σ αij 27 jnt dzt − Lm Σijk (x.52) and Hooke’s law. Excellent agreement is observed between the numerical and analytical solutions.40) Using the equations : 27 ui (y) = α=1 3 ˆm βα uαi (z1 .
Now. are constant.4.n = S. z2 . and the constants k1 and k2 are deﬁned in Table 4. 2.n (y)linear ≡ α=13 ˆm βα uαi. 27 m m m m m β α uα. z3 ) α=4 ui. α = 1.1 © 2005 by Taylor & Francis Group.1 β α−3 uα (z1 . z2 . Let 27 27 ui. a vector displacement shape function is deﬁned as: uα = [uα1 . 12 ui.. z3 ) Now.3.1 α=13 = 2 β 13 + β 20 u4 + 2 β 15 u5 + 2 β 18 u6 m m +. In the above. z3 ) ˆm βα uαi. NUMERICAL RESULTS 91 Appendix B Proof of Equation (4.n (y) = since uαi .n (z1 .N β m i where the last equality is obtained by observing equation (4. with n = 1. z3 ) Similar expressions can be obtained for n = 2.n (z1 .42) 27 ui (y) = α=1 27 ˆm βα uαi (z1 .. LLC . z2 . 3.4...41).n (z1 . z2 .. + k2 β 19 +k1 β 20 u12 12 m = α=4 R. z2 . . z3 ) = α=13 m β α uαi.. uα2 .n (y)constant ≡ α=4 ˆm βα uαi. uα3 ] T (where T denotes the transpose of the vector).
y)uαi ] nj dS m 3 m R.43) m I 2= 12 − α=4 3 Sm [Uik (x. LINEAR ELASTICITY Appendix C Conversion of Equation (4. y)nj dS Applying Kaplan [69]’s formula (see [116].41)). while.26)): m I 22 = 3 − α=1 Lm m Dαjk dzj S.N β α−3 The remaining terms cancel on a singular element (see equation (4. on a nonsingular element. LLC .N β α − α=1 uα.n (x) Sm Σαjk (x. y)nj dS S.92 CHAPTER 4.N β −β α P © 2005 by Taylor & Francis Group. one gets (see equation (4. to the ﬁrst term above (its integrand is O(1/r)). one gets the contour integral: m I 21 = 12 − α=4 Lm (σ αij Uik − uαi Σijk ) m jst zs dzt R. y)σ αij − Σijk (x. [109]).N β α−3 + α=1 Sm Σαjk (x.
Further details are available in [100. is the subject of this chapter.1.Chapter 5 SHAPE SENSITIVITY ANALYSIS Shape sensitivity analysis of surface and internal displacements and stresses. i. 103]. LLC . ξ(b). b)nj (y)dS(y) ∂B ≡ (5. y. b)]] nj (y)dS(y) Fjk (ξ. b) − Σijk (ξ. at an internal point ξ ∈ ∂B is rewritten as: 0 = ∂B [Uik (ξ. y)σij (y. y(b).1 Sensitivity of the BIE Equation (4.1 Sensitivities of Boundary Variables The starting point in this chapter is the standard BIE (1. From equation (5. 5.1). b is a shape design variable and the spatial coordinates of the source and ﬁeld points depend on b .1).) The sensitivity (total or material derivative) of this equation is taken with respect to a design variable b. The resulting sensitivity BIE is split into three parts.1) As mentioned above. obtained from the BCM for 3D linear elasticity.e.26) collocated at an internal point ξ . (This regularized BIE is valid both at an internal point ξ ∈ B as well as at a boundary point x ∈ ∂B. b) − ui (ξ. one has: 93 © 2005 by Taylor & Francis Group. y)[ui (y. The ﬁrst part vanishes and the surface integrals in the second and third parts are systematically converted into contour integrals by using Stokes’ theorem. 5.
y.94 CHAPTER 5. LLC ∗ ∗ ∗ (5. for ui (y. is considered ﬁrst. b) − ui (ξ. y.7) ∂F (ξ.y. 5.8) . in a materials derivative sense. b). b)− ui (ξ.3) while its partial sensitivity is deﬁned as : f≡ It should be noted that.2) Now the (total) sensitivity of a function f (ξ(b).k y k .b) Each of these surface integrals will be converted to line integrals in the next two sections. b) nj (y)dS(y) ∂ξr ∂B ∂Fjk (ξ. From Bonnet and Xiao [15] (see also Petryk and Mr´z [128]): o nj = − y r. y)(ui (y. b).m nr nm nj © 2005 by Taylor & Francis Group.1) vanishes for all values of ξ ∈ B. As mentioned before. b) − Σijk (ξ. b) ∂B d [nj (y)dS(y)] db Uik (x. ∂ ∂yk . SHAPE SENSITIVITY ANALYSIS Fjk (ξ. The ﬁrst integral on the righthand side of equation (5. b) − Σijk (ξ. b) ∗ y r nj (y)dS(y) + + ∂yr ∂B + ∂B ∗ 0 = ξr Fjk (ξ.4) σ ij =σ ij −σij. where . inside the body. is deﬁned as : f≡ ∗ ∂f ∗ df ∂f ∗ ∂f yk + = ξk + db ∂ξk ∂yk ∂b ∂f ∂b ∗ ∗ (5. y.2 The Integral Ik The surface integral Ik is converted to a sum of contour integrals in this section.6) It should be noted that the last integrand above is jk∂b .1) with respect to b. y(b). ∗ ∗ (5. b): ui =ui −ui. y)[ui (y. y)σij (y. k ≡ Taking the sensitivity (total derivative) of equation (5.5) and similarly for ui (ξ. b)] (5.6) is zero because the integral in equation (5. one gets: ∂Fjk (ξ.j nr + y r. y) σ ij (y.1. b) = Uik (ξ.k y k (5. Thus: Ik + Jk = 0 (5. y. a point ξ. b)) nj (y)dS(y) (5. Let the second and third integrals together be called Ik and the last integral Jk .
r − y r.j nr ) dS(y) = Sm Lm qjr F dyq (5.here ξ is an internal point).6).r y r nj + Fjk y r.5.12) Let ∂B = ∪Sm and let Lm be the bounding contour of the boundary element Sm .j nr dS db Using equation (5.17) © 2005 by Taylor & Francis Group.r nj dS = 0 ⇒ m=1 ξr Lm qjr Fjk dyq =0 (5. I = 0. Therefore.13) with F ≡ Fjk y r . One can show that: M qjr Fjk dyq = m=1 Lm ∂B M ∗ Fjk.r nj − y r. SENSITIVITIES OF BOUNDARY VARIABLES ∗ 95 ∗ dS ∗ =y r.r nj − Fjk y r.1. a surface integral I (over a closed surface ∂B) of any divergencefree vector function F(y) .9) so that: ∗ ∗ d [nj dS] = y r.15) has the sensitivity expression : I= m=1 Lm ∗ ∗ M jnt Fj (y) y n dyt ∗ (5. from Gauss’ theorem. of the form: I= ∂B Fj (y) nj (y) dS(y) (5.j nr dS (5. one gets: M ∗ Ik = m=1 Lm qjr Fjk y r dyq ∗ (5. one has: (5.j nr dS ∗ ∗ ∗ (5. Using Stokes’ theorem in the form (see Toh and Mukherjee [168]): (F.14) It is useful to state here that formula (5.m nr nm dS (5.10) Ik = ∂B Fjk. I = 0. in the sense that.r nj − F.11) Since Fjk. the above expression can be written as: Ik = ∂B Fjk y r ∗ .16) Of course.r nj − Fjk y r ∗ .j = 0 (except at a point of singularity .14) is a general one. LLC .10) in (5.
Therefore. z) = a0 (x) + a1 (x)z + a2 (x)z 2 ˆ ˆ ˆ where: a0 ˆ a1 ˆ a2 ˆ = a0 + a1 x + a2 x2 = f (x) d f (x) = a1 + 2a2 x = dx = a2 (5.2) into (5. Also. dyq = dzq since dxq = 0 at a ﬁxed source point.1. are derived next. b))] jnt z n dzt (5. Let: f (y) = a0 + a1 y + a2 y 2 With the change of variables: y =x+z one can write: f (x. the above expression is completely regular.a simple example Series expressions for the partial sensitivities of displacements and stresses.1.1 Shape functions for partial sensitivities . y)(ui (y. LLC .96 ∗ ∗ ∗ CHAPTER 5.3. Substituting equation (5. b) − Σijk (ξ.22) It is easy to show that taking sensitivities of the above equations results in: f (y) =a0 + a1 y+ a2 y 2 + a1 y +2a2 y y ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ (5. one gets: M ∗ Ik (ξ) = m=1 ∗ Lm [Uik (ξ. y)σij (y. b) − ui (ξ.20) (5.14) (with y replaced by z) .14). It is useful to start with a very simple example.19) (5. 5.18) Since z n is O(r) as r = y − x → 0.23) f (y) = f (y) − f. 5. one can replace y r with z r in equation (5.24) © 2005 by Taylor & Francis Group. it remains valid at a boundary point x ∈ ∂B. SHAPE SENSITIVITY ANALYSIS ∗ ∗ Since y r =ξ r + z r . in terms of global BCM shape functions.y y = a0 + a1 y+ a2 y 2 ∗ ∗ ∗ (5.3 The Integral Jk The surface integral Jk is converted to a sum of contour integrals in this section.21) (5.
the partial sensitivities for the stresses can now be expressed as: 27 ∗ 27 σ ij = α=1 β α σ αij (y1 . the partial sensitivity of f (x) is its sensitivity at a ﬁxed point in space. are related to ak in the same manner as ak are related ˆ to ak in equation (5. x2 .3 The ﬁnal form of Jk The conversion procedure is entirely analogous to the derivation of the primary BCM equation presented in Chapter 4.27. it is easy to show that: 27 ∗ 27 ui = α=1 β α uαi (y1 . k = 4.26) and (5.20). y2 . 5.2 BCM interpolation functions The displacement and stress interpolation functions for the BCM are considered next.3. n = k − 3 α=1 ∗ ∗ (5. k = 0. z3 ) (5. LLC ..1.1. x3 ) β α . Finally.11 and 4.3. z2 ... 5.12). β α .25) where ak .10.27) = (5.29) = βα . z) =a0 (x)+ a1 (x)z+ a2 (x)z 2 ∗ (5.30) are ﬁrst substituted into the last integral on the righthand side of the sensitivity BIE (5. 3 Rnα (x1 . 27 Of course.12. The result is: © 2005 by Taylor & Francis Group. . 14.6). . Series expansions (5. are related to β α in the same manner as βα are related to βα (see equations 4. α = 13.: 27 ∗ βi βk βα = α=1 27 Siα (x1 . x2 . The potential functions are the same as before.26) ˆ where. x3 )uαi (z1 . y3 ) = α=1 β α (x1 . z2 .22). 2.30) 5. SENSITIVITIES OF BOUNDARY VARIABLES 97 As expected. x2 . i = 1. Following the procedure outlined above for the simple example. 4. 1.. one has: f (x..28) (5.e. One starts with equation (4. i. 2. with the change of variables (5. α = 1.8) for the displacements. x3 ) β α . z3 ) ∗ (5.5. x2 . x3 )σ αij (z1 . the surface integral Jk is converted into a sum of contour integrals.. y3 ) = α=1 β α (x1 .1. 2. y2 . . Now.
9) for σij into (5. the explicit expression for Ik (obtained from (5.8) for ui and (4.4. 5. Finally.12) with (5.7) evaluated at a general boundary point x ∈ ∂B. 3. k = 1. Next. k = 1. and (5. For example.29).1. Comparing equations (4.32) in which it is convenient to evaluate the sensitivity of T −1 from the formula: (T −1 )∗ = −T −1 T T −1 ∗ ∗ (5. β k =uk (P ). .29) for β α are substituted into (5.18)) and the explicit expression for Jk (obtained from (5.31)) are input into the BIE sensitivity equation (5. LLC .31) + α=1 Lm βα − βα m P P It is interesting to comment on the physical meaning of the quantities β α for the case when the surface source point P is a regular oﬀcontour boundary point (ROCBP).32 . 2.18).6) (see also 5.98 CHAPTER 5.27 ..4 The BCM Sensitivity Equation An explicit form of the BCM sensitivity equation is now derived.36). The result is: M 27 m 0 = m=1 α=13 M 12 Lm (σ αij Uik − uαi Σijk ) + m=1 α=4 Lm jnt z n dzt z n dzt ∗ ∗ T −1 a m α (σ αij Uik − uαi Σijk ) jnt R T −1 a m m α−3 © 2005 by Taylor & Francis Group. 3. 2. the quantities β α can be easily interpreted in terms of the partial sensitivities of displacements and their derivatives from P P equations (4.10 . etc. SHAPE SENSITIVITY ANALYSIS Jk (x) = 1 2 m=1 α=13 M 12 M 27 (σ αij Uik − uαi Σijk ) Lm m jnt zn dzt βα m + m=1 α=4 M 3 Lm (σ αij Uik − uαi Σijk ) Dαjk dzj m∈S / m=1 jnt zn dzt βα (5.4. On any boundary element: β= T −1 a +(T −1 )∗ a ∗ ∗ (5. 27. it is clear ˆ that βk =β k . an explicit expression for Ik is obtained by substituting the series expressions (4. Now.5..33) Expressions (5.27 .5.32) is used to write β in explicit form.31).
34) Comparison of the above sensitivity equation (5.2 Sensitivities of Surface Stresses ∗ The ﬁrst step is to use equation (5. its discretized form can be written with the same coeﬃcient matrix A as for the standard BCM. i. the usual switching of columns leads to: A x= B y +h ∗ ∗ ∗ ∗ (5. Therefore.36) where x contains the unknown and y the known values of boundary sensitivities.5. y = 0 5.32) to ﬁnd β on each element. ∗ In many applications. © 2005 by Taylor & Francis Group. SENSITIVITIES OF SURFACE STRESSES M 3 99 m P P − m∈S / M m=1 Σαjk α=1 27 Lm jnt z n dzt ∗ S T −1 a − T −1 a α m + 1 2 m=1 α=13 (σ αij Uik − uαi Σijk ) Lm jnt zn dzt × T −1 a M 12 m ∗m + T −1 m ∗ m a α + m=1 α=4 Lm (σ αij Uik − uαi Σijk ) R T −1 a M 3 m ∗m jnt zn dzt × + T −1 m ∗ m a α−3 + m∈S / m=1 Dαjk dzj × α=1 Lm m P m ∗ P ∗ P S T −1 a ∗m − T −1 a + T −1 ∗P m a − T −1 a α (5. LLC .e. There are at least four ways to ﬁnd stress sensitivities on the surface of a body.25) reveals that the integrals in its last three terms are identical to those in the standard BCE.34) with the standard BCE (4.35) where the righthand side vector h = − K a can be computed from equation (5.34) by using the boundary values of the primary variables a that are known at this stage. : K a= h ∗ ∗ (5.2. Finally.
x2 . x3 ) β α + α=1 Rnα (x1 .2. y 2 . x3 ) Finally.37) Note that σ αij (y 1 .40) © 2005 by Taylor & Francis Group.k xk Now: vij.12.41) (5.. from equation (4. 5. by diﬀerentiating equation (4.j are computed by diﬀerentiating equation (4. y3 ). y 3 ) ∗ ∗ ∗ ∗ (5. Hooke’s law is used to determine the stress sensitivities from the sensitivities of the displacement gradients. x2 . SHAPE SENSITIVITY ANALYSIS 5. y3 ) + α=1 ∗ 27 βα σ αij (y 1 .2.38) x where.2. y2 . x3 )βα (5.1 Method One Equation (4. one has: ∗ 27 ∗ 27 ˆ βk = α=1 Rnα (x1 .. x2 .11).k = ui.3 Method Three ∗ ∗ One writes: v ij = v ij +vij.jk Also. 5.33). y2 .100 CHAPTER 5. x3 ) involves sensitivities of the source point coordinates (x1 . 5. x2 .9) is diﬀerentiated to give: σ ij = α=1 ∗ ∗ ∗ ∗ ∗ 27 β α σ αij (y1 .2 Method Two Sensitivities of displacement gradients vij ≡ ui. LLC . The result is: v ij (x) ∗ ˆ β4 ∗ = β ˆ 5 ∗ ˆ β 6 ∗ ˆ β7 ˆ β8 ˆ β9 ∗ ∗ ∗ ˆ β 10 ∗ ˆ β 11 ∗ ˆ β ∗ 12 (5. . y 2 . n = k − 3 ∗ ∗ ∗ ∗ Note that Rnα (x1 .39) ∗ ∗ ∗ ∗ k = 4.33): (5. y 3 ) involves sensitivities of the ﬁeld point coordinates (y1 .
3 .2. x3 ) β α .. equation (5. 2. It is ﬁrst observed that the operators .j = (β i ). one has: P v ij = (ui.. The term vij.4 Method Four The starting point is.21). Further details are available [103].j It follows from equation (5. j = 1.27) that: P 27 (5. x3 ) β α (5. 5. .1 Sensitivities of Displacements The starting point is equation (1.40). from equations (4. 5.j = α=1 Siα. k P and commute and that ui (P ) =β i for i = 1.34 . Therefore. 3 (5. ∗ i.3.k on the righthand side of equation (5. 3. 5. x2 ..3. from equations (4. 12. α = 13.28): 27 ∗ ˆ βk = β k = α=1 Rnα (x1 .. 2.27.42) x with. the displacement BIE at an internal point ξ.40) are available in terms of βα . x2 . The other term.45) 5. LLC .j (x1 . is treated diﬀerently. 14. SENSITIVITIES OF VARIABLES AT INTERNAL POINTS v ij (x) ˆ β4 = β ˆ 5 ˆ β 6 101 ˆ β7 ˆ β8 ˆ β9 ˆ β 10 ˆ β 11 ˆ β 12 (5. n = k − 3 The curvature expressions needed in equation (5.43) k = 4. . again.j ) = (ui ).5.40) is treated in the same fashion as in Section 5. v ij .4.44) v ij = (β i ).11) and (5.36)..3 Sensitivities of Variables at Internal Points Boundary contour integral equations of the sensitivities of internal displacements and stresses are derived in this section. This equation is written as: © 2005 by Taylor & Francis Group.2.
SHAPE SENSITIVITY ANALYSIS uk (ξ. b) ∗ y r nj (y)dS(y) + ∂yr ∂B d + Gjk (ξ. b) − Σijk (ξ.2. y. b)] nj (y)dS(y) (5. the second and third integrals together be called J2k and the last integral be J3k . y)σij (y. b) ∗ (5.48) Let the ﬁrst term on the righthand side of equation (5.r (ξ) ξ r ∗ ∗ (5. b) = J1k + J2k + J3k It is obvious that: J1k = uk.47) = ξr ∂Gjk (ξ. b) nj (y)dS(y) (5. y)σij (y. b) = ∂B [Uik (ξ.47) into (5. y)ui (y.53) ∗ © 2005 by Taylor & Francis Group. 5.p (ξ) ξ p M ∗ + m=1 Lm [Uik (ξ.50.1. y)ui (y.49).52 and 5. y)ui (y. one has: uk (ξ. y)σij (y.51) Finally. b) nj (y)dS(y) ∂ξr ∂B ∂Gjk (ξ. b) [nj (y)dS(y)] db ∂B + ∂B ∗ Uik (ξ. b)] jnt z n dzt + uk (5. b) ∗ = uk. one obtains the expression: uk (ξ.50) Using exactly the same procedure described in Section 5.49) (5.102 CHAPTER 5. y) ui (y.46) Deﬁne: Gjk (ξ. LLC . one gets: uk (ξ. one gets: M J2k = m=1 Lm jnt Gjk z n dzt ∗ (5. b) Taking the sensitivity of equation (5. y.52) Substituting equations (5.48) be called J1k . y. Therefore. b) − Σijk (ξ. 5. b) = Uik (ξ. b) − Σijk (ξ. b) − Σijk (ξ.51. y. y) σ ij (y. J3k =uk (5.46).
8 and 4.r (ξ. b) = − ∂B [Uik.3 for details of the treatment of the partial sensitivity uk .55) Now. the boundary contour integral form of the displacement sensitivity equation is: ∗ ∗ uk (ξ. b) − Σijk. b) ≡ uk.3. Finally.54) 5.3.9). the starting point is the displacement gradient BIE (1.r (ξ.r (ξ) ξ r M 27 + m=1 α=13 M 12 Lm (σ αij Uik − uαi Σijk ) (σ αij Uik − uαi Σijk ) m=1 α=4 M 3 Lm jnt z n dzt ∗ ∗ T −1 a m m α + − m=1 α=1 Lm jnt m z n dzt m α RT m −1 m a α−3 Σαjk 1 2 m=1 α=13 M 27 jnt z n dzt ∗ S T −1 a + (σ αij Uik − uαi Σijk ) Lm jnt zn dzt × T −1 a M 12 m ∗m + T −1 m ∗ m a α + m=1 α=4 Lm (σ αij Uik − uαi Σijk ) R T −1 a M 3 m ∗m jnt zn dzt × + T −1 S T m ∗ m a α−3 + m=1 α=1 Lm Dαjk dzj m m −1 ∗ a + T −1 m ∗ m a α (5.5. b)] nj (y)dS(y) (5.2 Sensitivities of Displacement Gradients and Stresses This time. which is written as: vkr (ξ. Please refer to Section 5. SENSITIVITIES OF VARIABLES AT INTERNAL POINTS 103 An explicit form of equation (5.1. y)σij (y.53) is obtained by writing the displacements and stresses in terms of their interpolation functions from equations (4. LLC .r (ξ. one deﬁnes: © 2005 by Taylor & Francis Group. b) = uk.28). y)ui (y.
This is: ∗ ∗ v kr (ξ. b) One has Hjkr.r (ξ. one gets: (uk ). b) − Σijk. b) = uk. (z n ). b) − Σijk. b) nj (y)dS(y) (5. Diﬀerentiating equation (5. b) nj (y)dS(y) (5. ∂Gjk (ξ.rp (ξ) ξ p +uk. b)] m=1 Lm ∗ ∗ ∗ ∗ (5.rp (ξ) ξ p M − m=1 Lm [Uik. b) − Σijk. (see Section 5.59) Normally. y) = − and Gjk. y)σij (y.j = 0.r (ξ. SHAPE SENSITIVITY ANALYSIS Hjkr (ξ.58) from another point of view.2) z n =y n − ξ n .r (ξ.1 leads to an equation for the sensitivities of displacement gradients at an internal point. LLC .56) (5.60) However.r = −(ξ n ). y)σij (y.r M ∗ ∗ − m=1 Lm [Uik. y) σ ij (y.r (ξ.r and: M (ξ n ). b) − Σijk.r (ξ.3.57) The exact same reasoning as in the previous section 5. y)ui (y. y) ui (y. one would expect another term on the righthand side of the above equation. b) − Σijk (ξ. y)ui (y.p (ξ)(ξ p ).1.58) It is very interesting to verify (5. y. namely: M ∗ jnt (z n ).j = 0 at an internal point since: Hjkr (ξ. y)σij (y.53) with respect to ξr .r (ξ.r (ξ. y) σ ij (y. b)] jnt z n dzt ∗ − ∂B Uik. y)ui (y.r (ξ.104 CHAPTER 5.r dzt ∗ [Uik (ξ. b) jnt dzt =0 (5. y) ui (y.r m=1 Lm ∗ Gjk (ξ. y)σij (y. b) − Σijk.61) © 2005 by Taylor & Francis Group. b)] jnt z n dzt ∗ − ∂B Uik.r (ξ.r (ξ) ∗ = uk. y) ∂ξr (5. y)ui (y.r (ξ. y) = Uik.
r jnt z n dzt ∗ S T −1 a m m α − m=1 α=13 Lm m (σ αij Uik − uαi Σijk ) T −1 a M 12 ∗m jrt dzt × + T −1 m ∗ m a α + m=1 α=4 Lm (σ αij Uik − uαi Σijk ) R. one writes the displacements and stresses in terms of their interpolation functions in order to obtain an explicit form of equation (5. is analogous to an expression for the displacement gradient at an internal point. It should be noted that the second term in the righthand side of equation (5. LLC . Now.p (xp ).58) and (5.rp (ξ) ξ p M 27 − m=1 α=13 M 12 Lm (σ αij Uik.58) is analogous to the integral on the righthand side of equation (5.r )∗ = (φ). Finally.58).59) are consistent.r − φ.5. while the last term in equation (5.58). [63]): (φ.r ) (σ αij Uik. for example. using the wellknown formula (which is valid for any suﬃciently smooth function φ . The ﬁnal result is: ∗ ∗ v kr (ξ. The displacement gradient BCE (4.r − uαi Σijk. Therefore.60) vanishes.3.17)).r − uαi Σijk.see.52) is very useful for obtaining an explicit expression for this integral.62) with φ = uk . SENSITIVITIES OF VARIABLES AT INTERNAL POINTS 105 since Gjk is divergence free (see (5. v kr .r ∗ ∗ (5.r T −1 a M 12 m ∗m jnt zn dzt m × + T −1 m ∗ a α−3 − m=1 α=4 Lm (σ αij Uik − uαi Σijk ) jrt dzt × © 2005 by Taylor & Francis Group. b) = uk. it is easy to show that equations (5. the expression in (5. Haug et al.53) (with G replaced by H).r ) m=1 α=4 Lm jnt z n dzt ∗ ∗ T −1 a m m α − jnt z n dzt × R T −1 a M 3 m m α−3 + m=1 α=1 M 27 Lm Σαjk.
rp (ξ) = m=1 α=13 M 12 Lm (σ αij Uik.4 Numerical Results: Hollow Sphere Sensitivity results are presented in this section for a thick hollow sphere under internal pressure (see Section 4.p T −1 a m m α−3 + m=1 α=4 M 3 Lm (σ αij Uik.3.rp T −1 a m α ∗ (5. LLC . stress sensitivities can be easily obtained from v kr by using Hooke’s law. medium and ﬁne.4).3 and the internal pressure is 1 unit.2. respectively. © 2005 by Taylor & Francis Group. A generic surface mesh on a oneeighth sphere is given in Figure 4. the inner and outer radii of the sphere are 1 and 2 units.r ) Σαjk m=1 α=1 M 3 Lm jrt dzt jpt dzt R T −1 a m α−3 + − m=1 α=1 M 3 Lm S. As before. Three levels of discretization.r T m m −1 ∗ a + T −1 + T −1 m m ∗ m a α jrt dzt + S T −1 a m ∗m ∗ m a α (5. The design variable is the inner radius a of the sphere. 5.106 CHAPTER 5.rp in the above equation can be obtained from (see [103]): M 27 m uk.r − uαi Σijk. are used in this work.p T ST m m −1 m a α Σαjk.r ) − m=1 α=4 M 12 Lm jpt dzt T −1 a m m α (σ αij Uik − uαi Σijk ) jrt dzt R. coarse. Mesh statistics are shown in Table 4.64) Finally. the Poisson’s ratio ν = 0.r − uαi Σijk.r jpt dzt m −1 m a α + m=1 α=1 Lm Dαjk dzj S. the shear modulus G = 1. SHAPE SENSITIVITY ANALYSIS R T −1 a M 3 m ∗m + T −1 m ∗ m a α−3 + m=1 α=1 M 3 Lm Dαjk dzj Σαjk m=1 α=1 Lm S.63) The curvatures uk.
65) b−a is used here.4.7 0.2).1.9 2 0.5 0.2 0.5. However.6 0. along various lines on the sphere surface.38 . x2 = x3 . Numerical solutions: coarse mesh ◦ ◦ ◦◦.1 0 1 1. NUMERICAL RESULTS: HOLLOW SPHERE 107 Sensitivity of radial displacement Sensitivity of radial displacement 0.1.5 0.3 0.1 1.4 1.6 1. is somewhat tricky on the inner surface R = a.4 1.5 1. however.3 1. The situation.2 0. Exact solution —. they do appear to converge to the exact solution with increasing mesh density. It is seen that the numerical results for the coarse mesh exhibit large errors. (Stress proﬁles for this example appear in Figure 4.1 5.6 0.1. for diﬀerent discretizations.1 Sensitivities on Sphere Surface Displacement sensitivities Displacement sensitivities (from equation (5. agree well with the exact solution (Figure 5.3 0. LLC .34)). Numerical solutions from the medium mesh.7 1.8 1.2 1. ﬁne mesh ++++ (from [100]) 5.4 0.4 0.2 1.9 2 Radial coordinate Radial coordinate (a) (b) Figure 5.39).) A linear design velocity proﬁle: b−R (5.4.5.2 Stress sensitivities Stress sensitivities on the surface of the sphere are calculated from equations (5. for stress sensitivities on the outer surface R = b.5 1.1 1.1: Sensitivity of radial displacement along (a) the x3 axis and (b) along the line x1 = 0. (A typical displacement proﬁle for this example appears in Figure 4.4. one has: © 2005 by Taylor & Francis Group. The reason for this is under investigation. R = ∗ 5.3 1.7 0.7 1.8 1. appear in Figure 5.1 0 1 1. medium mesh ∗ ∗ ∗∗.6 1.4. Here.3.4). Please see Chandra and Mukherjee [22] for a discussion of analytical solutions for design sensitivities for various examples.
66) R ∂R (and similarly for the other components of stress). this ∗ ∗ situation makes it quite diﬃcult to calculate σ θθ and σ φφ accurately. 5.2 5. LLC .8 1 0 5 10 15 20 25 30 35 40 Nodes on the outer surface R=b Figure 5.6 0. are quite high.4. ∗ ∗ ∗ Numerical solutions from the medium mesh: σ RR + + + +. respectively. which is negative. These results are obtained from equation (5.1 shows percentage root mean square errors in σ θθ and σ φφ . which is positive.35.108 1 0. Exact solutions —.1 Sensitivities at Internal Points Sensitivity of radial displacement Figure 5. those on the inner surface. The exact solution for σ θθ =σ φφ is 120/49.2 0 0. σ θθ ◦◦◦◦.4 0.6 0. SHAPE SENSITIVITY ANALYSIS Stress sensitivities 0.54) which is used after ﬁrst © 2005 by Taylor & Francis Group.67) It is seen that while the errors on the outer surface are very low.2 0. As can be seen from Figures (5. Perhaps further work on this problem. needs to be carried out.4 0.4). even with the ﬁne mesh. σ φφ ∗∗∗∗ (from [100]) ∂σθθ ∗ (5.5(a) shows the sensitivity of radial displacement along the line x1 = x2 = x3 .2. because one must calculate the diﬀerence between two numbers that are reasonably close. These are deﬁned as: σ θθ = σ θθ + ∗ = 100 fexact 1 n n (fexact − fi i=1 2 numer ) (5. while the convected term in equation (5.4. ∗ ∗ Table 5.8 CHAPTER 5. especially when accuracy is measured in terms of percentage errors.66) is −12/7. including the development of a diﬀerent algorithm for calculation of surface stress sensitivities.2: Stress sensitivities on the outer surface R = b.
49 ﬁne mesh 1.01 15.3: Stress sensitivities on the inner surface R = a. Numerical solutions from (a) the medium mesh and (b) the ﬁne mesh: σ θθ ◦ ◦ ◦◦.62 2.4.5.88 Table 5.1: Percentage root mean square errors in sensitivities of stress components (from [100]) 4 4 Stress sensitivities Stress sensitivities 3 2 1 0 1 2 3 0 5 10 15 20 25 30 35 40 3 2 1 0 1 2 3 0 10 20 30 40 50 60 70 Nodes on the inner surface R=a Nodes on the inner surface R=a (a) (b) Figure 5. LLC . ∂σθθ ∂R R+ ∗ + + + .45 26.99 1. NUMERICAL RESULTS: HOLLOW SPHERE 109 R=b R=a σ θθ ∗ σ φφ ∗ σ θθ ∗ σ φφ ∗ medium mesh 1. Exact solutions: —.15 14.77 25. σ θθ ∗ ∗ ∗∗ (from [100]) ∗ © 2005 by Taylor & Francis Group.
5. Oneeighth of the block is modeled in order to take advantage of symmetry.4: Stress sensitivities on the inner surface R = a.2 Sensitivities of stresses Sensitivities of the stress components σθθ and σRR .4. Of course. Again. 5. along the line x1 = x2 = x3 .5(a) that the agreement between the numerical and analytical solutions is very good. 5.65) is used here.5 5.5(b). are presented in Figure 5. Again. ∂σφφ ∗ ∂R R + + + + . The BCM model is fully threedimensional but the imposed boundary conditions are chosen such that a state of plane strain prevails in the block and the numerical results obtained from the BCM can be compared with Kirsch’s analytical solution for the corresponding 2D plane strain problem. σ φφ ∗ ∗ ∗∗ (from [100]) ∗ solving the sensitivity boundary value problem (5. loaded in uniform remote tension. SHAPE SENSITIVITY ANALYSIS 4 4 Stress sensitivities 2 1 0 1 2 3 0 5 10 15 20 25 30 35 40 Stress sensitivities 3 3 2 1 0 1 2 3 0 10 20 30 40 50 60 70 Nodes on the inner surface R=a Nodes on the inner surface R=a (a) (b) Figure 5.110 CHAPTER 5.34).2. It is seen from Figure 5. The side of a square face of the full block is 20 units. the hole diameter is 2 units and the thickness (in the x3 direction) is 6 units.1 Numerical Results: Block with a Hole Geometry and Mesh This example is concerned with a rectangular block with a cylindrical hole of circular crosssection. The mesh on the front and © 2005 by Taylor & Francis Group. the linear velocity proﬁle (5. the analytical solution is only available for an inﬁnitesimal hole in a slab and this is referred to as the “exact” solution in this section of this chapter. very good agreement is observed between the numerical and analytical solutions. Numerical solutions from (a) the medium mesh and (b) the ﬁne mesh: σ φφ ◦ ◦ ◦◦. Exact solutions: —. LLC .
8 1. NUMERICAL RESULTS: BLOCK WITH A HOLE 111 0.5 0.6 1 Sensitivity of radial displacement 0.4 1.5: (a) Sensitivity of radial displacement along the line x1 = x2 = x3 .2 1.3 1.3 1.5 1 1.9 2 1 1. (b) Design velocities for this example (from [103]) © 2005 by Taylor & Francis Group. Numerical solutions ∗ ∗ from the ﬁne mesh: σ θθ ∗ ∗ ∗∗ σ RR ◦ ◦ ◦◦ (from [103]) X2 10 9 8 7 6 5 4 3 2 1 0 0 1 2 3 4 5 6 7 8 9 10 X2 L L R a φ X1 X1 (a) (b) Figure 5. (b) Sensitivity of stresses along the line x1 = x2 = x3 .5 Stress sensitivities 0. Exact solution —.3 0.1 1.4 1. Numerical solution from the ﬁne mesh: ◦◦◦◦.5 1. Exact solutions —.5 1.6 1.2 0 0.7 1.7 1.8 1.1 1.6 1.5.4 0.2 1. LLC .9 2 Radial coordinate Radial coordinate (a) (b) Figure 5.1 0 0.5.6: (a) Mesh on a quarter of the front and back faces of a rectangular block with a circular cylindrical hole.
It is concerned with stress sensitivities for the situation depicted in Figure 5. and along the line x1 = 3x2 . The numerical and analytical results for this problem are shown in Figure 5. x3 = 1. © 2005 by Taylor & Francis Group. boundary conditions on the symmetry planes are applied in the computer model of the oneeighth block in the usual way.7(a) and (b). In other words. which are obtained from independent calculations. The chosen design velocity distribution in the slab is linear along any radial direction in any square section normal to the x3 axis and is independent of the x3 coordinate.2 Internal Stresses Comparisons of numerical and analytical solutions for internal stresses in the block are presented in Figures 5. 5.3 Sensitivities of Internal Stresses This last example is a diﬃcult one.68) for φ ≥ π/4 (5.7(a). The internal point numerical results. one has: R= and R= ∗ ∗ L/ cos φ − R L/ cos φ − a L/ sin φ − R L/ sin φ − a for φ < π/4 (5. referring to Figure 5.5.5 for uniaxial remote loading in the x1 direc√ tion.6(b). show acceptable agreement with the analytical solutions for the stress sensitivities. respectively) of the oneeighth block is shown in Figure 5. The design variable in this example is the hole radius a.5. Two cases are considered: stresses along the line x1 = x2 . even though the analytical solutions vary rapidly near the hole. Of course.112 CHAPTER 5. Four layers of triangles (in the thickness direction) constitute the mesh on the remaining surfaces of the oneeighth block. The complete block is loaded by uniform remote tensions in the x1 and x2 directions while u3 = 0 on the faces normal to the x3 axis in order to simulate plane strain conditions.69) 5.8.5 for equal biaxial loading. x3 = 1. The agreement between the analytical and numerical results is again seen to be very good. SHAPE SENSITIVITY ANALYSIS back faces (x3 = 3 and x3 = 0.6(a). LLC .
Exact solutions —.2 0 Stress components 1 2 3 4 5 6 7 8 9 10 1. x3 = 1. Numerical ∗ ∗ solutions: σ 11 ◦ ◦ ◦◦.2 0.4 0.8: Stress sensitivities along the line x1 = x2 . σ 22 − − −.7 0.3 0.6 0.2 1.8 0. NUMERICAL RESULTS: BLOCK WITH A HOLE 113 1. x3 = 1.6 0. LLC .2 0.3 0.5 0 1 2 3 4 5 6 7 8 9 10 0.8 0.1 Stress sensitivities 0 0.4 1. σ 22 ∗ ∗ ∗∗ (from [103]) © 2005 by Taylor & Francis Group.5 for equal biaxial loading.5 for uniaxial loading in .7: Stresses (a) along the line x1 = x2√ x3 = 1. Numerical solutions: σ11 ◦ ◦ ◦ ◦.1 0.9 0.4 0 1 2 3 4 5 6 7 8 9 10 X1 = X2 Figure 5.2 0.1 1 0. σ22 ∗ ∗ ∗ ∗ (from [103]) 0.5.2 1 1.5 for uniaxial ∗ ∗ loading in the x1 direction. Exact solutions: σ 11 — .5 1. the x1 direction and (b) along the line x1 = 3x2 .4 0 X1 = X 2 X1= 3 X2 (a) (b) Figure 5.4 1.3 Stress components 0.5.
is the subject of this chapter... while the constraints are expanded in a ﬁrst order Taylor series. the optimization problem is approximated by expanding the objective function in a second order Taylor series about the current values of the design variables. of the design variable bk ...1) (6.4) in which b = b1 .. Nh k = 1. N (6.. with the BCM.4) gives side constraints that are used to limit the search ( ) (u) region of an optimization problem. Ng j = 1.2) (6. Its general form can be stated as follows: Minimize Subject to gi (b) ≥ 0. . . ( ) (u) f (b) i = 1. 115 © 2005 by Taylor & Francis Group. respectively... used in gradientbased optimization algorithms. . Equation (6. The most common mathematical programming approaches.3) (6. LLC .. f (b) is the objective function. hj (b) = 0. b2 . are the successive linear programming (SLP) and successive quadratic programming (SQP) methods.Chapter 6 SHAPE OPTIMIZATION Shape optimization of 3D elasticity problems.1 Shape Optimization Problems An optimal shape design problem can be stated as a minimization problem under certain constraints.. 6. respectively. the parameters bk and bk denote the lower and upper bounds. and gi (b) and hj (b) are the inequality and equality constraints. .. bk ≤ b k ≤ b k . bN T are the design variables. In the SQP method.. Further details are available in Shi and Mukherjee [150]. Here.
This is the axisymmetric version of the planar problem described in Phan et al. The ﬁrst example is that of optimizing the shape of a ﬁllet in a tension bar whose volume is selected as the optimization function. Of course. if a2 /a1 = S2 /S1 . in 3D linear elasticity. The variable boundary ED is modeled as a cubic spline (using the IMSL subroutine DCSDEC) which is deﬁned by the ﬁxed end points E and D and by the variable points Ck .116 CHAPTER 6. The required gradients of the objective functions. [22]) with the BEM. The Young’s modulus. k = 1. An optimal shape is sought that minimizes the volume (and therefore the weight when the bar material has constant density) without causing yielding anywhere in the bar. also. are calculated internally by the optimization subroutine by the ﬁnite difference method. The end circles of this surface of revolution are kept ﬁxed. k = 1.3 and σ (V M ) = 120 psi. these sensitivities could also have been obtained by the direct diﬀerentiation approach described in Chapter 5. 2. is as follows. then the tangential stress around the elliptical cutout is uniform ! This problem has been solved in 2D by the BCM in [132] and an elasticplastic version of this 2D problem has been solved in [169] (see.2 Numerical Results Two illustrative shape optimization examples. 3 are the design variables. Now. Also. LLC . The bar is loaded by a uniform axial tensile traction of 100 psi . k = 1. 2. This subroutine.1. Let the semimajor and semiminor axes of the elliptical hole be a1 and a2 . 132]. is based on the FORTRAN subroutine NLPQL developed by Schittkowski [146]. are solved using the BCM coupled with the IMSL optimization subroutine mentioned above. Poisson’s ratio and allowable von Mises stress are taken as ˆ E = 107 psi. The design surface is the surface of revolution (initially a cone) obtained by revolving the curve ED about the symmetry axis AB. ν = 0. respectively. [132] . SHAPE OPTIMIZATION The subroutine DNCONF from the IMSL library is coupled with a 3D BCM code for elastic stress analysis in order to carry out the shape optimization examples that are described in this chapter. the full 3D BCM code is used here. let S1 and S2 be the remote tensile loadings in the coordinate directions. 6. loaded in remote biaxial tension. 3 and D have equally spaced projections on the axis of symmetry while the radii rk . 6. The second problem is concerned with a cube with an ellipsoidal cavity loaded in remote triaxial tension. 2. Ck . An interesting result in 2D [5. that uses the SQP method.3.2. for an inﬁnite elastic plate with an elliptical cutout. Of course. 3.Section 5. The 3D elasticity problem is described in this chapter and some interesting results are obtained. The points E. © 2005 by Taylor & Francis Group.1 Shape Optimization of a Fillet The initial crosssection of the axisymmetric bar is shown in Figure 6. constraints etc. respectively. and the coordinate axes be centered at the center of the ellipse with the x1 and x2 directions along the major and minor axes. respectively.
1: Modeling of a bar with a ﬁllet (from [150]) Quadratic CIM9 elements (see Figure 4.5 B z 117 C3 4.2(a) while Figure 6. r3 . LLC .6) (V where σi M ) are the values of the von Mises stress at the centroids of certain elements on the curved surface of the bar. respectively. σ 1 ≤ i ≤ ns (6. This is: [σ (V M ) ] = (3/2)sij sij 2 (6.5 and the design surface. There are 16 elements on each of the end circles (on the planes z = 0 and z = 20. which is obtained after just 1 iteration. so that ns = 16. This is: zD φ(r1 . since the physical problem is axisymmetric. respectively) of the bar. 3 are constrained to lie within the triangle EGD. The usual deﬁnition of the von Mises stress is used.2(b) shows the objective function as a © 2005 by Taylor & Francis Group.1) are used in both the numerical examples described in this chapter. A total of 160 elements are used to discretize the surface of the bar.5 D r3 r1 r2 Figure 6. the elements chosen are the ones whose centroids lie along FEDC.6. 2. the cylindrical surface r = 4. z) − (4. The objective function is the volume of the axisymmetric object bounded by the plane z = 9. In addition.7) in terms of sij .5)2 dz (6.0. r2 . NUMERICAL RESULTS r 9 F u3=0 9 τ1=0 τ2=0 A 20 E C 1 G C2 τ3=100psi C 4. and the design surface has 64 elements. is shown in Figure 6. Here. the one described by DC has 32 elements. one has: (V σi M ) /ˆ (V M ) ≤ 1. the deviatoric components of the stress σij .5) The axisymmetric shape of the body is maintained during the optimization process and the design nodes Ck . k = 1. The ﬁnal converged solution. These are distributed as follows. r2 . r3 ) = zG π r2 (r1 .2. The surface of revolution described by the edge FE has 32 elements.
5 6 5. LLC . some suitable measure of stress) on its surface for a given remote loading. and.5 4 9 10 11 12 13 C1 C2 C3 14 15 D 16 140 120 100 80 60 40 r φ/π 0 1 2 3 4 5 z Iteration number (a) (b) Figure 6. is shown in Figure 6. Three cases of remote loadings (cases (1). loaded in remote triaxial tension.3. The cube is of size 30×30×30. SHAPE OPTIMIZATION 9 8. the cavity shape is sought that would make all the stress components on it uniform. which are not shown in this ﬁgure. The cavity surface is the design surface with the ellipsoid semi axes a1 and a2 as the design variables (a3 = 1). The mesh consists of 8 identical CIM9 elements on each surface of the cube and 72 CIM9 elements on the surface of the cavity. The ﬁnal (converged) value is 150. are suitably restrained. is a spherically symmetric problem. The initial value of the objective function is 551. The cube surface is ﬁxed.85% of its initial value. The cavity mesh evolves with the changing shape of the cavity during the optimization process. As stated at the start of this section.5 7 6. respectively.2: (a) Optimal shape of the ﬁllet (b) Objective function φ as a function of iteration number (from [150]) function of iteration number. This. the ﬁnal volume of the design portion of the bar is about 27. the shape of an ellipsoidal cavity that would have uniform stress (i.2. as expected. © 2005 by Taylor & Francis Group. Consistent units are used.118 E CHAPTER 6.e.2 Optimal Shapes of Ellipsoidal Cavities Inside Cubes A cube with a centrally located ellipsoidal cavity. Of course. the design area reduces to 49. is applied.3. Starting with an ellipsoidal cavity with axis ratios a1 : a2 : a3 = 3 : 2 : 1. In the corresponding 2D problem [132].5 5 4. of course.5 180 160 8 7. (2) and (3)) are considered : S1 : S2 : S3 = 1 : 1 : 1. As a test case.e. the objective here is to ﬁnd. S1 : S2 : S3 = 1 : 1 : 1. if possible.5 : 1. 6. the faces of the cube. the ﬁrst case of remote loading.35 with one stress constraint being violated.28 % of its initial value. i.39 when two stress constraints are active. 2 : 2 : 1 and 2 : 1. with S3 = 105 in all cases. Thus. This time the shear modulus is taken as G = 105 and the Poisson’s ratio ν = 0.
of course. © 2005 by Taylor & Francis Group. and an overbar denotes the mean value of the appropriate quantity over the design surface. NUMERICAL RESULTS 119 x3 S3 a3 a1 a2 x2 S2 x1 S1 Figure 6. Uniform σkk on a surface means that σnn + σss + σtt (where n is normal and s and t are any two orthogonal directions.2. Since the cavity surface is unloaded. The criterion σkk is uniform on a cavity surface is treated as a generalization of the 2D case in which the tangential stress on a cutout surface was made uniform [132]. the trace of the stress tensor. this implies that σss + σtt is uniform for all points on the cavity surface. The more interesting cases. ns is the number of elements on the design surface. and the second is the von Mises stress. a2 ) = ns ns (V σi M ) 2 − (σ (V M ) ) 2 2 (6. a2 ) = σkk − σkk i=1 (i) (6. LLC . tangential to the surface at a point on it) is also uniform on the surface.3: A cube with an ellipsoidal cavity under remote loading (from [150]) the optimal shape of the cavity is found to be a sphere. The ﬁrst is σkk .6. are loading cases (2) and (3). The corresponding objective functions to be minimized are deﬁned as: 1 ns ns 2 φ1 (a1 . two diﬀerent scalar measures of stress on the cavity surface are considered.8) 1 φ2 (a1 . a superscript (i) denotes evaluation of the appropriate quantity at the centroid of the ith element on the design surface. This time.9) i=1 In the above equations.
the optimal a1 : a2 : a3 does not match the load ratios). The corresponding objective function.9. it can be seen from Figure 6.6.5. is shown in Figure 6.5(b). 1 Start 2. respectively. are shown. 1 0. LLC .4(a) shows the stress σkk .4086 Optimal 3. while the optimal values of a1 and a2 are equal (i.b.e.1. 1.0489 (2) Table 6.07% of its initial value.1. 1 0.1 that. 1 u. 2.1 3.1 1.1 again shows that the optimal cavity geometry is not compatible with the loading in this case. It does. 2. and its mean value. The stress σkk and the objective function for this case appear in Figures 6.4105.654. 1 Start 2. the crosssection of the optimal shaped cavity in the x3 = 0 plane is a circle). eventually settle down to about 2.1 Uniform trace of the stress tensor over the cavity surface The load cases (2) and (3) are tried for the ﬁrst objective function φ1 .b. In each case. It is seen from Figure 6. 2.3. initially and at the end of the optimization process. 1.7.1 0.1 in which the side constraints on the design variables. however. 3.11057 Optimal 3. 7.654.4(b). as a function of iteration number. It can be seen from Table 6. 2.044 Optimal 2. The loads are normalized with S3 = 1 (from [150]) 6. Upper bound is denoted as u.b. 1.002286 Uniform σ (V M ) on cavity surface 2. the starting geometry of the ellipsoid is taken to be compatible with the loads.2.4105. 1 0. SHAPE OPTIMIZATION Ellipse semiaxes Objective function Uniform σkk on cavity surface 2. Figure 6. 2.5(a) that the initial distribution of σkk (dots) is quite widespread while the optimal distribution (asterisks) is concentrated within a band around its mean value σkk . 1 u.1: Results for cube with ellipsoidal cavity under remote loading.e. © 2005 by Taylor & Francis Group.b. Examination of Table 6.2.5. It is seen that the initial distribution of σkk (dots) is quite widespread while the optimal distribution (asterisks) is concentrated within a band around its mean value σkk .4 . 1 u.5.5(a) and 6.120 Load case (2) (3) Loads CHAPTER 6.5608. the loading and geometry are not compatible as in the 2D case (i. inspired by the 2D case.5(b) that in this case the behavior of the objective function is initially oscillatory. The more diﬃcult case is load case (3). 9. Again. 3. Details are shown in Figures 6.0027 2. 1 Start 2. as well as the initial and optimal values of the design variables and the objective function.8022. Lower bounds for ellipse semiaxes are 1. The results appear in Table 6.
2 5 4..8 σkk 4. LLC . NUMERICAL RESULTS 121 5.5: (a) The stresses σkk and σkk for load case (3). Initial : σkk .8 0 0.4: (a) The stresses σkk and σkk for load case (2).2 0.4 4 3.2.8 4.4 0 1.8 σkk φ1 0.6 0 5 10 15 20 25 Nodes at element centroids Iteration number (a) (b) Figure 6.6. σkk − −− Optimal : σkk ∗ ∗∗.2 0 10 20 30 40 50 60 70 80 0 1 2 3 4 5 Nodes at element centroids Iteration number (a) (b) Figure 6...4 0..6 4.8 5.4 1.2 0 10 20 30 40 50 60 70 80 φ1 0..8 4.4 1.4 1.4 5. σkk — (b) Objective function φ1 as a function of iteration number for load case (3) (from [150]) © 2005 by Taylor & Francis Group.6 5.. σkk − −− Optimal : σkk ∗ ∗∗.2 1 0.6 4. σkk — (b) Objective function φ1 as a function of iteration number for load case (2) (from [150]) 5 4. Initial : σkk .2 1 0.6 0.
5 6 4 3. It is conjectured that a more general cavity shape. 6.2.6: (a) The stresses (σ (V M ) ) and (σ (V M ) ) for load case (2).5 5 4.6(a) and 6.5:1) did not converge.. needs to be considered for this diﬃcult problem. Initial : (σ (V M ) ) . Figure 6.43% of its initial value. It is known that © 2005 by Taylor & Francis Group.2. The results are summarized in Table 6.5 1 0. LLC . The next example considered is the optimal shape of an ellipsoidal cavity in a cube loaded by remote tensions. This time the converged value of the objective function is 1.122 CHAPTER 6..5 4 3.5 (σ (VM))2 φ2 5. As expected.5 7 6. The sought after cavity shape is the one that makes some measure of the stress uniform on it.5 3 2. rather than an ellipsoidal one. the optimal cavity shape has a circle in the x3 = 0 plane..6(a) shows a signiﬁcant diﬀerence between 2 the initial and ﬁnal values of the mean value of (σ (V M ) ) .6(b).2. SHAPE OPTIMIZATION 7.5 0 10 20 30 40 50 60 70 80 2 1.5 0 0 1 2 3 4 5 6 Nodes at element centroids Iteration number (a) 2 2 (b) Figure 6. The case of φ2 with the load case (3) (2:1. compared to about half in the corresponding 2D problem. (σ (V M ) ) − −− Optimal : (σ (V M ) ) ∗ ∗∗.5(a) . (σ (V M ) ) — (b) Objective function φ2 as a function of iteration number for load case (2) (from [150]) 6. Also.4(a) and 6.2 Uniform von Mises stress over the cavity surface 2 2 2 2 The ﬁnal example considers the load case (2) with the second objective function φ2 .1 and details are shown in Figures 6. the optimal (V M ) 2 distribution of (σ ) (asterisks) lies in a wider band around its mean value compared to the situations in Figures 6.3 Remarks The optimal shape of a ﬁllet in an axially loaded bar is obtained quickly within just one iteration ! The size of the design region reduces to nearly a quarter of its initial value in this 3D problem.
LLC .2. This is true if either no two of the remote tractions are equal. and/or if one tries to make the von Mises stress uniform on the cutout surface.6. © 2005 by Taylor & Francis Group.5(b)) one sees initial oscillations of the objective function as a function of iteration number. for an elliptical cutout in an inﬁnite plate. Also. Sometimes (Figure 6. optimization of this class of problems is often more diﬃcult than for the previous bar with a ﬁllet example. NUMERICAL RESULTS 123 in the corresponding 2D problem. the tangential stress on the cutout surface is uniform when the loading and geometry are compatible. (The precise meaning of compatibility is explained in the body of this chapter). It is found that such is not the case in the 3D problem.
applied to the BCM. The idea of using hypersingular residuals. LLC (7. to obtain local error estimates for the BIE. [123]. real positive constants c1 and c2 exist such that: c1 r ≤ ≤ c2 r 125 © 2005 by Taylor & Francis Group. under certain favorable conditions.44) in order to obtain the hypersingular residuals vkn in the displacement gradients uk. [127]). [96] and Paulino et al. the stress residuals are obtained from Hooke’s law: skn = λvmm δkn + µ(vkn + vnk ) (7.1 Hypersingular Residuals as Local Error Estimators The usual BCM equation (4. a scalar measure r of the residual. was ﬁrst proposed by Paulino [122] and Paulino et al. evaluated at the centroid of a triangular surface element. this value of a is input into the righthand side of equation (4. Next.Chapter 7 ERROR ESTIMATION AND ADAPTIVITY The subject of this chapter is error analysis and adaptivity with the BCM.1) Finally. 7. has appeared in Mukherjee and Mukherjee [111]. is postulated based on the idea of energy. The main idea. and has been discussed in detail in Chapter 2 of this book. Menon et al. Next.2) It has been proved in [96] and [127] for the BIE that.n . [123]. This is: r = skn vkn (7.3) . This idea has been applied to the collocation BEM (Paulino et al. and is presented in this chapter.25) is solved ﬁrst for the boundary variables (tractions and displacements) a.
The remeshing strategy is based on the values of the error estimator e at each element centroid. A possible criterion for stopping cell reﬁnement can be: e ≤ eglobal ¯ (7.2 Adaptive Meshing Strategy The ﬂow chart for adaptive meshing is shown in Figure 7. 7. but diﬀerent in detail from. ERROR ESTIMATION AND ADAPTIVITY Generate mesh Solve BCM equations yes Calculate HBCM residuals Need to subdivide elements? no STOP Obtain error estimators Figure 7. It should be mentioned here that the deﬁnitions of the residuals used in [96] and [127] are analogous to. In the rest of this chapter. This strategy is shown in Figure 7. LLC . Thus.126 CHAPTER 7.1. deﬁned in equation (7. a hypersingular residual is expected to provide a good estimate of the local error on a boundary element.2 in which e is the ¯ average value of the error estimator e over all the boundary elements. is the hypersingular residual. the ones proposed in this chapter. e = r.4) where eglobal has a preset value that depends on the level of overall desired accuracy. © 2005 by Taylor & Francis Group.2) (and evaluated at an element centroid). and e is the local element error estimator that is used to drive an hadaptive procedure with the BCM. where r.1: Flow chart for adaptive meshing (from [111]) where r is some scalar measure of a hypersingular residual and is a scalar measure of the exact local error.
LLC _ _ _ _ γ=e/e 1<γ<2 2<γ<3 3<γ . It is seen from Figures 7. As a consequence (Figures 7. It is known ([37]. the shear modulus of the cylinder material is 1.3 and 7.3 (b) shows that these element error estimators (denoted by vertical bars at element centroids) are largest on the elements near the boundary of the clamped face. Element error estimators are obtained from equations (4. The initial mesh on the top (loaded) and bottom (clamped) faces of the cylinder are identical and are shown in Figure 7.3 (b).44).4 (b). but exhibits sharp gradients near its boundary. The behavior of the shearing stress component σzr (here r.3.4 that this behavior is captured well by the adaptive scheme.3 7. the normal stress component σ33 varies slowly over much of the clamped face. This stress component becomes singular on the boundary of the clamped face.25) within each element and then using these averaged traction values.Short Clamped Cylinder under Tension This ﬁrst example is concerned with a short cylinder which is clamped at the bottom and subjected to unit tensile traction on the top surface (Figure 7. (c).3 (in consistent units).2) after ﬁrst averaging the traction results from (4.1 Numerical Results Example One .3 (b) while the initial uniform mesh on its curved surface is shown in Figure 7. Figure 7. The radius and length of the cylinder are each 2 units. for this problem.1) and (7.0 and the Poisson’s ratio is 0.3(a)). Also. θ. NUMERICAL RESULTS 127 One element is split into 2 elements 3 elements 4 elements Figure 7. (d)) the region near the boundary of the clamped face is reﬁned most while the mesh on the loaded face of the cylinder is left unaltered.7. z ≡ 3 are the usual polar coordinates) on the clamped face is qualitatively similar to that of σ33 . Figures 7. The stresses are uniform on the loaded face.3. (7. (d) show that some mesh reﬁnement © 2005 by Taylor & Francis Group.4 (c). [137]) that.2: Remeshing strategy (from [111]) 7.
LLC .5 1 1.5 0 0.5 1 0.5 0 1 2 2 1.128 CHAPTER 7.5 0 2 1 2 X2 0 1 1 2 4 2 2 1 0 X1 (a) 2 1.5 0 2 1 (b) 2 1.5 2 1.5 1 0.5 1 0.5 2 X1 X1 (c) (d) Figure 7.5 X3 p=1 1 0. ERROR ESTIMATION AND ADAPTIVITY 2.5 0 0.5 1 2 X2 1.5 0 2 1 X3 X3 X2 0 1 2 2 1.3: Adaptive meshing of the top and bottom faces of a clamped cylinder under tension: (a) geometry and loading (b) initial mesh with element error estimators (c) mesh at the end of the ﬁrst adaptive step (d) mesh at the end of the second adaptive step (from [111]) © 2005 by Taylor & Francis Group.5 1 0.
5 0 0.5 0 2 1 0 (b) 2 1.5 0 2 1 0 X3 X3 X2 1 2 2 1 0.5 0 0.3.5 1 1.5 X1 (a) 2 1.7. LLC .5 X3 p=1 1 0.5 0 2 2 1 0 X2 1 2 2 1 0.5 X1 (c) (d) Figure 7.5 1 0.5 2 X1 1.5 0 0.5 2 4 1.5 2 1.5 1 1. NUMERICAL RESULTS 129 2 1.4: Adaptive meshing of the curved surface of a clamped cylinder under tension: (a) geometry and loading (b) initial uniform mesh (c) mesh at the end of the ﬁrst adaptive step (d) mesh at the end of the second adaptive step (from [111]) © 2005 by Taylor & Francis Group.5 X2 1 2 2 1 0.5 1 1.5 1 0.
in Paulino et al. that the radial stress gradients are large at points on the hole surface. subjected to external radial tensile loading. (The visible elements are shown in Figure 7. In this axisymmetric problem. are only slightly reﬁned in order to maintain mesh compatibility.1. The situation on the curved surface FADG is particularly interesting. The bars in Figure 7. while the rest of the mesh on it remains unaltered. of course.5 and the hidden ones are not). ERROR ESTIMATION AND ADAPTIVITY mesh initial after 1st adaptive step after 2nd adaptive step # of elements 144 192 246 # of nodes 290 386 494 e ¯ 0. respectively. The corresponding 2D case is discussed. on which the stresses are independent of the x3 coordinate. as well as on the surface FADG of the hole. please see the discussion in the following paragraph). when the exact © 2005 by Taylor & Francis Group.1: Mesh statistics and e for the clamped cylinder under tension (from ¯ [111]) takes place on the bottom layer of the curved surface of the cylinder. in plane strain.2 Example Two .3.2 above. As expected ([124]. The next (and ﬁnal) mesh. LLC . while the symmetry planes ABCD and EFGH.3 and the external radial traction is 3 (in consistent units). A quarter of the cylinder is modeled and the initial mesh on the quarter cylinder is shown in Figure 7.0042723 Table 7. It is important to note.the Lam´ Problem for a Hollow e Cylinder This example is concerned with a thick hollow cylinder. e is seen to decrease with mesh reﬁnement. also.0048994 0. the average value of the error ¯ estimator e over the entire surface of the cylinder. ¯ 7.6). together with e. always zero.6. [124]. It is important to check the behavior of the actual errors.0086799 0. obtained from the adaptive strategy outlined in Section 7. the mesh statistics. the error estimators are largest on the surface of the hole and on the elements on the upper and lower surfaces (EBAF and HCDG) of the cylinder that lie near the hole.130 CHAPTER 7.5. the Poisson’s ratio is 0. As expected. is shown in Figure 7. which is nearest to the clamped face.0. Of course. The inner and outer radii of the hollow cylinder are 1 and 3 units. Finally. The shear modulus of the material is 1. and this fact leads to large error estimators and signiﬁcant reﬁnement of the boundary elements on the surface FADG. appear in Table 7. the tangential gradients of the stress ﬁelds in the θ direction are. reﬁnement of the surfaces EBAF and HCDG is expected in view of the presence of radial stress gradients on these surfaces. It is seen that mesh reﬁnement is carried out vigorously on the upper and lower surfaces EBAF and HCDG of the cylinder (the hidden elements are not shown in Figure 7. however.5 are the error estimators evaluated at the centroids of the boundary elements. in some detail.
6 0.5 1 0.5 F B A 1 X2 G 0.ﬁnal mesh on quarter cylinder (from [111]) e © 2005 by Taylor & Francis Group.5 2 2.2 0 3 E X3 H 2 F 2.6: Lam´ problem .5: Lam´ problem .5 2 2. LLC .5 B X2 G D1 1.3.5 C D1 1.2 0 3 H 2.5 0 0 0.4 0.8 E X3 0.8 0.5 A 1.7.6 0. NUMERICAL RESULTS 131 1.5 C 3 X1 Figure 7.5 2 1.5 0 0 0.5 3 X1 Figure 7.4 0.2 1 0.initial mesh on quarter cylinder together with e element error estimators (from [111]) 1 0.
g.6 2. Exact solution: —. n is the number of nodes and σθθ is the average value of the exact solution for σθθ (here ¯ 4.1) on the boundary elements on the symmetry face ABCD in Figures 7. © 2005 by Taylor & Francis Group.83% for the initial and ﬁnal mesh.5 and 7.7 is deﬁned as: = 100 σθθ ¯ (i) n 2 i=1 ( i ) n (i) (7. from the initial and the ﬁnal mesh. BCM solution from initial mesh: ◦ ◦ ◦◦. The adaptive meshing procedure is seen to reduce the error signiﬁcantly in one step. in.7.2 1.8 2 2.5 4 3. as well as by examining other numerical results from the BCM.5 σθθ 5 4.5) where the pointwise error i = (σθθ )numerical − (σθθ )exact at node i.4 1. The tangential stress σθθ . ERROR ESTIMATION AND ADAPTIVITY solution is available.6.5 6 5.8 3 r Figure 7. for example.3 (open circles). This can be seen. Mukherjee et al. obtained from a reasonably ﬁne mesh.5 3 1 1. respectively. LLC . e. also. by observing the BCM results for the Lam´ e problem for a hollow sphere under internal pressure in Figure 4.6 1. [167]) while the numerical results. 7 6. as a function of the radial distance r from the center of the cylinder. respectively. The resulting values of the L2 errors are 9. are designated by open circles and plus signs. for example.7: Lam´ problem for a hollow cylinder. in adaptive meshing problems such as this example. The inaccurate results from the initial coarse mesh is a consequence of the chosen mesh.83% and 3. Tangential stress σθθ as a e function of radial distance r. The numerical results for the tangential stress are obtained from the calculated tractions at the traction nodes Ii (see Figure 4.2 2. not the method itself.132 CHAPTER 7. [109]).5).7 is the exact solution (from. BCM solution from ﬁnal mesh: ++++ (from [111]) The L2 error in a numerical solution in Figure 7. Chapter 4 (see.4 2. The solid line in Figure 7. is shown in Figure 7.
Part III THE BOUNDARY NODE METHOD 133 © 2005 by Taylor & Francis Group. LLC .
For 3D linear elasticity. 8. One deﬁnes : m u(s) = i=1 m pi (s − sE )ai = pT (s − sE )a pi (s − sE )bi = pT (s − sE )b i=1 τ (s) = (8. [72] (elasticity) (see. s2 ). 80. is suitable for the BNM. [25]. one deﬁnes surface curvilinear coordinates (s1 . is presented as well. The 2D problem. 52. For problems in potential theory. of ongoing work on the BNM with Cartesian coordinates [79. let u be the unknown potential function and τ ≡ ∂u/∂n (where n is a unit outward normal to ∂B at a point on it). also. [108. LLC .Chapter 8 SURFACE APPROXIMANTS A moving least squares (MLS) approximation scheme. 77]). using curvilinear coordinates on the 1D bounding surface of a 2D body.1 Moving Least Squares (MLS) Approximants It is assumed that. potential theory. s2 ). (Chati and Mukherjee [26]. the bounding surface ∂B of a solid body is the union of piecewise smooth segments called panels. This procedure is described below. which uses the curvilinear coordinate s on the boundary of a body. On each panel. Chati et al. The 3D problem requires the curvilinear surface coordinate s with components (s1 . or on the 2D bounding surface of a 3D body. 164].1) 135 © 2005 by Taylor & Francis Group. 163. for 3D problems. elasticity). let u denote a component of the displacement vector u and τ be a component of the traction vector τ on ∂B. A brief discussion. is discussed in detail in Mukherjee and Mukherjee [107] (potential theory) and in Kothnur et al.
136
CHAPTER 8. SURFACE APPROXIMANTS
The monomials pi (see below) are evaluated in local coordinates (s1 − sE , s2 − 1 sE ) where (sE , sE ) are the global coordinates of an evaluation point E. It is 2 1 2 important to state here that ai and bi are not constants. Their functional dependencies are determined later. (The name “moving least squares” arises from the fact that the quantities ai and bi are not constants). The integer m is the number of monomials in the basis used for u and τ . Quadratic interpolants, for example, are of the form: pT (˜1 , s2 ) = [1, s1 , s2 , s2 , s2 , s1 s2 ], s ˜ ˜ ˜ ˜1 ˜2 ˜ ˜ m = 6, si = si − sE ; ˜ i i = 1, 2 (8.2)
The coeﬃcients ai and bi are obtained by minimizing the weighted discrete L2 norms:
n
Ru =
I=1 n
wI (d) pT (sI − sE )a − uI ˆ wI (d) pT (sI − sE )b − τI ˆ
I=1
2
Rτ =
2
(8.3)
1 w 0.5 0 0
1 0.8
BNM
4
EBNM
5
w
0.5 d
1
0.6 0.4 0.2
1
0
2
E
3
1 0.5 0
_
1 0.5 0.5
_1 2 _ _
0 0.5 s1
s
1
(a)
(b)
Figure 8.1: Domain of dependence and range of inﬂuence. (a) The nodes 1, 2 and 3 lie within the domain of dependence of the evaluation point E. The ranges of inﬂuence of nodes 1, 2, 3, 4 and 5 are shown as gray regions. In the standard BNM, the range of inﬂuence of a node near an edge, e.g. node 4, is truncated at the edges of a panel. In the EBNM, the range of inﬂuence can reach over to neighboring panels and contain edges and/or corners  see, e.g. node 5 (b) Gaussian weight function deﬁned on the range of inﬂuence of a node (from [163]) where the summation is carried out over the n boundary nodes for which the weight function wI (d) = 0 (Weight functions are deﬁned in Section 8.3). The
© 2005 by Taylor & Francis Group, LLC
8.1. MOVING LEAST SQUARES (MLS) APPROXIMANTS
137
quantity d = g(s, sI ) is the length of the geodesic on ∂B between s and sI . These n nodes are said to be within the domain of dependence of a point s (evaluation point E in Figure 8.1(a)). Also, (sI − sE , sI − sE ) are the local 1 1 2 2 surface coordinates of the boundary nodes with respect to the evaluation point sE = (sE , sE ) and uI and τI are the approximations to the nodal values uI and ˆ ˆ 1 2 τI . These equations above can be rewritten in compact form as: ˆ ˆ Ru = [P(sI − sE )a − u]T W(s, sI )[P(sI − sE )a − u] ˆ ˆ Rτ = [P(sI − sE )b − τ ]T W(s, sI )[P(sI − sE )b − τ ] ˆ u ˆ ˆ where uT = (ˆ1 , u2 , · · · , un ), whose kth row is: (8.4) (8.5)
ˆ τ T = (ˆ1 , τ2 , · · · , τn ), P(sI ) is an n × m matrix τ ˆ ˆ
(k) (k) (k) (k)
[1, p2 (s1 , s2 ), ...., pm (s1 , s2 )] and W(s, sI ) is an n × n diagonal matrix with wkk = wk (d) (no sum over k). The stationarity of Ru and Rτ , with respect to a and b, respectively, leads to the equations: a(s) = A−1 (s)B(s)ˆ , u where A(s) = PT (sI − sE )W(s, sI )P(sI − sE ) B(s) = PT (sI − sE )W(s, sI ) b(s) = A−1 (s)B(s)ˆ τ (8.6)
(8.7)
It is noted from above that the coeﬃcients ai and bi turn out to be functions of s. Substituting equations (8.6) into equations (8.1), leads to:
n n
u(s) =
I=1
ΦI (s)ˆI , τ (s) = u
I=1
ΦI (s)ˆI τ
(8.8)
where the approximating functions ΦI are:
m
ΦI (s) =
j=1
pj (s − sE )(A−1 B)jI (s)
(8.9)
An alternative form of (8.9) is: Φ(s) = pT (s − sE )(A−1 B)(s) where Φ(s) is 1 × n.
© 2005 by Taylor & Francis Group, LLC
(8.10)
138
CHAPTER 8. SURFACE APPROXIMANTS
ˆ ˆ As mentioned previously, u and τ are approximations to their real values u and τ . Matrix versions of (8.8) can be written as: [H]{ˆ } = {u} , u [H]{ˆ } = {τ } τ (8.11)
Equations (8.11) relate the nodal approximations of u and τ to their nodal values.
Remarks
• Remark 1: Invertibility of A The matrix A is an m × m matrix, composed of the matrices P and W (equation 8.7). It needs to be invertible for the construction of the shape functions. It is also desirable that A be well conditioned. From a wellknown fact in linear algebra about ranks of products of matrices, it is necessary that the rank of matrix P be m. However, if n < m i.e. the number of nodes n in the domain of dependence of an evaluation point is less than the order of the polynomial basis m, then matrix A would be rank deﬁcient and would become noninvertible. So, it is essential to choose the parameter which controls the range of inﬂuence of a node, ˆ namely d, such that n ≥ m. However, even if the condition n ≥ m is satisﬁed, but the n nodes in the domain of dependence of the evaluation point E lie on a straight line on the surface, then the matrix A becomes singular. Also, it has been observed that choosing n ∼ m may lead to an unacceptably large condition number of the matrix A. • Remark 2: Matrix H As noted above, the matrix H relates the actual nodal values to their nodal approximations. It is observed through numerical experiments that the matrix H has m eigenvalues equal to unity. The associated m eigenvectors are described by the monomials used in the bases for constructing the approximation. Thus, when looking for solutions that cannot be spanned by the monomials used in the bases, the matrix H plays a signiﬁcant role in the success of the method. • Remark 3: Boundary conditions The H matrix plays a crucial role in the satisfaction of essential boundary conditions in the EFG method [108] and in the satisfaction of all boundary conditions in the BNM [107]. • Remark 4: Deﬁnition of a panel Curvilinear coordinates (s1 , s2 ) are used to measure distances over curved surfaces. However, real life objects consist of piecewise smooth surfaces, referred to as panels in this work, and deﬁning curvilinear coordinates across edges and corners is a formidable task. In this work, collocation nodes are placed inside panels, and, in order to circumvent the problem
© 2005 by Taylor & Francis Group, LLC
8.2. SURFACE DERIVATIVES
139
of “reaching over edges,” it was decided that the range of inﬂuence (ROI) of each node would be truncated at an edge or corner (Figure 8.1(a)). It will be seen through numerical experiments that restricting the range of inﬂuence of a node to the panel to which it belongs still yields acceptable results. An alternative to truncation of ROIs at edges is to use Cartesian coordinates (see Figure 8.1(a) and Section 8.4). • Remark 5: The nature of s1 , s2 The coordinates (s1 , s2 ) are the curvilinear coordinates measured along the bounding surface ∂B. These coordinates are local and not global. In other words, these are constructed with the origin at the evaluation point E i.e. these curvilinear coordinates will always be (0,0) at the evaluation point E. This simpliﬁes the computation of the shape functions to some extent. Since, (s1 = 0, s2 = 0), one has p1 = 1 and pi = 0 for i = 2, · · · , m. This further implies that the shape function is just the ﬁrst row of the matrix C.
8.2
Surface Derivatives
Surface derivatives of the potential (or displacement) ﬁeld u are required for the HBIE. These are computed as follows. With C = A−1 B equations (8.8) and (8.9) give:
n m
u(s) =
I=1 j=1
pj (s − sE )CjI (s)ˆI u
(8.12)
and the tangential derivatives of u can be written as: ∂u(s) = ∂sk
n m
I=1 j=1
∂pj ∂CjI (s) uI ˆ (s − sE )CjI (s) + pj (s − sE ) ∂sk ∂sk k = 1, 2 (8.13)
The derivatives of the monomials pj can be easily computed. These are: ∂pT (s1 − sE , s2 − sE ) = [0, 1, 0, 2(s1 − sE ), 0, (s2 − sE )] 1 2 1 2 ∂s1 ∂pT (s1 − sE , s2 − sE ) = [0, 0, 1, 0, 2(s2 − sE ), (s1 − sE )] 1 2 2 1 ∂s2 (8.14) (8.15)
After some simple algebra (Chati [23]), the derivatives of the matrix C with respect to sk take the form:
© 2005 by Taylor & Francis Group, LLC
140
CHAPTER 8. SURFACE APPROXIMANTS
∂C(s) ∂B(s) ∂B(s) = −A−1 (s) P(sI − sE )A−1 (s)B(s) + A−1 (s) ∂sk ∂sk ∂sk k = 1, 2 (8.16) with ∂B(s) ∂W(s, sI ) = PT (sI − sE ) ∂sk ∂sk In deriving equation (8.16), the following identity has been used: ∂A−1 (s) ∂A(s) −1 = −A−1 (s) A (s) , ∂sk ∂sk k = 1, 2 (8.18) (8.17)
Tangential derivatives of the weight functions (described in Section 8.3) are easily computed (Chati [23]). The ﬁnal form of the tangential derivatives of the potential (or displacement) u, at an evaluation point E, takes the form: ∂u E (s ) = ∂sk
n m n m
I=1 j=1
∂pj (0, 0)CjI (sE ) uI ˆ ∂sk ∂B E (s ) I − P(sI − sE )A−1 (sE )B(sE ) ∂sk uI ˆ
jI
+
I=1 j=1
pj (0, 0) A−1 (sE )
(8.19) with k = 1, 2. In the above equation, I is the identity matrix. One also needs the spatial gradient of the function u in order to solve the HBIE. For problems in potential theory, this is easily obtained from its tangential and normal derivatives, i.e. ∂u/∂sk and ∂u/∂n (see (1.14)). For elasticity problems, however, one must also use Hooke’s law at a point on the surface ∂B. Details of this procedure are given in Chati et al. [27] and in Chapter 1 of this book. Equation (8.19) can be rewritten in compact form as: ∂u E (s ) = ∂sk
n
ΨI (sE )ˆI u
I=1 (k)
(k)
;
k = 1, 2
(8.20)
where the approximating functions ΨI
m
are:
ΨI (sE ) =
j=1
(k)
∂pj (0, 0)CjI (sE ) ∂sk
© 2005 by Taylor & Francis Group, LLC
8.3. WEIGHT FUNCTIONS
m
141 ∂B E (s ) I − P(sI − sE )A−1 (sE )B(sE ) ∂sk
+
j=1
pj (0, 0) A−1 (sE )
jI
(8.21)
8.3
Weight Functions
The basic idea behind the choice of a weight function is that its value should decrease with distance from a node and that it should have compact support so that the region of inﬂuence of a node is of ﬁnite extent (Figure 8.1(b)). Possible choices of weight functions [26] are: • Gaussian (referred to as  WFA) : e−(d/dI ) 0
2
wI (d) =
for d ≤ dI for d > dI
(8.22)
• Exponential (referred to as  WFB) : e−(d/c)2 −e(dI /c)2 2 1−e(dI /c) wI (d) = 0
for d ≤ dI for d > dI
(8.23)
• Cubic Spline (referred to as  WFC) : ˆ2 ˆ3 2/3 − 4(d) + 4(d) wI (d) = ˆ ˆ2 ˆ3 4/3 − 4(d) + 4(d) − (4/3)(d) 0 • Quartic Spline (referred to as  WFD) : 1 − 6(d)2 + 8(d)3 − 3(d)4 ˆ ˆ ˆ wI (d) = 0
ˆ for d ≤ 1/2 ˆ for 1/2 < d ≤ 1 for d > 1 (8.24)
ˆ for d ≤ 1 ˆ for d > 1
(8.25)
ˆ where d = d/dI and c is a constant. Here d = g(s, sI ) is the minimum distance, measured on the surface ∂B, (i.e. the geodesic) between a point s and the collocation node I. In the research performed to date, the region of inﬂuence of a node has been truncated at the edge of a panel (Figure 8.1(a)) so that geodesics, and their derivatives (for use in equation (8.17)), need only be computed on piecewise smooth surfaces. Finally, the quantities dI determine the extent of the region of inﬂuence (the
© 2005 by Taylor & Francis Group, LLC
142
CHAPTER 8. SURFACE APPROXIMANTS
compact support) of node I. They can be made globally uniform, or can be adjusted such that approximately the same number of nodes get included in the region of inﬂuence of any given node I or in the domain of dependence of a given evaluation point E. Such ideas have been successfully implemented in Chati and Mukherjee [26] and Chati et al. [25].
8.4
Use of Cartesian Coordinates
One of the drawbacks of using curvilinear surface coordinates as described above in Section 8.1 is the need to truncate the range of inﬂuence of a node at an edge or corner (see Remark 4 in Section 8.1). Another is the need to compute geodesics on general surfaces. The more straightforward approach, namely the use of Cartesian coordinates, suﬀers from the disadvantage that the matrix A deﬁned in equation (8.7) becomes singular if all the nodes in the domain of dependence of an evaluation point lie on a plane (see [117, 79]). Li and Aluru have suggested, in two recent papers [79, 80], ways to use Cartesian coordinates in a modiﬁed version of the BNM which they call the boundary cloud method. (The acronym BCLM is used for the boundary cloud method in this book). Nice results for 2D problems in potential theory are given in [79, 80]. This idea is discussed below for 3D problems in potential theory with linear approximants. Extension to 3D elasticity is relatively straightforward.
8.4.1
Hermite Type Approximation
For the 3D Laplace equation (see [79] for the 2D case), one writes: ∂pT (x)a (8.26) ∂n Collocation is not allowed at a point on an edge or a corner. For a linear approximation: u(x) = pT (x)a, τ (x) = pT (x) = [1, x1 , x2 , x3 ] ∂pT ∂x1 ∂x2 ∂x3 (x) = 0, , , = [0, n1 , n2 , n3 ] ∂n ∂n ∂n ∂n
(8.27)
where n is the unit outward normal at a boundary point. The coeﬃcients ai are obtained by minimizing the weighted discrete L2 norm:
n 2 n
J=
I=1
wI (xt , xI ) pT (xI )a − uI ˆ
+
I=1
wI (xt , xI )
∂pT I (x )a − τI ˆ ∂n
2
(8.28)
In [79], where a ﬁxed least squares approach is adopted, xt are the coordinates of a ﬁxed point inside a cloud (chosen to be the center of a cloud),
© 2005 by Taylor & Francis Group, LLC
as before. ∂n T(x) = ∂pT (x)(A−1 D) ∂n (8. they are constant within each cloud but vary from cloud to cloud.4. the Euclidean distance between xt and xI is used in the weight function. A disadvantage of this approach. N = pT (x)(A−1 D) (8.33) B(x) = PT (xI )W(xt . USE OF CARTESIAN COORDINATES 143 and xI .31) S(x) = with: ∂pT (x)(A−1 B). xI ) ∂n (8. LLC . is that discontinuities in the normal derivative of the potential function.normal derivatives are modeled with continuous approximants. The weight functions are piecewise constant. across edges and corners. 8.e. Very recently.4. Telukunta and Mukherjee [163.29) n n τ (x) = I=1 SI (x)ˆI + u I=1 TI (x)ˆI τ (8. This time. i. 164] have combined © 2005 by Taylor & Francis Group.2 Variable Basis Approximation Li and Aluru [80] present a variable basis approach for solving the 2D Laplace’s equation with the boundary cloud method (BCLM). The stationarity of J with respect to a leads to the equations: n n u(x) = I=1 MI (x)ˆI + u I=1 NI (x)ˆI τ (8. are not addressed properly .8. xI ) ∂PT I (x )W(xt .35) It is proved in [79] that the matrix A in (8.33) is nonsingular.34) D(x) = (8. are the coordinates of node I.32) A(x) = PT (xI )W(xt . This is an elegant approach in which reduced bases are appropriately employed in order to avoid singularity of the matrix A = PT WP. xI )P(xI ) + ∂PT I ∂P I (x )W(xt .30) where: M(x) = pT (x)(A−1 B). xI ) (x ) ∂n ∂n (8. even across edges and corners. as well as that of the standard BNM.
is described next. Finally.4.36)) for a cloud must satisfy two competing requirements . These results have been obtained from a fresh implementation of the equations presented in [80]. Straight cloud: x1 = c1 Basis [1. a broken cloud contains at least one corner in 2D and at least one edge or corner in 3D problems. 80]). As mentioned before. The approximants for u are identical in the variable basis BCLM [80] and in the EBNM [163].46 .3 is. below. in general. LLC .36) A basis (i. in a new approach called the extended boundary node method (EBNM). SURFACE APPROXIMANTS the advantages of the variable basis approach [80]. The ﬁrst step is to distinguish between singular and nonsingular clouds (DODs and ROIs are sometimes called clouds in this work . τ (x) = qT (x)b (8.2. a ﬂat cloud (for 3D problems) is one in which the nodes lie on a plane. 80] while the variable basis approach adopted in [163. the functions in p or q in (8. 164] uses moving least squares for both the BCLM and the EBNM.1) in terms of Cartesian coordinates x: u(x) = pT (x)a . It is noted here that the smoothness of the ﬁnal approximating functions for u and τ (see equation (8. 80] that adopts a ﬁxed least squares approach.e. The starting point is to write the approximations (8. for 2D and 3D potential theory.it must be broad enough to include all cases. still the geodesic between x and xI in the EBNM. 8. depends on the choice of the weight function wI [10]. d is taken as the Euclidean distance between xt (a ﬁxed point inside a cloud) and xI in the work of Li and Aluru [79. It is important to mention here that the ﬁxed least squares approach has been adopted by Li and Aluru in [79.1 Twodimensional problems The following bases are used for the various cases listed Bases for u and τ .144 CHAPTER 8. diﬀerent for τ .the term is taken from the work of Li and Aluru [79. at a regular point on the bounding surface of a body. A curved cloud is a smooth curve in 2D and a smooth surface in 3D problems.47)) are nonsingular.8. yet it must be narrow enough such that the matrices A = PT WP and C = QT WQ (see (8. A straight cloud (for 2D problems) is one in which the nodes lie on a straight line. The title “BCM” used in ﬁgures depicting numerical results in [163] refer to a variable basis BCLM with moving least squares. They are.44)). The speciﬁc approximants for the EBNM are given below . The EBNM. together with allowing dis∂u continuities in τ = ∂n . x2 ] for u and τ © 2005 by Taylor & Francis Group.the approximants for τ are continuous (same as those for u) in [80] but allow for jumps in τ across corners and edges in the EBNM [163]. Similarly. however. The quantity d in the argument of a weight function wI in Section 8. adapted to the moving least squares formulation.
y) = a0 + a1 x + a2 y which justiﬁes the chosen basis. x1 . © 2005 by Taylor & Francis Group.8. x1 .o. x2 ] for u [n1 . use the basis curved cloud segment. x1 . x2 ) ∈ C ∞ in B. n2 x1 . x2 ] ∪ n2 [1. n1 x1 . LLC . x1 or x2 ].38) (8. x1 ] Basis n1 [1. n2 x2 ] for τ with at least one segment curved Explanation for choice of basis for u on a broken cloud. n1 x1 + n2 x2 = c3 It is clear that the recommended reduced basis for τ covers all these cases. x20 ) which. Assuming ∇u ∈ C ∞ in B. Straight line: n1 x1 = c1 Straight line: n2 x2 = c2 Straight line: Basis Basis n1 [1. x2 ] for u and τ 145 [1. y0 )(x − x0 ) + uy (x0 . x2 )n1 + u.1 (x1 . Invertibility of C = QT WQ for τ on a broken cloud. x2 ] n2 [1. x2 ] on a general Special cases of straight segments of a broken cloud are as follows. A general curved cloud segment has the equation f (x1 . for simplicity. n2 . therefore. A Taylor series expansion for u about (x0 . n1 x2 . It is assumed that u(x1 . (8. is written as (x0 .t. n2 . Let a corner C on ∂B have coordinates (x10 . one has: τ = u.2 (x1 . y) = u(x0 . y0 )(y − y0 ) + h.39) (8. x2 ) = 0. USE OF CARTESIAN COORDINATES All other straight clouds: Basis [1. y0 ) is of the form: u(x.37) n1 [1. The linear approximation of u about C is of the form: l(x. x2 )n2 One can. x1 . y0 ) + ux (x0 . Explanation for choice of basis for τ on a broken cloud.4. n2 x1 ] for τ with both segments Broken cloud: Basis straight [n1 . x1 or x2 ] ∪ n2 [1. x1 ] for u and τ All curved clouds: Basis [1. y0 ). n1 x2 .
one gets the matrix [C] that is displayed on the next page. qx]. however. n2 ] → [p. L1 is chosen to be part of the y axis). q.2: Boundary segments of 2D regions with corners (from [163]) Straight segments with reduced basis. L2 is straight. q3 = q4 = −1 With the reduced basis [p. x2 ) → (x. it is easy to show that: x1 = x2 = 0. y) and [n1 . y3 = y4 = 0 q1 = q2 = 0. with β a constant. This matrix is nonsingular if the segment L2 is curved. For this case. Straight and curved segments with full basis. LLC . Clouds must be chosen to fulﬁll this requirement. row 5 = β × row 2 and the matrix becomes singular! © 2005 by Taylor & Francis Group. qx. px. py. qy].41) wi xi wi xi L2 2 L2 wi xi 0 0 This matrix is nonsingular provided that each cloud segment contains at least two nodes. py. q.40) (8. Figure 8. In this case. Referring to Figure 8. one has qi = βpi for all nodes on L2 . SURFACE APPROXIMANTS L1 = y axis 1 2 L 1 = y axis 1 L 2 = x axis 2 3 o (b) 4 O 3 (a) 4 L2 Figure 8. using the full basis [p. p1 = p2 = −1 p3 = p4 = 0. q]. If. Let (x1 .2(a) (the worst case scenario). (Without loss of generality. one has: 0 L1 wi L1 wi yi 2 0 L1 wi yi L1 wi yi [C] = 0 0 L2 wi 0 0 L2 (8.2(b) shows a broken cloud with one straight and one curved segment.146 CHAPTER 8.
8. USE OF CARTESIAN COORDINATES Σ L w i + Σ L p2 wi i 1 2 Σ L p2 wi x i i 2 Σ L 2 p2 wi x 2 i i Σ L 1 wi yi + Σ L 2 p2 wi yi i Σ L pi q i wi 2 Σ L 2 pi q i wi x i Σ L 2 p i q i w i yi Σ L2 Σ L 2 pi q i wi x i Σ L 2 pi q i wi x 2 i Σ L p i q i w i x i yi 2 Σ L2 ΣL pi q i wi yi 2 Σ L 2 pi q i wi x i yi Σ L 2 pi q i wi yi2 Σ L 2 q2 wi yi i Σ L q2 wi x i yi i 2 Σ L 2 q2 wi yi2 i Σ L 2 p2 wi x i yi i 2 Σ L wi yi2 + Σ L pi wi yi2 1 2 symmetric q2 wi i q2 wi x i i Σ L 2 q2 wi x 2 i i The matrix [C] 147 © 2005 by Taylor & Francis Group. LLC .4.
x3 ] n2 [1. x1 . below. x3 ] on a general curved cloud segment. x1 ] n3 [1. x1 or x3 . x1 . x2 . Special cases of ﬂat segments in a broken cloud Plane P1 : n1 x1 = c1 Plane P2 : n2 x2 = c2 Plane P3 : n3 x3 = c3 Basis Basis Basis n1 [1. x2 . x2 .47)) to be invertible. a ﬂat cloud (or a ﬂat segment of a broken cloud) must always contain three or more points. x3 . x2 . x2 or x3 ] ∪ n3 [1. x1 . x3 )n3 (8. x3 ] ∪ n3 [1. x1 or x2 . x1 . x3 ] ∪ n2 [1. x1 . x2 . x3 ) = 0. x1 or x3 . Bases for u and τ .3 (x1 . x3 ] ∪ n2 [1. x2 . x2 ] ∪ n1 [1. x3 ] n2 [1. x2 . one has: τ = u. x3 .4. x1 ] for u and τ for u and τ for u and τ for u and τ x1 = c1 or n1 x1 + n2 x2 = c4 x2 = c2 or n2 x2 + n3 x3 = c5 All other ﬂat clouds: Basis [1. x1 . x2 ] All curved clouds: Basis [1.148 8. LLC . x2 or x3 ] n3 [1. x2 ] Plane P4 : n1 x1 +n2 x2 = c4 Plane P5 : n2 x2 +n3 x3 = c5 Plane P6 : n3 x3 +n1 x1 = c6 © 2005 by Taylor & Francis Group. Choice of basis for τ on a broken cloud. x3 ] Broken cloud: Basis [1. x2 . x2 ] Basis Basis Basis n1 [1. Flat cloud: The following bases are used for the various cases listed Basis [1.1 (x1 . x1 or x2 .2 CHAPTER 8. x3 )n1 + u. Assuming ∇u ∈ C ∞ in B.46 . x2 . x3 )n2 + u. use the basis n1 [1. The arguments given for the 2D case can be easily extended to the 3D case as well. x1 .42) One can. therefore. x2 . x1 . a word of caution. In order for the matrices A and C (see (8. x3 ] for u see below for τ Explanation for choice of basis for u on a broken cloud. x1 .2. A general curved cloud segment has the equation f (x1 . and all the points on it must not lie on a straight line [164].8. SURFACE APPROXIMANTS Threedimensional problems First.2 (x1 . x3 ] Basis [1. x2 .
n2 x1 . x2 . n2 x3 . n2 x2 . two of (x1 . n1 x2 . n1 x2 . n3 x3 ] In the above. n3 x2 ] [n1 . two of (x1 . n2 x3 . fk . one has. n2 x1 . n2 x3 . n1 x3 . n2 x2 . n1 x1 . n2 x3 ] [n2 . n3 x2 . x3 ) = 0 Basis [n1 .2(b). n1 x2 . n2 . LLC . n1 x3 . n3 . n2 x3 . x2 ) = 0 f2 (x2 . n2 . 2. one must use a basis that is a union of the bases for its segments. x3 )] On a broken cloud. For example: P1 ∪ S1 : P1 ∪S2 : Basis Basis [n1 . are nonlinear functions of their arguments. n2 x2 . n1 x2 . n2 x1 ] [n1 . n2 x3 ] [n1 . two of (x1 . n1 x3 . n3 x2 . n2 x1 . n1 x1 . x3 )] ∪ n2 [1. x2 . n3 x2 . n2 x3 . On a broken cloud. n2 x1 . n1 x1 . n2 x3 . x2 . n1 x1 . n3 . n2 . for example: P1 ∪ P2 ∪ P3 : Basis [n1 . n1 x3 . n1 x2 . n3 x2 ] For a union of three ﬂat segments. n3 x3 ] f4 (x1 . x3 )] ∪ n3 [1. n3 x2 ] Special cases of curved segments in a broken cloud S1 : S2 : S3 : S4 : f1 (x1 . n3 . one has. 4. n1 x2 . n3 x3 ] Note that the case P1 ∪ S1 is a 3D version of the 2D example in Figure 8. n2 . n2 . n2 . n1 x2 . k = 1. n2 x3 . n1 x2 . n1 x3 . n3 x1 . x2 . n3 . for example: P1 ∪ P2 : P1 ∪ P4 : P1 ∪ P5 : P1 ∪ P7 : Basis Basis Basis Basis [n1 .8. n1 x3 . n3 x1 . n1 x3 . x3 ) = 0 f3 (x3 . n3 . n2 x2 .4. n2 x1 . For a union of two ﬂat segments. n2 x3 . n2 x1 . n2 x1 ] [n1 . x1 ) = 0 Basis Basis Basis [n1 . n2 . n3 x3 ] [n1 . n1 x2 . n1 x2 . USE OF CARTESIAN COORDINATES 149 Plane P7 : n1 x1 + n2 x2 + n3 x3 = c7 Basis n1 [1. n2 . n3 x1 . n1 x3 . n3 . © 2005 by Taylor & Francis Group. n3 x2 . one must use a basis that is a union of the bases for its segments. n2 x1 . n1 x3 . n3 x1 . n1 x3 . n2 x1 . n3 x1 . 3. n3 . n3 x1 . n3 x1 . n2 x2 . n2 .
44) where: α(x) = pT (x)(A−1 B)(x). xI )Q(xI ).. with: A(x) = PT (xI )W(x.3). It is important to mention here that the rest of this part of this book presents the boundary node method with curvilinear surface coordinates (s1 . p2 (xk ). P(xI ) is an n × m matrix whose kth row is: [p1 (xk ).45) C(x) = QT (xI )W(x. 164]. analogous to (8.47) In the above.2. τ (x) = u I=1 βI (x)ˆI τ (8. 8. SURFACE APPROXIMANTS Invertibility of C = QT WQ for τ on a broken cloud with ﬂat segments. . are most encouraging. xI ) (8. Use of Cartesian coordinates in the 3D BNM is an interesting subject of continuing research.. s2 ). one ﬁnally gets: n n u(x) = I=1 αI (x)ˆI . D(x) = QT (xI )W(x..4.46) β(x) = qT (x)(C−1 D)(x) (8. LLC . let: n n Su = I=1 wI (d) pT (xI )a − uI ˆ 2 .150 CHAPTER 8. pm (xk )] similarly for Q(xI ). Numerical results from the EBNM. A proof of this fact is available in [164]. xI ) is an n×n diagonal matrix with wkk = wk (d) (no sum over k). © 2005 by Taylor & Francis Group. Sτ = I=1 wI (d) qT (xI )b − τI ˆ 2 (8. The matrix C becomes singular if and only if all the points on any ﬂat cloud segment are colinear. B(x) = PT (xI )W(x. for 2D and 3D problems in potential theory [163.3 Determination of approximating functions for u and τ Next. xI )P(xI ). and W(x.43) Following the same steps as in Section 8.. xI ) (8.1.
LLC .1.1 Potential Theory in Three Dimensions BNM: Coupling of BIE with MLS Approximants Coupled equations The bounding surface ∂B of a body B is partitioned into Nc cells ∂Bk and MLS approximations for the functions u and τ (8. ny nodes lie in the domain of dependence of the ﬁeld point y and nx nodes lie in the domain of dependence of the source point x. Numerical results for selected examples follow. y) k=1 ∂Bk I=1 ΦI (y)ˆI − F (x.1. When x and y belong to the same cell.7). y) τ I=1 ΦI (y)ˆI − u I=1 ΦI (x)ˆI u dS(y) (9. with moving least squares (MLS) interpolants (see Chapter 8). to obtain the singular and hypersingular boundary node method (BNM and HBNM) equations. the cell is treated as a singular cell and the special techniques described in the next subsection 151 © 2005 by Taylor & Francis Group.1.1 9.see Chapter 1).1 9. Potential theory is presented ﬁrst. The result is: 0= Nc ny ny nx G(x.8) are used in the usual BIE for 3D potential theory (1. then linear elasticity.Chapter 9 POTENTIAL THEORY AND ELASTICITY This chapter presents a procedure for coupling of 3D singular and hypersingular integral equations (BIE and HBIE .1) where ΦI (x) and ΦI (y) are the contributions from the Ith node to the collocation point x and ﬁeld point y respectively. 9. Also.
(b) . y z x . POTENTIAL THEORY AND ELASTICITY .152 CHAPTER 9. regular Gaussian integration can be used as long as the source and ﬁeld points are on diﬀerent cells. The details follow. as y → x. the kernels become singular as the source and collocation points approach each other. i. . .1. However. .1.t . The strongly singular integral in (9. 9. y) is strongly singular (O(1/r2 )) while the kernel G(x. y) in (9.1(a). . . .e. The kernel F (x.P . P (a) .2 Weakly singular integration scheme As mentioned above.1) over a (in general curved) surface cell as shown in Figure 9. Consider evaluating the integral with G(x. The method suggested by Nagarajan and Mukherjee [114] has been used to carry out the weakly singular integration in the BNM.s η 2 J (c) P I η 1 ξ2 (d) θ=π/2 1 ξ1 θ (e) 1 1 1 Figure 9. ρ ρ=1 .1: Mapping scheme for weakly singular integrals (from [26]) are used to carry out the integrations. y) is weakly singular (O(1/r)) as r → 0 (here r is the Euclidean distance between x and y). y) by the O(r) function u(y) − u(x). LLC . Otherwise. Various methods have been proposed in the literature to handle weakly singular integrals arising in the BEM. This integral can be represented © 2005 by Taylor & Francis Group.1) is regularized in the usual way by multiplying the kernel F (x. . regular Gaussian integration is used.
the ρ in the numerator cancels the O(1/r) singularity. This involves a Jacobian and the integral I takes the form: t=1 s=1−t I= t=0 s=0 O(1/r)J1 ds dt (9. can be easily extended for various other kinds of geometric interpolations. Quadratic (T6) triangles are used to describe the geometry of the bounding surface.7) where J3 is the Jacobian of the ﬁnal transformation. the cell is mapped into the parent space (Figure 9. In this example. Now.6) As ρ is a measure of the distance between the source point and the ﬁeld point.1(b)) using the wellknown shape functions for T6 triangles. regular Gaussian integration can be used to evaluate the above integral I.1(c)) using the mapping for linear (T3) triangles.5) which maps the ﬂat triangle from the η1 − η2 coordinate system into a rectangle in the ρ − θ space (Figure 9.6). Finally. © 2005 by Taylor & Francis Group. consider the mapping [114]: η1 = ρ cos2 θ . The method described here. POTENTIAL THEORY IN THREE DIMENSIONS as: I= ∂B 153 O(1/r) dSQ (9.4) where J2 is the Jacobian for each triangle.3) Now. however. regular Gaussian integration can be used. the integral I is now regularized. in the parent space the triangle is divided into six pieces.1. The ﬁnal mapping involves the use of quadratic (Q4) shape functions to map the rectangle from the ρ − θ space into the standard square in ξ1 − ξ2 space (Figure 9. The integral I now takes the form: 6 θ=π/2 θ=0 ρ=1 ρ=0 I= i=1 O(1/r)J1 J2 ρ sin θ dρ dθ (i) (9. The integral I can now be written as: 6 η2 =1 η2 =0 η1 =1−η2 η1 =0 I= i=1 (i) O(1/r)J1 J2 dη1 dη2 (i) (9.1(e)). to evaluate the integral I in equation (9.1(d)). η2 = ρ sin2 θ (9. Now. so that the ﬁeld point y could coincide with the collocation point x.2) The cell shown contains the source point x. In other words. Each individual triangle is mapped into the parametric (η1 − η2 ) space (Figure 9. LLC . First.9. The ﬁnal form of the integral I is: 6 ξ2 =1 ξ2 =−1 ξ1 =1 ξ1 =−1 I= i=1 O(1/r)J1 J2 J3 ρ sin θ dξ1 dξ2 (i) (9.
Also let {ˆ } y and {ˆ } be their corresponding approximations. Let the vector {¯ } contain the prescribed boundary conditions y and {x} contain the rest.1.34)). (9.11).8) where {ˆ } and {ˆ} contain the approximations to the nodal values of u and τ u τ at the NB boundary nodes. A special procedure is needed here since. for example. y) become nearly singular and nearly hypersingular. and for the BCM in Section 4.1. This gives rise to a system of equations of the form: [H1 ]{ˆ} = {¯ }.10) ˆ Equations (9. let {ˆ} = ({ˆ } ∪ {ˆ }).3.1) becomes: [K1 ]{ˆ } + [K2 ]{ˆ} = {0}. z NB equations (9. together with (8. the nodes with prescribed quantities are considered ﬁrst.9). An analogous procedure is used here for problems in potential theory. [104]).1) contains approximations to the nodal values of the primary variables rather than the variables themselves [107. respectively.8) is now written as: [M]{ˆ} = {0}. is a consequence of the fact that the kernels G(ξ. Satisfaction of boundary conditions is carried out next. sometimes referred to as the boundary layer eﬀect. y) and F (ξ. It is assumed that in a general mixed boundary value problem.2) while the potential gradient is obtained from (1. 9. as ξ → y.1. in Kane [68] (equations (17. u τ NB equations (9. (as a postprocessing step) consideration of the rest of the boundary nodes (those without prescribed boundary conditions) results in the equations: {x} = [H2 ]{ˆ}.3 CHAPTER 9. This phenomenon. due to the lack of the delta function property of MLS approximants.154 9. It is well known that direct use of these equations can yield poor results for the potential and terrible results for the potential gradient at internal points that are close to the boundary of the body. The explicit equations for potential theory are available. POTENTIAL THEORY AND ELASTICITY Discretized equations and boundary conditions The discretized assembled form of equation (9.10) are now solved together for 2NB unknowns zk . A possible remedy for the BEM for elasticity problems is discussed in Section 1. It should be noted here that the above mentioned equations in [68] were used to regularize the BIE and © 2005 by Taylor & Francis Group.1. z y NB equations (9.25) and (17. Finally. LLC .11) which yield the required boundary values xk . z NB equations (9.3 of this book (see. Each of these vectors is of length NB .8) in each case. either u or τ is prescribed at each boundary node. 108]. x z u τ Equation (9. also.9) Referring to equations (8.9) and (9.4 Potential and potential gradient at internal points The potential at a point ξ inside the body B is obtained from equation (1. Finally.
k (x) ∂Bi ∂G(x. y) ∂n(x) ΦI (y)ˆI − u I=1 I=1 ΦI (x)ˆI − u. y) nk (y) − nk (x) dS(y) ∂xm ny nx − ∂Bi ∂F (x. MLS approximations for the functions τ and ∂u/∂s (8. 9. is a new application of these equations. respectively. respectively.9.8. is used in equations (9.2 9. 8. This approach. where ΦI (x) and ΦI (y) are the contributions from the Ith node to the collocation point (x) and ﬁeld point (y).13). The resulting HBNM equations are: Nc 0 = i=1 ∂Bi ∂G(x. while use of these equations to obtain the primary variable and its derivative. © 2005 by Taylor & Francis Group. POTENTIAL THEORY IN THREE DIMENSIONS 155 HBIE.13).2.k (x) ∂Bi ∂G(x.1.1 HBNM: Coupling of HBIE with MLS Approximants Coupled equations This time. y) ∂n(x) ny nx ΦI (y)ˆI − τ I=1 I=1 ΦI (x)ˆI dS(y) τ − u.12) and (9.3 of this chapter. LLC . The discretized form of the potential gradient equation (1.13) respectively. in global coordinates.20) are used in the HBIEs for 3D potential theory (1. at internal points close to the boundary of a body.k (x)(yk − xk ) dS(y) u (9. y) ∂xm ΦI (y)ˆI − u I=1 I=1 ΦI (x)ˆI − u.1.2.14) (2) The gradient of u from equation (9.1. is discussed later in Section 9.k (x)(yk − xk ) dS(y) u (9. y) ∂xm ny nx ΦI (y)ˆI − τ I=1 I=1 ΦI (x)ˆI dS(y) τ − u.1.14) at a source point x is: nx nx nx ∇u(x) = n(x) I=1 ΦI (x)ˆI + t1 (x) τ I=1 ΨI (x)ˆI + t2 (x) u I=1 (1) ΨI (x)ˆI u (9. y) nk (y) − n(x) dS(y) ∂n(x) ny nx − ∂Bi ∂F (x.12. with nx and ny nodes in their respective domains of dependence. 1.14). for the BNM for linear elasticity.12) Nc 0 = i=1 ∂Bi ∂G(x.
8) and boundary conditions are imposed in the manner described above in Section 9.1. Exact solutions.1. 1 2 3 © 2005 by Taylor & Francis Group. The complete sphere is modeled in all cases.1.18) (9. and the normal derivatives of the potential are computed on the sphere surface.13) are evaluated by Gaussian quadrature while weakly singular integrals are evaluated by the procedure outlined in Section 9.17) (9. POTENTIAL THEORY AND ELASTICITY Discretized equations Equation (9.13) is used to solve boundary value problems in potential theory. (Results for example problems on cubes are available in Chati and Mukherjee [26]). • Linear solution u = x1 + x2 + x3 • Quadratic solutionone u = x1 x2 + x2 x3 + x3 x1 • Quadratic solutiontwo u = −2(x1 )2 + (x2 )2 + (x3 )2 • Cubic solution u = x3 + x3 + x3 − 3x2 x2 − 3x2 x3 − 3x2 x1 1 2 3 1 2 3 • Trigonometric solution u= 2r2 2r2 1 cos2 φ − − 2 R 3R2 3 (9.15) where R is the radius of the sphere.3 Numerical Results for Dirichlet Problems on a Sphere The BNM [26] and the HBNM [27] are used to solve Dirichlet problems on a sphere. LLC .1. 9.2. As described above in Section 9. Dirichlet problems are posed with these solutions imposed (in turn) on the surface of a solid sphere. The exact solutions presented below have been used to evaluate the performance of the various parameters of the BNM and the HBNM.1. φ is the angle measured from the x3 axis and r2 = x2 + x2 + x2 . nonsingular integrals in (9.16) (9.19) (9.2 CHAPTER 9.3.13) has the same form as (9. The discretized version of equation (9. Potential gradients at internal points are also computed in some cases.2.156 9.1.2.1.1.
© 2005 by Taylor & Francis Group.22) In the above.using linear (T3) and quadratic (T6) triangles. placement of the collocation node at the centroid of the triangle in the parent space yields excellent results. This nodal placement has been used in all the examples presented below. Numerical results have been obtained using linear (T3) triangles and quadratic (T6) triangles for interpolating the geometry. each of these triangles are mapped into a unit triangle in the parent space (see Figure 9. These are given below. Several error measures have been employed in order to assess the accuracy of the numerical solutions.1. respectively. As mentioned above. • Local φi (num) εlocal = − φi (exact) φi (exact) × 100 % (9. LLC . as expected. One node per cell has been used for all the numerical examples presented in this section. The crude mesh contains 72 and the ﬁne mesh 288 cells (over the surface of the sphere). POTENTIAL THEORY IN THREE DIMENSIONS 157 Error measures. φ is a generic function. φi is its nodal value at node i and φ(num) is its numerical value. The usual curvilinear coordinates θ and φ are used. The results have been obtained for four diﬀerent meshes. Dirichlet boundary conditions corresponding to the exact solution have been imposed on the surface of the sphere. 27] that. 9.used for HBNM results) (φ(num) − φ(exact) )2 dA × 100 (φ(exact) )2 dA A (φ) = A % (9. To carry out the integration.3.21) • Global Two (integrated L2 error .used for BNM results) NB εglobal = 1 φi max 1 NB (φi i=1 (num) − φi (exact) 2 ) % (9. These results are taken from [26]. Position of collocation nodes. It has been shown (from numerical experiments) in [26.1 Results from the BNM A variety of problems have been solved on a sphere.20) • Global One (nodebased global percentage L2 error .1.2). Each of the meshes has two versions .9.
3 that using a higher order basis has only a marginal eﬀect on the solution. For the optimum choice of dI . It can be seen that a crude mesh with 72 quadratic (T6) cells already yields acceptable results. It can be seen from Table 9. Various weight functions. are given in equations (8. except for the case with the ﬁne mesh with T6 triangles.0 Parent space 1/3 z Linear (T3) triangle y . © 2005 by Taylor & Francis Group. x . while. POTENTIAL THEORY AND ELASTICITY Quadratic (T6) triangle y .21)) in τ has been recorded for increasing dI . the Gaussian weight function seems to have an edge over the other weight functions used.c .2: Mapping of linear (T3) and quadratic (T6) triangles onto the parent space (from [26]) Choice of compact support of weight function. the global one error (global percentage L2 error from (9. Choice of weight function and polynomial basis. The parameter dI (the compact support of a weight function) needs to be chosen such that a ‘reasonable’ number of nodes lie in the domain of dependence of an evaluation point E. which is about 2m − 3m for a quadratic polynomial basis m = 6. for various weight functions proposed in the literature. The results for the linear and quadraticone solutions have been obtained using a polynomial basis m = 6. various researchers feel that it is best to use lower order polynomial bases in meshfree methods.22 8.25). for the cubic solution. respectively (see (9. 9.0 s z Figure 9. LLC . x t 1. In fact.9.18)).1. . proposed in the literature. Tables 9. there are about 12 − 14 nodes in the range of inﬂuence of each node. . Also. .158 CHAPTER 9.15 . 1/3 1. It is observed that choosing the smallest possible dI yields the lowest L2 error. quadraticone and cubic solutions.3 present a convergence study that has been carried out for imposed linear. .2 and 9.3. polynomial bases m = 6 and m = 10 have been used. . In Figure 9. .
9.199 1.237 5.057 Fine Mesh (T3) 2.789 1.314 5.086 2.722 2.2: L2 error in τ for the quadraticone Dirichlet problem (from [26]) © 2005 by Taylor & Francis Group.788 1.3: Eﬀect of changing dI on the global L2 error in τ for a prescribed linear solution (from [26]) Weight Functions WFA WFB WFC WFD Crude Mesh (T3) 4.391 Fine Mesh (T3) 1.253 Crude Mesh (T6) 1. POTENTIAL THEORY IN THREE DIMENSIONS 159 4.1.5 2 1.5 3 L error in 2 τ 2.886 1.5 5 dI Figure 9.539 0.022 5.5 2 2.022 4.103 Fine Mesh (T6) 0.069 1.584 0.067 1.5 1 1.785 1.237 4. LLC .942 Table 9.011 1.031 1.029 4.5 3 3.246 Fine Mesh (T6) 0.497 0.5 4 4.757 Table 9.697 1.893 Crude Mesh (T6) 1.104 2.696 1.5 4 3.1: L2 error in τ for the linear Dirichlet problem (from [26]) Weight Functions WFA WFB WFC WFD Crude Mesh (T3) 4.083 2.
007 5.019 2.053 3.5.0 3.413 1.650 2.998 3.0.414.0.1.847 1.485 1.4 presents the results for τ at a few points on the boundary that are not the collocation nodes.944 5.963 2.949 3.0.799 3.954 Table 9. The Dirichlet © 2005 by Taylor & Francis Group.1.225. Results at internal points.732.4 (a) and (b) show variation in the potential and its directional derivative at points inside the sphere.002 0. Upon solving a Dirichlet boundary value problem using the BNM.284 2.0) (0.733 0.484 0.513 0.0.0.732) (1.4: Local error in τ at boundary points for the quadraticone Dirichlet problem (from [26]) Results at boundary points other than at nodes.0.796 Table 9.5.962 2.018 2.665 2.944 6.570 Fine Mesh (T3) m = 6 m = 10 2.0. Table 9.1.0.0.0.1. to compute τ at any boundary location other than the collocation nodes.000 2.414) (1.318 2.0) τnum 0.707.414.254 3.1.949 Local Error 0.707. need to be used. The results are obtained upon imposing a quadratic1 solution on a crude mesh (72 T6 cells) with the Gaussian weight function.0.1.0) (0.345 2.912 τexact 0.485 1.318 Fine Mesh (T6) m = 6 m = 10 0.301 2.909 0.1. the value of τ is known at the NB boundary nodes.3: L2 error in τ for the cubic Dirichlet problem (from [26]) Coordinates (2.225.160 CHAPTER 9.475 2.732 2.8).866. POTENTIAL THEORY AND ELASTICITY Weight Functions WFA WFB WFC WFD Weight Functions WFA WFB WFC WFD Crude Mesh (T3) m = 6 m = 10 5.456 3.896 5.897 0.852 3.732) (1.1.866.120 0.574 1. However.003 2.0 3.828 1.0.0) (1.641 5.285 2.298 2. It can be seen that the accuracy of the numerical solution is well within acceptable limits.641 5. LLC .118 3.0) (1.674 Crude Mesh (T6) m = 6 m = 10 2.1.689 1. the MLS approximants.525 7.319 2. equation (8. Figures 9.
© 2005 by Taylor & Francis Group.3 of this book. the serial BNM is considerably slower than the serial BEM and this is because the approximation functions in the BNM need to be generated at each point unlike in the BEM. Certainly.5 2 3 6 8 u 4 5 6 7 8 10 12 14 BNM solution without modification BNM solution with modification Exact solution 0 0. Table 9. Values of u and ∇u.5 compares the wallclock times from a serial BEM code with a serial and a parallel version of the BNM code.9. One possible remedy is an accelerated BNM (Kulkarni et al. 16. not the solution phase. It is clear that this modiﬁcation is essential for success of this method at computing derivatives of the potential function at internal points that are close to the boundary of a body. along the line x1 = x2 = x3 . Also. has been parallelized.4: Variation of the potential u and tangential derivative of u.5 2 Radial coordinate Radial coordinate (a) (b) Figure 9. are obtained with the usual BNM (without modiﬁcation) and by a new application of equations (17. for the BEM for elasticity problems. and for the BNM for elasticity in Section 9. and 32 processors using the message passing interface (MPI) standard on the IBM SP2 (R6000 architecture. The results obtained by the BNM have also been compared with those from the conventional BEM for a Dirichlet problem on a sphere with the exact trigonometric solution as given in equation (9. The parallel BNM code is run on 4. 0 1 0 2 2 4 ∂ u/ ∂ s BNM solution without modification BNM solution with modification Exact solution 0 0.5 1 1.3 later in this chapter. This issue.19).1. [77]).2.25) and (17. The gradient is dotted with the diagonal (x1 = x2 = x3 ) in order to get the directional derivative along this line. at internal points that are close to the surface of the body. The parallel version is an early one in which only the assembly phase of the BNM matrices.1.5 1 1. 120 MHz P2SC Processor). with a prescribed cubic solution (from [26]) Comparison of BEM and BNM.34) in Kane [68] (with modiﬁcation). POTENTIAL THEORY IN THREE DIMENSIONS 161 boundary value problem is solved upon imposing the cubic solution on a crude mesh (72 T6 cells) with the Gaussian weight function and a quadratic basis (m = 6). the BNM is very easy to parallelize. is discussed in detail in Section 1. (Similar observations have been made by other researchers regarding the performance of the EFG compared to the FEM). LLC .
3.162 Meshes 72 T6 cells 128 T6 cells 288 T6 cells CHAPTER 9.3 secs 47. Figure 9.23) It can be clearly seen that the two methods yield comparable results and have almost identical rates of convergence.1.6 secs 249.4 secs 35. LLC .3 secs 690. Once again.0 secs 53.0 secs 2. POTENTIAL THEORY AND ELASTICITY Serial BEM 3.3. The L2 error in τ is deﬁned here as: e(τ ) = (num) (exact) 2 NB − τi ) i=1 (τi (exact) 2 NB ) i=1 (τi (9. parallel versions drastically reduce wallclock times.2 Results from the HBNM This section follows the layout of Section 9.5 secs Serial BNM 29. as shown in Table 9.5 secs 7. Dirichlet problems have been solved with various forms of the potential function (from © 2005 by Taylor & Francis Group.3 secs 4.4 secs 27.5 secs 106.5.7 secs Parallel BNM 4 Procs 16 Procs 10.0 secs 32 Procs 1.5: Convergence of the BNM and the BEM with a prescribed trigonometric solution (from [26]) 9.5: Comparison of wallclock times for the BNM and BEM for a Dirichlet problem on a sphere (from [26]) and.1 above.3 secs 9.5 shows a comparison of the L2 error for the BEM and the BNM as a function of the number of boundary nodes.1.3 secs Table 9. 10 0 BNM solution BEM solution 10 1 e(τ) 10 2 10 3 10 1 10 2 10 3 NB Figure 9.
Another important feature of the MLS approximants is the range of inﬂuence associated with each node. Computation of geodesics can get quite cumbersome on a general curved surface described by splines. θQ . For this speciﬁc example. Linear/Quadratic/Cubic). Thus. In this work the parameter dI is chosen to be nonhomogeneous in the sense that each evaluation point has an identical number of nodes in its domain of dependence.15 . it is expected that the computation of geodesics on complicated shapes will not be a hindrance towards using the present methodology.22) is used in all cases. Also. These results are taken from [27]. Overall. However.2). for a given polynomial basis (e. cubic and trigonometric solutions imposed on the sphere. θP .9. φP ) Field point : Q (x1Q . φQ ) Curvilinear coordinates between P and Q : s1 = R(φQ − φP ) . with one node per cell placed at the centroid of each triangle in the parent space (Figure 9.19)) prescribed on the surface of a sphere. Now. while Table 9. with idealized geometry and boundary conditions. x3Q ) ≡ (R. x2P . x2Q . for all the numerical examples presented in this section. quadratic. this arrangement yields the best numerical results. In order to construct the interpolating functions using MLS approximants. the exact geodesic between two points x (collocation node) and y (ﬁeld point) is the length of the arc between these points on the great circle containing them. Table 9. LLC .1. POTENTIAL THEORY IN THREE DIMENSIONS Collocation point : P (x1P . 72 T6 triangular cells have been used on the surface of a sphere. The parameter which controls the socalled “compact support” associated with each node is dI . For the sphere. s2 = R(θQ − θP ) ˜ ˜ Exact geodesic Approximate geodesic Ψ : Angle between P & Q d = (˜2 + s2 ) s1 ˜2 Ψ = arccos(¯P · ¯Q /R2 ) r r d = RΨ 163 Table 9. it is necessary to compute the geodesic on the bounding surface of a body. x3P ) ≡ (R.6 summarizes the procedure for computing the exact and approximate geodesic on a sphere.6: Exact versus approximate geodesics on the surface of a sphere (from [27]) equations (9. Range of inﬂuence of nodes.7 presents a comparison in the (global two) L2 errors for linear. © 2005 by Taylor & Francis Group.9. The Gaussian weight function (8. a very simple approximation to the geodesic would be to use the “Euclidean” distance between points x and y.g. it can be seen that the errors remain reasonably small even if the approximate geodesic is used to replace the exact one. Geodesics.
The Dirichlet boundary value problem is solved upon imposing the © 2005 by Taylor & Francis Group. Figures 9. respectively.059 % 0. The values of dI for all nodes in S are set equal to dmax . for the cubic and trigonometric solutions. 6 25 Cubic function Trigonometric function 20 m=3 5 Cubic function Trigonometric function m=6 4 L error in ∂ u / ∂ n 15 L error in ∂ u / ∂ n 3 10 2 2 5 2 1 0 4 6 8 10 Number of nodes (n) 12 14 0 10 12 14 Number of nodes (n) 16 18 (a) m=3 (b) m=6 Figure 9. of the node in S which is farthest from E.109 % 0.279 % Table 9.164 CHAPTER 9. This fact has also been observed for the BNM. It is observed that the lowest value of L2 errors is obtained for n ∈ (2m.665 % 0. For a given n.172 % Approximate geodesic 0. (b) quadratic polynomial basis (m = 6) (72 T6 cells with one node per cell) (from [27]) Results at internal points. Figures 9. let S be the set of nodes in the domain of dependence of a particular evaluation point E.6: Global two L2 error in τ for varying number of points in the domain of dependence of an evaluation point E: (a) linear polynomial basis (m = 3). along the geodesic.7: Global two L2 error in τ for Dirichlet problems on a sphere with the exact and an approximate computation of geodesics (from [27]) the number of nodes n in the domain of dependence of each evaluation point becomes the parameter of interest.6 (a) and (b) show the eﬀect of varying the number of nodes n for a linear basis (m = 3) and a quadratic basis (m = 6). 3m). The parameters dI are chosen as follows.0717 % 0.7 (a) and (b) show variation in the potential and its x1 derivative. for points along the x1 axis inside the sphere. LLC . where dmax is the distance from E. POTENTIAL THEORY AND ELASTICITY Exact solution Linear Quadratic Cubic Trigonometric Exact geodesic 0.715 % 4. respectively.313 % 1.
23)) for the HBNM and BEM. Comparison of BEM and HBNM.8 presents a comparison in the L2 error in τ (as deﬁned in (9. as functions of the (global) number of nodes. The results obtained by the HBNM have also been compared with those from the conventional BEM for a Dirichlet problem on a sphere with the trigonometric solution (equation (9. An analogous method for potential theory has been employed in order to get accurate results at internal points that are close to the surface of the sphere in Figure 9.2.1.1. and in Section 9.9.2.19)).5 2 (a) (b) Figure 9. 9. is more accurate than the BEM solution for this example.2.1. the bounding surface ∂B of a body B is partitioned into Nc cells ∂Bk and MLS approximations for © 2005 by Taylor & Francis Group. Figure 9.2 9.1. The HBNM solution. Once again. This ﬁgure shows that the two methods yield comparable results and have similar rates of convergence.3 (for the BNM) later in this chapter.3 (for the BEM).1 9.7. It is seen from these ﬁgures that the HBNM solutions match the exact solutions within plotting accuracy for both u and ∂u/∂x1 . the reader is reminded that a technique for dealing with nearly singular integrals in linear elasticity is described in detail in Section 1.5 1 x1 1.5 2 u x 10 _6 _3 _7 _4 _ 8 9 _5 _ _ 6 7 0 _ 10 _ x1 0.1 Linear Elasticity in Three Dimensions BNM: Coupling of BIE with MLS Approximants Coupled equations As before for the case of potential theory (Section 9.1). LLC .5 1 1. LINEAR ELASTICITY IN THREE DIMENSIONS _ 165 3 4 0 _ HBNM solution Exact solution _1 HBNM solution Exact solution _5 _ 2 ∂ u/ ∂ x 1 x 10 0 0. however.2.7: Variation of (a) potential and (b) ∂u/∂x1 along the x1 axis for a sphere with the prescribed trigonometric solution (from [27]) trigonometric solution on a cell conﬁguration consisting of 72 T6 cells with one node per cell and a quadratic basis (m = 6).
25).8) are used in the usual BIE for 3D linear elasticity (1. y) I=1 ΦI (y)ˆiI − u I=1 ΦI (x)ˆiI u dS(y) (9.1. as before.1. 3NB of which are prescribed by the boundary conditions. The ˆ ˆ other 6NB quantities of interest are ui .166 CHAPTER 9. ny nodes lie in the domain of dependence of the ﬁeld point y and nx nodes lie in the domain of dependence of the source point x. 2. 3.2 Discretized equations and boundary conditions This discussion closely follows that on potential theory in Section 9.1. τi . i = 1. The discretized form of (9.24) where ΦI (x) and ΦI (y) are the contributions from the Ith node to the collocation point x and ﬁeld point y respectively. at NB boundary nodes. 2.2. LLC .8: Comparison of e(τ ) (L2 error in ∂u/∂n) for the HBNM and the conventional BEM for a Dirichlet problem on a sphere (from [27]) the functions ui and τi (8. 9.8). The remaining 6NB equations.3. τi .1. i = 1. regular Gaussian integration is used.24) has the same form as (9. that relate the nodal approximations of ui and τi to their nodal © 2005 by Taylor & Francis Group.2 are used to carry out the integrations. Also. When x and y belong to the same cell.1. at NB boundary nodes. but this time with 3NB equations for the 6NB quantities ui . the cell is treated as a singular cell and the special techniques described in Section 9. 3. y) I=1 ny ΦI (y)ˆiI τ nx − Tik (x. Otherwise. The result is: Nc ny 0= m=1 ∂Bm Uik (x. POTENTIAL THEORY AND ELASTICITY 10 1 BEM solution HBNM solution 10 2 e(τ) 10 3 10 4 10 5 10 1 10 2 10 3 NB Figure 9.
[27]).50) and (1.3 Displacements and stresses at internal points The displacement at a point ξ ∈ B is obtained from (1. y) I=1 ΦI (y)ˆkI − uk (ˆ ) − uk. The complete system of equations is solved in a manner that is very similar to that described in Section 9.17) or (1. the BNM versions of continuous equations such as (1.2. These are [104]: uk (ξ) Nc = uk (ˆ ) x ny ny + m=1 ∂Bm Uik (ξ. For this purpose. 4.52) are needed. LLC .28). (ˆ )(y − x ) dS(y) u x x ˆ (9. y) I=1 ny ΦI (y)ˆkI − σkm (ˆ )nm (y) dS(y) τ x − ∂B Sijk (ξ. 9.21).40) and (1. evaluation of displacements and stresses at internal points. and the stress (directly) from (1.1.26) Equation (9. also. As mentioned several times before (see Sections 1. that are very close to the bounding surface of a body.2.3. requires special care. In each case. Displacement gradients on the surface of the body are obtained as part of the BNM solution of the original boundary value problem. of course.1.1 of this book (see.9.30).25) ny σij (ξ) = σij (ˆ ) + x ∂B Dijk (ξ. have the same form as (8.11).2. The HBNM equations are: Nc ny nx 0 = l=1 ∂Bl Dijk (x.2.26)) requires the displacement gradients and stresses at the ˆ target point x (in addition to the usual displacements and tractions at ﬁeld points y). adapted from Lutz et al. is described in detail in Section 1.41).27) or (1.3 for potential theory.2.8) must be used. This procedure.1.3).2 9.2. LINEAR ELASTICITY IN THREE DIMENSIONS 167 values.2.1 HBNM: Coupling of HBIE with MLS Approximants Coupled equations The starting point here are the HBIEs (1. y) τ I=1 ΦI (y)ˆiI −ui (ˆ ) u x dS(y) (9. (8. y) I=1 ΦI (y)ˆiI − Tik (ξ. [89] for the BEM. y) I=1 ΦI (y)ˆkI − τ I=1 ΦI (x)ˆkI τ dS(y) © 2005 by Taylor & Francis Group. the displacement gradient from (1. 9.
Displacement gradients at a boundary point x are obtained by applying the procedure outlined in Section 1.39) and (1.3 Numerical Results Numerical results from the standard and the new BNM. are presented in this section. The procedure followed for discretization of (9.2.2). y)(nm (y) − nm (x)) dS(y) ∂Bl ny nx − ∂Bl Sijk (x. LLC . The discretized version of equation (9.25) and (9.168 − σkm (x) CHAPTER 9.28) are the HBNM equations for linear elasticity.8). Three illustrative examples are considered in this section : a hollow sphere under internal pressure.27) and Nc ny nx 0 = l=1 ∂Bl Dijk (x.27) (see (1.2.2. y) I=1 ΦI (y)ˆkI − u I=1 ΦI (x)ˆkI − uk.27) and (9.1. while the weakly singular integrals are evaluated using the procedure outlined in Section 9.17) and (1. The resulting equation is an alternate version of (9.28) has the generic form shown in (9. 9.40)).30) while the new BNM results are obtained from equations (9. a bimaterial sphere and a cube with a spherical cavity under tension (the 3D Kirsch problem).1.m (x)(ym − xm ) dS(y) u (9.1.2. Nonsingular integrals are evaluated using the usual Gauss quadrature over surface cells. [104] and [27]. as expected. taking the limit x ∈ ∂B → x ∈ ∂B in the continuous equation (9. as well as from the HBNM. POTENTIAL THEORY AND ELASTICITY Dijk (x.28) is quite analogous to the BNM case described before in Section 9. 9.2. These equations are fully regularized and only contain nonsingular or weakly singular integrands.m (x)(ym − xm ) dS(y) u (9. © 2005 by Taylor & Francis Group.28) ˆ Note that. y)nj (x) ∂Bl I=1 ΦI (y)ˆkI − u I=1 ΦI (x)ˆkI − uk.2. The standard BNM results are obtained from (the BNM versions of) equations (1.26).2 Discretization Equations (9.26) yields a corresponding regularized HBNM. These numerical results are taken from [25]. y)nj (x) I=1 ΦI (y)ˆkI − τ I=1 ΦI (x)ˆkI τ dS(y) − σkm (x) ∂Bl Dijk (x. y)nj (x)(nm (y) − nm (x)) dS(y) ny nx − Sijk (x.1 and the stress components at this point are then obtained from Hooke’s law.2.
with tractions prescribed over both the inner and outer surfaces of the sphere. The resulting singular matrices are regularized using the procedure described in [25] (see. LLC . 72 quadratic T 6 triangular cells (with one node at the centroid of each cell) are used on each surface of the hollow sphere.1 Hollow sphere under internal pressure The inner and outer radii of the sphere are 1 and 4 units. LINEAR ELASTICITY IN THREE DIMENSIONS 169 0. as a function of the radial coordinate.3. respectively.5 3 3.1 0 HBNM solution Exact solution 1 1. It is observed that the stress solutions from the standard BNM are meaningless at © 2005 by Taylor & Francis Group.9 (a) while corresponding results from the HBNM appear in Figure 9.1 0 Radial displacement Radial displacement 0.3 0.2. [90]). Finally.6 0.3 0. The agreement between the exact and numerical solutions are observed to be excellent. together with the exact solutions from [167]. appear in Figures 9.5 0. It is observed that the results from the standard BNM and HBNM are excellent as long as an internal point is not very close to one of the surfaces of the sphere. also. while those from the new BNM are very accurate everywhere.9: Radial displacement as a function of the radial coordinate in a hollow sphere: (a) from the standard and the new BNM (b) from the HBNM.11. Internal stresses.10 (a) and (b). The radial displacement. are shown in Figure 9.4 0.8.6 0. from the standard as well as the new BNM. The exact solution from [167] is also included in these ﬁgures.9 (b). the internal pressure is 1.5 4 Radial coordinate Radial coordinate (a) (b) Figure 9. The radial and circumferential stresses from the new BNM and from the HBNM.2. including at internal points that are close to the bounding surfaces of the sphere. the stress solutions from the standard BNM are shown in Figure 9.2 0.5 3 3. Internal displacements.25.5 2 2.9.4 0.5 2 2. The entire surface of the sphere is modeled here. together with details near the inner surface of the sphere in Table 9. together with the exact solution (from [104] and [27]) 9. E = 1 and ν = 0.5 4 1 1. respectively.5 BNM standard method BNM new method Exact solution 0.2 0.
5 2 2.5 4 _1 1 1. LLC .5 4 Radial coordinate Figure 9.5 3 3.5 0.5 2 2.10: Radial and circumferential stresses as functions of the radial coordinate in a hollow sphere: (a) from the new BNM (b) from the HBNM.5 Stress components 0 Stress components 0 0. from the new and standard BNM.5 BNM solution for σ θθ BNM solution for σ rr Exact σ θθ Exact σrr _ 0.5 2 2.5 4 Radial coordinate Radial coordinate (a) (b) Figure 9.5 3 3.170 CHAPTER 9. together with the exact solutions (from [104] and [27]) 5 4 Stress components 3 2 BNM solution for BNM solution for BNM solution for BNM solution for Exact σ θθ Exact σrr σθ θ σ rr σ θθ σrr (standard method) (standard method) (new method) (new method) 1 0 1 2 1 1.11: Radial and circumferential stresses as functions of the radial coordinate in a hollow sphere.5 3 3. POTENTIAL THEORY AND ELASTICITY 0. together with exact solutions (from [104]) © 2005 by Taylor & Francis Group.5 HBNM solution for σ θθ HBNM solution for σ rr Exact σθ θ Exact σrr 1 1 1.
98020 0.74915 0.51168 0.72693 standard 3. For material 2: u(2) = r r B2 A2 (1 − 2ν2 ) − 3 (1 + ν2 ) E2 2r B2 r3 .50887 0.18719 0.2.01546 0. σtt = A2 − (2) (9.43126 3.49365 0.40084 0. LLC . For material 1: u(1) = A1 r r 1 − 2ν1 E1 (9. Numerical results for this model have been obtained by prescribing displacements on the outer boundary.76287 1. the exact solution for the radial displacement and stresses in each material is given below.03 1.38727 Table 9. such as that presented in Section 9.86168 0. One could also prescribe tractions over the entire outer surface and then appropriately modify the scheme presented in [25] for solving traction prescribed problems.99057 5.01 1.30) (1) σrr = σtt = A1 (1) where the constant A1 is deﬁned below.85789 0.43050 0.8: Radial and circumferential stresses as functions of the radial coordinate in a hollow sphere.35304 0.24717 exact 0.37928 171 r 1.08215 0. This is planned for the future.12 (a) shows a schematic of two perfectly bonded spheres of two diﬀerent materials.1.2 Bimaterial sphere The BNM has been extended to solve problems involving material discontinuities.9.48071 0.07 1.45465 0. 9.91379 0.29) (9.40809 0. Figure 9. and an algorithm for improving these results.36737 0.2.31) (9.79589 0.09 1.33) © 2005 by Taylor & Francis Group. together with the exact solutions.42834 0.2. Upon prescribing a radial displacement u0 on the outer surface of the sphere. at points very close to the inner surface (from [104]) internal points very near the inner surface of the sphere.71229 σθθ new 0.05 1. from the new and standard BNM. is absolutely essential in this case.3.39290 0.3.68714 1.90056 exact 0.11 standard 4.81338 0. LINEAR ELASTICITY IN THREE DIMENSIONS σrr new 0.32) (2) σrr = A2 + B2 2r3 The constants are: A1 = A2 + B2 3 R1 (9.97012 0.92753 0.76857 0.46201 0.
3 3D Kirsch problem The 3D Kirsch’s problem consists of a cube with a small spherical cavity subjected to far ﬁeld uniform tension (Figure 9. The material and geometric parameters are : E = 1.35) (9. ν = 0.5 1 1.3.33.0.0). A constant radial displacement is prescribed on the outer boundary of material 2 (u0 = 1.172 CHAPTER 9. β= .0.6 0. E1 = 1.9 Radial displacement E 2 ν2 R1 0. POTENTIAL THEORY AND ELASTICITY x2 1 0. LLC .0. 9. ν1 = 0.14 (a)).4 0.5 0.3 0.0.36) A2 = u 0 E2 1 + β 3 R2 R1 C(1 + α) C= (1 − 2ν2 )(1 + β) 1 + ν2 + 3 3 R1 (1 − α) 2R2 The material and geometric parameters chosen for the two materials are : • for material 1.0 • for material 2. cutout radius a = 1.2 0.5 4 Radial coordinate (b) Figure 9. Figures 9.34) (9. Figure 9.12: Bimaterial sphere subjected to uniform external radial displacement : (a) conﬁguration (b) radial displacement along the x1 axis from the new BNM together with the exact solution (from [25]) α= E1 (1 − 2ν2 ) E2 (1 − 2ν1 ) .2. R1 = 1.5 2 2.25.0.0.8 0.28. along the line x1 axis.7 0.12 (b) shows a comparison of the numerical solution and the exact solution for the radial displacement within the two materials.13 (a) and (b) show the radial and circumferential stress. respectively. Again. E2 = 2. the loading is applied without restraining any rigid body © 2005 by Taylor & Francis Group. It can be seen that the jump in the circumferential stress at the bimaterial interface is very well captured by the new BNM. E1 (1 + ν2 ) 2E2 (1 − 2ν1 ) B2 = − u 0 E2 R2 C (9.5 3 3. R2 = 4. ν2 = 0.1 x1 R2 E 1 ν1 BNM solution Exact solution x3 (a) 0 0 0. side of cube 2b = 20.
is given as [167]: σ33 = σ0 1 + 4 − 5ν 2(7 − 5ν) a r 3 + 9 2(7 − 5ν) a r 5 (9. for points in the plane x3 = 0. it can be clearly seen that the BNM and HBNM solutions are in excellent agreement with the analytical solution.2 1.5 2 2.8 1. The exact solution for the normal stress σ33 .3 is essential for obtaining accurate values of stresses near the surface of the cavity.4 Circumferential stress 1. The cell structure consists of 96 Q4 cells modeling the cube and 72 T6 cells modeling the spherical cavity.2. © 2005 by Taylor & Francis Group.5 4 Radial coordinate Radial coordinate (a) (b) Figure 9.4 1.9. HBNM and the exact solution for the (normalized) normal stress σ33 /σ0 along the x1 axis.13: Bimaterial sphere subjected to uniform external radial displacement : (a) radial stress σrr (b) circumferential stress σθθ along the x1 axis from the new BNM. LLC .5 173 1.3 1.1. LINEAR ELASTICITY IN THREE DIMENSIONS 1.2.5 4 0.5 1.1 1 Radial stress 1. again with one node per cell.6 1. It should be noted that the algorithm presented in Section 9.5 3 3.37) Figure 9.5 2 2.3 1.5 3 3.1 BNM solution Exact solution 1 BNM solution Exact solution 0. together with the exact solutions (from [25]) modes and the scheme presented in [25] is used to obtain meaningful numerical results.9 0 0. Again.5 1 1.7 1.5 1 1.14 (b) shows a comparison between the new BNM.9 0 0.2 1.
27]) © 2005 by Taylor & Francis Group. LLC .4 1. together with the exact solution (from [25. POTENTIAL THEORY AND ELASTICITY X3 2.174 CHAPTER 9.14: 3D Kirsch problem : (a) conﬁguration (b) σ33 /σ0 along the x1 axis from the new BNM and the HBNM.2 2 1.8 1.6 1.8 0 1 2 x1 3 4 5 (a) (b) Figure 9.2 1 X2 BNM solution HBNM solution Exact solution X1 σ /σ 33 0 0.
2) (10. while applications in 3D elasticity are discussed in the next chapter . The idea of using hypersingular residuals.1 10. for problems in 3D potential theory and linear elasticity [28]. the HBNM (equation (9.1 Hypersingular and Singular Residuals The Hypersingular Residual Let the BNM (equation (9. to obtain local error estimates for the BIE. has appeared in Mukherjee and Mukherjee [111]. was ﬁrst proposed by Paulino [122] and Paulino et al. This idea. Menon et al. [96] and Paulino et al.Chapter 10 ADAPTIVITY FOR 3D POTENTIAL THEORY The subject of this chapter is error analysis and adaptivity with the BNM.1) and its numerical solution be (u . LLC (10. 10. This idea has been applied to the collocation BEM (Paulino et al. applied to the BCM.j is deﬁned as. This idea has also been applied to the BNM.1)) for potential theory be written in operator form as: LBN M (u. and has been discussed in detail in Chapter 2 of this book. τ ). [127]). as applied to problems in 3D potential theory. τ ) = 0 ∗ ∗ (10.3) . [123]. τ ∗ ) 175 © 2005 by Taylor & Francis Group.the last one in this book. τ ) = 0 The hypersingular residual in the potential gradient u. Applications in 3D potential theory is the subject of this chapter. and is presented in Chapter 7 of this book. rj ≡ residual(u. [123].1. Also.j ) = LHBN M (u∗ .12)) is written in operator form as: LHBN M (u.
the ones proposed in this chapter. LLC .5) © 2005 by Taylor & Francis Group.1: Interchange of BIE and HBIE (a) hypersingular residual (b) singular residual (from [28]) and is calculated from equation (9. one gets from equation (9. [123]).1(a) It has been proved in [96] and [127] for the BIE that.13): LHBN M (uo . Thus.1): (10.4) where r is some scalar measure of a hypersingular residual and is a scalar measure of the exact local error. but diﬀerent in detail from. τ o ) = 0 and from equation (9. In this case. a hypersingular residual is expected to provide a good estimate of the local error on a boundary element. one can reverse the above procedure to deﬁne singular residuals by ﬁrst solving the HBIE and then iterating with the BIE. Paulino et al. for potential theory. 10. under certain favorable conditions.176 CHAPTER 10. ADAPTIVITY FOR 3D POTENTIAL THEORY BIE BVP Solution HBIE BVP Solution HBIE Residual Error estimate (a) BIE Residual Error estimate (b) Figure 10.12). This idea is illustrated in Figure 10.2 The Singular Residual The argument for using the residuals as error estimates is symmetric (see Paulino [122]. Therefore. real positive constants c1 and c2 exist such that: c1 r ≤ ≤ c2 r (10.1. It should be mentioned here that the deﬁnitions of the residuals used in [96] and [127] are analogous to.
τ o ) (10. The BNM equation used for solving the BVP is (9.10. the latter set of methods provides substantially more ﬂexibility in the (re)discretization process than the former ones. while the current work presents a uniﬁed residual formulation.2: Typical selfadaptive iterative BNM algorithm (h−version) according to the scheme of Figure 10.2. 124]) and meshless methods.2 Error Estimation and Adaptive Strategy There are similarities between adaptive techniques (e. and the HBNM equation used for residual computation is (9. hversion) for meshbased methods (see Paulino et al. The above formulation for singular and hypersingular residuals is a generalization of the earlier work by Menon et al. © 2005 by Taylor & Francis Group.Figure 10. 10. However.g. LLC .2.1).12) (from ([28]) r ≡ residual(u) = LBN M (uo .1(a). ERROR ESTIMATION AND ADAPTIVE STRATEGY Start Initial cell configuration 177 Redefinition of cells. [126. in as simple a manner as possible. [96] in the sense that Dirichlet. The h−version iterative selfadaptive procedure employed in this work is presented in the ﬂowchart . The goal is to eﬃciently develop a ﬁnal cell conﬁguration which leads to a reliable numerical solution.6) This idea is illustrated in Figure 10. Neumann and mixed problems require separate prescriptions in [96]. nodes and regions of influence No Convergence Yes Final configuration BNM simulation HBNM residuals Iterative Cell Design Cycle Global error estimates Local error estimates Figure 10.1(b).
respectively.178 CHAPTER 10. for problems in potential theory.2. are used as error estimators.8) and (10.11) Here. From equation (10.13) This error measure is only used in Figure 10.2. at each node.j .j (exact) − u. In all the adaptivity examples presented in this chapter. The scalar residual measures.3): rj = residual(u.12)) is also used for ∂u/∂n at points on the surface of a cube (see examples of Section 10. e = u(exact) − u(numerical)  (10. evaluated at nodes. These equations are presented here for the sake of completeness.12) are used to obtain the singular residual and exact error.6) and (10. 10.12) (10.10) Equations (10.j . at each node. deﬁned above.j (numerical) (10. respectively. r and are themselves scalar measures of the residual and exact error.2 Local Residuals and Errors . is deﬁned as: j (10.Hypersingular Residual Approach From equation (10. respectively. © 2005 by Taylor & Francis Group.6): r = residual(u) and the exact local error in u is deﬁned as. Equations (10. This quantity is deﬁned as: e ∂u ∂n = ∂u (exact) ∂u (numerical) − ∂n ∂n (10. LLC . in the potential u.Singular Residual Approach The singular residual is deﬁned in an analogous fashion. ADAPTIVITY FOR 3D POTENTIAL THEORY 10. u. in the gradient u.1 Local Residuals and Errors .7) (10.j ) A scalar residual measure is deﬁned as: r = rj rj The exact local error in the gradient.8.10) are used to calculate the hypersingular residual and exact error. The local error measure (equation (10.3).9) and the corresponding scalar measure is deﬁned as: = j j (10.8) = u. for problems in potential theory.
These global errors are used in many of the tables that are presented later in this paper.0. then the number of cells to be reﬁned is less than with γ = 1. If γ < 1.0.3).2. The advantage in this case is that the reﬁnement rate increases. the computational eﬃciency may decrease owing to likely generation of an excessive number of cells. ERROR ESTIMATION AND ADAPTIVE STRATEGY 179 one node is used for each cell and is placed at its centroid. 10.14) where Nn is the total number of nodes.4 of this chapter. or over the whole boundary ∂B. © 2005 by Taylor & Francis Group. In this work. is presented in Section 10.” The standard procedure consists of using γ = 1.0. The parameter γ in equation (10. for a ONEstep reﬁnement. Of course.2.2. 10.4 Global Error Estimation and Stopping Criterion Global L2 error.16) is ¯ only useful for test examples in which the exact solution is known. If the inequality: r>γ r ¯ (10. An alternative procedure. A global L2 error. is deﬁned as ¯(φ) = A (φ(exact) − φ(numerical) )2 dA 100% (φ(exact) )2 dA A (10.3 Cell Reﬁnement Criterion A simple criterion for cell reﬁnement consists of subdividing the cells for which the error indicator is larger than a certain reference value.14)) or ¯ from equation (10.0. then the cell is subdivided into four pieces (see Figure 10. According to Figure 10.0. equation (10. If γ > 1.2. the increase on the total number of cells is comparatively small when γ > 1. the reference quantity is taken as the average value of the error indicator (here the average residual) given by: r= ¯ 1 Nn Nn r(i) i=1 (10.15) is satisﬁed. however.16). however. A comparison of the residual r and exact error demonstrates the eﬀectiveness of residuals as error estimates.16) where φ is a variable of interest and A is the area of a panel or of the whole surface ∂B. then the number of cells to be reﬁned is larger than that with γ = 1. An indication of overall convergence may be obtained by evaluating either r (equation (10.0. The scalar residual measure at this centroidal node is used as an error estimator for that cell. LLC .10. on a panel. the numerical solution of the next iterative step is expected to be more accurate than that of the current step.15) is a weighting coeﬃcient that controls the “cell reﬁnement velocity.
LLC . .3 Progressively Adaptive Solutions: Cube Problem The adaptive process illustrated by Figure 10. . Several aspects are investigated such as the quality of the adaptive solution obtained for scalar (potential theory) and vector ﬁeld (elasticity theory) problems.1(a). Quadrilateral cell . Ideally. The goal of the adaptive procedure is to obtain welldistributed (i. Subdivided cell Figure 10.g. Triangular cell . Laplace’s equation is solved using the BNM. performance of the method on problems with either pure or mixed boundary conditions.3: Cell reﬁnement for quadrilateral and triangular cells with one node per cell (from ([28]) Stopping criterion. Subdivided cell . This example. .17) where rglobal has a preset value. ADAPTIVITY FOR 3D POTENTIAL THEORY . which depends on the overall level of accuracy desired. the error estimates should decrease both locally and globally.3) can be stopped when: r ≤ rglobal ¯ (10. permits assessment of various parameters of the adaptive strategy for meshless methods based on BIE techniques.180 CHAPTER 10. as the iterative cell reﬁnement progresses. evaluation of the quality of error estimates obtained by means of hypersingular or singular residuals. 10. For generic problems where the exact solution is not available (e. .e. near optimal) cell conﬁgurations.2 is applied to the representative example of a Dirichlet problem on a cube. cell reﬁnement (see Figure 10. . . together with an elasticity example discussed in Chapter 11. according to Figure 10. and the (hypersingular) residuals are obtained using the HBNM. sensitivity © 2005 by Taylor & Francis Group. most engineering problems).
A quadratic basis is used for the construction of the MLS interpolating functions.0 in equation (10. It can be seen from Figure 10. It can be clearly seen that the hypersingular residual tracks the exact error perfectly.2)).18) Note that the solution (10.3) and (10. This comparative procedure tests the sensitivity of the adaptive scheme with respect to the initial conditions. the adaptivity results in this section have been obtained starting with two diﬀerent relatively coarse initial cell conﬁgurations.3. the boundary value problem is solved again using the BNM. The idea behind the adaptive procedure is to start with a rather crude cell conﬁguration and carry out cell reﬁnement in the region where the residual is large according to a certain criterion.3. This is an indication that the procedure for error estimation and adaptivity is moving in the right direction.e.4(a) shows a discretization consisting of 54 rectangular cells with one (centroidal) node per cell.j (from equations (10. LLC .9) and (10.1)).5 shows a comparison between the hypersingular residual (from equations (10.1 summarizes the various output parameters of the adaptivity procedure. Figure 10.4(b) that the cell reﬁnement occurs only at the corners where the exact error is the largest. The boundary value problem is solved using the BNM (equation (9.10)) computed for the initial conﬁguration # 1 (Figure 10. m = 6 (see equation (8.1 that excellent numerical results are obtained in a single step of the adaptivity process and hence the adaptive procedure is not continued further.8)) and the exact local error in u. Cell reﬁnement is carried out using γ = 1. i.1 Exact Solution The following exact solution.18) is symmetric with respect to y and z but that its dependence on x is diﬀerent from its dependence on y or z. The appropriate value of u is prescribed on ∂B (Dirichlet problem) and ∂u/∂n is computed on ∂B.3. this is a proper test of the meshless method and the adaptivity procedure. 10. Hence. Because the exact solution cannot be represented in terms of polynomials. Then the results are used in the HBNM (equation (9.4(b). and the resulting reﬁned cell conﬁguration consisting of 126 cells is shown in Figure 10.10. which satisﬁes the 3D Laplace’s equation. Now.2 Initial Cell Conﬁguration # 1 (54 Surface Cells) Figure 10.4(a)) at each node on the surface.12)) to obtain the hypersingular residual. and convergence properties. PROGRESSIVELY ADAPTIVE SOLUTIONS: CUBE PROBLEM 181 of the “ﬁnal” solution with respect to the starting cell conﬁguration (initial condition of the selfadaptive problem). is used in this example : u = sinh πx1 sin 2 πx2 √ 2 2 sin πx3 √ 2 2 (10. It can be seen from Table 10. Table 10. 10. © 2005 by Taylor & Francis Group.15).
5 0 0. obtained with γ = 1 (from ([28]) Output parameters Number of cells x = ±1 y = ±1 z = ±1 All faces Average residual (¯) r Maximum residual rmax Initial 54 1.4: Cell conﬁgurations on the surface of a cube. ADAPTIVITY FOR 3D POTENTIAL THEORY 1 0.4(a)) and the conﬁguration obtained at the end of the ﬁrst step of the adaptivity process using γ = 1.9) and (10.12)).6 shows a comparison between the hypersingular residual (from equations (10.5 0. computed for the initial conﬁguration # 2 (Figure 10.3068E01 Table 10. It can be clearly seen that the hypersingular residual tracks the exact error very accurately.4(b)) (from [28]) 10. Figure 10.7605E02 0.5 _ z _ 1 1 0.5 0 0.7(a)) with. In fact.4209 % 7.j (from equations (10.5 _1 _1 0 _ 0.2578 % 0.3 Initial Cell Conﬁguration # 2 (96 Surface Cells) The initial conﬁguration # 2 is for the same physical cube with 16 uniform cells on each face (Figure 10. the results for the ﬁner cell conﬁguration #2 are very similar to those shown in Figure 10.0 (Figure 10. as always.1450 % 0.5 _ 1 0.3. one node at the centroid of each cell.5 0 _ z _ 1 1 0.1: ¯(∂u/∂n) (see (10. (b) First adapted step : 126 cells.16)) and residuals r.5 0 _ 1 0.1)). © 2005 by Taylor & Francis Group.5 _ 1 0.6911 % 2. the boundary value problem is solved using the BNM (equation (9. (a) Initial conﬁguration # 1: 54 surface cells.7(a)).5 0. LLC . and the results obtained are used in the HBNM (equation (9.5197E01 Final 126 0.2366E01 0.8)) and the exact local error in u.0238 % 0.3) and (10. rmax for the initial cell ¯ conﬁguration (Figure 10. As before.6911 % 7.5 0 _ 0.2773 % 0.5 for the coarser initial cell conﬁguration # 1.182 CHAPTER 10.5 x y 1 _1 x y (a) (b) Figure 10.10)).0519 % 0.
one node per cell).5: Comparison of hypersingular residual and exact local error in u.6 0. These quantities have been normalized by their respective maximum values. one node per cell).5197 × 10−1 and max = 0.j for the initial conﬁguration # 2 (96 cells. where rmax = 0.6223 × 10−1 (from ([28]) © 2005 by Taylor & Francis Group.6 0.2 Hypersingular residual and exact error Hypersingular residual Exact error 1 0. These quantities have been normalized by their respective maximum values.2 0 0 20 40 60 Node number 80 Figure 10. where rmax = 0. LLC .3.1829 × 10−1 and max = 0.4 0.10.2 Hypersingular residual and exact error Hypersingular residual Exact error 1 0. PROGRESSIVELY ADAPTIVE SOLUTIONS: CUBE PROBLEM 183 1.8 0.6: Comparison between hypersingular residual and exact local error in u.4 0.2051 (from ([28]) 1.j for the initial conﬁguration # 1 (54 cells.8 0.2 0 0 10 20 30 Node number 40 50 Figure 10.
(d) third step (1164 cells) (from ([28]) © 2005 by Taylor & Francis Group.5 0 0.5 0 1 1 0.5 0 1 1 0.5 _1 1 0.5 x 0.5 x 0.5 _1 1 0.5 0 _ 0.5 _1 x 0 y (a) (b) 1 0.5 0 1 (c) (d) Figure 10.5 _1 _ 0.5 1 0.5 0 _ 0.5 0 z _ 0.5 y _1 _1 _ 0.184 CHAPTER 10.5 y 1 1 0. LLC .5 x 0.5 0 _ 0.7: Cell conﬁgurations on the surface of a cube (a) initial conﬁguration # 2 (96 surface cells).5 0 _ z z _ 0.5 0 0.5 y _1 _1 _ 0. ADAPTIVITY FOR 3D POTENTIAL THEORY 1 0.5 _1 1 0. (b) ﬁrst adapted step (168 cells). (c) second step (456 cells).5 0 _ 0.5 1 1 0.
7(c)) and 3 (Figure 10. Table 10.15) and starting from the initial conﬁguration # 2 given in Figure 10.7(b)) is not sensitive to this fact. which provides a comparison between the hypersingular residual (from equations (10. It is noted that the cell reﬁnement should begin at the corners of the cube where the error in ∂u/∂n is the largest. ¯(∂u/∂n) ﬁnally decreases from Step 2 to Step 3 and reaches its lowest value at this step.2 is repeated three times using γ = 0. Comparing the contour plots of Figures 10.7(a). (c) and (d). The resolution of these and subsequent contour plots is much ﬁner than the corresponding cell discretization because the error is actually evaluated at a large number of points on the boundary (panels) of the body. reﬁnement occurs close to the edges and corners where the error in ∂u/∂n is largest.9) and (10.7(a)). Thus. on the faces x = ±1. however.8(b) shows a contour plot for the exact local error e in ∂u/∂n on the y = −1 face of the cube for adapted cell conﬁguration of Figure 10.3) and (10.3.8(a) and (b)).13)) is calculated on each of the faces of the cube.10. Note that.10)). respectively. Figures 10. However.7(b).7(d). For the ﬁrst step of the adaptive procedure (see Figure 10. Step 1 (Figure 10. the local error e in ∂u/∂n (from equation (10. especially at the corners (cf Figures 10. The resulting reﬁned cell conﬁgurations are shown in Figures 10. ¯(∂u/∂n) (from equation (10.16)) increases from the initial conﬁguration to Step 1. These results conﬁrm the observation made at the end of the previous paragraph regarding regions of large errors which demand a ﬁner discretization. It is interesting to note that the absolute value of the exact solution (equation (10. It can be seen from Figure 10.9 (c) and (d)).2 summarizes the results of the adaptive process for the cube problem starting with the initial cell conﬁguration consisting of 96 cells (Figure 10. On the faces y = ±1 and z = ±1.18)) has the same functional dependence on the y = −1 and z = 1 faces and diﬀerent on the x = −1 face of the cube.8)) and the exact local error (from equations (10. The underlying cell structure on the face is also shown in the color plots.7(b).8(a)(d).7(b)). This is a tribute to the quality of residuals as error estimates. Figure 10.7(d)) of the adaptive procedure are. PROGRESSIVELY ADAPTIVE SOLUTIONS: CUBE PROBLEM 185 Adaptivity results. Figure 10. and from Step 1 to Step 2. the error in ∂u/∂n has reduced substantially. due to the reﬁnement procedure. Steps 2 (Figure 10.8 shows contour plots of the exact local error e in ∂u/∂n on the y = −1 face of the cube.9. which has the sophisticated cell pattern of Figure 10. LLC . selected results are shown in Figure 10. In order to obtain a better understanding of the adaptivity procedure.9 that the hypersingular residual tracks the exact error reasonably well.8(c) and (d) show the exact local error e in ∂u/∂n on the y = −1 face of the cube for the adapted cell conﬁgurations consisting of 456 cells and 1164 cells. The iterative cell design cycle of Figure 10. one can readily verify that the error in ∂u/∂n decreases substantially during the adaptive process. Note that.5 in equation (10. ¯(∂u/∂n) monotonically decreases as the number © 2005 by Taylor & Francis Group. respectively (see Figures 10. The results are shown on the x = −1 and z = 1 faces as a representative sample of the results over the 168 nodes. Other relevant comments are in order.
2 0 z 0.8 0.05 0.8 0.6 0.2 _ _ 0.2 z 0.01 _ 0.6 0.4 0. (b) ﬁrst adapted step (168 cells).12 0.6 0.06 0.02 _ 0.01 _ 0.01 0.04 0.1 0.4 0.5 0 x 0.03 0.07 0.5 0 x 0.05 0.02 _ 0.04 0.5 (c) (d) Figure 10.02 0.035 0. ADAPTIVITY FOR 3D POTENTIAL THEORY 0.015 0.8 0.8 z _ z _ _ _ 0.4 _ 0.04 0.2 0.04 0.2 0 0. (d) third step (1164 cells) (from ([28]) © 2005 by Taylor & Francis Group.06 0.4 0.5 _ 0.06 0.186 CHAPTER 10.4 0.4 0.2 0.4 0.6 0.6 0.03 0.2 0.02 _ 0.8: Error in ∂u/∂n (e(∂u/∂n)) on the face y = −1 of the cube (a) initial conﬁguration # 2 (96 surface cells).5 0 x 0. LLC .07 0.2 0 0.03 0.8 _ 0.005 0 x 0. (c) second step (456 cells).6 0.8 0.5 (a) (b) 0.8 _ 0.6 0.6 0.4 0.08 0.5 _ _ 0.025 0 0.5 _ _ 0.8 _ 0.045 0.
2 Hypersingular residual and exact error Hypersingular residual Exact error 1 0. LLC .4 0. PROGRESSIVELY ADAPTIVE SOLUTIONS: CUBE PROBLEM 187 1.6 0.10.2 0 0 10 20 30 Node number 40 50 Figure 10.3.9: Comparison of hypersingular residual and exact local error in u. where rmax = 0.j on the faces x = −1 and z = 1 of the cube of Figure 10.3645 × 10−1 (from ([28]) © 2005 by Taylor & Francis Group.1567 × 10−1 and max = 0.8 0. The quantities have been normalized by their respective maximum values.7(b) (ﬁrst step of the adaptive procedure).
2: L2 error in ∂u/∂n (¯(∂u/∂n)) and residuals r. decrease as the adaptive process progresses. The ﬂowchart of Figure 10. where h is a characteristic cell size in the area covered by the cells. LLC .19) where hi denotes the initial cell size.1567E01 Step 2 456 0.0551 % 0.15)). and setting the error estimate equal to η = r/(γ r) ¯ (see equations (10.0438 % 0.3661E02 0.1269 % 0.4 OneStep Adaptive Cell Reﬁnement The previous section has dealt with an iterative adaptive technique for cell reﬁnement (h−version).20) From equations (10.5 is used for the cell reﬁnement of the cube (from [28]) of adaptive cycles increases. as expected. rmax for the various ¯ steps of the adaptivity process starting with the initial cell conﬁguration consisting of 96 cells with one node per cell (Figure 10.5643E03 0. Figures 10.2781 % 0.19) and (10.1062 % 0. residual rmax Table 10.0451 % 0.7(a)).2537E02 Output parameters Number of cells x = ±1 y = ±1 z = ±1 All faces Average residual r ¯ Max.10 illustrates this idea which is based on the concept of reﬁnement level (RL) employed by Krishnamoorthy and Umesh [74].1829E01 Step 1 168 0.0759 % 1. Moreover.2785 % 0.11(a) and (b) show that diﬀerent degrees of reﬁnement are carried out for diﬀerent values of the reﬁnement level. the expression relating the ﬁnal cell size hf to the reﬁnement level RL is: hf = hi 2RL (10. 10.2089 % 0.1899 % 0. Here γ = 0.188 CHAPTER 10. From these ﬁgures. which are of order p. the global ¯(∂u/∂n) for “all faces.20).4963E02 0.3579E02 Step 3 1164 0. one obtains: hf = hi η 1/p (10. Here the interest is on developing a simple ONEstep algorithm for cell reﬁnement in the meshless BNM setting.0696 % 0.14) and (10.0551 % 0.2091 % 0.1811E03 0. Reﬁnement strategy. Assuming that the rate of convergence of the error is O(hp ).0654 % 1. the RL is given by: © 2005 by Taylor & Francis Group.1247 % 0. ADAPTIVITY FOR 3D POTENTIAL THEORY Initial 96 0.1135 % 0.” as well as the average and maximum residuals.
nodes.4. and regions of influence Error estimates Figure 10. The bold lines illustrate the idea of cell structure embedding (from ([28]) © 2005 by Taylor & Francis Group.10.11: Reﬁnement level RL using (a) rectangular and (b) triangular cells.10: ONEstep adaptive BEM algorithm based on multilevel cell reﬁnement (from ([28]) RL = 0 # of cells = 40 = 1 RL = 1 # of cells = 41 = 4 RL = 0 # of cells = 40 = 1 RL = 1 # of cells = 41 = 4 RL = 2 # of cells = 4 = 16 2 RL = 3 # of cells = 43 = 64 RL = 2 # of cells = 42 = 16 RL = 3 # of cells = 43 = 64 (a) (b) Figure 10. ONESTEP ADAPTIVE CELL REFINEMENT 189 Start Initial Configuration BNM simulation Multilevel cell refinement HBNM residuals Final Configuration Redefinition of cells. LLC .
5 0 0. The hypersingular residual (from equations (9. The cell structure obtained in ONEstep is shown in Figure 10.12 and (10.5 x 0 _ 0.1)) by imposing the exact solution in equation (10. The boundary value problem is solved using the BNM (equation (9.5 _1 1 0.3.12(b). p = m. which consists of 438 cells with one node per cell.1 Initial Cell Conﬁguration # 1 (54 Surface Cells) The ONEstep multilevel strategy is implemented on the cube of Figure 10.5 0 _ 1 0. Table 10. (b) Adapted conﬁguration with 438 cells using γ = 0. LLC . 10. Errors are again estimated by means of hypersingular residuals.12: ONEstep multilevel cell reﬁnement for the cube problem.5 _1 _1 1 0.5 0 0.15 (from ([28]) © 2005 by Taylor & Francis Group.5 0 0. This idea of ONEstep reﬁnement is applied below to the cube problem of Section 10.8)) is obtained and then the multilevel reﬁnement procedure is carried out using γ = 0.4.5 x y y (a) (b) Figure 10. The second condition in equation (10.21) is enforced because cell structure coarsening is not considered in this work.18) as Dirichlet boundary conditions.12(b). ADAPTIVITY FOR 3D POTENTIAL THEORY RL = log η p log 2 for η ≥ 1 for η < 1 0 (10. which consists of 54 cells with one node per cell.190 CHAPTER 10.12(a) and ending with the conﬁguration of Figure 10.5 0 _ 0. (a) Initial conﬁguration # 1 with 54 cells. 1 0.12(a).15. For the interpolation procedure used in this work.5 _ z _ z _ 1 1 0.21) where p is order of the interpolating function.5 0. This table shows that ¯(∂u/∂n) and the residual consistently decrease with reﬁnement.3 shows a comparison of the results from the ONEstep multilevel reﬁnement scheme starting with the conﬁguration of Figure 10.5 _ _1 _1 1 0.
4.5 x 0.1450 % 0.2618E01 191 Table 10.13(a)).5 1 1 0.4209 % 7.5 y 1 1 0.13(b).5 0 1 z (a) z (b) Figure 10.10.18) as the exact solution.6911 % 2.5 0 0. ONESTEP ADAPTIVE CELL REFINEMENT Output parameters Number of cells x = ±1 y = ±1 z = ±1 All faces Average residual r ¯ Maximum residual rmax Initial 54 1. rmax for the initial conﬁguration (Fig¯ ure 10.5 x 0.3: ¯(∂u/∂n) and residuals r.5197E01 Final 438 0. and the hypersingular residual is obtained by means of equations (9.5 0 0.0403 % 0.2 for the progressive adaptive reﬁnement.1)) is used to solve the boundary value problem using equation (10.5 1 1 0.12(b)) obtained by the multilevel reﬁnement strategy with γ = 0.13: ONEstep multilevel cell reﬁnement for the cube problem (a) initial conﬁguration # 2 with 96 cells (b) adapted conﬁguration with 1764 cells using γ = 0.5 0 0.5 0 0.15.8).5 y 1 1 0.15 (from ([28]) Table 10.4.5 0 1 1 0.2366E01 0.15 (from [28]) 10. Multilevel reﬁnement is carried out using γ = 0. which consists of 1764 cells with one node per cell.0339 % 0.4 summarizes the results for ¯(∂u/∂n) and the residual for the multilevel reﬁnement strategy starting from the initial conﬁguration # 2 of 96 cells (see Figure 10.0343 % 0.2 Initial Cell Conﬁguration # 2 (96 Surface Cells) The BNM (equation (9.0411 % 0. LLC .6911 % 7. 1 0. The cell structure obtained in ONEstep is shown in Figure 10. which includes the peculiarity © 2005 by Taylor & Francis Group. Qualitatively.12) and (10.12(a)). and the ﬁnal conﬁguration (Figure 10.4615E03 0. these results are analogous to those of Table 10.
LLC . Moreover.13(b)) obtained by the multilevel reﬁnement strategy with γ = 0.0696 % 0.2400 % 0.2400 % 0.0654 % 1.4: ¯(∂u/∂n) and residual for the initial conﬁguration (Figure 10.4963E02 0.13(a)). © 2005 by Taylor & Francis Group. for further explanations.15 (from [28]) observed on the x = ±1 faces of the cube.0759 % 1.4.2 also hold for explaining the results of Table 10. the remarks concerning Table 10.1034 % 0.1535E01 Table 10.3370E03 0. and the ﬁnal conﬁguration (Figure 10. the reader is referred to Section 10. ADAPTIVITY FOR 3D POTENTIAL THEORY Output parameters Number of cells x = ±1 y = ±1 z = ±1 All faces Average residual (¯) r Maximum residual (rmax ) Initial 96 0. Therefore.192 CHAPTER 10.1899 % 0.1169 % 0.3.1829E01 Final 1764 0.
real positive constants c1 and c2 exist such that: c1 r ≤ ≤ c2 r (11. under certain favorable conditions.27)) as. Thus. 3 (11.1. k = 1. k = 1. ∗ rij ≡ residual(σij ) = LHBN M (u∗ .1) ∗ with the numerical solution (u∗ .4) where r is some scalar measure of a hypersingular residual and is a scalar measure of the exact local error. k = 1. 3 (11. τk ). 11. τk ) = 0 . Also. 2. Please see the introduction to Chapter 10.24)) for elasticity be written in operator form as: LBN M (uk . τk ) k . It should be 193 © 2005 by Taylor & Francis Group. 2. 3 (11. as applied to 3D linear elasticity. 2.1 Hypersingular and Singular Residuals The Hypersingular Residual Let the BNM (equation (9. τk ) = 0 . a hypersingular residual is expected to provide a good estimate of the local error on a boundary element.27)) is k written in operator form as: LHBN M (uk .1(a). LLC .1 11. It has been proved in [96] and [127] for the BIE that.3) This idea is illustrated in Figure 10. the stress residual is deﬁned from the stress HBNM (equation (9.Chapter 11 ADAPTIVITY FOR 3D LINEAR ELASTICITY The subject of this chapter is error analysis and adaptivity with the BNM.2) This time. the HBNM (equation (9.
10) are used to compute the hypersingular residual and exact error.24) o ri ≡ residual(ui ) = LBN M (uo . τk ) = 0 k .5) and from equation (9. © 2005 by Taylor & Francis Group.2 The Singular Residual From equation (9. the ones proposed in this chapter. The goal is to eﬃciently develop a ﬁnal cell conﬁguration which leads to a reliable numerical solution.8) and (11. in as simple a manner as possible.8) = σij (exact) − σij (numerical) (11.2.1 Local Residuals and Errors .10) Equations (11. τk ) k . 2.3) : rij = residual(σij ) A scalar residual measure is deﬁned as: r = rij rij The exact local error in stress is deﬁned as: ij (11. while the current work presents a uniﬁed residual formulation. for problems in linear elasticity. LLC .2.7) (11. ADAPTIVITY FOR 3D LINEAR ELASTICITY mentioned here that the deﬁnitions of the residuals used in [96] and [127] are analogous to. in the stress σij .Hypersingular Residual Approach From equation (11.1(b).2 Error Estimation and Adaptive Strategy The h−version iterative selfadaptive procedure employed in this work is presented in the ﬂowchart .6) This idea is illustrated in Figure 10. k = 1.9) and the corresponding scalar measure is deﬁned as: = ij ij (11. but diﬀerent in detail from.28): o LHBN M (uo . [96] in the sense that Dirichlet. 3 (11. 3 (11. Neumann and mixed problems require separate prescriptions in [96]. at each node. The above formulation for singular and hypersingular residuals is a generalization of the earlier work by Menon et al. 11.194 CHAPTER 11. respectively. 11. 11.1. k = 1.Figure 10. 2.
respectively.2. evaluated at nodes. 11. The scalar residual measures. the roles of the BNM and HBNM are reversed.3.Singular Residual Approach From equation (11. 11.2.11) (11.1 of Chapter 7.6): ri = residual(ui ) so that a scalar residual measure is: r = r i ri The exact local error in ui is deﬁned as: i (11.2.11.4 in Chapter 10.e. at each node. This time. The scalar residual measure at this centroidal node is used as an error estimator for that cell. In all the adaptivity examples presented in this chapter. the HBNM (9.14) are used to obtain the singular residual and exact error. in the displacement ui . one node is used for each cell and is placed at its centroid.12) and (11. PROGRESSIVELY ADAPTIVE SOLUTIONS: PULLING A ROD 195 11.14) Equations (11. i.28) is used to solve the boundary value problem while the singular residuals are obtained from the standard BNM (9.24) (see Figure 10.2 Local Residuals and Errors . A comparison of the residual r and exact error demonstrates the eﬀectiveness of residuals as error estimates. deﬁned above.12) = ui (exact) − ui (numerical) (11.3 Cell Reﬁnement Global Error Estimation and Stopping Criterion The algorithms used here are the same as those described in Sections 10.1(a)).2.3 Progressively Adaptive Solutions: Pulling a Rod The problem under consideration here is the stretching of a cylindrical elastic rod clamped at one end (see Figure 11.3 and 10. © 2005 by Taylor & Francis Group.13) with a corresponding scalar measure: = i i (11.3. LLC .1(b)). This same problem has been addressed with the BCM in Section 7. are used as error estimators. for elasticity problems.
ADAPTIVITY FOR 3D LINEAR ELASTICITY z σ0 L x D (a) 2.5 _ _ 3 2 0 θ (radians) 1 2 3 2 _ 2.0 (b) and (c) initial cell conﬁguration with 144 cells (one node per cell) .5 2 1.0.(b) clamped and loaded faces (c) curved surface of the cylindrical rod mapped onto the (z. ν = 0. LLC .5 0 0. E = 1.25.196 CHAPTER 11. θ) plane (from [28]) © 2005 by Taylor & Francis Group.0. σ0 = 1. D = 4.1: Stretching of a short clamped cylindrical rod by a uniform tensile load (a) physical situation : L = 2.5 1 0.0.5 0 _ _ _1 y _ 1 z _ 1.5 _ 2 1.5 2 _ 1 0 x 1 2 (b) (c) Figure 11.5 1 0.
1(b). Figure 11. (11. However. This is to be expected considering the physical nature of the problem at hand which has a singularity on the bounding circle of the clamped face [37.” according to the scheme of Figure 10. the boxes for the “BNM simulation” and “HBNM residuals” are replaced by “HBNM simulation” and “BNM residuals.28). 11. and D = 4.1(b) and (c) show the initial cell conﬁguration on the clamped and loaded faces and on the curved surface of the rod.0.2.6). Since the singular residual is higher on the clamped and curved faces.11. than on the loaded face. The residual has been normalized with respect to its maximum value. LLC .0.8 Singular residual 0.0.3.2 Adaptivity Results The adaptive strategy is carried out according to the ﬂow chart of Figure 10.1 Initial Cell Conﬁguration The geometric and material parameters chosen are: E = 1. Upon obtaining the solution to the boundary value problem. σ0 = 1. ν = 0. The boundary value problem is solved by prescribing tractions on the top face of the cylinder. Figures 11.0.3.25.2419 × 10−2 (from [28]) 11. PROGRESSIVELY ADAPTIVE SOLUTIONS: PULLING A ROD 197 Traction ( σ ) 0 1 Clamped Traction free 0.6 0. and the curved surface traction free. 137]. the singular residual is obtained at each node. It can be seen that the residual is considerably higher on the clamped face and on the curved surface near the clamped face. The curved © 2005 by Taylor & Francis Group. with the bottom surface completely clamped.4 0. rmax = 0.1 (b) and (c). and (11.2: Singular residual for the initial conﬁguration of 144 cells on the cylindrical rod of Figure 11.2 0 20 40 60 80 Node number 100 120 140 Figure 11. most of the subdivision of cells occurs on those faces.3.12)) obtained for the initial cell conﬁguration (144 cells).2 shows the singular residual (from equations (9. L = 2.
First. Nevertheless. 11. The present study also employs singular residuals (equation (9. Note that. is not required at all in the present meshless methods (BNM and HBNM).1(b) and 11. a signiﬁcant mismatch in the number of subdivisions is created at every adaptive step. ADAPTIVITY FOR 3D LINEAR ELASTICITY surface of the cylinder near the clamped surface is reﬁned due to the singularity at the edge of the clamped face.1(b).3).3(c). for the initial conﬁguration (Figure 11.1(a) and (b) are given in Figure 11. The results for the multilevel reﬁnement for the original problem of Figure 11.1(b) and (c).4(c). and 11.5(a) show that the loaded face is not reﬁned at all and remains as in the initial conﬁguration (Figure 11. 11. 11. and such freedom in modeling is expected to be especially advantageous in analyzing problems with complicated geometry. 11.5(b) indicates that cell reﬁnement only takes place near the edge of the face.6(a)).5. such gradation. On the clamped face. The resulting reﬁned cell conﬁgurations are shown in Figures 11. However.15) and starting from the initial conﬁguration of Figures 11.1(b)). the number of subdivisions along the edge of the top and bottom faces is the same as the number of subdivisions along the edge of the curved surface (12 subdivisions).24)) for error estimation (see Section 11. while it is not reﬁned at all in the adapted conﬁgurations shown in Figures 11.4 is demonstrated here for the same problem as above (Section 11. The initial conﬁguration is the same as that of Figure 11. which is essential for meshbased methods.25 in equation (10.198 CHAPTER 11.6).3) rather than hypersingular residuals (equation (9. a comparison of Figures 11. the top face (the loaded face) is NOT reﬁned at all and so the cell structure on that face remains as shown in Figure 11. when adaptivity is carried out.5(a). Three steps of adaptivity are pursued using γ = 1. Second. 11.4(a) and 11.3(b).1(c).1(b) and (c). two main diﬀerences are noticeable. © 2005 by Taylor & Francis Group.5). Figures 11.311. and 11.11. This does not present any problem for the meshless method.27)). the loaded face is reﬁned (a little) here (see Figures 11. 11.5(c) show the progressive reﬁnement on the curved surface of the cylinder.1 (b) and (c)). Comparing these results with the ones of Section 11.3 (Figures 11. However. as expected. LLC .3 . One can observe that reﬁnement primarily occurs along the curved surface near the clamped edge of the cylindrical rod. the multilevel cell reﬁnement does not allow the smooth cell gradation which occurs in progressively adapted cell conﬁgurations (see Figures 11.4(a) and 11.3(a).1(b). 11. This is similar to the study carried out for the same problem by means of iterative adaptive cell reﬁnement. As expected. which is the region where gradients in stresses are largest.6. Figures 11.4 OneStep Adaptive Cell Reﬁnement The ONEstep procedure outlined in Section 10.5 and 11. one veriﬁes that the overall trends are quite similar in both situations.4(b).3(a). However.
5 1 0.4.5 _ _ 1.5 0 0.5 _ y _ 1 y _ 1 1. ONESTEP ADAPTIVE CELL REFINEMENT 199 2.25.5 1 0.5 2 _ 1 0 x 1 2 (a) (b) 2 1.5 0 _ _ _ z 3 2 1 0 θ (radians) 1 2 3 (c) Figure 11.11. (a) loaded face (b) clamped face (c) curved surface of the rod (from [28]) © 2005 by Taylor & Francis Group.5 0 0.3: Short clamped cylindrical rod Step 1 : Adapted conﬁguration consisting of 228 cells obtained with γ = 1.5 _ _ 2. LLC .5 Clamped face 2 1.5 Loaded face 2 1.5 1 0 x 1 2 2.5 _ _ _ 2 2 _ 2 _ _ 2.5 1 0.
5 2 _ 1 0 x 1 2 2.4: Short clamped cylindrical rod Step 2 : Adapted conﬁguration consisting of 324 cells obtained with γ = 1.25.5 Clamped face 2 1.5 0 _ _ _ z 3 2 1 0 θ (radians) 1 2 3 (c) Figure 11.5 0 0.5 1 0.5 Loaded face 2 1.5 _1 _ 1.5 _2 _ 2.5 _2 _ _ 2.5 1 0.200 CHAPTER 11. (a) loaded face (b) clamped face (c) curved surface of the rod (from [28]) © 2005 by Taylor & Francis Group. ADAPTIVITY FOR 3D LINEAR ELASTICITY 2.5 _ 2 _ 1 0 x 1 2 (a) (b) 2 1. LLC .5 _ y _ y _ 1 _ 1.5 1 0.5 0 0.
5 0 _ _ _ z 3 2 1 0 1 θ (radians) 2 3 (c) Figure 11.11.5 _ 2 _ _ 2 _ 2.5 0 0.25.5: Short clamped cylindrical rod Step 3 : Adapted conﬁguration consisting of 576 cells obtained with γ = 1.5 1 0.5 0 0.5 _ 2. ONESTEP ADAPTIVE CELL REFINEMENT 201 2. LLC .5 _ _ y _ 1 y _ 1 _ 1.5 2 _1 0 x 1 2 2.5 Loaded face 2 1.5 2 _ 1 0 x 1 2 (a) (b) 2 1.4. (a) loaded face (b) clamped face (c) curved surface of rod (from [28]) © 2005 by Taylor & Francis Group.5 _ _ 1.5 1 0.5 1 0.5 Clamped face 2 1.
5 _1 _ y _ 1 1.5 _ 2 _ _ 2 _ _ 2.1(b) and (c) (from [28]) © 2005 by Taylor & Francis Group.5 2 _ 1 0 x 1 2 (a) (b) 2 1.5 0 _ _ z 3 2 _1 0 θ (radians) 1 2 3 (c) Figure 11.5 1 0.5 1 0.6: ONEstep multilevel cell reﬁnement for short clamped cylinder: (a) loaded face (b) clamped face (c) curved surface of cylinder.5 Clamped face 2 1. LLC . The initial cell conﬁguration is shown in Figure 11.5 2 _1 0 x 1 2 2.5 Loaded face 2 1.202 CHAPTER 11.5 _ y _ 1.5 _ _ 2.5 0 0.5 1 0. ADAPTIVITY FOR 3D LINEAR ELASTICITY 2.5 0 0.
(2002). Elementfree Galerkin methods.A. Meshless methods: An overview and recent developments.. Engineering Analysis with Boundary Elements. S. Lu. S. and Li. Problems and Methods of Optimal Structural Design. Fleming. Plenum. Electrostatic BEM for MEMS with thin beams. M. [6] Bao. [7] Bao. Encino. M. P. ASME Journal of Applied Mechanics 71:551559. (2005). and Krysl. (1997). S. Mukherjee. N. Y. Berkshire. T.. 203 © 2005 by Taylor & Francis Group. G. New York. An eﬃcient numerical technique for electromechanical simulation of complicated microelectromechanical structures.R. Z. [8] Bao. (1996). N. Z. (1992).. Finite cloud method: a true meshless technique based on a ﬁxed reproducing kernel approximation. Y. (2004). and White. McGraw Hill Europe. N. J.V. P. N. Sensors and Actuators A 58:111. and Mukherjee. [9] Becker. Roman. 28:14271435. Nonlinear vibrations of beams strings plates and membranes without initial tension. The Boundary Element Method in Engineering.. S. Singapore.Y. McGraw Hill International. Maidenhead. A.. and Shen. Krongauz. (2004). [2] Aluru.N. [11] Belytschko. Organ. The Boundary Element Methods in Engineering. [5] Banichuk. Electrostatic BEM for MEMS with thin conducting plates and shells. International Journal for Numerical Methods in Engineering 50:23732410. (1994). CA.R.. Computer Methods in Applied Mechanics and Engineering 139:347. and Mukherjee. S. LLC . [4] Banerjee. International Journal for Numerical Methods in Engineering 37:229256. (1994). UK. (1983).K. The Meshless Local PetrovGalerkin (MLPG) Method. (2001). [10] Belytschko. T. Tech Science Press. Z.Bibliography [1] Aluru. and Gu. [3] Atluri. and Aubry. L. In press. Communications in Numerical Methods in Engineering. D.
Eﬃciency of a posteriori BEM error estimates for ﬁrst kind integral equations on quasiuniform meshes. (2001). Computer Methods in Applied Mechanics and Engineering 190:43194337. [21] Carstensen. Estep. New York. UK. Boundary Elements: An Introductory Course. and Stephan. M. and Mukherjee. S. e e [18] Carstensen. Evaluation of gradients on the boundary using fully regularized hypersingular boundary integral equations. (1996). Marcel Dekker.K. Dissertation. [14] Bonnet.K. [22] Chandra. Wiley. (1995). C. [24] Chati. C. M.P. E. Adaptive boundary element methods and adaptive ﬁnite element methods and boundary element coupling. [16] Brebbia. (1995). [23] Chati. S. (1997). NY. [15] Bonnet. 4758.D. Boundary Integral Equation Methods for Solids and Fluids. C. C. Nicaise eds. H. and Mukherjee. © 2005 by Taylor & Francis Group. M. C. Shape sensitivity analysis and shape optimization in planar elasticity using the elementfree Galerkin method. (1999).P.D. S. NY. (1999).. [17] Bui. Mathematics of Computation 65:6984. S. (1975). and Mukherjee. Meshless Standard and Hypersingular Boundary Node Method .. Meshless approach to shape optimization of linear thermoelastic solids.Applications in ThreeDimensional Potential Theory and Linear Elasticity. Ithaca. (1995). A. Comptes Rendus de l’Acad´mie des Sciences Paris S´rie A 280:11571160. D. and McGraw Hill. and Xiao. LLC . Boundary Element Methods in Manufacturing. M. [20] Carstensen. Computational Mechanics Publications. and Dominguez. J. and Stephan. Southampton. F. M. [19] Carstensen. Computation of energy release rate using material diﬀerentiation of elastic BIE for 3D elastic fracture. Chichester. Mathematics of Computation 64:483500. [13] Bobaru. and Mukherjee. (1992). A posteriori error estimates for boundary element methods. Boundary Value Problems and Integral Equations in Nonsmooth Domains. Ph. F. International Journal for Numerical Methods in Engineering 53:765796. Computational Mechanics 15:372383. M. Oxford University Press. UK. 2nd edition. Engineering Analysis with Boundary Elements 15:137149. hadaptive boundary element schemes.A. (1995).204 BIBLIOGRAPHY [12] Bobaru. Costabel. E. Application des potentiels ´lastiques ` l’´tude des ﬁssures e a e planes de forme arbitrare en milieu tridimensionnel. Dauge and S. (1995). Acta Mechanica 135:4155. H. (2002). Cornell University.
The meshless hypersingular boundary node method for threedimensional potential theory and linear elasticity problems. Boundary Element Methods.. Review of dual boundary element methods with emphasis on hypersingular integrals and divergent series. J. S. Computers and Mathematics with Applications 43:379391. (1999).T. [26] Chati.. Number 167 in Lecture Notes in Pure and Applied Mathematics. Ciarlet and J. Evaluation of J integrals and stress intensity factors in a 2D quadratic boundary contour method. S. [28] Chati. J. Finite Element Methods (Part 1) .H. (2000). 2 of Handbook of Numerical Analysis. Lions eds. Mukherjee. The meshless standard and hypersingular boundary node methods . Fiszdon and K. (1999). P. The boundary node method for threedimensional linear elasticity. [32] Chen. (1992).H. G. S. (1999). M.. S.K. (1991). [29] Chen. Elsevier Science Publishers. M. L.. and Kwak. Wilmanski eds. 15.Vol. and Stephan.X.. ASME Applied Mechanics Reviews 52:1733. [34] Ciarlet. M. Y. and Zheng. M. M. International Journal for Numerical Methods in Engineering 47:15231547. M. and Tanaka.K. (1995). [35] Costabel. [31] Chen. The boundary node method for threedimensional problems in potential theory. Academic Press. G. International Journal for Numerical Methods in Engineering 50:22332269. W. LLC . Boundary integral equation method for shape optimization of elastic structures. and Nicaise. boundaryonly RBF technique. [36] Costabel. (1988). W. eds. Mukherjee. G.G.X.K. G. and Hong. W. and Paulino. H..M. and Mukherjee. Engineering Analysis with Boundary Elements 25:639653. Communications in Numerical Methods in Engineering 15:91100.applications to error estimation and adaptivity in threedimensional problems. [27] Chati. [33] Choi. International Journal for Numerical Methods in Engineering 26:15791595. (2001). Leng. S. Mukherjee. J. (2001). X. Tang. W. Boundary Value Problems and Integral Equations in Nonsmooth Domains. Mathematical Models and Methods in Mechanics Vol. Marcel Dekker Publications. M. © 2005 by Taylor & Francis Group. Zhou. Basic error estimates for elliptic problems.J. and Mukherjee. Banach Center Publications..Y. and Paulino. Dauge. International Journal for Numerical Methods in Engineering 46:11631184. E. [30] Chen.. M.H. P. integrationfree. S. S. and Zhou.K. (2002)..BIBLIOGRAPHY 205 [25] Chati. B. A meshless. (1985).K. Boundary integral equations for mixed boundary value problems in polygonal domains and Galerkin approximation.
The method of ﬁnite spheres.C. (1999). (1988). Cambridge University Press. C.206 BIBLIOGRAPHY [37] Cruse.. Hansbo. Heidelberg.. M. Acta Numerica Chapter 3. Hsiao. International Journal of Solids and Structures 37:71617183. (1996). C.. J. Estep. Asymptotic and a posteriori error estimates for boundary element solutions of hypersingular integral equations. S. (1995).T. T. International Journal for Numerical Methods in Engineering 39:32733304. Iserles ed. [48] Ghosh. K¨gl. T. and Lapina. Boundary Element Analysis in Computational Fracture Mechanics. Dordrecht. (1996). and Oden. M. LLC .. An hp adaptive method using clouds. [39] Cruse.R. and Belytschko. Englewood Cliﬀs. D. (2003). [50] Glushkov. Glushkova.M. K.E. Kluwer. International Journal of Solids and Structures 36:11051128.A. S. [46] Fung. Acta Mechanica 67:107119. L. S. NJ. C. 3D coupled electromechanics for MEMS . (1969). (1987). [45] Feistaur. N. Proceedings of IEEE MEMS pp. and Richardson. (2000). [40] De. R. N.M. 105158. [49] Gilbert. R.D.an hp meshless method. J. Germany. M. (1996). © 2005 by Taylor & Francis Group. Numerical Methods in Partial Diﬀerential Equations 12:673705. A new boundary element method formulation for three dimensional problems in linear elasticity. G. [47] Gaul. and Mukherjee.. Introduction to adaptive methods for partial diﬀerential equations.122127. E. A. YC.J. J. K.applications of CoSolve EM. J. [38] Cruse. [43] Duarte. Boundary Element Methods o for Engineers and Scientists. Modeling fracture in o MindlinReissner plates with the extended ﬁnite element method.A. (2000). UK. The Netherlands. 3D elastic stress analysis at polyhedral corner points.. Computational Mechanics 25:329345.D.T. and Kleinman. and Johnson. Springer Verlag. M¨es. (1965).A. P. [44] Eriksson.. [41] Dolbow. [42] Duarte. (1996). and Senturia.A. Legtenberg. SIAM Journal of Numerical Analysis 32:666685. Nonsingular Somigliana stress identities in elasticity. O. (1995). Foundations of Solid Mechanics. International Journal of Solids and Structures 5:12591274.. Prentice Hall. and Bathe. Hp clouds . T.A. Numerical solutions in threedimensional elastostatics. and Wagner. N. Computer Methods in Applied Mechanics and Engineering 139:237262. T. J. and Oden.
Hypersingular integrals in boundary element fracture analysis. F. J. and Paulino.T.H. (1994). Computational Mechanics 20:2633. NJ. G. A “pure” boundary node method in potential theory. K. [59] Guidera. H. [61] Harrington. (1998). © 2005 by Taylor & Francis Group. L. New York. and Kane. and Lardner. (1997). [55] Gray. Communications in Numerical Methods in Engineering 18:411427. and Mukherjee. (2002). (1994).. Balakrishna.BIBLIOGRAPHY 207 [51] Goldberg. Introduction to Boundary Elements : Theory and Applications. L.F. A numerical quadrature for nearly singular boundary element integrals. (1997). Introduction to Maple. Choi. Boundary element method for regions with thin internal cavities. A.J. G. Engineering Analysis with Boundary Elements 13:169179. [64] Hayami.. Design Sensitivity Analysis of Structural Systems.K.H. LLC . M. J.H. R. Symmetric Galerkin fracture analysis.I. R. and Matsumoto. H.F. (1995). Hypersingular formulation for boundary stress evaluation. R. L. [62] Hartmann.H. L. Symmetric Galerkin boundary integral fracture analysis for plane orthotropic elasticity. [60] Guiggiani. International Journal for Numerical Methods in Engineering 29:11351158.J. [52] Gowrishankar. E. IEEE Press.W.J.J. and Bowman. Piscataway. [58] Gray. and Ingraﬀea. Academic Press. S. V. International Journal for Numerical Methods in Engineering 40:30853101. (1989).. Boundary Element Communications 8:230238.J. K.A. [63] Haug. Journal of Elasticity 5:5973. A. Engineering Analysis with Boundary Elements 13:143154. G. and Komkov.R. Engineering Analysis with Boundary Elements 6:180184. New York.J.J. M. [65] Heck. Symmetric Galerkin boundary integral formulation for interface and multizone problems. New York. SIAM Journal of Applied Mathematics 58:428455. Berlin. Field Computation by Moment Methods. (1990). [56] Gray. (1993). and Paulino. Springer Verlag. (1975). Engineering Analysis with Boundary Elements 15:103109. [54] Gray. Superconvergence and the use of the residual as an error estimator in the BEM . (1993). (1989). L. Martha. C. [57] Gray. Penny shaped cracks. (1998). Theoretical Development. and Paulino. [53] Gray. Crack tip interpolation revisited. L. Springer Verlag. (1986). L.
International Journal for Numerical Methods in Engineering 40:29532974. and Saigal. [75] Krishnasamy. Banerjee and S. © 2005 by Taylor & Francis Group.R. (2003).An Introduction. and Kamiya. R.. (1989).C. P. Kim. F. Schmerr. Curve and Surface Fitting . W. S. interpretation. L. F. Recent studies on adaptive boundary element methods. [77] Kulkarni. (1994). Y. Sensors and Actuators A 103:130134. K. T. S. Prentice Hall. S.Linear Static and Dynamic Finite Element Analysis. S.. An improved element free Galerkin formulation. (1990). Reading. [70] Kita.. LLC . Developments in Boundary Element Methods7 .. [74] Krishnamoorthy. and Rudolphi.C. Micromachined piezoelectric membrane acoustic device. Boundary Element Analysis in Engineering Continuum Mechanics. [68] Kane. London.H. 207252. Rudolphi. T. ed.H.. S. Hypersingular boundary integral equations: their occurrence. and Mukherjee. (2000). pp.208 BIBLIOGRAPHY [66] Hughes.K. Computational Mechanics 32:240249. Rizzo. (1994). T.J.J.J. C. S. and Salkauskas.. The Finite Element Method . Englewood Cliﬀs. Kobayashi eds. Lee. MA. G.. [71] Ko. [72] Kothnur. and Mukherjee.J.S. Twodimensional linear elasticity by the boundary node method. (1997). NY. Elsevier Applied Science. AddisonWesley. regularization and computation. J. Computers and Structures 48:121133. Application of an accelerated boundarybased meshfree method to twodimensional problems in potential theory. Hypersingular boundary integral equations : some applications in acoustic and elastic wave scattering. K. and Rizzo. [69] Kaplan.. Advanced Calculus 3rd. (1990). E. N. (2003). Mukherjee. S. (1993). Choi. [67] Kaljevic. Telukunta.X. ASME Journal of Applied Mechanics 57:404414. G. Y. (1984). V.R. (1992). (1999). [78] Lancaster. London. Mineola.. Dover. Adaptive mesh reﬁnement for twodimensional ﬁnite element stress analysis. Linear Integral Equations.S. [76] Krishnasamy.S. [73] Kress. International Journal of Solids and Structures 36:11291147. S. Academic Press. Advances in Engineering Software 19:2132.W. and Umesh. and Kim.J. Springer Verlag. P. NJ. I. S.R.
and Gray. Y. and Chang.BIBLIOGRAPHY 209 [79] Li.K. S. Archives of Computational Methods in Engineering 3. [85] Liu. and Aluru.R. N. Boundary Elements XV. International Journal of Solids and Structures 35:44274436. (2003).Moving beyond the Finite Element Method. Barking.. [83] Liu. Y.K. N. Engineering Analysis with Boundary Elements 14:315323. [81] Liapis. (1996). Computer Methods in Applied Mechanics and Engineering 107:131144. [87] Liu. A boundary cloud method with a cloudbycloud polynomial basis.. Cornell University. Elimination of rigid body modes from discretized boundary integral equations. (1993). (2002)..J.R. W. (1992). Overview and applications of the reproducing kernel particle methods. W. NY. G.J.J. C. Brebbia and J. (1991).D.J..T.D. [80] Li. © 2005 by Taylor & Francis Group. Ph. UK. Engineering Analysis with Boundary Elements 27:5771. Computer Methods in Applied Mechanics and Engineering 139:91157.R.. and Rizzo. Ye. Jun.. Mesh Free Methods ..D. Ingraﬀea. A. [86] Liu. L. Numerical Methods for Hypersingular and NearSingular Boundary Integrals in Fracture Mechanics. A review of error estimation and adaptivity in the boundary element method. Zhang. [88] Lutz. Boca Raton. International Journal for Numerical Methods in Engineering 41:541558. Y.A. E.J. R. Pan. E. LLC .R. Chen. Computer Methods in Applied Mechanics and Engineering 191:23372370. and Aluru. Y. D. Belytschko. Rencis eds. G. [90] Lutz.J. Y. Computational Mechanics Publications. C. Use of ‘simple solutions’ for boundary integral methods in elasticity and fracture analysis. FL. and Mukherjee.T. [89] Lutz. Boundary cloud method: a combined scattered point/boundary integral approach for boundary only analysis. Hypersingular boundary integral equations for radiation and scattering of elastic waves in three dimensions. E. (1998). and Rizzo. Uras. T. (2002). Southampton and Elsevier. [84] Liu. Dissertation. (1995). F. R. 453468.A. Essex.. 380. Chen. (1993). S. CRC Press. Analysis of shelllike structures by the boundary element method based on 3D elasticity : formulation and veriﬁcation. C. Uras. [82] Liu. (1996).. W. F. and Chang. Nearly singular and hypersingular integrals in the boundary element method. S. G. International Journal for Numerical Methods in Engineering 35:17371751.. (1998). Ithaca. Generalized multiple scale reproducing kernel particle methods.J. C.A..
S. Dolbow. Engineering Analysis with Boundary Elements 24:767776. Surface variables and their sensitivities in threedimensional linear elasticity by the boundary contour method. (1996).A. T. [97] Mo¨s. [92] Martin. (2000). Hypersingular Error Estimates in Boundary Element Methods. X. Finite Elements in Analysis and Design 15:177188. A new formula for the Cmatrix in the Somigliana identity.A. J. and Mukherjee.X. NY. [99] Mukherjee. M. Y. [93] Martin. The hypersingular boundary contour method for threedimensional linear elasticity. and Rizzo.S. c [94] Manti˘. (1999). (2000). (1993).A. [96] Menon. and Belytschko... G. A ﬁnite element method e for crack growth without remeshing. London. T. and Mukherjee. Paulino. S. International Journal for Numerical Methods in Engineering 42:885906. [98] Mukherjee. Computer Methods in Applied Mechanics and Engineering 173:387402. Y. Cornell University.. Computational Methods in Applied Mechanics and Engineering 173:449473. Smoothnessrelaxation strategies for singular and hypersingular integral equations. X. Internal variables and their sensitivities in threedimensional linear elasticity by the boundary © 2005 by Taylor & Francis Group. (1993).a bibliography (19901993).. Rizzo.J. Finite parts of singular and hypersingular integrals with irregular boundary source points. (1998). J. S. G. and Mukherjee. F. S. (2000). and Mukherjee. International Journal for Numerical Methods in Engineering 46:131150. F. and Cruse. G. LLC . [102] Mukherjee. S. (1996).210 BIBLIOGRAPHY [91] Mackerele. International Journal of Solids and Structures 37:66236634. [103] Mukherjee. [95] Menon. Journal of Elasticity 33:191201. Mesh generation and reﬁnement for FEM and BEM .H. ASME Journal of Applied Mechanics 65:300309. P.J. Hypersingular integrals : how smooth must the density be ? International Journal for Numerical methods in Engineering 39:687704. V. X.X. (1998). S. (1999). S. P. (1999). CPV and HFP integrals and their applications in the boundary element method. Shi. Y. Analysis of hypersingular residual error estimates in boundary element methods for potential problems. [100] Mukherjee. Boundary Element Methods in Creep and Fracture. Thesis. Shi. (1982). [101] Mukherjee. Ithaca.. N. Elsevier.
LLC . K. Chati. S. Mathematics and Mechanics of Solids 6:4764. [111] Mukherjee. S. and White. Evaluation of nearly singular integrals in boundary element contour and node methods for threedimensional linear elasticity. Engineering Analysis with Boundary Elements 23:8996. The boundary node method for potential problems. and Mukherjee. Computational Mechanics 19:264270. A. A mapping method for numerical evaluation of twodimensional integrals with 1/r singularity.. [114] Nagarajan.X. Engineering Analysis with Boundary Elements 26:839844. [105] Mukherjee. and Mukherjee. [109] Mukherjee. S. [107] Mukherjee.X.. (1991). S. (1994). (1997). [104] Mukherjee. A. (1997). (2000). S.D. A. M. Shi. Engineering Analysis with Boundary Elements 20:3544. (1993). J. and Mukherjee. White. Computational Mechanics 12:1926. X..X. K. (1997). S. S. Kim. and Mukherjee.. Computer Methods in Applied Mechanics and Engineering 187:289306. and Senturia. MA. E. On boundary integral equations for cracked and for thin bodies. S. Y. and Mukherjee. FastCap User’s Guide. (2001). International Journal of Solids and Structures 37:76337654. Lutz. International Journal for Numerical Methods in Engineering 40:797815. (1992). S. A novel boundary element method for linear elasticity with no numerical integration for 2D © 2005 by Taylor & Francis Group. (1999). Thermoelastic fracture mechanics with regularized hypersingular boundary integral equations. K. IEEE Transactions on ComputerAided Design of Integrated Circuits and Systems 10:14471459. and Mukherjee. S. and Nagarajan. Shah. FastCap: a multipole accelerated 3D capacitance extraction program. (2002). The boundary contour method for threedimensional linear elasticity with a new quadratic boundary element. Error analysis and adaptivity in threedimensional linear elasticity by the usual and hypersingular boundary contour method. S. [113] Nabors. [112] Nabors.X. International Journal of Solids and Structures 38:161178. Y. On boundary conditions in the elementfree Galerkin method..K. Mukherjee. Y.BIBLIOGRAPHY 211 contour method. [115] Nagarajan. J. Regularization of hypersingular boundary integral equations : a new approach for axisymmetric elasticity.. [110] Mukherjee. X. S. Y.. Y.X. (2001). and Shi. Cambridge. [108] Mukherjee. MIT. [106] Mukherjee.
(2001). A methodology for adaptive ﬁnite element analysis: towards an integrated computational environment. G. (1999). [120] Oden. F. E. Boundary Element Method : Fundamentals ıs. Menon. J. International Journal for Numerical Methods in Engineering 39:20052029. O.A. Nodal sensitivities as error estimates in computational mechanics. ASME Journal of Applied Mechanics 61:264269. (1999). and Villon. Novel Formulations of the Boundary Element Method for Fracture Mechanics and Error Estimation..M. [125] Paulino. (1997).H.J.F. Engineering Analysis with Boundary Elements 23:97105. K. V. [124] Paulino. Engineering Analysis with Boundary Elements 25:523534. and Lutz. ASME Journal of Applied Mechanics 63:278286.H. C. A Galerkin boundary contour method for twodimensional linear elasticity.. and Zarkian. and Gray.. and Martha. Menezes. Yoshida. N.. The boundary contour method for threedimensional linear elasticity. [126] Paulino. B. n and Applications.D. Mukherjee.D.. Ithaca. Galerkin residuals for adaptive symmetricGalerkin boundary element methods. [119] Novati. (1999). Cavalcante Neto. Cornell University. (1998). NY.. P.M. Computational Mechanics 23:5362. and Zienkiewicz. Generalizing the ﬁnite element method: diﬀuse approximation and diﬀuse elements. S.. G. G. L. Ph. Oxford University Press. Gray. Computational Mechanics 23:361388..T. P. A fast multipole boundary integral equation method for crack problems in 3D. UK. (1996). L. G. [123] Paulino. G. [116] Nagarajan. G. and Springhetti.H. (1997). Computer Methods in Applied Mechanics and Engineering 153:117126. J. Dissertation. Touzot. Duarte. I. and Kobayashi. (1999). Shi. [122] Paulino. L. J. Hypersingular residuals .212 BIBLIOGRAPHY and line integrals for 3D problems. G.H. and Ca˜as. A new cloud based hp ﬁnite element method. (1992). Error estimation using hypersingular integrals in boundary element methods for linear elasticity. Mukherjee.H. Acta Mechanica 121:191213. ASCE Journal of Engineering Mechanics 125:575585. G. S. and Mukherjee S. S.F. and Ramesh. © 2005 by Taylor & Francis Group.. (1996).a new approach for error estimation in the boundary element method.J. LLC . G. [127] Paulino. R.H.B.C. [118] Nishimura.. (1995). A. [121] Par´ F. Computational Mechanics 10:307318. [117] Nayroles.
T. S. Y.S. T.V.R. S. National Congress of Applied Mechanics. A.. and Mukherjee. [136] Phan. and Phan. [137] Pickett. A. Archives of Mechanics 56:697724. (1994).R. Computational Mechanics 20:310319.V. A. (1944). (2002). LLC .R. S. A. Archive of Applied Mechanics 69:481489.BIBLIOGRAPHY 213 [128] Petryk. (1998).F. Detroit. G. and Mayer. [138] Poon.. Z. Computational Mechanics 28:425433. The pinwheel tilings of the plane. Engineering Analysis with Boundary Elements 23:195199. (1998).N. Mukherjee. [135] Phan. (1999).V. H. J. and Liu. (1999). and Phan. Use of simple solutions in regularizing hypersingular boundary integral equations in elastoplasticity.V. and Mayer.V.R. J.J. Mukherjee. The boundary contour method for twodimensional linear elasticity with quadratic boundary elements.R. T. Application of the Fourier method to the solution of certain boundary value problems in the theory of elasticity.R.R. [133] Phan. Transactions of the ASME (Journal of Applied Mechanics) 66:A176A182. International Journal for Numerical Methods in Engineering 42:13911407. International Journal of Solids and Structures 35:19811999.. Acta Mechanica 130:209225. Stresses. The boundary contour method for twodimensional Stokes ﬂow and incompressible elastic materials.V. [129] Phan. and Mayer. © 2005 by Taylor & Francis Group. The hypersingular boundary contour method for twodimensional linear elasticity. [139] Radin. ASME Journal of Applied Mechanics 65:3945.. Boundary contour analysis of thin ﬁlms and layered coatings. Annals of Mathematics 139:661702. S.J. H. A... A. Structural shape optimization system using the twodimensional boundary contour method. [134] Phan. Kaplan. Mukherjee. C. (1986). A. (1998). and Mr´z. [130] Phan. (1997). stress sensitivities and shape optimization in twodimensional linear elasticity by the boundary contour method. J. J. Gray. (1998). On design sensitivity analysis in linear elasticity by the boundary contour method. (2001). MI. A boundary contour formulation for design sensitivity analysis in twodimensional linear elasticity. A. Sixth U. Time derivatives of integrals and funco tionals deﬁned on varying volume and surface elements. Mukherjee. M.. Mukherjee. S. L. and Mayer. and Ahmad.V. [132] Phan.V.R. S. [131] Phan.N.
Mathematical and Computer Modeling 15:269278. The use of simple solutions in the regularization of hypersingular boundary integral equations. Dynamic analysis of microelectromechanical systems. and Mukherjee. Local boundary integral equation (LBIE) method for solving problems of elasticity with nonhomogeneous material properties. and Atluri. (1988). [147] Senturia. Johnson. (2000). [142] Roman. Design and fabrication of electrically actuated synthetic microjets. Adaptive h. Engineering Analysis with Boundary Elements 23:627637. M. N. E. NLPQL: A FORTRAN subroutine solving constrained nonlinear programming problems. P. p. [144] Saigal S. [146] Schittkowski. P. S. S. (1990).. © 2005 by Taylor & Francis Group.D. S. International Journal for Numerical Methods in Engineering 28:13351349.N. [151] Sladek. M. AIAA Journal 28:12031204. [149] Shi. Computational Mechanics 24:456462. (1999). (1991). S. Shulman. R.. Computers and Structures 56:769783.. V. S.. (1992). Acta Technica CSAV 31:523531. F. Shape optimization using the boundary element method with substructuring. Ramesh. An integral equation approach to boundary value problems of classical elastostatics. New York. Kim. T.. [145] Sandgren. (1967).H. F. LLC .K. (1989). and Wu. International Journal for Numerical Methods in Engineering 39:41194139. ASME Paper No. Shape optimization in threedimensional linear elasticity by the boundary contour method. S. (1986). [143] Rudolphi. J.. J.J. [148] Shi. American Society of Mechanical Engineers.J. Ramesh. Quarterly of Applied Mathematics 25:8395. Sladek. Nabors.M. and Aubry. J.214 BIBLIOGRAPHY [140] Rank.versions for boundary integral element methods.. K. [141] Rizzo. Computation of stresses by BEM in 2D elastostatics. and Mukherjee.P. V. A computeraided design system for microelectromechanical systems (MEMCAD). Boundaryelement shape optimization system for aircraft structural components. J. E. and Kane. (2003). Simulation methods for microelectromechanical structures (MEMS) with application to a microtweezer. Journal of Microelectromechanical Systems 1:313. X. and Sladek. [150] Shi. S. Annals of Operations Research 5:485500. K. and hp.. B.A. and Mukherjee.J. F. (1986). Harris. and White. [152] Sladek. (1995). (1996). IMECE200341579 . International Journal for Numerical Methods in Engineering 26:19131924.
Computer Methods in Applied Mechanics and Engineering 133:183208. and Mukherjee. S. B. Regularization techniques applied to boundary element methods. (2004). K. (1987). Generalpurpose computer program for shape optimization of engineering structures using the boundary element method. (1976). 3570 Plenum Press. Computers and Structures 56:713720. A selfadaptive coordinate transformation for eﬃcient numerical evaluation of general boundary element integrals. A.BIBLIOGRAPHY 215 [153] Sloan. A. © 2005 by Taylor & Francis Group.F. (1995). Goldberg ed. [158] Sukumar. (2000).. and Copps... Numerical Solution of Integral Equations Chapter 2. Error analysis in boundary integral methods. E. Moran. International Journal for Numerical Methods in Engineering 24:959973. [155] Sloan. M. and Belytschko. J.. Computers and Structures. and Fenner.dimensional problems. Superconvergence. Moran. and Belytschko. S. Extended boundary node method for threedimensional problems in potential theory. ASME Applied Mechanics Reviews 47:457499. [156] Stephan. [154] Sloan. International Journal for Numerical Methods in Engineering 43:839887. The natural element method.. V. Submitted. (2004).H. (1998). Improvement by iteration for compact operator equations. Semenov. S. Natural neighbour Galerkin methods. T. [161] Tafreshi.. Iserles ed. The hp boundary element method for solving 2and 3. N. [157] Strouboulis. [159] Sukumar. Moran. UK. (1994).C. Computer Methods in Applied Mechanics and Engineering 181:4369. S.H.P. A.. N. I. (2000). I. (2001). and Sladek. International Journal for Numerical Methods in Engineering 50:127. New York. I. International Journal for Numerical Methods in Engineering 48:15491570. R. and Mukherjee. Extended o ﬁnite element method for threedimensional crack modeling. J. A.T. Engineering Analysis with Boundary Elements 28:10991110. Acta Numerica Chapter 7. N. [164] Telukunta.Y. LLC . (1996). M¨es. Mathematics of Computation 30:758764. An extended boundary node method for modeling normal derivative discontinuities in potential theory across corners and edges.. [165] Telles. (1992). N.H. The design and analys sis of the generalized ﬁnite element method. [162] Tanaka. [160] Sukumar. T. Babuˇka. B. T. M. (1990). I. 287339 Cambridge University Press. Sladek. [163] Telukunta. B.. T. and Belytschko.
[177] Yamazaki. J. [170] Wendland. R. (1970). Y. G. [178] Yang. LLC . L. Communications in Numerical Methods in Engineering 14:505517. Leu. (1979). [167] Timoshenko.F. © 2005 by Taylor & Francis Group. [168] Toh. International Journal of Solids and Structures 36:10451079.P.J. (1998). S. and Saigal. International Journal of Solids and Structures 31:533550. [176] Xu. Shape optimization in elasticity and elastoviscoplasticity by the boundary element method. Elementfree Galerkin study of steady quasistatic crack growth in plane strain tension in elasticplastic materials. and Aliabadi.. Numerische Mathematik 53:539558. and Yu. [174] Xu. A. S.P... Mathematical Methods in the Applied Sciences 1:265321. and Oliveria. Threedimensional shape optimization using the boundary element method. Y. International Journal of Solids and Structures 31:22992312. (1998). Hypersingular and ﬁnite part integrals in the boundary element method. DH. A. S. [169] Wei. Journal of Computational Mathematics 10:273289. (1994). S. Computers and Structures 37:561568. S. Stephan. [175] Xu.N. An elementfree Galerkin analysis of steady dynamic growth of a mode I crack in elasticplastic materials. Theory of Elasticity 3rd. McGraw Hill. and Hsiao. R. W.216 BIBLIOGRAPHY [166] Telles. J. Chandra. X. [171] Wendland. A posteriori local error estimates of boundary element methods with some pseudodiﬀerential equations on closed curves. Component shape optimization using BEM. M.C. and Saigal. and Yu.H. Computational Mechanics 21:276282. S.L. W. An elementfree Galerkin formulation for stable crack growth in elastic solids. (1994). and Kitano. (1992). AIAA Journal 32:12951301. (1999).. [172] Wendland.F.C.J.L. Engineering Analysis with Boundary Elements 13:135141. DH. ed. New York. Computer Methods in Applied Mechanics and Engineering 154:331343. and Mukherjee. (1998). On the integral equation method for the plane mixed boundary value problem for the Laplacian.L. and Saigal.C. Third degree polynomial transformation for boundary element integrals : Further improvements. Y. K. E.J. K. and Goodier. (1988). Direct evaluation of boundary stresses in the 3D BEM of elastostatics. Sakamoto. (1994). W. (1994). J. [173] Wilde. (1990). and Mukherjee. Adaptive boundary element methods for strongly elliptic integral equations. M.
The dual boundary contour method for twodimensional crack problems. The Finite Element Method. Maidenhead. [182] Yu. and Lan.R. LLC . and Atluri. and Taylor. [186] Zhou. [190] Zienkiewicz. The traction boundary contour method for linear elasticity. Mathematical and Computational Aspects Vol. Z.Y. International Journal of Solids and Structures 27:983998. A posteriori error estimates and adaptive approaches for some boundary element methods. S.. C.N. Finite Element Structural Analysis. Springer. and Mukherjee. u [183] Zhang. (1991). Self adaptive boundary element methods. R..X. Brebbia. A new meshless regular local boundary integral equation (MRLBIE) method. N. Boundary stress calculation . (1994). S.H. International Journal for Numerical Methods in Engineering 46:12371252. T. Berkshire. Z. New York.J. UK.. Nishimura. Computational Mechanics Publications and Springer Verlag. [180] Yoshida. Application of a new fast multipole boundary integral equation method to crack problems in 3D. International Journal for Numerical Methods in Engineering 46:18831895. Shape Design Sensitivity Analysis and Optimization Using the Boundary Element Method.. (1998). Q. Engineering Analysis with Boundary Elements 25:239247.J. Design sensitivity coeﬃcients for linear elastic bodies with zones and corners by the derivative boundary element method. Vols.Y. International Journal of Fracture 92:201212. (1987). and Cao. [185] Zhao. 241256. O. T. JD. (1999). Zhang. S.a comparison study. [188] Zhu T.BIBLIOGRAPHY 217 [179] Yang. 1 of Boundary Elements IX. Computers and Structures 71:7785.C. McGraw Hill. [187] Zhou. Sun.. S.Y. and Kobayashi. S. L. D. (1999).X. D. and Sun. A local boundary integral equation (LBIE) method in computational mechanics. [189] Zhu. Kuhn eds.Y. (1998). (1999). 1 and 2. Computational Mechanics 21:223235. Cao. Z. S. S. W. Zeitschrift f¨r Angewandte Mathematik und Mechanik 68:T435T437. Wendland and G. Z. [184] Zhao. Prentice Hall. Englewood Cliﬀs. S. [181] Yu. NJ. © 2005 by Taylor & Francis Group. K. and a meshless discretization approach. (1986). (1988).H.L. (2001). (1991).