Chapter 1

ELEMENT METHOD
1.1

Historical perspective: the origins of the nite element method

The nite element method constitutes a general tool for the numerical solution of partial dierential equations in engineering and applied science. Historically, all major practical advances of the method have taken place since the early 1950s in conjunction with the development of digital computers. However, interest in approximate solutions of eld equations dates as far back in time as the development of the classical eld theories (e.g. elasticity,

DR

electro-magnetism) themselves. The work of Lord Rayleigh (1870) and W. Ritz (1909) on variational methods and the weighted-residual approach taken by B. G. Galerkin (1915) and others form the theoretical framework to the nite element method. With a bit of a stretch, one may even claim that Schellbachs approximate solution to Plateaus problem (nd a surface of minimum area enclosed by a given closed curve in three dimensions), which dates back to 1851 is a rudimentary application of the nite element method. Most researchers agree that the era of the nite element method begins with a lecture presented in 1941 by R. Courant to the American Association for the Advancement of Science. In his work, Courant used the Ritz method and introduced the pivotal concept of spatial discretization for the solution of the classical torsion problem. Courant did not pursue his idea further, since computers were still largely unavailable for his research. More than a decade later Ray W. Clough, Jr. of the University of California, Berkeley, 1

AF

INTRODUCTION TO THE FINITE

Introduction

and his colleagues essentially reinvented the nite element method as a natural extension of with coining the term nite element, has spent the summers of 1952 and 1953 at Boeing under the supervision of M.J. Turner working on modeling of the vibration in a wing structure and it is this work that he led to his formulation of nite elements for plate structures. An apparently simultaneous eort by John Argyris at the University of London independently led to another successful introduction of the method. It should come as no surprise that, to a large extent, the nite element method appears to owe its reinvention to structural engineers. In fact, the consideration of a complicated system as an assemblage of simple components (elements) on which the method relies is very natural in the analysis of structural systems. In the few years following its introduction to the engineering community, the nite element method has attracted the attention of applied mathematicians, particularly those interested in numerical solution of partial dierential equations. In 1973, G. Strang and G.J. Fix authored the rst conclusive treatise on mathematical aspects of the method, focusing exclusively on its application to the solution of problems emanating from standard variational theorems. The nite element has been subject to intense research, both at the mathematical and technical level, and thousands of scientic articles and hundreds of books about it have been authored. By the beginning of the 1990s, the method clearly dominated the numerical solution of problems in the elds of structural analysis, structural mechanics and solid mechanics. Moreover, the nite element method currently competes in popularity with the nite dierence method in the areas of heat transfer and uid mechanics.

DR
1.2 tion
ME280A

Introductory remarks on the concept of discretiza-

The basic goal of discretization is to provide an approximation of an innite dimensional system by a system that can be fully dened with a nite number of degrees of freedom. To clarify the notion of dimensionality, consider a deformable body in the three-dimensional Euclidean space, for which the position of a typical particle with reference to a xed coordinate system is dened by means of a vector x, as in Figure 1.1. This is an innite dimensional system with respect to the position of all of its particle points. If the same body is assumed to be rigid, then it is a nite dimensional system with only six degrees of freedom. Version: August 31, 2010, 13:41

AF

matrix structural analysis and published their rst work in 1956. Clough, who is credited

Structural analogue substitution method

Figure 1.1: An innite degree-of-freedom system

A dimensional reduction of the above system is accomplished by placing a (somewhat severe) restriction on the admissible motions that the body may undergo. Finite dimensional approximations are very important from the computational standpoint, because they often allow for analytical and/or numerical solutions to problems that would otherwise be intractable. There exist various methods that can reduce innite dimensional systems to approximate nite dimensional counterparts. Here we consider three of those methods, namely the physically motivated structural analogue substitution method, the nite dierence method and the nite element method.

1.2.1

Structural analogue substitution method

Consider the oscillation of a liquid in a manometer. This system can be approximated (lumped) by means of a single degree-of-freedom mass-spring system, as in Figure 1.2.

DR
manometer walls). Version: August 31, 2010, 13:41

Clearly, such an approximation is largely intuitive and cannot precisely capture the complexity of the original system (viscosity of the liquid, surface tension eects, geometry of the

Figure 1.2: A simple example of the structural analogue method

AF

Introduction The structural analogue substitution method, whenever applicable, generally provides

also the delity of its results) can vary widely. The network analysis of Kron in the 1930s and 1940s is generally viewed as a typical example of the structural analogue approach.

1.2.2

Finite dierence method

Consider the ordinary dierential equation k

AF
u(L) = uL , d2 u . ul+1 2ul + ul1 = , dx2 x2
l

d2 u = f in (0, L) , dx2 u(0) = u0 ,

T
l+1 N N +1 ; 0 1 1 ,
l + 1 l

coarse approximations to complex systems. However, its degree of sophistication (hence,

(1.1)

where k is a constant and f = f (x) is a smooth function. Let N points be chosen in the interior of the domain (0, L), each of them equidistant from its immediate neighbors. An algebraic (or dierence) approximation to the second derivative may be computed as (1.2)

with error o(x2 ). Indeed, assuming that the solution u(x) is at least four times continuously 0 x

DR
x x 0 1 l1 l point l in Figure 1.3, write ul+1 = ul + x
l l

Figure 1.3: The nite dierence method in one dimension

dierentiable and employing twice a Taylor series expansion with remainder around a typical

du x2 d2 u x3 d3 u x4 d4 u + + + dx 2! dx2 3! dx3 4! dx4 du x3 d3 u x4 d4 u x2 d2 u + + dx 2! dx2 3! dx3 4! dx4

(1.3)

ul1 = ul x

; 0 2 1 .
l 2

(1.4)

ME280A

Version: August 31, 2010, 13:41

Finite element method Adding the above equations results in ul+1 2ul + ul1 x2 d2 u = dx2 x2 4!
l

T
d4 u dx4 +
l + 1

d4 u dx4

(1.5)

l 2

so that ignoring the second term of the right-hand side, the proposed approximation to the second derivative of u is recovered. Applying the dierence equation (1.2) to nodal points 1, 2, ..., N , and accounting for the boundary conditions (1.1)2,3 gives rise to a system of N linear algebraic equations u 2 2u 1 =

ul+1 2ul + ul1

AF
2uN + uN 1 f1 x2 u0 , k fl x2 = , k fN x2 uL , = k qle du = k dn
e l

l = 2, ..., N 1 ,

(1.6)

with unknowns ul , l = 1, 2, . . . , N . Again, an innite-dimensional problem with respect to the value of u in the domain (0, L) is transformed by the above method into an N -dimensional problem. Clearly, the state equations are (approximately) satised only at discrete points 1, 2, ..., N . Also, the boundary conditions are enforced directly when writing the discrete counterparts of the state equations in the nodes that reside next to the boundaries. It is easy to see that nite dierence methods run into diculties when dealing with complex boundaries due to the need for spatial regularity of the grid.

DR
1.2.3 Finite element method
domain is equal to and qle+1 du = k dn
e

Revisit the problem in the previous section and consider the same discretization as in Figure 1.3. Consider the line segment between points l and l + 1. This is now the domain of the nite element e. In this domain, we assume that u varies linearly, as shown in Figure 1.4,
du The normal ux q = k dn , where n denotes the outward unit normal to the element

and attains values ul at point l and ul+1 at point l + 1.

ul ul+1 . = k x ul+1 ul . = k x

(1.7)

(1.8)
ME280A

l+1

Version: August 31, 2010, 13:41

Introduction

e1 l1 l

Figure 1.4: A one-dimensional nite element approximation

at points l and l + 1, respectively. These two equations can be written in matrix form as

AF
k x 1 1 1 ul 1 ul+1 = qle qle+1 . k x 1 1 1 u l 1 ul 1 =
1 qle 1

T
e l l+1 qle1 .
e l

(1.9)

An analogous matrix equation can be written for element e1, whose domain lies between points l 1 and l, and takes the form

(1.10)

Now, adding the rst equation of (1.9) to the second equation of (1.10) yields k (ul+1 2ul + ul1) = qle + qle1 . (1.11) x To approximate the right-hand side of (1.11), rst note that the terms qle and qle1 represent uxes on opposite sides of the same point l, namely, recalling (1.7) and (1.8), qle + qle1 = k du dx k du dx
e 1

(1.12)

DR
ME280A

du = f dx. Hence, if dx L the total force ftotal = 0 f dx is distributed to the points 0, 1, . . . , N + 1 so that to point l du corresponds a force f at l is exactly f l , then the jump in the normal derivative k l , therefore dx (1.11) attains the form k (ul+1 2ul + ul1 ) = f (1.13) l . x Then, the complete nite element system becomes f 1 x u0 , u 2 2u 1 = k f l x ul+1 2ul + ul1 = , l = 2, ..., N 1 , (1.14) k f N x uL , 2uN + uN 1 = k At the same time, one may rewrite the dierential equation as k Version: August 31, 2010, 13:41

Classications of partial dierential equations

This is the so-called direct approach to formulating the nite element equations. Upon to within the denition of the force term. Yet, these equations were derived by way of fundamentally dierent approximations. It will be established that in nite element methods the state equations are satised in an integral sense over the whole domain with respect to a set of (simple) admissible functions. Also, it will be seen that boundary conditions can be handled trivially.

1.3

Classications of partial dierential equations

Consider a scalar partial dierential equation (PDE) of the general form F (x, y, ...u, u,x , u,y , ...u,xx , u,xy , u,yy , ...) = 0 , (1.15)

where x, y , . . . are the independent variables, and u = u(x, y, . . .) is the dependent variable. In addition, write

AF
u,x = u x , u,xx = 2u x2 , etc .
3u,x +u,y u = 0 1 u,yy 3u = 0 xu,xx + y + u, u,2 yy = 0 xx 2 u,x u,xxx +u,yy = 0

T
(linear rst order) , (linear second order) , (non-linear second order) , (non-linear third order) .

comparing (1.6) and (1.14), it is concluded that the two sets of equations are identical

(1.16)

The order of the PDE is dened as the order of the highest derivative of u in (1.15). Also, a PDE is linear if the function F is linear in u and in all of its derivatives, with coecients depending on the independent variables x, y, . . ..

DR
Examples: dierential equations of the general form Version: August 31, 2010, 13:41

For the purpose of the forthcoming developments, consider linear second-order partial

au,xx + bu,xy + cu,yy = d ,

(1.17)

where not all a, b, c are equal to zero. In addition, let a, b, c be functions of x, y only, whereas d can be a function of x, y, u, u,x , u,y . Equations of the form (1.17) can be categorized as follows:
ME280A

Introduction (a) Elliptic equations (b2 4ac < 0) A typical example of an elliptic equation is the two-dimensional version of the Laplace (Poisson) equation used in modeling various phenomena (e.g., heat conduction, electrostatics), namely u,xx +u,yy = f for which a = c = 1 and b = 0. ; f = f (x, y ) ,

The equation of transient linear heat conduction in one dimension, ku,xx = u,t ; k = k (x) ,

where a = k and b = c = 0, is a representative example of a parabolic equation. (c) Hyperbolic equations (b2 4ac > 0) The one-dimensional linear wave equation,

where a = 2 , b = 0 and c = 1, falls in this class of equations.

DR
the dierential equation exhibit discontinuities. by the nite element method.
ME280A

Extension of the above classication to more general types of partial dierential equa-

tions than those of the form (1.17) is not always an easy task. The elliptic, hyperbolic or parabolic nature of a partial dierential equation is associated with the particular form of its characteristic curves. These are curves along which certain derivatives of a solution to The type of a partial dierential equation determines the overall character of the expected

solution. Broadly speaking, elliptic dierential equations exhibit solutions which are as smooth as its coecients allow. On the other hand, the solutions to parabolic dierential equations tend to smooth out any initial discontinuities, while the solutions to hyperbolic partial dierential equations preserve any initial discontinuities. To a great extent, the type of the partial equation dictates the choice of methodology used in its numerical approximation

AF
2 u,xx u,tt = 0 ,

(b) Parabolic equations (b2 4ac = 0)

Version: August 31, 2010, 13:41

Suggestions for further reading Remarks:

dimensional convection-diusion equation of the form u,t + u,x = u,xx ;

The above equation is of hyperbolic type if = 0 and > 0 (i.e., when the diusive term is suppressed), since

implies that its solution satises the previously mentioned wave equation. On the other hand, for > 0 and = 0 the convective part vanishes and the equation is purely parabolic and coincides with the previously mentioned one-dimensional transient heat conduction equation. The dominant character in the convection-diusion equation is controlled by the relative values of parameters and .

 The type of a partial dierential equation may be spatially dependent, as with the following example: u,xx + xu,yy = 0 ,

where a = 1, b = 0 and c = x, so that the equation is elliptic for x > 0, parabolic for

DR
1.4
Section 1.1 [1] C.A. Felippa. Version: August 31, 2010, 13:41

x = 0 and hyperbolic for x < 0.

Suggestions for further reading

solution of problems of equilibrium and vibrations, 1943. Int. J. Num. Meth. Engr., 37:21592187, 1994. [This reference contains the original article on the nite

element method by Courant, preceded by an interesting introduction by C. Felippa.]

[2] R.W. Clough, Jr. The nite element method after twenty-ve years: A personal view. Comp. Struct, 12:361370, 1980. [This reference oers a unique view of the
nite element method by one of its inventors]. ME280A

AF
= (u,x ),t = (u,x ),t = (u,t ),t = u,tt

2 u,xx = (u,x ),x = (u,t ),x

An appreciation of R. Courants Variational methods for the

T
0 , 0.

 Partial dierential equations of mixed type are possible, such as the classical one-

10

Introduction [3] P.G. Ciarlet and J.L. Lions, editors. Finite Element Methods (Part 1), volume II
article in this handbook presents a comprehensive introduction to the history of the nite element method, authored by J.T. Oden].

Section 1.2

[1] O.C. Zienkiewicz and R.L. Taylor. The Finite Element Method; Basic Formulation and Linear Problems, volume 1. McGraw-Hill, London, 4th edition, 1989. [Chapter 1 of this book is devoted to an introductory discussion of discretization]. [2] G. Kron. Numerical solutions of ordinary and partial dierential equations by means of equivalent circuits.
of dierential equations].

interesting use of an electrical circuits analogue method to obtain approximate solutions

Section 1.3

[1] F. John. Partial Dierential Equations. Springer-Verlag, New York, 4th edition, 1985. [Chapter 2 contains a mathematical discussion of the classication of linear
second-order partial dierential equations in connection with their characteristics].

DR
ME280A

AF

J. Appl. Phys., 16:172186, 1945. [This is an

of Handbook of Numerical Analysis. North-Holland, Amsterdam, 1991. [The rst

Version: August 31, 2010, 13:41