An Introduction to the

Finite Element Analysis

Agenda
PART I
Introduction and Basic Concepts

1.0 Computational Methods

1.1 Idealization
1.2 Discretization
1.3 Solution
2.0 The Finite Elements Method
2.1 FEM Notation
2.2 Element Types
3.0 Mechanichal Approach
3.1 The Problem Setup
3.2 Strain Energy
3.3 External Energy
3.4 The Potential Energy Functional

Joint Advanced Student School

St.Petersburg 2005
Agenda

PART II
Mathematical Formulation

4.0 The Mathematics Behind the Method

4.1 Weighted Residual Methods
4.2 Approxiamting Functions
4.3 The Residual
4.4 Galerkin’s Method
4.5 The Weak Form
4.6 Solution Space
4.7 Linear System of Equations
4.8 Connection to the physical system

Joint Advanced Student School

St.Petersburg 2005
Agenda

PART III
Finite Element Discretization

5.0 Finite Element Discretization

4.1 The Trial Basis
4.2 Matrix Form of the Problem
4.3 Element Stiffness Matrix
4.4 Element Mass Matrix
4.5 External Work Integral
4.6 Assembling
4.7 Linear System of Equations
6.0 References
7.0 Question and Answers

Joint Advanced Student School

St.Petersburg 2005
 PART I
Introduction and Basic Concepts
1.0 Computational Methods

Joint Advanced Student School

St.Petersburg 2005
1.1 Idealization
• Mathematical Models
• “A model is a symbolic device built to
simulate and predict aspects of behavior of
a system.”
• Abstraction of physical reality
• Implicit vs. Explicit Modelling
• Implicit modelling consists of using existent
pieces of abstraction and fitting them into
the particular situation (e.g. Using general
purpose FEM programs)
• Explicit modelling consists of building the
model from scratch
Joint Advanced Student School
St.Petersburg 2005
1.2 Dicretization

1. Finite Difference Discretization

• The solution is discretized
• Stability Problems
• Loss of physical meaning
2. Finite Element Discretization
• The problem is discretized
• Physical meaning is conserved on elements
• Interpretation and Control is easier

Joint Advanced Student School

St.Petersburg 2005
1.3 Solution

1. Linear System Solution Algorithms

• Gaussian Elimination
• Fast Fourier Transform
• Relaxation Techniques

2. Error Estimation and Convergence

Analysis

Joint Advanced Student School

St.Petersburg 2005
2.0 Finite Element Method
• Two interpretations
1. Physical Interpretation:
The continous physical model is divided into
finite pieces called elements and laws of
nature are applied on the generic element.
The results are then recombined to
represent the continuum.
2. Mathematical Interpretation:
The differetional equation reppresenting the
system is converted into a variational form,
which is approximated by the linear
combination of a finite set of trial functions.
Joint Advanced Student School
St.Petersburg 2005
2.1 FEM Notation

Elements are defined by the following

properties:
1. Dimensionality
2. Nodal Points
3. Geometry
4. Degrees of Freedom
5. Nodal Forces
(Non homogeneous RHS of the DE)

Joint Advanced Student School

St.Petersburg 2005
2.2 Element Types

Joint Advanced Student School

St.Petersburg 2005
3.0 Mechanical Approach

• Simple mechanical problem

• Introduction of basic mechanical concepts
• Introduction of governing equations
• Mechanical concepts used in mathematical
derivation

Joint Advanced Student School

St.Petersburg 2005
3.1 The Problem Setup

Joint Advanced Student School

St.Petersburg 2005
3.2 Strain Energy

• Hooke’s Law:

 ( x)  E ( x)

where
du
 ( x) 
dx
• Strain Energy Density:
1
   ( x) ( x)
2

Joint Advanced Student School

St.Petersburg 2005
3.2 Strain Energy (cont’d)

• Integrating over the Volume of the Bar:

1 1 L 1 L
U    dV   p dx   ( EAu ' )u ' dx
2 V 2 0 2 0
1 L
U   u ' EAu ' dx
2 0
• All quantities may depend on x.

Joint Advanced Student School

St.Petersburg 2005
3.3 External Energy

• Due to applied external loads

1. The distributed load q(x)
2. The point end load P. This can be
included in q.

• External Energy:
L
W   qu dx
0

Joint Advanced Student School

St.Petersburg 2005
3.4 The Total Potential
Energy Functional
• The unknown strain Function u is found by
minimizing the TPE functional described below:

  U W
or
[u ( x)]  U [u ( x)]  W [u ( x)]

Joint Advanced Student School

St.Petersburg 2005
 PART II
Mathematical Formulation
4.0 Historical Background

• Hrennikof and McHenry formulated a 2D

structural problem as an assembly of bars
and beams
• Courant used a variational formulation to
approximate PDE’s by linear interpolation
over triangular elements
• Turner wrote a seminal paper on how to
solve one and two dimensional problems
using structural elements or triangular and
rectangular elements of continuum.

Joint Advanced Student School

St.Petersburg 2005
4.1 Weighted Residual Methods

The class of differential equations containing also the one

dimensional bar described above can be described as
follows :

d du
L[u ]   ( p ( x) )  z ( x )u  q ( x ), 0  x 1 (1)
dx dx

u (0)  u (1)  0.

Joint Advanced Student School

St.Petersburg 2005
4.1 Weighted Residual Methods

It follows that:
L[u ]  q  0
Multiplying this by a weight function v and integrating over
the whole domain we obtain:
1
 ( L[u ]  q)v dx  (v, L[u ]  q)  0
0
(2)
For the inner product to exist v must be “square integrable”
Therefore:

v  L2 (0,1)
Equation (2) is called variational form

Joint Advanced Student School

St.Petersburg 2005
4.2 Approximating Function

• We can replace u and v in the formula with their

approximation function i.e.
N
u ( x )  U ( x )   c j j ( x )
j 1

N
v( x)  V ( x)   d j j ( x)
j 1
• The functions fj and yj are of our choice and are meant to be
suitable to the particular problem. For example the choice of
sine and cosine functions satisfy boundary conditions hence it
could be a good choice.

Joint Advanced Student School

St.Petersburg 2005
4.2 Approximating Function

• U is called trial function and V is called test function

• As the differential operator L[u] is second order

u  C 2 (0,1)  U  C 2 (0,1)
• Therefore we can see U as element of a finite-diemnsional
subspace of the infinite-dimensional function space C2(0,1)

U  S N (0 ,1)  C 2 (0 ,1)
• The same way

V  Sˆ N (0 ,1)  L2 (0 ,1)

Joint Advanced Student School

St.Petersburg 2005
4.3 The Residual

• Replacing v and u with respectively V and U (2) becomes

(V , r )  0, V  Sˆ N
r ( x)  L[U ]  q
• r(x) is called the residual (as the name of the method suggests)
• The vanishing inner product shows that the residual is orthogonal to
all functions V in the test space.

Joint Advanced Student School

St.Petersburg 2005
4.3 The Residual
N
• Substituting V ( x)  d 
j 1
j j ( x) into (V , L[U ]  q)  0

and exchanging summations and integrals we obtain

N

d
j 1
j ( j , L[U ]  q )  0 d j , j  1, 2 , ..., N

• As the inner product equation is satisfied for all choices of V in

SN the above equation has to be valid for all choices of dj
which implies that

( j , L[U ]  q)  0 j  1, 2 , ..., N

Joint Advanced Student School

St.Petersburg 2005
4.4 Galerkin’s Method

• One obvious choice of  j would be taking it equal to 

j
• This Choice leads to the Galerkin’s Method

( j , L[U ]  q )  0 j  1, 2 , ..., N
• This form of the problem is called the strong form of the
problem. Because the so chosen test space has more
continuity than necessary.
• Therefore it is worthwile for this and other reasons to convert
the problem into a more symmetrical form
• This can be acheived by integrating by parts the initial strong
form of the problem.

Joint Advanced Student School

St.Petersburg 2005
4.5 The Weak Form

• Let us remember the initial form of the problem

d du
L[u ]   ( p( x) )  z ( x)u  q( x), 0  x 1
dx dx

u (0)  u (1)  0.
1
(v, L[u ]  q )   v [( pu ' )' zu  q ] dx
0
• Integrating by parts
1 1

 v [( pu' )' zu  q] dx   (v' pu'vzu  vq) dx  vpu'

1
0
0
0 0

Joint Advanced Student School

St.Petersburg 2005
4.5 The Weak Form

• The problem can be rewritten as

A(v , u )  (v , q)  0
where
1
A(v , u )   (v' pu ' vzu ) dx
0
• The integration by parts eliminated the second derivatives
from the problem making it possible less continouity than the
previous form. This is why this form is called weak form of the
problem.
• A(v,u) is called Strain Energy.

Joint Advanced Student School

St.Petersburg 2005
4.6 Solution Space

• Now that derivative of v comes into the picture v needs to have more
continoutiy than those in L2. As we want to keep symmetry its
appropriate to choose functions that produce bounded values of
1
A(u , u )   ( p (u ' ) 2  zu 2 ) dx
0
• As p and z are necessarily smooth functions the following restriction is
sufficient 1

 
2 2
(u ' ) u dx
0

• Functions obeying this rule belong to the so called Sobolev Space and
they are denoted by H1. We require v and u to satisfy boundary
conditions so we denote the resulting space as H 1
0

Joint Advanced Student School

St.Petersburg 2005
4.7 Linear System of Equations

• The solution now takes the form

A(v , u )  (v , q) v  H 01
• Substituting the approximate solutions obtained earlier in the
more general WRM we obtain
U ,V  S 0N  H 01
A(V ,U )  (V , q) V  S 0N
• More explicitly substituting U and V (remember we chose
them to have the same base) and swapping summations and
integrals we obtain
N

 c A(
k 1
k j , k )  ( j , q) , j  1, 2 , ..., N

Joint Advanced Student School

St.Petersburg 2005
4.8 Connection to the Physical System

Mechanical Formulation Mathematical Formulation

1
1 L
U   u ' EAu' dx A(v , u )   (v' pu'vzu) dx
2 0 0
L
W   qu dx (v , q )
0

  U  W  0 A(v , u )  (v , q)  0

Joint Advanced Student School

St.Petersburg 2005
 PART III
Finite Element Discretization
5.0 Finite Element Discretization

• Let us take the initial value problem with constant coefficients

 pu ' ' zu  q ( x ), 0  x 1
p, z  0
u (0)  u (1)  0.
• As a first step let us divide the domain in N subintervals with
the following mesh

0  x0  x1  ...  xN  1
• Each subinterval ( x , x ), j  1 : N is called finite element.
j 1 j

Joint Advanced Student School

St.Petersburg 2005
5.1 The Trial Basis

• Next we select as a basis the so called “hat function”.

 x  x j 1
 x j 1  x  x j
 x j  x j 1
 x x
 j ( x)   j 1 x j  x  x j 1
 x j 1  x j


 0 otherwise

Joint Advanced Student School

St.Petersburg 2005
5.1 The Trial Basis

• With the basis in the previous slide we construct our

approximate solution U(x)

• It is interesting to note that the coefficients correspond to the

values of U at the interior nodes
Joint Advanced Student School
St.Petersburg 2005
5.1 The Trial Basis

• The problem at this point can be easily solved using the

previously derived Galerkin’s Method

 c A(
k 1
k j , k )  ( j , q ) , j  1, 2 , ..., N

• A little more work is needed to convert this problem into

matrix notation

Joint Advanced Student School

St.Petersburg 2005
5.2 Matrix Form of the Problem

• Restricting U over the typical finite element we can write

U ( x)  c j 1 j 1 ( x)  c j j ( x) x  [ x j 1 , x j ]

• Which in turn can be written as

 j 1 ( x) c j 1 
U ( x)  [c j 1 c j ]   [ j 1 ( x)  j ( x)]  x  [ x j 1 , x j ]
  j ( x)   cj 
in the same way

 j 1 ( x) d j 1 
V ( x)  [d j 1 d j ]
 ( x )   [ j 1 ( x)  j ( x)] d  x  [ x j 1 , x j ]
 j   j 

Joint Advanced Student School

St.Petersburg 2005
5.2 Matrix Form of the Problem

• Taking the derivative

 1 / h j  c j 1 
U ' ( x)  [c j 1 c j ]   [1 / h j 1 / h j ] c  x  [ x j 1 , x j ]
 1 / h j   j 
h j  x j  x j 1
• Derivative of V is analogus

 1 / h j  d j 1 
V ' ( x)  [d j 1 d j ]   [1 / h j 1 / h j ]  x  [ x j 1 , x j ]
 1 / h j  d
 j 

Joint Advanced Student School

St.Petersburg 2005
5.2 Matrix Form of the Problem

• The variational formula can be elementwise defined as

follows:
N

 [ A (V ,U )  (V , q)
j 1
j j ]0

A j (V , U )  A Sj (V , U )  A Mj (V , U )
xj
A (V ,U )  
S
j pV 'U ' dx
x j 1

xj
A (V ,U )  
M
j zVU dx
x j 1

xj
(V , q )   Vq dx
x j 1

Joint Advanced Student School

St.Petersburg 2005
4.3 The Element Stiffnes Matrix

• Substituting U,V,U’ and V’ into these formulae we obtain

 x j  1 / h j   c j 1 

A (V , U )  [d j 1 d j ]  p 
S
[1 / h j 1 / h j ] dx   
j  x j1  cj
  1/ hj   
 x j p  1  1  c j 1  c j 1 
S 
A j (V , U )  [d j 1 d j ]  dx   [d j 1 d j ]K j  
 x j1 h 2  1 1    c j 
 j    cj 

p  1 1 
Kj   1  1
hj  

Joint Advanced Student School

St.Petersburg 2005
4.4 The Element Mass Matrix

• The same way

 x j  j 1   c j 1 
A (V , U )  [d j 1 d j ]  z  [ j 1  j ] dx   
M
j  x j1  j  cj
    
c j 1 
A j (V ,U )  [d j 1 d j ]M j  
S

 cj 

zh j 2 1 
Mj 
6 1 2

Joint Advanced Student School

St.Petersburg 2005
4.5 External Work Integral
• The external work integral cannot be evaluated for every
function q(x)
xj
(V , q )   Vq dx
x j 1

• We can consider a linear interpolant of q(x) for simplicity.

q ( x)  q j 1 j 1 ( x)  q j j ( x), x  [ x j 1 , x j ]

• Substituting and evaluating the integral

xj  j 1  q j 1 
(V , q) j   [d j 1 , d j ] [ j 1 ,  j ]  dx  [d j 1 , d j ] l j
 j   qj 
x j 1

h j 2q j 1  q j 
lj    Element load vector
6 q j 1  2q j 
Joint Advanced Student School
St.Petersburg 2005
4.6 Assembling

• Now the task is to assemble the elements into the whole

system in fact we have to sum each integral over all the
elements
• For doing so we can extend the dimension of each element
matrix to N and then put the 2x2 matrix at the appropriate
position inside it
• With all matrices and vectors having the same dimension the
summation looks like
 2 1 
 c1   d1   1 2  1 
N c  d   
 A j  d Kc c   2 
S T
 ... 
d  2 
 ... 
p
K 
h
 1 2 1
... ... ...


j 1 
      1 2  1
c
 N 1  d
 N 1   
  1 2 

Joint Advanced Student School

St.Petersburg 2005
4.6 Assembling

• Doing the same for the Mass Matrix and for the Load Vector
4 1 
1 4 1 
 
zh  1 4 1 
N

 j
A M

j 1
 d T
Mc M 
6 

... ... ... 
 1 4 1
 
 1 4

 q0  4q1  q2 
 
N
h  q1  4q2  q3 
 (V , q)
j 1
j  d Tl l
6 ... 
 
q N  2  4q N 1  q N 

Joint Advanced Student School

St.Petersburg 2005
4.7 Linear System of Equations

• Substituting this Matrix form of the expressions in

N

 [ A (V ,U )  (V , q)
j 1
j j ]0

we obtain the following set of linear equations

d T [(K  M )c  l ]  0
• This has to be satisfied for all choices of d therefore

(K  M )c  l  0

Joint Advanced Student School

St.Petersburg 2005
References

• Carlos Felippa

http://caswww.colorado.edu/courses.d/IFEM.d/IFE
M.Ch06.d/IFEM.Ch06.pdf

• Joseph E Flaherty,Amos Eaton Professor

http://www.cs.rpi.edu/~flaherje/FEM/fem1.ps

• Gilbert Strang, George J. Fix

An Analysis of the Finite Element Method
Prentice-Hall,1973
Joint Advanced Student School
St.Petersburg 2005