AN INTRODUCTION TO THE FINITE ELEMENT METHODMcGraw-Hill Series in Mechanical Engineering Consulting Editors Jack P. Holman, Southern Methodist University John R. Lloyd, Michigan State Univesity Anderson: Modern Compressible Flow: With Historical Perspective ‘Arora: Introduction to Optimum Design Bray and Stanley: Nondestructive Evaluation: A Tool for Design, Manufacturing, sand Service Culp: Principles of Energy Conversion Dally: Packaging of Electronic Systems: A Mechanical Engineering Approach Dieter: Engineering Design: A Materials and Processing Approach Eckert and Draket Analysis of Heat and Mass Transfer Edwards and McKee: Fundamentals of Mechanical Component Design Gebbart: Heat Conduction and Mass Transfer 1: Internal Combustion Engine Fundamentals Howell and Buckius: Fundamentals of Engineering Thermodynamics Hutton: Applied Mechanical Vibrations uvinall: Engineering Considerations of Stress, Strain, and Strength Kane and Levinson: Dynamics: Theory and Applications Kays and Crawford: Convective Heat and Mass Transfer Kimbrell: Kinematics Analysis and Synthesis Martin: Kinematics and Dynamics of Machines Modest: Radiative Heat Transfer Norton: Design of Machinery Phelan: Fundamentals of Mechanical Design Raven: Automatic Control Engineering Reddy: An Introduction to the Finite Element Method Rosenberg and Karnopp: Introduction to Physics Schlichting: Boundary-Layer Theory ‘Shames: Mechanics of Fluids ‘Sherman: Viscous Flow Shigley: Kinematic Analysis of Mechanisms Shigley and Mischke: Mechanical Engineering Design Shigley and Uieker: Theory of Machines and Mechanisms Stier: Design with Microprocessors for Mechanical Engineers Stoecker and Jones: Refrigeration and Air Conditioning Ullman: The Mechanical Design Process ‘Vanderplaats: Numerical Optimization: Techniques for Engineering Design, with Applications Viscous Fluid Flow ‘eld: CAD/CAM Theory and PracticeAlso Available from McGraw-Hill Schaum’s Out Most outlines include basic theory, definitions, and hundreds of solved problems and supplementary problems with answers, ‘Titles on the Current List Include: ¢ Series in Mechanical Engineering Acoustics Basic Equations of Engineering Continuum Mechanics Engineering Economics Engineering Mechanics, 4th edition Fluid Dynamics, 2d edition Fluid Mechanics & Hydraulics, 24 edition Heat Transfer Introduction to Engineering Calculations Lagrangian Dynamics Machine Design Mathematical Handbook of Formulas & Tables Mechonical Vibrations Operations Research Statistics & Mechanics of Materials Strength of Materials, 2d edition Theoretical Mechanics ‘Thermodynamics, 2d edition Schaum’s Solved Problems Books Each ttle in this series is a complete and expert source of solved problems containing thousands of problems with worked out solutions. Related Titles on the Current List Include: 3000 Solved Problems in Caleulus 2500 Solved Problems in Differential Equations 2500Solved Problems in Fluid Mechanics and Hydraulics 1000 Sotued Problems in Heat Transfer 3000 Solved Problems in Linear Algebra 2000 Solved Problems in Mechanical Engineering Thermodynamics 2000 Solved Problems in Numerical Analysis 700 Solved Problems in Vector Mechanics for Engineers: Dynamics £800 Solved Problems in Vector Mechanics for Engineers: Statistics Available at your College Bookstore. A complete list of Schaum titles may be obtained by writing to: Schaum Divi McGraw-Hill, Ine Princeton Road, S-1 Hightstown, NJ 08520AN INTRODUCTION TO THE FINITE ELEMENT METHOD Second Edition J.N. Reddy Oscar 8. Wyatt Chair in Mechanical Engineering Texas Ad&M University College Station, Texas 77843 McGraw-l Hill, I Inc. New Yok tous San Rrancico Auckland, Bogott Caracas Lisbon London Madrid Milan Montceal Now Delt Pus Sun Juan Singupre Syanoy Tokyo Teron“TA pet “This book was set in Times Roman. ‘The editors were John J. Corrigan and John M. Morrss;, Ro aa the production supervisor was Louie Karan "The cover wat designed by Joseph Gillian {0 Project supervision was done by The Universities Press R. R. Donnelley & Sors Company was printer and binder. TAILFSRE1993.—wuretsomamy Aa arodacten fe tte semeat inna 269154 [AN INTRODUCTION 10 THE FINITE ELEMENT METHOD Copytight © 1993, 1984 by MeGraw-Hil, In. All ights reserved. Printed i the United ‘States of America, Except as permitted under the United States Copyright Act of 1976, to part ofthis publication may be reproduced o distributed in any form or by any Ticens of stored in a data base or retrieval sytem, without the prior written permission ‘of the publisher. 294567890 DOCMOC 99876549 ISBN 007-051355-4 Library of Congress Cataloging edly, 1. N. Joba N.) "An introduction t0 the finite element method/J. N. Reddy.—2nd ed. pe cm—{MeGraw-Hil series in mechanical engineering) Incies bibliographical references and index. ISBN 0-97.051355-4 1, Finite element method, 1. Series. TAMTESRE 1993 - (6207 0013153520 9.29532ABOUT THE AUTHOR N, Reddy is the inaugural holder of the Oscar S. Wyatt, Jr, Endowed Chair ‘Mechanical Engineering at Texas A&M’ University. Prior to the current ‘appointment, he was the Clifton C. Garvin Chaired Professor of Engineering Science and Mechanics at Virginia Polytechnic Institute and State University, After receiving his Ph.D. in Applied Mechanics in 1974, he joined Lockheed Missiles and Space Company in Huntsville, Alabama, as a rescarch scientist. In 1975, he joined the School of Aerospace, Mechanical and Nuclear Engineering at the University of Oklahoma. In 1980, he was appointed as a professor of Engineering Science and Mechanics at Virginia Polytechnic Institute and State University Dr. Reddy has authored and coauthored over 150 journal papers and seven books on the theory and applications of the finite element method in solid and structural mechanics, fiuid mechanics and heat transfer, and on the development of refined theories and finite element models of laminated composite plates and shells. Dr. Reddy has taught many short courses on the finite element method to induistry and government, and has received Certificates of Teaching Excellence from Virginia Polytechnic Institute and State University. Dr. Reddy serves on the editorial boards of a dozen journals, including Journal of Applied Mechanics, Meccanica, International Journal for Numerical Methods in Engineering, International Journal of Numerical Methods in Fluids, Journal of Non-Linear Mechanics, and Journal of Composites Technology and Research, He has received the College of Engineering Research Award from the University of Oklahoma and the Alumni Research Award from the Virginia Polytechnic Institute and State University. Dr. Reddy was awarded the 1989 Walter L. Huber Research Prize of the American Society of Engineers and 1992 Worcester Reed Warner Medal of the American Society of Mechanical Engineers. He is a Fellow of the American Academy of Mectan- , the American Society of Mechanical Engineers, the American Society of ivil Engineers, and the Aeronautical Society of India. viiTo My FAMILY ‘Aruna, Anita, and AnilCONTENTS Preface to the Second Edition xvii Preface to the First Edition xix Part 1 Preliminaries 1 Introduction 3 1.1 General Comments 3 1.2. Historical Background 5 5 5 6 13 1 Concept of the Finite Element Method General Comments Approximation of the Circumference of a Circle Approximate Determination of the Center of Mass 8 Solution of Differential Equati 10 Some Remarks 3 1.4 The Present Study 15 1.5. Summary 15 References for Additional Reading 16 2. Integral Formulations and Variational Methods 18 2.1 Need for Weighted-Integral Forms 18 2.2 Some Mathematical Concepts and Formulze 20 2.2.1 Boundary, initial, and Eigenvalue Problems 20 2.2.2 Integral Relations 2 2.23 Functionals 26 2.2.4 The Variational Symbol 2 2.3. Weak Formulation of Boundary Value Problems 28 23.1. Introduction 28 2.3.2 Weightcd-Integral and Weak Formulations 28 2.3.3 Linear and Bilinear Forms and Quadratic Functionals 33 234 Examples 35x ‘contests 2.4 Variational Methods of Approximation 2.4.1 Introduction 2.4.2. The Rayleigh-Ritz Method 24:3. The Method of Weighted Residuals 25 Summary Problems References for Additional Reading Part 2. Finite Element Analysis of One-Dimensional Problems: 3 Second-Order Boundary Value Problems 3.41. Introduction 3.2. Basic Steps of Finite Element Analysis 3.2.1 Model Boundary Value Problem 3.2.2. Discrotization of the Domain 3.2.3 Derivation of Element Equations 3.2.4 Connectivity of Elements 32.5 Imposition of Boundary Conditions 3.2.6. Solution of Equations 3.2.7. Postprocessing of the Sotution 3.2.8 Radially Symmetric Problems 3.3 Applications 33.1 Heat Transfer 33.2 Fluid Mechanies 3.3.3. Solid Mechanics 3.4. Summary Problems References for Additional Reading. 4 Bending of Beams 4.1 Introduction 42. The Euler-Bernouli Beam Element 42.1 Governing Equation 42.2. Disertization ofthe Domain Derivation of Element Equ: ‘Assembly of Element Equatio ‘of Boundary Conditions 427 42:8 Examples 4.3. Plane Truss and Euler—Bernoulli Frame Elements 44 The Timoshenko Beam and Frame Elements 44.1 Governing Equations 44.2 Weak Form Postprocessing of the Solution 443. Finite Element Model 4.5. Inclusion of Constraint Equations 4.6 Summary eesesss oa 0 9 95 95 103 105 uy 13 wt 128 ua 3 3 143 143 144 14a 1st 154 156 158 160 167 a 7 m 178 187 11comms xi Problems 192 References for Adéitional Reading 198 Finite Element Error Analysis 199 5.1. Approximation Errors 19 5.2. Various Measures of Erors 200 533. Convergence of Solution 21 514 Accuracy ofthe Solution 202 5.5. Summary 207 Problems 207 References for Additional Reading 208 Eigenvalue and Time-Dependent Problems 209 6.1. Eigenvalue Problems 209 6.14 Introduction 208 642. Formulation of Eigenvalue Problems 210 6.13 Finite Element Models 213 6.1.4 Applications 216 6.2 Time-Dependent Problems ns 6.2.1 Introduction 224 62.2. Semidscrete Finite Element Models 225 62.3 Time Approximations 7 6.2.4 Mass Lomping m2 62.5 Applications 23 63 Summary eos Problems 261 References for Additional Reading 245 Numerical Integration and Computer Implementation 246 2A Tzoparametric Formulations and Nomeres! Integration 246 7AM Background 246 712 Natural Coordinates 248 74.3 Approximation of Geometry 29 714 Isoparametsic Formulations 251 7.1.5 Numerical Integration 251 72. Computer Implementation 238 72:1 Introductory Comments 258 General Outtine 239 Preprocessor 260 CCaleulation of Flement Matrices (Processor) 262 ‘Assembly of Element Equations (Processor) 265 Imposition of Boundary Conditions (Processor) 261 Solution of Equations and Postprocessing 26 13. Applications of the Computer Program FEMIDV2 200 TBA General Comments 20 73.2 Mlustrative Examples m 7.4 Summary 286 Problems 281 References for Additional Reading 291xii cones Part 3. Finite Element Analysis of Two-Dimensional Problems 8 Single-Variable Problems 8.1 Introduction 8.2 Boundary Value Problems 8.2.1. ‘The Model Equation 8.2.2 Finite Flement Discretization ‘Weak Form Finite Element Model Interpolation Functions Evaluation of Element Matrices and Vectors. ‘Assembly of Element Equations Postprocessing. ‘Axisymmetric Problems ‘An Example 8.3 Some Comments on Mesh Generation and Imposition ‘of Boundary Conditions 83.1 Discretization of a Domain 83.2 Generation of Finite Element Data 8.3.3 Imposition of Boundary Conditions 84 Applications 1 Heat Transfer 8.4.2 Fluid Mechanics 8.4.3. Solid Mechanics 8,5. Eigenvalue and Time-Dependent Problems 85.1 Introduction 8.5.2 Parabolic Equations 8.5.3 Hyperbolic Equations 8.6 Summary Problems References for Additional Reading 9 Interpolation Functions, Numerical Integration, and Modeling Considerations 9:1. Library of Elements and Interpolation Functions 9.1.1 Introduction 9.12 Triangolar Elements 9.43 Rectangular Elements 911.4 The Serendipity Elements 9.2. Numetial Integration 9.2.1 Preliminary Comments 922 Coordinate Transformations 9.2.3 Integration over a Master Rectangular Element 9.2.4. Integration over a Master Teiangular Blement 9.3. Modeling Considerations 93.1 Preliminary Comments 9°32 Element Geometries 93.3 Mesh Generation 9.3 Load Representation 94 Summary 33410 i Problems References for Additional Reading Plane Elasticity 10.1. Introduction 10.2, Governing Equations 10.2.1 Assumptions of Pane Elasticity 102.2 Basie Equations 10.3. Weak Formulations 103.4 Preliminary Comments 10:32. Peiniple of Virtual Displacements in Matrix Form 10.33 Weak Form ofthe Governing Differential Equations 10.4. Finite Blement Model 10.4.1. Matrix Form ofthe Modet 1042. Weak Form Model 10.4.3 Bigenvalue and Transient Problems 10.5. Evaluation of Integrals 10.6 Assembly and Boundary and Initial Conditions 10.7 Examples 10.8 Summary Problems References for Additional Reading Flows of Viscous Incompressible Fluids AL1 Preliminary Comments 11.2 Governing Equations 11.3 Velocity-Pressure Finite Blement Model 11.4 Penalty-Finite Element Model 11.4.1 Penalty Function Method 11.4.2. Formulation of the Flow Problem as a Constrained Problem References for Additional Reading Bending of Elastic Plates 12.1 Introduction 122. Classical Plate Model 12.2.1 Displacement Field 12.2.2 Virtual Work Statement 12.2.3 Finite Element Model 12.3 Shear Deformable Plate Model 12.3.1 Displacement Field 32.3.2 Virtual Work Statement 12.3.3 Finite Element Model 12.3.4 Shear Locking and Reduced Integration 12.4 Bigenvalue and Time-Dependent Problems 516 516 317 518 S19xiv comers 12.5. Examples sa 12.6. Summary 52 Problems 329 References for Adsitional Reading, 331 13 Computer Implementation 533 BAL Introduction 533 13.2. Preprocessor 38 13.3._Elemont Computations: Processor 535 13 Applications ofthe Computer Progrom FEM2DV2 540 134.1 Inoduetion 540 15.42 Description of Mesh Genrators 548 13.43 Applications (Hvstrative Examples) ssi 13.5. Summary 563 Problems 570 References for Additional Reading 575 Part 4 Advanced Topics 14 Weighted-Residual Finite Element Models, and Finite Blement Models of Nonlinear and Three-Dimensional Problems 319 14.1 Introduction 51 142. Alternative Formulations 5380 14.2.1. Introductory Comments 580 1422 Weighted-Residual Finite Element Models 530 Mixed Formulations 590 ar Problems 594 ‘General Comments 504 Large Deftection Bending of (Euler~Bernoull) Beams 595 14.3.3 Solution Methods for Noalinear Algebraic Equations 397 14.34 The 2-D Novier-Stokes Equations 598 14.4, Three-Dimensional Problems 599 14.5. Summary oo Problems a References for Additional Reading 606 Appendixes Appendixes 1 Fortran Listing of FEMIDV2 oo 2 Fortran Listing of FEM2DV2 640 Index - 679PREFACE TO THE SECOND EDITION ‘This second edition has the same objectives as the first, namely, an introduction to the finite element method as applied to linear, one- and two-dimensional problems of engineering and applied sciences. ‘The revisions are mainly in the form of additional details, expansion of the topics discussed, and the addition of a few topics to make the coverage more complete. ‘The major organizational change from the first edition is the division of its five chapters into fourteen chapters here. These chapters are grouped into four parts. This reorganization should aid instructors in selecting suitable material for courses. Other organizational changes include putting problem sets at the ends of the chapters, providing a chapter summary for each, and reviewing pertinent cquations and text in each chapter instead of referring to several chapters back. In addition, example problems in Chapters 3 and 8 are presented in separate sections on heat transfer, fluid flow, and solid mechanics. ‘Additional details are provided on the construction of the weak forms, time approximations (0.g., accuracy and stability of schemes, and mass lumping), alternative finite element formulations, and nonlinear finite element models. The new topics include sections on trusses and frames, the Timosh- enko beam clement, eigenvalue problems, and classical plate bending ele~ ments. All these changes are also reflected in the revised computer programs FEMIDV2 and FEM2DV2 (revised versions of the FEMID, FEM2D and PLATE programs in the first edition). Therefore the sections on computer implementation and applications of FEMIDV2 and FEM2DV? have also been modified extensively. These changes are accompanied by the addition of several figures, tables, and examples. ‘These extensive changes have resulted in a second edition that is 60% larger. In the interest of keeping the cost of the book within reasonable limits Poxvi PREFACE 10 THB SECOND EDITION retaining the basic approach and technical details, certain portions of the original manuscript have been omitted. More specifically, answers to selective problems have been included at the end of the problem statements themselves, rather than in a separate section. Interested readers and instructors can obtain ‘copy of the excutable programs on a diskette from the author. Fortran source programs can also be purchased from the author. ‘There is no doubt that this edition is more complete and thorough than the first. It can be used as a textbook for an introductory and/or intermediate level course on the finite element method at senior undergraduate as well as graduate levels. Students of engincering and applied sciences should feel ‘comfortable with the coverage in the book, ‘The author gratefully acknowledges help in reading the manuscript and suggestions for constructive changes from several colleagues. ‘These include: Hasan Akay, Purdue University at Indianapolis, Norman Knight, Jr, Clemson University; J. K. Lee, Ohio State University; William Rule, University of ‘Alabama; Martin Sadd, University of Rhode Island; John Whitcomb, Texas A&M University, and the author’s research students: Ronald Averill, Filis Kokkinos, Y. S. N. Reddy, and Donald Robbins. It is a great pleasure to acknowledge typing of the manuscript by Mrs Vanessa McCoy, without whose patience and cooperation this work would not have been completed. JN. ReddyPREFACE TO THE FIRST EDITION ‘The motivation which led to the writing of the present book has come from my many years of teaching finite-element courses to students from various fields of engincering, meteorology, geology and geophysics, physics, and mathematics. ‘The experience gained as a supervisor and consultant to students and colleagues in universities and industry, who have asked for explanations of the various mathematical concepts related to the finite-element method, helped me introduce the method as a variationally based technique of solving differential equations that arise in various fields of scicnce and engineering. The many discussions I have had with students who had no background in solids and structural mechanics gave rise to my writing « book that should fill the rather unfortunate gap in the literature. “The book is designed for senior undergraduate and first-year graduate students who have had a course in linear algebra as well as in differential ‘equations. However, additional courses (or exposure to the topics covered) in mechanics of materials, fluid flow, and heat transfer should make the student feel more comfortable with the physical examples discussed in the book, In the present book, the finite-element method is introduced as a variationally based technique of solving differential equations. A. continuous problem described by a differential equation is put into an equivalent variational form, and the approximate solution is assumed to be a linear combination, ¥ ¢@;, of approximation functions g,. The parameters c, are determined using the associated variational form. The finite-element method provides a systematic technique for deriving the approximation functions for simple subregions by which a geometrically complex region can be repre- sented. In the finite-element method, the approximation functions are piece- wise polynomials (i.e., polynomials that are defined only on a subregion, called an clement).PRRPACE To THE FIRST EDITION ‘The approach taken in the present book falls somewhere in the middle of the approaches taken in books that are completely mathematical and those approaches that are more structural-mechanics-oriented. From my own ex- perience as an engineer and self-taught applied mathematician, I know how Unfortunate outcomes may be arrived at if one follows a “formula” without deeper insight into the problem and its approximation, Even the best theories lead ultimately to some sort of guidelines (e.g., which variational formulation is suitable, what kind of element is desirable, what is the quality of the approximation, etc.). However, without a certain theoretical knowledge of variational methods one cannot fully understand various formulations, finite- clement models, and their imitations. Tin the present study of variational and finite-element methods, advanced mathematics are intentionally avoided in the interest of simplicity. However, a minimum of mathematical machinery that seemed necessary is included in Chapters 1 and 2. In Chapter 2, considerable attention is devoted to the construction of variational forms since this exercise is repeatedly encountered in the finite-clement formulation of differential equations. The chapter is ‘concerned with two aspects: first, the selection of the approximation functions that meet the specified boundary condtions; second, the technique of obtaining algebraic equations in terms of the undetermined parameters, Thus, Chapter 2 not only equips readers with certain concepts and tools that are needed in Chapters 3 and 4, but it also motivates them to consider systematic methods of constructing the approximation functions, which is the main feature of the finite-element method. In introducing the finite element method in Chapters 3 and 4, the traditional solid mechanics approach is avoided in favor of the “differential equation” approach, which has broader interpretations than a single special casc. However, when specific examples are considered, the physical back- ground of the problem is stated, Since a large number of physical problems are described by second- and fourth-order ordinary differential equations (Chapter 3), and by the Laplace operator in two dimensions (Chapter 4), considerable attention is devoted to the finite-element formulation, the derivation of the interpolation functions, and the solution of problems described by these equations. Representative examples are drawn from various fields of engineer- ing, especially from heat transfer, fluid mechanics, and solid mechanics, Since this book is intended to serve as a textbook for a first course on the finite-element method, advanced topics such as nonlinear problems, shells, and three-dimensional analyses are omitted, Since the practice of the finite-clement method ultimately depends on one’s ability to implement the technique on a digital computer, examples and exercises are designed to let the reader actually compute the solutions of various problems using computers. Ample discussion of the computer im- plementation of the finite-element method is given in Chapters 3 and 4. Three model programs (FEMID, FEM2D, and PLATE) are described, and their application is illustrated via several examples. The computer programs are very ‘easy to understand because they are designed along the same lines as thePREEACE TO THE FiRST EDITION XIX theory presented-in the book. The programs are available for mainframe and IBM PC compatibles from the author for a small charge. Numerous examples, most of which are applications of the concepts to specific problems in various fields of engineering and applied science, are provided throughout the book. The conclusion of the examples are indicated by the symbol ll. At approprate intervals in the book an extensive number of exercise problems is included to test and extend the understanding of the concepts discussed. For those who wish to gain additional knowledge of the topics covered in the book, many reference books and research papers are listed at the end of each chapter. ‘There are several sections that can be skipped in a first reading of the book (such sections are marked with an asterisk); these can be filled in wherever needed later. The material is intended for a quarter or a semester course, although it is better suited for a semester course. ‘The following schedule of topics is suggested for a first course using the present textbook: Undergraduate Graduate Chapter 1 Self-study Chapter 1 Self-study Chapter 2 Section 2.1 (self) Chapter 2. Section 2.1 (self) Section 2.2 Section 2.2 Sections 2.3.1-2.3.3, Section 2.3 Chapter 3. Sections 3.1-3.4 Chapter 3 Sections 3.1-3.7 Sections 3.6-3.7 Chapter 4 Sections 4.1-4.4 Chapter 4 Sections 4.1-4.8 Section 4.7 Sections 4.8.1-4.8.4 Chapter 5 Term Paper Due to the intimate relationship between Sections 3.5 and 4.6, 3.6 and 4.7, and 3.7 and 4.8, they can be covered simultaneously. Also, it is suggested! that Sections 3.6 and 3.7 (hence, 4.7 and 4.8) be covered after Section 3.2. The author wishes to thank all those students and colleagues who have contributed by their advice and criticism to the improvement of this work. The author is also thankful to Vanessa McCoy for skillful typing of the manuscript, to Mr, N. S, Putcha and Mr. K. Chandrashekhara for proofreading the pages, and to the editors Michael Slaughter and Susan Hazlett for their help and cooperation in publishing the mauscript. JN, Reddy Tejashowina vadheetamasthe (May what we study be well studied)PART. 1 PRELIMINARIESCHAPTER 1 INTRODUCTION 1.1 GENERAL COMMENTS Virtually every phenomenon in nature, whether biological, geological, or mechanical, can be described with the aid of the laws of physics, in terms of algebraic, differential, or integral equations relating various quantities of interest. Determining the stress distribution in a pressure vessel with oddly shaped holes and numerous stiffeners and subjected to mechanical, thermal, and/or acrodynamic loads, finding the concentration of pollutants in seawater or in the atmosphere, and simulating weather in an attempt to understand and predict the mechanies of formation of tornadoes and thunderstorms are a few examples of many important practical problems. ‘Most engineers and scientists studying physical phenomena are involved with two major tasks: 1, Mathematical formulation of the physical process 2, Numerical analysis of the mathematical model The mathematical formulation of a physical process requires background in related subjects (€.g., laws of physics) and, most often, certain mathematical tools, The formulation results in mathematical statements, often differential 34 etnananres equations, relating quantities of interest in the understanding and/or design of the physical process. Development of the mathematical model of a process is achieved through assumptions concerning how the process works. In a humerical simulation, we use a numerical method and a computer to evaluate the mathematical model and estimate the characteristics of the process. While the derivation of the governing equations for most problems is not unduly difficult, their solution by exact methods of analysis is formidable task. In such cases, approximate methods of analysis provide alternative means Of finding solutions. Among these, the finite difference method and the variational methods such as the Rayleigh~Ritz and Galerkin methods are most frequently used in the literature. Tn the finite difference approximation of a differential equation, the derivatives in the latter are replaced by difference quotients (or the function is expanded in a Taylor series) that involve the values of the solution at discrete mresh points of the domain, The resulting algebraic equations arc solved, after imposing the boundary conditions, for the values of the solution at the mesh points. Tn the solution of a differential equation by a variational metbod, the equation is put into an equivalent weighted-integral form and then the approximate solution over the domain is assumed to be a linear combination (Ech) of appropriately chosen approximation functions ¢, and undetermined cosffcients, ¢, ‘The coefficients c, are determined such that the integral Statement equivalent to the original differential equation is satisfied. Various variational methods, e.g., the Rayleigh-Ritz, Galerkin, and least-squares methods, differ from each other in the choice of the integral form, weight functions, and/or approximation functions. A more complete discussion of variational methods will be given in Chapter 2. They suffer from the Gisadvantage that the approximation functions for problems with arbitrary domains are difficult to construct. "The finite element method overcomes the disadvantage of the traditional variational methods by providing a systematic procedure for the derivation of the approximation functions over subregions of the domain, The method is endowed with three basic features that account for its superiority over other competing methods. First, a geometrically complex domain of the problem represented as a collection of geometrically simple subdomains, called finite tlements, Second, over each finite element, the approximation functions are derived using the basic idea that any continuous function can be represented by a linear combination of algebraic polynomials. Third, algebraic relations among the undetermined coefficients (i.e., nodal values) are obtained by satisfying the governing equations, often in a weighted-integral sense, over ‘each element. Thus, the finite element method can be viewed, in particular, as an element-wise application of the Rayleigh-Ritz or weighted-residual meth- ‘ods. In it, the approximation functions are often taken to be algebraic polynomials, and the undetermined parameters represent the values of the olution at a finite number of preselected points, called nodes, on the boundary and in the interior of the element. The approximation funetions arewtropverion derived using concepts from interpolation theory, and are therefore called interpolation functions. One finds that the degree of the interpolation functions depends on the number of nodes in the element and the order of the differential equation being solved. 1.2 HISTORICAL BACKGROUND ‘The idea of representing a given domain as a collection of discrete parts is not unique to the finite element method. It was recorded that ancient mathe- maticians estimated the valuc of sr by noting that the perimeter of a polygon inscribed in a circle approximates the circumference of the latter. ‘They predicted the value of z to accuracies of almost 40 significant digits by representing the circle as a polygon of a finitely large number of sides. In modern times, the idea found a home in aircraft structural analysis, where, for example, wings and fuselages are treated as assemblages of stringers, skins, and shear panels, In 1941, Hrenikoff introduced the so-called framework method, in which a plane clastic medium was represented as a collection of bats and beams. The use of piecewise-continuous functions defined over a subdomain to approximate an unknown function can be found in the work of Courant (1943), who used an assemblage of triangular elements and the principle of minimum total potential energy to study the St Venant torsion problem, Although certain key features of the finite element method can be found in the works of Hrenikoff (1941) and Courant (1943), its formal presentation is attributed to Argyris and Kelsey (1960) and Turner, Clough, Martin, and Topp (1956). The term “finite element” was first used by Clough in 1960. Since its inception, the literature on finite element applications has grown exponentially, and today there are numerous journals that are primarily devoted to the theory and application of the method. A review of the historical developments and the basic theory of the finite element method can be found in mote than three dozen textbooks that are exclusively devoted to its introduction and application. The selective finite clement books listed in References for Additional Reading at the end of this chapter are only for additional information on certain topics (c.g., three-dimensional problems, shells, structural dynamics, plasticity, and mathematics of finite elements). For the beginner, it is not necessary to consult these; the present book provides complete details of the method as applied to linear field problems, with ‘examples from fuid mechanics, heat transfer, and solid mechanics, 1.3 THE BASIC CONCEPT OF THE FINITE ELEMENT METHOD 1.3.1 General Comments ‘The most distinctive feature of the finite element method that separates it from others is the division of a given domain into a set of simple subdomains, called finite elements, Any gcometric shape that allows computation of the solution or its approximation, or provides necessary relations among the values of the6 reecmunaxnes solution at selected points, called nodes, of the subdomain, qualifies as a finite element. Other features of the method include seeking continuous, often polynomial, approximations of the solution over cach element in terms of nodal values, and assembly of element equations by imposing the interelement continuity of the solution and balance of interelement forces. Here the basic ideas underlying the finite element method are introduced via two simple examples: 1, Determination of the circumference of a circle using a finite number of line segments 2, Determination of the center of mass (or gravity) of an irregular body ‘The first example is an expansion of an article written by the author in 1978 for fa student magazine at the University of Oklahoma. Ideas. expressed in the second can be found in books on statics of rigid bodies. 1.3.2. Approximation of the Circumference of a Circle Consider the problem of determining the perimeter of a circle of radius R (see Fig. 1.la). Ancient mathematicians estimated the value of the circumference @ Element Node o Blement fength ke eURE LA Approximation of the circum ference ofa ciel by line elements (@) Gitele of radius (0) vaiform ‘and nonuniform meshes used to represent the circumference of the circ; () a typieal element.emopucrion 7 by approximating it by line segments, whose lengths they were able to measure. The approximate value of the circumference is obtained by summing the lengths of the line segments used to represent it. Although this is a tri ‘example, it illustrates several (but not all) ideas and steps involved in the finite element analysis of a problem. We outline the steps involved in computing an approximate value of the circumference of the citcle. In doing so, we introduce certain terms that are used in the finite element analysis of any problem. 1. Finite element discretization, First, the domain (j.e., the circumference of the circle) is represented as a collection of a finite number n of subdomains, namely, Tine segments. This is called discretization of the domain, Hach subdomain ie segment) is called an element. ‘The collection of elements is called the finite element mesh. ‘The elements are connected to each other at points called nodes. In the present case, we discretize the circumference into a mesh of five (n= 5) line segments. The line segments can be of different lengths. When all elements (i, line segments) are of the same length, the mesh is said to be uniform; otherwise, it is called a nonuniform mesh (see Fig. 1.15). Element equations, A typical element (i.e., line segment, Q*) is isolated and its required properties, ie., length, are computed by some appropriate means. Let fi, be the length of element 2° in the mesh, For a typical element 2, h, is given by (see Fig. 1.1c) ‘A, = 2R sin 40, aay where R is the radius of the circle and @, <1 is the angle subtended by the line segment. The above equations are called clement equations. Ancient mathematicians most likely made measurements, rather than using (1.1), to find fh, 3. Assembly of element equations and solution. The approximate value of the circumference (or perimeter) of the circle is obtained by putting together the element properties in a meaningful way; this process is called the assembly of the element equations. It is based, in the present case, on the simple idea that the total perimeter of the polygon (assembled elements) is equal to the sum of the lengths of individual clements: : m= She a2) 2 Then P, represents an approximation to the actual perimeter, p. If the mesh is uniform, or h, is the same for each of the elements in the mesh, then 6,=2n/n, and we have anlar’) as 4. Convergerice and error estimate. For this simple problem, we know the exact solution: p =2nR. We can estimate the error in the approximation and show that the approximate solution P, converges to the exact p in the8 enenaneanies limit as n>. Consider the typical element 9%. The error in the approximation is equal to the difference between the length of the sector and that of the line segment (see Fig. 1-1c): Be=|Se— hel 4) Where 5, = RO, is the length of the sector. Thus, the error estimate for an ‘element in the mesh is given by Qn =R(2-2sin2 5 #,=R(2-2sin) as) The total error (called global errar) is given by multiplying E, by 1: a” B= 2R(n-nsin2) =24R ~ F, (1.6) We now show that goes to zero as n>, Letting x = 1/m, we have and cos ax" S38) a2nR (1.7) ti = te (202222) =n (2 Hence, E, goes to zero as n>, This completes the proof of convergence. In summary, it is shown that the circumference of @ circle can be approximated as closely as we wish by a finite number of piecewise-tinear functions. As the number of elements is increased, the approximation improves, i.e., the error in the approximation decreases. 1.3.3 Approximate Determination of the Center of Mass ‘Another elementary example to illustrate the finite element concept is provided by the calculation of the center of mass of a continuous body. It should be recalled, from a first course on statics of rigid bodies, that the calculation of the center of an irregular mass or the centroid of an ieregular volume makes use of the so-called method of composite bodies, in which a body is conveniently divided (mesh discretization) into several parts (elements) ‘of simple shape for which the mass and the center of mass (element properties) can be computed readily. The center of mass of the whole body is then obtained using the moment principle of Varignon (a basis for the assembly of element properties): (my tims be. bm KS mye taka tt Ok (8) where X is the x coordinate of the center of mass of the whole body, m, is the mass of the eth part, and %, is the x coordinate of the center of mass of the ethprooucrion 9 att, Similar expressions hold for the y and z coordinates of the center of mass of the whole body. Analogous relations hold for composite lines, areas, and volumes, wherein the masses are replaced by lengths, areas, and volumes, respectively. , When a given body is not expressible in terms of simple geometric shapes (elements) for which the mass and the center of mass can be represented mathematically, it is necessary to use a method of approximation to represent the properties of an element, As an example, consider the problem of finding the centroid (%, 7) of the irregular area (region) showa in Fig. 1,2. The region can be divided into a finite number of rectangular strips (elements), a typical element having width h, and height b,. The area of the eth strip is given by Ae=hzb., The atca A, is an approximation of the true area of the element because b, is an estimated average height of the clement. ‘The coordinates of the centroid of the region are obtained by applying the moment principle: where % and J, are the coordinates of the centroid of the eth element with respect to the coordinate system used for the whole body. When the center of ‘mass is required, A, in the above equations is replaced by the mass m, = p.A,, p, being the mass density of the eth element; for a homogeneous body, pe is the same for all elements. It should be noted that the accuracy of the approximation will be improved by increasing the number of strips (decreasing their width) used. Rectangular elements are used in the present discussion for the sake of simplicity only; one may choose to use elements of any size and shape that approximate the given area to a satisfactory accuracy. For example, a trapezoidal element will require two heights to compute the area: Ae = thee + bert) > ante iva if / bl : [fs F Ly FIGURE 1.2 Approximate determination ofthe mass or geometric eentoid ofan ieregular reghon by dividing it into a set of retangular or trapezoidal subregions.10 rretnsesanues where b, and Bq; ate the left and element. “The two examples considered above illustrate how the idea of piecewise approximation is used to approximate irregular geometries and calculate required quantities. In the first example, the circumference of @ circle is approximated by a collection of line segments, whose measure is available. In the second, the geometric centroid or mass centroid of an irregular domain is focated by approximating its geometry as a collection of strips that allow computation of their arcas. Rectangles and trapezoids provide examples of the element geometries. ‘Thus, subdividing a geometrically complex domain into parts that allow the evaluation of desired quantities is a very natural and practical approach. ‘The idea can be extended to approximate functions fepresenting physical quantities. For example, the temperature variation in a {wo-dimensional domain can be viewed as @ curved surface, and it ean be approximated over any part of the domain, i.c., over a subdomain or element, by a function of desired degree. Figure 1.3 shows a cutved surface over a triangular subregion approximated by a planar surface, i.c., a linear polyno- mmial, Such ideas form the basis of finite element approximations. ‘The next example illustrates this idea for a one-dimensional continuous system. it heights, respectively, of the eth 1.3.4 Solution of Differential Equation Consider the temperature variation in a composite cylinder consisting of two coaxial layers in perfect thermal contact (see n from a (Qwith two insulations) carrying an electric current and heat flow across & thick-walled composite circular cylindrical tube are typical examples. The temperature T is a function of the radial coordinate r. The variation of T with r is, in general, nonuniform. We wish to determine an approximation T(r) to T(r) over the thicknesses of the cylinder, The exact solution is determined by solving the differential equation 1d( di 1S (ma (1.98) coved suntce L FIGURE 13 = ‘Anprowimeton of curved surface (or @ nonuniform ‘hreton) over 2 angular region by = planar sreenemopuction 11 FIGURE 1. (@) Coaxial (composite) cylinder made of ‘eo diferent materials. (6) Finite element representation of a radial line of the liner. subject to appropriate boundary conditions, for example, insulated at r= and subjected to a temperature 7, at r= R,: ar. ke ar atr=Ri TW)=% atr=R, (a.95) where k is the thermal conductivity, which varies from layer to layer, Ry and R, are the inner and outer radii of the cylinder, and q is the rate’ of energy generation in the medium. Note that the temperature is independent of the circumferential coordinate (because of the axisymmetric geometry, boundary conditions, and loading), and it has the same variation along any radial line When it is difficult to obtain an exact solution of the problem (1.9), either because of complex geometry and material properties or because q(r) is a complicated function that does not allow exact evaluation of its integral, we seck an approximate one. In the finite element method, the domain (Rj, R,) is divided into N subintervals, and the approximate solution is sought in the form Ti) S Divi) (Ri THOM) (Rit hy +++ + hyySr Uy = ey(4? = 2e) + 6360 —3x) +1 which satisfies the boundary conditions (2.26) of the problem for any values of ¢ and ¢;. ‘The constants c) and ¢> are to be determined such that the ‘Uy in (2.3) satisfies (2.22) in some sense. If we require Uy to satisfy (2.2a) in the exact sense, we obtain Wy PUy 4 Uy = —2ex(e ~ 1) = Bea(x? = 1) —2epx ~ 6,27 de de 0 Since this expression must be zero for any value of x, the coefficients of the various powers of x must be zero: 1420, +3 (6, +3e:)=0 =9er=0 @=0 4 ey(x?=2r) + 0,00 — 32) +1 ‘The above relations are inconsistent; hence, there is no solution to the equations. On the other hand, we can require the approximate solution U to satisfy the differential equation (2.2a) in the weighted-integral sense, f wRde=0 (24a) where X is called the residual, #Uy oo Wy20 eneunnanses and ww is called a weight function. From (24a), we obtain as many linearly independent equations as there arc independent functions for w. For example, wwe take w=1and w =x, we obtain 0 f AR de = (14 2e1 + 3ea) + 4(-6ey — 3¢2) + Hes — 9a) + Ae om f de =Se1 26,496) + H6e~36) # Mere +e or Joti =1 Jet Ber=t . which provide two linearly independent equations for cy and ¢z (giving =F and c= ~'$). ° ‘Thus, integral statements of the type in (2.4q) provide means for obtaining as many algebraic equations as there are unknown coefficients in the ‘approximation. This chapter deals with the construction of different types of integral statements used in different variational methods, A variational method js one in which approximate solutions of the type u ~ 3 ¢j¢ + @o are sought, ‘and the coefficients c, are determined, as shown above, using an integral Statement. The variational methods differ from each other in the choice of the ‘weight function w and the integral statement used, which in turn dictates the Choice of the approximation functions ¢, In the finite element method, a given domain is viewed as an assemblage of subdomains (j.c., elements), and an ‘approximate solution is sought over each subdomain in the same way as in variational methods. Therefore, it is informative to study variational methods before we study the finite element method. ‘Our goal in this chapter is to illustrate the basic steps in the integral formulations and the associated approximations of various boundary problems. ‘Toward this goal, we first introduce necessary terminology and notation. (2.4b) 2.2 SOME MATHEMATICAL CONCEPTS AND FORMULAE 2.2.1 Boundary, Initial, and Eigenvalue Problems DOMAIN AND BOUNDARY. The objective of most analyses is to determine ‘unknown functions, called dependent variables, that satisfy a given set of differential equations in a given domain or region and some boundary ‘conditions on the boundary of the domain. A domain is a collection of points in space with the property that if P is a point in the domain then all points sufficiently close to P belong to the domain, ‘This definition implies that a domain consists only of internal points. If any two points of the domain can be joined by a line lying entirely within it then the domain is said to be convexINTEGRAL FORMULATIONS AND VARIATIONAL NETiODS 21 and simply connected, The boundary of a domain is the set of points such that, jn any neighborhood of each of these points, there are points that belong to the domain as well as points that do not. Note from the definition that the points on the boundary do not belong to the domain. We shall use the symbol to denote an arbitrary domain and T' to denote its boundary. ‘A function of several variables is said to be of class C"(Q) in a domain @ if all its partial derivatives up to and including the mth order exist and are continuous in . Thus, if fis of class C° in two dimensions then f is continuous (i.c., af /ax and 5f/ay exist but may not be continuous). The letters x and y will always be used for rectangular coordinates of a point in two dimensions. When the dependent variables arc functions of one independent variable (ay, x), the domain is a line segment (je., one-dimensional) and the endpoints of the domain are called boundary points. When the dependent variables are functions of two independent variables (say, x and y), the (two-dimensional) domain is a surface (most often a plane) and the boundary is the closed curve enclosing it. It is not uncommon to find problems in which the dependent variable and possibly its derivatives are specified at points interior to the domain (e.g., bending of continuous beams) A differential equation is said to describe a boundary value problem it the dependent variable and possibly its derivatives are required to take specified values on the boundary. An initial value problem is one in which the dependent variable and possibly its derivatives are specified initially (ie., at time ¢=0). Initial value problems are generally time-dependent problems. Examples of boundary and initial value problems are given below. BOUNDARY VALUE PROBLEM (2.5) wad (05)] 26 {TIAL VALUE PROBLEM pT raunf for 03) will be zero. Figure 2.4 shows a comparison of the Rayleigh-Ritz solution with the exact solution.INTEGRAL FORMULATIONS AND VARIATIONAL wETHODS 49) 20: Three parameter solution s+ Two-parameter solution ‘One parameter solution 10) ao wx 10? 9 oe an) ~ He -20 ft it -20 00 02 04 0s 08 0 a2 aw rigune 24 Comparison of the Rayleigh-Ritz solution with tho exset solution of a cantilever beam under a uniform transverse load (Euter-Bernouli beam theory). The next example deals with two-dimensional heat conduction in a square region. Note that the dependent variable, namely the temperature, is denoted by 7, consistent with the standard notation used in heat transfer books. Example 2.6. Consider the Poisson equation in a unit square region: SEPT = qo in 2= (2, 9):0<@%y)<1) 2.940) T=0 onsides x=1 and y=1 ai 2.940) 0 onsides x=0 and y=0 'n where qy is the rate of uniform heat generation in the region. The variational problem is of the form (see Example 2.3) Bw, 7) = 160) (2.952) sar and linear functionals are where the bili won foe ih Fvaoaray We consider an N-parameter approximation of the form axdy @.98b) Ta= S cvemoxcoray, aa li-ne 2.96)50° rrsuowsanuss Note that (2.96) involves @ double summation. Since the boundary conditions are tromogencous, we have $o=0. Incidentally, @ also satisfies the natural boundary conditions of the problem. While the choice 4,=sin ix sininy meets the essential boundary conditions, it is not complete, because it cannot be used to generate the solution that daes not vanish on the sides x =O and y=0. Hence, are not admissible. “The coefficients By and F; can be computed by substituting (2.96) into (2.95b). Since the double Fourier series has two summations [see (2.96)], we introduce the notation Baoan [for sin an 05 a9) sin ax 28019) + (ay cos ex sin a,y)(az 608 ax sin ay)] dr dy -( it LAK or f4t Ak(@} + aj) it isk and j=! @sta) saa, [os ir cos yy de dy = 2 sn a sn 2.07) In evaluating the integrals, the following orthogonality conditions were used it ij iti it i4j it i=} [[seonsnandea? [ovaonnntl ‘Owing to the diagonal form of the coefficient matrix (2.97a), we can readily solve for the coefficients ey: Ft snavsing eqn Bia Mo sinasin Book (a+ eae os ‘The one- and two-parameter Rayleigh-Ritz solutions are 3240, cs cos bce cos tay (2.99) {10.3285 cos ix cos Lny ~ 0.0219(c0s 4x cos Bxy + c08 dtr cos bry) +0.0081 c0s Zax cos 27y)] (2.100) I algebraic polynomials ae to be used in the approximation of T, one can choose $:= (1-20 -y) of gr = (1-21-99), both of which satisfy the’ (homogeneous) essential boundary conditions. However, the choice ¢,= (1 —x°)(1—y*) also meets the natural boundary conditions of the problem. The one-parameter Ritz solution for the choice @=(1—\(1-y?) is 1,9) -a-z0-¥) @.10) ‘The cxact solution of (2.94a, b) is & (=i 0s any cosh Tey) <2 [ayy +4 3 ress cny concer (2.102)INTEGRAL FORMULATIONS AND VARIATIONAL METHODS SI na. Sins — hay 30) Oheranntr ~ Ongar = romance Jor Three p 00 00 02 04 06 o8 10 12 FIGURE 2.5 CCompatison of the Rayleigh-Ritz solutions with the analytical sol (2.98) in two dimensions ‘of the Poisson equat where a, =4(@n~ 1). The Rayleigh-Ritz solutions (2.99), (2.100), and (2.101) are compated with the exact solution (2.102) in Fig. 2.5. Tho analytical solution is evaluated using 20 terms of the series (2.102). 2.4.3 The Method of Weighted Residuals As noted in Section 2.3.2, one can always write the weighted-integral form of a differential equation, whether the equation is linear or nonlinear (in the dependent variables). The weak form can be developed if the equations are second-order or higher, even if they are nonlinear. However, it is not always possible to construct ‘a functional whose first variation is equal to the variational form. The Rayleigh-Ritz method can also be applied to all problems, including nonlinear problems, that have weak forms. In this method, the weight functions are necessarily equated to those used in the approximation. The weighted-residual method is a generalization of the Rayleigh-Ritz method in that the weight functions can be chosen from an independent set of functions, and it requires only the weighted-integral form to determine the parameters. The method of weighted residuals can be used to approximate the weighted-integral form of any equation. Since the latter form does not include any of the specified boundary conditions of the problem, the approximation functions should be selected such that the approximate solution satisfies both the natural and essential boundary conditions. In addition, the ‘weight functions can be selected independently of the approximation functions, but are required to be linearly independent (so that the resulting algebraic52. pReuinvants equations are linearly independent). This flexibility is advantageous in certain nonlinear problems. In this section, we discuss the general method of weighted residuals first, and then consider certain special cases that are known by specific names (e.g., the Galerkin and least-squares methods). Although a limited use of the wweighted-residual method is made in this book (see Chapter 14), it informative to have a knowledge of this method for use in the formulation of certain nonlinear problems readers might encounter in their work. The method of weighted residuals can be described in its generality by considering the operator equation A(u)=f in Q, (2.103) where A is an operator (linear or nonlinear), often a differential operator, acting on the dependent variable u, and f is a known function of the independent variables. Some examples of such operators are provided by (of) 400 ale) LebS) S65) e109 4. Alu) = -4 (4) 5. Alu 2) = 02 4 yy FH ae 8) Buy Ou, Fu, 8 (du, av "Bet Pay tat * ay ay * Be For an operator A to be linear in its arguments, it must satisfy the relation A( ait + fv) = a@A(u) + BAY) (2.105) for any scalars a and fi and dependent variables w and v. It ean be easily verified that all operators in (2.104), except for 4 and 5, are linear. When an operator does not satisfy the condition (2.105), it is said to be nonlinear. The function uw is not only required to satisfy the operator equation (2.103), it is also required to satisfy the boundary conditions associated with the operator equation, From the examples considered so far, the boundary conditions associated with the operators defined in 1, 2, and 3 of (2.104) are obvious [see Examples 2.1-2.3], __ Inthe weighted-residual method, the solution w is approximated, in much the same way as in the Rayleigh-Ritz method, by the expression w= 3 691+ bo 2.106)INTEGRAL FORMULATIONS AND VARIATIONAL METHODS $3 coxcept that the requirements on o and @ for the weighted-residual method fare more stringent than those for the Rayleigh-Ritz method. Substitution of the approximate solution uy into the left-hand side of (2.103) gives a function ‘fv= Aun) that, in general, is not equal to the specified function f. The difference A(uy) ~f, called the résidual of the approximation, is nonzezo: Renate) -foal 3 of, +) -F40 e107 Note that the residual R is a function of position as welt as of the parameters ¢ In the weighted-residual method, as the name suggests, the parameters cj are determined by requiring the residual R to vanish in the weighted-integral sense: [vi nRee sep dedy=0 ¢ Dyer ND (2.108) where @ is a two-dimensional domain and y; are weight functions, which, in general, are not the same as the approximation functions g,. ‘The set {1,} must be a linearly independent set; otherwise, the equations provided by (2.108) will not be linearly independent and hence will not be solvable. ‘The requirements on go and ¢ for the weighted-residual method are different from those for the Rayleigh-Ritz method, which is based on the weak (integral) form of the differential equation. The differentiability requirement ‘on @; in the weighted-residual method is dictated by the integral statement (2.108), as opposed to the weak form in the Rayleigh-Ritz method. ‘Thus, dj must have nonzero derivatives up to the order appearing in the operator equation (2.103). Since the weighted-integral form (2.108) does not include any of the specified (either essential or natural) boundary conditions, we must also require uy in (2.106) to satisfy all specified boundary conditions of the problem, Consequently, go is required to satisfy all specified boundary conditions, and %, are required to satisfy the homogeneous form of al specified boundary conditions of the problem. These requirements on gy and 4 will increase “the order of the polynomial expressions used for the wweighted-residual method. In general, the g, used in this method are higher-order functions than those used in the Rayleigh-Ritz method, and the functions used in the latter may not satisfy the continuity (j.e. differentiability) requirements of the weighted-residual method. Various special cases of the ‘weighted-residual method are discussed in the following paragraphs. ‘THE PETROV-GALERKIN METHOD. ‘The weighted-residual method is re- ferred to as the Peirov-Galerkin method when y,% $,, When the operator A is linear, (2.108) can be simplified to the form E[[ vanes |e=[ wr—acootacay(54 pRELINGNARIES or x SAy=F (2.109) fs ‘Note that the coefficient matrix [A] is not symmetric: Ay= | WAG) dedy tay e110) ‘THE GALERKIN METHOD. For the choice of weight function , equal to the approximation function ¢, the weighted-residual method is better known as the Galerkin method, The algebraic equations of the Galerkin approximation are @.1la) where A= [oatendsdy, A= { olf -ac@olded 2.1110) Once again, we note that Ay is not symmetec. In general, the Galerkin method is not the same as the Rayleigh~Ritz method. This should be clear from the fact that the former uses the weighted-integral form whereas the latter uses the weak (or variational) form to determine the coefficients cj. Consequently, the approximation functions used in the Galerkin method are required to be of higher order than those in the Rayleigh-Ritz method, If the equation permits, and one wishes, the differentiation can be transferred from the solution w to the weight function w = @; and one thereby obtains the weak form to relax the continuity requirements on the approxima- tion functions and include the specified natural boundary conditions of the problem. “The Rayleigh-Ritz. and Galerkin methods yield the same solutions in two cases; (i) when the specified boundary conditions of the problem are all of the ‘essential type, and therefore the requirements on @, in the two methods become the same and the weighted-integral form reduces to the weak form; and (ii) when the approximation functions of the Galerkin method are used in the Rayleigh-Ritz method, The reader is urged to keep the distinction between the Rayleigh-Ritz and Galerkin methods in mind, THE LEAST-SQUARES METHOD. In this method, we determine the para- rieters ¢; by minimizing the integral of the square of the residual (2.107): a EL Reo aera =o[RCTEGRAL FORMULATIONS AND VARIATIONAL METHODS SS; or aR Ga Ravay=0 (2.1128) Comparison of (2.112) with (2.108) shows that y= 9R/ae. If A.is a linear ‘operator, Y= A(¢), and (2.1122) becomes BU aconacey ae ay]e= [ AC@att -Aleolee dy or Saget e128) i where 4a=[a(enace) des R= [ AC@atE —AC¢)] dey 1120) Note that the coefficient matrix Ay is symmetric, but it involves the same order of differentiation as in the governing differential equation, ‘THE COLLOCATION METHOD. In the collocation method, we seek an approximate solution uy to (2.103) in the form of (2.106) by requiring the residual in the equation to be identically zero at N selected points x! = (x', y') (=1,2,...,.N) in the domain @: Re‘ yg)=0 G=1,2,...,N) (2.113) The selection of the points x‘ is crucial in obtaining a well-conditioned system of equations and ultimately in obtaining an accurate solution. The collocation method can be shown to be a special case of (2.108) with y= 8(x—x'), where (x) is the Dirae delta function, which is defined by [s0v80-H ara =1@) en With this choice of weight fun ns, the weighted-tesidual statement becomes : f o0-29R (4 6) ardy =0 or RG, G)=0 (2.115) We consider an example to illustrate the use of various cases of the weighted-residual method.56 raeunananues Consider the differential equation [see Example 2.4 with Set 2 boundary ~Ginus=0, w=0 wl @u6) For a weighted-tesidual method, gp and g; should satisfy the following condi #0)=0, Gi{1)=1 (satisfy actual boundary conditions) (0) =0, ${(1)=0 (satisfy homogeneous form of the specified boundary conditions) For a choice of algebraic polynomials, we assume ga(x)=a+ bx and use the two conditions on g, to determine the constants a and b, We obtain Ole) Since there sre two homogeneous conditions, we must assume at least a three- parameter polynomial to obtain a nonzero function, y=, bx+ cx’, Using the conditions on @,, we obtain $.=-er@—2) ‘The constant c can be set equal to unity because it will be absorbed into the parameter cy. For $5, we can assume one of the forms: dratbrtd? of geatattde with d#0; 45 docs not contain alhorder terms in either case, but the approximate Solution is complete because {¢,, 4.) contains all terms up to degree three. For the fust choice of g2, we obtain stead b2-a%(1— Br) a inthe approximation of th equation is n= -(04 S08) (aoe Sew) so(2-2e tx) 4e(-2+4r +B) ent ‘The x ‘We next consider various methods, ‘The Petrov-Gulerkin method. Let the weight functions be Wer ee @.us) ‘Thea or Bet Be-h=0, Herta -b=0 @.119) Solving for ¢,, we obtain c, = 8 and c, = ~ 28; the solution becomes lupe = 1.302053r ~ 0.173021x* ~ 0.014663? @.120)DDITEGRAL FORMULATIONS AND VARIATIONAL METHODS S77 “the Galerkin method. Taking y= gy, we have [xe-nrac~o, [va-sorae=0 fet Ba-b=0, let Be k<0 e121) Hence, the solution becomes (with ¢, = 885, ¢3 = ala), tg = 1.2894 ~0,1398x* — 0.00325"? 12) “The least-squares method. Taking i= @R/ac,, we have *-26+x)Rde 4 “Q—4e ta" WIRA l l Be-Be-B=0, et Bath=0 ey “The least-squares approximation of (2.116) is given by (with c,— 8 and cy ~ 3) tts = 1.2601x ~ 0,08017x* — 0.033252? (2,124) ‘The collocation method. Choosing the points x= and x=4 as the collocation wwe evaluate the residuals at these points and set them equal to zero: Ife ~ 6c, = 18 900, + 34e, = 18 (2.125) “The solution is given by (cy = #29 and c= #8) We =1.3612x ~ 0.129273? ~ 0.03422" @.126) ‘The four approximate solutions are compared in Table 2.3 with the exact solution (2.84), For this problem, the Petrov-Galerkin method gives the most accurate solution. 2.5 SUMMARY In this chapter, we have studied two major topics that are of immediate interest in the study of the finite element method in the forthcoming chapters 1, Weighted-integral and weak formulations of differential equations 2, Solution of boundary value problems by the Rayleigh-Ritz and weighted- residual (¢.g., the Galerkin, least-squares, and collocation) methods ‘The weighted-integral statements are required in order to generate the necessary and sufficient number of algebraic equations to solve for the parameters ¢; in the approximate solution, Thus the algebraic equations are equivalent to minimizing the error introduced in the approximation of the differential equation in a weighted-integral sense, In studying the two topics, a three-step procedure for developing the ‘weak form of a differential equation is presented, and procedures for obtaining58° pexnavanies TABLE, Comparison of the Rayleigh-Ritz, weighted-esidual, and exact solutions of the boundary value problem in @.116) Solution, u(x)t x te tine roots Me 0 6.0000 0.0000 0.0000 0.0000 9.0009 0.0000 O41 0.1262 0.1280 0.1285 0.1275 0.1252 0.1348, 02 02513 02529 0.2535 0.2523 02185 02668 03 0378 039 03751 037k 0.3659 0.2958 04 04963 04038 04941 0.4932 O48 05216 05 06112 0.6097 0.6096 0.6003 0.6058 0.6440 06 (OTH 0.726 O72 07m 9.7 0.708 07 O8M0 08324 08317 O82) ORBLE 0.8718, 08 09402 0.9393 9384 O.9W4 0.9207 0.9887, 09 1.0833 LOL LOM Lous 1L0m9 Lost Oo 189 L497 46s L467 1.1977 ‘Sobcripts ato as follows: RR, Reyleigh-Ritz; PO, Pettov-Galekin; 6, Galerkin; 1, lest squares C cllocaton algebraic equations in terms of the unknown parameters of the approximate solution are developed. These topics are immediately applicable in the finite clement method, which is a piecewise (or element-wise) application of a variational method. Thus, the material covered in this chapter constitutes the heart of the finite element method. A few remarks are in order on the variational methods of approximation studied here. The traditional variational methods (e.g., the Rayleigh-Ritz, Galerkin, and least-squates) presented in Section 2.4 provide a simple means of finding lly continuous approximate solutions to physical problems. The ap. proximate solutions obtained via these methods are continuous functions of position in the domain. ‘The main disadvantage, from the practical point of view, of variational methods that prevents them from being competitive with traditional finite difference methods is the difficulty encountered in selecting the approximation functions. Apart from the properties the functions are required to satisfy, there exists no unique procedure for constructing them. The selection process becomes more dificult or even impossible when the domain is geometrically complex and/or the boundary conditions are complicated. From the preceding discussion, it is apparent that the variational methods can provide a powerful means of finding approximate solutions—provided one can find a way to systematically construct approximation functions, for almost any geometry, that depend only on the differential equation being solved and not on the boundary conditions of the problem. This property enables one to develop a compater program for a particular class of problem (each problem in the class differs from the others only in the data), that is, a general-purpose[INTEGRAL FORMULATIONS AND VARIATIONAL MetHoDs 59 computer program. Since the functions must be constructed for a geometrically complex domain, it seems that (recall the discussion of the method of composites for the determination of the center of mass of an irregular shape from Chapter 1) the region must be represented (or approximated if required) fas an assemblage of simple geometric shapes for which the construction of approximation functions becomes simpler. ‘The finite element method to be discussed in the forthcoming chapters is based on these ideas. In this method, a given domain is represented (discretized) by a collection of geometrically simple shapes (elements), and on each element of the collection, the governing equation is formulated using any one of the variational methods. The approximation functions are systematically generated for cach (typical) element using the essential boundary conditions. ‘The elements are connected together by imposing the continuity of the dependent variables across the interelement boundaries. The remaining chapters of this book are devoted to the introduction of the finite element method and its use in the analysis of several model differential equations representing mathemati- cal models for many physical processes. PROBLEMS Secs 21-23 tn Problems 21-29, const the weak fom ang, whenter pose, quate 2s Ove denon ht cndecton omen: 4 (a h(t sanmq tor det wom [etfene-na]] =a. eet ‘where @ and q are functions of x, ang By 6, tay and Qp are constants, 2.2, Beam on elastic foundation: éw Gan) Mea where b = El and fare functions of x, and k is a constant (foundation modulus) 23. Longitudinal deformation of a bar with an end spring: -£ (=a tr ocect (o$f +40) where a and q are functions of s, and & and P are constants.60 PReLnanantes: 24 The Tnasento(erforabe bean ery ~aloa(Z+)]-1 -£ (or) + cxalitsy) 0 w(0) = w(L) =0, (a) for 0 -ysoes OH “¢ 389 BJo VonEIaO}" [RY “z 21989 © Jo woROgRP SHIOASUEAL E 0 aterm rpms ssroxsuen pamngunsiq, 2jqe> WY Woo, won2ayop aeroASUEAE, 6 > = smuay sommes ayqeyes Grew Pe °O =7-"|(xp/np 2) swomspuos Arepunog feampenl “Ont = °-“|m1 :worrpuos Azepanog jenUassT, T>x>0 105 b= (Fo) Pe 3B) P "p UuoSsuaUIEp auo uy suopenbs 29p10-puco9s amp Jo sopdurexs owiog re aTavL"TR PINTe ELEMENT ANALYSIS OF ONE-DIMENSIONAL PROBLEMS when c(x)=0. The mathematical structure common to apparently different fields is brought out in this table. Thus, if we can develop a numerical procedure by which (3.1) can be solved for all possible boundary conditions, the procedure can be used to solve all field problems listed in Table 3.2, as well as many others. This fact provides us with the motivation to use (3.1) as the ‘model second-order equation in one dimension. A step-by-step procedute (see Table 3.1) for the formulation and solution of (3.1) by the finite element method is presented next. 3.2.2 Discretization of the Domain ‘The domain of the problem in the present case consists of all points between and x = L: Q= (0, L); see Fig. 3.1(b). ‘The domain Q is divided into a set of line elements, a typical element being of length h, and located between points A and B, The collection of such elements is called the finite element ‘mesh of the domain (see Fig. 3.1c). The reason for dividing the domain into finite clements is twofold: first, to represent the geometty of the domain; and, second, to approximate the solution over each element of the mesh in order to better represent the solution over the entire domain, Approximation of the domain in the present case is not a concern, since it is a straight line. If the domain is a curve then approximation by a set of straight or curved line elements is necessary to represent it. Approximation of the solution over each element of the mesh is simpler than its approximation over the entire domain. Recall that in the traditional variational methods, the solution is required to satisfy the boundary conditions of the problem, This places severe restrictions con the choice of approximation functions, especially when discontinuities exist in the geometry, material properties, and/or loading of the problem (see Chapter 2 for details) ‘To connect the elements and impose continuity of the solution at nodes common to elements, we identify the endpoints of each line element as the element nodes. Depending on the degree of polynomial approximation used to represent the solution, additional nodes may be identified inside the element. The nodes play the role of interpolation points, as will be seen shortly, in constructing the approximation functions over an element. "The number of elements used in a problem depends mainly on the ‘element type and accuracy desired. Whenever a problem is solved by the finite clement method for the first time, one is required to investigate the convergence characteristics of the finite element approximation by gradually refining the mesh ({.e., increasing the number of elements) and comparing the solution with those obtained by higher-order elements. The order of an clement refers to the degree of polynomial used to represent the solution over the element. 3.2.3 Derivation of Element Equations ‘The derivation of finite element equations, i.c., algebraic equations that relate the primary variables to the secondary variables at the nodes of the elements,FINITE ELEMENT ANALYSIS OF ONE-DIMENSIONAL PROBLEMS 73 involves three steps: 1. Construct the weighted-residual or weak form of the differential equation. 2. Assume the form of the approximate solution over a typical finite element. 3. Derive the finite element equations by substituting the approximate solution into the weighted-residual or weak form. A typical clement 9° = (%4, x2), whose endpoints have the coordinates x=x, and x=Xg, is isolated from the mesh (see Fig. 3.2a), We seck an approximate solution to the governing differential equation over the element, using the Raylcigh-Ritz method discussed in Chapter 2. In principle, any method that allows the derivation of necessary algebraic relations among the ‘nodal values of the dependent variable can be used. In this book we develop the algebraic equations using the Rayleigh-Ritz method, which is based on the weak form of the differential equation, The equations resulting from the application of a variational method are relations between the primary variables (i.e., those involved in the specification of the essential boundary conditions) and the secondary variables (j.¢., those involved in the specification of the natural boundary conditions). The three steps in the derivation of finite element equations of a typical element of the mesh are discussed next. STEP 1: WEAK FORM. In the finite element method, we seck an approximate solution to (3.1) over each finite element. ‘The polynomial approximation of aa) = uf ley) = of a2), SA a6) w FIGURE 32. Finite element dseretization of » one-dimensional donsain for the model problem in (31). (@) A ‘ypical finite clement from the finite element mesh in Fig. 3.1(@); x= global coordinate, #-local coordinate. (6) A typical element, with the definition of the primary (u) and secondary (Q) variables at the element nodes,"TA. vnurs ELEMENT ANALYSIS OF ONE-DIMENSIONAL FROBLEDS the solution within a typical finite element * is of the form usaf) 3.3) where uf are the values of the solution at the nodes of the finite element and {yj ate the approximation functions over the element. The coefficients uf are Qeeermined such that (3.1) is satisfied in a weighted-integral sense. As Gjscussed in Chapter 2 (which should be consulted for additional details), the necessary and sufficient number of algebraic relations among the uj can be obtained by recasting the differential equation (3.1) in 2 weighted-integral form: af af-Zleft)vord]ae ( @ where w(x) denotes the weight function and Q* = (x4, x5) is the domain of a typical element (see Fig. 3.2a). For u =U" and each independent choice of w, wwe obtain an independent algebraic equation relating all uf of the element. A total of 1 independent equations are required to solve for n values wf. When Ww js selected to be yf and (3.4) is used to obtain the ith equation of the required n equations, the resulting finite clement model (i.e., system of algebraic equations among the nodal values) is termed the Galerkin finite element model. Since (3.4) contains the second derivative of U", the approximation functions ‘vf mast be twice differentiable. In addition, if the secondary variables are to be included in the model, ~pf must be at least cubic. Similar arguments apply for cases of the weighted-residual methods discussed in Chapter 2. For additional details of the weighted residual finite clement models, see Reddy (1986) and Chapter 14. "To weaken the continuity required of the functions yj(x), we trade the differentiation in (3.4) from w to w such that both « and w are differentiated equally, once each in the present case. The resulting integral form is termed the weak form of (3.1). This form is not only equivalent to (3.1) but it also contains the natural boundary conditions of the problem. The three-step procedure of constructing the weak form of (3.1) was presented in Chapter 2, and is revisited in the next few paragraphs. “The first step is to multiply the governing differential equation with a ‘weight function w and integrate over a typical element. The second step is to trade differentiation from u to w, using integration by parts. This is achieved as follows. Consider the identity Leds) ott on ‘which is simply the produet rule of differentiation applied to the produet of two functions, a du/de and w. Integrating this identity over the element domain,SECOND.ORDER BOUNDARY VALUE PROBLEMS 75 we obsn Ja [ Sle GE) ae + fo Ea = [oeits fiattar si ‘Substituting (3.56) into (3.4), we arrive at the result [ef. (2.34)] 0 (eR eon) ac— |e)” oo ‘The third and last step is to identify the primary and secondary variables of the variational (or weak) form, This requires us to classify the boundary conditions of cach differential equation into essential (or geometric) and natural (ot force) boundary conditions. The classification is made uniquely by examining the boundary term appearing in the weak form (3.6), (wail, a Asa rule, the coefficient of the weight function in the boundary expression is called the secondary variable, and its specification constitutes the natural boundary condition. The dependent unknown in the same form as the weight function in the boundary expression is termed the primary variable, and its specification constitutes the essential boundary condition. For the model ‘equation at hand, the primary and secondary variables are In writing the final form of the variational (or weak) statement, we assume that all boundary conditions at the element level are of the natural type, so that they can be included in the variational statement: (Bl, e649 ‘The primary and secondary variables at the nodes are shown on the typical element in Fig. 3.2(b). Students of engineering recognize that this figure is the free-body diagram of the typical element, with its internal forces reactions) Qf and (5. The quantities QF = Q, and Q$=Q, have the meaning of forces in the axial deformation of bars; Qf is a compressive force while Q5 is a tensile force (algebraically, both are positive, as shown in Fig. 3.26). For heat conduction problems Qf and Q§ denote the heats conducting info the body. The arrow on the second node should be reversed for heat transfer problems, because the Fourier heat conduction law relating the gradient of tomperature to the heat flux contains a negative sign (implying heat flows from -2 B7276 ruse ELEMENT ANALYSIS OF ONEDIMENSIONAL FROBLESS hot to cold). For additional details on heat transfer, see Section 3.3.1. With the notation in (3.7), the variational form becomes w(xn)Qo G.8) ‘This completes the three-step procedure of constructing the weak fon ‘The weak form in (3.8) contains two types of expressions: those containing both w and u; and those containing only w. We group the former type into a single expression, called the bilinear form: Bw, u) “[ (ee em) dx (3.94) We denote all terms containing only w (but not «) by I(w), called the linear form: Iw) = c wg de + w(64)Qa-+ wen) Qe (2.90) ‘The variational statement (3.8) can now be expressed as Biw, u) =O) G.10) which is called the variational problem associated with (3.1). As will be seen later, the bilinear form results directly in the element coefficient matrix, and the linear form leads to the right-hand-side column vector of the finite element equations, Those who have a background in applied mathématics or solid and structural mechanics will appreciate the fact that the variational problem (3.10) is nothing but the statement of the minimization of a quadratic functional or of total potential energy 1(u): 61=0 where 6 is the variational symbol (see Section 2.3.3) and Tis the quadratic funetional defined by [see (2.436)] Iu) = 3B(u, w) ~ Hu) Gary Equation (3.11) holds only when (x) is linear in ts, and B(W, u) is bilinear and symmetric in wand w, Bw, u) = Blu, w) ‘When (3.1) deseribes the axial deformation of a bar, 4B(u, u) represents the clastic strain energy stored in the bar, {(u) represents the work done by applied forces, and [(u) represents the total potential energy of the bar clement. It is important to note that finite element formulations do not require the existence of the functional 1(u). What is needed is a way to obtain exactly rm algebraic equations among the uf of (3.3) such that the governing differential equation is satisfied over the element in some meaningful way. In the presentSECOND-ORDER BOUNDARY VALUE PROBLENS 77 study, we use the Weak form of the differential equation, ie., (3.8) or (3.10), and the Rayleigh—Ritz method to obtain the n algebraic equations among the nodal variables uf and OF. STEP 2: APPROXIMATION OF THE SOLUTION, Recall that the weak form fover an element is equivalent to the differential equation and the natural boundary conditions of the element. The essential boundary conditions of the clement, Say U(X4)=2q and u(t») = ts, are not included in the weak form. Hence, ‘they must be included in the approximation of u(x). Thus, the approximation of u(x) must be an interpolant, i, must be equal to tg at x4 and uy at X,. Since the weak form contains the first-order derivatives of u, any continuous function would be a candidate for the finite element solution. Let us denote the finite clement solution over element 9° = (x4, X) by US Then wwe seck the approximate solution U* in the form of algebraic polynomials. The reason for this choice is twofold: first, the interpolation theory of numerical ‘analysis can be used to develop the approximation functions systematically ‘over an clement; second, numerical evaluation of integrals of algebraic polynomials is easy. ‘As in variational methods, the approximation solution U* must fulfill certain requirements in order that it be convergent to the actual solution w as the number of elements is increased. These arc: 4, The approximate solution should be continuous over the ele- ment, and differentiable, as required by the weak form. 2, It should be a complete polynomial, i., include all lower-order terms up to the highest order used. 43, Tt should be an interpolant of the primary variables at the nodes of the finite element, @.12) “The reason for the first requirement is obvious; it ensures a nonzero coefficient matrix. ‘The second requirement is necessary in order to capture all possible states, ie., constant, linear and so on, of the actual solution. For example, if a linear polynomial without the constant term is used to represent the tempera ture distribution in a one-dimensional system, the approximate solution can never be able to represent a uniform state of temperature in the element. ‘The third requirement is necessary in order to satisfy the essential boundary conditions of the element and to enforce continuity of the primary variables at points common to several elements. For the variational statement at hand, the minimum polynomial order is linear. A complete linear polynomial is of the form Ur=atbe G13) where a and b are constants. This expression meets the first two requirements in (3.12). To satisfy the third Ua) i» U"(en) (14)78 vintre HLENENT ANALYSIS OF ONE-DIMENSIONAL FROBLENS we express the constants and in (3.13) in terms of u{ and w§, Equations (3.14) provide two relations between (@, b) and (uf, u3): ulsatbxy +bxp af) xalfa (Lit) ease) Inverting (3.15b) by Cramer’s rule, we obtain (a a 5 Xp | fal 1 ty where he=p—X4 and a= (xj, B=(-1S Hite, HEX In (3.15d), fand j permute in a natural order: (B.15a) oor, in matcix form, 1 x vie) Cats + a8) 1 “oe G.15c) us ui) = Bit + Bis) 5 te 1 uf “lt us| he x» @.lSd) if G=1 then j=2; if i=2 then j=1 ‘The a7 and ff are introduced to show the typical form of the interpolation functions, Substitution of (3.15c) into (3.13) yields 1 U(x) = 5 [erius + aud) + (Bias + Baus] =o + Bin + (08+ Bos Thats UG)= vibes + wsbME= 3 HE 6.162) where WG) plots py =Z2*, ya) =F Catt By = G.165) which are called the linear finite element approximation functions. For the linéar interpolation (3.16), we label the endpoints as nodes 1 and 2, and rename the secondary variables as Qn= 25 On=O G7SUCOND-ORDER HOUNDARY VALUE pROSLEMS 79 ‘The global node numbers for elements connected in series can be related to the element node numbers. For linear elements, the global node numbers of clement 2" are e and ¢+1, and the global coordinates of the clement nodes are x, and Xer1 (Kee, Xq =e aNd Xp = p51). ‘Note that the element interpolation functions wf in (3.16b) are expressed in terms of the global coordinate x (Le., the coordinate of the problem), but they are defined only on the element domain 2° = (x4, ¥2)= (te, Xu). If we choose to express them in terms of a coordinate £ with origin fixed at node 1 of the element, #7 of (3.16b) take the forms 3.18) The coordinate £ is termed the local or element coordinate. The functions pf are shown in Fig. 3.3(a). Note that 1f is equal to 1 at node 1 and zero at node 2, and fis equal to 1 at node 2 and zero at node 1. These properties of yf are known as the interpolation properties. ‘The global interpolation functions ®, can be defined in terms of the element interpolation functions corresponding to the global node J (see Fig. Win 1 = ath, ve ah, @ Ors TY } i. 7 Global node numbers Tors. Nat ‘Typical element { WE torn ees WA fore reins o FIGURE 33 (@) Local and (6) global interpolation functions for the two-aode (lines#) element (&4= x,80 Fire ELEMENT ANALYSIS OF ONE-DIMENSIONAL, PROBLEMS 3.36). Since U*(x) of (.16a) is an interpolant of u(x) over the clement 2°, VF are also called interpolation functions. Interpolation functions derived using the dependent unknown—not its derivatives—at the nodes (j.c., interpolation functions with C® continuity) are called the Lagrange family of interpolation junctions. When the dependent unknown and its derivatives at the nodes are ‘used to derive the interpolation functions, the resulting interpolation functions fare known as the Hermite family of interpolation functions (sce the classical ‘beam element in Chapter 4). ‘Note that ‘pf are derived systematically; starting with an assumed degree of algebraic polynomial for the dependent unknown and determining the coefficients of the polynomial in terms of the primary degrees of freedom, we expressed the dependent variable as a linear combination of approximation functions and the primary nodal variables. The key in the procedure is ta select the number and the location of nodes in the element so thaf the geometry of the latter is uniquely defined. The number of nodes must be sufficient to allow the assumed degree of interpolation of the solution in terms of the primary variables, For a linear polynomial approximation, two nodes with one primary Unknown per node arc sufficient to define the geometry of the clement, provided the two nodes are the endpoints of the element. Since a quadratic polynomial is uniquely defined by three parameters, a total of three nodal points must be identified in the element. To define the geometry of the flement, two of the nodes must be the endpoints of the element, The third ean be identified inside the element. Returning to the linear approximation (3.13), which is recast as (3.161), wwe note that the true solution is approximated over each element by a linear polynomial U"(e) (Gee Fig. 3.4a), The error in the approximation, B= u(x) — UG), can be reduced by either decreasing the element size fh, oF increasing the degree of the approximation (see Fig. 3.4). ‘A quadratic approximation is of the form ue at bx tex® G.19) Which requires three nodes in order to rewrite U" in terms of the values of u(x) at the nodes. Two of the nodes are identified as the endpoints of the element to define the geometry, and the third node is taken interior to the element. In theory, the third node can be placed at any interior point. However, the midpoint of the element, being equidistant from the end nodes, is the best choice. Other choices (e-g., quarter-point) are dictated by special considera- tions (e.g., to have a certain degree of singularity in the derivative of the solution). Thus, we identify three nodes in the element of length /. (see Fig. 3.5a) and rewrite U‘(x) in terms of the thrce nodal values, (uj, #5, u5)- We have U(x) = at bt + cet)? at bxs + ces)” (3.200) + bxS + (x5)DANITE ELEMENT ANALYSIS OF ONEDIMENSIONAL PROBLEMS 81 = Fract solution Sielineer-element solution ‘Trrecclinear clement solution ut) ——Linearslement solution Quadratc-clement solution ) o FIGURE 34 Refiemens of finite element solutions; (a) mesh refinement using inear elements; (6) quadratic ‘element solution using three elements. wi) [i a oP ]fa wpa] 1 xt GH ye (3.200) w) Lt xs @s7tle where 27 is the global coordinate of the ith node of the clement 2*, Inverting the above relations, we obtain or, in matrix form, af = x70)? — 407 Bi= (a? — Ge? 21) = -Gi-xi),(82. MITE ELEMENT ANALYSIS OF ONE-DIMENSIONAL PRODLEMS a ae a ees ~2 7-1 Pee ethetemem—7 Let er Dihelement © FIGURE 35 ‘One-dimensional Lagrange quadratic element anc its interpolation fonctions: (a) geometry of the ‘lement; (6) interpolation functions; (c) global interpolation functions corresponding © the ‘quadratic interpoletion functions. Here 7 denotes the global node number, e the clement number, tnd fthe element node number, and (3.19) takes the form Ute) = wiles + vse es + wsledus = 3 vias 6.22) where #f are the quadratic Lagrange interpolation functions, (of+ Bix + ix?) (= 1,2,3) (3.23)SECONDORDER BOUNDARY VALUE PROBLEMS 83 Here D® denotes the determinant of the matrix in (3.20b), and af, i, and yf are defined by (3.21). The subscripts used in (3.21) permute in a natural order: if G1 then j=2 and k=3 if @=2 then j=3 and k=1 3.24) if G=3 then j=1 and k=2 For example, af, iS, and yf are given by v aay ray, B=G- ON, 1 ‘The quadratic interpolation functions can be expressed in terms of a local coordinate £, with origin fixed at node 1, the left end of the element. The global coordinate x is related to the local coordinate # by the relation xomte 3.25) where x=.x, is the global coordinate of the first node of the clement Q*. For a quadratic element with the interior node, node 2, located at ¥= ah, we have vila) (1 wO=a ail) 6.28 150) = Gel aa) where 02, the weak form in @.8) must be modified to include nonzero secondary variables, if any, at interior nodes: o=[" (0GfSE ter) ae [" wg ade 3 mene @.20) where xf is the global coordinate of the /th node of element Q* If nodes 1 and ‘n denote the endpoints of the element then Of and Qs represent the unknown point sources, and all other Qf are always known (j.e., applied point sources), Following the Rayleigh—Ritz procedure developed in Section 2.4.2, we substitute (3.29) for v and Wi, 5, ..., YS for w into the weak form (3.30) to obtain 7 algebraic equations: 0= f° [SE (SB) revel wiviea)— via] ae 3 weenoy ie o= f° [9S (SB) evs vic) - via] ae—3 viene ia (Su ie “B) «cvi(S esto) —wia|ac~ 3 weeny (ith equation) of Es BE) + ons(S suse) — via ae 3 veep Vet (3.31a)‘86 raNITe ELENEKE ANALYSIS OF ONE-DIMENSIONAL FROBLENS Note that the numbering of the algebraic equations follows that of the primary variables in the element. The ith algebraic equation can be written as o= 3 Kgus—ff-OF (@=1,2, 0.57). (3.31b) f im (Te evi) de= D008, wim [" avidr=ttw G.31e) Note that the interpolation property (1) of (3.28) is used to write Zwiene-or (32) Equations (3.31) can be expressed in terms of the coefficients Kj, f%, and Qj as Kiwi + Kiuit...+ Kiwis fit OF Kaui + Kaus +... + Kau =f 3 + Kia. + Kah =f +O (3.332) Kiauh + Kia ts. + Kitt fat Oe Jn matrix notation, the linear algebraic equations (3.332) can be written as IK*Hu"} = {f"} + {O°} (3.336), ‘The matrix [K*] is called the coefficient matrix, or stiffness matrix in structural mechanics applications. The column vector {f°} is the source vector, or force vector in structural mechanics problems. Note that (3.33) contains 2 unknowns: (uf, ws,..., w6) and (Q§, Q5,..., 5), called primary and secon- dary element nodal degrees of freedom; hence, it cannot be solved without having an additional n conditions. Some of these are provided by the boundary conditions and the remainder by balance of the secondary variables Q7 at nodes common to several elements. This balance can be implemented by putting the elements together (i.c., assembling the clement equations). Upon assembly and imposition of boundary conditions, we shall obtain exactly the same number of algebraic equations as the number of unknown primary and secondary degrees of freedom. The ideas underlying the assembly procedure are discussed in the next section. ‘The coefficient matrix [K], which is symmetric, and source vector (f*} can be evaluated for a given element and data (a, ¢, and q). For element-wise~ constant values of a, c, and q (say, de, ce, and q.) the coefficients Kj and f7 can easily be evaluated for a typical element._ SECONDORDER BOUNDARY VALUE PROMLENS 87 Linear element. For a mesh of linear elements, the element Q° is located between the global nodes x4= x, and x»=Ax,44 (see Fig. 3.2). Hence, Ki fe conn) as, = [" aartae or, in the local coordinate system £, “/ dvi i= ["(0 Set 5 coyrei) i= [ganas where x = x84 and dvi due dx de ‘The wf can be expressed in terms of as fsee (3.18) VilX)=1~3/h., WR) =3/h, We can compute Kj and ff by evaluating the integrals. We have wf PBI de= di, wo [PDirb-Dile 4, (by symmetry) ‘Similarly, fe [at-f)dtmtatn s8=[aZaemtacr ‘Thus, for constant q,, the total source qeh, is equally distributed to the two nodes. The coefficient matrix and column vector are Cs rem ad Fa 048 ry ef} (3.340) Ia =a.x and ¢=¢., the coelfcient matrix [K*] can be evaluated as ILE ox88. FMire ELEMENT ANALYSIC OF ONE-DIMENSIONAL FROBLENS FIGURE 3.6 Approximation of an element with linearly varying cross-section by an equivalent clement, vith constant erossection. ‘The reader should verify this. Note that when a is a lincar function of x, this is ‘equivalent to replacing a in the coefficient matrix with its average value compare (3.34a) with (3.35)]: Gang = Ae + Feet )e (3.36) For example, in the study of bars with linearly varying cross-section eagey=#(4,42e4=Ae,) hth this amounts to replacing the varying cross-section with a constant cross- section within cach element, the cross-sectional area of the constant section being the average area of cross section of the linearly varying element (see Fig. 3.6). Here A, denotes the cross-sectional area at x, and A,,, is that at XA test ‘When a, c, and q are algebraic polynomials in x, the evaluation of Kj and f7 is straightforward. When they are complicated functions of x, numerical evaluation of the integrals in [X*] and {f°} will be sought. A complete discussion of the numerical evaluation of integrals is presented in Chapter 7. When a and q are element-wise-constant and c=0, the finite element equations corresponding to the linear element are él 1 he 3 ot (+ (3) (3.374) Me ye Beye ad 7 Fae Mele + OF “a 8.37) Wem AM haat OS Quadratic clement. For a quadratic-clement mesh, the element Q* is located between global nodes x4—2.-1 and Xy = X2ey1+ Hence, Ceeftfsnm) abl rn hy fi viaeds= [vias ae where the Lagrange quadratic interpolation functions pi(e) (SECOND-ORDER BOUNDARY VALUE PROBLEMS 89 given in (3.27). Evaluating the integrals in (3.38), we obtain Cerin ef. H0a(2)']s oan Pf R Dee) elere(Q eae the and 30 on. Similarly, f= [ali ea) je > Poli lo-g Note that, for quadratic elements, the total source q.h, is not distributed equally between the nodes. The distribution is or equivalent to that of two linear elements of lengths }h,, Therefore, the computation of f7 should be based on the interpolation functions of that element. The sum of ff for any element should always be equal to the integral of q(x) over the clement: Lael Q(z) dx 6.39) In summary, for element-wise-constant values of a, c, and q, the clement matrices of a quadratic element are “1 2 (3.402) 7-8 [k}=2+]-8 16 3.400) 5 (by symmetry) & i = = Hel yg 3.2.4 Connectivity of Elements In deriving the element equations, we isolated a typical clement (the eth) from the mesh and formulated the variational problem (or weak form) and(90. rnsTe FLEMIENT ANALYSIS OF ONE-DIMENSIONAL PROBLENS developed its finite element model. To solve the total problem, we must put the elements back into their original positions. In doing this before discretiza- tion, we impose the continuity of the primary variables and balance of the secondary variables at the connecting nodes between clements. Continuity of the primary variables refers here to the single-valued nature of the solution; balance of sccondary variables refers to the equilibrium of point sources at the junction of several elements. Thus, the assembly of elements is carried out by L. Continuity of primary variables at connecting nodes: ugeut? Gla) i.e., the last nodal value of the element Q is the same as the first nodal value of the adjacent element 91, x 2, Balance of secondary variables at connecting nodes: 0 if no external point source is applied 05+ 0% =| Qo ian external point source of magni- ‘tude Qp is applied G.41b) In writing (3.41), it is assumed that elements are connected in a sequence. The continuity of primary variables u3=u{*' and balance of secondary variables Q5+ Of" for a mesh of linear elements is illustrated in Fig. 3.7. The balance of secondary variables can be interpreted as the continuity of adu/dx (not adU"/ds) at the point common to elements Q* and Q"** (when no change in oy Ms FIGURE 3.7 ca oo 05 Assembly of two fincar Lagrange ele- tment: (a) continuity of the primary vari able; (@) balance of the secondary o variables,s8coND.ORDER BOUNDARY VALUE FROBLENS OM adu/ds is imposed externally): (a) or (3.42) 0:4 0F"'=0 6") ‘The interelement continuity of the primary variables i imposed by renaming the two variables uf, and w{*! at x=2y as one and the same, namely the value of w at the global node N: us In 6.43) where V = (n—1)e +1 is the global node number corresponding to node of the element {2 and node 1 of the element Q°*!, For example, for a mesh of E linear finite elements (n= 2), we have w= U; w=ul=U; ub=ul=U; G44) wt sur= Us ub= Ure ‘To enforce balance of the secondary variables Qf, (3.416), it is clear that wwe can set 0% + Of"? equal to zero or a specified value only if we have such ‘expressions in our equations. To obtain such expressions, we must add the nth equation of the element 9° to the first equation of the clement Q**"; that is, wwe add 2 Kyun fat On and SS xsptup ap sor B to give 3 (Big uj PKG UI) = fat fi + (Oe OT) =fEt fi + Qo G.45) This process reduces the number of equations from 2 to E+ 1. The first equation of the first clement and the last equation of the last element will92 rave ELEMENT ANALYSIS OF ONF-DIMERSIONAL PROBLEMS remain unchanged, except for renaming of the —primary variables. The left-hand side of (3.45) can be written in terms of the global nodal values as (Kini + KoauS +. ob KSqu8) + (RGD ET + RST + REST) = (KinUn + KoaUer tos + KinUren-s) + KG Uy ait KGS Wyant + KGS Uyrvan-a) = KinUn + KiaUwert so + Kit Uresn-2 + (Kin KEE Wanna + KEE Upon be A KG Uyennna 46) where N=(n ~ De +1. For a mesh of E linear elements (1 =2), we have KhU, + KinU2=fi+ Qi (unchanged) KAU, + (Kh + K})U2+ KhaUs = [3 +S + O3+O% KAU: + (Kh + KU; + KhUs = f3+ fi + O34 Qi 47a) KECUp-c+ (Ki + Ki Ue + KRU ee =f + FE + OF t+ OF KiU + KaUru =f? +Q§ (unchanged) ‘These are called the assembled equations. They contain the sum of coefficients and source terms at nodes common to two elements. Note that the numbering. of the global equations corresponds to the numbering of the global primary degrees of freedom, U;. This correspondence carries the symmetry of element matrices to the global matrix. Equations (3.474) can be expressed in matrix form as Ki Kh uy Kh Kh+Kh Kh 0 Us Ki Kit Ki Us 0 KE‘+KE KG[] Ue KE KE} Wes, K ot Fathi 3+ OF Gr O84 Oh @.476) Pept] lors oF ff oF Recall that all the above discussion of assembly is based on the assumption that elements are connected in series. In general, several elements, can be connected at a node, and the elements do not have to be consecutively numbered. In that case, the above idea still holds, with the change that coefficients coming from all elements connected at one node will add up. ForSECOND-ORDER BOUNDARY VALUE YRODLEMS 93 Rigid bar (constrained to move Horjzonally) HIGURE 3.8 ‘The geometry and finite clement mesh of a bar 4a hy = hy structure. example, consider the structure consisting of three bar elements shown in Fig. 3.8. Suppose that the connecting bar is rigid ({e., not deformable) and is constrained to remain horizontal at all times. Then the continuity and force balance conditions for the structure are B= Us, O)+Q}+Q3=2P (3.48) To enforce these conditions, we must add the second equation of element 1, the first equation of clement 3, and the second equation of element 2: (Réuei + Kiaua) + (Kinteh + Kigu) + (Kaui + Ku) =fit fit fit Qh+OR+ OF (3.49) We note the following correspondence of local and global nodal values (see Fig. 3.8): weu= w= Uy ub=U;, wbaulaub= Us, wb= Uy Hence, (3.49) becomes KAU, + KAU, + (Kia + Ki, + KR)U, + KRU, =f) + fi +f3+ Q5+ Q2+ 03 fit fit fit2P ‘The other equations remain unchanged, except for renaming of the primary variables. The assembled equations are Kh 0 Kh 07/4 ft at OK, KR Out J oR oe Ky KR [Ul ]AsreAl7 orrorsorf 6%? 0 0 Kh Keilu fi Q3 where K = Kh+ Kh + Kh. ‘The coefficients of the assembled matrix can be obtained directly. We note that the global coefficient K,, is a physical property of the system, relating global node J to global node J. For axial deformation of bars, Ky, denotes the force required at node J to induce a unit displacement at node J, while the displacements at all other nodes are zero. Therefore, Ky, is equal to the sum of94 FINITE ELEMENT ANALYSIS OF ONE-DIMENSIONAL PROBLEMS | all Ki for which { corresponds to I and j corresponds to J, and i and j are the local nodes of the element ©. Thus, if we have a correspondence between clement node numbers and global node numbers then the assembled global coefficients can readily be written in terms of the element coefficients. The correspondence can be expressed through a matrix (B], called the connectivity ‘matrix, whose coefficient by has the following meaning: bby is the globel node number corresponding to the jth node of element { For example, for the structure shown in Fig. 3.8, the matrix (B] is of order 3x2 (3 elements and 2 nodes per element): 13 (BJ=|2 3 34 . ‘This array can be used in a variety of ways—not only for assembly, but also in the computer implementation of finite element computations. The matrix [B] is used to assemble coefficient matrices as follow: Kh=Ky, _ because local node 1 of element 1 corresponds to global node 1 Kh=Kjy, because local nodes 1 and 2 of element 1 correspond to global nodes 1 and 3, respectively and so on, When more than one element is connected at a global node, the clement coefficients are to be added. For example, global node 3 appears in all three rows (.e., elements) of the matrix [B], implying that all three elements ‘are connected at global node 3. More specifically, it indicates that node 2 of clement 1, node 2 of element 2, and node 1 of element 3 are the same as global node 3. Hence Kh+ Kh+ Ki=Kss ‘Assembly on paper can be carried out by examining the finite element mesh of the problem. For the mesh shown in Fig. 3.8, we have Kqy=K%, because global node 2 is the same as node 1 and global node 3 is the same as node 2 of element 2 Ky=0, because global nodes 2 and 4 do not belong to the same element Ky=Kh+Kht+ Ki and so on. ‘Tn summary, assembly of finite elements is carried out by imposing interelement continuity of primary variables and balance of secondary vari- ables [see (3.41)]. Renaming the elemental primary variables in terms of the global primary variables and using the correspondence between the local and global nodes allows the assembly. When certain primary nodal values are not required to be continuous (by the variational formulation) across elements, such variables may be condensed out at the element level before assembling elements.SECOND-ORDER BOUNDARY VALUE PROBLEMS 95 3.2.5 Imposition’ of Boundary Conditions Up to this point, the specific nature of the problem has not been used in the development of the finite element model or in the assembly of finite elements In other words, the discussion. in Sections 3.2.1-3.2.4 is valid for any differential equation that is a special case of the model equation (3.1), One particular probiem differs from others in the specification of the data and boundary conditions. Here we discuss how to impose the boundary conditions of a problem on the assembled set of algebraic equations. To this end, we use the problem in Fig. 3.8, Its boundary conditions are evident from the structure, The known primary degrees of freedom (i.e., displacements) are uJ=U,=0, u}=U;=0, = U,=0 @.sia) ‘The known secondary degrees of freedom (i.e., forces) are O34 024 Q}=2P (3.516) The forces Q1, QF, and OF are unknown (reaction forces), and they can be determined in the post-computation, i.e., after the primary degrees of freedom are determined. Imposing the boundary conditions (3.51) on the assembled system of ‘equations (3.50) and for fi, we obtain Kh 0 Kh oT u=0 f 0 Kh Kp 0 |} u;=0l_Jo% Kh Ky KhtKa+kh Ka |) am [")2e( 6? oo Ka Khllu=0) (03 ‘This contains four equations in four unknowns: Us, Q1, Q2, and Q3 3.2.6 Solution of Equations ‘As a standard procedure in finite element analysis, the unknown primary degrees of freedom are determined first by considering the algebraic equations corresponding to the unknown primary variables. ‘Thus, in the present case, we consider the third equation in (3.52) to solve for Us: KU, + Ki Us+ (Kho + Kia + KU )US + KYU, = 2P- or (Bla + Kha + Ki)Us= 2P ~ (KAU, + KU, + KU) (3.53) Equation (3.53) is called the condensed equation for the unknown Us. The term in parentheses on the right-hand side is zero because all specified displacements are zero in the present problem. Hence, the solution is given by Us=2P (Kin + Kin + Kis) (3.54) ‘The unknown secondary variables are determined by considering the 1g equations of (3.52), ic., those that contain the unknown secondary remai96 rusrre eLessar ANALYSIS OF ONE: variables: of) fxn o Ke oTfe ip=| 0 Ki Kh oO U; 3 3 a, |) U 2) Lo o Kh kad KhUs =1 KAU; }, because Uy, Up and U,ate zero (3.54) KiUs It is possible, although net common with computer programs, to move all the unknowns to the left-hand side in (3.52) and solve for them all at once. But this process requires more computational time in practical problems. ~ ‘in general, the assembled finite element equations can be partitioned conveniently into the following form: : fe) pee) _ Line} veto = Ley} 6.55) where (U!) is the column of known primary variables, (U?) is the column of Unknown primary variables, (F!) is the column of unknown secondary variables, and (F?) is the column of known secondary variables. Writing (3.55) as two matrix equations, we obtain [KU] } + [KP 07} = (FY (3.564) {(K2](U") + [K?(U7} = (F?} (3.565) From (3.56), we have {U?) = (KPC?) = (XU) G.56e) Once (U?} is known, (F*} can be computed from (3.562). 3.2.7. Postprocessing of the Solution ‘The solution of the finite clement equations gives the nodal values of the primary unknown (¢.g., displacement, velocity, or temperature), Postprocess- ing of the results includes one ot more of the following: 1, Computation of any secondary variables (e.g.. the gradient of the solution), 2. Interpretation of the results to check whether the solution makes sense (an understanding of the physical process and expericnce are the guides when other solutions are not available for comparison). 3. ‘Tabular and/or graphical presentation of the results.SECOND-ORDER BOUNDARY VALUE PROBLEMS 97 ‘To determine the solution u as a continuous function of position x, we return to the approximation (3.29) over each element: uI@)=S uve cr ataynq FOZ VIO) G57) ore = S wryie where A is the number of elements in the mesh. Depending on the value of x, the corresponding element equation from (3.57) is used. The derivative of the solution is obtained by differentiating (3.57): mde 3.58) aus due de de ‘The derivative calculated from different elements meeting at a node is always iscontinuous in all C° approximations (i.e., approximations in which only the function values are interpolated), unless the approximate solution coincides with the actual solution, The secondary variables Qj can be computed in two different ways. In (3.545), we determined the unknown secondary variables Qj, Q3, and 0} from the assembled equations of the problem in Fig. 3.8. Since the assembled equations often represent the equilibrium relations of a system, the Qf computed from them will -be denoted by (QjJequi. The QF can also be determined using the definitions in (3.7), replacing u with U. We shall denote QF computed in this way by (Qfeer Since (QF)eer are calculated using the approximate U', they are not as accurate as (Qj)equy. However, in finite clement computer codes, (Qi)ase are calculated instead of (Qieqa This is primarily because of computational aspects. Recall that, in arriving at the98 nite ELEMENT ANALYSIS OF ONE-DIMENSIONAL PROBLEMS result (3.54b), we used part of the assembled coefficient matrix. In the numerical solution of simultaneous algebraic equations in a computer, the original assembled coefficient matrix is often modified, and therefore the coefficients needed for the determination of the secondary variables are not available, unless they are saved in an additional array. For the problem in Fig. 3.8, we have dua (HAD )L =~ me dua-=~( (@da=(24)| hye . EAU. xy, , 659) where ty and hy are the lengths of elements 1 and 3, respectively. ‘The Qs computed using the definitions (3.7) are the same as those derived from the assembled equations for the problem in Fig. 3.8. This equality is not fo be expected in general. In fact, when the source veetor q is not zero, the secondary variables computed from the definitions (3.7) will be in ‘etror compared with those computed from the assembled equations. The error decreases as the number of elements or the degree of interpolation is increased. This completes the basic steps involved in the finite element analysis of the model equation (3.1). A few remarks are in order on the steps described above for the model equation, Remark 1, Although the Raylcigh-Ritz method was used to set up the element equations, any other method, such as a weighted-residual (e.g., the least-squares or Galerkin) method, could be used. Remark 2, Steps 1-6 (see Table 3.1) are common for any problem, The derivation of interpolation functions depends only on the element geometry, and on the number and position of nodes in the element, The number of nodes in the element and the degree of approximation used are related. Remark 3. The finite element equations (3.31) are derived for the linear operator equation A@)=q, where A= Hence, they are valid for any physical problem that is described by the operator equation A(u)=4 or its special cases. One need only interpret the uantitics appropriately. Examples of problems described by this operator areSECOND.ONDER BOUNDARY VALUE PROBLEMS 99 listed in Table 3.2. Thus, a computer program written for the finite element analysis of (3.1) can be used to analyze any of the problems in this table. Also, note that the data a=a(x), ¢= (x), and q=q(x) can be different in each of the elements. Remark 4. Integration of the element matrices in (3.31c) can be implemented ‘on a computer using numerical integration (see Chapter 7). When these integrals are algebraically complicated, one has no other choice but numerical integration. Remark 5. As noted in (3.48) and (3.516), the point sources at the nodes are included in the finite element model via the balance of sources at the nodes. ‘Thus, in constructing finite element meshes, one should include nodes at the locations of point sources. If a point source does not accur at a node, it is possible to “distribute” it to the element nodes. Let Qo denote a point source at point x», 4 (uri) = 3 vee) w where ,(x) (11, 2, 3,4) are the piecewise-continuotis global interpolation functions, EPG) for tx SH) WG) — for sx = 2101 oa)={ o Substituting (iv) for « and v = ©, into ( o= [felt (S022) -aa]er—oee. , we obtainSECOND.ORDER BOUNDARY VATE PRonLEMS 103 1e &; is nonzero only between x, and x;,,, the integral becomes LOR Ge Os Gin BE) ea] tea 60906 619 and we have (or a thre lement nh) ress of felt (A2s ult) ng] as—esen04 fat (ut Sts va ) eg] e090 ers ott su) ag] er-ogon c (iii) “le afb (y, ey sai). ea] de ~@(3)Q6 ‘These equations, upon performing the integrations, yield (3.47), with the last column Gontaining Qs) in the latter replaced by 0 Qo 0 () 0 Although this procedure gives the assembled equations directly it is algebraically complicated (especialy for two-dimensional problems) and not amenable to simple ‘computer implementation. 3.2.8 Radially Symmetric Problems ‘The equations governing physical processes in cylindrical or spherical ge- ometries are described analytically in terms of cylindrical or spherical coordinates. When the geometry, loading, and boundary conditions are dependent only on the radial direction and independent of the other two coordinates, the governing equations are one-dimensional. ‘The equations governing radially symmetric problems in cylindrical geometries are of the form {an analogue of (3.1)} ld te a ]- (0) for RAq + PBT., we can immediately write the finite element model of (3.77) from (3.31): (KE) = (97) + (O°) G.8la) where = c (ca WMG. ppysy))ae, £1 de ds h, . @ ¢ ai-(-‘F) (2) L. where Qf and Q5 denote heat flow into the element at the nodes. Tuations (3:78) and (3:79) are also special cages of the model boundary value problem, However, in developing the weak forms of (3.78) and (3.79), {he integration mast be carried over a typical volume element of each system, te lstrated in Section 3.2.8 for a radially symmetric cylindrical problem. The wreak form of (3.78) is given in (3.65), and the finite clement model is given by 5) with (u") = (7°), Similarly, the weak form of (3.79) can be developed Using a volume element of a sphere: [ vicag + rorya (810) a AV ardrdody for 0<0<2n, 0<$<20 ‘The weak form of (8.79) is 2 ( (dnd a 8 " om any [” (KE gw)r ar Orw(ra) Oia) 82) and the finite element model of (3.79) is [KVTY = (F594 (29 (3.83)SECOND ORDER BOUNDARY VALUE PROBLEMS 109 where Kj= Qn)? [tate as, fi tr dr af -emr(reZ) = ay (v«D) In the following examples, we consider some typical applications of the finite element models (3.81) and (3.66). (3.836) Example 3.1, Consider a slab of thickness L and constant thermal conductivity & (Wm !C~), Suppose that encrgy at a uniform rate of go (W m™) is gencrated in the ‘wall, We wish to determine the temperature distribution in the wall when the boundary surfaces of the wall are subject to the following three different sets of boundary conditions: sti TO=T, Th)=h (3.84a) sar (eZ) men cwars, [eEeaer-9]] 846) sas T=, (KGB) =e (Wo) G84) ‘The governing differential equation for this problem is given by (3.77) with Hence, the finite element model in (3.81) is applicable here. We must select the order ‘of approximation (or type of element) to evaluate the coeficients Kj and ft in (3.81) For the choice of linear elements and the data a=kA™=constant and constant, (3,812) takes the form [see (3.37@)] RAP 1 1/73) _ Agohef1} , (0%) welt alle} “S*C} (23) For a uniform mesh of N elements, ie., a mesh of same-size elements, dyshyoe++=LIN=h, the essembled equations are 1-t u, ama Offa wap nto 6, M4). |) ¢ 0 in al] ou =a] | ty 1 a l+i + OF gn frei, | +08 + 3.85)110 wire eLeMer ANALYSIS OF ONE-DIMENSIONAL PRODLEMS where Up ((=1,2)...,N1) denotes the temperature at global node 1. The tssembled equations (3.85) are valid for al three sets of boundary conditions in @.84). We now specialize the finite element equations (3.85) for each set of boundary conditions. Set 1. The boundary conditions (3.84a) imply that UT, Uno h (G.86a) and the balance ofthe heats at global nodes 2, 3, ..., NV requires OF'+Qi=0 for e=2,3,...,N GB.86b) “The condensed equations for the unknown primary variables Us, Us Uy are cobtsined by considering the 2nd, 3rd... , Nth equations of (3.85) (and omitting those onesponding tothe known temperatres): 2-1 9 Tr) 42 0 kA Ml : en 0 22 ° “12 -1, ‘The unknown secondary variables Q} and OY are computed from the first and last of the assembled equations (3.85): (Oaqa= baad AYU), Saga back +E Urns Us)_ 880) ‘Tey can also be computed from the definition (3.816): (oow=-s@)-taSin( Hern (3.886) acm (eag] wna rE) =P arr) ‘The values computed in (3.885) are in error compared with those in (3.882) by the nodal sources, igghA, for k= LIN. For any number of elements, the solution can be computed from (3.87). For this set of boundary conditions and ‘a linearelement mesh, the minimum number of ~ i For N=2(h=4L), kA q+ (E-T) ra olf |. 2 ee o- ole and the solution is ut eset (Dew kA . GA GT, (ODSeCOND-ONDER BOUNDARY VaLuE pReaLEMS 111 For N=3 (= SL), 1-10 6 oF KAP -1 0 2 -1 0 o-1 2-1 “yo oon 1 @ oe let alla}-ae{) CG) ri lode leo ie, % ‘The solution is given by alt sory ail Ly MH sen+n), Wass +2n) u (Oda Hh + ECT = 1), DBs Rash +B, ny The exact solution of (3.77) (with € @.84a), is }, subject to the boundary conditions re [F_() Jamison Oe) tat ie 1) Note that the finite clement solution at the nodes coincides with the exact solution for ‘any number of elements. In fact, for constant a= kA, it can be shown that the finite ‘clement solution at the nodes is exact [see Reddy (1986), pp. 403, 404}; the only error ‘would be that due to the round-off error introduced in electronic computations. The finite element solution, being linear, will not be exact at points other than the nodes. If we use quadratic elements, we ean improve the solution at points between the nodes, Let us consider a mesh of one quadratic element. The element coeficient matrix and source vector are given in (3.40). We have (ce= @__atA(, 32) Mg where h =. The solution is Aggh 3h 6 Ak aw ++ 5) hn + 2) (Qieawn = ~AQoh +a, EB) (Qian = —HAgoh +a n)112 axrre ELEMENT ANALYSIS OF ONE.DIMENSIONAL PROBLESS If-we compute the QF using the definition (3.810), we obtain orm (ea) (B)| =-lo pg) aG-e) OCG), = Lael +a, “2 which coincides with the exact value. Recall that (Q!)aq for a Finear-element mesh does not coincide with the exact value. Also, the values of T(x) at x = 42, for example, from the two linear- and one quadratic-clement meshes can be computed using the interpolations ren = ie ‘ 4AGT+7) (linear element) ret) 3 E4404) (quadratic element) ‘The quadratic interpolation gives the exact value, Thus, the finite clement solution given by the quadratic element is exact at all points, because the exact solution varies ‘quadratically. Set 2, In this case, the surface at x=0 is subjected to a uniform heat flux go (Wim) (if itis insulated, g,=0), and heat is dissipated by convection into a fluid of temperature T. at the boundary surface at x = L. These boundary con ~ApT Tyra where A is the cross-sectional area normal to heat flow and f is the heat transfer coefficient. Equations (3.866) are also valid for the present case. For a one-element .), the finite element equations are : AP apc aon) Lt alt= (eh ee) kal ‘Their solution issECOND-ORDER BOUNDARY VALUE rRomLENs 113 where a= 1+ Bh/k. ‘The exact solution is : ol? (, , 22) ole re Se( aa) The finite element solution agrees with this at the nodes. ‘Set 3. In this case, the boundary surface at x =O is maintained ata temperature ‘while atthe other boundary surface, x= L, a heat fx atthe rate of go is removed. These boundary conditions can be expressed in terms of the primary and secondary vatiabes as | UT, Om Ago ‘and (3.866) are stil valid. For a two-element mesh with fiy=/y =AL =f, we have 1 olf{u 1 at 2 -1flatastlal}s ont aslo”? LL on kA KAT 2 -1]{Us) _ Agoh {2’ 4 “a ase} 0 eau Mach +4 (u,— Uy We nave 3qok* | gol Oe eth we (Qe (2AQoh + Age) | For a three-element mesh (t= 4L), we have 2-1 Vf) af) £°) eal m2 rift LL obama ot le? US Lad Using Cramer's rate, we can solve for Us, Us, and Uy: , iia o UatG|R 2-1) Pig a where414 sexsrs ELEMENT ANALYSIS OF ONEDIMENSIONAL FROBLENS and (F)= (BB B)T is the righthand-side vector: Ageh AAgah + Ago Evaluating the determinants, we obtain UE +R) =e aah +60) % EA y= 46 § h k h : pa fe 2B +26) (Agel + 23s) + Ts A A Pit 2R +38) —F aah +360 4% (Oaga= Hager +44 (UB) = Algo 308) (Qua ‘The exact solution of (.77) with €=0 and subject to the boundary conditions (.84c) is given by Algo Soh) rea Z[@Qt-5) +8] on ow Ot) Note that the finite clement solution at the nodes, for any number of elements, coincides with the exact solution. The exact value of Of is a (2dawu= (ba )] = 010) ~AGGoL +80 ‘The value of Q computed from the assembled equations is the same as ‘the exact one; Tewever, hen computed using the deftition Q} = ~kA(AU,/ds) ca iis in error by dramount fi=tAgsh. As the number of elements is increased, the value of (O)ax sapronces the enact valve, Of coume, when quadratic clements are used, we shall btn a more accurate (or even exact) value of (Oa ‘The next example deals with radially symmetric heat transfer in @ oylinder. Example 3.2. Consider a long solid eylinder of radius R, in which energy is generated ‘at a constant rate qo (Wm), The boundary surface at r=Ro is maintained at @ Constant temperature To. We wish to calculate the temperature distribution T(r) and heat fix g(r) = -kaT fdr (or heat Q = —AkdT/dr).SECOND-ORDER BOUNDARY VALUE ProsLeMs 115 ‘The governing equation for this problem is given by (3.78) with @= qo, The boundary conditions are rng (one) 6a ‘The zero-flux boundary condition at r=0 is a result of the radial symmetry at r=0. Tf the cylinder is hollow with inner radius R; then the boundary condition at r= Ry can be specified temperature, specified heat flux, or convection boundary condition, depending on the situation. ‘The finite element model of the governing equation is given in (3.66) (for unit height of the eylinder and a =r) [KUTI=(F9 +10 510) where Kenan feet ar, seman" vaorde ‘anu * @ sane ° opm ani orm2nt(rT)L, and (4) ae the coordinates ofthe element & = (rr). or the choice of linear interpolation functions 1 a [se (3.16D) and (3.66e)] Wi=C—Nhe VEC —rIIhe the element equations fora typical linear element are Qekrat hel LUT yg, hel2e tree) , [OF ees Ly tewetac {a} om ‘The element equations for individual elements are obtained from these by giving the clement length fi, and the global coordinates of the element nodes, 714 and Fey For the mesh of one linear element, we have r, and aL MG) -S Ge) (2) ‘The boundary conditions in (3.90) imply U;~ T, and Q}=0. Hence the temperature at (global) node 1 is U,= qRI/3k+T and the heat at r= Ro is Qi = mk(U;~ Ui) ~ FageRe= —AG0Re ‘The negative sign indicates that heat is removed from the body (because d7/dr <0). ‘The one-element solution as a function of the radial coordinate ris 3.93) and the heat flux is G94)U6 sore ELEMENT ANALYSIS OF ONEDIMENSIONAL PROBLEMS ‘The exact solution of the problem can be obtained by integrating (3.78) and evaluating the constants of integration with the help of the boundary conditions in 6.30): rot (g-* (on) For a mesh of two linear elements, we take fy =, and 115 hy +lig= Ro, The tworelement assembly gives 1-1 O14) care f 3% ot ak} -1 143 3]. MBM) Rota. p+] OF OFF 0.96) 0-3 3tluy, 4Ro+ 2Ro, @ 8.95) g(r) = Saar (in Wm), O(Re) = mgoRe Ao, rh em t= Boy Trposing the boundary conditions Us, and Q!=0, the condensed equations are ah 1-170) mkt} {so of od) + ond es ‘Thee solution is o-sten, yaa en 58a) k t From equilibrium, Qf s computed as Qh —shaqyh}-+ 3ak(Uy— U)=—ngoR3 8.986) “The finite element solution becomes Tran?) = for 0< r= 4Re - ~ 6.99) 8 (2) 4, for 1Rysr0, we obtain 2 (aayep= pact 3.120) 3 = PA ae : where fis the body force per unit length ber Ne Ae + doy) (A+ aa) FIGURE 3.12 ‘Axially loaded member and a typical portion of & ser member of length dr with its axial fores.seconb.onDER BOUNDARY VALUE PROBLEMS 125 From the strain-stress relation €,=(0,—v0,)/E, we have (because 5-9) au Es fe a Substituting for @, from here into (3.120), we obtain or @.121) For static problems, this reduces to 4 (_,de ~2 (ca$)=70) G.122a) Equation (3.1224) can be used to determine the displacement u(x) of a material point located at a distance x along the axis of a uniaxially loaded ‘member, called a bar. It should be recalled that (3.122a) is derived under the assumption that all material points on the line x = constant move by the same distance u(x) (Le., the stress at any cross-section is uniform). Equation (3.1222) is the same as the model equation (3.1), with a= EA and q =f(x). ‘The average transverse deflection u(x) of a cable made of elastic material is also governed by an equation of the form (3.1224): -E(r2) =s() (3.1226) where T is the uniform tension in the cable and fis the distributed transverse force. Example 3.4, A bridge is supported by several concrete piers, and the geometry and loads of a typical pier are shown in Fig. 3.13. The lord 20 KN m™~ represents the weight of the bridge and an assumed distribution of the tratic on the bridge. The concrete weighs approximately 25kN m”* and its modulus is E =28% 10°KN mr. We wish to analyze the piet for displacements and stresses using the finite element method. ‘The pier is indeed a three-dimensional structure, However, we wish to approxim- ate the deformation and stress fields in the pier as one-dimensional. ‘To this end, we ropresent the distributed force at the top of the pier as a point force F = (0.5%0.5)20= 5 kN ‘The weight of the concrete is represented as the body force per unit length. The total force at any distance x is equal to the weight of the concrete above that point. The ‘woight at a distance x is equal to the product of the volume of the body above x and the126 mare ELEMENT ANALYSIS OF ONEDIMENSIONAT, PROBLENS [ Fe a lide Robe of anne specific weight of the concrete: 5# 0.50), 5.0 625(1 40.52) 0s wey =05°°4 ($05 ; ‘The body force per unit length is computed from aw. fa Groat) ‘This completes the load representation of the problem. “The governing differential equation for the problem is given by (3.122e), with = 28% 10° KN m™ and cross-sectional area A(x): A(x) = (0.5 + 0.5x)0. 4a+x) Thus snail] saan =) ste oe bun oon [ware ‘The faite element mode is [Hu) = 9 + {2 (G.124a) u@=o 6.123) [i eatittias, ri= ["seowiae “ v (3.1245) du) BAS ele) For the choice of linear interpolation functions (3.165), we have y= [eta taf) deen Hie teedl Fin [2st + adyt de 625K Aor #280)SECOND-ORDER NOUNDARY VALUE PRonLiMS 127 ‘Similarly, other components can be evaluated: eagle taal | i] the iyo 62s®((]e fete ) Let us consider a twoelement mesh with hy ~/ig= 1m. We have wil 25 15, 9 SEEM -[9] 3.125) 6 5.208 ebrst fess} Te seb equine 0375 0375 0.000] (U4) [4.167 a E] 0.375 1.000 =0.625 |) U, F = 12.500} +4 O84 OF e000 -0.625 a.6as|lu) (8333, Q ‘The boundary and equilibrium conditions require Us=0, Oi+Qi=0, O=SKN ‘The condensed equations are Loss iealle}{asm) 2 ‘and the solution is given by Ux 2111 X10-'m, Uy=1238X10%m, QE=—30kN 3.126) Glas 1s 342 ef 25-25 “Glas sh x’) y= 0.6255U,~ 8.333 Hence the stress at the fixed end is given by )3/A = —30/0.75 = —40 KN m"* ‘The exact solution of (3.123) is wt) =f[ sas 6asa+sy-2sn(**2)] xn “The exact values of w at nodes 1 and 2 are u(0)=2.08 x 10°, ut) = 1.225 x 10" m ‘The four-element model gives 2.008 x 10-tm and 1.228 x 10-*m, respectively. ‘The finite clement solution at the nodes is not exact because a= EA is not a constant in the problem, 3.4 SUMMARY In this chapter, the finite clement formulation of a second-order differential ‘equation in a single variable has been presented systematically, following a step-by-step procedure. The basic steps of the formulation and analysis of a128 nie ELEMENT ANALYSIS OF ONEDIMENSIONAL, PROBLENS typical equation are described in Table 3.1, The model equation is repre~ sentative of the equations arising in various fields of engineering (see Table 3.2). The finite element model is developed following three steps: 1. Weak formulation of the differential equation over an element. 2. Finite element interpolation of the primary variables of the weak fomulation. 3. Finite element formulation over a typical element. ‘The weak formulation itself involves a three-step procedure, which enables identification of primary variables (j.c., variables that are required to be continuous throughout the domain, including the nodes at which elements are connected). The finite element interpolation functions havg been developed here on the basis of continuity, completeness, and linear independence. The finite element model has been developed by substituting appropriate interpola- tion of the primary variable into the weak form of the differential equation. ‘Applications of the model to the solution of problems of heat transfer, fiuid mechanics, and solid mechanies have been presented. To aid the reader, a brief review of the basic terminology and governing equations of cach of the three fields has also been given. The numerical examples should aid the reader in deeper understanding of the steps involved in the finite etement analysis of ‘one-dimensional second-order differential equations. Tt has been shown that the secondary variables of a problem can be computed using either the global algebraic equations of the finite element mesh (ie., condensed equations for the secondary variables) or by their ‘original definition through finite element interpolation. The former method fives more accurate results, which will satisfy the equilibrium at interelement nodes, whereas the latter gives less accurate results, which are discontinuous at such nodes. ‘The secondary variables computed using the Lagrange linear Clements are element-wise-constant, while they are element-wise-linear for the Lagrange quadratic elements PROBLEMS Many of the following problems are designed for hand calculation while some are intended specifically for computer calculations using the program FEMIDV2. This should give the student deeper understanding of what is involved in the formulation and Solution of a problem by the finite element method. The hand calculations can be verified, in most cases, by solving the same problems using FEMIDV2, which is described in Chapter 7. basin the wok om ad the ie comet ode of he flowing eet dp day & (bu - Ect 50S) earns tr secrenSECONDONDER BOUNDARY VALUE rRoDLEMS 129 where a,b, ¢, and fare known functions of postion x. Ensure that the element cocficient matrix (K*] is symmetric, What is the nature of the interpolation functions for the problem? 3.2, Construct the weak form and the finite element model of the differential equation A (att) -oMtes orocect over a typical element 9° = (x4, x5). Here a, b, and fare known functions of x, and u i the dependent variable. ‘The natural boundary condition should not involve the function (2). What type of interpolation functions can be used for u? 3.3. Construct the weak form and associated finite clement model of the equation ay ae -f (elt) verey tr 0x uch ha he tra ond condon othe pe te Ft ku u)= 4s included in the weak form. Here a, ¢, and fare known functions of x, while k, uy, and ate constants. 344, Derive the Lagrange cubic interpolation funetions for a foursnode (one- dimensional) element (with equally spaced nodes) using the alternative procedure based on interpolation properties (3.28). Use the local coordinate & for simplicity. 35. Verify (3.340,b) by actual evaluation of K3= Kj and fi. 3.6, Evaluate the following coofficiont matrices and source vector using the linear Lagrange interpolation functions: «otto xf" er any dtha my f ot ext vE Uf a, fim [otavnean here a, a1 Go er» do and gy ate constants, 3:7. Verity the coefficients in (3.404, 0) by actual evaluation of Kij= Ki and ff pw bor , re u; ew E oj tek te [Ser ba ew jogo A ‘v i Typical element FIGURE P38130 38, 39. 3.10, FINITE ELEMENT ANALYSIS OF OWE IMENSIONAL, PROBLEMS Consider the system of linear clastic springs shown in Fig, P3.8. Write the forco-displacement relationship for a typical (single) spring element, and assemble the element equations to obtain the force-displacement relations for the entire system. Use the boundary conditions to write the condensed equations for the unknown displacements and forces. Consider the hydraulic network (the flow is assumed to be laminar) shown in Fig. P3.9. A typical element (which is a circular pipe of constant diameter) with (wo nodes is also shown in the figure. The unknown primary degree of freedom at teach node is the pressure P, and the secondary degree of freedom is the flow (or ‘ischarge) Q. ‘The element equations relating the primary variables to the secondary variables are given by La lla mes where d, is the diameter of the pipe, hy is its length, and 4 isthe viscosity of the Rid. Write the condensed equations for the unknown pressures and flows (use the minimum number of elements.) ‘Answer: P,= Qa, Py=#Qa, r= 1iQa. gee am Rp=M [6° J Po FIGURE P39 Consider the direct current electric network shown in Fig, P3.10. We wish to determine the voltages V and currents 1 in the network using the finite element rmethod. A typical finite element in this case consists of a resistor R, with the primary degree of freedom being the voltage and the secondary degree of freedom being the current. The element equations are provided by Ohm's law: apo a -1yvs my ala illvd-{e) ‘he cminaly cons a th intlenent ode ese tat tenet caren flow into any junction (node) always be zero in a closed loop. Set up the algebraic equations (/¢., condensed equations) for the unknown voltages and currents.SECOND-ORDER SOUNDARY VALUE PRoBLENS 131 Beg pf vi 1a nn hes Re ‘Typical clement FIGURE P3.10 3411. Consider the composite structure of axfally loaded members shown in Fig, P3.11. Write the continuity conditions (ie., the correspondence of element nodal values to global nodal values) and the equilibrium conditions (i.c., the relationships between Qf at the interelement nodes) for the structure, Derive the assembled coefficient (stiffness) matrix for the structure, and set up the condensed equations for the unknown displacements and forces. Aluminum [Ey = 107 tb in-* Ey = 3 X10" Ib in z | [in a 2 | Aluminum 30 ii ip dam 20 in Vin di he 16 in Rigid bar FIGURE P3.11 3412, Use the finite element method to solve the differential equation ~ Gio etx =0 tor 0 10 psi) beam shown in Fig, P3.31. Determine the transverse deflection using two linear elements. Exploit the symmetry of the beam.SECOND.ORDER BOUNDARY VALUE PRonLents 141 3.32. Determine the axial deformation of a nonuniform bar, A= Ag+ A,x, under its ‘own weight (fy per unit length). Use two linear elements, The bar is fixed at r=0, 3.38, Turbine disks are often thick near their hub and taper down to a smaller thickness at the periphery. The equation ‘governing a variable-thickness 1 =1(r) disk is S10) ~ 105+ 190°? =0 where a is the angular speed of the disk and ond(itey!), oe ‘ dr) (0) Construct the wes intel form of We goog equation ssh thatthe Sincar fom i symetie end the et unde) endian res spelg te qany rr (6) Bentop ine fit cen’ model ested ih the wk orm died pane) 513, Determine the aia deformation of «varying oszeston menber (ee Fig 0 une sown wig Use one unas See FIGURE P34 REFERENCES FOR ADDITIONAL READING Fluid mochanies rd, R. B., W. E, Stewart, and E. N. Lightfoot: Transport Phenontena, John Wiley, New York, 1960. Duncan, W. ., A. S. Thom, and A. D. Young: Mechanics of Fluids, 2d ed, Ekevier, New York, 1970. Hare, M. Ex: Ground Water and Seepage, McGraw-Hill, New York, 1962 Shames, I. H-: Mechanics of Fluids, MeGeaw-Hill, New York, 1952, Vallentine, H.R Applied Hydrodynamics, Butterworths, London 1959. Heat transfer Kreith, Fz Principles of Heat Transfer, d ed., Harpet & Row, New York, 1973. Myers, G. Eu: Analytical Methods in Conduction Heat Transfer, McGraw-Hill, New York, 1972 Nagotow, E. Pe: Applications of Numerical Methods to Heat Transfer, McGraw-Hill, New York, 1978,142. re ELEMENT AKALYSIS OF ONE-DIMENSIONAL, PROBLEMS Solid mechanics Boresi, A. Bo, and P. P. Lyon ‘cits, 1, 1974. Dym, €. Lo, and [Hl Shanes: Solid Mechanics: A Variational Approach, McGraw Hil, Now York, 1973. Reddy, 1 Ne: Energy and Variational Methods ix Applied Mechanice, Joho Wiley, New York, 1988 “Timenshenka, S. P., and J. N. Good stcty in Enginering Mechanics, Prentice-Hall, Englewood 1 Theory of Elasticity, McGraw-Hill, New York, 1970 ‘Variational methods (Gee aso References for Additional Reading in Chapter 2) Crandall 8. Hi: Engineering Analysis, McGraw-Hill, New York, 1956, Mithta.'s. Gs Varaonal Methods i Mathematical Physics, Pergamon Press, New York, 1964, Tis Mumerieal Performance of Varialonal Methods, Wolters-Noordhott, Groningen, wn, ? . ‘oden, 3 T, and J, N. Reddy: Variational Methods in Theoredcal Mechanis, Springer Vers, New York, 1976; 2d ed, 1983 IReddyn I. Ne Brergy end Varlafonal Methods in Applied Mechanie, John Wiley, New York, 1984 Applied Functional Anayss and Veretional Methods in Enginering, MeGrawHil, New ‘York, 1986; Krieger, Melbourne, FL, 1991. Onn ML, Rasmussen: Advenced Engineering Analysis, John Wiley, New York, 1985 Keieger, Melbourne, FL, 1950 Roktonja Ke! Variadonal Methods in Mathematics, Selence and Engincerng, Retéel, Boston, 1977. Schechter R. S.: The Variational Methods in Engineering, McGraw-Hill, New York, 1967 Manton Ke Variational Methods in Elasticity ond Plascty, 20 ed, Petgamon Press, New York, 1975; 34 ed, 1982.“CHAIR 4 BENDING OF BEAMS 4. INTRODUCTION Here we consider the finite clement formulation of the one-dimensional fourth-order differential equation that arises in the Euler—Bernoulli beam theory and the pair of one-dimensional second-order equations associated with Timoshenko beam theory. The formulations of a fourth-order equation and two coupled second-order equations. involve the same steps as described in Section 3.2 for a second-order equation, but the mathematical details are somewhat different, especially in the finite clement formulation of the equations. 4.2 THE EULER-BERNOULLI BEAM ELEMENT 4.2.1 Governing Equation Jn the Euler-Bernoulli beam theory, it is assumed that plane cross-sections perpendicular to the axis of the beam remain plane and perpendicular to the axis after deformation. In this theory, the transverse deflection w of the beam 143,44) suwrre ELEMENT ANALYSIS OF ONE-DIMENSIONAL PROBLEDS fo Ae) . woos a a cue mad vat Bleu ete me te & force-bending — momerit—deflec- ae! vention. is governed by the fourth-order differential equation £08) fle) for 0 eon (S| =o ceo ‘These can be stated in compact form as (i, j= 1, 2) () ly PAD) = 8, O34G)=0, DO (| -« where %,= element Q* = (t,, X41). (4.11a) (4.11) and #,=h, are the local coordinates of nodes 1 and 2 of the150 nar ELEMENT ANALYSIS OF ONE-DIMENSIONAL PROBLEMS It should be noted that the order of the interpolation functions derived above is the minimum required for the variational formulation (4.4), If a higher-order (., higher than cubic) approximation of w is desired, one must tither identify additional primary unknowns at each of the two nodes or add additional nodes with the two degrees of freedom (Ww, ~dw/dx). For example, if we add dw/de? as the primary unknown at cach of the two nodes, or add @ third node, there will be a total of six conditions, and a fifth-order polynomial is requited to interpolate the end conditions (see Problems 4.27 and 4.28). However, continuity of div?/dx? is not required, in general, and such elements should be used only in problems where d’w/dx* is continuous everywhere. FINITE ELEMENT MODEL. The finite element model of the Euler-Bernoulli eam is obtained by substituting the finite element interpolation (4.9a) for w and the @ for the weight function v into the weak form (4.44). Since there are four nodal variables uf, four different choices are used for v, v= @%, v= 5, v= 4% and v=9%, to obtain a set of four algebraic equations. The ith algebraic equation of the finite element model is (for v= #7) on 3 (fog eae)ur [wae or rz or D Kiwj-Fr=0 (4.126) Bi co deh gh mn ga [oe Ga, rie [" araeror | Gato where Note that the coefficients Kj are symmetric: Kj= Kj, In matrix notation, (4.125) can be written as {Kis Ke KUT) (1) (OF s. Kia Kis Kis |} us| _ Js, ] 8 = (4. 5 Kin Kis KS [Jus] 45") 95 ee Ki, Ka Ky Kullu) Ud los ‘This represents the finite element model of (4.1). Here [K‘] is the stiffness matrix and (F*) is the force vecior of a beam clement. When 2 transverse point force Fis applied at a point xo inside the element, iis distributed to the ‘element nodes by the relation [see Remark 5 in Chapter 3: (3.60)}: 7 Fail) (4.14) which contains both transverse forces (F§ and F) and bending moments (F5 and F 9).BENDING OF MEANS ISI For the casein which & (= ED) and f are constant over an clement, the clement stiffness matrix [K*] and force vector {F*} have the specific forms (see Fig. 4.2 for the element displacement and force degrees of freedom) th —6 —3h 2h? 3h? ON | = r=constan Woah he as en + & (/-=constant) hn Qs, It can be verified that the generalized force vector in (4.15) represents the “statically equivalent” forces and moments at nodes 1 and 2 due to the uniformly distributed toad over the element. For any given function f, (4.12c) provides a straightforward way of computing the components of the general- ized force vector {f*}. 4.2.4 Assembly of Element Equations ‘The assembly procedure for beam elements is the same as that used for bar elements, except that we must take into account the two degrees of freedom at each node, Recall that the assembly of elements is based on: (a) interelement continuity of the primary variables (deflection and slope) and (b) interelement ‘equilibrium of the secondary variables (shear force and bending moment) at the nodes common to elements. To demonstrate the assembly procedure, we select a two-element model (see Fig. 4.4). There are three global nodes and a total of six global generalized displacements and six generalized forces in the problem. The continuity of the primary variables implies the following relation between the element degrees of freedom uf and the global degrees of freedom U, (eee Fig. 4.4); W=U, WU, wana; 4.16) ul=ud=U,, = Us ad In general, the equilibrium of the generalized forces at @ node between two connecting elements 2° and Q/ requires that Q5 + Of = applied external point force Qi + O5= applied external bending moment ain If no external applied forces are given, the sum should be equated to zero. In equating the sums to the applied generalized forces (j.e., force or moment), the sign convention for the element force degrees of freedom (see Fig. 4.2)152. sasire ELEMENT ANALYSIS OF ONE-DIMENSIONAL PROBLEMS U,=0, Qf=unkrown Un OL GF=O Ur =0.08 mnknown, Sarat Element 2 a ya a 2 FIGURE 44 Nesibly of two Buler-Bemoulil (or casscl) beam finite elements, and the fate clement aAtaion ts linear combination of the nodal values and interpolation functions should be followed. Forces are taken positive acting upward and moments are taken positive acting clockwise. “To impose the equilibrium of forces in (4.17), it is necessary to add the thied and fourth equations (corresponding to the second node) of @ to the te a cecond equations (corresponding to the first node) of 9% Consequently, the global stifinesses Ky, Kyu, Keo, and Ku associated with slobal node 2 are the superpositions of the element stifinesses: Ky=Kb+Kiy Ko Kict Ki, Ko= Kit Ky Kun Kit Kb (4.18) In general, the assembled stiffness matrix and force vector for beam elementseNpING OF BEAMS 153 connected in series have the following forms: Global nodes 1 2 3 Kh Kh Kh Kh [k]=| Ki, Kh Kh Kb K+ Kit Kit Kix Ku+Kh | Ku+Kh [Kb | ee pp (4.192) Fy str {Fy= Hae (4.198) FE Fi ‘The connectivity matrix [3] (which will be used in computer implementa- tion) for the two-element mesh is 12 tl=[, 3] Since there are two primary degrees of freedom per node, repetition of a number in (B] indicates that the coeticients associated with both degrees of freedom will add up. For example, the repetition of the global node number 2 (which corresponds io global degrees of freedom 3 and 4) in rows 1 and 2 indicates that the global Ks3, Ks., Kg, and Ky, have contributions from both elements 1 and 2. ‘The assembled system of equations for a mesh of two elements with hy=h,=4L=h, and constant EI and f [hence, [K*] and {f*} are given by1154 sore ELEMENT ANALYSIS OF ONEDIMENSIONAL PROBLEMS (4.15) is 6 3h 6 3h OT 3h 2h? 3h we 0 0 | 2g) -6 3k 646 3h—3h -6 —3h| Us [3h he 3h—-3h 2242? 3h MP | Ue o 0 6 th 63h | Us 0 0 3h ho LO 6 a -h a fa j6re h+ OF |, Bynnpyolver, 6 |e] @ a a ‘The reader is cautioned that (4.20) does not repiesent the assembled equations of any two beam clements; it is based on the assumption that /hy= lay (ED. =(EDa, and (f= (Pla Equations (4.19) are more general. 4.2.5 Imps ions [At this step of the analysis, we must specify the particular boundary Conditions, i.e., geometric constraints and forces applied, of the particular problem to be analyzed. The type of essential (also known as geometric) boundary conditions for a specific beam problem depend on the nature of the ‘geometric support. Table 4.1 contains a list of commonly used goometric supports for beams. For the sake of completeness, the boundary conditions on the axial displacement u arc also included, The natural (also called force) boundary conditions involve the specification of genéralized forces when the corresponding primary variables are not constrained. Here we consider a ‘cantilever beam (Je., a beam fixed at one end and free at the other) of length ‘L, flexural rigidity EY (=constant), and subjected to a uniformly distributed force f,, end force Fy, and end moment My (see Fig. 4.4), First, we write the equilibrium conditions for the generalized forces. At global node 1, Q} and Qi (the shear force and the bending moment, respectively; i.e», the reactions at the fixed end) are not known. At global node 2, there are no externally applied shear forces and bending moment. Hence, 23+ 0) Qi+ Qi=0 (4.214) ‘At global node 3, the shear force is given as fa, and the bending moment as My (note the sign convention for F and My from Fig. 4.2) aD. n of Boundary Con ‘My (4.21)BevonG OF BEANS 155 ABLE 4 ‘Types of commonly used support conditions for beams and frames Displacement Force oundry “ype of support tonitons "4 None Al a apcied FREE zt PINNED ef pS-. Transverse fore and moment ROLLER are specified Geica) S * wad ‘Horizontal force and bending ROLLER Troment are speciiod (horizontal) None specied FIXED or CLAMPED awlas~0 Next, we identify and impose the specified generalized displacements ‘Since the beam is clamped at global node 1, it follows that the deflection w and the slope dw/dr are zero there: ut= wh (4.22) a Q 6 0 Us 6 BR Us hy} lM156 Nire ELEMENT ANALYSIS OF ONEDIMIENSIONAL FRODLENS 4.2.6 Solution Equation (4.23) contains six algebraic equations in six unknowns: (G1, O, Us, Us, Us, Ue). Because the algebraic equations for the unknown feneralized. displacements (Us, Ux, Us, Us) do not contain the unknown generalized forces (Q4, O'), equations 3, 4, 5, and 6 of the system (4.23) can be solved independently; the known values of the displacements Us and Uz are used in equations 3 and 4. This provides us with the motivation to partition (own by dashed lines) the matrix equation (4.23), which ean be recast in the form vey coc) aos [hey tents} Cr} , ea where (U"} denotes the column of known generalized displacements, (U") the column of unknown generalized displacements, (F'} the column of Unknown forces, and {F*} the cohimn of known forces. Equation (4.24) can be written, after carrying out the matrix ‘multiplication, in the form [cf 655) [KM] U!) + (KAVO?} = (FY [K™](U) + [KU = (FY or (K™|(U} = (8) - [K*](U'} (4.25) {FI} = (KM](U} HKU, (4.256) ‘These are the condensed equations for the generalized displacements and forces, respectively. Since (U'} and (F?} are known ({U"} = (0}), we can use (4.25a) to solve for (U2), and then use (4.256) to compute the unknown reactions (F') Us 12 0 -6 -3h]"( oh u{_m | 0 4 3h 0 yd (4 wy u,f 261] -6 3h 6 3h Rot Yoh 426) Us, 3h 3h 2h? | Mo tefl” Inverting the matrix (say, by Gaussian elimination) and performing the matrixexpiva oF wens 157 multiplication (jc, solving the equations), we obtain 2h? -3h Sh? —3h Soh mh |-3h 6, 9h 6 0 (U7 FEF| sy? on 16K? 12K |) B+ doe =3h 6 ~12h 12 JL-Mo+ dy, SHR + 3hMy + Ffah™ h_ | —9hR—6My~Tfoh? BET | 10h"F + 13hMy + 17H? 4.27) = 12KK, = 12My ~ 8h? ‘The reactions Q} and Q} can be obtained by substituting (4.27) into (4.256). ‘The Qf obtained from the clement (equilibrium) equations are more accurate than those obtained from the definitions (4.3), wherein the derivatives of w are obtained by differentiation of its finite element interpolation. ‘The reactions from equilibrium (Qf)equi are Us wm {y uf Lad Us, — Us + 2foh) fons + fh) + ‘u) 428) It can be verified that the reactions Q{ and Q3 in (4.28) satisty the static equilibrium equations of the beam: Qi+ (H+ fA)=0, O1— Fh + Boh? + Mo) =0 ‘The reactions Q} and Q} can also be computed using the definitions (4.3): From (4.106), we note that the second derivative of the Hermite cubic interpolation functions is linear over the element and the third derivative is constant over the element. Therefore, the reactions, i.e., bending moment and shear force, computed using the definition (4.3) are element-wise linear and constant, respectively. Further, at nodes connecting two elements, they yield discontinuous values because the second and third derivatives of w are not (4.294) ied (er)1158 sustre ELEMENT ANALYSIS OF ONE-DIMENSIONAL PROBLEMS ‘made continuous across the interelement nodes. Substituting (4.105) into (4.292), we obtain the values apne tis u8i)| = ai[u(-7g) +o(—_)]-- e209 (4.296), os=B1( 04+ Vez) = We +28 + BiH) he Which are in error by f= Hoh and f$=bfoh” compared with those in (4.28). 4.2.7 Postprocessing of the Solution “The finite element solution as a function of position x is given by . egy a [Ua + Ud for 0 eo VG) area) tg). aea-Q)G-Y om x)? 3 2 2 seas ay ails) CH] -The exact solution of (4.1) subject to the boundary coniition (4.21) ean be obtained by direct integration, and is given by ‘Hot — AR + oh) + A Mo + Fol + Mol? Ape? +H + fol x? — (Mot E+ AGL?) for O< ask M(x) = Sfx? - (Fs tLe + Mo + Fab + Mol® (431) “The finite element solution (4.30) and the exact solution (4.31) are compared in Table 4.2 for the data f=24kNm', H=60KN, L=3m, Mp=0kNm £=200%10KN mm, 1=29% 10% mm* (EL = 5800 KN m") ‘As expected, the finite element solution for w and @ coincides with the exact Solution at the nodes. At other points, the difference between the finite ‘element and exact solutions is less than 2%. “This completes the finite element formulation and analysis of the fourth-order differential equation (4.1) governing the Euler-Bernoulli beam theory. Whenever the flexural rigidity 6 = / is a constant in each element, theBENDING OP azANes 159 ‘TABLE 42 Comparison of the finite element solution with the exact solution of the cantilever beam of Fig. 4.4 (2 elements; L=3m, B/= 5800 KN m?, fo= 24 kN m™!, y= 60KN, Mg =0KN m) w@) 0 = dwlae MIEI (kN m) x(n) FEM Exact KEM Exact FEM Exact 0.00 0.0000 0.0000 0.00005 0.0000 0H; 0.0197 O.187s 0.0008 0.0008 0.0088 0.0059 0.0452 0.0455 03750 0.0083 0.003306 I7! ots ood 05625 G.0oT1 0.007244 eats .0878 03S 07500 0.0124 0014 = OT Oa. O.0338 05375 G01 0188403727 0.08S 0.0301 11250 0.0263 0.0263 0.026 OS n6S 0.02665 1312s GoM? 0.0347 Ou oeT oa .23 1500 0.089} 0.0139 0.05124. 0.0512 oLDH 0.0202 1687S 0.0539 0.0539 0.0546 0.0547 .0169 IT E8750 0.0644 O.0684 ©0573. 0876. ODIO. 2.0525 0.0754 0.0755 0.0600 0.0600 018.01 2.2500 0.0868 0.0869 0.0620 0.0620 a.0093 0.0089. 2ABIS G98 0.0987 0.0635 0.0634 0.0058 0.0065 218230 0.1106 0.1107 L085. 0.0013 v.00 Dams 0.108 0.1228 0.0651 0.0650 0.0017 0.0020, 3.0000 0.1350 0.1350 0.0652 0.0652 0.0098) 0.0000, 1 Nodal vues aloes are computed by interpolation, clement stiffness matrix (4.15) can be used directly. The finite element solution for the generalized displacements at the nodes is exact for all problems for which Ef is constant and the applied transverse load f is a polynomial expression. Further, the solution is exact at all points if the distributed load is such that the exact solution is a cubic. The bending moment at any point in the element ° of the beam can be ‘computed from the formula (4.324) For beams of rectangular cross-section with height H (and width B) the maximum bending stress is, MH _,EH@w_ EHS és Py get FS Ge (4.326) O ‘The minus sign is for the top and the plus sign for the bottom of the beam.1160 intr ELENENT ANALYSIS OF ONEDIMENSIONAL, FRODLENS AAG) = f= 400 Tb 1 10,0001 [es wee y= 2x 107 tb set PAD eal 12 te —e-6 — " o} Wy = 0,02 Us Ue Us Ue Ur ——————e 1 2 3 4 o 10,090 1 , fh 1 i A Aa AP ga G3 + Q} = ~10,000 tb ao G4 $=) (a) ot atonwo ra a © FIGURE 45 Fini clement mesh and equilibrium conditions for the team bending problem considered in Eeample 42: (@) physical problem; (6) finite element mesh of three elements (©) equihsarn ‘ondifons among the generalized forces (.c., secondary variables), 4.2.8 Examples [Example 4.1. Consider the beam shown in Fig. 45(@). The differential equation (6-1) js valid with the following discontinuous data: _ (2x10 be? for 0U.=0; wen=0>t,=0, (H)] =0>u0 (635% Using the known forces and displacements in (4.38), we ean partition the global system of equations to obtain the condensed equations for the unknown generalized displacements and forces. Since the specified generalized displacements are all zero, the condensed stiffness matrix associated with the unknown generalized displacements can be obtained by deleting the columns and rows corresponding to the known U, [see the submatrix enclosed by the dashed lines in (4.34): i £000 1.200 4.000 0.000 0.000") (| u) ( 20,000 : 0.309 0.783 -0.069 -0.17) | u,| | 12,000 i iw 11.333 0417 1.667] | U, b= 4 =20,000 > (4.36) i symmetric os ~1.250] | u,} | 10,000 soon] | u °162. UNITE ELEMENT ANALYSIS OF ONE-DIMENSIONAL TROBLENS ‘The unknown reactions can be computed from the remaining equations, ‘equations 1, 7, and 8 of (4.34): 4 (Ue Qt) [12,000 ~o12 004 -012 00 09 Vy, a@t=1 o S41 oo 00 0.0 0.0556 0.1667} u, Us @ 6 0.0 0.00.0 0.1667 0.3333. aan “The algebraic equations (4.36) ace solved frst for the generalized displacements, ‘and then (4.37) are used to obtain the unknown generalized forces of the problem. “ — Fast ons = Geld (FEM My FIGURE 46 oto ervorem ‘Comparison of the finite element 0- "tions for defection, slope, and © 5 10 152-25 3 ering moment of 'a_ cantilever xi ” ample 4.2), beam with the exact solutions (Ex-ENDING OF weaMs 163 Equations (4.36) are solved with the help of a computer; the solution is U,= 0.03856, U,=—0.2808 ft, Us = 0.01214 (4.380) Us= 0.1101, Up -0.02752 ‘The reaction forces, from the element equilibrium equations, ate Qi 18,565,541, Q}=15,434.461b, QI=92,165ftIb (4.380) Plots of the finite element solution for the transverse deflection w, rotation @=~dw/de, and bending moment M = El d?w/de? are shown in Fig. 4.6. Because of the discontinuity in the loading and the flexural rigidity, the exact solution w(x) is also defined by three separate expressions for the three regions Rix? ~ bAgtt — Ayr for 0=x=10ft we) q BR? — BAe — Mi) Ane +A, for 10f 1, the spring acts essentially like a sigid support. This completes the development and applications of the Euler—Bernoulli beam element. 4.3 PLANE TRUSS AND EULER-BERNOULLI FRAME ELEMENTS Structures composed of bar elements and beam elements are classified as truss and frame structures, respectively. By definition, bars can only carry axial Toads and deform axially, whereas beams can take transverse loads and bending moments about an axis perpendicular to the plane of the member. All members of a truss are subjected to only axial loads, and no transverse sheating forces and bending moments are experienced by any member. All members are connected to each other through pins that allow free rotation about the pin axis. On the other hand, members of a frame are connected by a rigid connection (¢.g., welded or rivetted), so that axial and transverse forces and bending moments can be developed in the members. A truss can be looked upon as a special case of a frame structure. Thus, a typical truss member can be modeled using the bar finite element developed in Section 3.3.3. A member of a frame structure can be modeled by a superposition of the bar element with the beam clement of Section 4.2. ‘The objective of this section is to formulate, with the help of the information from Sections 3.3.3 and 4.2, the truss finite element and frame finite element. The formulation will be based on matrix notation. Since a truss finite clement is a special case of the frame element, the derivation is presented for the frame element only. In many truss and frame structures, the bar and beam structural elements ace found in many different orientations (see Fig. 4.10). Analysis of such structures for displacements and stresses requires the setting up of a global coordinate system and referencing of all quantities (i.e. displacements, forces, and stifinesses) of individual elements to the common (global) coordinate system in order to assemble the elements and impose boundary conditions on the whole structure. When a truss element is oriented at an angle from the global axis, its axial displacements at the nodes have components along the global axes. Thus, every node of a truss will have two displacements in the alobal coordinates: one along the global x axis and another transverse to the x axis. Therefore, the element will have two degrees of freedom per node in the global coordinate system, In order to facilitate the transformation from element equations to global equations, we append the element equations for the axial displacements to those associated with the transverse displacements. Since there are no stiffnesses associated with the transverse displacements168 anmre ELEWENT ANALYSIS OF ONE-DIMENSIONAL FROBLENS t, Apin connection Fy Members rotate freely about the pin axis @ z ip frame are welded 30 that they cannot rotate freely about the axis perpendicular 7 . to the plane of the structure me |\« © FIGURE 4.10 ‘Typical exomples ofthe plas truss and plane frame: (a) 2 pane tre severe (all element erry ep eal toads) (0) atypical plane feeme strctre (all members may cary axial and transverse Toads and bending moments). (because the element experiences axial deformation only), the entries in the Stiiness matrix corresponding to the transverse displacements are set equal to goro. For example, when the linear Lagrange interpolation is used for the axial displacement, the element stifiness matrix in the element coordinates for 8 trans clement with two displacements (axial and transverse) per node can be written as 1 0-10 AE| 0 0 00 UF] -1 0 10 a9) 0 0 00 Similarly, a superposition of the bar element of Section 3.3.3 with the beam Sloment of Section 4.2 gives three primary degrees of freedom per node (see Fig. 4.11a) of the frame clement: (1 w, ~dw/dx). When the axial stiffness (AB, bending stiffness EL, and axial distributed force q and transverse Gisteibuted force f are element-wise-constant, the superposition of the Tincar bar clement with the Hermite cubie Euler-Bernoulli beam element gives the following element equations: (k*}{u} = (FY (4.504)BENDING OF BEAMS 169 % as R 8 a iy fs - ay A Fy a Be Generalized displacements Genetalized forces @ Element degiees of freedom Global degrees of freedom o FIGURE 4.11 “The frame clement with primary and secondary variables (or degree of freedom) in the local and ate systems. (a) The generalized displacements and forces in the element coordinate system (8, §,2)- (6) The generalized displacements in the element ccordinate system and global coordinate system (, , 2); the y axis is into the plane ofthe paper. The angle of orientation is rmeesured counter-clockwise from the global x axis to the element £ axis, The tras element is fbtained from the frame element by omiting the rotation and moment degrees of freedom at the odes (#60, = i= 0 forthe trus element). where aye aah) (or a fh i 4% —ane| Jos » os 05) wf O° dak PY 0% Gsm) % +m | los 0, af) (0% and q = qo and f =f, are constants over an element. "The element stiffness matrix (K‘] is of order 6x 6: nu 0 0 0 0 0 6 -3 0 -6. -3h mm 0 3h An? 0 uw 0 of’ #@ ar | 45 3h 0 6 Bh 0) 3h 2h4170 pore ELENENT ANALYSIS OF ONEDINENSIONAL FRODLEMS In the following paragraphs, we develop transformation relations to express the element equations (4.50c)—valid in the element coordinate system—t0 the ‘global coordinate system. ‘Let d and w denote the axial and transverse displacements referred to the local coordinate system (Z, 2). The local coordinates (, 7, 2) are related to the ‘global coordinates (x, y, z) by (see Fig. 4.116) [cose 0 sina | (x z j o 1 0 |yy (451) 2) L-sine 0 cosatle where the angle a is measured counter-clockwise from x axis to ¥ axis. Note that the y and coordinates are paral to each other, and they are io the plane of the paper. The same transformation relations hold/for displacements {us v, w) along the global coordinates (x,y, 2) and displacements (@, 2) in the local coordinates (f, §, 2). Since we are considering 2-D structures, we have v= 0-0, The rotation 6 = —dw/dx about the y axis remains unchanged: 0B. Hence, the relationship between (x, w, 6) and (2, #, 8) can be written as a cosa sine 0] (x wh=| sine cosa 0) (4.52) a o 0 ile ‘Therefore, the six element nodal degrees of freedom aj in the (&, 7, 2) system are related to the six degrees of freedom uf in the (x, y, 2) system by (a= a am) [cose sine 0} uy) ay ~sina cosa 0 0 ua al | oo 4 easy - sie sma’ oyu CP? ais 0 — f-sina cose 0] ] us fe, 0 0 Ashes or {ay = (Tu) (4.536) ‘The inverse of (4.53) is {ul} = (TY) RYE and it can be shown that [rJ= (07 (4.54) i.e., the inverse of [T*] is equal to its transpose.enomna oF peAMS A7L Analogously to (4.536), the element force vectors in the local and global coordinate systems are related according to {FP} =(r](re} (4.55) ‘To obtain expressions for the element stiffness matrix and the force vector referred to the global coordinates in terms of the clement stiffness matrix and force vector in the local coordinates, we use (4.53)-(4.55) in the ‘element equations, ay = (Fy (4.56) Substituting the transformation equations into (4.56), we obtain [RT e"} = (TF Premultiplying both sides with [7*]-"=[7"]", we obtain (TVR Tu) =F) (4.57) which gives TVIRUTL (FSI) ] (4.580, 6) [k‘]= ‘Thus, if we know the element matrices [K*] and (F*} of an element Q* in the local coordinate system (Z, J, 2) then the element matrices of the element n the global coordinate system, which is obtained by rotating the element coordinate system through an angle a in the counter-clockwise direction about the y axis, are given by (4.58). Note that the angle a is measured in the counter-clockwise direction from the positive global x axis. Inserting the element stiffness from (4.50e) for [R*] into (4.584) and carrying out the indicated matrix multiplications, we arrive at the clement stiffness matrix [K*] referred to the global coordinates: 2 oot a ia Ge-Ocosesine pri a bor sree hain “hee (wos! a+ sista) ~(u~<)sin wens ~Ihsin ar jcas! a 6a a ef) asin (usin? ab coH a) Vcose (4 Oosasine msn at bcos w asin oe 3h cos a we “Ah sin oe heos@ we (4.594) where w= ARQ (4.596)72 were ELEMENT ANALYSIS OF ONEDIMENSIONAL. FHOBLENS Equation (4.580), after multiplication, yields RY (Ficosa~ Risin a) | | Raina + Roose B & = (4.60) B71 Riese Asin Go | | fsina+ Koos K i ‘which is the element force vector referred to the global coordinates, ‘We next consider an example each of a truss and frame structure, sample 4.3. Consider the three-member trss shown in Fig, 4:12(e). All members of these have identical cross-sectional area A and modulus E, The hinged supports ot points 1, 2, and 3 allow fee rotation ofthe members about the axis (taken a5 Post Pi the plane of the paper). We wish to determine the horizontal and vertical displacements atthe joint 3 and the forces in each member ofthe structure ‘Gus al joints are hinged, and the applied forces are acting at the nodes, the members are subjected to only axial forces. Hence, the structure i a truss, We use Treas Enite elements 10 model the structure, Any further subdivision of the members FIGURE 4.12 Geometry and finite element dis- cevetization of » plane teuss: (a) ge fometry and loading; (0) clement fpumbers, global node numbers of lement nodes, and element nodal forces in the element coordinates.sexpina oF weans 173 oes not add to the aéeuracy, because for all truss problems the finite clement solutions fre exact. This ig a consequence of the fact that all truss members with constant ‘rosssetion are governed by the homogeneous differential equation eu ae whose solution is of the form u(x)=car ep. Thus, linear interpolation of the Uisplacements should give the exaet result, “The global aode numbers and clement numbers are shown in Fig. 4.12(8). There ave two degrees of freedom, horizontal and vertical displacements, at each node of the element, The element stiffness matrix in the local coordinate system is 1 0-1 0 A&.| 0 0 0 0 ae |-1 0 1 0 oo 0 0 (61) [kJ = (4.62) “The (cransformed) stifness matrix of the element Q, in the global coordinate system is ven by which is obtained from (4.59) by deleting the rows and columns corresponding to the bending degrees of freedom and setting all bending stiffnesses to zero, The element stifiness matrix is 4% 4 in the global coordinate system because of the horizontal and vertical displacement degrees of freedom at each node. The element data for the problem are as follows. Global nodes Element of the Geometric Material number clement properties proerty Orientation 1 2 30 AARL ane 2 10300 Ahj=VBL OE anss 3 120 Am=b OE a= ‘The assembled stifness and force cocfficients are given by 1 2 3 4 5 6 Kut KhtKh Kho Ky Kok [a KbtKh Kh Ky Kh Ku [2 KatKh Kt Kh Kis 3 (kl~ eK Kh x symmetric KUtKh Kh OKA [4 m Kytky K+ KL Kit Kil 6 (4.64){174 ontrs ELEMENT ANALYSIS OF ONEDIMENSIONAL PROBLEDIS| ‘The force vector can be written directly by including the known applied forees: (4.65) Substituting the clement data into (4.63) and the result into the assembled ‘equations, we obtain 03536 0.3536 0.0 0.0 ~0.3536 ~-0.3536 1.3536 0.0 1.0 -0.3536 --0.3536 1000-1000 10 00" (00 symmetric. ™ 1.3536 0.3536 0.3536, “The specified displacement degrees of freedom are Un U,=H=U.=0 (4.6) ‘The first two correspond to the horizontal and vertical displacements at node 1, and the last two to the horizontal and vertical displacements at node 2. ‘The condensed ‘equations for the unknown displacements and forces are a BAp1as36 0.3536)(06) _ {2 Blass asl eal“ U2") won a+ ot 0.3535 -0.3836 ai ai| a) ~aas6 -0asce |p Q3+Qi[ L| -10 0.0 U, 1+ 8 09 00 where, for example, Q?+ Q? is the horizontal force and Q3 + Q? the vertical force at node 1. Solving (1.67) and computing the reaction forces (4.68), we obtain 3PL. (4.68) PL geal pa G42v5 esse Us=(G+2V2) BA 8B » Re-P, B=3P, B= ‘The ses in each element can be computed frm the elation % A where Qs is the axial force at node 2 of the element @". Note that, for a truss element vith constant cross ection, the following relations bold =0, Gia0, Oi=-Os (4.69)BENDING OF BEAMS 175, “The Of ean be deterinined from the element equilibrium equations oF 10-1 07 fa5 O3(_ EA) 0 0 0 olf O(~h ]a 0 1 olla om) Os. 0 0 0 odke, Hence, Fromm the global displacements Us and U, we have gn Pe a jnw=o Pe gs BaF Be vycosatUsinan2e, ao any ‘Therefore, we have Pp Viz, a0 (4.72) ‘The next example deals with a frame structure. It also illustrates how to incorporate a point load between nodes. Example 4.4. The frame structure shown in Fig. 4.13 is to be analyzed for displace- ‘ments and forces. Both members of the structure are made of the same material (E) and have the same geometric properties (A, 1). The element stiffness matrices and force vectors in the global coordinate system (x, y, 2) can be computed from (4,59) and (4.60). The geometric and material properties of each element are as follows. Element 1 Eetddin, An10H5 I=Iink cosa=00, snent0 4 E=10' psi, f= 4Pibin* > lament 2 L=180in, A=10in’, 1=10in’, cosa=0.8, sine =0.6 180i, a ine ar Exp, f=0 ‘The load &,=4P at the center of the element is distributed to the nodes according to 4.19) The assembled stiffness matrix and force vectors are obtained by superposing the last three rows and columns of element 1 on the fist three rows and columns of element 2; i.e, the 3 x3 submatrix associated with rows and columns 4, 5, and 6 of element 1, ‘and the 3%3 submatrix associated with rows and columns 1, 2, and 3 of element 2 overtap in the global stiffness matrix. ‘The known geometric boundary conditions are U=0, =O, Uy=0, U0, Und, =o (4.750) ‘The force boundary conditions are Qi+Oi=0, Ol+0F=-2P, Os 08-0 (4.750)4176 snsns ELEMENT ANALYSIS OF ONE-DIMIENSIONAL PROBLEMS z woe y FIGURE 4.13 Geometry, loading and finite clement di cretization of & plane frame structure: feeometey and loading; (6) finite clement Geeretization, clement mumbers, and ele- iment fores in the element coordinates, Since all specified values of the known boundary eonsitions on the primary variables sscmro, the condensed equations for the unknown global daplacement degrees of freedom are ose 0.2665 —0.0178] (Us) “f 1 wo] 0.2655 0.8846 —0.0148 |) Usp =) 4 FP 4.16) -ooirs -o.0148 5.0000 }Lu) Las, ‘whose solution is U,=0.859% 10-*P (in), Uy 0.681 x 10™4P (in), Ug 0.961 = 10-4? (ead) 7 ‘the reactions and forces in each member in the global coordinates can be computed from the element equations (OYA 1K UY (4.782) “The forces (Q°) can be transformed to those in the element coordinate system by means of (4.55): O12? (4.780)BENDING OF BEAMS 177 We obtain ant 41.458) ons 0.180] an. 710.9001 21.550 {Oy yagi QPe Dd gl P 79) 1.275 0.180] 50.450, 10.870 44 THE TIMOSHENKO BEAM AND FRAME ELEMENTS 4.4.1 Governing Equations Recall that the Euler~Bernoulli beam theory is based on the assumption that plane cross-sections remain plane and perpendicular to the longitudinal axis after bending. This assumption implies that all transverse shear strains are zero. When the normality assumption is not used, i.e., plane sections remain plane but not necessarily normal to the longitudinal axis after deformation, the transverse shear strain &,, is not zero. Therefore, the rotation of a transverse normal plane about the y axis is not equal to —dw/dx. Beam theory based on these relaxed assumptions is called a shear deformation beam theory, most commonly known as the Timoshenko beam theory. We denote the rotation about the y axis by an independent function (x), The governing equilibrium, ‘equations of the ‘Timoshenko beam theory are A [oan (u+3)]+s=0 (4.80) é (a) ~GAK (w+ *) =0 (4.808) where G is the shear modulus and K, is the shear correction coefficient, which is introduced to account for the difference in the constant state of shear stress in this theory and the parabolic variation of the actual (i.e., as predicted by equilibrium equations) shear stress through the thickness. When the second equation is substituted into the first for GAK,(W + dw/dx), and ¥ is replaced with —di/dx, we obtain governing equation (4.1) of the Euler-Bernoulli beam theory, 44.2 Weak Form ‘The weak form of (4.80) over an element 2° = (x4, x2) is developed using the sual procedure, now applied to each equation. We multiply (4.802) wtih a Weight function ~v; and (4,805) with a weight function —w,, and integrate{178 rastre ELEMENT ANALYSIS OF ONE-DIMENSIONAL PROBLEMS over the element length: o= fomfs [ca,(w +2) ] +4} a o= [ole (e128) -cax.(w +3) ]ae Integrating the first term of each integral once by parts, we obtain 0=['[atonn(ot)-mz]a-[noax(we $2)” an o-[" [arte emcan,(w+ We) ac~[mer | (4.816) “The coefficients of the weight functions w, and Ww, in the boundary integrals are where V is the shear force and M is the bending moment; ‘constitute the secondary variables of the weak form. The weight functions and we, must have the physical interpretations that give wV and woiM units of work. Clearly, wy, must be equivalent to (the variation of) the transverse defiection w, and Ww, must be equivalent to (the variation of) the rotation function Vs ww Hence, the primary variables of the formulation are w and '¥. Denoting the shear forces and bending moments at the endpoints of the element by the expressions (4.83) the weak statements in (4,81) can be written in the final form 0= [oa (+ 82) — ma] te ontead@r~ lends aa (4.84) = fC [Ge cael +) | de m.005- Gendt 4.43. Finite Element Model A close examination of the terms in (4,84) shows that both w and W are differentiated only once with respect t6 x. Since the primary variables are thewenpwva oF meaMs 179 dependent unknowns themselves (and do not include their derivatives), the Lagrange interpolation of w and W is appropriate here. ‘The minimum admissible degree of interpolation is linear, so that dw/dx #0 and d¥V/dx #0. The variables w and W do not have the same physical units; they can be interpolated, in general, with different degrees of interpol: Let us consider Lagrange interpolation of w and W in the form S wo? sv (4.85) where pf! and yf? are the Lagrange interpolation functions of degree m —1 and n—1, respectively. In general, m and n are independent of each other, although m=n is most common, However, when m=n=2 (ie., linear interpolation is used), the derivative of w is (a) de) hy which is element-wise-constant. The rotation function W, being linear, is not consistent with that predicted by w(x). For thin beams, the transverse shear deformation is negligible, and we have V = —dw/dx, which requires eXe"k, at w= Wi tL (4.860) or, equivalently (by equating like coefficients on both sides), Sita sizq=—(WE-w), s$—s5=0 (4.860) which in turn requires (4.87) ‘This implies that (x) is a constant Wa)=s5 =i (=82) (4.88) However, a constant state of W(x) is not admissible, because the bending energy of the element, Pa BL ate)? f B (Sy a (4.894) Would be zero. This numerical problem is kaown as shear locking. To circumvent this, two alternative procedures have been developed in the literature: 1. Use a consistent interpolation for w and W such that dw/de and W are polynomials of the same order (i.e., m=n +1).180 nave ELEMENT ANALYSIS OF ONE-DIMENSIONAL, PROBLEMS 2, Use equal interpolation ((.c. m =n) for w and W, but evaluate the bending energy with actual interpolation of Wf and the shear energy (“oe (@ + wy de (4.896) with a polynomial that is one order lower. “The latter can be achieved computationally without using different interpola- tions of W in (4.89a, b). In the numerical evaluation of the integral (4.89), we use a quadrature rule (See Section 7.1 for details) that is necessary to calculate the integral f * can (M) ex (ase) exactly, For example, when the linear interpolation of w and W is used, we use sean: point quadrature rale 40 evaluate (4.899) because a one-point quadrax fase would give the exact value of the integral (4.89) wien GA = constant, A two-point quadrature is needed to evaluate the integral in (4.896) exactly. The tee of reduced integration, ic., one-point quadrature on the integral (4-896), We sesult in the finear term in the approximation of W not contributing to Tea Ghear energy, For illustrative purposes, we take a detailed look at the expression Gat [*(Geew) ae 2) a where x=x4-+Mte is the midpoint of the clement and h, is its length. Substituting (4.85) into this expression (with m =n=2), we have GAKhe 2 ( PE age <4) | The Ihe Nenxutna GAKshe (08 — wi, sitsi)? 2 he 2 ) (4.90) which is a weaker requirement than (4.87), ic., if (4.87) holds then (4.90) also holds, but (4.50) does not imply (4.87), We note that (4.87) must hold only for problems for which the transverse shear energy (4.89) is ne igible. ‘in summary, we use either consistent interpolation (m =n +1) oF equal interpolation with reduced integration in the evaluation of the transverse shear stiffness coefficients in the Timoshenko beam clement. ‘We consider both forms. of elements here. ‘Substitution of (4.85) for w and W, and = pl? and w= yf into the weak forms (4.84), we obtain the finite element equations: 02 3 Kit 3D Kify FP @= 12-2) Bs A (4.914) O= > K#m + D Kis fa morpenoine oF peas 181 where af" avi avi? wef AK, EE a Kp [" oan, ee yp ara Ke fay dy (4.91) EePAt onan [PP deen, Fn Equations (4.914) can be written in matrix form as [K™] (KPI (00}) 29 ley [Kk] (ole let (4.92) ‘A CONSISTENT INTERPOLATION ELEMENT (Cis). To illustrate the use of consistent interpolation, we select pf? to be Lagrange quadratic polynomials and ip? Lagrange linear polynomials. For this choice of interpolation, [K™] is 3x3, [K"] is 3X2, and [K™] is 2X2. The explicit forms of the matrices, when B/ and GAK, are constant, are 7-8 1 “5-1 iy = AR) 8 16 8], txe= BB] 4 -4 “Lor -8 7, 1s} (493) am EL{ 1-1], GAKA,[2 1 (x? et i+ 6 C 2 ‘The elemient equations become 2 3h _* 3h where a= EI/GAK,, 8 “3h 16 ah Los MTS 3 4 6 7 4 m8 1oayh 6 hop seh 6 h6 =I BIR ain ols ale ole + ols Ws Me is any specified transverse force at e fQ,)= & o fp tyQs oO} | Oo} [a (4.94) node 3, and1182 are BLEMENT ANALYSIS OF ONE-DIMENSIONAL, PROBLEMS Ou, Qs) axe shear forces and (>, Q.) bending moments at the end modes } and 2 of the element. Note that node 3, which is the middle node of the element, is not ‘connected to other elements, and the only degree of freedom there is the teaneverse deflection, Thus, there are different number of degrees of freedom i fferent nodes of the element, and this therefore complicates the assembly of elements and its implementation on a computer. Hence, we eliminate the fiode 3 dependence in the system of element equations by condensing out Ws- We obtain —iGAK, -4GAK, wit AGAK, 4GAK, We . -icak, yoaK, toaKa+% toaxn—F Y's Htcax, oax, icaKh—Z! aoakneSt| | (4.95) fiesale Ql frthl | Os -pi{ “Jo. aan Q where f= f+ O. This is obtained by solving the second equation of (4.94) for vos and substituting for m4 in the first four equations of (4.94). The forces fare obtained using j= [stevie ve-LyPY= quadratic interpolation funtion (4.98) “The element equations (4.95) can be assembled in the same manner as the Glassical beam equations, with two degrees of freedom (ws) per node, However, it should be noted that w is interpolated with quadratic polynomials. ‘For constant f = fy, the specified Joad vector in (4.95) takes the form [because f= Yoh, FE bts f3= Mol, Oo =0; see (3.406)] Afoh Boh 1 = 2 4.97) C9) sn hl? which is precisely the same as that obtained in the classical beam element [see (4.15)]-BENDING OF BEANS 183 AN EQUAL INTERPOLATION, REDUCED INTEGRATION ELEMENT (RIE). When equal interpolation is used (m =n), all submatrices in (4.92) are of the same order: Xn, The element coefficient matrices K}' and Kj? are ‘evaluated exactly, as is the first part of KP. ‘The second part of K7 is to be evaluated using reduced integration. For the choice of linear interpolation functions, and for constant values of GAK, and EI, the matrices have the following explicit values: GAK, we SEL a} ver SE ey ey 1 } 24m! H| 98) Ael-1 1 4 11, where one-point integration is used to evaluate the second part of [K”]. Note that [K"'], [K"], and the first part of [K®] are also evaluated exactly with one-point quadrature when EI and GAK, are constant. Hence the uniform one-point integration for [K*] satisfies all requirements. ‘The element equations are AK, iggy, AK: i h ~4GAK, sean t AGAK, ioaxn 2 st -104K, |[m|* OAK h 4GAK, 4AGAK, We HIGAK, YAK A-@ AGaK, YAK eT hye (2) = t + 3 (4.99) o} lo It is interesting to note that the element stiffness matrix in (4.99) is the same as that in (4.95) obtained from the consistent interpolation with quadratic approximation of w and linear approximation of W, except that the nodal iables are listed in a different order. The only difference is the load representation. In the consistent interpolation, the load vector is equivalent to that of the Euler-Bernoulli beam theory, whereas, in the equal interpolation with reduced integration element, the load vector does not contain any moment components due to the distributed load. ‘The quadratic interpolation of both w and \F with full integration of the element coefficient matrices also suffers slightly from the shear-locking phenomenon, A uniform two-point quadrature rule has the desired effect on {K""], [K], and (KI; ie., (K"], (KJ, and the first term of [K?] will. be4184 Nire ELEMENT ANALYSIS OF ONE-DIMENSIONAL, FROBLESS, evaluated exactly and the second term of [K™] approximately. As the degre ‘of approximation and/or the number of elements in the mesh is increased, the Shear tocking wil disappear and reduced integration is not necessary- Example 45, Here we reconsider the indeterminate, beam probler! ot Example 4.2 {eae te, 47}, and analyze i wth the reduced integration clement (RIE) ond ‘consistent {Secpolaion element (CIE). Unlike the Buler-Bernoulli beam element the Timo- interpoliion Cement doesnot yieid exact values at the modes, even when Ey and see ans constant, This is because ofthe coupled nature of the equations seating and ¥. ye beam is modeled using two (fr=4, r= 6) and four (I= 2, Ite= 2, In 9s ha=3) linear elements with reduced integration. ‘The element equations 6 © typical Fn nee ne ame as those in (.99, where forthe elements with distributed Fores are given by : . cof eps went, am where seach tent +4 We otin aya) int) for the two-element case and ne AE) on AEG) {for the four-element case (ht = hy) 0.02 Braet solution -o- FE solution @2L) © FB solution (HL) LH = 10, k = 09 0.00 ‘Transverse deflection (2) FIGURE 4.14 Tre cof the transverse defection obiained wing Timoshenko eam elements wih he exit (etection ofa cantilever beam (see Fig. 4.7). -ENDING OF MANS 185 Figure 4.14 shows plots of transverse deflection versus x for the ease with k=0 (ic., a centlever beam), beam length-o-height ratio L/H = 10, and EI = 10°f'Ib. ‘The satio L/H is» measure of the thinness of the beam. When itis large (ie, for very thin beams), the transverse shear deformation is negligible end the Euler—Bernoulli and ‘Timoshenko beam theories give the same results. For small values of L/H, say 10 GAK, =4% 10 Ib), Timoshenko beam theory predicts larger deflections, and for the problem at hand the diference is not noticeable enough to be seen if the deflections from both theories were included in Fi. 4.14. The finite element solutions shown in the figure are obtained withthe reduced integration elements Figure 4.15 contains results of a convergence study, which includes two and four linear and quadratic element meshes for the case of K/EI=1.0 and L/# = 10, The reduced integration elements are used, ‘The scale used for plotting the deflection such thatthe difference between the finite element solutions and the exact solution can be seen clearly. The mesh of four quadratic elements gives the converged solution. To show the effect of the transverse shear deformation, the exact solution for the thin beam (L/H = 10°) is also shown in Fig, 4.15. It is clear from the results that the convergence is rapid and the quadratic elements yield faster convergence ‘The accuracy of the reduced integration clement (RIE) relative to the full integration element (FIE) and the consistent interpolation element (CIE) ean be seen fiom the results shown in Fig. 4.16 (for L/H= 10 and k/EI=1). Clearly, the CIE, element gives more accurate results, followed by the reduced integration element Recall that the stifness matrices of the two elements are identical, the only difference being in thei force vectors [see (4.95) and (4.99). OF course, in the CIE element, the transverse deflection is interpolated using quadratic functions. The plots shown in Fig. 417 indicate that the reduced integration is necessry even for quadtatic elements, although they are not as sensitive to locking. 0.001 — Exact (L1H = 10, WEI = 1) = 1 = 4 | FE solutions > Se 20f WET = 1, LiH = 10) 2 > 40, i 9.000 Exact (LH = 10°, AIET = 1) E -oo 0.002 0 2 4 6 & wb BR FIGURE 4. Convergence of the finite clement solution tothe exact solution ofthe beam in Fig. 47 (k/EI = 1); the Timoshenko beam clement is used.4186 FNiTe ELEMENT ANALYSIS OF ONEDIMIENSIONAL, FROBLENS 0.001 TS. FE (@L) % -o— RIE (21) —o CIF @ elements) x Ton BIE GL) Tt, SE (4 cements) 3 opm (ay Eat 2 coo 2 3 3 2 -voo . . 002 7 re Distance « (8) FIGURE 4.16 crepe ofthe tonsverte detection w(x obtained with reused iteration clement (RYE): Garraagration elements (FTE), and consistent interpolation elements (CIE) with the exact olution: 2, two linea elements; AL, four linear elements 001 —— Reduced integration — Ful integration FES-2L 9.000 Sgrcee- 0,001 “Transverse defection w(=) FIGURE 4.17 Feo et full and reduced integration on the accuracy of the finite element solution (FES) ‘Obtained usin the Timoshenko beam elementenna oF neans 187 ‘The following general observations can be made about various finite clement models based on the Timoshenko beam theory: 1. The reduced integration element (RIE) exhibits less locking compared with the full integration element (FIE). 2, As the number of elements in the mesh is increased or the degree of approximation is increased (i.c., higher-order elements are used), the finite clement solutions obtained by both RIE and FIE elements improve; iuc., the effect of locking is reduced with mesh refinements and higher-order elements. 3, The consistent interpolation element (j.e., quadratic approximation of w and linear approximation of W) with full integration yields a more accurate solution than that predicted by the RIE element. This is due to a better representation of the distributed load. 4, The element with quadratic approximation of both w and W and reduced integration of the coefficients yields more accurate results than the consistent interpolation element with quadratic approximation of w and Jinear approximation of Y and with full integration of the coefficients, The frame element based on shear deformation beam theory can be obtained by superposing the bar and beam 4.8 INCLUSION OF CONSTRAINT EQUATIONS When the support plane of a roller support is at an angle to the global coordinate system (see Fig. 4.18), the boundary conditions on the displace- ‘ments and forces at the roller are w=, O5=Oo (4.101) where uj is the normal component of the displacement, Qf is the tangential ‘component of the force at node 1 of the element 2%, and Qs is any specified tangential force. These conditons must be expressed in terms of the global noo p+ win, FIGURE 4.18 “ 6 6 “Transformation of specified boundary conditions from a Local 208 8 — using coordinate sysem 10 the global coordinate system.4188 are ELEMENT ANALYSIS OF ONE-DIMENSIONAL FROBLENS components of displacements and forces by means of the transformation (4.53) and (4.53) us = —uj sin B + u§ cos B=0 (4.1024) Qf = Qi cos f+ Q5sin B= Qo (4.102) where (wf, wf) and (Qf, 8) are the x and z components of the displacements and forces at the support. Equations (4.102) can be incorporated into the global system of equations as follows. Consider the assembled system of equations, Ky, + Kats KyUi+ Kath+* 4 KU = ++ KoUn= Be (4.103) Kast Kyat + +--+ KunUn = Fo where U, are the global generalized displacement degrees of freedom and Fare the sums of the applied (f7) and internal (Qf) generalized force degrees of freedom, Suppose that the roller suppott is at the first node of the element 2°, and that the corresponding displacement degrecs of freedom are U;, Uj, and Up. The forces Q§, O3, and Q$ corresponding to these displacements end up in'the forces F), F,, and Fx, respectively. Hence, to include the force boundary condition of (4.1026), we must add cos f times the /th global equation to sin B times the Jth equation: (cos BK Uy + KnUat.. + Ky Uy + KyU; +... + KinUy) = (cos B)F;, (sin B)(Kn Uh + KinUt oe + KU + KyyUy ++ Kan U) = (in BE and obtain (Ky.cos B+ Ky sin B)U +... + (ny 008 B + Kr sin B)Un = (Fjcos B+ B sin) = ({icos B+ F$sin B) + (Qi e0s f+ Q3sin B) = (ffc0s B + Fésin B)+ Qo (4.104) We replace the Jth equation with this and replace the th equation with (4.1022): (sin )U, + (—c0s B)U,=0 (4.105) These modifications of the global set of algebraic equations violate the symmetry of the resulting global stiffness matrix. ‘To retain the symmetry of the global stiffness matrix, an alternative method of incorporating the conditions (4.102) is presented here. Equations (4.1022) o (4.105) can be viewed as the constraint equations among the global displacements, which have a companion relation among the associated forces, ive., (4.102b). In the following paragraphs, we present a general procedure byENDING OF BEAMS 189 which constraint equations of the type {u's (teal among the primary variables (4.106) r= {ies ‘can be implemented in the assembled system of equations (4.103). In (4.106), {U'} denotes the 2X1 vector of generalized nodal displacements that are selected as the independent nodal variables, and {U*} denotes the m x1 vector of generalized nodal displacements that are selected as the dependent nodal variables. The matrix operator [A] is of order (w +m) xn. For example, consider a frame problem with NV nodes and a total of 37 (=m +n) degrees of freedom in the problem. If one of the supports is a roller support, there will be one constraint equation of the type (4.1022). If the global node number corresponding to the roller is f then p =3(/ ~ 1) +1 is the first degree of freedom at the support. Hence, the constraint condition (4.1022) can be expressed in terms of the generalized global displacement degrees of freedom U, and Uy yx as —U, sin B+ Up: 008 B= 0 (4.107) Suppose that U, is selected as the independent variable and U,., as the dependent variable. Then the number of independent variables is 1’ = 3N +2, and there is only one dependent variable (ie., m=). For this case, the constraint (4.1062) has the form (4.1066) Yu) fi0..0 0 0 u| jor..o0 0 0 u,| Joo... 0 “ 0s Ur] o 0 (4.108) Unea} LO 0... 0 tanp 0 to Con coum Equation (4.106a) represents a transformation similar to (4.552), between two sets of global generalized nodal displacements, with [A] being the transformation matrix. Therefore, the discussion presented in the paragraph following (4.58) can be used to transform the equations in m +n variables to those in only 1 variables. Consider the assembled equations [K]{U} = {7} (4.109)190 swore ELENENT ANALYSIS OF ONEDIMENSIONAL, PROBLEMS whieh canbe rearranged a8 of} I, where (U') and (U2) denote the vectors defined in (4.1064). Using the ‘transformation (4.106a), we obtain [RAMU = (PF (111) \- (4.110) -To obtain a symmeteic coefficient matrix, we premultiply both sides with [A’, and arrive at (AT RIAU} = AP RMU} = (PF . (4.1124) where (RI=(AV TRA), (F) = FATTY (4.1126) Equations (4.112) are now ready for the implementation of boundary conditions and solution. "As an example, consider the frame clement in Fig. 4.18. We have one constraint equation, U,= Ustan B ‘The transformation equation (4.108) has the form uv) [100 0 O7f, U, o10 0 0 u ul_joor 0 of uf-}ooo 1 of}, Ju} |.9.0.90 Wy Us a0 0 tang OILY. ‘and the stiffness matrix [R] is the same as [K], with the fifth and sixth rows and columns interchanged: Kye Ku Kis Kis (Rle| Kar Ku Ke Kes Ko Ker Keo Kes Kao Ku Kio Kesenn OF neaws 19% The transformed stiffness matrix and force vector are, from (4.112b), of order 5x5 and 5X 1, respectively. ‘The elements of [KJ and {F} are given by Ry= Ry oi=1,2,3,5) Ru=Rat Ratan) Ru Rut Ketan ] ornare siee Ray + (Rag + Res) tan B+ Ree tan? B FaR G=12,3,5, fh y+ Fetan B During the imposition of the boundary conditions of the problem, the force A, will be replaced by Qo. 4.6 SUMMARY In this chapter, finite element models of the classical (i.e., Euler~Bernoulli) and Timoshenko beam theories have been developed. The classical beam theory is governed by a fourth-order differential equation, and therefore results ina weak form whose primary variables contain’ the transverse deflection and its first derivative. Therefore, Hermite interpolation of the transverse deflection is required in order to impose the continuity of the deffection and its derivative at the nodes between elements. In the case of the ‘Timoshenko beam theory, there are two second-order equations governing the transverse deflection and the rotation. The weak forms of the equations require Lagrange interpolation of the transverse deflection and rotation. Since the rotation function is like the (negative of the) derivative of the transverse deflection, the degree of the interpolation used for the rotation should be one less than that used for the transverse deflection. Such selective interpolation of the variables is called consistent interpolation. When the same interpolation functions are used to approximate the transverse defiection and the rotation, the resulting stiffness matrix is often too stiff—especially when the number of elements used is small—to yield good solutions. This is due to the inconsis- tency of interpolation of the variables, and the phenomenon is known as shear locking. It is overcome by the use of reduced integration to evaluate the stiffness coefficients associated with transverse shear strains. Both reduced integration elements (RIE) and consistent interpolation elements (CIE) have been developed here. It has been shown that consistent interpolation with linear approximation of the rotation and quadratic interpolation of the transverse deflection yields the same stiffness’ matrix as that obtained with a reduced integration clement with finear approximation of the rotation as well as the deflection. However, the load vector of the consistent interpolation element, for a uniformly distributed transverse load, is equal to that of the Hermite cubic element of the classical beam theory. ‘The plane truss element and frame elements of the classical and Timoshenko beam theories have also been discussed. A plane truss element is4192 wre ELEMENT ANALYSIS OF ONE-DIMENSIONAL PROBLENS ‘a bar element that carries only axial loads and is oriented, in general at on langle from the horizontal axis (the global x axis is taken to Be horizontal). The saa dimont is first modified by adding columns and rows of zeros correspond ing to the transverse displacements, so that it has two degrees of freedom {arial and transverse displacements) per node. ‘The arbitrary plane Wwe crenent is then obtained by transforming the stiffness matrix and foree vector Fame mlement coordinates to global coordinates, which are taken t© be Kotrontal and vertical, Thus, a plane truss element has two degrees of freedom (horizontal and vertical displacements) per node, and carries only set loads. ‘The frame element is @ superposition of the beam and bar ‘Slements, and has three degrees of freedom (axial displacement, transverse stration, and rotation about an axis perpendicular to the plane of axial and ttaneverse coordinates). The general plane frame element js oriented. at an angle from the horizontal position, and its ‘equations ‘are obtained by transforming the equations of the frame element in local coordinates, nally, a procedure for including constraint cénditions among the displacements and/or forces has been presented, PROBLEMS 4,1-4.16, For the beam problems shown in ‘beam element, and give: (a) the assembled stiffness matrix and foree vector; (5) the specified global displacements and fosces, and the equilibrium condiions, £2) the condensed matsix equations for the primary unknowns ie, generleed UTeplacements) ‘and the secondary unknowns (Ke, generalized fores) separately, Solve for the unknown displacements if there are ess than four unknown poacements (use Cramer rue), and evaluate the Bending moment M' druejds? at point C using the finite element interpolation of w (when the ci trensed equations are not solved, express MT in terms of the nodal values of jigs. P4.1-P4.16, use the Buler-Bernoulli fh inh in — EI = constant; Point Cisatx= jh FIGURE Past fo = 400 No TT ae a el Point Cis atx = 7.5m 200 GPa, G = 79 GPa, A = 2.86 x10" ms = FIGURE, P42 20 x 10°F mtBENDING OF BEAMS 193 me) m Moment of inertia J bh Modus E Figure Psa fu TUTTI th — ——-— his Point C is at x = He; EI = constant (Ib in?) FIGURE PA XG 12% 10% psi k= Bin B=30% 106 bin A= ad%; Point-C is atx = iT FIGURE P45 168) ~ fosin J 7 2000 10 Br = constant yl, $6 ft t ‘Bt f, = 1000 £21 = conetaat Point C i at x=10 FIGURE P46 FIGURE P47 1000 1 4 & 400 N mr? 2400 N san (ee | ___s00 8 wt ——— x = oe om maps Fistfstepsnee sity EL = 2041x107 Nm? EL = constant Point Cis atx = 3m Point Cisaty = 250 FIGURE PA. FIGURE P49)194 pure ELEMENT ANALYSIS OF ONEDIMENSIONAL PROBLESS x soon ect eg MQM ag ict 1906 heen whe (CTT La = x UU £ sow ! Fm om tin rT p= 10 8 mt fet & N 2500 N EEE 5 mf — 5 m4 EL = 2% WSN mt; Point Cis atx = 75 m FIGURE P42 eT * a I teats — J HIGURE PA13 fy = 1000 Nm! Rs: | guts oS Lee a S00 N frome are emis (fitch « ze Point Cis j-— ot one FIGURE 4.15 - —4e Moment of inet, 1 10-4 £1 (N m4) = 10 he?BENDING OF REAMS 195 fy kn EE Point Cis at th 7 je EL = constant i ‘Spring stfiness & FIGURE P4.16 the appropriate element). Use the minimum number of elements required in each problem. “Answer: (Problem 4.1) Us = fch'/ABEL, Us= (fe /96B). 47-421. Repeat Problems 4.1-4.5 using the ‘Timoshenko beam element (with reduced integration). Use a value of # for the shear correction factor. Note that accurate results can be obtained only with a suficiently large number of elements, ‘when compared with the Eules~Beraouli element. 4.22. Consider a simply supported beam on an elastic foundation (with foundation modulus £) and subjected to uniform transverse loading. Determine. the transverse displacement at midspan using one Euler-Bernoulli beam element. 4.23. Consider the axisymmetric bending of a linea elastic circular plate of constant thickness. ‘The governing difleremil equation according to. the thin plate assumption i (see Section 7.3 for additonal details) 1 ( ew 2) 14 éw Pa) rap Pw ae Pe ae) rie Oa a where Djs, Dia, and Dzz are the plate material stifnesses (constant), w is the transverse deflection, and r is the radial coordinate. Develop (@ the weak form of the equation over a typical element "= (r4,r»); the ‘quantities in parentheses should not be integrated by parts; (6) the finite element model of the equation in the form (te) = 79 +10) Make sure that (K‘] is symmetric (ie., the bilinear form in (a) should be symmetric). Comment on the interpolation functions that are admissible for the clement. 4.24, The differential equations governing axisymmetric bending of circular plates according to thick plate theory are booed Dea) east where Ay, Diy, Dr, and Ds, are plate material stiffnesses, W is the rotation function, w is the transvetse deflection, and fis the transverse load. Develop (@) the weak form of the equations over an element; (6) the finite element model of the equations,196 rir eLeMeNT ANALYSIS OF ONE-DIMENSIONAL FROBLEDS 44a8, Solve the problem of a thin (radiustothickness ratio r/t= 100) clamped ipotropic (Y= 0.3) ciular plate with uniformly distributed load using the fervent Neveloped in Problem 4.23. Exploit symmetry, and use (wo thin plate (Euler— Bernoulli) clements in the computational domain, 1426, Solve the circular plate problem of Problem 4.25 using a two-element mesh of ‘Timoshenko elements (use & = 2). 427. Consider the fourth-order equation (4.1) and its weak form (4.4), Suppose that 8 come element is employed, with three primary variables at each node: (8,8), where Pw/de?. Show that the sssociated inter- polation (Hermite) functions are given by aos?h? 220sth = 52s! _ Bie = 07H + AGTH = 6 o oy + = De ath? — 30h + 3x4 p= > _ ae PP = Math? 46h sath be oa a where x is the clement coordinate with the origin at node 1, Also compute the element stiffness matrix and force vector. 4428. Consider the weak form (4.4a) of the Evler-Bernoulli beam element. Uss a Tceinode clement with two degrees of freedom (Ww, 8), where O~ —dwidx. Derive the Hermite interpolation functions for the element. Compute the clement stifiness matrix and force vector. 1429-4:36. For the truss and frame problems shown in Figs. P4.20-P4.36, give (a) the erformed element matrices; (b) the assembled clement matrices; (@) the condensed matrix equations for the unknown generalized forces. isplacements_ and 2. Py a0 kins Ay = 3 ink ‘int 30,000 ksi 3 jon I FIGURE P4.29 Py = 8 hips EA isthe same for all Pee 8kips members = 30 x 106 psi, v= 03 yon A= 308 4 S—— eto t-afet0 4 FIGURE P4.30 .ENDING OF meas 197 Py = Dips F230 WF ps v 2 03 T Aye Ag= Ay = 5? aol Are Acm de = An = Gin? Ay = Ay = in? BA aR P, = 19 kips| Leas eo IS APL ts ad FIGURE Psat si | es E ya30 = 200 Gre ow A= 10 om? El, EA are the J = 10 mam? same for the to members FIGURE P432 v= 03 B= 30% 10° pst Skips 1-1 (43.78 KN mi) T= 100 iat A= 100 in? FIGURE P444198 Aire ELEMENT ANALYSIS OF ONE-DIMENSIONAL FROBLENS 1.5 hips #71 (21.9 KN me) 203 Boao x 6 pt 6 (1.83 On sips G56 1] Anion ‘Game forall) FIGURE P4.35 El, BA are the samo for both members v=03 anton FIGURE P6436 “Answer: (Problem 4.29) Uy=0.022973in, U, 9428 Ib, Pi= 7454 1b. =0.002169in, Pi= REFERENCES FOR ADDITIONAL READING sudynes, RG: Adeonced Srongih end Angled Ses Analyt, Metiraw 3) New Yoee 1977. ame Tad LH. Shames Sold Mechanic: A Variatonal Approach, Mera lily New ‘York, 1973. Rediy 0s Energy and Variational Methods i Applied Mecha, Sohn Wily» New Yate ‘98. ‘Tiomonto, S.P.,and J. N. Goodies Theory of Blastcty, MeGraw-Hil, New Yorks 1, TA and SK. Retr: Advanced Srengh and Applied Etat, Pues, New “York, ists Voter, and 3H Gaines: Advenced Seng of Matera, Prentice Eal, Englewood CS, Ni, 1931 ang Shes and CG. Selon: Introductory Senna Ansys, Prentice Hal, Englewood NS, 1988 Wiens ner wed WM, Losas, 2 Stucural Anais for Engineers, MeGrew ill, New Yor 1978.CHAPTER 5 FINITE ELEMENT ERROR ANALYSIS 5.1 APPROXIMATION ERRORS ‘The errors introduced into the finite element solution of a given differential equation can be attributed to three basic sources: 1, Domain approximation error, which is due to the approximation of the domain, 2. Quadrature and finite arithmetic errors, which are due to the numerical evaluation of integrals and the numerical computation on a computer. 3. Approximation error, which is due to the approximation of the solution Gee (iv) in the Note to Remark 10 in Section 3.2.7): uem=3 ve, (1) where U; denotes the value of u at global node J, and @, denotes the global interpolation function associated with global node J (sec Fig. 3.36), In the one-dimensional problems discussed thus far, the domains considered have been straight lines. Therefore, no approximation of the 199200 nine ELEMENT ANALYSIS OF ONE-DIMENSIONAL PROBLEMS domain has been necessary. In two-dimensi nal problems involving nonrectan- gular domains (as will be seen in Chapter 8), domain (or boundary) eproximation exrors are intfoduced into the finite element solutions, Tn general, these can be interpreted as errors in the specification of the data of the problem because we are now solving the given differential equation on a modified domain. As we refine the mesh, the domain is more ‘accurately represented, and, therefore, the boundary approximation errors are expected to approach zero. When finite element computations are performed on a computer, round-off errors in the computation of numbers and errors due to the humerical evaluation of integrals are introduced into the solution. In most linear problems with a reasonably small number of total degrees of freedom in the system, these errors are expected to be small (or zero ‘when only a certain decimal point accuracy is desired). o “The error introduced into the finite element solution U* because of the approximation of the dependent variable w in an element ‘Q¢ is inherent to any problem wry = 2D MVE D Uhr 6.2) where uy is the finite clement solution over the domain (14 Ut in Q°), Nis the number of elements in the mesh, M is the total number of global nodes, and nis the number of nodes in an clement. ‘We wish to know how the error =u ty, measured in a meaningful way, behaves as the number of elements in the mesh is increased. It can be shown that the approximation error is zero for the single second-order and fourth-order equations with element-wise- constant coefficients [see (5.30)-(5.35)]- 5.2 VARIOUS MEASURES OF ERRORS ‘There are several ways in which one can measure the “difference” (or distance) between any two functions wand uj. The poinovise error is the Gifference of u and 1, at each point of the domain. One can also define the difference of w and tu, to be the maximum of all absolute values of the differences of u and 1, in the domain 2= (a, b): le ln ma, jeu(e) — weal) 6.3) ‘This measure of difference is called the supmetric. Note that the supmetric is & real number, whereas the pointwise error is a funetion and does not qualify as iniistance of norm in a strict mathematical sense. ‘The norm of a function is a ‘non-negative real number. ‘More generally used measures (or norms) of the difference of two functions are the energy norm and the L> norm (pronounced "'L-two norm”). For any square-integrable functions w and us defined on the domain @=RTE ELEMENT ERROR ANALYSIS 201 (a, ), the two nornis are defined by encrny norm — hid = (S| 6-4) norm Ike sto= (f° uu? de) 65) where 2m is the order of the differential equation being solved. The term “energy norm” is used to indicate that this norm contains the same-order derivatives as the quadratic functional (which, for most solid mechanics problems, denotes the energy) associated with the equation, Various measures of the distance between two functions are illustrated in Fig. 5.1. These definitions can easily be modified for two-dimensional domains. 5.3. CONVERGENCE OF SOLUTION ‘The finite element solution wy in (5.1) is said to converge in the energy norm to the true solution w if lle tlle eh? for p>0 6.6) where ¢ is a constant independent of uv and u,, and fh is the characteristic length of an element. The constant p is called the rate of convergence, Note that the convergence depends on t as well as on p; p depends on the order of the derivative of w in the weak form and the degree of the polynomials used to approximate u [see (5.15) below]. Therefore, the error in the approximation can be reduced either by reducing the size of the elements or increasing the degree of approximation. Convergence of the finite element solutions. with mesh refinements (ie., more of the same kind of elements are used) is termed A-convergence. Convergence with increasing degree of polynomials is called P-convergence. (a 400) _ Ms) aly Maximam [5 La norm (area shaded) FIGURE 8.1 Dilflerent measures of error B= uu, besween the exact solution and the finite element solution 1, _ g The maximum norm and the Dy 3 oem are illustrated,202. nuts ELEMENT ANAL 5.4 ACCURACY OF THE SOLUTION Returning to the question of estimating the approximation error, We consider a Reet. differential equation in one dimension (m= 1, second-order ‘equations; m= 2, fourth-order equations): S cows lost)=r for 0 T(u) (5.14) ‘The equality holds only for w=, Equation (5.14) implies that the conver- gence of the energy of the finite clement solution to the true energy is from above. Since the relation ia (5.14) holds for any u,, the inequality also indicates that the true solution « minimizes the energy. A similar relation can be established for the fourth-order equation (1m =2) Now suppose that the finite element interpolation functions ©; (= 1,2,...,M) are complete polynomials of degree k. Then the error in the energy norm can be shown to satisfy the inequality [see Reddy (1986, 1991), p. 401} lel where ¢ is a constant. This estimate implies that the error goes to zero as the pth power of has h is decreased (or the number of elements is increased). In other words, the logarithm of the error in the energy norm versus. the logarithm of ft is a straight line whose slope is k+1—m. ‘The greater the degree of the interpolation functions, the more rapid the rate of convergence. Note also that the error in the energy goes to zero at the rate of k-+1—m; the error in the L, norm will decrease even more rapidly, namely, at the rate of +1, i.¢., derivatives converge more slowly than the solution itself. Error estimates of the type in (5.15) are very useful because they give an idea of the accuracy of the approximate solution, whether or not we know the true solution. While the estimate gives an idea of how rapidly the finite element solution converges to the truc solution, it does not tell us when to stop refining the mesh. This decision rests with the analysts, because only they know what a reasonable tolerance is for the problems they are solving. ‘As an example of estimating the error in the approximation, i.e. (5.15), consider the linear (two-node) element for a second-order equation (m = 1). We have for an element [etal Sch’, p=kti-m>o (5.15) y=. -5) + us 6.16) where s=/h and & is the local coordinate. Since uz can be viewed as a funetion of u, via (5.16), one can expand us in a Taylor series around node 1 to obtain zany tult dels... 6.17)1204 NITE, ELEMENT ANALYSIS OF ONEDIVERSIONAL THOBEENS where u’ = du/ds. Substituting into (5.16), we obtain tig = uy tus + huis + (5.18) Bapanding the true solution ina Taylor series about node 1, we obtain uy tuis huis +. (6.19) ‘Therefore, we have, from (5.18) and (5.19), i ah —s%) max |“ jy uh<3 6-9) ax | (9-9)? max iS (5.20) 2 oes 4 (ge-w)<3 ‘These lead to wus cake mal seat 6.2) where the constants ¢; and e2 depend only on the length L of the domain, “ie ceader may carry a similar etror analysis for the fourth-order ‘equation, Example 5.1. Here we consider a computational example to verify the #05 estimates in (22). Consider the differential equation 2 for 0, LE) Gi If the finite element approximation mis an interpolant of the true soluton vor sean ape enor in the solution and the energy of 3-1) when (a) is Tinea Spterpotant and (D) ti @ quadratic interpola. as tigen reesences at the end ofthe efpter to determine the inverplation ero in tums ofthe mesh parameter (spacing) hy and use the result Teltermine the error inthe energy in the form NEW = coh” where ey and pare constants; p determines the rate of convergence (P > 0) oF divergence (p <0) REFERENCES FOR ADDITIONAL READING iat, P. Ges The Finite Element Method for Eliple Problems, Nos-Hollany Amsterdam, 17. Davies nis The Finite Element Method, A Fist Approach, Clarendon Press, Oxford, 19, Ra hand J. N- Redys An Invroducin tothe Mathematcol Theory of Finte Elerion, Wiley Interscience, New York, 1982, eddy") Me Apia’ Penton! Arabs and Velational Methods gineering, McGraw Xi ‘New York, 1986; Krieger, Melbourne, FL, 1991 Rektone, Kes Variational Methods in Matiemaice, Science and Engineering, Reich, Rosco, is77. sora. and G. 5.x An Anas of he Fine Element Method, Prentice Hal, Enalovond ‘Clif, NI, 1973. Way Re abd AR. Nchel: Fine Element Anas and Appictions, John Wily, New Yor 18.CHAPTER 6 EIGENVALUE AND TIME-DEPENDENT PROBLEMS 6.1 EIGENVALUE PROBLEMS 6.1.1 Introduction Determination of the values of the parameter A such that the equation Alu) = AB(w) 6.1) where A and B denote linear differential operators, has nontrivial solutions w is called an eigenvalue problem. The values of 4 are called eigenvaities and the associated functions u are called eigenfunctions. For example, the equation Pu @ ~The, with A= S, a th Anas which arises in connection with the axial oscillations of a bar or the transverse oscillations of a cable, constitutes an eigenvalue problem. Here 4 denotes the square of the frequency of vibration . In general, the determination of the eigenvalues is of engineering as well as mathematical importance. In structural problems, the eigenvalues denote either natural frequencies or buckling loads. In fluid mechanics and heat transfer, eigenvalue problems arise in connection with the determination of the homogeneous parts of the solution. In these cases, eigenvalues often denote Bol 209210 anire ELEMENT ANALYSIS OF ONE-DINIENSIONAL FROBLENS amplitudes of the Fourier components making up the solution. Eigenvalues are also useful in determining the stability characteristics of temporal schemes, as discussed in Section 6.2. Tn this section, we develop finite element models of eigenvalue problems. In view of the close similarity between the equations of eigenvalue and boundary value problems, the steps involved in the construction of their finite element models are entirely analogous. Differential eigenvalue problems are Seduced to algebraic eigenvalue problems by means of the finite element approximation. ‘The methods of solution of algebraic eigenvalue problems are then used to solve for the cigenvalues and eigenvectors. 6.1.2 Formulation of Eigenvalue Problems Consider the parabolic partial differential equation u_ 3 (pq 3! . pea Zi 5-(kaZ) = e659 62) a plane Wp Here u denotes the temperature, K the thermal conductivity, p the Meaviy, A the cross-sectional area, ¢ the specific heat, and q the heat Generation per unit length, The Homogeneous solution (i... the solution when wert) of (62) is often sought inthe form of a product of a function of x and a {ction of f (-e., through the separation of variables technique) u(x, )= SOTO (63) Substitution of this assumed form of solution into the homogeneous form of (6.2) gives at 4 (,,8\q cash £ (kaSl)r=0 Separating the variables (assuming that peA and KA are functions of x only), dT 114 (4S) Tae ped Sax ae (64) Note that the left-hand side of this equation is a function of ¢ only while the righthand side is a function of x only. For two functions of pwo independent variables to always be equal, both must be equal to the same constant, say —Ai 1ar_ 11d (,,45 ot" pease) sa) or a 4,44 oa, ~ 4 (ka Gh) ~Aecas =0 6.5)EIGENVALUE AND TIMEDEPENDENT PROBLEMS 211 ‘The negative sign in the constant 2 is based on the physical requirement that the solution $(z) be harmonic in x while 7(/) must decay exponentially with increasing t. The solution of the first equation is 7'= Tje"™. When k, A, p, and ¢ are constants, the solution of the second equation is, S(x) = By sin Vix +B, cos Vix, a k ‘The constants 2, 7, By, and B, are determined with the help of initial and boundary conditions. In View of the above discussion, the solution of (6.2) is of the form uC, #) = UGe™ (6.6) This form is consistent with the solutions we derived above, with U(e)=S(x)To. Substituting (6.6) into the homogeneous form of (6.2), we obtain = ApcAU(xye™=0 or wy ka) — Apcau=0 61 We wish to determine 4 and nonzero U(x) such that (6.7) holds and the boundary conditions of the problem are met. Equation (6.7) describes an eigenvalue problem, & being the eigenvalue and U(x) the eigenfunction. ‘The axial motion of a bar can be described by the hyperbolic equation Fu_3/,, du AS pate = ale.) (6.8) Here « denotes the axial displacement, E the modulus of elasticity, A the ‘cross-sectional area, p the density, and q the axial force per unit length. ‘The natural axial oscillations of the bar arc periodic, and they can be determined by assuming a solution of the form ue, = U@)e™, with i= V1 (69) where w denotes the frequency of natural axial motion (or vibration), and U(x) denotes the configuration of the bar, called the mode shape, during the vibration. For each value of a, there is an associated mode shape. Substitution of (6.9) into the homogeneous form of (6.8) gives [-eaorw (ea!) J212 anit: eLENENT ANALYSIS OF ONEDIMENSIONAL PROBLEMS (6.10) where A= 0% Equation (6.10) is an eigenvalue problem, which involves Yctermining the square of natural frequencies 2 and mode shapes U. “This equation also arises in the solution of (6.8) by means of separation Of siabtes, as discussed in connection with the parabolic equation. Note (hat (6.7) and (6.10) are of the same form. Only the coefficients are different Equations similar to (6.10) can be derived forthe transverse vibrations of a beatn using the Fuler-Bernoulli or the Timoshenko beam theories. For the Euler-Bernoulli beam theory, we assume we n= woe “where a isthe frequency of natural transverse motion and W(x) is the mode Sues of the tansveise motion, Substittion of this form into the equation of notion of the Euler-Bernoulli beam theory aw, ® (p 8W atl )=0 EL ISS, il) gives (6.12) here A= 0, For the Timoshenko beam theory, we assume we, D= Woe, WO, A= Sexe" ‘and substitute into the equations af motion of the theory pA oe e [oax(Se+ w)|-0 as ort 2 (ert) cak(S4¥)= (6.13) oF x to obtain the eigenvalue problem ~ Zl cax( GE +s)]-Aeaw=0 eo — A (er'8) « cax(Ge+s)— A015 0 ‘The study of buckling of beam-columns also leads to an eigenvalue problem, For example, the equation governing the equilibrium of beamEIGENVALUE AND TIME-DEPENDENT FROSLENS 213 subjected to an axial force P, according to the Euler-Bernoulli beam theory, is (6.15) which is an eigenvalue equation with A= P as the eigenvalue, which represents the buckling load, Often one is interested only in the smallest value of 2, called the critical buckling load. For the Timoshenko beam theory, the buckling equations are ~ A feax( tts) vee ae de an -f(a B+ oax(? +w)-0 This completes the formulation of eigenvalue problems associated with the model problems studied in this book. In the next section, we develop the finite clement models of (6.7), (6.10), (6.12), and (6.14)-(6.16). 6.1.3 Finite Element Models ‘An examination of the eigenvalue equations derived in the previous section shows that they are a special case of the equations studied in Chapters 3 and 4. For example, consider ay dw ~E(eP)-1 17 ‘The eigenvalue equation associated with this is ay au ~E(eQ)=2000 (6.18) where a and ¢o arc quantities that depend on the physical problem: for heat transfer, a=kA, c= pcA where c is the specific heat, while, for a bar, ‘A, = pA a= Similarly, the eigenvalue equations associated with the transverse vibrations and buckling of beams are special cases of their static counterparts. Therefore, the finite element models of the eigenvalue equations can readily be developed. It is important to note that the spatial derivative operators of the static (i.e., non-time-dependent) and eigenvalue equations are the same. The difference between (6.17) and (6.18) is that, in place of the source term f, we hhave AcgU in the eigenvalue equations. This difference is responsible for214 pore: ELEMENT ANALYSIS OF ONEDIMIENSIONAL PROBLEMS ‘another coefficient matrix, in addition to the usual coefficient matrix [K°}, in the eigenvalue problems. The derivation of the finite element models of eigenvalue equations is presented next. ‘Over a typical element 2°, we seek a finite clement approximation of w in the form U -2 iyi) (6.19) "The weak form of (6.18) is o= [aM regu) dx -Q.wles)— Outta) (6208) where w is the weight function, and Q; and Q, are the pusual secondary variables (Q/=0,i #1, ") Be an Substitution of the finite element approximation into the weak form gives the finite element model of the eigenvalue equation (6.18): KH) - AME) = (2) (21) where Ce P coveviar (6.200) Equation (6.2la) represents the finite element model of the eigenvalue equations (6.18) and (6.10). ‘The finite element model of (6.12) is (K*Hut} — AM }u*) = {0°} (6.22) where {u*) and (Q*) ate the columns of nodal generalized displacement and force degrees of freedom of the Euler-Bernoulli beam element: my (ieler*)), (eG), where the subscripts 1 and 2 refer to element nodes 1 and 2 (at x =x, and aia). The matrices [K"] and (€M'], known as the stifiness and mass matrices,EIGENVALUE AND TIME-DEFENDENT PromLes 215 are defined by oo [228 a, en [oaorge Kj fare Gas, Mia ["“nasiogds (6220 where @f are the Hermite cubic interpolation functions (sce Chapter 4 for details). ‘The finite element model of (6.14), with equal interpolation of w and S, CR"] TREN Ow)) PTO) fe) _ Vt (her tellisod ch (0) welliss} {is} 6.230) where Avi Kp [can Mas, Kp [capac Kp [" (can yiyjs ee) a dx dx oi=["oavivras, Mp=["oivividy (622) vi [-eax(G+s)]., vs=[oax(S'+s)] ut (-2), Note, that for sufficiently large ratios of length L to height H of the beam, the Timoshenko beam element gives the results of the Euler-Bernoulli beam element. For example, for L/H 100, the effect of shear deformation is negligible, and both elements give approximately the same solution. ‘The finite clement models of (6.15) and (6.16) arc the same as those in (6.22a) and (6.23a), respectively, with A=P, and [M*] and [4f"] (and [M] = [0]) replaced by [G"] and [G"], respectively, where sgt das ody do [oe sa, [eta (6.24) ‘i ‘The coefficient matrix (G*] is known as the stability matrix. The numerical form of the stiffness matrices [K*] were given in the previous sections [see, e.g., (3.34a), (3.40a), (4.15), and (4.98)]. Expressions for the mass matrices Mf and stability matrices Gj for the Lagrange elements are also available from the previous derivations [sec, e.g., (3.34a) and (3.35)]. ‘The mass and stability matrices for the Hermite cubic interpolation, _ se apedes Mi= |" eeiofan, j= [a Seas (625)216 NITE ELEMENT ANALYSIS OF ONEDDINENSIONAL FROBLENE can be evaluated numerically for element-wise-constant values of ¢, and de: 186 -22h 54 13h h| 22h 4h? 13h 3h? (I= 750) 54 13h 156 22h (6.260) 13h 3h? 2h 4h? 36 -3h -36 -3h a | n3h ah IP C1=305| 36 3h 36 3h (6.260) 3h =k? 3h Ane ‘The assembly procedure and imposition of boundary conditions‘in the eigenvalue analysis remains the same as for static problems. The boundary conditions are necessarily homogeneous. , ‘The standard crror estimates for the fundamental eigenvalues _and eigenfunctions of the problems discussed here are [see Strang and Fix (1973)] HOSA SAO 4 Chae MIAO|RMIM Ij! hg cn MAO] (627) where (A®, w°) is the exact solution, (A, w) is the finite element solution, m is the order of derivatives appearing in'the weak form, k is the degree of polynomials used in u*, and ft is the characteristic length of the element (see Chapter 5 for the notation and definitions of errors). 6.1.4 Applications Here we consider a couple of examples of eigenvalue problems to illustrate the concepts described in the previous section. We consider one example of heat-transfer-type problem and one of free vibration of beams. Example 6.1. Consider a plane wal, ini both surfaces suddenly exposed to a equation ly at a uniform temperature Zy, which has at temperature 7. The governing differential (6.28) T(x, 0)=% where ais the diffusion coefficient, & the thermal conductivity, p the density, and ¢, the specific heat at constant pressure, Equation (6.28) is also known as the diffusion equation. We consider two sets of boundary conditionsTIGERVALUE AND -TIME.DERENDEY PRoRLENS 217 T0d=t, TMad=h. bere0 ro) root, [iaZepar-n]] 20 an Equation (6.28) can be normalized while making the boundary conditions ‘homogeneous. Let ou wT ae The differential equation (6.28), boundary conditions (6.29), and the become (6200) (608) pL & (6) Where the bars over x and f are omitted in the interest of brevity. Solution of (6.30a,b) by separation of variables leads to solution of the eigenvalue problem u@= uy =0 ‘The finite element mode! of this equation is given by (6.21), with a= 1 and ¢ the choice of linear and quadratic interpolation functions, the element equations (6.21a) become {see (3.34) and (3.40) Gls TEL DOD =(E} mecenenn 7-8 4 4 2 -1]\fur ga Eft ae af ea] 2 6 2] L TTT usa samery “Li -8 7, 12 4, us, Q%. For a mesh of two linear clements (the minimum needed), with h assembled equati if? V\fa] fai 14 1lfiat=i0 oi 2sla) lo ons U(0) =0 and U(1) =0 requite Uj = U,=0, Hence, the eigenvalue problem reduces to the single equation G-H)U,=0, or A= 120, #0248 myire ELEMENT ANALYSIS OF ONEDINENSIONAL ROBLENS For a mesh of one quadratic element, we have (= 1.0) B—hal6=0, or 4,=10.0 ‘The coscesponding eigenfunction amplitude is U,= 1.0 (or any nonzero constant), 50 that the eigenfunctions are as follows: for linear elements (h = 0.5), for 05x=0.5 ue@)= Uy or 050, itis valid for = 4, and t= és ('] = [MDs EM) (4). + [KL ade = (Fhe (6.412) IM {it} t Khar ldeer = Pon (6.416) where it is assumed that the mass matrix is independent of time, After multiplying (6.39a) throughout by Af,,1, we premultiply both sides of the ‘equation with [M) and obtain Atees LM) i) saa + Ategs CL a LMI}, = AAC deoa— (D2) Substituting for [M]{W), «1 and [M]{d}, from (6.414) and (6.41b), respectively, in the above equation, we obtain Atser O((Fhovs [Kh aaf}oas) + Aten = a) (Fs (I (.) = [MIC eos— (0)EIGENVALUE AND TIME-DEPENDENT PROBLEMS 229 Rearranging the terms into known and unknown ones, we obtain (6.422) where Rha }+atKha, (&h {Phovos = tess la(P}oo2 4 (1 a){F}s] (6.426) MOA a2=(1~ a) Atyyy Note that, in deriving (6.42), it has been assumed that (M'] is independent of time and that the time step is nonuniform. Equations (6.42) are valid for a typical element. ‘The assembly, imposition of boundary conditions, and solution of the assembled equations are the same as described before for static (or steady-state) problems. Calculation of [8] and {f) at time ¢ = 0 requires knowledge of the initial conditions {u}o and the time variation of (F). Note that, for a~0 (the forward difference scheme), we obtain (&]=[M"]. When the mass matrix [§M*} is diagonal, (6.422) become explicit, and one can solve for {1,1 directly without inverting [£]. ‘The mass matrix obtained according to the weak form, called the consistent mass matrix, is not diagonal. There are several ways to diagonalize mass matrices; these will be described in Section 6.2.4 Stability and accuracy, Since (6.392) represents an approximation, which is used to derive (6.422), error is introduced into the solution (1},,1 at each time step. In addition to the truncation error introduced in approximating the derivative, round-off errors can be introduced because of the finite arithmetic used in our computations, Since the solution at time t,4, depends on the solution at time ¢,, the erzor can grow with time. If it grows unboundedly with lime, the solution scheme is said to be unstable. If it is bounded (ic., it increases for one time step and decreases for another time step, but never exceeds a certain finite value), the solution scheme is said to be stable. The numerical scheme (6.42) is said to be consistent with the continuous problem (6.38) if the round-off and truncation errors go to zero as At—>0. Accuracy of @ numerical scheme is a measure of the closeness between the approximate solution and the exact solution, whereas stability of a solution is a measure of the boundedness of the approximate solution with time. As one might expect, the size of the time step can influence both accuracy and stability. When we construct an approximate solution, we like it to converge to the true solution when the number of elements or the degree of approximation is increased and the time step Ar is decreased. A time approximation scheme is said to be convergent if, for fixed t, and At, the numerical value {u}, converges to its true value (u(t,)} as At—0. Accuracy is measured in terms of the rate at which the approximate solution converges. If a numerical scheme is stable and consis- also convergent [see Isaacson and Keller (1966)].. A numerical scheme is said to be conditionally stable if it is stable only230 nine ELEMENT ANALYSIS OF ONE DIBIENSIONAL FRODLENS when certain restrictions on the time step are satisfied. For all numerical Schemes in which a' (Ata (6.55) 6.2.5 Applications Here we consider two examples of applications of finite element models of one-dimensional problems. Problems are taken from heat transfer and solid mechanics. Other field probiems can be related to heat-transfer-type problems Example 6.3. Consider the transient heat conduction problem with equ ou a 0 for 00 (Le. 1> 0) while the initial condition requires (Wont, (Udo=t for snd since the intial condition should be consistent with the boundary conditions, we take {ue-0.. Using the boundary conditions, we can write for the one-clement ‘model (h=1.0) (+ 0! Udan[H- 6-95 | 59)EIGENVALUE AND TINE-DEPENDENT PRORLEMS 235 which can be solved repeatedly for Us at different times, s=0, 1, -..+ Repeated use of (6.59) can cause the temporal approximation error to grow with time, depending on the value of a. As noted eatlier, the forward difference scheme (a=0) isa conditionally stable scheme. The critical time step is given by te= 2 ‘where Aas the maximum eigenvalue associated with (6.57): ~AMY(U) + [KIO = (0} For the model at hand, this reduces to RU, +AU; =0, or L=3fh?=3 Hence At,,= 0.6667. Thus, in order for the forward difference solution of (6.59) to be stable, the time step should be smaller than A‘.,= 0.6667; otherwise, the solution will be unstable, as shown in Fig. 6.7. For unconditionally stable schemes (a=), there is no restriction on the time step. However, to obtain a sufficiently accurate solution, the time step must be taken as a fraction of Afg. Of course, the accuracy of the solution also depends on the mesh size h. As this is decreased (Le., the number of elements is increased), At decreases. Figure 6.8 shows plots of w(1, :) versus time for a =0.5 and At=0.05. Solutions predicted by meshes of one and two linear elements and the mesh of one quadratic element are compsred with the exact solution. ‘The convergence of the solution with increasing number of elements is clear. The finite element solutions obtained with diferent methods, time steps, and meshes are compared with the exact solution in Table 63, 7 ‘Time # FIGURE 6.7 ‘Transient solution of a parabolic equation accocding to the forward difference scheme (At 0.675, ‘one linear element). The solution is unstable because the time step i lrger than the cial time step,1236 wire: ELEMENT ANALYSIS OF ONE-DIMENSIONAL FROBLEDSS wh Analytical 10 iL nears 10 es ar = 0.05 ) anos 0.6: 04 02 0.04 —.——- oon, Time 1 FIGURE 68 Tra i solution of a parabolic equation according to Tinear and quadratic finite elements TADLE 63 A Womparison of the finite clement solutions obtained using various time ‘approximation schemes and meshes with the analytical solution of a parabolic in arising in conductive heat transfer 0.05 08500 0.8696 0.8605 1.0359 0.9951 0.9933 oes Oon08 07561 O70 0.9279 0.9588 0.9354 19819 9550 O9519 0.9993 GIS Outil 6575 0.6371 0.8169 0.8639 0.8707 0.8683 0.8801 O.8725 0.8642 Gap 95220 OSTE 0.5482 0.7176 0.7557 0.7694 0.7679 0.7689 OTTSL 0.7723 Coe Dads7 04972 OATIT 0.6300 0.6759 0.6828 0.6780 0.6933 0.6855 0.6854 Ga 02771 04523 4059 0.5533 0.5905 0.6037 0.5987 0.6006 0.4070 0.6068 D3 03206 03759 MP2 0AR5H 0.5250 0.5325 0.5286 0.5394 0.5958 0.5367 oo 0.2725 114266 0.4608 OATS 0.4658 OATIO DATAL 0.4745 043. 0.2316 03786 04083 04158 0.4121 0.4201 0.4188 0.419 050 0.1969 0.228) 0.3592 0.3676 0.3639 0.3687 0.3701 0.3708 055 0.1673 0.2888 0.3176 03247 03213 0.3275 (0.3213 0.3277 Gad 01422 0.1969 O.16K7 0.2536 0.2798 0.2868 0.2837 0.2883 0.2890 0.2877 Des 1209 01625 LIS 0.2227 0.2472 0.2535 0.2505 0.2556 0.2556 0.2501 Ooo 0.1028 OMS 1220 0.1985 0.2180 0.2238 0.2212 0.2253 0.2258 0.2264 O15 O84 0.229 0.1050 0.1717 0.1924 0.1979 0.1953 0.1995 0.1996 0.2001 Gad 0.0743 0.1069 0.03 0.1508 0.1697 0.1747 0.1725 O.ATEL 0.1764. 0.176) GSS oot 0.0909 0.0777 0.1324 0.1498 0.1584 0.1523 0.1557 0.1559 0.1563 50 00535 0.0608 0.0669 0.1162 0.1322 0.1363 0.1345 0.1375 0.1378 0.1382 Dots 0.0003 0.0575 0.1020 0.1166 0.1205 0.1187 0.1216 0.1218 0.1222 DesBs 0.0611 0.0195 010896 0.1029 0.1065 0.1048 0.1074 0.1076 0.1080[IGERVALUE AND TIME-DEPENDENT PROBLEMS 237 ‘The next example concerns a hyperbolic equation. Example 6.4, Consider the transverse motion of @ beam clamped at both ends, according to the Euler-Bernoulli beam theory, 6 for o Hv) 2) where F; denotes the value of F(x) at the /th point of the interval [x4, x9] and (x) are polynomials of degree N~1. The representation can be viewed as the finite element interpolation of F(x), where F; is the value of the function at the Ith node. The interpolation can be of the Lagrange type or the Hermite type. Substitution of (7.2) into (7.1) and evaluation of the integral gives an approximate value of $. For example, suppose that we choose linear interpolation of F(x). Then N= (xy ~x)/h, r= (x—xa)/h, and S=1K +R), R=F0u), Ao F(a) 7.3) Thus, the value of the integral is given by the area of a trapezoid used to approximate the area under the function F(x) (see Fig. 7.1). Equation (7.3) is known as the trapezoidal rule of numerical integration. If we use the Lagrange quadratic interpolation of F(x), we obtain SatA+4E+K), F= FO), Fy= FQ, +4h), F(xp) which is known as Simpson’s one-third rule. Equations (7.3) and (7.4) represent the form of numerical quadrature formulae. In general, a quadrature formula has the form (4) s2[" Fede S reap, @s) Fe) FIGURE 7.1 ‘Approximate et tegral using the tcopezcidal rule: (2) ‘Two-point formula; (6) three-point formula248 rNire ELEMENT ANALYSIS OF ONE DIMENSIONAL PROBLEMS where x; ate called the quadrature points and W, are the quadrature weights. ‘These formulae require funetional evaluations, multiplications, and additions to obtain the numerical value of the integral. They yield exact values of the integral whenever F(x) is a polynomial of order r— 1. In this section, we describe several numerical integration techniques and formulations in which the geometry as well as the dependent variables are approximated using different degrees of polynomials. We begin with the discussion of a local coordinate system. 74.2 Natural Coor Of all the quadrature formulae, as will be discussed in the subsequent sections, the Gauss-Legendre one is the most commonly used. The details of the method itself will be discussed shortly. The formula requires the integral to be cast as one to be evaluated over the interval (~1,1]. This requires the transformation of the problem coordinate x to a local coordinate & such that (ee Fig. 7.2): when x=x4, & ates 5; when x=xp, & ‘The transformation between x and & can be represented by the linear “stretch” transformation xsatbe where a and 6 are constants to be determined such that the above conditions hold: x4=atb(-l), xe=@+(1) Solving for a and b, we obtain = Hep —4) =the, =n $44) =A + Hie Hence the transformation takes the form xen tht 8) 6) where x, denotes the global coordinate of the loft end node of the element and hi is the element length (see Fig. 7.2). The local coordinate & is called the normal coordinate or natural coordinate, and its values always lie between —1 and 1, with its origin at the center of the element. ‘The local coordinate & is useful in two ways: (j) it is convenient in constructing the interpolation functions; (ii) it is required in numerical integration using Gauss-Legendre quadrature, p-1 tat be FIGURE 722 1» Global coordinate x, local coordinate ¥, and €= 1° gormalized local coordinate &NUMERICAL INTEGRATION AND COMPUTER IMPLEMENTATION 249 ‘The derivation of the Lagrange family of interpolation functions in terms of the natural coordinate & is made easy by the property 1 in (3.28) of the interpolation functions: 1 itis, (Ei) = ‘ (7. where & is the & coordinate of the jth node in the clement, For an element with n nodes, (= 1,2, ...,n) ate functions of degree » — 1. To construct ‘yrsatistying (7.7), we proceed as follows. For each yi, we form the product of n—1 linear functions E-8 Um 1Q..5f- LIL... Jd: w= olG— BYE ~ 8) (E~ BE Bid EE) Note that ; is zero at all nodes except the ith. Next we determine the constant Gysuch that y= at = &: (Gi ENG 8) BBG Gard > ‘Thus, the interpolation function associated with node (E — SiS ~ &2) - =» (E ~ FE ~ Seg) -* (E— &) i EME: — Ba) = + Er Gis) Bins) Ge Sn) Interpolation functions that satisfy (7.7) are said to belong to the Lagrange family of interpolation functions, and the associated finite clements belong to the Lagrange family of finite elements. The interpolation functions, in (3.166) and (3.18) provide an example of the Lagrange interpolation functions (n = 2). Figure 7.3 shows the linear, quadratic, and eubie Lagrange interpolation functions expressed in terms of the natural coordinate (for equally spaced nodes). (&-&)1* a is wi(5)= t (78) 7.1.3 Approximation of Geometry Recall from (3.31c) and (4.12c) that the element matrices involve the derivatives of the interpolation functions with respect to the global coordinate hee o weU+Ol-H we wrD ” al = 6S + OG = 8) SURE 7.3, =H + OC OG 6) Legeange family of LD ioerpol Y= 80+ OU OE +) tion functions in tems of the nor = HGH GB +g) malted coordinate,250) rIsITe ELEMENT ARALYSIS OF ONE-DININSIONAL, PROBLEMS x, A transformation of the form xf) (7.9) js required in order to rewrite the integrals in terms of § (-1<§=1). The fanction fis assumed to be @ one-to-one transformation. An example of f(E) is provided by (7.6): FG) = Ha + tell +6), In this case, f(6) is «linear function of & Hence, a straight line is transformed into a straight line. When f is a nonlinear function, a straight line is mapped into a curve of the same degree as the transformation. it is natural to think of approximating the geometry in the same way as ‘we approximated a dependent variable. In other words, the, transformation (G) can be written as dS HiO ‘ (7.10) where xf is the global coordinate of the ith node of the element 9° and 1f are the Lagrange interpolation functions of degree m—1, Equation (7.10) represents the shape (or geometry) of an clement, and the Uf are therefore called shape functions. Equation (7.10) maps a geometric shape from & space into x space, ie, for any given &, (7.10) gives the corresponding x. When the clement is a straight line, (7.10) is exactly the same as (7.6) “The transformation (7.10) is useful in evaluating integrals by the Gauss quadrature. It should be noted that the transformation is not used to change the actual geometry of the element; the transformation is used to write integral expressions involving x in terms of expressions involving &: [free [fees a.) so that the Gauss quadrature can be used to evaluate the integral over [—1, 1]. “The differential element dx in the global coordinate system x is related to the differential element d& in the natural coordinate system & by dx a dé = §, dl i ea where §, is called the Jacobian of the transformation. We have. (Sasi) = 3 =: axtyf)= Dx (7.12) (2 0)” 2 as ¢ Fora linear transformation, i.e., when ‘pf are the linear Lagrange interpota- tion functions (rm =2), we have B=30-8), te=40+8) Fe=X(-) 4 250) = HOS x1) = the 7.13)NUMERICAL INTEGRATION AND COMPUTER IMFLENENTATION 251 It can be shown that J=4h, whenever the element is a straight line, irrespective of the degree of interpolation used in the transformation (7.10), 7.1.4. Isoparametric Formulations Recall that a dependent variable uw is approximated in the element @* by expressions of the form 2 4ivie) (1.14) ute JIn general, the degree of approximation used to describe the coordinate transformation (7.10) is not equal to the degree of approximation (7.14) used to represent a dependent variable, ie., pies. In other words, two independent elements can be used in the finite element analysis: one for the approximation of the geometry x and the other for the interpolation of the dependent variable «. Depending on the relationship between the degree of approximation used for the coordinate transformation and that used for the dependent variable, the finite element formulations are classified into three categories: 1. Subparametric formulations: m (7.15) In subparametric formulations, the geometry is represented by lower-order elements than those used to approximate the dependent variables. An example of this category is provided by the Euler-Bernoulli beam element, where the Hermite cubic element is used to approximate the transverse deflection, while the geometry is approximated, when straight beams are anglyzed, with linear interpolation functions. In isoparametric formulations (the most common in practice), the same element is used to approximate the geometry as well as the dependent unknowns: y:(x) = ¥,(g). In the superparametric formulations, the geometry is represented with higher-order clements than those used to approximate the dependent variables. This formulation is seldom used in practice. 7.1.5. Numerical Integration ‘The evaluation of integrals of the form [rea 16) by exact means is either dificult or impossible owing to the complicated form of the integrand F. Numerical integration is also required when the integrand is to be evaluated inexactly (as in the ‘Timoshenko beam element) and when252. nynf SLEMENT ANALYSIS OF ONE-DIMENSIONAL, PROBLEMS the integrand depends on a quantity that is known only at diserete points (e.g., in nonlinear problems). ‘The basic idea behind all numerical integration techniques is to find function P(x) that is both a suitable approximation of F(x) ‘and simple to integrate. The interpolating polynomials of degree », denoted by Fy, which interpolate the integrand at 2 + 1 points of the interval (a, b], often produce a suitable approximation and possess the desired properly of simple in- tegrability. An illustration of the approximation of the function F(x) by the polynomial P,(x) that exactly matches the function F(x) at the indicated base points is given in Fig. 7.4(a). The exact value of (7.16) is given by the area under the solid curve, while the approximate value [reoae . @ Fis) ao) FIGURE 74 Numerical integration by the Newton-Cotes quadrature: (¢) 3p- proximation of a function by Py); a) (6). the trapezoidal rules» (e) © Simpson's re.NUMERICAL INTEGRATION AND COMPUTER IMFLEMENTATION 253 is given by the aréa under the dashed curve. It should be noted that the difference (i.e., the error in the approximation) £ = F(x) ~ P,(x) is not always of the same sign, and therefore the overall integration error may be sina (because positive errors in one part cancel negative errors in other parts), even when P, is not a good approximation of F. ‘The commonly used integration methods can be classified into two basic groups: 1. The Newton-Cotes formulae that employ values of the function at equally spaced base points 2, The Gauss quadrature formula that employs unequally spaced base points. ‘These are described here. ‘THE NEWTON-COTES QUADRATURE. For r equally spaced base points, the Newton-Cotes integration formula is given by f F(x) dv =(6—a) s Foamy iy) where w; are the weighting coefficients and x, are the base points, which are equally spaced. For r= 1, (7.17) gives the rectangle formula, For r=2, it gives the familiar trapezoidal rule, in which the required area under the solid curve in Fig. 7.4(b) is approximated by the area uncer the dotted straight line [i.e., F(x) is approximated by Pi(x)]: [ F(x) de =4A[F@) + Fe), -E= 08) (7.18) where E denotes the error in the approximation and h two base points. The notation (1), read as “order of h the order of the error in terms of the spacing fk. For ‘Simpson's one-third rule (see Fig, 7.4c): the spacing between is used to indicate , (7-17) gives the [O° re ac= snr +4Fea)+ Fe), E=0(8) (7.19) ‘The weighting coefficients for r= 1,2, .., 7 are given in Table 7.1. Note that DABw,=1. The base point location for r=1 is x:=a+4(b-a)= 4a+). For r> 1, the base point locations are sey Gat (r— 1) Arab =a, a=atAx, where Ax = (b~ a)M(r~1)- ‘A comment isin order on the use of the Newton-Cotes integration formula (7.17). For &n r (i.e., when there is an even number of intervals or ‘an odd number of base points), the formula is exact when F(x) is @ polynomial of degree r+1 or less; for odd r, the formula is exact for a polynomial of degree r or less. Conversely, a p"-order polynomial is integrated exactly by choosing r=p +1 base points.254 NITE ELEMENT ANALYSIS OF ONE-DIMENSIONAL FROBLENS TABLE 74 ‘Weighting coefficients for the Newton—Cotes formula (7.17) ee 14 204 4 a4 4 a 48 i i 4 5s & g 8 2 % 6 Be & a = Bs & 7 th as & B & 8 ‘THE GAUSS-LEGENDRE QUADRATURE, In the Newton-Cates quadrature, the base point locations have been specified. If the x, are not specified then there will be 2r +2 undetermined parameters, the weights jw, and base points Xp which define a polynomial of degree 2r+1. ‘The Gauss-Legendre quadrature is based on this idea. The base points x; and the weights w; arc chosen so that the sum of the r-+1 appropriately weighted values of the function yields the integral exactly when F(x) is a polynomial of degree 2r + 1 or less. The Gauss-Legendre quadrature formula is given by (see Fig. 7.5) f FG) dr= B®) ae-f rem | 7.20 where w; are the weight factors, & are the base points [roots of the Legendre - ge -057135 (@) 05735 FIGURE 7.5 ¢ 0 t ‘The two-point (a) and three-point (6) Gauss 0774597 0.714597 Legendre quadrature,NUMERICAL IVTEGRATION AND COMPUTER IMPLEMENTATION 255 apie 72 Weights and Gauss points for the Gauss Legend quadrature [ Faas Seem ana & 7 ©.00c0000000—One-point formula +£0,5773502492_Two.point formule 6.0000000000 Three-point formula 0. 8888888889 +£0.7745965592 013855555585 £0.5209810195 Four formula 06521451548 $20.8611369116 0.3478588451 0.0000000000 -Five-point formula 05688888889 +20.5384693101 ‘0.4786286705 0.9061798459 0.2369258850 40.2386191861 —Sixpoint formula O46TOT306 ++0.6612003865 0.3607615730 2O.932465142 ITis2928 polynomial P,,,()], and P is the transformed integrand PEO) = FEN FE), (1.21) ‘The weight factors and Gauss points for the Gauss-Legendre quadrature (7.20) are given, for r=1,..., 6, in Table 7.2. ‘The Gauss-Legendre quadrature is more frequently used than the Newton-Cotes quadrature because it requires fewer base points (hence a saving in computation) to achieve the same accuracy. The error in the approximation is zero if the (2r+2)th derivative of the integrand vanishes. In other words, a polynomial of degree p is integrated exactly by employing 1 =4(p +1) Gauss points. When p +1 is. odd, one should pick the nearest larger integer: be+n) (7.22) In finite element formulations, we encounter integrals whose integrands F are functions of x, ,(x) and derivatives of ¥j(x) with respect to x. For the Gauss-Legendre quadrature, we must transform F(x) dx to P(E) dé to use (7.20). For example, consider the integral Lf odvtdut Ki f a) EY ae 723) Using the chain rule of differentiation, we have dwilx) _ dyilE)dé _ 5. dwilé) de dé de? de oH256 rinire ELEMENT ANALYSIS OF ONE-DIMENSIONAL. PRORLEMS ‘Therefore, the integral can be written, with the help of (7.10), as * Lai avy gag ae) 5 dé $e Fag (7.25) = 3 Fugom (126) where havtaye 8 ati Fae as’ Fag 7 For the isoparametric formulation, we take if = >i. The transformation from x to £ is not required for the Newton-Cotes quadrature. . ‘As noted earlier, the Jacobian matrix will be the same (J. = ih.) when the element is straight, even if the coordinate transformation is quadrs cubic. However, when the element is curved, the Jacobian is a function of § for transformations other than linear. It is possible to determine the number of Gauss points required to evaluate the finite element matrices gn [dui vi Ke a ag [vows as, [via 28) exactly using linear, quadratic, and cubic interpolation functions. For interpolation functions, the integrand of Kj, is constant, requiring only one-point Gauss quadrature. The integrand of the mass matrix Mj is quadratic 2), requiring [r=4(p +)1=3] two-point quadrature. The coefficients ff are evaluated exactly by one-point quadrature. Similarly, for quadratic and cubic elements, we can estimate the number of Gauss points needed to evaluate Kj, Ms, and ff exactly. The results are summarized below. Note that, in estimating the quadrature points, it is assumed that the Jacobian is a constant, which holds true when the element is a straight line, — ‘Number of Gauss quadrature points needed Element type Ky Linear 1 Quadratic 2 Cubic 3 oe If the matrices in (7.28) have variable coefficients or the elements are curved {and hence J, = f.(8)], the degree of the variation of the integrands changes and the number of Gauss points needed to evaluate the integral cexactly changes. If the elements are straight, and the coefficients a= a(x) and c= (x) together with f =/(x) are no more than lincar in x, then the number[NUMERICAL U’TEORATION AND COMPUTER IMELEMENTATION 257 ‘of Gauss points for evaluating the coefficients ava a ee [eyeye Kye a a, ag" cvinj de (7.29) remain the same as listed in the above table. However, the evaluation of f requires one point more than before. Conversely, the two-point quadrature for linear elements, three-point quadrature for quadratic elements, and four-point quadrature for cubic clements would exactly evaluate Kj with a quadratic variation of a(x), Mj with linear variation of c(x), and f with quadratic variation of f(x) The use of Gauss quadrature on (7.28) yields the following values (exact up to the fifth decimal place) when the clement is straight and the isoparametric formulation is used: Quadratic (three-point formula) 2.33333 -2.66667 0.33333 [k}=;, | ~2-66667 5.38333 -2,66667 *L 0.33333 -2,66667 2.33333, 1.33333 0.66667 —0.33333 0.66667 5.33333 0.66667 (7.30) 0.33333 0.65667 1.333433. 0.166667 wf -» tia 0.166667. h IM)=55 Cubie (four-point formula) 3.700 ~4.725 1.350 0.325 4.725 10.800 7.425 1.350 1.350 -7.425 10800 —4.725 0.525 1.350 ~4.725 3.700. 0.761905 0.589286 0.214286 0.113095 0.589285 3.857143 -0.482143 ~0.214286 0.214285 —0.482143 3.857143 0.589286 0.113095 -0.214286 0.589286 0.761905. 0.125 0375 0.375 0.125 (7.31) {Fy = he In Section 7.2, we study the computer implementation of the steps involved in the finite element analysis of one-dimensional problems. AS a part258 faNITe ELENENT ANALYSIS OF ONEDIMENSIONAL PROBLEMS ‘of the element calculations there, the computer implementation of the numerical integration ideas presented in this section will be studied. A model finite clement program (FEMLDV2) for the solution of one-dimensional problems is also described, and its application is demonstrated via several examples. Appendix 1 contains a source listing of program FEMIDV2. 7.2. COMPUTER IMPLEMENTATION 7.2.1 Introductory Comments Chapters 3-6 were devoted to the finite element formulations of two classes of boundary value, initial value, or eigenvalue problems in one dimension: 1. Second-order differential equations (¢.g., heat transfer, fluid mechanics, ED elasticity, bars, and the Timoshenko beam theory); 2. Fourth-order differential equations governing the Euler-Bemnoulli beam theory. ‘The frame element, obtained by superposing the bar element and the beam element, was discussed in Chapter 4 By now, it should be clear to the reader that the steps involved in the finite element analysis of a general class of problems (¢.g., single second-order, single fourth-order, and a pair of second-order equations) are systematic and an be implemented on a digital computer. Indeed, the success of the finite ‘element method is largely due to the ease with which the analysis of a class of problems, without regard to a specific problem, can be implemented on a igitel computer. For different geometries, boundary and initial conditions, ‘and problem data, « specific problem from the general class can be solved by imply supplying the required input data to the program. For example, if we develop a general computer program to solve equations’ of the form au, Pu 8 (au) F (, uw aos gla) tae (hae) trey ‘then all physical problems described by (3.1) and (4.1) and their time- dependent versions can be solved for any compatible boundary and initial conditions. ‘The purpose of this section is to discuss the basic steps involved in the development of @ computer program for second- and fourth-order onc- dimensional differential equations studied in the preceding chapters. The ideas presented here are used in the development of the model program FEMIDV2 (a revised version of FEMID from the first edition of this book), and they are meant to be illustrative of the steps used in a typical finite element program development. One can make use of the ideas presented here to develop a program of one’s own.NUMERICAL INTEGRATION AND COMPUTER IMPLEMENTATION 259) PREPROCESSOR + Read the input data + Echo the input dats + Waite the input data and plot the mesh PROCESSOR + Compute element coefficient matcices ‘and column vectors + Assemble element equations + Impose boundary conditions + Solve equations or ‘nd eigenvalues POSTPROCESSOR + Compute solution at points other than the nodes + Compute the gradient of the solution + Printplot the resus |} FIGURE 7.6 F The three ain functional units ofa finite element program. 7.2.2 General Outline A finite element program consists of three basic parts (see Fig. 7.6): 1, Preprocessor 2, Processor 3, Postprocessor In the preprocessor part of the program, the input data of the problem are read in and/or generated, This includes the geometry (e.g., length of the domain and boundary conditions), the data of the problem (¢.g., coefficients in the differential equation), finite element mesh information (e.g., element type, number of elements, element length, coordinates of the nodes, and connec: tivity matrix), and indicators for various options (¢.g., print, no print, type of field problem analyzed, static analysis, eigenvalue analysis, transient analysis, and degree of interpolation).260 yarn ELENENT ANALYSIS OF ONE-DIMENSIONAL, FRODLESS In the processor part, all steps in the finite element method, except for postprocessing, discussed in the preceding chapters are performed, These include the following: 1. Generation of the element matrices using numerical integration. 2, Assembly of element equations. 3, Imposition of the boundary conditions. 4, Solution of the algebraic equations for the nodal values of the primary variables. In the postprocessor part of the program, the solution is computed by interpolation at points other than nodes, secondary variables that are derivable from the solution are computed, and the output data are processed in a desired format for printout and/or plotting. “The preprocessor and postprocessors can be a few Fortran statements to read and print pertinent information, simple subroutines (¢.g., subroutines to generate mesh and compute the gradient of the solution), or complex programs linked to other units via disk and tape files. The processor, where typically large amounts of computing time are spent, can consist of several Subroutines, each having a special purpose (¢.g., @ subroutine for the aleulation of element matrices, a subroutine for the imposition of boundary conditions, and a subroutine for the solution of the equations), The degree of Sophistication and the complexity of @ finite element program depend on the teneral class of problems being programmed, the generality of the data in the quation, and the intended user of the program. Tt is always desirable to describe, through comment statements, all variables used ia the computer program. A flow chart of the computer program FEMIDV2 is presented in 7-1. In the following sections, a discussion of the basic components of a fal finite element program is presented, and then the basic ideas are lystrated via FORTRAN statements (see Appendix 1), 7.2.3 Preprocessor ‘The preprocessor wnit consists of reading input data and generating finite element mesh, and printing the data and mesh information. The input data to ‘i finite element progcam consist of element type IELEM (ie., Lagrange ‘lement or Hermite element), number of clements used (NEM), specified ‘boundary conditions on primary and secondary variables (number of boundary conditions, global node number and degree of freedom, and specified values of the degrees of freedom), the global coordinates of nodes .%,, and element properties (¢.8,, coefficients a, Bey Ce for ete). Ifa uniform mesh is used, the Tength of the domain should be read in, and global coordinates of the nodes can be generated in the program. “The preprocessor portion that deals with the generation of finite element mesh (when not supplied by the user) can be separated into a subroutineNUMERICAL INTEGRATION AND COMPUTER wLeMeNrATiON 261 seesa aeaxt0s “anans couaisa saxaa09 seanmoxgns oy) te Buore ZANE wesoud samndwuen ap jo EHD HOLE cL man vane262 rine: ELEMENT ANALYSIS OF ONEDIRIENSIONAL PROBLEMS (MESHID), depending on the convenience and complexity of the prog'am. Mesh generation includes computation of the global coordinates X; and the Connectivity array (B]=[NOD]. Recall that the connectivity matrix deseribes the relationship between element nodes to global nodes: NOD(, J) = global node number corresponding to the Jth node of element J “This array is used in the assembly procedure as well as to transfer information from element to the global system and vice versa. For example, to extract the ‘element nodal coordinates x? of the element Q" from the global coordinates X;, we can use the array NOD: xfeX, 1=NOD(, i) ‘ ‘The arrays (ELX) and (GLX) ate used in FEMIDV2 for {x7} and (X7), respectively. “ 7.2.4 Calculation of Element Matrices (Processor) ‘The most significant part of a processor is where we generate’ element matrices, ‘The element matrices are computed in various subroutines (COEFNT and TRSFRM), depending on the type of problem being solved. ‘These subroutines typically involve numerical evaluations of the element matrices [K*] and [M'] (program variables ELK and ELM) and the element vector (f*} (program variable ELF) for various field problems. The Gauss {quadrature described in Section 7.1.5 is used to evaluate element matrices and Tectors, and the arrays are assembled as soon as they are computed. Thus, a Joop on the number of elements in the mesh (NEM) is used to compute ‘lement matrices and assemble them (subroutine ASSMBL). It is here that the connectivity array NOD plays a crucial role. By putting element matrices into flobal locations one at a time, we avoid the computation of all element matrices at once. Element matrices for different model equations (MODEL) and type of problem (NTYPE) are generated by assigning values 2s follows: 4. MODEL=1, NIYPE=0: all field problems described by the model equation (3.1), including radially symmetric heat-transfer-type problems, 2. MODEL =1, NIYPE= 1: radially symmetric elastic disk problems (sec Problem 3.33). 3, MODEL=2, NTYPE=0 (RIE) or 2 (CIE): the Timoshenko theory of beams. 4, MODEL=2, NTYPE=1 (RIE) or 3 (CIE): the Timoshenko theory for bending of circular plates. 5, MODEL=3, NTYPE-=0: the Euler-Bernoulli theory of beams. 6. MODEL=3, NTYPE=1: the Euler-Bernoulli theory for bending of circular plates.[NUMERICAL TEGRATION AND COMPUTER INFLENENTATION 263 0: the two-node truss element. |, NTYPE = 1: the two-node Euler-Bernoulli frame eloment, |, NTYPE =2: the two-node, Timoshenko frame element. 9. MODEL =. ‘The time-dependent option is exercised through variable ITEM: ITEM= ITEM ITEM static analysis first-order time derivative ({.e., parabolic) problems second-order time derivative (j.e., hyperbolic) problems ‘The element matrices are evaluated using the Gauss quadrature, except for MODEL = 4, where the explicit forms of element coefficients are programmed in the interest of computational efficiency. The element shape functions SF and their derivatives GDSF are ‘evaluated at the Gauss points in subroutine SHPID. The Gaussian weights and points associated with two-, three-, four-, and five-point integration are stored in arrays GAUSWT and GAUSPT, ‘respectively. The nth column of GAUSWT, for example, contains the weights corresponding to the n-point Gauss quadrature rule: GAUSPT(, n) th Gauss point corresponding to the n-point Gauss rule The variable NGP is used to denote the number of Gauss points, which is selected to achieve good accuracy. As noted earlier, the linear, quadratic, and cubic interpolation functions require two, three, and four quadrature points, respectively, to evaluate the element coefficients exactly. ‘Thus, if IELEM is the element type, i for linear TELEM=4 2 for quadratic (Lagrange elements) 3. for cubic then NGP=IELEM +1 would evaluate Kj, Mf, and 7} [see (7.29)] exactly when (x) is linear, and a(x), b(x), and f(x) are quadratic functions. The Hermite cubic element is identified with IELEM = 0, for which case NGP is taken to be 4. The coefficients a(x)= AX, b(x)=BX, and e(x)=CX, together with f(s) =FX in the differential equation (7.32) are assumed to vary with x as follows: AX= AXO+AXL+X (a= ag +a,x) BX = BXO-+BX1#X (b= bot b,x) CX = CX0 + CX1#X (€=co+eix) FX=FXO+PXISX+EX2#X#X (f= ft fix thx’) For radially symmetric elasticity problems, (AX0, AX1) for (BX0, BX1) for circular plates) are used to input Young's modulus E and Poisson's ratio ¥.264 nine ELEMENT ANALYSIS OF ONEDINERSIONAL PROBLEMS ‘The Gauss quadrature formula (7.20) can be implemented on the computer as follows. Consider Ki of the form f [ovo St a + ecowtui| dx (0.33) Let us use the following program variables for the quantities in (7.33): BLK(,D=Ky, SF()= 95. vse =“ AX=a(x), CX=c(x), ELX()=af NPE=n, the number of nodes in the element After transforming x to & ‘ so Day Xa tthe + )} (7.34) the coefficients Ki in (7.33) take the form (see (7.25)] ea f fate LAL ceyvens y= [ [ow 5S cervins] at et) = 2 FED IM (7.356) where Fj denotes the expression in the square brackets in (7.35a), is the Jacobian, and (£,, W,) are the Gauss points and weights. Examination of (7.356) shows that there are three free indices: i j, and J We take the Gauss-point loop on [as the outermost one: Inside this loop, we evaluate Fs at the Gauss point & for cach i and j, multiply with the Jacobian $=4h, and the weights W,, and sum: ELKG, j) ELK(i, j) + FiEDIW: (7.36) “To accomplish this, we must initialize all arrays that are being evaluated using the Gauss quadrature: po 10. J NPE Bo to 1 = Lire ao EIK(L,d) = 060 ‘This initialization is made outside the do-loop on number of Gauss points. ‘The computation of coefficients Fy in (7.35b) requires evaluation of a, 0, ‘yn and dyy,fdE at the Gauss point &). Hence, inside the loop on J, we call subroutine SHPID to evaluate , dyi/dx = (dy,/d8)/J. We now have allNUMERICAL INTEGRATION AND COMFUTER mMRLEMENrATION 265 quantities needed to compute Kj (1.35b): bo 100 mz = 1 nce" XE = cause (ti, NoP) a1) subroutine suPiD to evaluate the interpolation functions (GP) and their global derivatives (GOSF) at the Caves point XT CARL, suPtD (Xt, NPE, SP ,GDSF,63) Cons = G3 scASSwE (ar er) x= 0.0 Do 20" Yan nee 20 Kak + GRE) ABLK(Z) bo s0.g~2,¥PE BLP(3) = ELP(O) + conewesr (g)4rx Do. $0)x = tae 30 BLxc(E,5)=ELK(T, 5) 4CONSTA (ANSEDEP (1) 4cDBF(I) 4exAOP(2) #8809) In the same way, all other coefficients (e.g., Mj and f1) can be evaluated, Recall that the element properties (i.c., Kj, Mj, and f3) are calculated by calling a suitable subroutine (COEFNT or ‘TRSFRM) for the field problem being analyzed within a loop on number of elements. As soon as the element properties are available for a particular element, they are put into their proper locations (j.e., assembled) with the help of array NOD. The assembly is explained in the next section. 7.2.5 Assembly of Element Equations (Processor) ‘The assembly of element equations should be carried out as soon as they are computed, rather than waiting till element coefficients of all elements are computed. The latter requires storage of the element coefficients of each element. In the former ease, we can perform the assembly in the same loop in which a subroutine is called to calculate element matrices. A feature of the finite element equations that enables us to save storage and computing time is the assembly of element matrices in upper-banded form When element matrices are symmetric, as is the case in most problems of interest in this book, the resulting global (or assembled) matrix is also symmetric, with many zeros away from the main diagonal. Therefore, it sufficient to store only the upper half-band of the assembled matrix. The randwidth of a matrix is defined as follows. Let N, be the number of elements between the diagonal element and the last nonzero element in the ith row, after which all elements in that row are zero; the half-bandwidth is the maximum of (N,-+1)XNDE, where NDF is the number of degrees of freedom per node: = max {(N, + 1) XNDF]266 INITe ELEMENT ANALYSIS OF ONE-DIMEASIONAL FRODLENS where n is the number of rows in the matrix (or equations in the problem). General-purpose equation solvers are available for such banded systems of equations. “The hatf-bandwidth NHBW of the assembled (ie., global) finite element satrix can be determined in the finite element program itself. The local nature af the finite element interpolation functions (i.c., F are defined to be nonzero only over the element $2) is responsible for the banded character of the Gssembled matrix. Recall from earlier discussions that if two global nodes do hhot belong to the same element then the corresponding entries in the global matrix are 2er0s: Ky =0 if global nodes I and J do not correspond to local nodes of the same element ‘This property enables us to determine the half-bandwidth ‘NHBW of the assembled matrix: NHBW= max. {abs [NOD(W, D)~NOD(N, J)]+ 1) XNDF_ (7 31a) where NEM = number of elements in the mesh NPE = number of nodes per element (137b) NDF =number of degrees of freedom per element Clearly, for one-dimensional problems with elements connected in series, the maximum difference between nodes of an element is equal to NPE —1. Hence, NHBW = [(NPE ~ 1) +1] x NDP-= NPE x NDF (7.38) Of course, NHBW is always less than or equal to the total number of primary degrees of freedom, iJc., the number of equations, NEQ, in the finite element mesh of the problem. “The logic for assembling the element matrices Ki into the upper-banded form of the global cocfficients K,y is that the assembly can be skipped whenever J<1 and J>NHBW. The main diagonal, T=J, of the assembled square mattix (je., full storage form), becomes the first column of the dusembled banded inatrix (i.e., banded storage form), as shown in Fig. 7.8. ‘The upper diagonals (parallel to the main diagonal) take the position of respective columns in the banded matrix. ‘Thus, the banded matrix has dimension NEQ X NHBW, where NEO denotes the total number of equations in the problem The element coefficients Ki and f? of a typical clement @" are to be assembled into the global coefficient matrix {K] and source vector (F), respectively. Ifthe ith node of the element is equal to the Ith global node, and the jth node of the element is equal to the Jth global node, we have F,= Fi (forNDE=1) (7.394)[NUMERICAL INTEGRATION AND COMPUTER IMPLEMENTATION 267 ull storage mode Upper-half-banded mode - Hamish = NHBW KK KK Al blanks Sx xx All entries above “<< this diagonal ex Symmetric y Main diagonal NEQ x NEQ NEQ = NHBW FIGURE 7.8 Finite element coefficient matrix storage in upper-hal-banded form, ‘The values of J and J can be obtained with the help of array NOD: 1=NOD(n, i), J=NOD(n, /) (7.396) Recall that itis possible that the same / and J may correspond to & pair of i and j of some other element 2". In that case, KY will be added to existing coefficients K,y during the assembly. For NDF> 1, the logic still holds, with the change Kemesc = Kire-nute-n (P19 = 1,2,...,NDF) (7.40a) where NR=(I-1)XNDF+p, NC=(@-1)xNDF+q (7.405) and J and J are related to i and j by (7.396). A Fortran listing of the subroutine ASSMBL can be found in FEMIDV2 in Appendix 1. 7.2.6 Imposition of Boundary Condi Imposition of boundary conditions on the primary and secondary global degrees of freedom can bé carried out through a subroutine (BONDRY), which remains unchanged for 2-D or 3-D problems. There arc three types of ‘boundary conditions for any problem: ions (Processor) J. Essential boundary conditions, i.e., boundary conditions on pi variables (Dirichlet boundary conditions). wary268 NITE ELEMENT ANALYSIS OF ONE-DIMENSIONAL, PROBLEMS 2, Natural boundary conditions, ie. boundary conditions on secondary variables (Neumann boundary conditions). 3. Mixed boundary conditions (Newton boundary conditions). ‘The procedure for implementing the boundary conditions on the primary variables involves modifying the assembled coefficient matrix (GLK) and right-hand column vector (GLE) by three operations: Step 1 moving the known products to the right-hand column of the matrix equation. Step 2 replacing the columns and rows of GLK corresponding to the known primary variable by zeros, and setting the coefficient on the main diagonal to unity. Step 3 replacing the corresponding component of the right-hand Column by the specified value of the variable, Consider the following N algebraic equations in Tull matrix form: Ku Ko Ky...) (Uy A Ka Koz Kay Ul JR Ky Kn Ky ---|)U(~) 8 where U; and F, are the global primary and secondary degrees of freedom, and K, are the assembled coefficients. Suppose that Us — Gs is specified. Recall that when the primary degree of freedom at a node is known, the correspond- ing secondary degree of freedom is unknown, and vice versa. Set Kss= 1 and F, = Os; further, set Ke = Kis =0 for I= 1,2,..., Nand 1#S. For §=2, the modified equations are Ku 0 Ky Ky... Ku] (0 o 10 0... of fy 0 Ky Ku Up = 0 Ka Ku Y, where R=R-Kal, (=1,3,45,...,.mi#2) ‘Thus, in general, if Us Ks=1, = 0s; R=R-KsOs; Ku = Kis where /=1,2,...,$—1, S+1,..., 2 @#S). This procedure is repeated for every specified primary degree of freedom. It enables us to retain the original order of the matrix, and the specified boundary conditions on the primary degrees of freedom are printed as part of the solution. Of course, the logic should be implemented for a banded system of equations. Os is known, we have[NUMERICAL INTEGRATION AND COMPUTER INFLEMENTATION 269 ‘The specified secondary degrees of freedom (Q,) are implemented directly by adding their specified values to the computed values. Suppose that the point source corresponding to the Rth secondary degree of freedom is specified to be fg. Then Fra = fin + Pe where fi is the contribution due to the distributed source (x); fy is computed as a part of the clement computations and assembled. Mixed-type boundary conditions are of the form ae k(u~i%)=0 (i and k are known constants) (7.41) which contains both the primary variable w and the secondary variable adufdx. Thus a du/dx at the node P is replaced by —Ky(Wtp ~ dp): Qp = —Ky(Up — Op) ing Kyp by adding k, to its existing value, Kp Kp ky ‘This amounts to mo and adding k,U, to Ey, Boh thy, All three types of boundary conditions are implemented in subroutine BONDRY for boundary, initial, and eigenvalue problems. The following are used in subroutine BONDRY (see Appendix 1): NSPV number of specified primary variables NSSV number of specified secondary variables NNBC number of Newton boundary conditions SPV column of the specified values 0, of primary variables VSSV column of the specified values fy of secondary variables (7.4) NBC column of the specified values k, UREF column of the specified values 0, ISPV array of the global node and degree of freedom at the node that is specified (ISPV(L, 1) = global node of the Ith boundary condition, ISPV(I, 2) = degree of freedom specified at the global node, ISPV(I, 1)] imitar definitions are used for ISSV and INBC arrays, 7.2.7 Solution of Equations and Postprocessing Subroutine SOLVE is used to solve a banded system of equations, and the solution is returned in array GLF. The program performs the Gaussian elimination and back-substitution to compute the solution, For a discussion of the Gaussian climination used to solve a set of linear algebraic equations, the270 vere ELEMENT ANALYSIS OF ONEDIMENSIONAL PROBLEMS reader is referred to the book by Carnhan, Luther, and Wilkes (1969). On ost computing systems, a vatiety of equation solvers are available, and one ‘can use any of the programs that suits the needs. See Appendix 1 for a listing of subroutine SOLVER. Postprocessing involves computation of the solution and its gradient at preselected points of the domain. Subroutine PSTPRC is used to evaluate the solution and its derivatives in any clement: vorSumeo (Gi -f6G), o69 for the Lagrange elements and ow § (270) (ne Su( ZH) oma y2,3) 0496) a for the Hermite cubic elements. ‘The nodal values uf of the element Q° are we) = 3 woe, deduced from the global nodal values U; as follows: U, 1=NOD(e, j), when NDF = For NDF> 1, [is given by 1 =[NOD(e, )~ 1] NDF and Ujep= Upp (P= 1,2 +-+y NDE) ‘The values computed using the derivatives of the solution are often inaccurate because the derivatives of the approximate solution become increasingly inaccurate with increasing order of differentiation. For example, the shear force computed in the Euler—Bernoulli beam theory, dj, _ 2d (, C0 vat (ot) => wt (oS 7. 420) -S aS (of) an will be in considerable error compared with the true value of V. The accuracy increases, rather slowly, with mesh refinement and higher-order elements. When accurate values of the secondary variables are desired, it is recom- ‘mended that they be computed from the element equations: 2-3 Kyi fe CA2 om) 0.45) However, this requires recomputation or saving of element coefficients Kj, and fi. 73 APPLICATIONS OF THE COMPUTER PROGRAM FEMIDV2 7.3.1 General Comments ‘The computer program FEM1DV2, which embodies the ideas presented in the previous section, is intended to illustrate the use of the finite element models Ueveloped in Chapters 3-6 to a variety of one-dimensional field problems, some of them not discussed in this book, FEMIDV2 is a modified version ofNUMBRICAL HCTEGRATION AND COMMUTER MMELENEIErATION 27 FEMID from the first edition of this book. It was developed as a learning computational tool for students of the course. In the interest of simplicity and ease of understanding, only the model equations discussed in this book and their immediate extensions are included in the program (see Appendix 1). Table 7.3 contains a summary of the definitions of coefficients of various model problems and their corresponding program variables. The table can be used as a ready reference to select proper values of AX0, AX1, and so on for different problems, 73.2 Ilustrative Examples Here we revisit some of the example problems considered earlier to illustrate the use of FEMIDV2 in their solution. Only certain key observations concerning the input data are made, but complete listings of input files for each problem are given. In the interest of brevity, the output files for most problems are not included. ‘Table 7.4 contains a description of the input variables to program FEMIDV2. ‘Example 7.1 Steady heat transfer in a fin. The problem is governed by £8 cand for 0: HopRind, NEYPE=D: Navepet: MODEL=a,, WIYPE=: prablen of MODE, EQUATION 1 Bitewlar DISK (PLATE STRESS) Gizeutar, DISK (PLANE STWATN) ‘iwoshenke BEAM (RIEA) problem Tmoshanke ELATE (RIE), problem ‘Nnoshonke REAM. (CIE##). problen ‘noshenko PEATE (CEB) probl Biler-Bernoull! BEAM problen Euler-Bernowlli Clroular plete plane TRUSS problem Buise-pernovlil ‘FRAME problen ‘Sincchenke’ (CIB) FRAME’ problen 4. ~ Reduced Integration Elenent (RIE) fa 7 Eineieten® tnterpoletion Elenent (CTE) rem - Indicator for transient enalyste: rrew-0, stoady-ctate solution Toetei, fransient analysis of PARAPOLIC eq Firte2, Transient analyess of HYPERBOLIC oq Yiives, Eigenvalue analysis + pata cara 35 Totem - Type of finite elenent: TebmI-o, Hermite cubic finite elexent. Hitec) Linese zagrange finite element TELeNs2! Guadratie Lagrange Einite element ue ~ Nusber of elenents in the mesh + pats cara 4 cow ~ Indicator for continuity of data for the problem: root, Date (AX,PX,/Ck,FX and mesh) Ls continuoas [eownad) bata ie elenent dependent NpRAT ~ tndieetor for peinting of elenent/global matrices: prvt=0, Not print elenent or global matrices, ‘Soe postprocese. the solution and print ueruret, Print’ Elenent 1 coatficient matrices, only Eat postprocsee tha solution and print | SKIP mcons that you ont the np data, It does ot mean that you eave a lank, tthe res format” sed [SE Mites of el dna card ae read frm the stm ac; Ite values are not found am the sare Hae, te rae eri lok fr them on te nest line(). Each dot card shoud start with ne ne. Note cht the sage es nc inte dataset of Tables 75-7.14 a only forte convenienes ofthe eader—they ae rot ead bythe computes.NUMERICAL INTEGRATION AKD COMPUTER IMPLEMENTATION 275 ‘TABLE 7.4 (Continued) pRt=2, Print Blenent 1 and global matrices but Tio? postpracess. the solution upRin>2, Not print elesent or global matrices end "Hor postprocess the solution SKIP Corde 5 through 15 for TRUSS/FRAME problens (HODRL oud Cards 3 through 15 only if MODEL.NE: in'data ie discontinuous (1e0uT * 9) 4), ana SKIP Cards 5 through 9 bata card 52 Px(1) ~ Array of elenent Lengths, Dx(4) denotes the globat Soordinate of Node 1 of the weahj DE(Z)_(1=2/1EMi) denotes ‘the length of the (I-i)at elesent. Here NENI=WEMGL, and NEMCnunber Se elements in the Hes. cards 6 through 9 detine the coefficients in the nodel equations MI gosteicients are expressed In terns of GLOBAL coordinate x, Sse Table 7.3 for the meaning of the coceficients, especially for deformation ‘of circular plates and Timoshenko elements. pata card 6: Axo - Constant tern of the cooftioient ax Ra > Linear” tern of the conftioient Ax Date cara 7: Bx0- Constant torn of the costficient nx Bul Linear” teem of the coatticient BX Deva cara a: CO - Constant term of the coefficient cx Gx" Linear” term of the coefficient ox SKIP Card 9 for eigenvalue problens (i.e, ITEN=3) pata card 9: FXO ~ Constant term of the coefficient Fx PH = Linear" tera of the coeftlolent Fx PR2 ~ quadratic term of the coefficient. x SKIP Cards 10 through 15 if date is continuous (ICONT.HE.0), carde 20 through 12 are read for each elanent (1-8.) NEN tines). AIL Coctficients are expressed in terns of the GLOBAL coord (GUX) ~ Globai x-coorainates of the FIRST and IAS? nodes of the alenent pata cara ai (00) = Connectivity of the element: HOD (Wt, )-cloval node number cofresponding to the T-th node of Blenent it (2-4/P5) NPE denotes the Husker of nodes Der Elenent {DCAK)~ Constant and Linear terns of the coefficient AX bata card 13: {DCBK}~ Constant and Linear terns of the coefficient Bx276 saxire: ELEMENT ANALYSIS OF ONEDIMENSIONAL PROBLEMS ‘TABLE 7.4 (Continued) + bata card a4 (Dcex)- constant and Linear terms of the coefficient ox * pata card 15: . (DcrX}- Constant, Linear and quadratic terms of FX READ Cards 16 through 21 for TRUSS/FRAME problens (WODEL = 4). SKID Garde 26 through 21 1f MODED.NE.4 * pata cara 16: nn) = Number of nodes in the finite elenent mesh SKIP Cards 17 through 19 for TRUSS problens (NTYPE = 0). cards 1S through 19 ere read for each element, ive, NEM tines + pate cara a7 PR = Poisson's ratio of the naterialy BE _ Young's modulus of the neterial * SE _ Length’ of the olanent 5k = Grogsveotional area of the element. Sr 7 Moment of Inertia of the element €6 7 Cosine of the angie of orientation of the elenent SH _ Sine of the angle of eriantation of the element Angle is nsasured counter-elock-wise fron x axis # PR is not used for the Euler ~Bernoulli elenent + pata cara 18: hr = Intensity of the horizontal distributed force YE = Intenaiey of the tranaversely distributed force Pr 7 Point Load on the elenent XB Distance from nede 1, along the length of the elenent, to the point of load application, PF cur Cosine Of the angie of orientation of the lod PF SHE 7 Sine” of the angie of erisntation of the lond PF Bngle ts measured counter-clock-wise tron the olenent x ais + pata card 19: (op; ~ connactivity of the elexent: Nop(ti,x) = Global node number corresponding te the T-th nods of Biosont N (21, NPE) READ Carde 20 and 21 only for TRUSS problems (1ITYPE = 0), cards 20 -and21 are read for each elenent;i-o, NEM tines, + pata card 20: SE Young's modulus of the material SL > Longth’of the elenent. SX 1 Orogsvectional ares 9€ the elenent G6 | Gonine of the angle of orientation of the elenent SK 7 Sine” of the angla of orientation of the element Angle $2 measured counter-clock-wise from x axis, ke = Intenaity'ef the horizontal distributed force pata cara 21: {10d} ~ connectivity of the etexent Nob(t1) = Global node Nunber corresponding te the Teh node of Elexont M1 (T=1,NPE)NUMERICAL IETEGRATION AND COMPUTER INFLENENTATION 277 ‘TABLE 7.4 (Continued) pata cara 221 sev - Nunber of specified PRIMARY dagress of freedon SKIP cara 23 if ne Gara 23, NSPY tino: prinaty variables is specified (NSPV=0). Repeat + pata cara 23: (r= 1 to NSPY) Tepv(z,1) = Node number at vhich the PY is specifies Tspv(1/2} = specttiee local prinary DoF at the node USEv(r) ~ specified value of the primary variable (PV) (Hill not Yend for olgenvalue problems) SKIP cad 24 for eigenvalue problens (1. J when ITEM) + pata card 24: Nssv - Mumber of specified (nonzero) SECONDARY variabley SKIP card 25 {f no secondary variable 1s specified (WSSv=0). Repeat Gard 25 NeSY tines + pata card 25: (I~ 2 to Nssv) esv(1,1) ~ Mode number at which the sv is specifies Essv(z/2) ~ Specified local’ secondary DOF at’ the node Ussv(z}, = Specifled value of the secondary variable (FV) + pata card 26: nee = Humber of the Newton (nixed) boundary conditions SKIP Card 27 if no mined boundary condition is spect tied (anisc=0) he ained boundary condition is Assumed to he of eho form: SV#VIEC Nleviunpey = 0. Repeat cara25 nse times + pata card 27: (1 = 1 to msc) Bise(z,1) ~ Node number at which the mixed 8.c. ts specitied [ipe(1/2) = Local bor of the PV and SV at, the aode wnne(z) = Velue af the coefficient of the PV in the B.C, UREF(Z} ~ Reference value of the PY SKIP Card 28 if ITEN-O (read only for time-dependent and eigen Syalue problens mecdependent and ¢ + pata card 201 ‘eto = Constant part.ot.Cr = cro + crix GM 7 Linear pare or ci = cro + c7iax SKIP renatning cards if eteady-state or oigenvalue analysis is to Bo perforned {riNe0 or TPsMe3) 1 pata cara 29) DP ~ Sime sncrenent (unifora) Zien = parsneter in the ine epproximation schene BETA _ Paraneter in the tine approximation schene » pata card 20: INcOMD- Tndioator for inft{a1 conditions: Incotb=9, Honodeneous (zero) initial conditions Ancol», Nonhowogeneous initial conditions nrrue - Runber of tine stepe for vhich solution 1s sought INIVL ~ Tine step intervals at which solution is to printed278 rire ELEMENT ANALYSIS OF ONEDIMENSIONAL PROBLEMS TABLE 7.4 (Continued) SKIP Cards 91 and 22 if initial conditions are zero (IHCOND=0) __ + pata cara 31: {000} - Array of initial valuce of the primary variables SKIP card 22 for parabolic equations (xrEm=) + pata card 32: {ou1) = Array of initial values of the first time-derive- fives of the primary variables For Set 2 boundary conditions, we have one essential boundary condi cone mixed boundary condition: : u=0, ot+bu-u)=0 Hence, NSPV=1, NSSV=0, and NNBC=1, and the values in the mixed boundary conditions are input as VNBC(1) = A/k=1.28, UREF(I)=U.=0.0, INBC(L,1)=5, INBC(1,2) ‘The output file for Set 1 boundary conditions is presented in Table 7.5(a), ‘The input files are echoed in the output, The input file for Set 2 boundary conditions is given in Table 7.5(6). Example 7.2 (Example 3.2) Radially symmettic heat transfer in a solid cylinder, The ‘governing equation of the problem is given by [see (3.78)] A (ote) ante for 00, agrange) « fio. of eye of frecdon per node, NDF. No. of alonsnts in the mesh, NES. No: of fotai DOF inthe medal, NB. No. of specitied prinary DOP, SPV, Tol of specified secondary DOF, NESV. Nel of specified Newton B. Cy: MmIBC. Boundary information on prinary variables! 21 o.20900R+03 Global coordinates of the nedes, (6x) cosfticiente of the aigferential aquation: Axo = 0.20008402 axa = Gxo = olaseoeros xi = ocon00Rse0 FRO = 010000800 Fxi= D10000E+00 Fx2 = 0,00008100 SOLUTION (values of pve) at the HODES: D.L0000Et02 0.25198EFO2 0,12504z+02 0485608401 0,303508+02,280 rete ELEMENT ANALYSIS OF ONE-DIMENSIONAL. PROBLEMS 5 (Continued) x P. Varieble 5, Variable 0.00000%+00 9.290908+02 ~0.103758104 Gizeiasee2 QloressEto2 -0.10375Es04 Olisezsz-ot olezveseso2 -0-1037sE000 Siditses-ot ol ysekerso2 -0.10379E008 GliizaoRso1 Oleru7onto2 -0:107n104 bissogsEeo1 Oledyaio2 ~0:10375R40¢ Didestan-o1 O.e1360B+02 -0.103798404 Olsieaap-aL 0.43263z+02 -0.103758+04 Dlezsoos-o8 G35iseE+02 ~0.10375Er04 Diezsooe-0i O1a5388e+02 01362450103 Oc7ost3e-01 Olgaaacueo2 01362058103 Giyeiase-01 Ol29io4zt02 -0.36285E03, Glassaa-01 Olaeeeieto2 -0-36249n003 : Sisazsoe-ol olzses1Ero2 ~0.362455+03 Ditoissesa0 0;20999er02 ~0.362(5E+03 Oliogasesoa 9,ieisosro2 -0.36245E+03 SlinzioEsoe 0. is3sesro2 ~0.362458103 . sTas00Es00 9,12s04Ero2 -0.362450109 Oi1as0ont00 0, 12308Era2 -0.322978t09 Diiszeietes 0:1154aB+02 -0.322378+09 Diidoszetea 0/10s92B+02 01122378409 Bilsniaesoo lseaezeven -0.122373~03 biiseasesoo olaseoazven 01122378109 Diieaossroo 9.77241Bt01 -0112237E109 Olitieseroe 9Le7ea1Etor ~0.1237E602 Siivsesnteo 9:se120E+01 0112237803 Siievsoztea al4esgoe+o1 -0.i2237E+03 DiteysoEtoo OldeseoEven -0.29136E102 Biissaipsoo Ol4e204Eto1 ~0.291368002 SidosizEre9 OLésooeEtor -0.20136Es02 Oldtestere9 oLet7aieso1 -0-29136E+02 Qibierseron ovazdssevo -0-29136E+02 DizzessRrog 0:371798+01 0291368402 bragsassusoa 0.34903EF01 ~0.291368402 Bigigiaieoo O.3ae26E+o1 ~0.291368+02 Olas00oHe00 O.20350Ero1 -0.29136Er02 eee eee {@) Set 2 boundary conditions (input data only) bxanple 7.11 steady heat transfer in a fin (sat 2 boundary conditions) ran) Wobel,, neve, TERM ia TELEA, HEH aa Eoow, MERIT oto 0.0625 0.0625 0.0625 0.0625 DK) 1a oL0 Reo, b10 o.0 Bx, BXL assid 0:0 xo, Cx B10 010 0.0 PRO, Fx Fx2 1 nee 1 1 100.0 Tepv(a,s), vsPvC2) 5 nssv i mec 5 1 1.28 0.0 INBG(4,3), YNBC(2), UREFCA)[NUMERICAL INTEGRATION AND COMPUTER INFLENENTATION 281 TABLE 76 Input file for the problem in Example 7.2 Reauple 7.2 Radially symmetric heat transfer in a solid cylinder Too : HobEL, NEYPE, ITEH 22 TELEX, Ee 2:0 0.005 0.005 bx bro S5r66a7 XO, AXE 0:0 12.663728 0.0 FXO, PRL, FX2 5 1 100.0 ESPY(L,3), VSPV(2) 6 issv ° tmiae ‘The input file of the problem is presented in Table 7.6 (ef. Table 3.3). Note that the finite element solution obtained with two quadratic elements is more accurate (essentially the same as the exect) than the solution obtained with four linear elements ‘The fluid mechanics problem in Example 3.3 is very simple, and the reader should be able to generate the input file very easily to solve it with FEMIDV2. Fxample 7.3 Deformation of a rotating disks For this case, we have MODEL=1, NIYPE= 1, snd TEM =0. For x mesh of two quadratic elements (ie., TEL=2 and NEM = 2), we use ICONT = | and {DX) = (0.0,0.25R, 0.25R, 0.258, 0.25R) where Ris the radius of the ditk. Fora uniform and homogeneous disk of thickness #7 and made of isotropic material, we take the moduli Ey~ By. Since we ate seeking results in nondimensional form, we take R10, E=10, ‘Thus, we have AXO(= 2) = 1.0, AXI BXO(=¥4)=03, BXI(=H)=10 CX0=00, Cx1=0.0 ‘The body force is f,= per = fy + fir +f0. Hones, t FXO=00, FXI=1.0, FX2=00 2) =1.0 ‘The boundary conditions are w(O)=0 (by symmetry) and ro,=0 at r= R (Gteessfree condition). Since the secondary variable is homogencous, there is no need to impose it—only the boundary condition on the primary variable isto be imposed. We have NSPV=1, NSSV=0, NNBC ISPVG,1)=1, ISPV(1,2)=1, _VSPV(1)=0.0 ‘The input file for this problem is presented in Table 7.7.282. FINITE ELEMENT ANALYSIS OF ONEDAENSIONAL PHOBLENS TABLE, 77 Input file for the problem in Example 7.3 peauple 7.31 Deformation of a rotating disk aa o Morn, wivex, 27= 22 TELEX, NEM oto 0.50 0.80 Dey ie sa Dx, a ars ale 3x0, Bx ero ato x0, cx eo 18 oo Fro FxL, #x2 1 sey L200 Eepy(2,3), vsPV(1) ° kes ° ase Example 7.4 (Example 4.2) Clamped and spring-supported beam ‘In Rig. 4.7. We solve the problem using the Euler-Bernoulli beam element (MODEL =3, NTYPE =0, TELEM=0) snd the Timoshenko beam element (MODEL =2, NIYPE=0 or 2, IELEM = 1,2, or 3). the loading is discontinuous, we set ICONT |. A minimum of two elements are required to model the problem (.e., NEM = 2). If we take El = 1,0E6 (i.¢., 10° ff 1b) then EpjS_ EIS CAR yy he av For L/H =10, we have H = 1.0 because L = 10 ft. For the choice v=0.25, we have GAK = 41H =4 x 10% ‘Thus, we use AX0=0.0, AXI=0.0, BXO (=£1)=1.086 BXi=0.0, CX0=0.0, Cxi=00 for the Euler-Besnoulli beam and AXO (=GAK)=4.086, AXI=0.0, BXO BX1=0.0, CX0=00, CXt for the Timoshenko beam, ‘The distributed transverse load is zero in element 1, and it is El) =1.086 0 fe= -" Poo in element 2. Hence, FxX0=0.0, FI FX0= 66.666, FX 0.0, FX2=0.0 inclement 1 ~16.666, FX2=0.0 inelement 2 The global coordinates of nodes and the connectivity matrix entries for each ‘element are obvious from the geometry. For the Euler—Bemoulli beam element the number of nodes is always equal to 2 (NPE =2), whereas for the Timoshenko beam clement, the number of nodes depend on the degree of intespotation (or element type) selected: NPE =IELEM + 1.NUMERICAL PETEGRATION AND COMPUTER IMFLEMENTATION 283 ‘The boundary Conditions for this problem are i woo, (2) =a wean) = ‘Therefore, we have (NSSV = 0) SPV=2, PVCAD=1, ISPVUA2)=1, YSPVO)=00 ISPV@,1)=1, ISPV@,2)=2, VSPV@)=00 NNBC=1, INBC(Q,1)=3, INBC(,2)=1 VNBC() .0E6 (for k/ET=1) and 0.0 (for UREF(1) =0.0 Tables.7.8 and 7.9 give the input fles for the Euler-Bernoulli and Timoshenko elements. Note that the Eulet-Bernoulli element is a Hermite cubic clement, whereas the Timoshenko element is only a Lagrange linear element. 10) Example 7.5 (Example 4.4) Analysis of a plane frame, Here we consider the two-member frame structure shown in Fig. 4.13(a). We shall analyze it using the Euler-Bernoulli frame element (MODEL = 4, NTYPE = 1) and the Timoshenko frame element (MODEL =4, NTYPE=2). The former gives an exact solution for all frame structures with constant cross-section members. The Timoshenko frame element, on the other hand, does not yield accurate results unless several elements per member of the structure are used. ‘The input files of the problem are presented in Tables 7.10 and 7.11. The cight-clement mesh of Timoshenko clements gives results comparable to the two clement mesh with the Buler-Bernoulli frame elements (results are not included here). The next example deals with the use of FEMIDV?2 for time-dependent problems (j.e., ITEM= 1 or 2). TABLE 78 Input file for the problem in Example 7.4 Example 7.4: Clanped and spring-supported bean (E-n elenant; k=1.084) 02 ELEM, wen” bt cone, Newer vo 4.0 uxt} 2 noo (3) % xo, AKL % Bx, Bx x0, xt 0.0 FXO, PX1, Px Guxin Nop (2,3) 2x0, AL Data for Exo, Bx Blement 2 16666) -16.666667 0.0 FXO, PAL, F2. nee TePY(2,2), vseveay spy(2,3), vsPv(2} sev NWgC (with transv. spring} 1 104 0.0284 FINITE PLEMENT ANALYSIS OF ONE-DIMENSIONAL FROBLENS TABLE 7.9 Input file for the problem in Example 7.4 ———————e—e——e—e ee Exeuplo 7.4; Clamped ane spring-supported bean (TIM element; k=1-084) rar) MODEL, NIYEE, T2EW ia TeeM, NEM ot Teowr, nprer 4.0 cox) 2 noDti,3) 20 AKO, i. pata for ° (ay = 10) Bxo, Bx Blenent 2 ° eto) ext 2 0.0 xo} ex x2 ° axa) Nop (2,3) 0 2X0, XL vata tor, ° xo) Bat Blonent. 2 ° xo, Gxt er668507 -16.666667 0.0 Propane | . Tspv(a,a), vsey(3) Ispv(2,o), ¥SPY(2) nesv INBE (with teane. spring) bork 0.0 ERBE(a 3), WHBC(1}, UREP(2) Example 7.6 (Example 6.4) Transient heat conduction in a plane wall. Consider the transient heat conduction problem of Example 6.4 [see (6.56)]: 7.490) (Set 1) (7.496) TABLE 7.10 Input file for the problem in Example 7.5 ———=—_—rvv Examplo 7,5: Analysis of @ plane frana (BB elenent) ver 6 WobeL, NIVPE, TTEH 02 ITeLeM, MEM ba Teown, NRT 3.3. 1,086 144.0 10.0 10.0 0.0 1.0 FRy SE, Sb, Shy ST, C8, Sit blo. -Oleasee6 0.0" G.0 0.0 0.0 He) VE, Pe, XB, cet, sit ve n0b(2,3) 0.2) 1.0K 199.0 10.0 10.0 0.8 0.6 9:0 Ole e410 010 OL6 0.8 sane az above for Elenent 2 2 4 é nsPy i 2 bo 201 oD roPv, vseV 302 blo 2 3 oe 202-20 qssv, vssv é mse[NUMERICAL INTEGRATION AND COMPUTER IMPLEMENTATION 285 TABLE 7.11 Input file for the problem in Example 7.5 Bxample 7.5: Analysio of @ plene frane (WIM element) v2 oO MODEL, MEYPE, TPE a a 0.3 1.0B6 0 10,0 20.0 0.8 0.6 ae) 1 2 oo ~ 23 oo a w0,9-0, sx)) -0 aa 649 w= In Example 6.4, only Set 1 boundary conditions were considered. We have MODEL = 1, NTYPE = 0, ITEM= 1, ICONT = 1, NSPV = 1, NSSV = 0, and NNBC=0 for Set 1 and NNBC-= 1 for Set 2. ‘The coefficients of the differentiat equations are ( 0, ¢= 0.0, ¢ = 1.0 and f =0) AX1=00, BX0=0.0, BX1=00 CX1=00, FXO=00, FXI=0.0 CT0=1.0, CTI=0.0 ‘The boundary and initial conditions (since INCOND = 1) are input as ISPVG,1)=1, ISPV(1,2)=1, _VsPV(1)=0.0 GUO) = (0.0,1.0, 1.0, ...} From the discussions of Example 6.4, we use At=0.01 (DT =0.01) and print the solution for every time step (Je., INTVI.= 1). ‘The input file of the problem for Set 1 boundary conditions is presonted in Table 702. n. For this problem, Example 7.7 (Example 6,2) Natural vibrations of a cantilever be: the boundary conditions are given by an w=0, (28) =0 forte Er Bamotbean ee w=, Y)=0 forte Tinatenks beam tery ‘The input data for all variables is the same as inthe static analysis, In addition, we must input cp and ¢q. For the Timoshenko beam theory [see (6.13)], ¢» denotes the inertia286 finite ELEMENT ANALYSIS OF ONEDIMENSIONAL, PROBLENS TABLE 7.12 Input file for the problem in Example 7.6 Beanple 7.6: "raneient Reat conduction in a plane wall aces fiones, WivPR, CTEM i ELEM, tie i cow, Nenn 508 oxy 2 2X0, Ax 28 Bx0, BL 0 exo, Cx 2 0.0 Fro, Pe 7x2 Neo 20 TSPV(1,3), YSPC) tssv Ne 0, om Bey ‘aura, sera ENonb, iertMe, HVE \ vot) 2 PA, and ¢, denotes the rotatory inertia pf. The eigenvalue solver used in FEMIDV2 requires the matrix [B] in [A}(x} = 4{B]{x} to be positive-definite. Hence, eo and én, should be nonzero, otherwise, the mass matrix coeiicients associated with W will be ‘The input files for the natural vibrations of the cantilever beam by the two types of elements are given in Tables 7.13 and 7.14. ‘The reader can investigate the convergence characteristics of the elements in improving the accuracy of the fundamen- tal frequency with the use of FEMIDV2. 7.4 SUMMARY In this chapter three main items have been discussed: numerical integration of finite clement coefficient matrices and vectors, logical units of a typical finite element program and their contents, and applications of the finite element program FEMIDV2. The numerical evaluation of the coctficients is required because of (a) variable coefficients of the differential equations modeled and (6) special evaluation of the coefficients, as was required for the Timoshenko TABLE 7.3, Input file for the problem in Example 7.7 Example 7.7: Natural vibrations of a cantilever beam (8-B; with RI) ero} Nobel, NOYPE, rTEN af ELEN, 101 ta cows, PRN? oro 0.25 0.25 0.28 0.25, xa bre are XO, Axi ie ole BxO, BX ole 810 cxo, ext 2 nord aa TSPV(1,3) 12 Tsev(2,3) ° fe $200.0 3-0 cro, ert (LyH=20)NUMERICAL ISTEGRATION AND COMPUTER IMELEMENTATION 287 TABLE 7.14 Input file for the problem in Example 7.7 7: Watural Vibrations of a cantilever beam (TIM 61 ° ‘MODEL, MEYPE, IN 50.25 0.25 0.25 aay 2X0, AXE Bro, Bx xo, CHL ay 1Pv(2,3) 12 ispy(2)3) ° reine 12.082 1.0 ere, ort (eyst-20) beam element with equal interpolation. The Newton-Cotes and Gauss Legendre integration rules have been discussed. The integration rules require the transformation of the integral expressions from the global coordinate system to a local coordinate system. This transformation requires interpolation of the global coordinate x. Depending on the relative degrees of interpolation of the geometry and the dependent variables, the formulations are classified as subparamettic, isoparametric, and superparametric. ‘The three logical units—preprocessor, processor, and_postprocessor— have been discussed. The contents of processor, where most finite clement calculations are carried out, have been considered in detail. Fortran statements for numerical evaluation of integral expressions, assembly of element coefficients, and imposition of boundry conditions have been discussed. ‘A description of the finite element computer program FEMIDY2 has been presented, and its application to problems of heat transfer and solid mechanics has been discussed. PROBLEMS Section 7.1 In Problems 7.1-7.5, compute the matrix coefficients using (@) the Newton-Cotes integration formula and (b) the Gauss-Legendre quadrature. Use the appropriate ‘umber of integration points, and verify the results with those obtained by the exact, integration, TA. [Porn ean, a [orovpae ‘where y, are the linear (Lagrange) interpolation functions. Ansiver: Kyy= It UN4 2(t4 +89), Ga ANCL 4 4H), 7.2, Repeat Problem 7.1 for the quadratic interpolation functions288 vive ELEMENT ANALYSIS OF ONE DIMENSIONAL PROBLEMS 73. aie Kun £ (Gay, where ate the Hermite cubic interpolation fan “Answer: 7=2: Kyo 12/0? (exe), Gy, = 039848 114, Repeat Problem 7. for the case in which the interpolation faacions are the fth-order Hermite polynomials of Problem 4.21. “Answer: Kay =P, Gy = ih (exact for five-potot quadeatre), “Ley where yr=sin[in(§ +3) Use three and five-point Gauss quadrature to compute Kx Section 7.2 CCompater exercises (ase FEMIDV2) 1 Solve the problem in Example 3.1 (Set 2) using two, four, and six linear clements, Tabulate the results along with the exact solution, Use the follo dati: L=002m, k=OWarC*, go=10 Wm, L=50C, p= ‘500 Wm!C 7.1. Solve the problem in Example 3.1 (Set 2) using one, two, and three quadratic ‘loments. Compare the finite element results for the temperature and heat flax With the exact solutions at the nodes 178. Solve the heat transfer problem in Example 3.2 (Set 2), using four linear elements and two quadratic elements, and tabulate them with the exact solution st nodes (Gee Table 3.3) 79. Solve the problem in Example 3.4 using four quadratic elements, and compare ‘the solution with that obtained using eight linea elements and the exact solution 7.10, Solve the one-dimensional flow problem of Example 3.4 (Set 2), for dP/dr ~24, using four linear elements and two quadratic elements. Compare the finite clement results with the exact oe. TAAL, Solve the Couette flow problem in Example 3.3 using (a) four linear elements and @) two quadratic elements. Compare the finite clement solution for the temperature with the exact solution. 7.42, Solve the problem of heat flow in a sod (Problem 3.19) using (a) four linear clemenis and (b) two quadratic elements, and compare the results with the analytical solution 7.13, Solve Problem 3.26 using (a) four linear elements and (6) two quadratic clements, and compere the finite element solution with the exact solution at nodes. 7.14, Solve the problem of axisymmetric deformation of a rotating circular disk using (@) four linear elements and (b) two quodratc elements (see Example 7.3) ‘Assume that the body fore is given by f = par 7.48. Solve Problem 3.28[NUMERICAL INTEGRATION AND COMPUTER IMPLEMENTATION 289) 71.16, Solve Problem 3.30. 77, Solve the problem in Example 3.4 using (2) four linear and (6) two quadratic elements, Determine the stress at the fixed end, 7.18-7.29. Solve the beam problems, 4.1-4.12 using the minimum number of Hermite ‘cubic elements, 7.30. Analyze a clamped circular plate under a uniformly distributed transverse load using the Euler-Bernoulli plate element. Investigate the convergence using two, four, and six elements by comparing with the exact solution ot sob-GT where D = EH?/12(1—v7), qo is the intensity of the distributed load, a is the radivs of the plate, His its thickness, and v is Poisson's ratio (v= 0.25). 7.31. Repeat Problem 7.30 with (a) two, four, and six linear and (6) one, two, and three quadratic ‘Timoshenko plate elements (RIE) for a/H= 10 and 100. 7.32, Repeat Probiem 7.30 with the Timoshenko plate element (CIE) for «/#7 = 10 and 10. 7.33, Consider an annular plate of outer radius a and and inner radius b, and thickness H. Ifthe plate is simply supported at the outer edge and subjected to a uniformly distributed moment M, at the inner edge (see Fig. P7.33), analyze the problem using the Euler-Bernoulli piate element. Investigate the accuracy using two and wr) four elements and comparing withthe analytical solution (v= 0.25) bMa?~P) ab?My ir WO~T yw Dy Gye)” () See Problem 7.30 for the definition of D. Radivs @ Line load slong the jnner edge FIGURE P7.33 17.34, Repeat Problem 7.33 with (a) two and four linear and (6) one and two quadratic ‘Timoshenko (RIE) elements for a/H = 10 and 100. 7.35. Consider the simply supported annular plate deseribed in Problem 7.33. Suppose that the inner edge is loaded with a uniformly distributed shearing force Qg, Use ‘meshes of two and four Euler—Bemoulli plate elements to analyze the problem, 7.36, Analyze Problem 7.35 with (a) two and four linear and (b) one and two quadratic ‘Timoshenko (RIE) plate elements for a/H = 10 and 100, 7.31. Repeat Problem 7.36 with the Timoshenko (CIE) plate element,290 rnstre ELEMONT ANALYSIS OF ONE-DIMENSIONAL, PROBLEMS 7.38, Consider a simply supported circular plate of radius a, loaded at the center with a load P. Analyze the problem with two, four, and six Euler-Bernoulli plate elements and compare with the analytical solution (v =0.25) “0= go lagen e-em (| Note that Q}= P/2x. 7.39, Analyze Problem 7.38 with four and eight linear Timoshenko (RIF) plate ‘elements for a/17 = 10 and 100. 7.40. Analyze the simply supported annular plate in Problem 7.33 when itis subjected to a uniformly distributed losd of intensity qo. Use the EulerBernoulli plate element, ‘TAL. Analyze the annular plate in Problem 7.40 using two and four Timoshenko (CIB) plate elements. ¢ 7.42-7.49. Solve the truss and frame problems in Figs. P4.29-P4.26. 7,50. Consider the axial motion of an elastic bar, governed by the second-order ‘equation eu_ au EM for Ox ag PAG fro with the following data: length of bar L=500mm, crossscctional area A Imm?, modulus of elasticity = 20,000N mm *, density p=0,008 N s* may ‘boundary conditions (0,9 bu O BAS (L041 and zero initial conditions. Using 50 linear elements and Af =0.002s, determine the exial displacement and plot the displacement as a function of position along the bar for ¢= 0.85. 7.81. Consider the following aondimensionalized differential equation governing the plane wall transient [see Myers (1971), p. 101): er er Fat Ga0 for 0 > ° 1 4d (a) 6) 4 3 2 : [| / : 1 4 ©) (a nouns $2 Finite elements in two dimensions: (2) a thrce-node element; (6) fournode elements; (c) a five-node clement; (d) a six-node element.smiouevaniante PRontens 305 terms of uf from (8.19): uysule, y= Hex tea t= Uta, Ya) = 1 + Cax2 + aye (8.20) Hs Us, Js) = 61+ eats # Cay where the clement label ¢ on the us, xs, ys, and ¢s is omitted for simplicity. Throughout the following discussion, this format will be followed. In matrix form, we have a Tou nfo wp=|1oa mw lye (8.21) Hs, 1 xs Shes, Solution of (8.21) for ¢; (i= 1, 2, 3) requires inversion of the coefficient matrix in (8.21). The inverse does not exist when any two rows or columns of (8.21) are the same. This happens only when all three nodes lie on the same line, Thus, in theory, as long as the three vertices of the triangle are distinct and do not lic on a line, the coefficient matrix is invertible. However, in actual computations, if any two of the three nodes ate very close relative to the third, oor the three nodes are almost on the same line, the coefficient matrix can be nearly singular and numerically noninvertible. Hence, one should avoid ‘elements with narrow geometries (sce Fig. 8.3) in finite element meshes. Inverting the coefficient matrix in (8.21), we obtain pf @ % lay" Bi Bz Bs}, 2A,=a,+a,t+ay (8.22) th Ya Ya and, solving for ¢; in terms of ty, fe} = [A] {a}, , ee Va te oo Cru a3 ‘Triangular “geometries that should be o avoided in finite element meshes.306 FIVrTE ELEMENT ANALYSIS OF TWO-DIMENSIONAL PROALEMS we obtain (@juy + eeatea + ests) 2A. 1 = Ay (Bites Bata + Bsus) (8.234) 1 (Crs + yatta + Yatts), where A, is the area of the triangle (or 2A- is the determinant of (A]), and «;, B,, and 7, arc the geometric constants , . = Ye ERY Bay Ye } (i#j#k; and 4, j and & permute ina natural order) = Oye) (8.236) Substituting for ¢; from (8.232) into (8.15), we obtain UG y)= Le Mavea teat asus) + Bre + Bate + Bats + (rit + yatta + Ystts)y) (828) FIGURE 84 Linear interpolation functions for the three-node triangular element.SWGLE-VARIABLE rRowLeNs 307 here ‘pf are the linear interpolation functions for the triangular element, (8.25) and a, By, and 7, are constants defined in (8.236). ‘The linear interpolation functions 7 are shown in Fig. 8.4. They have the properties VOR ID= By GI=1,2,3) a 2 ays 2 ays (8.26) DSviet, Ying Silo rot it Ox oy Note that (8.24) determines a plane surface passing through 1), ta, and ws. Hence, use of the linear interpolation functions wf of a triangle will result in the approximation of the curved surface (x,y) by a planar function Ur=Di1ufpi as shown in Fig. 8.5. We consider an example of the computation of yf. Example 8.1, Consider the triangular element shown in Fig. 8.6. Let Ue yactetey=(l x ofc} lay) = Balt Finite element mesa >*° of the domain FIGURE 85 Representation of a continuous function u(x, y) by linear interpolation functions of three-node triangular elements,308 NTT PLEMENT ANALYSIS OF TWODDIMENSIONAL PROBLEMS O38 FIGURE 86 ‘A triangular element with clement nodes and coordinates (for Example 8.1). Te} Gee put layt=3]}-1 3-2 2-1 4 where Substituting into U, we obtain ([A4]“*=[A*]) vien=0 x ns} a(n ofa] = Dvn te as where InAL+ Abe + Agy and Aj are the elements of the inverse matrix [A] Alternatively, fom the definitions (8.230), we have a a5x4-3x3211, ay =3x1-2x 1, Pa=4—1=3, 1-3=-2 Ha-B-2=-1, H=-2-H= art ent aya7 an2x3-Sx1=1 24, ‘The interpolation functions are WerMUL 2-29), VER HCSH3x—y), Sad -2e+3)) LINEAR (FOUR-NODE) RECTANGULAR ELEMENT. Here we consider an approximation of the form (8.16) and use rectangular element with sides a and b (see Fig. 8.7a). For the sake of convenience, we choose a local coordinate system (i, 7) to derive the interpolation functions. We assume that UG, F) =e teak + esF + cakF (8.27)SINGLE-VARIABLE PROBLEMS 309 o FIGURE 3.7 inear rectangular clement and its interpolation fonctions: (a) geometty of the element; (b) interpolation fonctions s and require m= UO, 0) =e, t= Ula, 0) =, 4,0 (8.28) t= Ua, b) =e + cpa + gb + cqab y= UO, b) =e, +636 Solving for ¢ (=1,..., 4), we obtain =m, una (8.29) oa tO310 rHoTe ELEMEKE ANALYGIS OF TWO-DIMENSIONAL PROBLEMS Substituting these into (8.27), we obtain aobtaa) ta; iS u@3)=m(1 where or, in concise form, wenn) 1-EF (8.316) where (i, j,) are the (f, 7) coordinates of node i. The interpolation functions are shown in Fig. 8.7(b). We again have WE N= Oy im be, Dvi (6.324) where (%, J) are the coordinates of node j in the (&, ¥) system. Note that the linear interpolation functions for the four-node rectangular element can also be obtained by taking the tensor product of the 1-D linear interpolation functions (3.18) associated with sides 1-2 and 2-3: “tes del yl ea ‘The procedure given above for the construction of the Lagrange interpolation functions involves the inversion of an 7x matrix, where n is the number of nodes in the element. When 1 is large, the inversion becomes very tedious. The alternative procedure discussed in Chapter 3 for one-dimensional elements proves to be algebraically simple. Here we illustrate the alternative procedure for the four-node rectangular element. Equation (8.324) requires that vi, =O (F=2,3,4), wiley) ‘That is, yi is identically zero on lines be of the form and y=. Hence, pi(@, j) must HE N=ela—HO-)SINGLEVARIABLE PRosLeNs 311 Using the condition {(x,, yi) = wi(0, 0) = 1, we obtain c= 1/ab, Hence, VG =F eH -Y)=(1-4)(1-F) ab Likewise, one can obtain the remaining three interpolation functions. 8.2.6 Evaluation of Element Matrices, and Veetors ‘The exact evaluation of the element matrices [K*] and {f*} in (8.140) is, in general, not easy. Therefore, they are evaluated using numerical integration techniques described in Section 74.5. However, when ay, a, and f are clement-wise constant, it is possible to evaluate the integrals exactly over the linear triangular and rectangular elements discussed in the previous section. ‘The boundary integral in (Q*) of (8.145) can be evaluated whenever q, is known. For an interior element (j.e., one that does not have any of its sides on the boundary of the problem), the contribution from the boundary integral cancels with similar contributions from adjoining elements of the mesh (analogous to the QF in one-dimensional problems). A more detailed discussion is given below. For the sake of brevity, we rewrite (K matrices [S°*] (a, B =0, 1, 2): in (8.14) as the sum of four basic {KJ = aog(S?] + aus[S4] + ayefS"] + 4,[87)" +4,{5"] (8.34) where [ }” denotes the transpose of the enclosed matrix, and SHP~ | Via Vpndray 625) With Yio = 9 Bay X=, and x2= 3 Yo = Yor All matrices in (8.34) and interpolation functions in (8.35) ate understood to be defined over an element; i.e., all expressions and quantities should have the element label e, but these ‘are omitted in the interest of brevity. We now proceed to compute the matrices in (8.34) and (8.14b) using the linear interpolation functions derived in the previous section. ELEMENT MATRICES FOR A LINEAR TRIANGULAR ELEMENT. For a triangle, the following exact integral formulae are available for evaluating the integrals. Let Ina [ ry" dedy (6.36)312 NITE ELEMENT ANALYSIS OF TWO-DIMENSIONAL PROBLENS ‘Then we have (the element label on A is omitted) Jon= A. (area of the triangle) I= Ak, £=1Y x, =A 9 y= A(S nino), wwe obtain yt oat ys Si = Tab SP= Zar =F [levy + C+ e+ Cony + ard 639) +4 [oBiBy + fa) + 1B) + Tava} In view of the identity aj +f. + y,9 = 4A [which follows from (8.235) and (8.37)), for an element-wise-constant value of f =f,, we have Fi MG + Bib + 19) = MeAe (8.40) This should be obvious, because for a constant source j,, the total magnitude of the source on the element is equal to f.A., which is distributed equally among the nodes. Once the coordinates of the clement nodes are known, one can compute a, f;, and y; from (8.235) and substitute into (8.39) to obtain the element matrices, which in turn can be used in (8.34) to obtain the element matrix (K*]. For example, when aja, aay, and ago are zero, and ay, and az, are clement-wise-constant, we have 1 Kiya qq CuBiBi + aay) (8.41) ELEMENT MATRICES FOR A LINEAR RECTANGULAR ELEMENT. When ay (é,7=0,1,2) and f are constants, we can use the interpolation functions of (8.31a) expressed in the local coordinates £ and 9, which are related to the global coordinates by xeitxh y= +y, de=ar, (8.42) where (x5, yi) are the global coordinates of node 1 of the element Q* withSINGLEVARIASLE PRORLENS 313 respect to the global (x, y) coordinate system. For example, we have se LP wmacdy= [ [vn aeay where a and 6 are the lengths along the £ and J axes of the clement. Consider the coefficient st=[Lomda=LE( -[(-Dal(-po-z Similarly, we can evaluate all the matrices [S°] with the aid of the integral identities. (8.43) We have “1 1 sm t 1 -1 [87] (8.44) {f}=Yab{1 101 17 EVALUATION OF THE BOUNDARY INTEGRALS, Here we coi evaluation of boundary integrals of the type ai= farted (45) the where qf is a known function of the distance s.along the boundary T*, It is not necessary to compute such. integrals when a portion of I* does not coincide with the boundary I of the total domain ©, On portions of I* that are in the intetior of the domain &, q$ on side (, /) of the element 2 cancels with 9 on side (p, q) of the element 2 when sides (, j) of 2° and (p, q) of Q! are the same (ie., at the interface of Q° and Q/), This can be viewed as the ‘equilibrium of the internal “flux” (See Figs. 8.8b,c). When T° falls on the314 Fre ELEMENT ANALYSIS OF TWO.DIMERSIONAL PROMLEMS ‘4 (specitied) 4 (specified) ‘= tive boundary of the domain ® © FIGURE B& Computation of boundary forces and equilibrium of secondary variables at interefement boundaries: (@) finite cloment discretization; (6) equilibrium of forces at interfaces; (€) computation of forces on the true boundary. boundary of the domain 2, g% is either known, in general, as a function of s, or is to be determined in the post-computation. In the latter case, the primary vatiable will be specified on the portion of the boundary where qq is not specified. ‘The boundary I” of linear two-dimensional elements is a set of linear ‘one-dimensional elements. Therefore, the evaluation of the boundary integrals al problems amounts to evaluating line integrals. It should not be surprising to the reader that when two-dimensional interpolation functions are evaluated on the boundary, we obtain the corresponding one-dimensional interpolation functions. For example, consider the linear triangular clement shown in Fig. 8.9. ‘The linear interpolation functions for this clement are given by (8.25). Now let us choose a coordinate system (s, £) with its origin at node 1 and the coordinate s parallel to the side connecting nodes 1 and 2. The two coordinate systems (x, y) and (s, f) are related by xeatbs tert yay t bs teat ‘The constants a4, bs, ¢1» @2, ba, and ¢2 can be determined with the followingSIxGLevARIADLE PROBLEMS 31S when s=0, 1=0, r=x, y=y1 whert when We obtain xG Dent Gan) e4[(E tr Ent (6.46) yo Dm n+ 0-954 [(E-a)n—Se tos] ‘These equations allow us to express v(x y) as W(6,4), which can be evaluated on the side connecting nodes 1 and 2 by setting ¢=0 in is, 1): ws) = ws, 0) = pCxGs, 0}, y(s, 0) AO=ntG IE Yat WE For instance, woman al(istie]o[0-Dneor]) i Blatant ay(1-2 where the definitions of a, Bx, and y; have been used to rewrite the entire expression, Similarly, n=", ys(s)=0 wals)= 7, e966) where a= hy. is the length of side 1-2. We note that Y(s) and als) are precisely the linear one-dimensional interpolation functions associated with the line element connecting nodes 1 and 2. Similarly, when y(t, y) are evaluated fon side 3-1 of the element, we obtain w= prs va=0, vals 5 where the s coordinate is taken along the side 3-1, with origin at node 3, and ‘hy is the length of side 1-3. Thus evaluation of Qf involves the use of appropriate 1-D interpolation functions and the known variation of g, on the boundary: O5=[ vasetds +f wisiaisys+ f wlsdat) as =QitQiat Vis (8.47)316 nine ELEMENT ANALYSIS OF TWO-DIMENSIONAL PROBLEMS FIGURE 89 ‘The linear triangular clement inthe glabel (x, 9) and local (5,2) coordinate systems. where J; denotes the integral over the line connecting node i to node j, the s coordinate is taken from node i to node j, with origin at node i, and Qf, is defined as the contribution to QF from gf on side J (see Fig. 8.9) of the element 2%: i= [wands 48) For example, Gimp ancd=[ Gdsrdro+! (deeds ‘The contribution from side 2-3 is zero, because y, is zero on side 2~3 of a ygular element, For a rectangular element, Q§ has contributions from sides 1-2 and 4-1, because is zero on sides 2-3 and 3-4, Example 8.2. Consider the evaluation of the boundary integral Qf in (8.48) for the four cases of q(5) and finite elements shown in Fig. 8.10. For each case, we must use the 4(6) and the interpolation functions associated with the type of boundary element linear or quadratic). Case f. q(s) = qo constant; linear element: Oi $ a0 demon [yds +040 =1.2.3) where s at - Walp, Waa wed We have Dix boke (= Oi), i= tach, (= OH), 5 Case 2. ¢(s) = qos/h, (linear variation); linear element: Of wenden B [suds 6=1.2.3)SNOLEVARIABLE PROBLEMS 317 a he ’ a ‘a0 is 1 y 1 Ss case 3 Case 2 ‘0 ‘ 4 1 ns 2 e . 0 1 case case 4 HIGURE 6.10 Evaluation of boundary integra in the nite element analysis (Example 8.2) where We have Os=Iaohe (= Oh), =O Oi= lack. (= Case 3, 4(6)= qo constant; quadratic element: Oi=$ gods (12.0048) “iri: 0-7), and Ys, Ys, and Ys are zero on side 1-2-3. We have Qi= lack, (= Oi), O3=dgehe (= ON), O5= Laake (= 05) Case 4. q(5) a8 shown in Fig. 8.10; linear element: a= $ aernae=[ aoind+ | ayas+o =Qr+02 (=)318 FINITE ELEMENT ANALYSIS OF TWODDIMENSIONAL PROBLEMS We obtain, ae J agi (i-Z)aroro= baohnn (= 08), ore [ agar a(t Z)aro lava thes (= 054-05) @5=04 | gz ds+0= tasks (= 0%) ps ° has 8.2.7 Assembly of Element Equations ‘The assembly of finite element equations is based on the same two princiles that were used in one-dimensional problems: 1. Continuity of primary variables . 2, “Equilibrium” (or “balance”) of secondary variables We illustrate the procedure by considering a finite element mesh consisting of f triangular element and a quadrilateral clement (see Fig. 8.11a). Let Ki (7=1,2,3) denote the coetficient matrix corresponding to the triangular element, and let K7 (i,/=1,..., 4) denote the coefficient matrix corresponding to the quadrilateral clement, From the finite clement mesh shown in Fi B.11(a), we note the following correspondence (j.e., connectivity relations) Global» Loeal Ky Kl Ki Kh ke Kut Kf Ky 0 Ks ° Ke Rib + ka Ku Kg ke Kio Kd Ke : Kea Kyl + Ki + Kil + Kit Kas Kit Ke Ks 0 Key Kat Kis FIGURE 8.11 Assembly of finite element coefficient matrices using the corcespondence between global and ‘clement nodes (one unknoséa per node): (a) assembly of two elements; (4) assembly of several elements,SINGLE-VARIABLE FROBLENS 319 between the global and clement nodes: 123 al (64 where X indicates that there is no entry. The correspondence between the local and global nodal values is (see Fig. 8.114) Wh= Uy, wb=ut=U, wW=uk=U;, W=Uy W=Us (8.50) which amounts to imposing the continuity of the primary variables at the nodes common to elements 1 and 2. Note that the continuity of the primary variables at the interelement nodes guarantees the continuity of the primary variable along the entire intetelerinent boundary. For the case in Fig. 8.11(@), the requirement ul=uj and uj=u guarantees UG) = U%s) on the side connecting global nodes 2 and 3. This can be shown as follows, The solution U*(¢) along the line connecting global nodes 2 and 3 is linear, and is given by wey ani(1—F) aude where s is the local coordinate along side 2-3 with its origin at global node 2 and h is the length of side 2-3 (or side 2). Similarly, the finite element solution along the same line but from element 2 is along the interface of the two elements. Next we use the balance of secondary variables. At the interface between the two elements, the fluxes from them should be equal in magnitude and ‘opposite in sign. For the two elements in Fig. 8.11(@), the interface is along the side connecting global nodes 2 and 3. Hence, the internal flux q}, on side 2-3 of clement 1 should balance the ffux q3 on side 4-1 of element 2 (recall the sign convention on q): @irs= Ger oF Gie3=(-as (51) In the finite element method, we impose the above relation in a weighted- integral sense: [amen [aves fatwa [aria @20)320 NITE ELEMENT ANALYSIS OF TWO.DIMENSIONAL PROBLEMS where As, denotes the length of the side connecting node p to node q of the element 2°. The above equations can be written in the form J aavids +f anwi ds = 0, givids+ | qhpids=0 (8.52b) hts hie 3 7 or Qt Oe Oat Qis=0 (8:52¢) where Qj denotes the part of Qf that comes from side J of element ¢ [see (8.48)]: oo= | aivrae “The sides of triangular and rectangular elements are numbered as shown in Fig. 8.11. These balance relations must be imposed in assembling the clement equations. We note that Qj is only a portion of Q; [see (8.47)]. ‘The element equations of the {wo elements are written first. For the model problem at hand, there is only one primary degree of freedom (NDF = 1) per node. For the triangular element, the element equations are of the form Khul + Khu + Kind =fi + Qi Kul + Khud + Kite (8.532) Kul + Kid + Kind = f+ 08 For the rectangular element, the element equations are given by Kiui+ Kind + Kis + Kina = fi+ QF Kid + Kanu + Ki + Khu = f3-+ OF ee ie 2 (6.538) Kul + Kid + Ki + Kha = 3+ 3 Kuh + Kind Kilt Kiu= f+ OF In order to impose the balance of secondary variables in (8.52), we must add the second equation of element 1 to the first equation of element 2, and also add the third equation of element 1 to the fourth equation of element 2: (Kul + Kul + Khu) + (Kul + Kiu-+ Kind + Kia) = (+O + (+08) (Kou) + Kjau3 + Khel) + (Rand + Riu + Kind + Kiel) =(f3 + Q3) + (fi + OD) Using the global-variable notation in (8.50), we can rewrite the above equations as follows {this amounts to imposing continuity of the primary variables in (8.50)]: Ka.U, + (Kia + Ki,)Us + (Kis + KYs)Us + KinUy + KigUs= ft + fi + (Q3+ 01) KU + (Kia + KR)U2 + (Kis + Kis)Us + KinUs + KigUs= f+ fi + (Q3+ 03)