Introduction to the Finite Element Method—Theory,

Programming, and Applications

Chapter 1 Introduction

The finite element method is a technique for obtaining approximate solutions to

boundary value problems. To introduce the method, we present an elementary
example of the deflection of tightly stretched wire under a distributed load. This
example is sufficient to refresh the reader on some aspects of differential equations
that will prove important for understanding approximate solutions, introduce the
concept of approximate solutions, and actually define and illustrate the finite element
method. This particular application was selected because the solution is easily
visualized. However, any number of other applications could have been used, such as
one-dimensional problems in heat transfer, porous flow, or electrostatics. The
variables given here can be interpreted to correspond to one of these applications, and
reader is encouraged to do so if that is helpful.

1.1 Governing Equation and an Exact Solution

Consider a tightly stretched wire as shown in Figure 1.1. Under the proper
circumstances, the deflection of the wire is accurately described by the solution of
d2y
T + w( x) = 0.0
dx 2
where

T = tension in the wire [F]

y = deflection of the wire [L]2
w = distributed load [F/L]

The bracketed terms are the dimensions of the variables, where F representes force
and L represents length.
The proper circumstances mentioned above are T should be large enough to ensure
that T remains nearly constant during the information and that
2
 dy  dy
  
 dx  dx

Before we obtain an exact solution, it is necessary to define a particular loading. The

loading selected is one that does not complicate the mathematics beyond that needed
to fulfill the purposes stated above, and is shown in Figure 1.2.

The governing equation now becomes

d2y
T − 3W = 0 for 0  x  L/ 2
dx 2

d2y
T − W = 0 for L/ 2  x  L
dx 2
with boundary conditions
y(0) = y(L) = 0
The loading is defined at all points but is discontinuous at x = L/2. The governing
equations tells us that we should expect the same for the second derivative of y with
respect to x. However, the first derivative must be continuous at x = L/2; otherwise the
second derivative would not be defined at this point.
There are several approaches avilable to arrive at the exact solution. The one chosen
here is to obtain the general solution for each of the two regions specified in Eq. 1.3,
and require each one to have the same value for deflection and the same value for
slope at x = L/2.
The general solutions for the two segments are
W 3 2 
y( x) =  x + C1 x + C 2  for 0  x  L/ 2
T 2 

W 1 2 
y( x) =  x + C3 x + C 4  for L/ 2  x  L
T 2 
The four constants of integration can be evaluated using the four conditions
y(x = 0) = 0

y(x = L) = 0

 L   L 
y x = −  = y x = +
 2   2 

dy  L  dy  L 
 x = −  =  x = +
dx  2  dx  2 
Evaluation of these constants gives us
WL2 3 x
2
5  x 
y( x) =    −   for 0  x  L/ 2
T  2  L 4  L 

WL2  1  x 2 1  x  1 
y( x) =    −   −  for L/ 2  x  L
T  2  L  4  L  4

which is the exact solution to our governing equation. The plot of this function is
shown in Figure1.3 at the end of the section.

1.2 Approximations to the Exact Solution

We now consider how we would arrive at an approximate solution to our governing

equation if the exact solution were not obtainable. Almost all methods for doing so
use an approximating function with a finite number of degrees of freedom, such as
 
y(x) = x( x − L) A0 + A1 x + A2 x 2 + A3 x3 
or
 x   2x   3x 
y( x) = A0 sin  + A1 sin  + A2 sin 
L  L   L 
Here the degrees of freedom (the number of undetermined parameters) equal 4 in
Eq.1.8 and 3 in Eq.1.9. Note that both approximations satisfy the boundary conditions
independent of the parametric values. This is an important requirement for selecting
approximating functions. Also worth noting is that the approximations can be made to
converge to any continuous function by extending the series indicated. Such series are
referred to as being complete or having the property of completeness, a topic covered
in the mathematics of infinite series.
There are other series we could consider, including a finite element approximation.
However, before we get to that, it is instructive to illustrate some of the methods used
to obtain values for the parameters in these series. To do so, we will use the truncated
polynomial series
y( x) = x( x − L)( A0 + A1 x)

Note that the dimensions of A0 and A1 must be [L-1] and [L-2], respectively. It is often
best to write a series in a way that makes the parameters nondimensional, or at least in
a way that gives them the same dimensions. However, to simplify the notation while
retaining the physical dimensions of all quantities, the above form was chosen.
The exact solution satisfies the differential equation for all values of x. However, there
are no values of A0 and A1 that would make our approximation capable of satisfying
our equation at all points. This is easily seen by noting that the exact solution has a
discontinuity in its second derivative, whereas all derivatives of our approximating
function are continuous. Because there are no values for our parameters that will give
us the exact solution, we must determine which values would be “best”. This would
not be a difficult task if we knew the exact solution. However, we are pretending not
to know this, but only the differential equation that it must satisfy. We need, therefore,
some technique for determining the best values for the parameters based only on the
governing differential equation.
There are several ways to determine, or define, these “best” values, most of which are
related to a function called the residual. This is the function obtained when the
approximating function is substituted into the governing equation. For our case, this
function is
R( x, A0 , A1 ) = 2TA0 + T(6 x − 2 L)A1 − 3W for 0  x  L/ 2

R( x, A0 , A1 ) = 2TA0 + T(6 x − 2 L)A1 − W for L/ 2  x  L

where we have indicated that the residual depends not only on x, but also on the
values of the yet undetermined parameters A0 and A1 such that the residual would be
zero for all values of x, then we would have the exact solution. However, as we have
already noted, this is not possible. Hence, a compromise must be made. We now
consider several well-known compromises.

1.2.1 Collocation
If we connot make R = 0 at all points between 0 and L, we will make it zero at as
many points as possible. Because we have only two undetermined parameters, we
expect at most to be able to satisfy the equation at two points. We choose x = L/4 and
x = 3L/4 as logical choices, and obtain
L 3W
2 A0 − A1 =
2 T

5L W
2 A0 + A1 =
2 T
which gives us
4W
A0 =
3T

2W
A1 = −
3LT
Hence, our approximate solution by collocation is
WL  2  x  4  x 
3 2
y x
= −   + 2  −  
L T  3  L  L 3  L 
 
Figure 1.3 at the end of this section, compares this solution with the exact solution as
well as the other approximate solutions that follow. The error between it and the exact
solution, is shown in Figure 1.4. Note that the maximum error is approximately 0.4%.

Two things are worth noting. First, the results are good because our approximation
was good—a cubic that was able to duplicate our exact solution very closely. That
might not be true for a more complex loading. In such a case, we would need to add
more terms (parameters) in our approximating function. Second, our solution matches
the exact solution at x = L/4 and x = 3L/4 (not easily seen from the graph, but it does).
This is not because these points were the collocation points; rather, it is more of an
accident peculiar to this particular problem. In general this will not be the case. We
satisfy differential equation at collocation points, not the solution.

1.2.2 Least Squares

Rather than making R(x, A0, A1) = 0 at two points, we might want to minimized its
magnitude in some average sense along the entire length from 0 to L. One way of
doing this is to minimize
L
J =  0
R 2 dx
Clearly, J is greater than zero except for the case where R is everywhere zero, i.e., the
exact solution. Therefore, of all possible functions that satisfy our boundary
conditions, the exact solution produces the lowest value of J. We know, because our
approximation does not contain the exact solution, that it cannot produce this
minimum value; thus, we settle for as low a value of J as it is able to produce. That is,
we find values of A0 and A1 that give J a stationary value. Thus, we solve
J J
= 0 and = 0
A0 A1
for A0 and A1.
Rather than integrate and then differentiate, it is easier to differentiate and then
integrate; therefore, we write
 L L R
Ai  0
R 2 dx =  0
2R
Ai
dx = 0

Equations 1.11 give us R, from which we obtain

R
= 2T for 0  x  L
A0

R
= T(6 x − 2L) for 0  x  L
A1
From this we obtain the following two equations:
J L/ 2

A0
= 0
[2T ][2TA0 + T(6 x − 2L)A1 − 3W ]dx

L
+  L/ 2
[2T ][2TA0 + T(6 x − 2 L)A1 − W ]dx = 0

J L/ 2

A1
= 0
[6T(6 x − 2L)][2TA0 + T(6 x − 2L)A1 + 3W ]dx

L
+  L/ 2
[6T(6 x − 2 L)][2TA0 + T(6 x − 2 L)A1 + W ]dx = 0
which gives, after integration and the solution of the resulting algebraic equations,
5W
A0 =
4T

W
A1 = −
2LT
Hence, our approximation solution found by at least squares method is
WL  1  x  5  x 
3 2
y 7 x
= −   +   −  
L T  2  L 4  L 4  L 
 
The solution is shown in Figure 1.3 and its error in Figure 1.4. It has a slightly higher
maximum error than the function found by collocation, but is still less than 1%.

1.2.3 Galerkin’s Method.

Another method for reducing the residual over the entire length is to make its
weighted average zero with respect to as many independent weighting functions as
there are undetermined parameters. That is,
L
0
W0 R( x, A0 , A1 , , An )dx = 0
L
0
W1 R( x, A0 , A1 , , An )dx = 0

L
0
Wn R( x, A0 , A1 , , An )dx = 0

The weighting functions must be independent in order for the resulting algebraic
equations to be independent. Galerkin’s method uses the independent terms of the
approximating function as the weighting functions; thus, there are always as many
weighting functions as there are independent parameters. For our example, the
method gives us
L/2
0
[ x 2 − Lx][2TA0 + T(6 x − 2 L)A1 + 3W ]dx
L
+ 
L/2
[ x 2 − Lx][2TA0 + T(6 x − 2 L)A1 + W ]dx = 0
 L/2
0
[ x 3 − Lx 2 ][2TA0 + T(6 x − 2 L)A1 + 3W ]dx
L
+ L/2
[ x 3 − Lx 2 ][2TA0 + T(6 x − 2 L)A1 + W ]dx = 0

Integration of these equations produces two algebraic equations in A0 and A1, the
solution of which gives
21W
A0 =
16T

5W
A1 = −
8TL
Thus, the approximation solution by Galerkin’s method is
WL  5  x  21  x 
3 2
 y 31  x 
  = −   +   −  
 L T  8 L 16  L  16  L 
 

1.2.4 The Ritz Method.

For our tight wire problem it is possible to write an expression for the total potential
energy of the system as a function of deflection, y(x). We know from the principles of
mechanics that this scalar function attains a minimum value corresponding to the
equilibrium position of the wire. That is, if the potential energy of the system is
calculated using the exact solution of the governing differential equation, its value
will be lower than that calculated when any other deflected shape is used for the
calculation. Thus, as we did for the least squares method, we will judge the best
values for A0 and A1 as those that give the smallest potential eneygy to the system, i.e.,
those values that produce a potential energy nearest to the potential energy that the
exact solution has.
The potential energy of the system is given by
 1  dy  2
L 
E =   T   − wydx
0
 2  dx  
for any loading w=w(x) and any deflection y=y(x). Similar expressions exist for other
applications. We will shortly show that the minimization of such an expression is
equivalent to satisfying the governing differential equation.
When an approximating function is substituted into the above integral, E becomes a
function of the undetermined parameters used for the function; thus, to obtain a
stationary value, we enforce
E  1
L 
Ai
=
Ai 0 
2
T( y)2 − wy dx

L  1 
= 
0 
Ai  2
T( y)2 − wy  dx

L  ( y) ( y )
= 
0  yT
 A i
− w
A
dx
i 

For our two-parameter approximating function,

y = ( x 2 − xL)A0 + ( x3 − x 2 L)A1

y = (2x − L)A0 + (3x 2 − 2xL)A1

( y)
= ( x 2 − xL)
A0

( y)
= ( x 3 − x 2 L)
A1

( y)
= (2 x − L)
A0

( y)
= (3x 2 − 2 xL)
A1

− 3W for 0  x  L/ 2
w = 
 − W for L/2  x  L
Substitution of these functions into Eq. 1.29 creates, as with the previous methods,
two linear algebraic equations for A0 and A1. Solving, we find that the resulting values
of our two parameters are identical to the ones we obtained by Galerkin’s method.
There is a reason for this, which will soon be explained. The plots of this function and
its error are the same a sthe plots for Galerkin’s method in Figs. 1.3 and 1.4.
For many of the boundary value problems associated with engineering analysis, there
is a corresponding scalar function such as we used here. These scalar functions of
functions are referred to as functionals. The procedures for determining the functional
corresponding to a given differential equation, or vice versa, are the subject of the
calculus of variations. The principle that the solution of a given boundary value
problem corresponds to a stationary value of a functional is referred to as a variational
principle. Finally, the concept of using a variational principle to judge the quality of
an aprroximate solution is referred to as the itz method. We will spend considerable
time exploring these ideas because they represent the foundations of the finite element
method.
1.2.5 Results.
In comparing the preceding methods, it is important not to place emphasis on which
one has the smallest error. The errors shown depend more on the particular problem
and the approximating function selected than on the method. All of the methods
converge to the exact solution as the number of terms used in the approximation
increase, provided that the series is complete. The question of convergence is
important and will be discussed at the conclusion of this chapter.
For now, however, let us see how each of the methods fared. Figure 1.3 illustrates the
deflection found for each method (the Ritz method is the same as Galerkin’s).
Figure 1.4 illustrates the difference between the exact solution and the approximate
solutions. Again, you should understand that the closeness of these approximations to
the exact solution is due to the fact that our approximating function, with only two
independent parameters, can be made very close to the piecewise parabola that is the
exact solution. In general, many more terms are necessary to obtain the accuracy
needed in most engineering analyses.

1.3 Finite Element Approximations

There is a very important difference between the Ritz method and the other three
methods illustrateed in the previous section, and that difference allows us to introduce
finite element approximations. Whereas the first three methods required that the
second derivative of the approximating function exist, the Ritz method required only
the existence of the first derivative. This weaker demand allows us to use piecewise
linear functions as our approximating function. Such approximations are foundations
of the finite element method, several of which are illustrated in Figure 1.5. You should
note three characteristics of these functions, which we will discuss in much more
detail in later chapters:

1. The functions are continuous, although their first derivatives (slopes) are not.
2. Each function is completely defined by its nodal point values; hence, these values
serve as the undetermined parameters of the approximating function.
3. The segments, or elements, need not be of equal length.
Figure 1.5. Finite element approximations.

Let us now use one such function to approximate he deflection of our tight wire. We
choose the particular one illustrated in Figure 1.6, given by
(3Y2 /L)x for 0  x  L/3

y( x) = (2Y2 − Y3 ) + (3/L)(Y3 − Y2 )x for L/ 3  x  2 L/3
 3Y − (3Y /L)x for 2 L/ 3  x  L
 3 3

The energy functional,

  dy  
2

E =  − wy + 0.5T   dx
L

0
  dx  
Requires the use of the derivative of this function, which is
 3Y2 /L for 0  x  L/ 3
dy 
= (3/L)(Y3 − Y2 ) for L/ 3  x  2 L/3
dx − 3Y /L
 3 for 2 L/ 3  x  L
Substitution of the function, its derivative, and the loading function into Eq. 1.32
gives us the potential energy as a function of the deflections, Y2 and Y3, which we
indicate by writing
E = E(Y2 ,Y3 )
Figure 1.6. Finite element approximation to the deflection of the tight wire.

The Ritz method states that the best approximation to the exact solution that these two
parameters can give are the two values that create the lowest value for E, that is, the
value of E that is cloasest to the value obtained by the exact solution. Thus, we seek
the two nodal values that will satisfy
E E
= 0 and = 0
Y2 Y3
As before, it would be easier to differentiate and then integrate, but this time we
choose to integrate first and obtain the potential energy in terms of our two nodal
values for displacement. This approach emphasizes that the potential energy is
quadratic function of the displacements; hence, it has one, and only one stationary
value.
A significant consequence of piecewise approximation is that the integration can be
performed element by element and summed to obtain the integral. This is the
approach we now use. The contribution to E from element A is
1 2
E A = −
L/ 3
− 3W[(3Y /L)x]dx + 
2
L/ 3
 2 T[3Y2 /2] dx
0 0
 

WL   3T 
EA =  Y2 +  Y22
 2   2L 
The contribution to E from element B is
EB = −  − 3W[(2Y − Y ) + (3/L)(Y − Y )x]dx
L/ 2

L/ 3
2 3 3 2

−  − W[(2Y − Y ) + (3/L)(Y − Y )x]dx

2 L/ 3
2 3 3 2
L/ 2

2 L/ 3 1 2
+ L/ 2  2 T[(3/L)(Y3 − Y2 )] dx
 
 5WL   3WL   3T 
EB =  Y2 +  Y3 +  (Y3 − Y2 )2
 12   12   2L 
The contribution to E from element C is
1 2
EC = − 
L
− W[3Y 3 − (3Y3 /L)x]dx + 
L/ 3
 2 T[−3Y3 /L] dx
2 L/ 3 0
 

WL   3T 
EC =  Y3 +  Y32
 6   2L 
The total energy is
E = E A + EB + EC

11WL   5WL   3T 
E =  Y2 +  Y3 +  (Y22 − Y2Y3 + Y32 )
 12   12   L 
which is quadratic function on Y2 and Y3. It thus has a single, stationary value where
E
= 0.0
Y2
and
E
= 0.0
Y3
Thus, we obtain
 11WL   6T   3T 
  +  Y2 −  Y3 = 0.0
 12   L   L 

 5WL   3T   6T 
  −  Y2 +  Y3 = 0.0
 12   L   L 
or, in matrix notation,
  6T   − 3T   11WL 
     −
  L   L  2 = 
Y  12 
  
 − 3T   6T   Y3   − 5WL 
 L      12 
   L  
The solution is
  WL2 
Y2   − 
= 4T 
   2
Y3  − 7WL 
 36T 

The solution is shown in Figure 1.7. The nodal point values in this case turn out to
coincide with the exact solution. However, this will not happen under different
circunstances (i.e., a different governing equation).

Figure 1.7. Finite element solution.

You have probably noted that this solution does not compare well with those found by
our two-term polynomial approximation. However, as more terms are added, the
polynomial and trigonometric approximations become ever more difficult to evaluate,
whereas the finite element solution simply repeats the same calculations, over and
over, for each element. This advantage becomes pronounced for partial differential
equations with irregular boundaries and nonconstant parameters.

1.4 Basis Functions

We have referred to our approximating functions as series. This is, of course, obvious
for the polynomial and trigonometric approximations; however, finite element
approximations do not, at first glance, appear to be a truncated series. Nevertheless, in
much of the literature associated with the finite element method the approximation is
treated as a series and written in the form
y( x) =  M (x)Y
i i

When this is done, the functions Mi(x) are called basis functions, with one such
function associated with each node. A set of these function for a four-node
approximation is shown in Figure 1.8.

Figure 1.8. Basis functions for a four-node approximation.

Note that the magnitude of each function at its respective node is unity and that is zero
at all other nodes. Hence, their sum would have a value of unity at each node, and the
function would simply be the straight line y = 1.0. If each basis function is multiplied
by a constant Ai, then the sum of these new functions would be a piecewise linear
function having a value at each node equal to the corresponding value of Ai for that
node. As an example, if
A1 = 0, A2 = 4, A3 = 3, A4 = −1

then
y( x) =  A M (x)
i i

Is the function shown in Figure 1.9.

Figure 1.9. ∑ MiYi for {Y} = {0, 4, 3, -1}.

1.5 Weak Formulation of Problem

As we have seen, the use of a piecewise linear approximation depends on being able
to express our problem in terms of the potential energy funcional, Eq. 1.28. We also
saw that for our polynomial approximation, the Ritz method produced the same
results as Galerkin’s method, and we noted that this was no accident. We now show
how to obtain the integral formulation from the differential equation—that is, how to
obtain the weaker formulation that requires only the existence of the first derivative
from the stronger formulation that requires the existence of the second derivative.
Both formulations are valid, and the solution to one is the solution to the other. Note
specifically that what follows is not directly relateed to approximation theory,
although we will certainly use the results for that pirpose. That is, the original
differential equation for our tight wire problem and the expression for the potential
energy of the wire are physical statements of the physics of the problem and have no
direct correlation with approximation theory.
Consider the original differential equation for the deflection of the tight wire,
d2y
T + w(x) = 0
dx 2
Because this equation must be satisfied at all points (i.e., must be zero at all points),
 d2y 
B( x)T 2
+ w( x) = 0
 dx 
must also be satisfied (i.e., equal to zero) for any function B(x). If so, then its integral
must likewise equal zero,
L  d2y 
0
B( x)T
 dx
2
+ w( x)dx = 0

Interpret this equation as a test that the trial function, y(x), must pass for all
permissible test functions, B(x), if it is the exact solution.
We now assume that the test functions, B(x), are sufficiently well behaved that the
integrand remains integrable. Thus, we integrate by parts to obtain
L d  dy  dB dy 
0

 dx
 BT

 − T
dx  dx dx
+ Bwdx = 0

L
dy  L  dB dy 
BT
dx  0
− 
0
 T
 dx dx
− Bw dx = 0

The integrate first term, in later problems, will allow us to specify boundary
conditions other than what we have. For now, however, we assume the originl
boundary conditions of our tight wire problem, i.e., y(x) known at both x = 0 and x = L,
and let Eq. 1.55 apply only to functions that meet these conditions. This being the
case, there is no need to test y(x) at these points, and we restrict our test functions to
those that are zero at x = 0 and x = L. Thus, for our specific problem,
L  dB dy 
0
 T
 dx dx
− Bw dx = 0.0

Note that Eq. 1.56 applies to the y(x) that satisfies our boundary conditions at x = 0
and x = L, whether y(x) is specified as zero or some other value at these points. On the
other hand, the test functions B(x) must be zero at these points; otherwise we would
have to include the first term in Eq. 1.55, which requires the value of dy/dx. This we
ordinary do not know when y is known. More on this point in the next chapter.