Sanjay FEM

CE222
Finite Element Methods

Course Notes
Sanjay Govindjee 1
Structural Engineering, Mechanics, and Materials
Department of Civil and Environmental Engineering
University of California, Berkeley
c
Copyright
by Sanjay Govindjee 2002, all rights reserved.
Associate Professor of Civil Engineering, University of California Berkeley, govindjee@ce.berkeley.edu
S. Govindjee
Preface
These notes are intended to serve as supplementary reading for an introductory graduate
level course in finite element methods. They are not intended to be a replacement for a
proper textbook in the subject. The goal here is to bring together a variety of viewpoints
that appear in several textbooks and research articles and present them in a more tutorial
manner than they normally appear for a selected number of issues. Thus, many important
topics to understanding and using finite element methods are not covered and the reader
is directed to comprehensive texts such as The Finite Element Method vol. I, II, and III
by O.C. Zienkiewicz and R.L. Taylor, The Finite Element Method by T.J.R. Hughes, Finite
Element Procedures by K.-J. Bathe, and An Introduction to the Finite Element Method by
J.N. Reddy. For a classical introduction from a structural engineering viewpoint the reader
is directed to Concepts and Applications of Finite Element Analysis by R.D. Cook, D.S.
Malkus, and M.E. Plesha. A very accessible undergraduate level introduction can also be
found in Introduction to the Finite Element Method by N. Ottosen and H. Petersson. For a
more mathematically rigorous introduction the reader is directed to the classical monograph
An Analysis of the Finite Element Method by G. Strang and G.J. Fix and Numerical Solutions
of Partial Differential Equations by the Finite Element Method by C. Johnson.
Draft Notes CE222 Spring 2002
Contents
1 The Basics of Problem Statements
2 FEM Approximation
3 Properties of the Solution uh
11
4 Convergence
12
5 Geometric and Energetic Interpretations of the Solution
13
6 Error Estimate
16
7 A Worked Example
17
8 Element View Point
22
9 Higher Order Elements
26
10 Calculations Over Standard Elements/Isoparametric Concepts
29
11 A 1-D Quadratic Isoparametric Element
32
12 Quadrature
34
13 Quadrature and Convergence
37
14 Introduction to Multi-Dimensional Problems
39
15 Isoparametric Triangles
43
16 Summary of Elasticity Equations
47
17 Strong, Weak, and Minimization Forms of Elasticity
48
18 Finite Element Approximations for Elasticity
51
19 Consistent Nodal Loads
54
20 Isoparametric Quadrilaterals
58
21 Mesh Refinement with Higher Order Elements
62
22 Stress Recovery Methods
65
23 Introduction to the Patch Test
68
24 Element Performance in Bending
72
S. Govindjee
25 Incompatible Modes
75
26 Incompressibility
79
27 Axis-symmetry
84
28 Introduction to Shear Deformable Plates and Beams

28.1 Plate Weak Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
87
91
29 Shell Modeling with Plates

29.1 Assemblages of Flat Plates . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29.2 Finite Element Arrays for Assemblages of Flat Plates . . . . . . . . . . . . .
95
95
97
30 Introduction to Dynamics
30.1 Weak Form Equations . . . . . . . . . .
30.2 Finite Element Approximation . . . . . .
30.3 Introduction to Numerical Integration by
30.4 A Scalar Example . . . . . . . . . . . . .
30.5 Convergence . . . . . . . . . . . . . . . .
30.6 Application to Heat Conduction . . . . .
30.7 Application to Elastodynamics . . . . . .
. . . . . . . . . .
. . . . . . . . . .
Finite Differences
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
99
100
101
103
103
105
110
113
The Basics of Problem Statements
The finite element method is a way to approximately solve boundary value problems (BVP).
The strong form of a boundary value problem is made up of an ordinary differential equation
(ODE) or partial differential equation (PDE) and a set of boundary conditions. The strong
form of a BVP is the most common form; however, the finite element method is based on
what is called the weak form of the BVP. A third form of a BVP is a minimization form.
Not all BVPs have minimization forms, but all BVPs have weak forms. BVPs dealing with
elastic bodies in mechanics have minimization forms.
To make the presentation concrete let us consider the problem of a 1-D elastic rod with
both force and displacement boundary conditions. The rod is subject to a prescribed displacement of u at the left end, x = 0, a prescribed force F at the right end, x = L, and
a prescribed body force b(x) per unit length of the rod. With these prescribed loads we
want to find the displacement at each point in the rod. Knowing that, we will also know
the strains and hence the stresses. Assume for simplicity that the cross-sectional area of the
rod, A, and the Youngs modulus, E, are both constants.
L
u
F
x
The strong form of
(S)
b(x)
the BVP says:

Find the function u(x) such that
AEu,xx + b
= 0
for x (0, L)
u(0)
= u
AEu,x (L)
= F
(1.1)
Certain approximation methods such as finite differences work directly with the strong form
of the BVP. However, we will need the weak form of the BVP. The weak form of the BVP is
found via a simple algebraic manipulation of the strong form plus an integration by parts.
In mechanics, the weak form corresponds to the virtual work equations.
To obtain the weak form, it proves useful to first define two different spaces or sets of
functions. The first set of functions is denoted by the symbol S. It is defined as the set of
all possible functions that still satisfy the displacement boundary conditions (the essential
or Dirichlet boundary conditions). It is written as
S = {u | u(0) = u}
(1.2)
The vertical bar, |, is read as such that. So the statement above is read as: The set of
all functions such that they are equal to u at x equals zero. Note that there are an infinite
number of functions in the set S. It will also prove useful to define the set of functions that
S. Govindjee
are equal to zero at the essential boundary conditions. This set is denoted by the symbol
V and is called the set of weighting functions, virtual displacements, variations, or trial
functions. It is written as
V = {v | v(0) = 0}
(1.3)
The statement above would be read as: The set of all functions such that they are zero when
evaluated at x = 0. Note that this corresponds to the concept in mechanics of an admissible virtual displacement. [Technically, there are certain differentiability and integrability
restrictions on these two sets. For the moment we will ignore them.]
The weak form can now be constructed. Pick an arbitrary weighting function v(x) from
the set V, i.e. v V. Multiply both sides of the ODE in the strong form by this function,
integrate both sides from x = 0 to x = L, then integrate by parts so that the number of
derivatives on v(x) is the same as the number of derivatives on u(x). The weak form of the
BVP then says:
Find the function u S such that for all v V

Z
Z L
L
(W )
v,x AEu,x dx =
vb dx + v(L)F
0
(1.4)
This form of the BVP is sometimes termed a variational form.

One should stop for a moment and consider the unusual form the BVP takes in the
weak form. An interpretation of it is that the solution to the BVP is found by choosing a
function in the set S, plugging it into the integral expression, and then testing whether the
integral expression is true for each and every function in the set V. Note that there are an
infinite number of functions to choose in the set S. Also, there are an infinite number of
functions in the set V. All the choices from S have to be tested against all the choices from
V in the integral expression in (W). The process continues until the element u S is found
that satisfies the integral expression in (W) for all the elements v V. Clearly, this is an
intractable task. The finite element method gives us a prescription for reducing this horribly
large search to a finite dimensional search.
Also note that by direct manipulation, it can be shown that the weak form (W) and
the strong form (S) are equivalent statements of the BVP; ie. the solution to (S) is also
the solution to (W) and the solution to (W) is also the solution to (S). By working from
the strong form to derive the weak form we have already shown that solutions of (S) satisfy
(W). To show that a solution of (W) satisfies (S), we need to show that the equations that
comprimise the strong form can be derived as consequences of the weak form. Starting with
the left-hand side of (1.4)
Z
0
(vAEu,x ),x vAEu,xx dx

L Z L

vAEu,xx dx
= vAEu,x
0
0
Z L
vAEu,xx dx
= v(L)AEu,x (L)
v,x AEu,x dx =
(1.5)
(1.6)
(1.7)
In the above manipulations, use has been made of the fact that v(0) = 0. Now take (1.7)
and combine it with the right-hand side of (1.4):
Z
v[AEu,xx + b] dx + v(L)F = v(L)AEu,x (L)
(1.8)
Since (1.8) must hold for any function in V, make the following choice:
v(x) = x(x L)[AEu,xx + b]
(1.9)
With this choice (1.8) becomes

Z
x(x L)[AEu,xx + b]2 dx = 0
(1.10)
Since x(x L) is always greater than zero in the interval (0, L) and [AEu,xx + b]2 0, it
follow that
AEu,xx + b = 0
(1.11)
Thus the ODE of the strong form problem statement is a consequence of the weak form
problem statement. The boundary conditions can also be shown to hold. First, the Dirichlet
boundary condition u(0) = u is automatically satisfied by construction of S. The Neumann
(or natural) boundary condition is seen to hold true by substituting (1.11) back into (1.8):
v(L)F = v(L)AEu,x (L)
(1.12)
After dividing both sides by v(L) the desired Neumann boundary condition is seen to hold.
S. Govindjee
FEM Approximation
The basic idea is to replace the infinitely large sets S and V by finite dimensional ones. The
approximation to S is denoted as S h . S h is a subset of S; ie. S h S. This is done so that
elements of S h are always elements of the original set that we were using to find the solution
to the BVP. The approximation to the set V is denoted by V h . Just as for S h , V h is chosen
to be a subset of the set it is approximating; ie. V h V. In the particular kind of finite
element methods we will use, elements uh S h are expressed in the particular form
uh = v h + g h
(2.1)
where v h V h and g h is a function equal to u at x = 0. Note that this form satisfies all
of our requirements for elements of the set S h . The superscript hs are to remind us that
our solution is associated with a discretization of our domain into elements of length h.
Pictorially, the relation between the real sets and the approximate sets is:
The finite element prescription continues by stating that any element of V h is given by a
linear combination of shape functions, such as the hat functions. Hence, we may write
vh =
n
X
NA (x)vA
(2.2)
A=1
where the NA (x)s are called shape functions and the vA s are constants. n is the dimensionality of the finite dimensional space V h . N1 (x) is the shape function associated with node
1; N2 (x) is the shape function associated with node 2; etc. It is a required property of the
shape functions that when one is evaluated at a node, it is zero except at the node with
which it is associated; ie.
NA (xB ) = 1 for A = B
NA (xB ) = 0 for A 6= B
Or more simply, NA (xB ) = AB , where AB is the Kroneker delta function.
For the hat functions this looks like:
(2.3)
(2.4)
9
g
N (X)
NA(X)
XA-1
XA
X1
XA+1
X2
X3
The elements of S h are fully specified by the definition for the elements of V h and the
definition that
g h = N g (x)
u
(2.5)
where N g (x) is the shape function associated with the node at which the essential boundary
condition is specified. Note that N g (0) = 1 and is zero at all other nodes; see the picture
above. Thus an element of S h is expressed as
h
u =
n
X
NA (x)uA + uN g (x)
(2.6)
A=1
Let us now examine the approximate statement of the weak form that the FEM tells us
to solve:
the function uh S h such that for all v h V h

Find
Z L
Z L
(2.7)
(W approximate)
h
h
v,x AEu,x dx =
v h b dx + v h (L)F
0
If we plug the expressions for uh and v h into the approximate weak form above, we get the
result that
" n Z
#

Z L
Z L
n
L
X
X
g
vA
NA,x AENB,x dx uB
NA b dx NA (L)F +
NA,x AEN,x u dx = 0
A=1
B=1
(2.8)
Since this equation is to hold for all elements of V , it must hold for arbitrary choices of the
constants vA . Hence the expression in the square brackets must be zero; ie.

Z L
Z L
n Z L
X
NA,x AEN,xg u dx = 0 (2.9)

NA,x AENB,x dx uB
NA b dx NA (L)F +
h
B=1
We now define an n by n matrix K with components

Z L
KAB =
NA,x AENB,x dx
(2.10)
a vector u with components uB and a vector F with components given by

Z L
Z L
FA =
NA b dx + NA (L)F
NA,x AEN,xg u dx
0
(2.11)
This allows us to rewrite our expression as

Ku = F
(2.12)
10
S. Govindjee
All the elements of K and F are known. u is called the vector of unknown nodal displacements; note that uh (xA ) = uA by our definition above. In structural mechanics K is called
the stiffness matrix and F is the load vector. Also observe that all we need to do is solve
this system of matrix equations to obtain the unknown values of the vector u. Once these
are known, we then have an expression for the solution to the approximate weak form of the
BVP.
There are several important properties of the matrix K.
1. K is positive definite. This means that we have a unique solution to our approximate
problem. Mathematically, positive definiteness is the condition that
u Ku > 0 for u 6= o
u Ku = 0 for u = o
(2.13)
(2.14)
Its importance here is that it assures us a unique solution for the vector u.
2. K is symmetric; ie. KAB = KBA . This is also expressed by saying that the transpose
of K is equal to K; ie. K T = K. One important consequence of this is that one only
needs to compute half of the matrix to know all the values.
3. For functions like the hat functions, K has a banded structure. In particular for the
hat functions shown above, KAB = 0 for |A B| > 1. The importance here is that
this allows us to use special storage and solution techniques when working with K in
the computer.
11
Properties of the Solution uh
Once we have an approximate solution, we can examine its character and try to assess its
quality. If we consider the hat functions, we can show the following facts:
1. Over a single element, uh is a linear function. If x is between nodes A and A + 1,
then only shape functions NA (x) and NA+1 (x) will be non-zero in the expression for
uh . Hence
uh = uA NA + uA+1 NA+1
(3.1)
If we expand the shape functions in this element, we find that
1
NA (x) = (x xA+1 )
h
1
NA+1 (x) =
(x xA )
h
(3.2)
(3.3)
where h = xA+1 xA is the length of the element. Plugging in above, we find that in
the element between nodes xA and xA+1

uA+1 uA
h
u (x) =
(x xA ) + uA
(3.4)
h
which is easily seen to be linear.
2. Because the displacements are linear functions in each element, the derivatives of the
displacements (the strains) are constants in each element. For the element between
nodes xA and xA+1 , this constant is equal to (uA+1 uA )/h.
3. uh is continuous over the whole rod but its derivative (the strain) is only piecewise
continuous. Sometimes this state of affairs is called C 0 continuity. When a function
and its first derivative are continuous, we say it is C 1 continuous. The pattern repeats
itself for higher derivatives; ie. a function that is C m continuous is one where the
functions and its first m derivatives are continuous.
12
S. Govindjee
Convergence
In order for this kind of approximate solution to approach the real solution as the size of
the elements is decreased (and consequently their number increased) we need to satisfy two
conditions.
Completeness: If the weak form of the problem contains derivatives of order m, then the
approximate solution over each element must be able to exactly represent solutions
with constant mth order derivatives and below.
Compatibility/Continuity: The approximate solution must be continuous of degree m1
where m is the order of differentiation present in the weak form; ie. uh must be at
least C m1 continuous over the whole domain and C m in element interiors.
The proof that these conditions will give us the property we want will not be given here.
If you are interested in more details, see the book by Strang and Fix [17]. In our example,
m = 1. So completeness requires uh to have both linear and constant terms when expressed
over each element. Compatibility requires that uh be at least C 0 continuous. Note both of
these conditions are satisfied by the hat functions.
13
Geometric and Energetic Interpretations of the Solution
Further assessment of the approximate solution uh sheds more light on how and why the
method works. If u is the real solution then
Z L
Z L
h
v h b dx v h (L)F = 0
(5.1)
v,x AEu,x dx
0
Note v h V h , but since V h V, it is also a member of V. Therefore, by (W) the above

expression is seen to be true for all v h V h . We also have that the approximate solution uh
satisfies
Z L
Z L
h
h
v AEu dx
v h b dx v h (L)F = 0
(5.2)
,x
,x
0
h
for all v V by construction. We can get a feel for the error e = u uh by subtracting
the two expressions above. This yields
Z L
h
v,x
AE(u uh ),x dx = 0
(5.3)
0
for all v h V h . To interpret this expression, first define the functional

Z L
a(f, g) =
f,x AEg,x dx
(5.4)
where f and g are functions of x. [A functional is just a function that takes functions for
arguments.] This allows us to write our expression as
a(v h , e) = 0 for all v h V h
(5.5)
The functional a(, ) can be interpreted as a kind of dot product operation between functions.
If the dot product between two vectors is zero, we say they are orthogonal (perpendicular).
In the Euclidean plane, vectors a and b are orthogonal when ab = 0 as shown in the picture
below. When they are not orthogonal, the dot product is non-zero.
y
y
b
b
a .b 0
a .b = 0
x
For our present case, we say that the error in our approximation is orthogonal to the set
(space) of functions V h . Pictorially, for zero essential boundary conditions (
u = 0), we can
draw the picture below
14
S. Govindjee
uh
where the set V h is represented by the line and the S by the plane. Each vector in the plane
represents a function in S and each vector lying along the line represents a function in V h .
Note that V h = S h because of the assumption in the drawing that u = 0.
A more intuitive physical consequence can be derived from this result. First note that
a(u, u) =
u,x AEu,x dx
(5.6)
is merely 2 times the strain energy stored in a bar with a displacement field u. Consider
uh S h to be our finite element solution and e = u uh to be the error from the exact
solution u. Pick any other element from S h and call it U h . Note that the error u U h can
be expressed in terms of the error e as follows
u U h = u (uh + wh ) = e wh
(5.7)
where wh is an element of V h . Now consider (twice) the strain energy of the error associated
with U h .
a(u U h , u U h ) = a(e wh , e wh ) = a(e, e) a(e, wh ) a(wh , e) + a(wh , wh )
(5.8)
The middle two terms are zero because of the orthogonality property shown above. The last
term is always non-negative by the definition of a(, ). Hence
a(u U h , u U h ) a(e, e)
(5.9)
The interpretation then is that the strain energy of the error in the finite element solution is
a minimum in comparison to the strain energy of the error associated with any other member
of S h . This property is known as the best approximation property.
In a similar fashion, for zero essential boundary conditions, one can easily show that
a(u, u) a(uh , uh )
(5.10)
In words, this says that the strain energy of the exact solution is always greater than the
strain energy of the finite element solution. In the same spirit, we note that the total potential
15
energy of the whole system is minimized by the exact solution u; therefore the approximate
solution uh represents a higher total potential energy. Thus
where
(w) =
Z
0
(uh ) (u)
(5.11)
1
2
AEw,x
bw dx w(L)F
2
(5.12)
These last results tell us that the finite element solution will appear stiffer than the real
solution; ie. that the real strains will be larger (energetically speaking) than the approximate
ones.
16
S. Govindjee
Error Estimate
Another important consideration is how fast will the error between uh and u decrease as the
number of elements is increased. A heuristic estimate can be made as follows: Consider the
exact solution u(x) over the element between nodes xA and xA+1 and let the length of the
element be h. The Taylor series with remainder for the exact solution expanded about the
point xA is
1
(6.1)
u(x) = u(xA ) + u,x (xA )[x xA ] + u,xx ()[x xA ]2
2
where is some point in the interval (xA , x). Consider any element U h S h (which is
linearly complete). This means that over the element it can be expressed as
U h = 1 + 2 x
(6.2)
for some constants 1 and 2 . Pick the U h for which the constants 1 and 2 exactly match
those in the Taylor series expression. Doing so and subtracting the two expressions gives
1
E = u U h = u,xx ()[x xA ]2
2
(6.3)
Since the finite element solution uh has been shown to be better than any other possible
solution in S h , we have the following estimate:
|e|max |E|max Ch2
(6.4)
where C is a constant that depends on the size of the element and the exact solution. The
main point is that the error in the displacement decreases with the square of the element
size. Hence, by halving the element sizes, we can reduce the error by a factor of 4. A
similar analysis for the derivative of the displacement shows that the error in the strains is
where C is another constant that depends on the same factors as C.
given by |e,x | Ch,
It is important to observe that the error in the strains and hence the stresses (through the
constitutive relation) decreases only linearly with a reduction in element size. For elements
with higher degrees of completeness, the errors decrease with higher powers of h. The error in
the strains, however, always decreases at the rate of one less than that of the displacements.
17
A Worked Example
Consider the axial rod problem shown below (which is the mechanical equivalent of Example
1 on page 164 of Ottosen and Peterson).
6 in
150 lbs
A = 10 in 2
6
E = 5 x 10 psi
b(x) = 100 lbs/in
The strong form for this problem reads:
Find the function u(x) such that
50 106 u,xx + 100 = 0

for x (0, 6)
(S)
u(0)
=
0
50 106 u,x (6)

= 150
(7.1)
The weak form of this problem reads:
the function u S = {u | u(0) = 0} such that for all v V = {v | v(0) = 0}

Find
Z 6
Z 6
(W )
6
v,x 50 10 u,x dx =
v100 dx v(6)150
0
(7.2)
The approximate weak form problem is found by approximating the spaces S and V, which
h
h
h
h
h
in
P this case are identical. The approximation is given by S = V = {w | w (0) = 0, w =
NA (x)wA }. Thus the approximate weak form problem reads:
the function uh S h such that for all v h V h

Find
Z 6
Z 6
(W approximate)
(7.3)
h
6 h
v,x 50 10 u,x dx =
v h 100 dx v h (6)150
0
The first thing to do is decide on a set of shape functions and level of discretization. For
this example, choose hat functions and discretize the domain into 3 elements with the node
numbering scheme shown below. Well let the node associated with the essential boundary
condition be called node 1. Thus, N g (x) will be written as N1 (x).
2 in
Remarks:
2 in
2 in
18
S. Govindjee
1. For our case,
NA (x) =
xxA1
xA xA1
xxA+1
xA xA+1
xA1 < x < xA

xA < x < xA+1
and NA,x (x) =
1
2
xA1 < x < xA

12 xA < x < xA+1
(7.4)
2. This choice satisfies the completeness and compatibility/continuity conditions. Therefore, mesh refinement will lead to a convergent solution.
By using the hat functions in the approximate weak form, we arrive at the matrix equations Ku = F as was shown before. We now need to calculate the components of the
stiffness matrix and the force vector. Start with the stiffness:
KAB =
NA,x 50 106 NB,x dx for A, B {2, 3, 4}
(7.5)
The symmetry of K permits us to only compute half of the matrix. Thus,

K22 =
=
K33 =
=
K44 =
K23 =
K24 =
K34 =
2
Z 4 2
1
1
6
50 10 dx +
50 106 dx
(N2,x ) 50 10 dx =
2
2
2
0
0
25 106 + 25 106 = 50 106
(7.6)
Z 6 2
Z 6
Z 4 2
1
1
50 106 dx +
50 106 dx
(N3,x )2 50 106 dx =
2
2
0
2
4
25 106 + 25 106 = 50 106
(7.7)
Z 6
Z 6 2
1
(N4,x )2 50 106 dx =
50 106 dx = 25 106
(7.8)
2
0
4

Z 6
Z 4
1
1
6
6
50 10
dx = 25 106
(7.9)
N2,x 50 10 N3,x dx =
2
2
0
2
0
(7.10)

Z 6
Z 6
1
1
N3,x 50 106 N4,x dx =
50 106
dx = 25 106
(7.11)
2
2
0
4
Z
Using the symmetry property, we can now write down the stiffness matrix as:
50 25 0
K = 25 50 25 106
0 25 25
(7.12)
We now need to compute the force vector F . Start by computing the part due to the
distributed load
Z 6
dist
FB =
NB 100 dx
(7.13)
0
19
By employing the definition of the hat functions

F2dist
F3dist
F4dist
Z
4
Z 4
1
1
x100 dx +
(x 4)100 dx = 100 + 100 = 200
2
2
2
Z 6
1
1
(x 2)100 dx +
(x 6)100 dx = 100 + 100 = 200
2
2
4
1
(x 4)100 = 100
2
(7.14)
(7.15)
(7.16)
The part from the end load is given by

FBend
load
0
= NB (6)150 = 0
150
(7.17)
The part from the displacement boundary condition (if it were not zero) would be given by
FBdisp =
Z
0
25
u
NB,x 50 106 N1,x u = 0 106
0
However, in our example u = 0. Adding the contributions together gives
200
FB = 200
50
The matrix problem now appears as:
200
50 25 0
u2
106 25 50 25 u3 = 200
0 25 25
u4
50
(7.18)
(7.19)
(7.20)
The solution is given as u = K 1 F . Since K is positive definite, we are assured that K

can be inverted and that the solution is unique. Typically, the inverse of K is not explicitly
formed as this is computationally wasteful. Usually, some sort of Gaussian Elimination or
iterative solution method is used. Either way, we get the solution:
u2
14
u3 = 20 106 inches
(7.21)
u4
18
P
We may now plot uh (x) =
NA (x)uA ; it is shown below on the left as the piecewise linear
function. The exact solution u(x) = (x2 + 9x) 106 is also shown below. On the right,
the approximate strains and the exact strains are shown.
20
S. Govindjee

!

"!#%$

6
45
2
/60
3
34
2
/10
3

"$#%#%& ' ()*+,-. !
&('') *!,+-.0/12, 0!%$

Remarks:
1. The approximate solution is nodally exact. This happens only for the 1-D linear case.
In general, FEM solutions are not nodally exact. For a proof of the nodal exactness
property, see, for example, Hughes [9].
2. Within elements, the solution is linear.
3. The strain approximation is much worse than the displacement approximation. The
strains are piecewise constant and will converge at a rate of one less than the displacements.
4. There are two ways to determine the reaction at the built-in end of the bar. The
inaccurate way is to use the approximate strain solution in the first element to calculate
an element stress and then multiply by the area to get force. This gives AEuh,x (0) =
50 7 = 350 lbs. The exact answer is AEu,x (0) = 50 9 = 450 lbs. Clearly,
using the element stresses is not a very good way of getting to the nodal reactions. A
more accurate and consistent way is to extract what the finite element method thinks
the reaction at node 1 is. This is done by assuming that the reaction is given as F1 , that
the deformation at node 1 is unknown, assembling the equation for node 1 and then
solving for F1 making use of the real boundary condition and the already computed
solution for the nodal displacements. The equation associated with node 1 is:
4
X
A=1
K1A uA = F1
(7.22)

where
F1 =
21
N1 (x)100 dx + F1 N1 (0) = 100 + F1
(7.23)
The stiffness elements can be computed as before to reveal that K11 = 25 106 ,
K12 = 25 106 , K13 = K14 = 0. Thus, we have that
(25 106 )u1 (25 106 )u2 = 100 + F1
(7.24)
But we know that u1 = 0 and that u2 = 14 106 . Thus F1 = 450 lbs a much
more accurate result.
5. An important characteristic of the finite element method is seen in Remark 4; viz. the
finite element approximation method respects the balance of forces on the body. The
sum of the reactions and the applied loads equals zero. Even though the solution is
just approximate, this fundamental principle of mechanics is preserved by the finite
element approximation.
22
S. Govindjee
Element View Point
In the last section, when constructing the stiffness matrix and the force vector, we were
continually faced with the evaluation of integrals over the body or domain. In particular,
for the elements of the stiffness matrix we had
Z
NA,x AENB,x dx
(8.1)
KAB =
where was the interval (0, 6) in the previous example. In the last section, we computed
each individual KAB fully and placed it in the stiffness matrix. When doing this, however,
a natural partitioning appeared; viz. the integration was always broken down into integrals
over individual elements. This leads us to consider reworking the assembly process into a
more natural procedure. First rewrite the integral above as
KAB =
Nel Z
X
NA,x AENB,x dx
(8.2)
e=1
where e is the element number, Nel is the number of elements, and e is the domain of the
eth element. Likewise for the force vector (with zero essential boundary conditions)
FB =
Nel Z
X
e=1
bNB dx + NB (
x)F
(8.3)
where there is a nodal force F applied at the node located at x = x. In the scheme being
introduced, the applied nodal loads will be handled separately. Only the integrations over
individual elements will be discussed in what immediately follows.
Let us now look at a single element in the discretization of the domain. For the linear
elements we have been using so far note that there are two nodes associated with each
element. Label the first node xe1 and the second xe2 . The superscript e is to remind us that
this is a local element level quantity; i.e. xe1 is the coordinate of node 1 on element e and
not global node 1. Associated with each of these local nodes is a local shape function. For
local node 1 this is N1e (x) = (x xe2 )/Le and for local node 2 this is N2e (x) = (x xe1 )/Le ,
where Le is the length of the element.
e
N (x)
1
e
N (x)
2
e
L
23
e
e
By noting that the local shape function derivatives are given by N1,x
= 1/Le and N2,x
=
e
e
1/L , we can now form the element stiffness matrix kab .
Z
e
e
e
kab =
Na,x
AENb,x
dx for a, b {1, 2}
(8.4)
e
Thus,
e
kab
AE
= e 2
(L )
Z
e
AE
dx = e
L
1 1
1 1
(8.5)
The element force vector is given in a similar manner as

bLe 1
e
fa =
1
2
(8.6)
(Note this only contains the contributions from the distributed loads.) After calculating
the element stiffness and force vector, we need to assemble them into the global stiffness
and force vector that we used previously. This is done with the aide of a table (sometimes
referred to as the LM array). This table gives the relation between local node numbers for
a given element and the global node numbers. As an example, consider the discretization
shown below for the 1-D rod in our example.
1
1
2
2
3
3
4
4
Element Number
5
5
Node Number
The LM array for this setup is given as:

Element Number
1 2 3 4 5
local 1 1 2 3 4 5
node # 2 2 3 4 5 6
The table is read as follows: When trying to assemble the contribution from, say element
2, the table says that the local node 1 is mapped to the global node 2 and that the local
node 2 is mapped to the global node 3. Therefore, the assembly process indicates that the
following operations should be performed:
K22
K23
K32
K33
=
=
=
=
2
K22 + k11
2
K23 + k12
2
K32 + k21
2
K33 + k22
(8.7)
(8.8)
(8.9)
(8.10)
24
S. Govindjee
and
F2 = F2 + f12
F3 = F3 + f22
(8.11)
(8.12)
If we assume that AE = 50 106 , all elements are of length 2, the body force b = 100, and
the end load is 150, then after this single assembly step the global system of equations
appear as:

F1
0
0 0
0 0 0 0
0
0 25 25 0 0 0 u2 100 0

0 25 25 0 0 0 u3 100 0
6
(8.13)
10
u4 = 0 + 0
0
0
0
0
0
0

0 0
0 0 0 0 u5 0 0
150
0 0
0 0 0 0
u6
0
The second vector on the RHS is due to the nodal point loads and point reactions at any
fixed nodes. Using the table we can add in the contributions from the other elements. For
instance, if element 3 is to be added in then local node 1 is mapped to global node 3 and
local node 2 is mapped to global node 4. Thus the global equations appear as:
0 0
0
0 0 0
0
0
F1
0 25 25 0 0 0 u2 100 0
0 25 50 25 0 0 u3 200 0
6
10
(8.14)
0 0 25 25 0 0 u4 = 100 + 0
0 0
0
0 0 0 u5 0 0
0 0
0
0 0 0
u6
0
150
If all the elements are
25 25
25 50
0 25
6
10
0
0
0
0
0
0
assembled, the global equations appear

0
0
0
0
0

25 0
0
0
u2

50 25 0
0
u3 =

25 50 25 0
u4
0 25 50 25
u5
0
0 25 25
u6
as:
100
200
200
200
200
100
F1
0
0
0
0
150
(8.15)
Remarks:
1. Normally, the rows associated with prescribed degrees of freedom are not assembled.
Also the columns associated with prescribed non-zero degrees of freedom are typically
moved to the right-hand-side.
2. Observe that for the hat functions, a tri-diagonal stiffness matrix is obtained. In terms
of bandwidth, the stiffness matrix has a half-bandwidth of 2.
3. Based on the preceding construction we can outline the basic structure of typical linear
finite element programs:
25
(a) Input: nodal locations.

(b) Input: element-node relations (connectivity).
(c) Input: body forces, nodal forces, prescribed displacements.
(d) Input: material properties.
(e) Loop over all elements e = 1, , Nel , form ke and f e , assemble into K and F .
(f) Solve for unknowns.
26
S. Govindjee
Higher Order Elements
So far we have only examined linear elements for 1-D problems. One of the most powerful
aspect of finite element analysis is that one can easily consider better (higher order)
approximations within the framework already developed. Higher order in this context refers
to the degree of completeness present in the approximation space. The hat functions allowed
us to exactly represent deformations of the form uh (x) = 0 + 1 x for x e . This implies a
solution that is linearly complete (all that was required for convergence in the rod problem).
A more accurate solution is however possible by considering an approximate solution that is
quadratically complete; i.e. one where arbitrary quadratic displacement fields can be exactly
represented over individual elements by the approximate solution. In this case we would be
able to write uh (x) = 0 + 1 x + 2 x2 for x e . Note that we are still referring to C 0
elements no higher degree of continuity over element boundaries is implied here.
The local (element) shape functions for the linear case were defined by requiring a shape
function to be 1 at its node and 0 at the other node. Since we were dealing with a linear
function, these two conditions defined a unique line; i.e. two points uniquely define a line.
For our quadratic element, we are faced with a shortage of conditions. To uniquely define
a parabola we need three conditions. This forces us to introduce an interior node in our
element. Thus the local node numbering scheme for the 1-D quadratic element is as shown
below:
Interior Node
xe
1
xe
2
xe
3
The element has three shape functions N1e , N2e and N3e . Each function is unity at its node
and zero at the other two nodes. Mathematical expressions for the shape functions are easily
derived using Lagrange interpolation formulas; viz.
(x xe2 )(x xe3 )

(xe1 xe2 )(xe1 xe3 )
(x xe3 )(x xe1 )
=
(x xe1 )(x xe2 )
=
N1e =
(9.1)
N2e
(9.2)
N3e
When plotted over the element these functions clearly meet the required conditions.
(9.3)
27
e
N (x)
1
e
N 2(x)
xe
1
e
N 3(x)
xe
2
xe
3
When viewed as global shape functions, they appear as shown below. Note that the shape
functions associated with the interior nodes are only non-zero in their associated elements.
N 3(x)
x
1
x
2
x
3
N 8(x)
x
4
x
5
x
6
x
7
x
8
x
9
The construction of the stiffness matrix and force vector for the quadratic elements follows
the scheme outlined in the previous section. The only difference is that the LM array is now
slightly larger. For the above discretization, the LM array is as follows:
Element Number
1 2 3 4
local 1 1 3 5 7
node # 2 2 4 6 8
3 3 5 7 9
Note that the element stiffness matrix is now a 3 3 matrix. The LM array, however, still
functions as before. For instance, if e = 4 then the one-one element of the local stiffness
4
matrix k11
maps over to K77 in the global stiffness matrix. The price paid for utilizing the
more accurate quadratic shape functions is that the global stiffness matrix is now pentadiagonal. Thus, solving for the unknowns will now be more expensive.
The error estimate for the quadratic element is obtained in the same way as for the linear
element. First the exact solution is expanded in a Taylor series over an element as
1
1
u(x) = u(xe1 ) + u0 (xe1 )[x xe1 ] + u00 (xe1 )[x xe1 ]2 + u000 ()[x xe1 ]3
2
6
(9.4)
where is some number is the range (xe 1, x). Now choose the element U h S h that is of
the form
U h (x) = 1 N1e (x) + 2 N2e (x) + 3 N3e (x)
= 0 + 1 [x xe1 ] + 2 [x xe1 ]2
(9.5)
(9.6)
where 0 = u(xe1 ), 1 = u0 (xe1 ), and 2 = 21 u00 (xe1 ). Note two things: 1) One can always
make this choice since S h is quadratically complete by construction. 2) U h in general will
not correspond to the finite element solution uh . Having chosen U h , now compute the error
1
E = u U h = u000 ()[x xe1 ]3
6
(9.7)
28
S. Govindjee
Since is a fixed point in the interval we can bound the error as

|E|max Ch3
(9.8)
where h is the length of the element. We may now argue heuristically that since U h is not
the best approximation to u in S h , the finite element solution (the best approximation)
uh will have an error that is bounded from above by E. Thus if e = u uh is the error, then
|e|max Ch3
(9.9)
A similar analysis shows that the error in the strains is given by

2
|e,x |max Ch
(9.10)
Remarks:
1. Halving the size of the elements reduces the error in the displacements by a factor of
8 and the error in the strains by a factor of 4.
2. We often use the notation e = O(h3 ) to express the convergence characteristics. This
notation is read e is order h cubed. Mathematically, it means that
e
constant
h0 h3
(9.11)
lim
The essence of this is that in the limit of small element sizes, the error decreases like
a cubic polynomial.
3. The convergence (error) behavior of elements is often plotted on log-log plots. The log
of the error is usually plotted against the log of the element size (as shown below).
log(e)
3
2
Linear Element
1
1
Quadratic Element
log(h)
10
29
Calculations Over Standard Elements/Isoparametric

Concepts
A useful extension of the local element view for calculating stiffness and force contributions
is the isoparametric concept. This concept is useful for a number of reasons and is perhaps
the most widely used family of elements. First, it standardizes many of the calculations
involved thus, lending itself to efficient implementation in a computer program. And
second, in multidimensional problems it (almost trivially) guarantees the satisfaction of the
completeness and compatibility/continuity conditions for convergence. We will first examine
this concept in 1-D and then later generalize it to multidimensions.
The main idea is to consider a coordinate transformation to a parent domain. In 1-D
the parent domain is the interval (1, 1). The element integrals are then carried out over
the parent domain. Let us first consider linear isoparametric elements. In this case, we wish
to consider a linear transformation from the element to the parent domain; i.e. the point xe1
maps to 1 and the point xe2 maps to 1.
(x)
x
x2e
x1e
-1
x ()
If we consider to be the coordinate in the parent domain, then we can write the transformation as
e

x2 xe1
xe + xe2
x=
+ 1
(10.1)
2
2
The fact that this gives the desired transformation is easily verified. To compute the finite
element integrals, we also need expressions for the shape functions in terms of . For the
linear hats, they may be defined as
Nae () = Nae (x()) = Nae (x) x()
(10.2)
Thus
N1e ()
N2e () =
x xe2
xe1 xe2
x xe1
xe2 xe1
x() =
xe2 xe1
2
xe1
1
x() = ( + 1)
2
xe1 +xe2
2
xe2
xe2
1
= ( 1)
2
(10.3)
(10.4)
30
S. Govindjee
Often these two results will be condensed into one equation by writing Nae () = 12 (a + 1)
where 1 = 1 and 2 = 1. Let us now consider the calculation of the element stiffness
matrix.
Z e
e
kab
=
x2
xe1
e
e
Na,x
AENb,x
dx
(10.5)
In order to properly transform this integration to one over the parent domain, we must
properly convert all parts of the integrand into functions of . This requires the following
results, which are merely consequences of the chain rule:
e
e
= Na,
,x
Na,x
dx = x, d
(10.6)
(10.7)
Using these two expressions, the element stiffness becomes

e
kab
1
Z 1
1
Z 1
1
e
e
Na,
,x AENb,
,x x, d
(10.8)
e
e
Na,
AENb,
,x d
(10.9)
a
b 2
AE
AE
d = e a b
e
2
2L
L
(10.10)
Remarks:
1. The force vector is worked in a similar fashion.
P
e
2. The coordinate transformation may be expressed as x() = 2a=1 Nae ()x
Pa 2. Thise is the
h
same form as the one used to express the displacement field u () = a=1 Na ()uea .
Because the same parameterization is used in both expressions the formulation is
termed an isoparametric formulation.
3. In more general cases, one starts by defining shape functions Nae () over the parent
domain using, for example, Lagrange interpolation formulae in terms of and then
generates shape functions over the element domain using the relation Nae (x) = Nae ()
(x). For higher order elements and elements in multidimensions, this generates shape
functions that are considerably different than those generated through the direct use
of Lagrange interpolation formulae on the element domain.
4. For this kind of formulation to make sense, the inverse transformation x1 = (x) needs
to exist and be unique and it needs to be differentiable of class C m when there are mth
order derivatives in the weak form. As long as the mapping x() is (a) one-to-one, (b)
onto, (c) C m , and (d) x, > 0, then (x) will exist and be C m . [This is a consequence
of the inverse function theorem; see e.g. Rudins book on Real Analysis.] For most
of our considerations, condition (d) on the Jacobian will be the most important; the
others are usually trivially satisfied.
31
5. With the conditions in Remark 4, it is relatively easy to show that the compatibility/continuity conditions hold; see e.g. Hughes[9]. The main question left to be answered as regards convergence is whether or not completeness is maintained. For the
1-D rod, completeness requires the approximate solution to be able to exactly represent arbitrary constant and linear displacement fields (constant strain and rigid body
motions) over single elements; i.e. if the exact solution is of the form u = constant
or u,x = constant, then uh should be able to exactly represent it. First, assume such
a field is given by u = 0 + 1 x, where the s are given constants. Second, assign
the nodal degrees of freedom in accordance with this field and, third, check that the
FEM interpolation formula exactly represents the assumed field. Thus, set the nodal
displacements of the element uea = 0 + 1 xea ; then
X
X
uh () =
Nae ()uea =
Nae ()[0 + 1 xea ]
(10.11)
X

X
= 0
Nae () + 1
Nae ()xea
(10.12)
By the isoparametric concept the last sum
the expression above is merely x().
P in
e
Thus completeness will be guaranteed if
Na () = 1. For the linear shape functions
considered so far
X
Nae () =
1 1
1 + 1 2 + 1
+
= + + + =1
2
2
2 2 2 2
(10.13)
Thus, completeness is satisfied and with the compatibility/continuity condition convergence is assured.
32
S. Govindjee
11
A 1-D Quadratic Isoparametric Element
As another example of the isoparametric concept, let us consider the formulation of the 1-D
quadratic element. As mentioned in the remarks above, we start by defining the parent
domain and shape functions over the parent domain. The parent domain is shown below,
with the interior node located at = 0.
(x)
x
x1e
x2e
x3e
-1
x ()
The shape functions are defined over the parent domain using Lagrange interpolation
formulae. Thus
( 1)
2
N2e () = 1 2
( + 1)
N3e () =
2
N1e () =
(11.1)
(11.2)
(11.3)
The mapping to the element domain is given by the standard formula

x() =
3
X
Nae ()xea
(11.4)
a=1
The element stiffness then becomes

e
kab
1
e
e
AENb,
,x d
Na,
(11.5)
A useful feature of this expression is that the shape function derivatives are given by standard
e
e
e
formulae (N1,
= 21 , N2,
= 2, and N3,
= + 21 ) that are independent of element nodal
locations. All of the information about the actual nodal locations is buried in the Jacobian
x, .
Remarks:
1. For the formulation to work, the Jacobian x, must be positive. For this to be the
case the node xe2 is required to be located in the middle half of the element; i.e.
3xe +xe 3xe +xe
xe2 ( 14 3 , 34 1 ).
33
2. The implied shape function over the element domain is given by Nae (x) = Nae () (x).
This does not correspond to the shape functions previously used for the element. In
particular, note that
(x xe2 )(x xe3 )
(11.6)
( 1)
( + 1)
+ 1 2 +
=1
2
2
(11.7)
N1e (x) = N1e () (x) 6=

except in the case when xe2 =
xe1 +xe3
.
2
3. The completeness condition is easily verified:

X
Nae =
34
S. Govindjee
12
Quadrature
Quadrature refers to the numerical evaluation of integrals; see books on numerical analysis
(e.g. Conte and de Boor [4]) for a thorough discussion of the material in this section.
The reason for studying quadrature rules stems from the fact that only in the most simple
elements can the necessary integrals be calculated exactly. The integrals we are referring to
are those that generate the stiffness and force/load vectors. As an example, we have for the
element stiffness of an axial rod:
e
kab
1
e
e
,x d .
AENb,
Na,
(12.1)
As another example, we have that the contribution of the body force to the element load
vector would be:
Z 1
e
fa =
Nae ()b(x())x, d .
(12.2)
1
Remarks:
1. Only when the inverse Jacobian ,x is a constant is a simple exact closed-form integration of the element stiffness possible. The force vector integration usually proves to
be relatively easy to perform in closed-form though this strongly depends on how
simple b(x()) is.
2. In most cases, the integrands will be rational functions in .
The generic question that arises is how one approximately or exactly computes an integral
of the form
Z 1
f () d .
(12.3)
1
Two common quadrature rules are the trapezoidal rule and Simpsons rule. The trapezoidal
rule says that
Z 1
f () d f (1) + f (1) .
(12.4)
1
The nomenclature is derived from the fact that the area under the curve f () is replaced by
a trapezoid, which is then integrated exactly.
35
f( )
-1
The error in this approximation can be written as a function that depends on the element
size h and a derivative of f at some point c in the interval (1, 1). For the trapezoidal rule
it can be shown that the error is a function of h3 and f 00 (c); i.e.
Etrap [h3 , f 00 (c)] .
(12.5)
Thus the rule is exact for linear functions (since f 00 0 for a line). Such error expressions
are derived using Newtons divided difference formulas.
Simpsons rule is somewhat more complex than the trapezoidal rule, and involves evaluating the function at three points. (The number of function evaluations is how one typically
computes the cost of a quadrature rule.) The rule is given as follows:
Z
1
4
1
f () d f (1) + f (0) + f (1) .
3
3
3
1
(12.6)
The error function for this approximation is given by

Esimp [h5 , f 0000 (c)] .
(12.7)
Thus it is exact for cubic functions (since f 0000 0 for cubics). Remarks:
1. The rate at which the error decreases upon choosing smaller and smaller elements is
higher for Simpsons rule than the trapezoidal rule.
2. Both of these rules can be written in the generic form
Z
f () d
N
int
X
Wl f (l ) ,
(12.8)
l=1
where Nint is the number of integration/quadrature points, Wl are known as the

weights, and l are known as the integration or quadrature points.
36
S. Govindjee
3. The trapezoidal rule makes use of two points. Thus we expect on intuitive grounds
that the rule should exactly integrate linear functions, since two points uniquely define
a line. Simpsons rule makes use of three points; thus, we expect on intuitive grounds
that the rule should exactly integrate quadratic functions. It does this, plus it exactly
integrates cubics. This indicates that our intuitive interpretation of the quadrature
rule is somewhat deficient. The deficiency lies in the fact that we have not counted the
weights as adjustable items or generalized points in the formulation of the quadrature
rule.
4. Thinking of the weights as generalized points, the trapezoidal rule involves four
points; so we would expect it to be exact for cubic functions. For Simpsons rule, we
have 6 points; so we would expect it to be exact for quintic functions. These two rules
are considered sub-optimal, since they do not perform as well as they should for the
number of generalized points used; i.e. for the number of function evaluations they do
not produce an optimal order of accuracy.
Optimal quadrature rules in 1-D are known as Gaussian quadrature rules. They have the
same form as in (12.8) but take advantage of the 2Nint generalized points that appear in the
rule; i.e. they exactly integrate polynomials of degree 2Nint 1. The choice of the weights
and quadrature(Gauss) points is made as follows: First consider an arbitrary polynomial of
degree 2Nint 1
2NX
int 1
p2Nint 1 () =
ai i .
(12.9)
i=0
Integrate this polynomial over the interval (1, 1):

1
Z 1
2NX
int 1
2
2
ai i+1
= 2a0 + a2 + +
a2Nint 2 . (12.10)
p2Nint 1 () d =

i+1
3
2Nint 1
1
i=0
1
P
The goal is to express this exactly in the form
Wl p2Nint 1 (l ). Making use of the generic
expression for p2Nint 1 , we then have that
N
int
X
l=1
Wl p2Nint 1 (l ) = a0
W l + a1
+ a2Nint 2
Wl l + a2
Wl l2 +
Wl l2Nint 2 + a2Nint 1
Wl l2Nint 1 .
(12.11)
The conditions under which the quadrature rule (12.11) is exactly equivalent to (12.10) can
be found by matching the coefficients of the ai s in the two expressions since the ai s are
arbitrary. This results in the following conditions:
P
P Wl = 2
Wl l = 0
P W 2 = 2
l l
3
.
(12.12)
Wl 2Nint 2 =
P l2Nint 1 2Nint 1
Wl
=0
l
37
These are a set of 2Nint non-linear equations in 2Nint unknowns (Wl , l ). Solving them
produces an optimal quadrature rule in 1-D. The error term for this rule is of the form:
E[h2Nint +1 , f (2Nint ) ] .
(12.13)
As an example, consider the case of Nint = 2. First, write out the conditions (12.12)
W1 + W2 = 2
W1 1 + W2 2 = 0
.
W1 12 + W2 22 = 23
W1 13 + W2 23 = 0
(12.14)
These equations are solved as follows: First it is argued that the quadrature rule should be
unbiased with respect to the line = 0, since it should produce the same result whether
integrating f () or f (). This implies that 1 = 2 and W1 = W2 . With these results,
(12.14)1 implies that W1 = W2 = 1 and (12.14)3 implies that 1 = 2 = 13 . Thus, the two
point Gaussian rule is
Z 1
1
1
f () d f ( ) + f ( ) .
(12.15)
3
3
1
The error in the rule depends on the derivative of order 2Nint = 2 2 = 4 and thus is exact
for cubics. The rule involves the same number of function evaluations as the trapezoidal rule
but is far more accurate.
13
Quadrature and Convergence
A natural question that arises is: How many Gauss points do we need in FEM so that
our error/convergence estimates (which were based on exact integrations over elements) still
hold? This is a somewhat delicate issue, and we will not discuss the theory in detail; see
Strang and Fix 4.3 for a comprehensive discussion. There are two conditions that need to
be satisfied for our previous estimates to hold. The first condition is that the quadrature
rule needs to exactly integrate polynomials of order k + k 2m, where m is the order of the
derivatives appearing in the weak form, k is the degree of the highest monomial appearing
in the shape functions and k is the degree of completeness of the shape functions. In 1-D we
typically uses elements where k = k. In multi-dimensions this is not always the case; 4-node
quads, for instance have, k = 2 and k = 1. The second condition that must hold is that the
global stiffness matrix needs to remain positive definite. Remarks:
1. In many cases, it is the second condition that governs the choice of a quadrature rule.
2. In multi-dimensions, optimal quadrature rules are not known, in general. For most
shapes, however, tables are available that define reasonably efficient rules. For quadrilateral and brick-shaped elements, sufficient rules are often generated by merely applying the Gauss rule for 1-D simultaneously in all directions; i.e. if and are the
38
S. Govindjee
coordinate directions in the 2-D parent domain, then we often use:
#
Z 1Z 1
Z 1 "N
int
X
f (, ) d d
Wj f (j , ) d
1
(13.1)
j=1
N
int N
int
X
X
Wl Wj f (j , l ) .
(13.2)
l=1 j=1
Thus, for example, a two point rule would generalize in 2-D to a four point rule with
unit weights with Gauss points located at ( 13 , 13 ).
14
39
Introduction to Multi-Dimensional Problems
Multi-dimensional problems are solved in an identical fashion to 1-D problems. One first
formulates the strong form of the governing equations, writes a weak form of the governing
equations, approximates the solution and variation spaces, and solves a set of matrix equations. As an introduction to the formulation of multi-dimensional FEM problems, consider
the problem of steady state heat conduction. Heat conduction proves to be an easier introduction to multi-dimensional problems than elasticity, since the unknown field is a scalar
field as opposed to a vector field. The domain of the problem is and it has a boundary
that is partitioned into two parts h and g . h is the portion of the boundary where heat
flux is prescribed and g is the portion of the boundary where temperature is prescribed.
g g = and g g = .
h
T=T
q.n = h
g
The governing strong form equation corresponds to the balance of energy in the material
and is given by
Find
T (x) such that
q(x)
=
r(x)
for x
(S)
(14.1)
(x)
T
(x)
=
T
for
x g
q(x) n(x)
=
h(x)
for x h
where q = T is the heat flux vector, is the matrix of conductivities (analogous to
moduli in mechanics problems), g is the portion of the boundary where the temperature is
prescribed as T(x), and h is the portion of the boundary where the heat flux normal to the
boundary is prescribed as h(x). The solution and variation spaces are seen to be
V = {v(x) | v(x) = 0 for x g }
S = {T (x) | T (x) = T(x) for x g } .
(14.2)
(14.3)
Note the choice of essential boundary conditions corresponds to prescribed values of the
unknown field variable (the temperature in this case).
To generate the weak form, multiply the strong form by an arbitrary member of the the
space of variations and apply the divergence theorem. For our purposes, we will use the
divergence theorem in the generic form of
Z
Z
[],i d = []ni d .
(14.4)
40
S. Govindjee
Thus we have that

Z
vr ,
(14.5)
vr ,
(14.6)
(vij T,j ),i v,i ij T,j =

vr ,
Z
Z
vij T,j ni + v,i ij T,j =
vr .
(14.7)
vqi,i =
v(ij T,j ),i =
(14.8)
Now make use of the boundary conditions. First, on g , v is equal to zero. Second, on h ,
qi ni = h. Using these two facts in (14.8) gives the weak form of the problem:
T S such
Find
Z
Z that Zv V
(W )
(14.9)
v,i ij T,j =
vr
vh .
The finite element approximation proceeds by replacing the infinite dimensional spaces
V and S by finite dimensional subspaces. For the types of finite element methods used in
this course, the following definitions suffice:
X
NA (x)vA } ,
(14.10)
V h = {v h (x) | v h (x) =
Ag
= {T (x) | T (x) = v h (x) + g h (x), v h (x) V h ,

X
NA (x)T(xA )} ,
g h (x) =
(14.11)
Ag
where
= {All Node Numbers} ,
g = {Node Numbers Corresponding To g } .
(14.12)
(14.13)
As before, we require that

NA (xB ) = AB =
1 A=B
0 A 6= B .
With these definitions in hand, the approximate weak form becomes
T h S h Zsuch thatZ v h V h
Find
Z
(W h )
v,ih ij T,jh =
vhr
vhh .
(14.14)
(14.15)
The corresponding matrix equations are found by using the shape function expansions
for the unknown temperatures and the variations. After a bit of algebraic manipulation and
41
placing the nodal temperature unknowns in a column vector, we arrive at a set of matrix
equations:
KT = F ,
(14.16)
where
T
1
T2
T =
,
...
Z
Z
KAB =
NA NB =
NA,i ij NB,j , and
Z
Z
Z
X
FA =
NA r
NA h NA
NC T(xC )
NA r
NA h
(14.17)
(14.18)
(14.19)
Cg
NA,i ij
NC,j T(xC ) .
(14.20)
Cg
Remarks:
1. The restriction of the integrations in (14.18) and (14.20) to be over individual elements
produces expressions for the element stiffness matrices and load vectors. For example,
the element stiffness becomes:
Z
Z
e
e
e
e
e
kab =
Na Nb =
Na,i
ij Nb,j
a, b {1, 2, , Nen } ,
(14.21)
e
where Nen is the number of element nodes.

2. Local node numbers are mapped to global node numbers using a LM-array, just as
in the 1-D case. Each row corresponds to a local node number and each column
corresponds to an element number. The entries in the array are the global node
numbers.
3. There exists a useful notation, used widely in FEM literature for the shape function
derivatives. For the case at hand, we write
e
Na,1
e
Ne
Na = B a =
.
(14.22)
a,2
e
Na,3
The vector B a has dimensions of Nsd by 1, where Nsd stands for number of spatial
dimensions. This allows us to write the components of the element stiffness matrix as
Z
e
B Ta B b
(14.23)
kab =
e
and the whole matrix as

e
k =
Z
e
B T B ,
(14.24)
42
S. Govindjee
where
B=
B 1 B 2 B Nen
(14.25)
4. All previous estimates about the quality of the FEM solution hold for this case as well.
Let T be the exact solution to the problem and T h the FEM solution.
Then, the error
R
h
in the solution is given by e = T T . If we define a(f, g) = f g, then
a(e, v h ) = 0
v h V h
a(T Th , T Th ) a(e, e)
a(T, T ) a(T h , T h ) .
Th S h
(14.26)
(14.27)
(14.28)
5. Also note that the global stiffness matrix K will be symmetric and positive definite,
as long as is symmetric and positive definite. (The most common case is isotropic
heat conduction where = 1.)
15
43
Isoparametric Triangles
The parent domain for triangular elements is a right triangle with vertices at (0,0), (1,0),
and (0,1); see figure.
x2
x( )
x1
For the so-called constant strain triangle (CST), there are three nodes located at the
vertices of the triangle. Associated with each node is a shape function. This element is the
generalization of the linear 1-D element to 2-D; thus, the shape functions are planes. A small
amount of consideration produces the following expressions for the shape functions in the
parent domain.
N1e = 1 ,
N2e = 2 ,
N3e = 1 1 2 .
(15.1)
(15.2)
(15.3)
The expressions define planes as desired; they are unity at their associated nodes and they
are zero at the remaining nodes. [Note, the node at (1,0) is considered node 1; the node at
(0,1) is considered node 2; and the node at (0,0) is considered node 3.] A convenience that
is often introduced is the notion of triangular or area coordinates. Using the definition given
in CMP [5, 5.2], it is easily seen that the 1 = t1 , 2 = t2 , and 1 1 2 = t3 ; (see figure).
Using these coordinates, the shape functions can be expressed as Ni (t1 , t2 , t3 ) = ti .
44
S. Govindjee
or t 2
t3= 0
t 3 = 1/2
t3= 1
1 or t 1
With the shape function definitions given, the isoparametric map over an element is defined
by:
3
X
xi =
Nae xeia .
(15.4)
a=1
Once the isoparametric mapping is defined one can compute the required FEM integrals over
the parent domain. Thus integrals
Z
e
f (x) dx 7
int
X
x
x
f (x())det[ ] d
W` f (x(` ))det[ (` )]
`=1
(15.5)
where ` represent integration points and W` are the weights. Key to using this formulation
is the proper calculation of the derivatives. Since shape functions for triangular elements
are more conveniently given in triangular coordinates, derivatives with respect to the parent
domain coordinates must be converted, through use of the chain rule, into derivatives with
respect to the triangular coordinates. Thus
t1
t2
t3
=
+
+
=
1
1 t1 1 t2 1 t3
t1 t3
t1
t2
t3
=
+
+
=
2
2 t1 2 t2 2 t3
t2 t3
The Jacobian matrix can then be expressed as
x1 x1
t3
x
t1
= x
2
2
x
t1
t3
x1
t2
x2
t2
x1
t3
x2
t3
(15.6)
(15.7)
(15.8)
For higher order triangles formulae (15.4)-(15.8) still hold, except that in (15.4) the
summation would include the added nodes. For instance, in the 6 node quadratic triangle,
45
there are additional nodes located at the mid-sides of the parent triangle. Note that in
triangular coordinates, these points are conveniently given by ( 12 , 12 , 0), (0, 12 , 12 ), and ( 12 , 0, 12 );
see figure.
1
1
The usefulness of triangular coordinates is now apparent as the proper shape functions can
be easily written down (by inspection) through the use of Lagrange interpolation formulae.
1
N1e = 2t1 (t1 )
2
1
e
N2 = 2t2 (t2 )
2
1
e
N3 = 2t3 (t3 )
2
N4e = 4t1 t2
N5e = 4t2 t3
N6e = 4t3 t1
(15.9)
(15.10)
(15.11)
(15.12)
(15.13)
(15.14)
Note, these shape functions are properly normalized to unity at their associated nodes and
are zero at the remaining nodes. Remarks:
1. When the unknown field is examined in parent domain coordinates, the CST produces
behavior of the form
T h = 0 + 1 1 + 2 2 .
(15.15)
This is a complete polynomial in two variables of the first degree. For the quadratic
triangle, one has that
T h = 0 + 1 1 + 2 2 + 3 1 2 + 4 12 + 5 22 .
(15.16)
This is a complete second degree polynomial inP

two variables. The last two expressions
h
are easily seen to be true by expanding T = Nae Tae .
46
S. Govindjee
2. A handy device for remembering the terms in the expansion is Pascals triangle. Each
row in the triangle represents the terms that need to be added to the previous rows to
generate a complete polynomial of one degree higher. The number of individual terms
used from the triangle also indicates the number of nodes needed in the element.
1
x
x
x
x4
xy
x3 y
y2
xy
xy
x2 y 2
(15.17)
y
xy 3
y4 .
3. Completeness is as easily verified for isoparametric elements in multi-dimensions as it

was in one-dimension. For the heat conduction problem, the number of derivatives on
the unknown in the weak form is 1; i.e. m = 1. Thus constant gradient and constant
temperature fields must be exactly representable by the FEM interpolation. Choose
an exact field
T = 0 + 1 x1 + 2 x2 ,
(15.18)
where the s are arbitrary constants. Set the nodal temperatures of an element in
accord with this exact field and verify that the interpolated field is identical to the
exact field everywhere. This gives:
X
Nae ()(0 + 1 xe1a + 2 xe2a ) .
(15.19)
Th =
a
P e e
The isoparametric
concept
says
that
x
=
Na xia ; therefore, completeness will be
i
P e
guaranteed if
NA = 1 just as in the 1-D case. As a simple example, consider the
CST:
3
X
Nae = t1 + t2 + t3 = 1 + 2 + (1 1 2 ) = 1
(15.20)
a=1
4. Compatibility for such elements can be relatively easily verified. For heat conduction,
m = 1 which implies that C 0 continuity is required across element boundaries. Clearly,
the FEM temperatures are continuous at the nodes since nodal data is shared among
adjoining elements. The question to be answered is whether the temperature restricted
to an inter-element boundary is the same when viewed from different elements. As an
example, consider the CST and examine T h restricted to the edge where t1 = 0. Along
this edge
(15.21)
T h |t1 =0 = t1 T1e + t2 T2e + t3 T3e = T3e + (T2e T3e )t2
where the relation t3 = 1 t1 t2 = 1 t2 has been utilized. Note that (15.21)
represents a linear function that only depends on nodal data at nodes 2 and 3. This
function is unique since two points uniquely define a line. Any element that is attached
at this edge will share the same nodal data. Thus when the temperature is viewed from
this other element it will be the same linear function along the shared edge. Hence C 0
continuity will be satisfied. The other edges of the element may be examined in the
same fashion.
16
47
Summary of Elasticity Equations
In this section the fundamental equations of small deformation linear elasticity theory are
summarized. The primary equation to be solved in an elasticity problem is the balance of
linear momentum. In the static case this reduces to
+ b = o.
(16.1)
ij,j + bi = 0i .
(16.2)
In indicial notation, this is written as
In the above, is the Cauchy stress tensor and b is the body force vector. The balance
of angular momentum reduces to the requirement that the stress tensor be symmetric; i.e.,
= T or ij = ji .
The displacement field u is related to the strain tensor through the relation
(u) = s u =

1
u + (u)T ,
2
(16.3)
which in indicial notation reads

1
ij (u) = (ui,j + uj,i ) ;
2
(16.4)
i.e., the strain tensor is the symmetric gradient of the displacement field.
The strains are related to the stresses through the stiffness tensor. In direct notation:
= C: ,
(16.5)
ij = Cijkl kl .
(16.6)
Cijkl = (ik jl + il jk ) + ij kl
(16.7)
or in indicial notation
For the case of isotropy,
where and are known as the Lame constants; is also known as the shear modulus. In
terms of Youngs modulus and Poissons ratio,
=
E
2(1 + )
E
.
(1 + )(1 2)
(16.8)
Thus, it can be written that

ij = 2ij + kk ij .
(16.9)
When inverted, this relation reads

ij =
1+
ij kk ij .
E
E
(16.10)
For convenience, Voigt notation is often employed when writing these relations. In this
notation, 2nd order tensors are mapped to vectors and 4th order tensors are mapped to
48
S. Govindjee
matrices. Strain tensors and stress tensors are mapped in slightly different ways. The strain
tensor becomes a strain vector with the following ordering:
11
22
33
=
(16.11)
212 .
223
213
Note the use of the 2s; i.e., engineering shear strain as opposed to tensorial shear strain is
used in the Voigt strain vector. In the stress vector, the 2s do not appear; i.e.
11
22
33
=
(16.12)
12 .
23
13
When working with Voigt notation, the stiffness tensor becomes a 6 6 matrix (for 3-D).
In the special case of Plane Strain, where x 3 () = 0 and u3 = 0,

11
2 +
0
11
22 =
2 + 0 22
(16.13)
12
0
0
212
and 33 = (11 + 22 ). In the special case of
non-zero),
11
1
E
22 =

(1 2 )
12
0
Plane Stress (where only 11 , 22 , and 12 are
1
0
11
0
22
0
1
(1 )
212
2
(16.14)
and 33 = (11 + 22 ). For both special cases, these relations are often abbreviated in
direct notation as = D.
17
Strong, Weak, and Minimization Forms of Elasticity
The basic field equations of elasticity are given by the momentum balance equations, the
strain-displacement equations, and the constitutive equations. To formulate a boundary
value problem, boundary conditions have to be provided. There are two primary types
of boundary conditions for elasticity; these are displacement and traction/force boundary
conditions. (Other more complicated boundary conditions such as elastic foundations will
not be discussed.) Displacement boundary conditions are given in the form
(x)
u(x) = u
x u
(17.1)
49
is a given (prescribed) displacement function on a portion of the boundary u .

where u
Traction/force boundary conditions are given in the form
(x)n(x) = t(x)
x t
(17.2)
where t is a given (prescribed) traction function on a portion of the boundary t . An

important consideration in stating boundary conditions for elasticity problems is that: At a
given point on the boundary in a given coordinate direction only a displacement or a force
may be prescribed but not both. Violating this rule produces boundary value problems for
which solutions do not exist. This consideration is expressed mathematically as u t = .
This construction also implicitly assumes that u t = , where represents the entire
boundary to the domain .
Collecting our results we have the strong form for the elasticity problem as:
Find
u(x) such that
+b
=
o
for x
(S)
(17.3)
(x)
u(x)
=
u
for x u
(x)n(x)
= t(x)
for x t ,
where = 12 (u + (u)T ), = C : , = T , u t = , and u t = . To obtain the
weak form equations from (17.3) we first define our function spaces. The unknown field variable in elasticity problems is the displacement. Thus, the prescribed displacement boundary
conditions are the essential boundary conditions and the traction boundary conditions are
identified as the natural boundary conditions. Knowing this, the space of weighting functions
(or admissible variations) and the solution space can be defined by:
V = {v(x) | v(x) = 0 for x u }
(x) for x u } .
S = {u(x) | u(x) = v(x) + g(x), v V, and g(x) = u
(17.4)
(17.5)
As before, there are also certain differentiability/integrability conditions on the functions in

these spaces. As we will be able to easily identify from the weak form equations, by counting
derivatives, these will require (for our purposes) that we have C 0 continuity.
The weak form is generated as before by taking an arbitrary function in V and multiplying
both sides of the balance equation with it and integrating over the domain. In this case, the
balance equation is not a scalar equation; it is a vectorial equation (has one free index). So,
instead of multiplying, a dot product is used. The divergence theorem (14.4) is then used to
make the integrand symmetric in the unknown displacements and the weighting functions.
Thus,
Z
ij,j vi + bi vi = 0
(17.6)
Z
(ij vi ),j ij vi,j + bi vi = 0
(17.7)
Z
Z
(17.8)
ij vi nj + ij vi,j + bi vi = 0
Z
Z
Z
vi ti + bi vi =
ij vi,j
(17.9)
t
50
S. Govindjee
Remarks:
1. This last equation, which is recognized to be the principle of virtual work, holds independent of material response; i.e. it is valid for general stress analysis problems. This
follows trivially, since no use was made of the constitutive properties of the material.
2. Because angular momentum balance (sum of the moments equals zero) implies that
is symmetric ( = T or ij = ji ), one can always replace vi,j by its symmetric part;
viz. 21 (vi,j + vj,i ). This allows us to use Voigt notation for the integrand:
ij vi,j = : (v) = T (v) =
11 22 12
v1,1
.
v2,2
v1,2 + v2,1
(17.10)
Note the abuse of notation in the second equality.

3. If we restrict ourselves to linear elasticity, we can rewrite the internal virtual work as
follows:
Z
Z
Z
Z
ij vi,j =
vi,j Cijkl kl (u) =
ij (v)Cijkl kl (u) =
T (v)D(u)
(17.11)
where D is the elasticity tensor expressed in Voigt notation.

4. The weak form statement of the problem is
u ZS such that
Find
Z
Z v V
(W )
vi ti + bi vi =
ij vi,j .
(17.12)
where ij (u) = 12 (ui,j + uj,i ) and ij = Cijkl kl (u).

5. The weak form equations can also be derived by considering the minimization of potential energy. For an elastic body, the potential energy is given by
(u) =
W ((u))
bu
t u
(17.13)
where W ((u)) = 12 (u) : C : (u) is the stored elastic energy density. The true displacement field is found by minimizing this potential. The condition for a minimum is
found by taking the directional derivative of (17.13) in the direction of an admissible
variation (an element of the set V) and setting it equal to zero. Thus we write:

d
(u + v) = 0 .
d =0
(17.14)
51
In some books, this is simply denoted by (u; v) = 0. Proceeding with the directional
derivative formula, we have

d
(u + v) =
d =0

Z
Z
d
1
=
jk (u + v)Cjklm lm (u + v) bi (ui + vi )
d =0 2
ti (ui + vi )
(17.15)
t
Z
Z
1
1
=
jk (u)Cjklm lm (v) + jk (v)Cjklm lm (u) bi vi
2
2
ti vi
(17.16)
t
Z
Z
Z
=
ij (v)Cijkl kl (u) bi vi
ti vi
(17.17)
Setting (17.17) equal to zero yields the weak form.
18
Finite Element Approximations for Elasticity
The finite element approximation for the elasticity problem proceeds in a fashion identical
to the problems examined to this point. First, the solution space and the weighting function
space are replaced by finite dimensional subspaces. Second, these subspaces are employed in
the weak form to generate matrix equations which are then used to solve for the unknown
displacements at the nodes. Because we are dealing with vector valued unknowns, the
notation becomes cumbersome and new notation is introduced in an effort to simplify the
look of the equations. Many of the symbols that are used have double meanings and their
precise definitions have to be inferred from the context in which they are used. An optimal
notation has yet to be discovered.
The space of admissible variations is defined as
V h = {v h (x) | vih (x) =
h
NA (x)viA
}
(18.1)
Au
where is the set of all node numbers, u is the set of node numbers where the displacements
h
refers to the vector vih at node A. In what follows the superscript
are prescribed, and viA
hs will usually be dropped to simplify the notation; the meaning should be clear from the
context. As an added convenience the summation contained in (18.1) is often expressed as
52
S. Govindjee
follows (note the abuse of notation)
v1 (x)
v2 (x)
NA (x)viA =
Au
N 1 0 N 2 0 Nn
0 N1 0 N2 0
v11
v21
v12
v22
..
.
Nn
v1n
v2n
= Nv
(18.2)
where n is the number of nodes in u . The N matrix is of great utility when working with
the FEM equations; as given above, it is strictly for 2-D problems but is trivially extended
to 3-D.
This type of notation also extends to the symmetric gradients of functions; viz. in 2-D
we have that
v11
v21
v12
v1,1
N1,1 0 N2,1 0 Nn,1
0
= 0 N1,2 0 N2,2
v2,2
0 Nn,2 v22 = Bv (18.3)
..
v1,2 + v2,1
N1,2 N1,1 N2,2 N2,1 Nn,2 Nn,1 .
v1n
v2n
Note the dimensions of the quantities above; v is 2n 1 and B is 3 2n. The pairs
of columns that are associated with each node are often referred to as B A so that B =
[ B 1 B 2 B n ]. In 3-D problems these matrices and vectors are expanded to handle
the added dimension. In many books, B is called the strain-displacement operator, because
given a displacement vector it generates a strain vector.
This notation can now be applied to the weak form equations for elasticity (17.12). This
gives
Z
Z
Z
T
T
(N v)T t
(18.4)
(Bv) DBu = (N v) b +
Note that the displacement boundary conditions have been assumed to be zero in order
to simplify the presentation. They can, however, be added without much difficulty. The
arbitrary vector of variation unknowns, v, is now factored out of this equation to yield

Z
Z
Z
T
T
T
T
B DB u N b
N t =0
(18.5)
v
Since v is arbitrary, the expression in the square brackets must be zero. Therefore by defining
Z
K =
B T DB
(18.6)
Z
Z
T
N b+
N T t
(18.7)
F =
53
we obtain the standard problem of Ku = F .

As before, the integrals are usually restricted to individual elements and then the contributions are added together through an assembly operation like that of direct stiffness. In
this case, the element stiffness matrix is given by
Z
e
B T DB
(18.8)
k =
e
where the dimension of B is now 32Nen for 2-D elasticity problems Nen being the number
of element nodes. In 3-D the dimension of B becomes 6 3Nen . The dimension of ke itself
is either 2Nen 2Nen or 3Nen 3Nen . The element force vector has a similar expression:
Z
Z
e
T
f =
N b+
N T t
(18.9)
e
et
where N will have dimensions 2 2Nen and 3 3Nen for 2-D and 3-D problems, respectively.
[Note that the last integral in (18.9) is a surface integral.]
54
19
S. Govindjee
Consistent Nodal Loads
As seen from the FEM formulation of elasticity theory, the nodal force vector is given by a
2n 1 dimension vector in 2-D problems, where n is the number of active nodes (number
of nodes where the displacements are unknown). The formal expression is
Z
Z
T
F =
N b d +
N T t dt .
(19.1)
In order to obtain proper results from the finite element method, load values must be input in
accordance with this formula. Doing so produces what are known as consistent nodal loads.
Note that this depends on the shape function matrix N and thus will change depending on
the element type for the same given physical problem.
Independent of element type the procedure for calculating consistent nodal loads is essentially the same. First the element point of view will be adopted. Thus, element consistent
nodal loads will be calculated and then assembled into the global load vector. It is this global
load vector that is input in the input file of a finite element program. For example, for the
3-node triangle shown below
0
N1 0
f11
e
f21
Z 0 e N1
Z 0
f12
0e det .
N2 0e 0 det =
(19.2)
fe =
=
f22
q
N2
et N2 q
et 0
0
N3e 0
f13
0 N3e
N3e q
f23
q
2
h
x2
1
x1
In the last equality, advantage has been taken of the fact that t1 = 0 on edge 2-3. Further
note, this implies that N1e = 0, N2e = t2 and N3e = 1 t2 on edge 2-3. Thus the integral on
et can be converted to an integral in terms of t2 :
Z 1 0
0
e h dt2
fe =
(19.3)
0 N2 q
0
N3e q
55
The h in the integrand represents the change of variables from the physical domain where
the edge length is h to the isoparametric domain where the edge length is 1. Plugging in the
expressions for the shape functions along the edge gives the final result of
e
f =
0
0
0
qh/2
0
qh/2
(19.4)
This is a general result for isoparametric elements with linear edges; viz., a uniform load on
the edge is equally distributed between the two nodes that make up the edge.
As a second example consider a 6-node triangle with straight sides, evenly spaced nodes,
and a uniformly distributed load q on edge 1-6-3.
q
6
1
3
h
5
4
x2
2
x1
In this case, the element load vector is a 12 1 dimension vector:
fe =
et
0
N1e q
0
0
0
N3e q
0
0
0
0
0
N6e q
e
dt .
(19.5)
In the above expression, advantage has been made of the fact that only shape functions
N1e , N3e and N6e will be non-zero along et (edge 1-6-3). Along the edge 1-6-3, t2 = 0 and
56
S. Govindjee
t3 = 1 t1 , thus
N1e = 2(t1 1/2)t1
N3e = 2(1/2 t1 )(1 t1 )
N6e = 4(1 t1 )t1 .
(19.6)
(19.7)
(19.8)
Using this information, it is straight forward to compute the non-zero entries in the element
load vector for this case where the edge is straight and the nodes are evenly spaced; viz.,
e
f21
2(t1 1/2)t1 qh dt1 =
0
e
f23
qh
6
2(1/2 t1 )(1 t1 )qh dt1 =
0
e
f26
4(1 t1 )t1 qh dt1 =
2qh
.
3
(19.9)
qh
6
(19.10)
(19.11)
Note the non-intuitive nature of the consistent nodal load, one-sixth of the load is given to
the corner nodes and two-thirds of the load is given to the mid-side node.
Remarks:
1. Consistent nodal loads from body forces are computed in an identical fashion; except,
the integrals involved are volume/area integrals over the elements themselves.
2. Most but not all codes automatically calculate consistent nodal loads for the cases of
uniform pressure on linear edges/surfaces and constant body forces. More complex
loading cases must often be manually calculated by the analyst.
The previous cases have been special cases of uniform normal pressure on an edge. In the
general case for this class of loads, the edge may be curved and the nodes need not be evenly
spaced. As an example of a more general case consider the general uniform normal traction
t = pn, where p is a constant pressure and n is the normal to a surface. The element force
vector is given by
Z
e
f =
N T pn det ,
(19.12)
et
where det is the integration parameter on the surface. To place the presentation in context,
consider the case shown in the figure below, where the node numbers given are the local
node numbers for the element.
57
1
Uniform normal pressure
acting on this face
4
x2
2
6
5
3
x1
The equation for the edge where the pressure is specified can be given parametrically in
terms of the isoparametric mapping as
xi (t1 , t2 , t3 )|t2 =0 = xi (t1 , 0, 1 t1 ) = xi (t1 ) .
(19.13)
From elementary calculus, we know that for a parametrically defined curve the normal vector
is given as

1
x2 /t1
p
n=
.
(19.14)
x1 /t1
(x1 /t1 )2 + (x2 /t1 )2
In a similar fashion, we know that the integration parameter is given by
p
det = (x1 /t1 )2 + (x2 /t1 )2 dt1 .
(19.15)
see the figure below for the construction that leads to this last relation.
(x1 / t 1) dt 1
(x2 / t 1) dt 1
d
t
Thus our final formula for the nodal loads is

Z 1
x2 /t1
e
T
f =
N p
dt1 .
x1 /t1
0
(19.16)
Note that for the case being considered

x1 = x11 N1 (t1 , 0, 1 t1 ) + x13 N3 (t1 , 0, 1 t1 ) + x16 N6 (t1 , 0, 1 t1 )
x2 = x21 N1 (t1 , 0, 1 t1 ) + x23 N3 (t1 , 0, 1 t1 ) + x26 N6 (t1 , 0, 1 t1 )
and the shape function matrix needs to be evaluated at N (t1 , 0, 1 t1 ).
Remarks:
(19.17)
(19.18)
58
S. Govindjee
1. An alternative and equivalent formula for computing such integrals that is often used
in continuum mechanics is
Z
Z
x X T
][
] nX d ,
(19.19)
f (x)n d = f (x(X))det[
X x
where x and X are different parameterizations of the same edge. x are the physical
coordinates along the edge and X are an alternative set of coordinates along the
edge = X(); in continuum mechanics the coordinates X are usually the Lagrangian
coordinates of the material points; in our case they will be the isoparametric coordinates. n is the normal to the edge in the physical coordinates and nX is the normal
to the edge in the X coordinates. For the case explored above this relation reduces to
Z
Z 1
x
f (x)n d =
f (x(t1 , 0, 1 t1 ))det[ (t1 , 0, 1 t1 )][ (t1 , 0, 1 t1 )]T nX dt1 ,
0
(19.20)
where nX is the normal to the edge in the parent domain; in this case nX = (0, 1)T .
2. In the case of uniformly applied tangential loads with magnitude the traction vector
can be given by a formula similar to the case of normal pressure:

1
t = x1 /t1 p
(19.21)
2
x2 /t1
(x1 /t1 ) + (x2 /t1 )2
in the case that the load is applied to the edge 1-6-3. The final expression for the nodal
load vector is

Z 1
x1 /t1
e
T
f =
N
dt1 .
(19.22)
x2 /t1
0
20
Isoparametric Quadrilaterals
A popular and commonly available class of elements are the isoparametric quadrilateral
elements. The starting element in this hierarchy is the 4-node quad which is also known
as the bilinear quad. In the spatial domain it can be any 4-sided quadrilateral where all
interior angles are less than 180 degrees. When an interior angle exceeds 180 degrees the
Jacobian determinant of the isoparametric map is no longer strictly positive thus leading to
a breakdown of the mathematics upon which the isoparametric concept is based.
The parent domain for all quadrilateral elements is the bi-unit square as shown below.
x( )
2
(-1,1)
4
(1,1)
3
1
1
(-1,-1)
2
(1,-1)
59
For the 4-node quadrilateral, one can employ a Lagrange interpolation idea to generate shape
functions for the element that are unity at their associated node and zero at the other nodes.
This leads to the following formulae
1
(1 1 )(1 2 )
4
1
=
(1 + 1 )(1 2 )
4
1
=
(1 + 1 )(1 + 2 )
4
1
=
(1 1 )(1 + 2 ) .
4
N1 =
(20.1)
N2
(20.2)
N3
N4
(20.3)
(20.4)
(The superscript es have been dropped for notational simplicity here and throughout the
remaining sections.) When plotted over the parent element these shape functions have a
pagoda shape. The displacement field for the element is given by
uhi (1 , 2 ) =
4
X
Na (1 , 2 )uia .
(20.5)
a=1
When fully expanded the two displacements can be written as

u1 (1 , 2 ) = 10 + 11 1 + 12 2 + 13 1 2
u2 (1 , 2 ) = 20 + 21 1 + 22 2 + 23 1 2 ,
(20.6)
(20.7)
where the s are constants that are expressible in terms of the nodal displacements.
Remarks:
1. Along the element edges 1 = 1 or 2 = 1 the displacement field is a linear function
just as it was for the CST triangle.
2. In the interior of the element the interpolative power of the element is greater than the
CST triangle because of the presence of the 1 2 term. In terms of Pascals triangle,
one has the terms shown in bold shown below.
1
1
12
2
1 2
(20.8)
22 .
Thus the element is complete to first order and contains some but not all second order
behavior (the bilinear part).
3. The approximation power of the strain field is
u1,1
11 + 13 2
=
.
u2,2
22 + 23 1
=
u1,2 + u2,1
12 + 21 + 13 1 + 23 2
(20.9)
60
S. Govindjee
Thus the element can exactly capture strain fields in which 11 is a linear function of
2 , 22 is a linear function of 1 , 12 is a linear function of 1 and 2 . This is a vast
improvement over the CST triangle with only a modest increase in cost (1 additional
node). Note, that when the element is distorted in the physical domain (not all interior
angles equal to 90 degrees) the interpolative power of the strain field (in terms of the
physical coordinates) changes. Nonetheless its superiority over the CST holds for many
practical situations.
Another common element in the quadrilateral hierarchy is the 9-node quadrilateral (the
quadratic quad). It is based on the 9 terms shown in bold below in Pascals triangle.
1
21
13
14
1
21 2
13 2
2
1 2
21 22
1 22
22
(20.10)
23
1 23
24 .
Thus in an undistorted case the displacement interpolation can exactly represent fully
quadratic fields and some cubic and quartic behavior. The strain behavior is similarly
enhanced over the 4-node quad. In practical terms, this means that fewer 9-node quads are
needed to capture a given spatially varying behavior than one would need 4-node quads.
The parent domain for the 9-node quadrilateral is shown below with the standard node
number scheme.
x( )
2
(-1,1)
4
8
(1,1)
3
1
(-1,-1)
2
5
(1,-1)
The corresponding shape functions can be generated in several ways. The primary methods
are (1) as shown in Table 6.6.1 in CMP[5] or (2) through direct application of Lagrange
61
interpolation formulae applied to the parent domain. In either case the net result is:
N1 =
N2 =
N3 =
N4 =
N5 =
N6 =
N7 =
N8 =
N9 =
1
(1 1 )(1 2 )1 2
4
1
(1 + 1 )(1 2 )1 2
4
1
(1 + 1 )(1 + 2 )1 2
4
1
(1 1 )(1 + 2 )1 2
4
1
(1 12 )(1 2 )2
2
1
(1 + 1 )(1 22 )1
2
1
(1 12 )(1 + 2 )2
2
1
(1 1 )(1 22 )1
2
(1 12 )(1 22 ) .
(20.11)
(20.12)
(20.13)
(20.14)
(20.15)
(20.16)
(20.17)
(20.18)
(20.19)
P
Note that the shape functions are unity at their respective nodes, Na ( b ) = ab , and
Na =
1. Further note that the displacement behavior on the edges of the element is quadratic in
terms of the isoparametric coordinate of the edge in question.
A third common quadrilateral element is the 8-node serendipity element. It is similar to

the 9-node quad but does not contain the term 12 22 in the interpolation field and is missing
the center node of the element. The shape functions for this element can not be generated
through a direct application of the multiplication
P of Lagrange interpolation polynomials;
doing so results in incomplete shape functions
Na 6= 1. Instead the following method is
applied: (1) write the bilinear shape functions for node 1, 2, 3, and 4, (2) use Lagrange
interpolation polynomials for the nodes 5, 6, 7, and 8, (3) correct the shape functions for
nodes 1, 2, 3, and 4 by subtracting off the value for the bilinear shape functions at the
62
S. Govindjee
mid-side nodes multiplied by the mid-side node shape functions:

1
1
(1 1 )(1 2 ) (N8 + N5 )
(20.20)
4
2
1
1
N2 =
(1 + 1 )(1 2 ) (N5 + N6 )
(20.21)
4
2
1
1
N3 =
(1 + 1 )(1 + 2 ) (N6 + N7 )
(20.22)
4
2
1
1
N4 =
(1 1 )(1 + 2 ) (N7 + N8 )
(20.23)
4
2
1
N5 =
(1 12 )(1 2 )
(20.24)
2
1
N6 =
(1 + 1 )(1 22 )
(20.25)
2
1
N7 =
(1 12 )(1 + 2 )
(20.26)
2
1
N8 =
(1 1 )(1 22 ) .
(20.27)
2
The above shape functions are seen P
to have the desired properties of being unity at their
Na = 1. The displacement field is quadratic on the
respective nodes, Na ( b ) = ab , and
element edges as was the case for the 9-node quadrilateral.
N1 =
21
Mesh Refinement with Higher Order Elements
The error estimates derived previously show that the error is bounded by terms of order hp
where h is the characteristic size of the element (radius of a circumscribing circle) and p is
a constant that depends on the order of the element. Thus to reduce the error, one either
reduces the element size (h-refinement) or one increases the order of p (p-refinement). Methods for grading meshes for h-refinement have already been discussed. Here issues associated
with grading for p-refinement are discussed.
Consider the transition from bilinear interpolation to quadratic interpolation. In this
case one would like to join a 4-node quad to a 8-node quad. If it is done as shown in the
figure below (where the node numbers are the global node numbers) it will result in an
incompatible displacement field.
4
6
9
Join
10
8
7
1
Element 1
2
Element 2
63
This violates one of the conditions needed for convergence of the finite element method and
will lead to un-trustworthy and likely un-physical predictions of behavior. To see that this
will give incompatible behavior consider the displacement field along the shared edge. When
the displacements are viewed from element 1 they will be linear functions that depend on
the nodal displacements at global nodes 2 and 3. When the displacements are viewed from
element 2 they will be quadratic functions that depend on the nodal displacements at global
node 2, 3, and 10. Except in special cases, where the displacements at node 10 satisfy
certain restrictions, these two functions will not coincide. Even a connection of the form
shown below will lead to a similar result. Elements 1 and 2 still have only linear edges and
element 3 has a quadratic edge.
3
Element 2
10
11
Join
9
Join
11
10
Element 3
10
10
Element 1
Possible deformation mode of shared

edge showing incompatible gaps
Connecting linear edges to higher order edges almost always leads to incompatible displacement fields along the common edges.
To correctly transition from linear to quadratic elements requires the use of transition
elements. For instance to transition from 4-node quads to 8-node quads, one would use a
5-node transition element as shown below.
[Node numbers are global node numbers]

4
Join
Element 2
12
Join
Element 1
Element 3
11
10
This element has only one quadratic edge with the remaining edges being linear. The shape
functions for this transition element are given in parent domain coordinates through the
same procedure used to construct the 8-node quadrilateral shape functions; the direct use of
Lagrange interpolation formulae can lead to shape functions that do not satisfy completeness
when constructing general transition elements for the transition element being presented,
64
S. Govindjee
however, this is not the case.

1
(1 1 )(1 2 )
4
1
=
(1 + 1 )(1 2 )
4
1
=
(1 + 1 )(1 + 2 )
4
1
=
(1 1 )(1 + 2 )
4
N1 =
N2
N3
N4
(21.1)
1
N5
2
1
N5
2
(21.2)
(21.3)
(21.4)
1
N5 = (1 + 1 )(1 + 2 )(1 2 )/(1 + 1)(1 + 0)(1 0) = (1 + 1 )(1 22 ) .
2
(21.5)
Note, it is easily verified that edges 1-4, 1-2, and 3-4 are linear and edge 2-5-3 is quadratic
(see figure below). Further note the quadratic edge only depends on nodal displacements at
local nodes 2, 3, and 5.
x( )
2
(-1,1)
4
(1,1)
3
5
1
(-1,-1)
2
(1,-1)
22
65
Stress Recovery Methods
After solving for the nodal displacements from Ku = F , the stresses can be determined
in several different ways. The most obvious (and consistent) way is to calculate the strains
from = Bu and then employ the constitutive relations = D. Such a method, though
simple, has several drawbacks. These detractions emanate from the fact that one typically
uses C 0 finite element approximations. The C 0 formulation gives discontinuous strain fields
as may be recalled from the discussion on the 1-D rod. As a more extreme example of
what can occur when recovering stresses in this fashion, consider a cantilevered beam with
a distributed load that is modeled with 8-node serendipity elements; see figure.
Below the beam in the figure, the shear strains are plotted along the neutral axis. The
exact solution is linear. The 8-node elements produce shear strains that vary quadratically.
At some locations in the elements the strains and hence the stress are even of the incorrect
sign. Note, however, that as in the 1-D rod problem, there are points in the elements at
which the strains are exact. The points where the strains are exact in an element are often
called Barlow Points.
Remarks:
1. Except in the simplest cases, it is impossible to predetermine the points at which the
strains(stresses) will be optimal (strain error equal to zero).
2. For undistorted 1-D elements, the property of nodal exactness for the displacement field
can be used to approximately locate these points. By approximately, it is understood
that the points correspond to the location where the leading term in the strain(stress)
error is zero. As an example, consider the 1-D linear element where the approximate
displacement field is of the form:
uh = 0 + 1 .
(22.1)
66
S. Govindjee
Since the element can exactly represent linear fields, errors will only be generated by
quadratic and higher order exact solutions. The leading term in the error will usually
be from the next lowest order term; i.e. the quadratic term. To that end, assume an
exact solution of
u = 2
(22.2)
and determine where the strain error is zero. Nodal exactness gives:
0 + 1 = 1
0 1 = 1
(22.3)
(22.4)
Solving for the s gives 0 = 1 and 1 = 0. Thus the strain error

d
d
e = (u uh ) = 2 1 = 2 0 ,
d
d
(22.5)
and the point where this is zero is at = 0 the center of the element. To determine
these points in a 1-D quadratic element, one would have to assume an exact field of
the form u = 3 .
3. For rectangular elements, the optimal stress points are usually associated with the
Gauss Quadrature rule of 1 order less than that required for full integration of an
element. For the 4-node element, 2 2 quadrature is required for full integration; thus,
the point of the 1 1 rule is the optimal stress point. For the 9-node element, 3 3
quadrature is required for full integration; thus the points of the 2 2 rule are the
optimal points. Once the elements are distorted these points are no longer optimal,
but are still good choices for locations to sample the stresses.
4. Most programs report stresses in one of a number of different ways. Usually an algorithm is used to smooth the stresses before plotting them; very rarely are the raw
stresses reported directly to the user. In FEAP, the raw element stresses can be
examined with the estr command in the Plot phase. The stre command in the Plot
phase involves the use of a stress smoothing algorithm.
As an example of a stress smoothing algorithm, considers FEAPs. After the displacement field has been solved for, the raw element stresses are calculated at the Gauss points
from the formula = DBu. These stresses are smoothed by generating a continuous
approximation to the raw element stresses. The smooth field is chosen to have the following
form
(x) = N (x)
(22.6)
are nodal stress values. The key to the
where (x) is the smoothed stress field and
algorithm is the determination of these nodal stress values. This is done by minimizing the
square of the difference between the smoothed and raw stress fields; i.e. the nodal values are
found from the following problem
Z
1
=
( )2 d .
(22.7)
min g()
2
67
Note that g() is a scalar valued function with a vectorial argument (the vector of nodal stress
unknowns). To find the minimum, take the variation of (22.7) in an arbitrary direction
Doing so gives
and compute g/ .
Z
g
=
N T ( ) d .
(22.8)
Setting this last result equal to zero and rearranging yields:

Z

Z
T
=
N N d
N T DBu d .
(22.9)
By defining the right-hand side above to be the vector P and the expression in the parenthesis
on the left-hand side to be the matrix M , the unknown smoothed nodal stresses are found
to be
= M 1 P .
(22.10)
These values are used in conjunction with (22.6) for plotting purposes.
Remarks
1. The preceding algorithm involves a considerable amount of manipulation of the FE
stresses fields; (most stress smoothing algorithms are of this nature). Fortuitously or
not, the smooth stress fields are usually better approximations to the actual stress
fields than the raw fields.
2. In the limit of proper mesh refinement, the raw stress fields should approach the
smoothed stress fields. This provides an excellent check on mesh refinement. The
raw fields can be compared to the smoothed ones, and refinement can be halted once
the two fields are within a certain tolerance of each other or similarly when the interelement stress jumps in the raw fields decreases below a certain tolerance.
3. A popular simplification to the algorithm above (and the one used in FEAP) is to
diagonalize the M matrix also known as lumping or mass lumping. By diagonalizing
the M matrix, the inversion shown above is trivial. There are many lumping schemes
employed in finite elements, but the most common is a simple row sum method. In
this method, a diagonal entry of row j is replaced by
n
X
Mjk
(22.11)
k=1
with all the other entries in the row set to zero afterwards.
4. This kind of stress smoothing is inappropriate at abrupt changes in modulus, thickness,
etc.; i.e. wherever actual stress discontinuities occur in the continuum equations.
68
S. Govindjee
23
Introduction to the Patch Test
The patch test is an evaluation of the validity of an elements formulation and implementation. In its original form [2], it allowed analysts to determine whether the elements they
were working with satisfied the completeness condition for convergence. In a later reincarnation, it also allowed one to determine the stability properties of an element [19]. This last
property is related to the convergence (or lack thereof) of a finite element mesh. When the
finite element integrals are not computed exactly (i.e., when using quadrature) the stability
of the element is an added requirement2 for convergence on top of the completeness and
compatibility conditions; see Laxs Equivalence Theorem [14, 3.5] for a similar situation
that arises in finite difference methods. For developers of new finite elements, this test is an
essential part of the development of good elements; for analysts and users of finite element
codes, this test is an essential part of the verification process of the validity of an analysis
based upon a certain element. This verification is the responsibility of the engineer NOT
the code developer.
In what follows, the most basic version of the patch test is presented for problems involving single derivatives in their weak forms (m = 1) such as linear elasticity and heat
conduction. The two parts of the test involve (1) the testing for completeness and (2) the
testing for stability. To provide a concrete framework for the discussion, consider 2-D plane
elasticity problems3 . For this class of problems, elements must be linearly complete in both
vertical and horizontal displacements. i.e., for arbitrary values of i , such elements must be
able to exactly represents fields of the following form:
u1 = 0 + 1 x1 + 2 x2
u2 = 3 + 4 x1 + 5 x2 .
(23.1)
(23.2)
To test an element for this property, one must individually test for the correct presence of
each constant i ; thus, six tests will be involved in this part of the patch test.
Begin by constructing a non-regular mesh as shown below for the case of 4-node quads.
(The irregular mesh is important because some elements pass the patch test for regular
meshes but not for distorted meshes.)
For mixed element formulations the stability test is required even for exactly integrated elements.
For problems posed in curvilinear coordinates such as axis-symmetric problems, the requirements for
the patch test are slightly different; their functional meaning, however, remains unchanged [19].
3
69
4
6
5
The test proceeds by imposing displacements on nodes {1 4} and {6 9} in accord with

a linear field from (23.2) and checking that (a) the displacement at node 5 is correct and
(b) the strains in the elements are correct. As an example of a test for the term 0 : Impose
on the boundary nodes the displacement u1A = 1 and u2A = 0; i.e., 0 = 1 all other i s
zero. Then check that (a) the displacement at node 5 is u15 = 1 and u25 = 0 and (b) the
element strains are all zero. As an example of a test for the term 4 : Impose the following
displacements on the boundary nodes u1A = 0 and u2A = 1 x1A , where x1A is the onedirection coordinate value for node A; i.e., 4 = 1 all other i s zero. Then check that (a)
the displacement at node 5 is u15 = 0 and u25 = x15 and (b) the element strains are 11 = 0,
22 = 0, and 12 = 1/2. Tests for the other i s can be constructed in a similar manner.
Remark:
1. A common third check that is added to this part of the patch test is a test on the
reactions. In this check, the nodal reactions are checked for exactness. This can be
done by taking the exact strain values multiplying by the elastic moduli to get the
exact stresses and then integrating on the element boundaries to get the exact nodal
reactions. These values are then compared to the nodal reactions reported by the FEM
code. This check ensures that the element handles tractions correctly.
The second part of the patch test involves checking the stability of the element. Here, one
calculates the eigenvalues of a single element and verifies that the number of zero eigenvalues
exactly matches the number of rigid body modes. For a plane elasticity element there should
be 3 modes with zero eigenvalues; these three modes correspond to rigid translation along the
x1 and x2 axes and rigid rotation about the x3 axis. For example, with a 4-node quadrilateral
the element stiffness matrix is an 8 8 matrix. For a properly formulated element there will
by 5 non-zero eigenvalues (volumetric expansion, uniform elongation, shear, bending mode
1, and bending mode 2) and three zero eigenvalues. The eigenmodes(vectors) are shown
below for all 8 cases of a 4-node rectangular element.
70
S. Govindjee
Rigid motions
T_2
T_1
R_3
S
V
UE
Non-Rigid motions
B1
B2
Remarks:
1. If we use the formula from Strang and Fix for the required quadrature rule for the 4
node quadrilateral we have k = 2 and k = 1; thus k+k2m
= 1. This implies that only
linear polynomials need to be integrated exactly; therefore only 1-point quadrature is
required. In 2-D problems, this translates into a single Gauss point located at (0, 0)
with a weight of 4. This element is very attractive from a cost standpoint being
4 times less costly than a fully integrated element (discussed in the next remark).
However in terms of the patch test it fails on the stability check. For part (1) of the
test, (a) and (b) check exactly; the test on reactions also checks exactly. However, for
part (2) of the patch test, we find that the element has 5 zero eigenvalues. The extra
two eigenvalues are associated with the two bending modes of the element. These
two modes are called zero-energy modes because they correspond to motions of the
element that can take place without strain energy (E = 21 uT ke u = 0). Their presence
introduces the possibility of spurious mechanisms in an analysis; for an example see
the figure below.
71
2. A fully integrated 4-node quadrilateral element has only 3 zero eigenvalues those of
the rigid body modes. Using a 2
2 Gauss
quadrature rule will produce this situation;
i.e. Gauss points located at (1/ 3, 1/ 3) with weights equal to 1. This element
passes all parts of the patch test.
3. Because of the cost savings, researchers have worked on the stabilization of the 1-point
element. Methods of this sort are termed hourglass stabilization methods. Hourglass
stabilized elements often perform better than fully integrated elements both in cost
and accuracy. The accuracy enhancement stems from the fact that the eigenvalues
of a fully integrated element grow under element distortion making the element too
stiff in response to bending modes. Stabilized 1-point elements are less susceptible
to this problem. The problem with hourglass stabilized elements is that they do not
completely eliminate the hourglass modes; thus, an analyst must carefully watch for
hourglassing in a solution. Solutions containing hourglassing should not be relied upon.
A conservative analyst would just stay away from such elements.
4. As another example, consider the 8-node quadrilateral. For this element k = 3 and
k = 2; thus, k + k 2m = 3. This implies that cubic polynomials need to be integrated
exactly. A 2-point Gauss rule will accomplish this. When extrapolated to 2-D this
becomes a 2 2 Gauss quadrature rule. This element exactly passes the three checks
in part (1) of the patch test (displacements, strains, and reactions). In part (2) of
the patch test we find 4 zero eigenvalues. The extra eigenvalue is associated with the
eigenmode shown below. This eigenmode is termed a non-communicable zero-energy
mode because in an assemblage of two or more elements this mode can not appear like
the hourglass modes did in the 1-point 4-node quadrilateral. This statement, though
often made, is not entirely correct. The mode can appear when one has meshes where
elements have vastly different material properties. To remove the mode, one needs to
use a 3 3 quadrature rule.
72
24
S. Govindjee
Element Performance in Bending
By looking at the performance of elements in bending we can gain an understanding of their

behavior in this technologically important area but more importantly we can gain insight
into the general performance of these elements. This understanding leads us to ways of
properly utilizing these elements and ways of improving their performance. Consider first
the problem of pure bending (no shear) on a beam of depth 2c as shown below.
x2
l
2c
x1
For the above problem choose M such that u1 (l, c) = u where u is a given quantity. For
such a moment the exact solution for the displacement field in the elasticity equations is
u
x1 x2
cl
u
=
(x21 x22 ) .
2cl
u1 =
(24.1)
u2
(24.2)
The corresponding exact strain field is given by

u
x2
cl
u
= x2
cl
= 0.
11 =
(24.3)
22
(24.4)
12
(24.5)
Using this solution we can examine the various elements for their ability to model this type
of solution.
The CST: To examine the constant strain triangle consider a single triangle as shown
below in a state of pure bending.
x2
l
2c
x1
Proceed by imposing at the nodes the displacements from the exact solution and examine
how well the element reproduces the exact strain field. This is calculated from the relation
73
= Bu, where B is the strain displacement operator for the CST:

11 = 0
22 = 0

1
c
12 =
2 u
c
4l
(24.6)
(24.7)
(24.8)
As can be seen by comparing these expressions to the exact strains, they are very poor. This
is one of the primary reasons that CST elements are not thought to be good elements for
stress analysis.
Note that this result does not imply anything about what happens if the domain is broken
up into many CSTs. In fact, through the use of a sufficient number of CSTs one can model
bending fields to any desired degree of accuracy. As an example consider a cantilever beam
with a shear load on the end. In the case that the load is applied as a parabolically varying
shear traction and the built-in end is allowed to warp, an exact solution to the elasticity
equations is available for comparison (see e.g. [20, Art. 21]).
B
h
A
4h
For comparison purposes assume a Poissons ratio of = 0.25 and consider the vertical
displacement at the point A and the bending stress at the point B. If the beam is modeled
by 128 CSTs the deflection error is 14.1% and the stress error is 14.6%; this model involves
160 degrees of freedom. If the beam is modeled by 512 CSTs the deflection error reduces to
3.9% and the stress error reduces to 4.4%; this model involves 576 degrees of freedom.
Remarks:
1. The ability of the CST to model the beam is discouragingly low. A large number of
elements are required to reduce the error to an acceptable level. The main difficulty
of the element can be seen from the polynomial nature of the exact solution to this
problem. The exact solution for the displacement field involves the following monomials: x3 , x2 y, xy 2 , y 3 , y, x, 1. The CST only has the following monomials: x, y, , 1.
Because of the high degree of mismatch many elements are needed to adequately model
the problem. In terms of our notation from 2, we can say that uexact is far from S h
until h becomes quite small. One way of improving the situation is to employ elements with better interpolatory power. If one uses the LST (6- node triangle which
has monomials through second order), then with 32 elements the displacement error is
only 0.4% and the stress error is just 1.4%; this model involves 160 degrees of freedom.
2. These performance measures tell us more than just how this element behaves in bending. It tells how the element is likely to behave in situations with stress fields that have
74
S. Govindjee
monomials similar to that of bending. For instance, we would expect that problems
that involve exact solutions with polynomial order 3 will be just as poorly approximated by CSTs even though the exact nature of the variation may not be the
same.
The 4-node quadrilateral Q4: We can perform the same kind of investigation on the
Q4 element. In this case consider a single element of depth 2c and length 2l. At the nodes
impose the exact displacements from the theory of elasticity solution. This results in a FEM
displacement field of
u
x1 x2
(24.9)
cl
u
u2 = (l2 + c2 ) .
(24.10)
2cl
Note that the element exactly captures the 1-direction displacements but fails to capture the
2-direction displacements. The resulting strain field from the element is given by
u1 =
u
x2
(24.11)
cl
22 = 0
(24.12)
u
12 =
x1
(24.13)
cl
As a result of the exact expression for u1 the bending strain in the element is exact. The
other strains are however incorrect; thus, to model say the cantilever beam used in examining
the performance of the CST many elements would be needed. Note this number would be
less than that for the CST but nonetheless more than would seem necessary from a cursory
look at the problem.
As another measure of element performance in bending, we can examine the ratio of the
moment needed to produce the displacement u in the Q4 to the exact moment from the
theory of elasticity. This ratio is given by
"
2 #
M
1
1
1 l
=
+
,
(24.14)
Mexact
1+ 1 2 c
11 =
where the RHS is computed by taking the ratio of the strain energy in the exact solution to
the Q4 strain energy for the same amount of displacement at the corner nodes.
Remarks:
1. In (24.14) the ratio l/c is known as the aspect ratio. From the equation we now see
why large aspect ratios lead to poor results. If = 0.3 then for aspect ratios 1, 2, and
4 the moment ratios are 1.5, 2.6, and 7.3 respectively.
2. Again it is remarked that if uexact is far from S h , then poor results will ensue.
3. As a matter of terminology, the spurious 12 in the FEM solution is called a parasitic
shear and the large stiffness of the elements in the limit of large aspect ratios is
referred to as locking.
75
4. Note that as before if h is small enough acceptable results can be obtained. For
instance, for the cantilever beam used in the example for CSTS, 64 fully integrated
elements with an aspect ratio of 4 leads to a displacement error at the tip of 10%.
5. There are many remedies for this poor behavior. Some common ones are: (1) use
higher order elements, (2) use under integrated elements with stabilization, or (3) use
a Wilson-Taylor incompatible modes element the QM6 [18] which is a particular
example of an enhanced strain element [15]. This last method is discussed in the next
section. These are not the only options for circumventing poor element performance in
bending; in later sections we will look at other methods that were designed for other
purposes that also help with the bending problem.
25
Incompatible Modes
The Wilson-Taylor incompatible modes element is the first element we will examine that
breaks some of the mathematical requirements that we setup for convergence of a finite
element method. This however does not mean that this element does not converge; rather
it means that the previous convergence theorems merely do not apply to such elements. For
our purposes we will only look at the formulation of these elements to gain an understanding
of what they do; we will not try and prove that they converge. Convergence proofs of such
elements are beyond the scope of this course. It is noted however that the element does pass
the patch test (an indicator of consistency and stability); interested readers are directed to
[15] (and references therein) where such techniques are interpreted with the aid of a 3-field
variational principle as in [7] or [21].
The requirement that this element breaks is the compatibility requirement. Which was a
condition that we imposed for the validity of the decomposition of integrals over the domain of
a body into integrals over individual elements. Without this condition the displacement field
is incompatible between elements; i.e. gaps form between elements. These gaps, however,
can be shown to tend to zero in the limit of mesh refinement. The nature of the element
and the reason for its success are best seen by examining the elements formulation.
The basic idea behind the QM6 element is the desire for a 4-node quadrilateral element
that behaves better in bending than the Q4 yet is less expensive than the 9-node quadrilateral
the Q9. In order to improve the performance in bending the monomials associated with
pure bending that are missing from the Q4 are added to it but no new nodes are added to
the element. The new form of the displacement field is given by
ui =
4
X
a=1
Na uia +
6
X
Na ia .
(25.1)
a=5
The uia are the nodal displacements and the ia represent internal element parameters that
are not shared between elements. For a {1, 2, 3, 4}, Na represent the standard bilinear
shape functions from the Q4 element. For a {5, 6} we define
N5 = 1 12
N6 = 1 22 .
(25.2)
(25.3)
76
S. Govindjee
These two expressions contain the monomials that are missing from the Q4 element in terms
of the pure bending solution.
Because the internal element parameters ia are not shared between elements, they can
be solve for on a local element level without assembling global equations. This process of
solving for element level variables on the local level before assembling global equations is
known as static condensation. To begin, look at the element stiffness matrix
Z
e
k =
B T DB ,
(25.4)
e
where
B=
B1 B2 B3 B4 B5 B6
312
The entries of the matrix above have the usual definition of
Na,x
0
B a = 0 Na,y
.
Na,y Na,x 32
(25.5)
(25.6)
The B matrix also serves as the strain-displacement operator; i.e.

u
=B
.
(25.7)
Note that the internal parameters are included in (25.7). For later use also note that the
element stiffness can be partitioned in the following form

kuu ku
e
k =
,
(25.8)
ku k
where kuu is the standard element stiffness for the Q4 element,
Z

T

T
B1 B2 B3 B4
ku = ku =
D B5 B6
(25.9)
and
k =
Z
e
T
Na b i +
B5 B6
B5 B6
(25.10)
The element loads are defined as

fia =
Z
e
Na ti
(25.11)
te
for a {1, 2, 3, 4}. There are no element loads on the internal degrees of freedom for the
element; i.e., it is assumed that fia = 0 for a {5, 6}. There are several reasons why this
choice which violates our previous rules of element construction is a good choice. The first
and perhaps simplest is the desire to maintain the situation where the sum of the element
loads is equal to the total load imposed on the element. Since this is already the case for the
Q4 element, the loads represented by Equation (25.11) already have this property; therefore
77
to ensure this state of affairs in general, we have to have the loads associated with the internal
degrees of freedom be equal to zero. As a second consideration, we do not want to have to
input internal element loads in an input file; this would greatly take away from the desired
simplicity of the element. This requirement can also be shown more rigorously by appealing
to enhanced strain methods [15].
The process of static condensation continues by writing out the element equations

u
f
e
k
=
(25.12)
o
and separating the last four equations from the first eight; i.e.
ku u + k = o .
(25.13)
This equation is then solved for the internal element parameters as

= k1
ku u .
(25.14)
This result is then inserted into the first eight equation for the element
kuu u + ku = f .
(25.15)

kuu ku k1
ku u = f .
(25.16)
This yields the final result for the
Remarks:
1. The expression in the parentheses represents the stiffness of a Q4 element that has been
modified slightly. As an intuitive interpretation of the new element stiffness, we can say
that it will behave in a more flexible manner than the Q4 since we are subtracting
from a Q4 stiffness matrix.
2. What has been described so far is known as the Wilson Q6 element [22]. It passes
the patch test only in rectangular and parallelogramic configurations and certain other
special cases. For arbitrary quadrilateral shapes it fails the consistency part of the
patch test. The element that passes the patch test for all configurations is a slight
modification of the element presented above. This element is known as the WilsonTaylor QM6 element [18].
To see the problem with the Q6 element consider the patch test. If the uia are set
in accord with a linear polynomial then the incompatible modes ia should not be
activated since the Q6 is built on top of the Q4 element which already passes the test.
Thus during a patch test, should be o. This implies that
ku u = o
(25.17)
78
S. Govindjee
during the patch test. If we expand this last equation, we have
Z

T

B5 B6
D B1 B2 B3 B4 u = o ,
|
{z
}
e
(25.18)
constant stress
where the last three items on the LHS of (25.18) represent a constant stress during a
patch test. This leads to the requirement that
Z
B a = o a {5, 6} .
(25.19)
e
This last requirement is satisfied by the Q6 element only when it a rectangle or parallelogram. Taylor proposed [18] that certain terms in the B-matrices for a {5, 6}
always be evaluated at the center of the element; B 5 is replaced by a B Taylor
where the
5
derivatives of the isoparametric mapping that appear from the chain rule are evaluated
at the element center and similarly for B 6 . This is a type of mixed 1-pt integration
rule for the incompatible mode terms and it allows for the satisfaction of (25.19) for
arbitrary shaped quadrilaterals. The element with this modification is known as the
QM6.
3. Such methods violate our rigorous conditions for convergence. Strang and Fix called
such techniques variational crimes because of these violations. Precise mathematical
justification for many methods of this type is however possible. These considerations
are however beyond the scope of this course.
26
79
Incompressibility
In linear elasticity the response of a material becomes incompressible in the limit as the
Poissons ratio 12 . This limit is important in the modeling of elastomeric materials and
other polymeric solids. Additionally, incompressible material response happens even when
the actual Poissons ratio of a material does not approach 12 ; typical case and very important
case is when one has plastic flow in metals. For such problems, the finite element formulation
that we have been working with (the displacement based standard Galerkin formulation) has
several well known difficulties. These difficulties arise from the fact that computers perform
calculation using a finite number of digits. The difficulty can be easily seen by considering
Frieds argument [6].
As given below, Frieds argument applies to the case of plane strain linear elasticity.
(Remember this is merely an argument and not a rigorous proof.) The argument begins by
decomposing the stiffness matrix of a finite element formulation into two parts as follows:
Z
Z
Z
T
T
(26.1)
K = B DB = B D B = B T D B ,
|
{z
} |
{z
}
K
K
where in 2-D plane strain
2 0 0
D = 0 2 0
0 0 1
and
(26.2)
1 1 0
D = 1 1 0 .
0 0 0
The Lame parameter is related to the Poissons ratio by =

global finite element equations
(K + K ) u = F .
(26.3)
2
.
12
Consider now the

(26.4)
1
,
2
As
. Thus in a finite precision setting when the two stiffness matrices are
added the information in K is lost and one is effectively left with the equations
K u F .
(26.5)
As 12 , the terms in K are heading to infinity and F is staying finite; the only possibility
for u is for it to approach o when the stiffness is positive definite.
Remarks:
1. This state of affairs is known as incompressibility locking.
2. The above argument does not apply to plane stress problems since in plane stress
the elasticity matrix does not have unbounded terms in the incompressible limit. For
reference, in this case,
1
0
E
.
1
0
D=
(26.6)
1 2
1
0 0 2 (1 )
80
S. Govindjee
3. The severity of the problem can be illustrated by considering a cantilever beam problem. Consider the case where the root section is permitted to warp and the beam is
subjected to a parabolically varying end shear. Let the beam have an aspect ratio of 4
to 1 and model it with 64 fully integrated Q4 elements with the same aspect ratio. For
the case of = 0.3 the tip displacement will be 90.4% of the exact elasticity solution;
for the case of = 0.499 the tip displacement will be only 33.4% of the exact elasticity
solution. This example is taken from [9, 4.4].
4. The difficulty can be seen in a variety of ways. A second illustrative example that
is more qualitative than the previous one is as follows [9, 4.3.6]: Consider the mesh
below that is constructed from CSTs.
F
If one looks at element 1 and considers incompressible motions only, then the possible
motions of the element are very restricted. First the two nodes on the base are fully
restrained. Next the node marked in the center of the mesh can only move horizontally
left or right. If it were to move vertically it would change the area of element 1
( 12 base height). If one looks at element 2 which is attached to the same node, then
a similar argument shows that the node in question can only move vertically and not
horizontally. To reconcile both of these requirements the node simply doesnt move,
u = o. This effect propagates throughout the mesh and leads to locking. Note that
this argument only applies to plane strain where 33 = 0; in plane stress the elements
can contract or expand perpendicular to the page to accommodate the motion.
There are several possible remedies for the difficulties illustrated above. Many of the
solutions that will be discussed were originally very ad hoc in nature and were primarily
numerical tricks. Over time some of these tricks have been given sound mathematical basis
in limited circumstances. In what follows, 5 different techniques from the many proposals are
presented. These comments are more of a general nature and interested readers are directed
to more comprehensive references [9, 23] and references therein.
81
1. Selective Reduced Integration (SRI). This numerical technique works for the case of
isotropic elasticity and is based on Frieds argument. The idea is as follows, when the
K tends to infinity and the stiffness remains positive definite, then the only possibility
for the displacements is for them to head to zero. In the case where K is singular,
the displacements can take values in the null space of K , where the null space of
K is defined as N (K ) = {u | K u = 0}; i.e. the displacements can have non-zero
values as the stiffness grows. The method of SRI exploits this situation by intentionally
lowering the rank of K . This is done in the simplest of fashions by under integrating
K . For the Q4 element, SRI means that one uses a 2 2 rule on K and a 1 point
integration rule on K . For the Q8 and Q9 elements, SRI means that one uses a
3 3 integration rule on K and a 2 2 integration rule on K . This rather ad hoc
method gives very good results for many cases. But because of its ad hoc nature it
has several problems and should be used with caution; interested readers are referred
to the previously given references and citations therein. As an example, consider the
cantilever beam used above. When using SRI on the Q4 element for = 0.3 the tip
displacement is 91.2% of exact and when = 0.499 the tip displacement is 93.7% of
exact.
2. The method of 1-point integration on the entire stiffness is also an option for this
problem. As discussed earlier this element must be used with hourglass control.
3. The method of incompatible modes, though not designed to combat incompressibility
problems, also shows reasonable performance here. In reference [23, 11.10], it is shown
that for a beam in pure bending the QM6 element with 22 and 33 quadrature gives
an exact result for the tip displacement in the undistorted configuration independent
of Poissons ratio. With distorted elements, one gets slightly less displacement than
the exact solution independent of Poissons ratio up to about a value of = 0.48 for
2 2 quadrature; above this value the displacement response dips to 83% of exact
in the limit 0.5. This reference also shows that if 3 3 quadrature is employed
on distorted QM6 elements that they lock in the incompressible limit. Qualitatively,
one can say that incompatible displacement fields are being activated that give the
element added ability to accommodate incompressible motions without having to lock.
The results from the variation of the quadrature rule also point out that a 2 2 rule
is additionally providing under integration of the incompatible modes thus lowering
some of the eigenvalues of the stiffness.
4. A fourth broad category of methods that addresses the incompressibility problem is
mixed methods. In a mixed formulation one interpolates not just displacements but
also stresses and possibly strains. The terminology mixed comes from the fact that
the unknowns are of mixed character; i.e. displacements and stresses. For the problem
at hand, one typically interpolates displacements and pressures. The word pressure
is placed in quotes because the added field isnt always strictly the pressure. The strong
form of such a problem would be find the displacement and pressure fields u(x) and
82
S. Govindjee
p(x) such that
ij,j + bi = 0i
p/ ui,i = 0
(26.7)
(26.8)
where ij = pij +2ij , ij = 21 (ui,j +uj,i ), and and are Lame parameters subject
to the usual displacement and traction boundary conditions. p is thought of as another
variable to be solved for and is equal to the pressure only in the limit as / .
The equation (26.8)2 provides the additional equation needed to calculate p. Thus in
the weak form, there will be two equations that need to be satisfied. And in the finite
element formulation we will need shape functions for p.
A crucial observation is that p appears in the differential equation (26.8)2 without
any derivatives. Thus in the weak form for Eq. (26.8)2 there will be no derivatives
on it either. This implies that we will only need continuity inside elements for approximations to p; i.e. there are no inter-element continuity requirements on p when
looking at convergent finite element schemes (according to our previously delineated
requirements). For this reason pressure nodes are typically interior to elements and the
pressure degrees of freedom are usually solved for on an element level through static
condensation. Common examples of such element are the Q4 element with constant
pressure (i.e. a single pressure node in the center of the element) and the Q9 element
with bilinear pressure where there are 4 interior pressure nodes located at the 2 2
Gauss points. The analysis of such elements is beyond the scope of this course and
interested readers may consult the above mentioned references. It is noted however
that these elements have a much stronger theoretical basis than the previous methods
for dealing with incompressibility. Thus mixed element technology is much preferred
to the other methods presented so far.
method. Formally, it
5. A related method to mixed methods is the so-called B-bar (B)
can be derived from a mixed method but its origins are far more ad hoc. In its original
form it is a pure displacement method [12]. The reasoning behind the element is as
follows. The standard Q4 element effectively attempts to enforce the incompressibility
constraint at each Gauss point in the element (wherever the constitution is evaluated).
In an ideal case, the incompressibility constraint would be satisfied pointwise in the
element. In a finite element however there are only a limited number of degrees of
freedom that are available to try and enforce the constraints. If there are too many
constraint the element will lock; thus if one tries to ensure that 11 +22 = 0 at too many
points in the element then there will not be enough remaining dofs for the element to
deform. The idea behind the B-bar element is to modify the terms associated with
volumetric strain in the element so that the incompressibility constraint is only satisfied
on average in the element and not at each Gauss point (or pointwise for that matter).
The way in which this is accomplished is to split the B matrix into 2 parts, B vol and
83
B dev . B vol is defined so that B vol u is equal to the volumetric strain
u11
N1,1 N1,2
Nn,1 Nn,2 u21
1
1
N1,1 N1,2 Nn,1 Nn,2 ... .

trace[]1 =
3
3
0
0
0
0
u1n
u2n
(26.9)
The deviatoric matrix is defined by B dev = B B vol and merely expresses the fact
that the strain can be additively decomposed into volumetric and deviatoric parts.
The B-bar method proceeds by replacing the volumetric strain operator by its average
= B dev + B
vol , where
value. Thus, one replaces the B operator by B
1,1 N
1,2
n,1 N
n,2
N
N
1,1 N
1,2 N
n,1 N
n,2 .
vol = 1 N
(26.10)
B
3
0
0
0
0
In the expression above
R
a,j = Re
N
Na,j
(26.11)
The net effect of this formulation is that it alters the pressure-volume terms by giving a
coordinate independent volumetric strain and hence pressure. Intuitively, this reduces
the number of pressure-volume constraints on the element. It is remarked that this
is a pure displacement formulation and does not involve pressure interpolations. The
In practical terms, this involves the addition
only thing that happens is that B B.
of only a single loop to a standard element and is thus simple to implement. The
matrix is full whereas the B matrix was sparse. In terms
disadvantage is that the B
of restrictions, this idea works only for linear kinematics but note that it is valid for
all constitutive responses.
84
27
S. Govindjee
Axis-symmetry
For problems that involve geometries that have an axis of symmetry and loads that are
also axially symmetric, we can conveniently formulate them in polar coordinates and nicely
reduce a 3-D problem into a 2-D problem. In polar coordinates, one defines the position of
points by coordinates (r, , z) instead of (x, y, z). The relation between the two systems is
x = r cos()
y = r sin()
z = z;
(27.1)
(27.2)
(27.3)
thus is measured from the x-axis. In polar coordinates, the displacements (ux , uy , uz )
are replaced by (ur , u , uz ), where ur is the displacement in the radial direction, u is the
displacement in the angular or hoop direction, and uz has the same meaning as when using
Cartesian coordinates. Note that these directions are functions of .
In the case of axial symmetry, no solution quantity is allowed to depend on the angular
coordinate . Thus
d
() 0 ;
(27.4)
d
(r, z) with no dependencies on . As a special of axis-symmetry in solid mechanics
i.e. u = u
we will only consider torsion-less axis-symmetry where in addition to (27.4) u = 0. In
this case, we have for the non-zero strains

rr

=
zz =
rz
ur
z
ur
r
ur
r
uz
z
uz
r
(27.5)
Such relations for the common cases of curvilinear coordinates can be found in most basic
books on mechanics; see e.g. Sokolnikoff [16, p. 182-184].
For isoparametric formulations, the displacements and the geometry are both interpolated
using the same parameterization. Thus
ur =
NA (1 , 2 )urA
(27.6)
NA (1 , 2 )uzA
(27.7)
uz =
and
r =
NA (1 , 2 )rA
(27.8)
z =
NA (1 , 2 )zA .
(27.9)
85
Using our previous machinery, we can now quickly construct a B matrix for the case at
hand; viz.
ur1
uz1
..
= Bu = B . ,
(27.10)
urn
uzn
where
B=
and
B1 Bn
BA =
NA
r
NA
r
0
NA
z
0
0
NA
z
NA
r
(27.11)
(27.12)
The constitutional relations for isotropic materials in curvilinear coordinates do not

change from the Cartesian ones as long as the coordinate system is orthogonal. Thus
rr

zz
rz
= D
2 +
2 +
2 +
sym.
(27.13)
0
rr
0
.
0 zz
rz
(27.14)
The equilibrium equation in polar coordinates is still expressed as the divergence of the
stress tensor plus the body forces equals zero. In terms of matrix notation the weak form is
still expressed as
Z
Z
Z
T
T
B DB = N t + N T b
(27.15)
For the volume integrals, we take advantage of the axis-symmetry condition to exactly integrate in ; thus,
Z
Z Z
Z 1Z 1
() = 2
()r dr dz = 2
()Jr d1 d2 ,
(27.16)
where J is the Jacobian determinant of the isoparametric mapping. The usual quadrature
rules can be applied to this last expression. For the surface integral the integral is likewise
exactly evaluated so that
s
2
Z
Z
2
r
z
() = 2 ()r
+
ds ,
(27.17)
s
s
s
where s is a parametric coordinate along the surface of loading in the r z plane.

Remarks:
86
S. Govindjee
1. The factor of 2 in the expressions above is omitted in some codes because is appears
on both sides of the FEM matrix equations. It is important to determine whether or
not the code contains the factor; otherwise, it is possible to obtain result that are off
by a factor of 2 in the stress, strains, and displacements.
2. In performing a patch test on such elements care must exercised in determining what
are the appropriate requirements for such an element. Interested readers are referred
to [19] for details on this matter. Here, it is merely pointed out that in terms of rigid
body modes for a single element there is only one rigid translation in the z-direction.
Thus a Q4 axis-symmetric element has only one zero eigenvalue in its element stiffness
matrix. In practical usage for such elements it also means than one need only restrain
the mesh with a single boundary condition to prevent mechanisms.
3. A similar setup can performed for problems with spherical symmetry by using spherical
coordinates.
28
87
Introduction to Shear Deformable Plates and Beams
In order to analytically or numerically model three (or two) dimensional structures, approximate theories are often employed for reasons of efficiency. In the case of bodies with one
characteristic dimension less than the other characteristic dimensions, plate and beam theories are often utilized. For such theories, we can apply our general rules of constructing
finite element approximations and create FEMs for them. In order to do this, however, the
governing equations are needed. In what follows a brief synopsis of the governing equations
is presented.
The basic idea behind beam and plate theories is to take advantage of the thinness of
the body as shown below.
x3
x3
Beam
Plate
L
x1
x2
t
L
x1
t
The thinness is exploited by making certain assumptions that seem reasonable given the
special geometry of the body. The validity of the assumptions can only be assessed through
solution of the elasticity equations or through physical experimentation. The particular theories to be presented are known as Timoshenko beam theory and Reissner-Mindlin plate
theory. Their non-shear deformable counterparts are the familiar Bernoulli-Euler beam theory and the Kirchhoff plate theory.
The underlying basis for the theories is the construction of assumptions on the character
of the motion of the material. These assumptions are known as kinematic assumptions and
there are two of them. Assumption 1 states that the vertical displacement u3 (x1 , x2 , x3 ) at
any point in the body can be expressed solely in terms of the (x1 , x2 ) coordinates of the
point in question; u3 (x1 , x2 , x3 ) = u3 (x1 , x2 ). Reference is usually made to the displacement
of the neutral axis or surface; see figure below. For beams simply drop the x2 s in all the
expressions here and below.
88
S. Govindjee
x3
u3
Assumption 2 states that for beams vertical plane sections remain plane but not necessarily
perpendicular to the neutral axis and for plates that vertical fibers remain straight but not
necessarily perpendicular to the neutral surface.
x3
This assumption allows one to characterize the motion in the plane of a plate in terms
of two rotation variables 1 and 2 as shown above (for beams in terms of just 1 ); i.e.
u1 (x1 , x2 , x3 ) = x3 1 (x1 , x2 )
u2 (x1 , x2 , x3 ) = x3 2 (x1 , x2 ) .
(28.1)
(28.2)
Note the convention in the figure where Greek subscripts are taken to range over {1, 2}.
Also note that 1 corresponds to a negative rotation about the 2-axis and 2 corresponds to
a positive rotation about the 1-axis.
Remarks:
1. By making the two kinematic assumptions above, for plates, the basic problem in
elasticity of finding, ui (xj ), the three displacements in three variables has been reduced
to a problem where one needs to find one displacement, u3 (x ), and two rotations,
(x ), in two variables. For beams, the problem is reduced to finding u3 and 1 as
functions of x1 . Even though there are still three unknowns in the plate theory the
problem is greatly simplified since the unknowns are only functions of two variables,
not three.

2. These assumptions can be used
ij = 21 (ui,j + uj,i ) one has
11
22
33
212
223
213
89
to examine the form of the strain field. By applying
x3 1,1
x3 2,2
0
x3 (1,2 + 2,1 )
2 + u3,2
1 + u3,1
(28.3)
For beams one only has 11 and 13 . For the non-shear deformable theories, one assumes
that the cross-section or fiber remains perpendicular to the neutral axis or neutral
surface; i.e. that the rotations are equal to the derivatives of the vertical deflection. For
plates, this results in 23 = 13 = 0, 11 = x3 u3,11 , 22 = x3 u3,22 , and 12 = x3 u3,12 .
For beams, this implies that all strains are zero except 11 = x3 u3,11 .
3. For structural theories based on kinematic assumptions, one typically expresses strains
in terms of curvatures. For plates the curvature tensor is defined as
=
1
(, + , )
2
(28.4)
and the two shear strains

3 = + u3, .
(28.5)
For beams there is only one curvature = 1,1 and one shear strain 13 = 1 + u3,1 .
Thus, for example, the strains in a plate can be expressed in matrix notation as
x3 11
x3 22
.
=
(28.6)
2x
3
12
2 + u3,2
1 + u3,1
The third primary assumption in beam and plate theories is that the through thickness
stress 33 = 0. This is clearly not true underneath loads and near supports. However, 33
can be shown to be negligible in comparison to the other stresses in such problems and thus
this actually make a reasonable assumption when one considers it on a basis of comparative
importance. One consequence, however, of this assumption is an inconsistency with the
kinematic assumption that leads to the result that 33 = 0. Here, one is lead to the result
11 +22 )
that 33 = (+2
; this comes directly from the standard 3-D linear isotropic elasticity
equations. Such inconsistencies are common in structural theories and are a consequence of
the nature in which they are developed.
In structural theories one commonally works with moments and shears stress resultants.
For plates the moments are defined by
Z t/2
M =
x3 dx3 .
(28.7)
t/2
90
S. Govindjee
The figure below illustrates these moments; note that they have units of moment per unit
length.
x3
x3
x2
x2
M_21
x1
x1
Q_2
M_22
M_11
Q_1
M_12
The shears are defined as

Q =
t/2
3 dx3
(28.8)
t/2
and are illustrated in the figure above; the units are force per unit length. The reason for
the introduction of these quantities is that the constitutive relations are typically recast in
terms of them and their relation to the curvatures and shears and the equilibrium equations
can be conveniently written in terms of them.
For the constitutive relations, the following results can be easily derived:

11
M11
D D
0
0 0
M22 D D
0
0 0
22

1
M12 = 0
0
2 D(1 ) 0 0
(28.9)
212 ,

Q1 0
0
0
t 0 23
0
0
0
0 t
13
Q2
3
Et
where D = 12(1
2 ) and is known as the flexural rigidity of the plate. Similar relations are
easily derivable for beams. As an example of how these relations are constructed consider
the expression for M11 .
M11 =
t/2
x3 11 dx3
(28.10)
t/2
Z t/2
x3
(28.11)
t/2
Z t/2
E
(11 + 22 ) dx3
(1 2 )
x3
E
(x3 11 x3 22 ) dx3
(1 2 )
(28.12)
t/2
2E
t3
(11 + 22 )
3(1 2 ) 8
= D(11 + 22 )
=
(28.13)
(28.14)
91
The equilibrium equations for beams and plates are written in terms of the stress resultants. They can be derived by considering small elements of a plate and performing force and
moment balance on them in the limit of vanishing size or they can be derived by taking the
3-D equilibrium equations and taking their 0th and 1st moments about the neutral surface.
In either case for plates the result is 3 equilibrium equations that need to be solved 1 for
vertical equilibrium and 2 for moment equilibrium:
Q1,1 + Q2,2 + p = 0
M11,1 + M12,2 Q1 = 0
M21,1 + M22,2 Q2 = 0 ,
(28.15)
(28.16)
(28.17)
where p is the normal pressure to the surface of the plate. In direct notation one has
Q+p = 0
M Q = 0.
(28.18)
(28.19)
Similar results also hold for beams.

In total for the plates, there are 13 equations (3 equilibrium, 5 constitutive, and 5
curvature-strain-displacement relations) and 13 unknowns (3 moments, 2 shear forces, 2
rotations, 1 displacement, 3 curvatures, and 2 shear strains). The key savings is that these
quantities are functions of two spatial variables not three. To specify a boundary value
problem four different types of boundary conditions can be utilized:
M n
Q n
u3
=
=
=
=
Q
u3
x M
x Q
x u
x
(28.20)
(28.21)
(28.22)
(28.23)
The () s represent different parts of the boundary and the barred quantities are given
boundary conditions. The full statement of the strong form of the boundary value problem
is: Find u3 , 1 , and 2 as functions of x such that (28.18) and (28.19) hold for given p(x )
and (28.20)-(28.23). The reduction to beams should be self-evident.
Remarks:
are not the corresponding values of the applied boundary
1. Note that the values M
moment vector.
Z t/2
M =
t x3 dx3
(28.24)
t/2
1 is the 2-component of the physically applied moment and M

2 is the negative
Thus, M
1-component of the physically applied moment.
28.1
Plate Weak Form
The weak form for the plate formulation of the last section can be constructed in two primary
ways: (1) Multiply the strong form equations by variations and integrate by parts or (2)
92
S. Govindjee
introduce our 3 assumptions into the weak form for 3-D elasticity and explicitly integrate
out the x3 -direction dependencies. In what follows method (1) is utilized.
Begin by defining the variations V = { | = 0, x } and u3 Vu =
{u3 | u3 = 0, x u }. The differentiability requirements on these functions are that they
must be at least once differentiable. By defining = +u3, and employing the usual
methods of weak form construction, a single weak form equation can be derived as
Z
A
, M + Q =
u3 p
+
M
,
u3 Q
(28.25)
where A is the 2-D domain of the plate itself.

Remarks:
1. The expression above has been organized so that the LHS can be interpreted as the
internal virtual work and the RHS as the external virtual work.
2. M involves single derivatives on the rotations and Q involves single derivatives on

the vertical displacement. Thus m = 1 and the requirement is for C 0 continuity in our
finite element approximations
3. When the additional Kirchhoff assumption is made, then = u3, and M involves
two derivatives on u3 . Thus m = 2 and the requirement is for C 1 finite element
approximations.
4. To generate a finite element approximation, one would begin by providing an expression

for the rotations and displacements; i.e.
MA A
(28.26)
NA u3A .
(28.27)
and
u3 =
93
Then define the following vectors and matrices:
11
= 22
212

13
=
23
M11
M = M22
M12

Q1
Q =
Q2
1
3
Et
1
Db =
12(1 2 )
0 0

t 0
Ds =
,
0 t
(28.28)
(28.29)
(28.30)
(28.31)
0
1
(1 )
2
(28.32)
(28.33)
so that
M = D b
Q = Ds .
Next define a vector of nodal unknowns
11
21
u31
..
.
d=
1n
2n
u3n
(28.34)
(28.35)
(28.36)
so that the following relations can be defined

= Bd
and
(28.37)
1
2 = N d .
u3
(28.38)
In the above,
B=
MA,1
0
0
0
MA,2
0
MA,2 MA,1
0
MA
0
NA,1
0
MA NA,2
(28.39)
94
S. Govindjee
and
MA 0
0
0 MA 0 .
N =
0
0 NA
(28.40)
Then one would employ the usual machinery to generate matrix equations and this
results in equations of the form
Kd = F ,
(28.41)
where the stiffness matrix is often decomposed into bending and shear contributions
as K = K b + K s .

Z
Db 0
T
Kb =
B
B
(28.42)
0 0
A

Z
0 0
T
Ka =
B
B.
(28.43)
0 Ds
A
The force vector is of the form

1
Z
Z
Z
0
M
0
2 +
NT 0 .
F =
NT 0
NT M
Q
A
M
Q
0
p
(28.44)
5. Note that in the limit as t 0, then kD s k >> kD b k. This is a situation somewhat

similar to the one that lead to incompressibility locking. Because of this, when using
shear deformable plate and beam elements in the thin limit, a certain amount of care
needs to be exercised. In other words, elements with modified integrations schemes
or mixed formulations should be used or bona fide thin plate elements with C 1 shape
functions should be used. Interested readers are referred to [9] and [24] and references
therein where some of these issues are discussed.
29
95
Shell Modeling with Plates
Plate-like structures are flat structures that carry load through bending and shear. This results in simple fabrication and construction costs at the expense of inefficient use of material.
More efficient use of material can be made through the use of shell-like structures (ie. those
structure wherein there is a coupling between the in- and out-of- plane forces). Shell-like
structures are basically those where the curvature of the structure forces this coupling (or
activation of the membrane action).
From the computational perspective one has three basic choices for the modeling of shelllike structures:
1. Use an appropriate shell theory for the problem at hand and generate a finite element
approximation to it for computational purposes. See [11] for a survey of some shell
theories and [24] for finite elements methods for shells.
2. One can simply employ 3-D brick elements. Typically, however, this is rather expensive
since adequate numbers of elements need to be used through the thickness of the shell.
With the advent of high-performance brick elements (those that bend well) this is
becoming more and more a viable possibility. Not many commercial codes to date
have good versions of these types of elements; they generally go by the names of
enhanced strain elements, and mixed-enhanced elements.
3. The last and most basic option is to approximate shell-like response through an assemblage of flat plate elements with membrane terms added. This is what is described
in the remainder of this section; see also [24].
29.1
Assemblages of Flat Plates
The basic flat plate element has three degrees of freedom at a node (u3 , 1 , 2 ). A membrane
element has two degrees of freedom at a node (u1 , u2 ) and is a plane stress element expressed
in terms of stress resultants.
Plate Element
x_3
Membrane Element
u_3
x_2
u_2
_1
_2
u_1
x_1
The membrane stress resultants are defined as forces per unit length and are simply the
in-plane stresses integrated through the thickness of the plate:
N =
t/2
t/2
dx3 .
(29.1)
96
S. Govindjee
The corresponding equilibrium equations are given as

N, + b = 0 ,
(29.2)
where b are the in-plane body forces (integrated through the thickness of the plate) of
dimension force per unit length-squared. The strain like variables for the membrane are the
usual planar strains = 21 (u, + u, ). The relevant constitutive equations follow from
the 3-D linear elastic constitution as
N11
1
Et
N22 =
1
(1 2 )
N12
0 0
0
11
22
0
1
(1 )
212
2
(29.3)
Adding the two degrees of freedom of the membrane to the plate element gives a five
degree of freedom element. Note however that the in-plane forces are decoupled from the
out-of-plane forces. Coupling in the linear theory can only take place due to geometric
curvature of the structure being analyzed. The flat plate has zero curvature and thus no
coupling. The coupling in an actual analysis arises when two such elements are joined at an
angle (not equal to 180 degrees). Such as is shown below:
x_3
Local Coordinate
System
x_2
x_1
X_3
x_1
Global Coordinate
System
X_1
JOIN
x_2
Local Coordinate
System
x_3
X_2
Note that in the local coordinate system for the element with node A that at each node
there will be 5 degrees of freedom associated with displacements in the three local coordinate
directions plus rotations about the local 1 and 2 directions. One has the same picture for
the element with node B. If we join the two elements together we will see the coupling.
In particular we have the following correspondence between degrees of freedom at nodes
97
A and B
uA
1
uA
2
uA
3
1A
2A
no correspondence
=
=
=
=
=
=
uB
2
uB
3
uB
1
no correspondence
1B
2B
(29.4)
REMARKS:
1. The desired coupling becomes evident by examining the second and third relations in
Eq. (29.4). From the second relation we see that transverse deflections in the element
with node B (which are caused by out-of-plane forces) are linked to in-plane deflections
in the element with node A (which are caused by in-plane forces). Thus there is a
coupling between in-plane and out-of-plane forces when two elements are joined at an
angle.
2. Note also that some of the degrees of freedom do not match up at all. To circumvent
this problem most elements of this type are given an extra degree of freedom for rotation
about the local 3-axis taken as positive about the 3-axis using the right hand rule.
Doing so give the additional correspondence of 1A = 3B and 3A = 2B .
3. Thus the shell like behavior (coupling) arises at the junctures of two elements. Even
though the convergence to true shell behavior in the limit of mesh refinement may be
slow, this is a very practical and common method of simulating shell-like behavior.
29.2
Finite Element Arrays for Assemblages of Flat Plates
Using the setup from the previous section we can set up the finite element array as follows.
For the nodal degrees of freedom at a single node, we can write in the local coordinate system
u1A
u2A
u3A
(29.5)
dA =
1A Plate .
2A
3A
The force vector for the element can also be expressed in the local coordinate system as
N1A
N2A
QA
FA =
.
(29.6)
M1A Plate
M2A
M3A
98
S. Govindjee
The element stiffness matrix can then be expressed on a per node basis as
22
K inplane 023 021

33
K plate
031 .
K A = 032
12
13
0
0
0
(29.7)
The final assembly into the global finite element arrays is performed just as it is in traditional
matrix structural analysis through the use of transformation matrices that relate the local
coordinate system to the global one. Thus we define the matrix of direction cosines as
eX1 ex1 eX1 ex2 eX1 ex3

R = eX2 ex1 eX2 ex2 eX2 ex3 ,
(29.8)
eX3 ex1 eX3 ex2 eX3 ex3
where exi is the unit vector in the i-th direction of the local coordinate system and eXi is the
unit vector in the i-th direction of the global coordinate system. Then the transformation
to the global coordinates is given via
dglobal = T dlocal
F global = T F local
K global = T K local T T ,
where
T =
(29.9)
(29.10)
(29.11)
R
R
..
.
R
(29.12)
6nen6nen
REMARKS:
1. This method of converting from local to global coordinates is only valid if the rotation
fields i are all small relative to unity. This restriction arises since in the above we
have treated the rotations as vectors.
2. If one uses the above formulation to model a structure that has sections of zero curvature then the zero diagonal element in the lower right-hand corner of Eq. (29.7) will
give rise to a singular global stiffness matrix and a spurious drilling mode in the solution. In such circumstance, most FEM program utilize an artificial drilling stiffness
value in place of the zero. The value is supposedly chosen such that it does not affect
any of the important parts of the solution. This is definitely a case of caveat emptor.
30
99
Introduction to Dynamics
In the previous sections, we have been dealing exclusively with static problems; ie. we have
been ignoring the dynamical terms that are present in the physical problems examined.
For many situations this is a valid approximation as these dynamical terms are of a small
magnitude in relation to the terms that we have retained. In the next few sections, we will
touch upon the basics of dynamic problems. First consider the dynamic extension of some
of the major problems with which we have dealt.
Heat Conduction
In the balance of energy equation for heat conduction we need to include the heat capacity
term. This gives us the following relation
cT + qi,i = r ,
(30.1)
where is the density of the material, c is the heat capacity, and we recall that qi = ij T,j
where ij is the thermal conductivity tensor. The only change from before is the first term
on the left-hand side of Eq. (30.1).
1-D Linear Elastic Rod

The static balance of forces or equilibrium equation now becomes the linear momentum
balance equation where the sum of the forces is now equal to the rate of change of the
momentum of the medium. Thus our governing equation is given by
AEu,xx + b = A
u,
(30.2)
where all terms are as before and is the material density.
Linear Elasticity
Here too the balance of forces or equilibrium equations now become the linear momentum
balance equations. Thus we have
ij,j + bi =
u,
(30.3)
where is the material density, ij = Cijkl kl and ij = 21 (ui,j + uj,i ).
In the previous sections all terms in Eqs. (30.1)(30.3) that contained a first or second time derivative were ignored. Clearly, for large problem classes this will not be true.
Such problems are classified as being dynamic problems. Within this classification, dynamic problems are sub-classified as either parabolic or hyperbolic. Parabolic problems are
characterized by infinite wave speeds; examples include heat conduction and ground water
diffusion. Hyperbolic problems are characterized by finite wave speeds; the main example is
elastodynamics in any dimension. The two different classifications lead to separate numerical
methods of solution though they both share a similar higher level formulation.
100
S. Govindjee
The method we will employ to deal with dynamic problems is known as the timecontinuous or semi-discrete method. In this formulation, one constructs the weak form
equations as we have seen previously. Then one approximates the solution spatially using a
finite element method with time as a continuous parameter in the nodal unknowns. Lastly
one approximates the solution temporally using a finite difference method.
30.1
Weak Form Equations
To generate the weak form expressions we proceed exactly as before: (1) define a suitable
space of trial solutions, (2) define a suitable space of test functions, (3) multiply the governing
balance equations by the test function, integrate over the domain, integrate by parts.

For concreteness consider a rod that is built-in at x = 0 and subject to dynamic body forces
b(x, t) and a dynamic end-load F (t). The solution space can be expressed as
S = {u(x, t) | u(0, t) = 0 and u(x, 0) = u(x) and u(x,

0) = u(x)}
,
(30.4)
where u(x) and u(x)

are known initial conditions. The space of test functions can be expressed as
V = {v(x) | v(0) = 0} ,
(30.5)
which is the same as in the static case. If we now multiply the balance Equation (30.2) by
an arbitrary test function and integrate as before we arrive at
Z
A
uv +
u,x AEv,x = F v(L) +
bv .
(30.6)
Heat Conduction
For the dynamic heat conduction case we will adopt all the previous notational conventions
from the static problem except that we will now allow for all boundary conditions to be time
varying. This leads to the following definitions for the space of solutions and the space of
test functions:
V = {v(x) | v(x) = 0 for x g }
(30.7)
S = {T (x, t) | T (x, t) = T(x, t) for x g and T (x, 0) = T(x) for x } , (30.8)
where T(x) are known initial conditions and T(x, t) are known boundary data on g . If
we now multiply the balance Equation (30.1) by an arbitrary test function and integrate as
before we arrive at
Z
Z
Z
Z
cT v + T,i ij v,j =
rv
hv .
(30.9)
101
Linear Elasticity
For the elastodynamic case we adopt the previous notational conventions of the static problem except that we now allow for all boundary conditions to be time varying. This leads to
the following definitions for the space of solutions and the space of test functions:
V = {v(x) | v(x) = 0 for x u }
(x, t) for x u
S = {u(x, t) | u(x, t) = u
(x)u(x,
and u(x, 0) = u
0) = u(x)
for x } ,
(30.10)
(30.11)
(30.12)
(x) and u(x)

(x, t) are known boundary
where u
and are known initial conditions and u
data on u . If we take the dot product of the balance Equations (30.3) by an arbitrary test
function and integrate as before we arrive at
Z
ui v i +
ij (u)Cijkl kl (v) =
bi vi +
ti vi .
(30.13)
The problem statement for each example would then read: Find the unknown function in
the space of solutions such that the corresponding weak form equation holds for all functions
in the space of test functions for all times in the time interval of interest.
30.2
Finite Element Approximation
To create a finite element approximation to the problems outlined above we choose subspaces of the space of trial solutions and the space of test functions. The choices are made
identical to the static case except that the vector of nodal unknowns for the solution is
take to be time dependent. The vector of nodal values of the test functions remains time
independent.
Heat Conduction
For the heat conduction case we have that
T h (x, t) =
NA (x)TA (t) = N T (t)
(30.14)
NB (x)vV = N v .
(30.15)
v (x) =
102
S. Govindjee
For convenience we can also define a gradient operator, B, just as before. If we know plug
into our dynamic weak form expression we will have
Z
X X Z
0 =
cNB vB NA TA (t) + NB,i vB ij NA,j TA (t)
A
rNB vB +
hNB vB
h
Z
X X Z
0 =
vB
cNB NA TA (t) + NB,i ij NA,j TA (t)
(30.16)
0 =
X Z
rNB +
hNB
Z
Z
Z
cNB NA TA (t) + NB,i ij NA,j TA (t) rNB +
Z
cN N T (t) +
T
(30.17)
0 =
Z
B B T (t)
rN +
If we now define the heat capacity matrix or thermal mass

Z
M=
cN T N ,
hNB
(30.18)
hN T .
(30.19)
(30.20)
the conductivity matrix

K=
B T B ,
(30.21)
(30.22)
and the generalized force vector as

F =
rN
hN T ,
then the problem can be expressed as a system of linear first order ordinary differential
equations
M T (t) + KT (t) = F (t) .
(30.23)

Following a similar procedure as with the heat conduction we will come to the following
system of linear second order ordinary differential equations:
(t) + Ku(t) = F (t) ,
Mu
(30.24)
where
MAB =
ANA NB
(30.25)
AENA,x NB,x
Z L
b(t)NA .
FA (t) = NA (L)F (t) +
KAB =
(30.26)
(30.27)
103
M is known as the consistent mass matrix.
Linear Elasticity
For linear elasticity we arrive a very similar system of linear second order ordinary differential
equations:
(t) + Ku(t) = F (t) ,
Mu
(30.28)
where
M =
N T N
(30.29)
K =
B T DB
(30.30)
F (t) =
N b(t) +
N T t(t) .
(30.31)
(30.32)
Here again, M is known as the consistent mass matrix.
All three examples lead to systems of ordinary differential equations in time for the nodal
unknowns. The only difference between the cases in a superficial sense is the number of time
derivatives involved. To solve these equations in time we will use finite difference methods.
30.3
Introduction to Numerical Integration by Finite Differences
The numerical integration of ordinary differential equations by the finite difference method is
a very developed branch of applied mathematics. A through discussion covering this material
in depth can be found in [14]; see also [13] for a more elementary treatment. For a discussion,
within the context of finite element approximations one can look at [9, 5] or [1]. For a peek
at the idea of using finite element methods instead of finite differences in time see [10].
30.4
A Scalar Example
To introduce the idea of numerical integration by finite differences consider the following
initial value problem
y = f (y, t)
y(0) = y ,
(30.33)
(30.34)
where f () is a known function. We desire a methodology for generating the solution to this
equation at discrete times t0 = 0, t1 = t, t2 = t1 + t, , where t is a given positive
number. We will denote the approximate solution at time tn as yn , where it is desired that
yn be, in some fashion, close to the exact value y(tn ). There are a large number of finite
difference methods that one could apply to this problem [14, 13], however, we will only
examine two of the simplest ones in order to introduce the basic ideas. The essential idea
behind finite difference methods is to approximate the derivatives appearing in the ODE
with an expression in terms of the values of the function y() and the time step t.
104
S. Govindjee
Forward Euler
In the forward Euler method we begin with a Taylor series expansion of the function y(t).
The expansion is formed about an arbitrary time step tn and evaluated at the next time step
tn+1 .
y(tn+1 ) = y(tn ) + y(t
n )t + O(t2 ) .
(30.35)
We now substitute for the first time derivative of y using Eq. (30.34). Thus
y(tn+1 ) = y(tn ) + f (y(tn ), tn )t + O(t2 ) .
(30.36)
If we now truncate the series and assume yn = y(tn ), then we come to the recursion relation
yn+1 = yn + f (yn , tn )t
(30.37)
which defines the forward Euler method. To utilize the method, one starts with a known
initial value y0 = y(0) = y and simply evaluates Eq. (30.37) to generate the sequence
of approximate values y1 , y2 , y3 , . . .. At the present we defer a critical examination of the
properties of the approximation but simply note that just as in finite element methods one
needs to consider certain issues in order to establish that what we have done will converge
to the true solution y(t). To this end we rewrite Eq. (30.37) in the completely equivalent
form
yn+1 yn
= f (yn , tn ) .
(30.38)
t
This also shows that the recursion relation can be viewed as an approximation to the ODE.
This will be important later on when we look at convergence.
Backward Euler
In the backward Euler method we also begin with a Taylor series expansion of the function
y(t). This time the expansion is formed about an arbitrary time step tn+1 and evaluated at
the previous time step tn .
y(tn ) = y(tn+1 ) y(t
n+1 )t + O(t2 ) .
(30.39)
We now substitute for the first time derivative of y using Eq. (30.34). Thus
y(tn ) = y(tn+1 ) f (y(tn+1 ), tn+1 )t + O(t2 ) .
(30.40)
If we now truncate the series and assume yn = y(tn ), then we come to the recursion relation
yn+1 = yn + f (yn+1 , tn+1 )t ,
(30.41)
which defines the backward Euler method. To utilize the method, one starts with a known
initial value y0 = y(0) = y and solves Eq. (30.41) for yn+1 to generate a sequence of approximate values y1 , y2 , y3 , . . .. Since an equation must be solved to determine yn+1 this is
considered an implicit method. Formally, it is only slightly different from the forward Euler
105
method but it does have very different approximation properties. If we write it in a form
similar to Eq. (30.38) then we find
yn+1 yn
= f (yn+1 , tn+1 ) .
t
(30.42)
This shows that the only difference between the two methods is where the derivative function
is evaluated.
Other Methods
There are clearly a wide variety of similar methods that could be developed using similar
ideas and keeping more terms in the Taylor series. Another category of methods can be
derived by using quadrature methods; see [13]. To keep things simple, these two will suffice
for a brief introduction.
30.5
Convergence
The main theorem that governs the properties of finite difference methods is call the Lax
Equivalence theorem. Loosely, it states that for convergence, a finite difference method must
be consistent and stable. If it has both these properties then the method will be convergent
is the sense that as the time step t is decreased the sequence of values y1 , y2 , y3 , . . . will
approach y(t1 ), y(t2 ), y(t3 ), . . .. For a more precise statement of this convergence property
see [14] and [13].
Consistency
Consistency is the statement that the finite difference equation is an approximation to the
initial ODE with a certain order of accuracy. The basic analysis is done by assuming that
the yn = y(tn ) and subtracting the ODE and the recursion relation from each other. With
the judicious use of Taylor series we can determine the discretization error, , involved in
replacing the ODE with the recursion relation.
For example, with the forward Euler method we have
=
yn+1 yn
f (yn , tn ) [y(t
n ) f (yn , tn )]
|
{z
}
t
|
{z
}
Finite Difference Recursion
ODE
yn+1 yn
y(t
n)
t
yn + y(t
n )t + O(t2 ) yn
=
y(t
n)
t
= O(t1 ) .
=
(30.43)
(30.44)
(30.45)
(30.46)
This result tells us that the finite difference recursion relation is consistent with the ODE; ie.
to a certain degree they are the same and more importantly that as the time step decreases
106
S. Govindjee
the difference between the two goes to zero. A similar analysis on the backward Euler method
gives us the same result; ie. = O(t1 ).
The requirement of consistency is that the exponent on t in the discretization error
be strictly greater than zero; ie. for = O(tk ), one must have k > 0 for consistency.
The standard terminology that is used says that the FD method is consistent (or accurate)
of order k. With this result one can show theoretically that for reasonably well behaved
functions f () that in the limit of t 0 that the finite difference sequence approaches
the exact solution. However, because of finite precsion round-off errors that occur in real
compuations this theoretical result is insufficient to ensure actual convergence. The added
condition for actual convergence is the other half of Laxs hypothesis viz. stability.
Stability
The notion of stability refered to in Laxs Theorem is associated with the properties of the
finite difference method itself. To test for this property, the FD method is normally applied
to the standard test problem of f (y, t) = y, where C and C denotes the space of complex
numbers. For this choice of f (), the exact solution to the IVP is
y(t) = y exp[t] = y exp[r t] (cos(i t) + sin(i t)) ,
(30.47)
where r and i are the real and imaginary parts of . When r < 0 it is seen that the
exact solution is oscillating-decreasing. A finite difference method is considered to be stable
if when it is applied to the test problem, that the sequence of values is decreasing for r < 0.
As an example consider the forward Euler method. If we apply it to the test problem we
have
yn+1 = (1 + t)yn .
(30.48)
The solution to this recursion relation is found by using the canonical guess of yn = Arn ,
where A and r are unknown. Inserting above gives the characteristic equation
Arn+1 = (1 + t)Arn
r = (1 + t) .
(30.49)
(30.50)
Thus yn = A(1 + t)n . If we now check the initial condition we come to the final answer of
yn = y(1 + t)n .
(30.51)
For stability we must have |1 + t| < 1. The combination of values of and t > 0 that
will satisfy this requirement is rather limited. The points in the complex t plane where
it is satisfied are depicted below as the disk of radius 1 centered at t = 1 + 0i.
107
-2

Region of Stability
for forward Euler
This type of finite difference method is called conditionally stable because for a given value
of with r < 0 the value of the time step is subject to a condition of being sufficiently small
for stability to hold. If || is large, then t will have to be very small. The consequence of
this is that to simulate the behavior of a system many time steps may be necessary. If the
time step is chosen to violate this restriction, then the finite difference solution to the test
problem will begin to increase without bound even though the exact solution is decreasing.
As a second example consider the backward Euler method. If we apply it to the test
problem we have
yn+1 = yn + tyn+1

1
yn+1 =
yn .
1 t
(30.52)
(30.53)
The solution to this recursion relation is again found by using the canonical guess of yn = Arn ,
where A and r are unknown. Inserting above gives the characteristic equation

1
n+1
Ar
=
Arn
(30.54)
1 t
.
(30.55)
r = 1
1 t
If we now check the initial condition we come to the final answer of

n
1
yn = y
.
1 t
(30.56)
1
| < 1. The combination of values of and t > 0 that
For stability we must have | 1t
will satisfy this requirement is quite large. The points in the complex t plane where it is
108
S. Govindjee

satisfied are depicted below as the entire region outside of the disk of radius 1 centered at
t = 1 + 0i.
t
Region of Stability
for backward Euler
This type of finite difference method is called unconditionally stable because for any
given value of with r < 0 the value of the time step is not subject to any restrictions
for stability to hold. The consequence of this is that in the simulation of the behavior of a
system time steps of any size may be taken (subject of course to the demands of accuracy).
Note the method even predicts decaying solutions for many values of t where r > 0 and
the exact solution is increasing without bound. Thus the method could be said to be overly
stable.
REMARKS:
1. For a given consistent finite difference method, if it operates within its region of stability
with respect to the test problem then it can be considered to be convergent.
2. Conditionally stable methods are normally suitable for calculations that range over
time intervals that are on the order of the characteristic time of the system under
investigation. This is primarily due to the severe time step restrictions.
3. Unconditionally stable methods (which are always implicit) are more expensive to use
than typical conditionally stable methods (which are normally but not always explicit);
however, one can use them for longer time simulations with larger time steps. Thus
they are better suited to longer duration simulations.
4. The forward and backward Euler methods are suitable for applying to first order ODEs.
In the next section they are applied to the heat conduction problem and a discussion
109
is presented on how to interpret the stability requirement when one has a system of
ODEs as opposed to just a single ODE.
110
30.6
S. Govindjee
Application to Heat Conduction
The finite element discretized heat conduction equations are given by

M T + KT = F .
(30.57)
In the next two sections we will apply the forward and backward Euler finite difference
methods to this system of ODEs and then in the following section we will examine the
relation of stability as discussed with in the previous section to the present problem that
involves a system of equations.
Forward Euler
The application of forward Euler to Eq. (30.57) gives

T n+1 T n
M
+ KT n = F n .
t
(30.58)
Rearranging yields
M T n+1 = M T n tKT n + tF n .
(30.59)
Thus knowing the state a time tn and the heat-load at tn we can compute the state at time
tn+1 by solving these linear algebraic equations.
REMARKS:
1. Normally, when working with an explicit method one modifies the thermal mass
matrix to make it diagonal so that no equations need to be solved to get the new
temperature field. The modified thermal-mass is usually called the lumped thermalmass and it is formed simply by taking the off diagonal terms in a row and adding them
to the diagonal element and then zeroing them. This obviously changes the problem
being solved in some fashion; for now we defer a discussion of this approximation until
we discuss mass lumping in the context of elastodynamics.
Backward Euler
The application of backward Euler to Eq. (30.57) gives

T n+1 T n
+ KT n+1 = F n+1 .
M
t
(30.60)
Rearranging yields
(M + tK) T n+1 = M T n + tF n+1 .
(30.61)
Thus knowing the state a time tn and the heat-load at tn+1 we can find the state at time
tn+1 by solving these linear algebraic equations.
REMARKS:
1. The matrix on the left-hand side in Eq. (30.61) is often called the dynamic heat conduction matrix, since formally it replaces the static FEM heat conduction matrix.
111
Stability
To understand the relation of the stability of the test equation to the present system of
equations requires a modal decomposition of our system of equations. We begin with the
homogeneous equations (ie. F = 0) and the general guess to the solution of the form
T (t) = exp[t] .
(30.62)
is an eigenvector or mode vector and is its corresponding eigenvalue. We can determine

and by inserting the guess into Eq. (30.57). This gives
M exp[t] + K exp[t] = 0 .
(30.63)
(K M ) = 0 .
(30.64)
Or equivalently
This is simply a generalized eigenvalue problem. For non-trivial solutions to Eq. (30.64), the
determinant of the matrix on the left-hand side needs to be equal to zero. Thus there will
be as many eigenpairs as there are degrees-of-freedom in the system. We will denote these
as (l , l ) for l = 1, . . . , n, where n is the number of degrees of freedom in the problem.
A very important property of the eigenvectors is that they are mass-orthogonal. Further
we usually normalize them so that they are mass-orthonormal. To see this property start
with two eigenpairs with differing eigenvalues (i , i ) and (j , j ). Then we have that
i M i = Ki
j M j = Kj .
(30.65)
(30.66)
If we now take the dot product of the first equation by j , the second by i , and subtract
the two results from each other we come to
(i j )j M i = 0 ,
(30.67)
where advantage has been made of the symmetry of both M and K. Thus if i j 6= 0
then we have j M i = 0; if i j = 0 we normalize the result so that both cases can
be expressed as
j M i = ji .
(30.68)
Note also that this implies that
j Ki = i ji
(no sum) .
(30.69)
From these properties we can conclude that the generalized eigenvectors form a basis for
Rn and we can expand the exact solution to Eq. (30.57) in them. In other word we can
write
n
X
T (t) =
T(i) (t)i ,
(30.70)
i=1
112
S. Govindjee
where T(i) (t) are known as the modal amplitudes of the solution. Inserting this expansion
into Eq. (30.57), gives
!
!
n
n
X
X
M
T(i) (t)i + K
T(i) (t)i = F (t) .
(30.71)
i=1
i=1
If we now take the dot product of both sides with an eigenvector j and make use of the
mass-orthonormal properties of the eigenvectors, then we end up for each value of j with the
following equation
T(j) (t) + j T(j) (t) = F(j) (t) ,
(30.72)
where F(j) = j F (t). Thus our system of linear first order ODEs turns into n decoupled
scalar ODEs in the modal amplitudes.
REMARKS:
1. When applying our finite difference methods to the heat conduction equations we are
essentially applying it each of these single degree-of-freedom problems simultaneously. The stability of the time stepping will then be governed by the magnitudes of
the eigenvalues j . In particular, if we use backward Euler then there will be no time
step restriction for stability. If we use forward Euler then,
t <
2
.
max j
(30.73)
j=1,...,n
Thus it is the fastest mode of the system that will govern the time step (even if we
are only interested in slower modes).
2. The practical issue of choosing a time step for conditionally stable methods relies
on being able to estimate the largest generalized eigenvalue of the system. Common
techniques include power iteration and inverse iteration with Rayleigh quotient; see [8]
for a discussion of these methods and their properties. Note that only an upper bound
to the largest eigenvalue needs to be determined.
30.7
113
Application to Elastodynamics
The application of finite difference methods to the equations of elastodynamics is a bit more
complex than the heat conduction case owing to the second order nature of the governing
equations. In this section we being with an illustration that direct methods of finite differences may not be applied to the second order system without care. Next we examine the
most popular finite difference method for elastodynamics and then we consider some useful
alternatives. Lastly we look at a number of practical issues that one needs to deal with in
the selection of an integration method.
An application of forward Euler

Let us consider the properties of the forward Euler method when applied to the elastodynamics problem. To simplify matters, we will work with the modally decomposed equations.
By choosing a general solution to the coupled matrix ODEs in the form
u(t) = exp[it] ,
(30.74)
we will arrive at the generalized eigenvalue problem

2 M + K = 0 .
(30.75)
This is of the same form which we had in the heat conduction case and the same conclusions
can be drawn:
2
, (j) ).
1. There will be one eigenpair for each degree of freedom in the system, ((j)
2. The eigenvectors will be mass orthogonal and we can further normalize them to be
mass orthonormal.
3. The eigenvectors will form a basis for Rn .
4. The general solution can be written as u(t) =
u(j) (t)(j)
5. The coupled system of ODEs can be written as the set of equivalent uncoupled ODEs
2
u(j) = F(j) (t)
u(j) + (j)
j = 1, 2, . . . , n ,
(30.76)
where F(j) (t) = (j) F (t).

With this background we concentrate on the behavior of a single mode and will drop the
subscript (j)s to simply the notation. To apply forward Euler we rewrite the second order
expression u + 2 u = F (t) as two first order equations:
d
dt
u
v
0 1
2 0

u
v
0
F
(30.77)
114
S. Govindjee
The application of forward Euler will lead to

un+1
un
0 1
un
0
=
+ t
+ t
vn+1
vn
2 0
vn
Fn

1
t
un
0
=
+
.
2
t 1
vn
tFn
(30.78)
(30.79)
This system of difference equations is consistent with the original system of ODEs to first
order.
The other issue that needs to be examined is the issue of stability. To that end consider
the case where the forcing function is zero. Then the solution to the difference equations
can be expressed as

n
u0
un
1
t
.
(30.80)
=
v0
vn
2 t 1
{z
}
|
An
For bounded behavior of the finite difference solution for large n, we must have that the
eigenvalues of the amplification matrix A be less than or equal to 1; (repeated roots must
be strictly less than one). For the amplification matrix at hand we find the eigenvalues to
be 1 it; both have a modulus greater than unity for finite values of the time step. This
tells us that a straight application of forward Euler to the elastodynamics equations will
generate a scheme that is unconditionally unstable.
Newmarks Method
Perhaps the most commonly used method to integrate the elastodynamics equations is Newmarks method. This is actually a two parameter family of methods; the two parameter
are typically denoted (, ). First we introduce velocity and acceleration approximations,
n ) and an u
(tn ), in addition to the displacement approximation un u(tn ).
v n u(t
Governing expressions for these vectors are given by the (approximate) momentum balance
equations and two finite difference approximations:
M an+1 + Kun+1 = F n+1
v n+1 v n
= (1 )an + an+1
t
un+1 un
t
= vn +
[(1 2)an + 2an+1 ] .
t
2
(30.81)
(30.82)
(30.83)
There are several different methods of applying these equations. A common one is based
on a predictor-corrector type layout. First one computes from a known state at time tn an
explicit predictor of the tn+1 displacements and velocities as:
n+1 = v n + t(1 )an
v
t2
n+1 = un + tv n +
u
(1 2)an .
2
(30.84)
(30.85)
115
This implies that the tn+1 accelerations are given as

an+1 =
1
n+1 ) .
(un+1 u
t2
Thus the (approximate) momentum balance equation can be written as:

1
1
n+1 .
M + K un+1 = F n+1 +
u
2
t
t2
(30.86)
(30.87)
We can now solve this last equation for the approximate displacements un+1 . The approximate velocities and accelerations are then easily computed as:
n+1 + tan+1
v n+1 = v
1
n+1 ) .
an+1 =
(un+1 u
t2
(30.88)
(30.89)
The cycle repeats itself for continuing time steps.

REMARKS:
1. Note that the solution involves the storage of not just the nodal displacements but also
approximations to the nodal velocities and accelerations.
2. The implementation described above is known as the d-form since it involves the
solution of equations for the nodal displacements. An alternative form known as the
a-form is also common place and involve a rewriting of the above expressions into
a form where the equation solving leads to the acceleration approximation and the
displacements and velocities follow from updating formulas.
3. The analysis of Newmarks family of methods can be most easily performed in the
modal setting.
Analysis of Newmarks Method

The analysis of Newmarks family of methods is normally performed in the modal setting.
In this setting (if we drop the subscripts associated with the particular modal equation in
question) we have the following Newmark equations:
an+1 + 2 un+1 = Fn+1
vn+1 vn
= (1 )an + an+1
t
un+1 un
t
= vn +
[(1 2)an + 2an+1 ] .
t
2
(30.90)
(30.91)
(30.92)
The discretization error can be found by substitution the modal momentum balance
expression into the second two Newmark equations and then subtracting the exact values
116
S. Govindjee
for the velocity and acceleration. The resulting difference (under the assumption of exact
values at time tn ) is the discretization error involved; ie.
un+1 un

u(t
n+1 )
vn + t
[(1 2)(Fn 2 un ) + 2(Fn+1 2 un+1 )]
t
2
= (t) .
vn+1 vn
u(tn+1 )
(1 )(Fn 2 un ) + (Fn+1 2 un+1 )
t
(30.93)
Through the judicious use of Taylor series and the exact modal momentum balance equation
one can show that = O(tk ), where k = 2 if = 12 and k = 1 for 6= 12 .
The stability of the method can be examined by creating an amplification matrix just
as was done in the forward Euler example. In the present case, one can show after a small
amount of algebra that
t2
1
A=
(12) 2
2
1+t2 2
t
1+t2 2
(1 ) +
1 t
(12) 2
2
1+t
2 2
t2 2
1+t2 2
(30.94)
The condition of stability requires that the eigenvalues of this matrix be less than or equal to
one and that repeated roots be strictly less than one. This lead to the following conclusions
[9]:
2
1
2
Unconditional Stability
12
< 2
Conditional Stability
t < ( 2 )1/2 /
(30.95)
(30.96)
REMARKS:
1. Once again we can see that just as in heat conduction it is the fastest mode (highest
frequency) in the problem that dictates the size of the time step in the conditionally
stable methods.
2. Several well known methods are contained in the Newmark family. From the preceding
analysis we can obtain their fundamental properties.
Name
Type
Central Differences Explicit
Trapezoidal Rule
Implicit
Linear Acceleration Implicit
Fox-Goodwin
Implicit
3. Only methods with =
1
2
0
1
4
1
6
1
12
1
2
1
2
1
2
1
2
Stability
Accuracy
t 2
t2
Unconditional
t2
t 3.464
t2
t 2.449
t2
are second order accurate.
4. The highest values of on the mesh can be bounded from above as the highest element
eigenvalue. This removes the need to compute global eigenvalue estimates. Usually
this estimate is not even directly computed, but rather further estimated as fraction
of the time it takes for an elastic wave to traverse the smallest element in the mesh.
117
5. Note that oscillations on the mesh near the Nyquist frequency of the mesh will have
very few elements per wave length and thus will not be very accurately represented.
As a rule of thumb one usually needs at least 10 elements per wave length for decent
accuracy. Thus in order to avoid having inaccurate modes polluting the computation
one normally wants methods with some sort of algorithmic damping in the high modes.
With the Newmark family this implies that > 12 . Thus one has to settle for first order
accurate methods. The maximum amount of damping is given by = ( + 12 )2 /4. The
downside of algorithmic damping is that it introduces amplitude errors.
6. The desire for second order accurate methods with high mode algorithmic damping has
led to the development of several alternative methods to Newmark. Two of the most
common are the Wilson- method and the HHT -method. For more on these two
methods the reader is directed to [9] and references therein. Here it is simply noted
that these methods give second order accuracy and high mode damping. Between
the two the HHT -method is somewhat superior to the Wilson- method, since the
Wilson- method is subject to spurious large amplitude overshoot phenomena at short
times.
7. A further alternative that is available in some finite element codes is a modal integration
option. In this method, the low mode eigenvectors and eigenvalues that are of interest
are computed with a method such as sub-space iteration [1]. The high modes are then
discarded and the modal equations are integrated exactly over each time step assuming
linear variations in the forcing function. This lead to good fidelity computations at a
fraction of the cost of complete system integration. For many structural applications
this is often the most economical method; see [3].
8. The method of choice is quite problem dependent. Generally, however, one can say that
for wave propagation problems explicit methods are more common and for structural
vibration problems implicit methods are more common. Another way of breaking this
down is that if the duration of the simulation is on the order of the system period
or less then an explicit method should be chosen; if the duration of the simulation is
greater than the period of the system then an implicit method should be chosen.
118
S. Govindjee
References
[1] K.-J. Bathe, Finite element procedures, Prentice Hall, 1996.
[2] G.P. Bazeley, Y.K. Cheung, B.M. Irons, and O.C. Zienkiewicz, Triangular elements
in plate bending conforming and nonconforming solutions, Proceedings of the First
Conference on Matrix Methods in Structural Mechanics, 1965.
[3] A.K. Chopra, Dynamics of structures, Prentice Hall, 1995.
[4] S.D. Conte and C. de Boor, Elementary numerical analysis, McGraw-Hill, 1980.
[5] R.D. Cook, D.S. Malkus, and M.E. Plesha, Concepts and applications of finite element
analysis, 3rd ed., John Wiley, 1989.
[6] I. Fried, Finite element analysis of incompressible materials by residual energy balancing,
International Journal of Solids and Structures 10 (1974), 9931002.
[7] Y. C. Fung, Foundations of solid mechanics, Prentice-Hall, 1965.
[8] G.H. Golub and C.F. Van Loan, Matrix computations, 3rd ed., Johns Hopkins University
Press, 1996.
[9] T.J.R. Hughes, The finite element method, Prentice Hall, 1987.
[10] C. Johnson, Numerical solutions of partial differential equations by the finite element
method, Cambridge University Press, 1987.
[11] A.W. Leissa, Vibration of shells, Tech. report, NASA, Washington, 1973.
[12] J.C. Nagtegaal, D.M. Parks, and J.R. Rice, On numerically accurate finite element
solutions in the fully plastic range, Computer Methods in Applied Mechanics and Engineering 4 (1974), 153178.
[13] D. Quinney, An introduction to the numerical solution of differential equations, Research
Studies Press (John Wiley), 1985.
[14] R.D. Richtmeyer and K.W. Morten, Difference methods for initial-value problems, John
Wiley, 1967.
[15] J.C. Simo and M.S. Rifai, A class of mixed assumed strain methods and the method
of incompatible modes, International Journal for Numerical Methods in Engineering 29
(1990), 15951638.
[16] I. S. Sokolnikoff, Mathematical theory of elasticity, McGraw-Hill, 1956.
[17] G. Strang and G. Fix, An analysis of the finite element method, Wellesley-Cambridge
Press, 1973.
119
[18] R.L. Taylor, P.J. Beresford, and E.L. Wilson, A nonconforming element fo stress analysis, International Journal for Numerical Methods in Engineering 10 (1976), 12111219.
[19] R.L. Taylor, J.C. Simo, O.C. Zienkiewics, and A.C. Chan, The patch test: A condition
for assessing finite element convergence, International Journal for Numerical Methods
in Engineering 22 (1986), 3962.
[20] S. Timoshenko and J. N. Goodier, Theory of elasticity, McGraw-Hill, 1970.
[21] K. Washizu, Variational methods in elasticity and plasticity, Pergamon Press, 1968.
[22] E.L Wilson, R.L. Taylor, W.P. Doherty, and J. Ghaboussi, Incompatible displacement
models, Numerical and Computer Models in Structural Mechanics (S.J. Fenves, N. Perrone, A.R. Robinson, and W.C. Schonbirch, eds.), Academic Press, 1973, pp. 4357.
[23] O.C. Zienkiewicz and R.L. Taylor, The finite element method, volume 1, 5th ed., Butterworth and Heinemann, 2000.
[24]
, The finite element method, volume 2, 5th ed., Butterworth and Heinemann,
2000.

Sanjay FEM

Hochgeladen von

Dokumentinformationen

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Sanjay FEM

Hochgeladen von

Copyright:

Verfügbare Formate

CE222

Finite Element Methods

Associate Professor of Civil Engineering, University of California Berkeley, govindjee@ce.berkeley.edu

Draft Notes CE222 Spring 2002

3 Properties of the Solution uh

5 Geometric and Energetic Interpretations of the Solution

8 Element View Point

9 Higher Order Elements

10 Calculations Over Standard Elements/Isoparametric Concepts

11 A 1-D Quadratic Isoparametric Element

13 Quadrature and Convergence

14 Introduction to Multi-Dimensional Problems

16 Summary of Elasticity Equations

17 Strong, Weak, and Minimization Forms of Elasticity

18 Finite Element Approximations for Elasticity

19 Consistent Nodal Loads

21 Mesh Refinement with Higher Order Elements

22 Stress Recovery Methods

23 Introduction to the Patch Test

24 Element Performance in Bending

28 Introduction to Shear Deformable Plates and Beams

29 Shell Modeling with Plates

Draft Notes CE222 Spring 2002

The Basics of Problem Statements

The strong form of

the BVP says:

Find the function u S such that for all v V

This form of the BVP is sometimes termed a variational form.

(vAEu,x ),x vAEu,xx dx

Draft Notes CE222 Spring 2002

v[AEu,xx + b] dx + v(L)F = v(L)AEu,x (L)

With this choice (1.8) becomes

x(x L)[AEu,xx + b]2 dx = 0

Draft Notes CE222 Spring 2002

the function uh S h such that for all v h V h

NA,x AEN,xg u dx = 0 (2.9)

We now define an n by n matrix K with components

a vector u with components uB and a vector F with components given by

This allows us to rewrite our expression as

Draft Notes CE222 Spring 2002

Properties of the Solution uh

Draft Notes CE222 Spring 2002

Geometric and Energetic Interpretations of the Solution

Note v h V h , but since V h V, it is also a member of V. Therefore, by (W) the above

for all v h V h . To interpret this expression, first define the functional

Draft Notes CE222 Spring 2002

Draft Notes CE222 Spring 2002

b(x) = 100 lbs/in

The strong form for this problem reads:

Find the function u(x) such that

50 106 u,xx + 100 = 0

50 106 u,x (6)

The weak form of this problem reads:

the function u S = {u | u(0) = 0} such that for all v V = {v | v(0) = 0}

the function uh S h such that for all v h V h

xA1 < x < xA

and NA,x (x) =

xA1 < x < xA

NA,x 50 106 NB,x dx for A, B {2, 3, 4}

The symmetry of K permits us to only compute half of the matrix. Thus,

Draft Notes CE222 Spring 2002

By employing the definition of the hat functions

The part from the end load is given by

However, in our example u = 0. Adding the contributions together gives

"$#%#%& ' ()*+,-. !

&('') *!,+-.0/12, 0!%$