Você está na página 1de 25
an Claes Johnson Numerical solution of partial differential equations by the finite element method Cho, 4 CAMBRIDGE UNIVERSITY PRESS New ya New Rochelle Melbourne Sydney Library of Congress cataloging in pubis Contents Preface 7 6. Introduction 9 0.1 Background 9 0.2 Difference methods ~ Finito element methods 10 0.3 Scope of the book 12 Remarks on programming 43 1.9 Remarks on finite element software 48 2. Abstract formulation ofthe finite element method for eliptic problems 50 2.1 Introduction. The continuous problem 50 2.2 Discretization, An error estimate 52 23 The energy norm SS 24 Some examples 55 3. Some finite element spaces 67 3.1 Introduction. Regularity requirements 67 3.2 Some examples of finite elements 68 33 Summary 79 4. Approximation theory for FEM. Error estimates for elliptic problems 84 4.1 Introduction 84 42 Interpolation with piecewise linear functions in two dimensions 84 4.6 Adaptive methods 94 4.7 An error estimate in the La()-nom 97 ‘8 Some applications to eliptic problems 101 5.1 The elasticity problem 101 5.2 Stokes problem 106 5.3 A plate problem 108 6, Direct methods for solving linear systems of equations 112 31 model problem 147 space 149 pace and time 152 9. Hyperbolic problems 167 9.1 Introduction 167 9.2 A convection-diffusion problem 168 9.3 General remarks on numerical methods for hyperbolic equations 171 94 Outline and preliminaries 173 9.5 Standard Galerkin 176 9.6 Classical artificial diffusion 181 equations 216 togral equation for an exterior Dirichlet problem using a single ial 219 Dirichlet problem with double layer potential 220 for a Fredholm equation of the first kind 224 2. FEM for a Fredholm equation of the second kind 227 finite element methods 232 troduction 232 some examples 234 12, Curved elements and numerical integration 239 12.1 Curved elements 239 12.2 Numerical integration (quadrature) 245 13. Some non-linear problems 248 fntroduction 248 2 Convex minimization problems 248 2.1 The continuous problem 248" 254 for convex minimization problems 255 problem 257 equations 258 13.6 Compressible flow: Bus References 269 Index 275 Preface ‘The purpose ofthis bookisto give anca: clement method as a general method differential equations in mechanics an: ‘types of equations, namely elliptic, parab ‘main part of the text is concerned indicating extensions also a chapter on fini elliptic problems. The ‘undergraduate courses at Chalmers Univer The first haf ofthe book (Chapt for the first time. ‘The emphasis of the toxt is on mathem: finite element method but many applic ‘mechanics and physics are also given. I have as simple as possible whi both mathematical questions on existence and regularity together with ‘numerical methods. Ihave found this to be a fruitful approach where, 0 hhand the numerical methods can be given the necessary mathematical background, and on the other hand, the fascinating and important techniques 1 ‘background is included and the reader is referred ‘complete account. ‘The list of references is limited and many contributions important for the development of the various subjects have been omitted. I have given just @ {few references leading into the very rich literature on finite element methods, want to thank Prof Raymond Chandler for the English, Tekn Lic eter Hansbo for supplying most of the nume Eriksson for reading parts of ‘Karlsson who swiflly typed a helped me with seemingly endless al Gateborg in July 1987 ‘Claes Johnson 0. Introduction ‘computers over the |: models have dramatically increased. Using comput ematical models, one can simulate and analyze com engineering and science. This reduces the need for expensive and time- consuming experimental testing and makes it poss different alternatives for optimization, etc. In fact, an intense activity has started in Computer Aided Design, Engine 1g" is emerging as a complement to theor ical models on a computer one needs numerical methods. Only in the very simplest cases is it possible to find exact analytical solutions of the equations in the model, and in general one has to rely on numerical techniques for finding approximate solutions. The finite element method (FEM) isa general technique for numeric n of differential andintegral i fe and engineering. ‘50's and early 60's for the mume ‘ructural engineering (el ‘the method was ‘engineering for beams, frames ar ‘where th structure was subdivided into small parts, so-called finite elements, with known simple behaviour. When the mathematical study of the finite ‘element method started in the mid 60's it soon became clear that in fact the ‘method is a general technique for numerical solution of partial differential ‘equations with roots in the variational methods in mathematics introduced in the beginning of the century. During the 60's and 70's the method was differential equation equations, ete). At of earlier methods in struct ° developed, by engineers, ‘general method for numeri integral equations with appli ‘Today, finite element meth thematicians and numerical anal of partial differential equat in many ateas of science and engineering. uséd extensively (often integrated in CAD suctural engineering, strength of materials, ‘iuid mechanics, nuclear engineering, electro-magnetism, wave-propagat scattering, heat conduction, convection-lffusion processes, integrated cir- cuits, petroleum engineering, reaction-diffusion processes and many other 0.2 Difference methods — Finite element methods ‘The basic idea in any numerical method for a differential equation is to discretize the given continuous problem with infinitely many degrees of freedom to obtain a discrete problem or system of equations with only finitely ‘many unknowns that may be solved using a computer. The classical numerical method for partial differential equations isthe difference method where the discrete problem is obtained by replacing derivat ifference quotients involving the values of the unknown ‘equivalent variarional problem. In the case of elliptic equations this variational problem in basic cases is a minimization problem of the form (Find ueV such, where Vis a given set of admissible functions and F:V— Ris a functional ie F()eR for all veV with R denoting the set of real numbers). The functions tac in an elastic body, a tempers vand (M) corresponds to an equi differential equation the considered system. In general the dimension of V is {functions in V cannot be described by a finite number of par ‘that can be solved on a computer the idea inthe finite clement method isto replace V by set Vi consisting of simple functions only depending on finitely many parameters. This leads to a finite-dimensional minimization problem ofthe form: 10 (My) Find uyeVy such that F(us)Me. also indicate how to see that if u isthe solution of (¥) then w s (D). Thus, we assume that ueV satisfies Ju'v'dx—Jfvdx=0 — WeV. a ° If we now assume in addition that u" exists and is continuous, then we can {integrate the first term by parts to get, using the fact that v(Q)=v(1)=0, —fcwrsovax-0 We. ‘But wit the assumption tat (us continuous this relation can only hold if (Problem 1.1) and it follows that w is the ‘Thus we have seen that 1 solution of (V) and in addition satisfies 1 regularity assumpton (u" is continuous), then w isthe solution of (D). Itis now possible to show that if w is the solution of (V), then u in fact satisfies the desired regularity assumption and that the three problems (D), (V) and (| below). are equivalent (cf Section 15 Problems 1.1 Show that if w is continuous on [0, 1] and fwdr-0 — Wev, a }. le assumptions the problem B above can be A). then w(x) =0 for xe 1.2. Show that under «given the formula 1.2 FEM for the model problem with piecewise linear functions ‘We shall now construct a finite-dimensional subspace Vy of the space V defined above consisting of piecewise linear functions. To this end let OexyI since in this ease for all xe ius, the matrix A is tri-diagonal, the main diagonal and the two adjoining diagon be different from zero. We have for j=1,... .. M, Note also that the matrix A is symmetric and positive definite since ro M n') and with v(x)= 19)(0), we have M mi, =( 2 i ‘with equality only if v'=0, that is since v(0)=0 onl a(Bu+6v, w)=Ba(u, w)+8: The bilinear form a(.,.) on VV is said to be symmetric if +0L(¥), 2 VEVR, . forall u, ¥, iinear form on VV ear in each argument, ale, j=a(w.y) Wy, weV. A symmetric bilinear form af. ,.) on VV is said to be a scalar product on Vit a(v,¥)>0 We, v0, ‘The norm |j-||, associated with a scalar product af defined by wev. Further if <. > isa scalar product with corresponding norm have Cauchy's inequality (1.28) || We further recall that if V ith a scalar product with ‘corresponding norm |||, then V is said tobe a Hilbert space if V is complete, ie, if every Cauchy sequence with respect to |]isconvergent, We recall hat a sequence v1, V2, ¥3. “ements v; inthe space V with norm | is said to be a Cauchy sequence if for all €>0 there isa natural number N such that liv-vlN. Further, v, converges to vif |Iv-vi| > Oasi—> «©. The reader unfamiliar with the concept of completeness may bypass this remark land think of a Hilbert space simply as linear space with a scalar product. ‘Wennow introduce some Hilbert spaces that are natural to use fr variational formulations of the boundary 8 we will consider. Let us start ‘an interval, we define the space uare integrable func 1o(l)=(0: vis defined on Tand {vidx<=) ‘The space L2(0) is a Hilbert space with the scalar product Oo, wef ds, and the corresponding norm (the Lz-norm): Iv2dx) 2 By Cauchy's inequality, if vand space is complete requires some familiarity withthe Lebesgue integral. In this book, however, itis sufficient to get an idea of La ‘mann integral; from this point of view we may veLa(I) as a piecewise continuous function, possibly unbounded, such that fu¥da0. 1.13 Prove Cauchy's inequality (1.28) 1.14 Consider the problem corresponding to (D) with an inhomogeneous boundary condition, ie, the problem -dusf in, =u) on. }=|x\* belongs to asp where f and up are given. Show that this problem can be given the following equivalent variational formulations: (¥) Find weV(uo) such that a(u,x)=(Ev) YweH{(@), (M) Find ue V(u) such that F(u)SF(W) Wwe V(t), where , V(as)= (eH): v=up on F} ‘Then formulate a finite element method for (1.31) and prove an error estimate, 1.6 A geometric interpretation of FEM ethod in geometric two elem 1d w in inear space with scalar product <.,.> are said to be orthogonal =0. us for simplicity consider the following variant of our previous problem -Autu=f in Q, 2) u=0 on, (cf Problem 2.5 below). The corresponding variational problem reads: Find ‘teF(Q) such that {Vu- Vude+ fav dx a 2 wertl(@), (1.33) =(6,¥) weH{(o), using the notation ={[Tu- Vwtuv}és. a ‘Note that <. ,.> isin fact the scalar product in the space HY(Q). sional subspace of Hi(Q), eg the space of piecewise 1.4, and consider the following finite element tyeVp such that (134) =(6¥) WeV), ince VxcH\@) we may choose veVs in (1.33) and on substraction from (1.34), we obtain 38 2.35) =0 WEY, ie the error u~us is orthogonal to Va with respect to <. ,.>. We may also express this fact as follows: The finite element solution uy is the projection. with respect to <. ,.> of the exact solution u om Vn, ie, uy is the element in )-norm | «|(o), oF in other words Wei Vi Fg 13 function f. Thus, we best approximation uy of u, without ‘knowing w itself, kno =Autusf in Q and u=0 on F. This remarkable fact eflerts the eli of the boundary value problem (132) Problem 1415 Prove that (1.35) and (1.36) are equivalent (cf the proof of Theorem Al). 1.7 A Neumann problem. Natural and essential boundary conditions ‘We shall now consider a problem with another type of boundary conc ‘namely the following Neumann problem (D): (37) —Autust in @ au, a asm) Bag r, where agin i Bounded domain wih boundary F and 2 denotes the outward normal derivative to T. The boundary condition is a Newmann condition while the boundary condition u=uy on T considered previously is said to be a Dirichlet condition. In mechenies or physics the Neumann condition (1.376) corresponds to a given force or flow g on ‘We can give the problem (1.37) the following variational formulation (V): Find uel(Q) such that (1.38) alu, =, v)+ WeHH(Q), where a, s)=f[Tu- Vvtwvlds, (6 v=f fd, e J Br ds This is equivalent to the following minimization formulation (M): Find ‘ucH(Q) such that F(u) Wev,. AS in Section 1.4 we see that this problem has a unique solution uy that can be determined by solving a symmetric, positive definite linear system of equations. We also have the following error estimate Wevi, tu—unlt(ay=Ch, if u is regular enough. The function uy will satisfy the Neumann condition rors, 38 whe sn aprorination og on (bln ‘Remark When formulating a difference method for (1.37) one meets severe difficulties due to the boundary condition (1.37b) unless © has a very simple shape such as a rectangle. On the other hand, in the finite element formula- tion, the same boundary condition does not cause any complication. Problems 116 ‘Show that the problem u(0)=u'(1)=0, ‘can be given the following variational formulation: Find ueV such that (@.v)=x) — Wev, where V= (veH(1): v(0)=0}. Formulate a finite element method for this problem usi wise linear functions. Determine the corre- sponding linear system of equations in the case of a uniform partition and study in particular how the boundary condition u'(1)=0 is approximated by the method. ‘Show that the problems (i) and (V) of this section are equivalent. 1.18 Let @ be a bounded domain in the plane and let be divided into two parts Ty and T>. Give a varia of the following problem: boundary T of mal formulation duet ing, =u) in Ty, au. au, onl, an * where f, up and g are given functions. Then formulate a finite element method for this problem. Also give an interpretation of this problem Wert), siven node, i=1,. .., M. Prove that Gj is given by (x) for 0