2/24/2020

Finite Element Method

Course 461603

1
 2/24/2020

Course Description and Objectives

• The course will cover an introduction to finite element “FE” method.
General idea is explained, and the methods for deriving the matrices
are discussed. Truss, beam and solid plane and 3D element matrices
are derived. Application using computer programs is demonstrated.

• Students should be able to:

- Distinguish between exact and numerical solutions, and recognize
the many numerical and approximate methods.
- Derive finite element matrices for elements
- Understand the concept and limitations of many FE programs

2
 2/24/2020

• Grading Policy
• Mid-term Exam : 30%
• Homeworks: 30%
• Final Exam : 40%
• Textbooks/References
• “A First Course in the Finite Element Method”, 6th Ed.,2016, by Daryl L. Logan
• “Introduction to the Finite Element Method”,2006, 3rd Ed., by J. N. Reddy
• “Introduction to Finite Elements in Analysis” , 3rd Ed. 2002, by T. R.
Chandrupatla and Ashok

3
 2/24/2020

Course Outline
Week Subject
1-2 Introduction, General Idea of F.E.M
3-4 Derivation of F.E. matrices for bar and truss element,
transformation matrices, assembly of elements, application of
boundary conditions, comparison to exact solution
4-7 Beam and 2D frame elements.
8 Midterm Exam
9-12 2D Plane elements, CST, Quadratic Elements, isoparametric
formulation, and numerical integration. Application of
boundary conditions.
12-15 3D solid elements, Plates and Shell
15-16 Computer Applications
4
 2/24/2020

Introduction , Methods of Analysis

Generally, there 3 methods to obtain and analyze data or information:
• Experimental, Empirical methods:
Gives Exact – Real data. Valid within the limits of the experiments, Very
Specific to the case. Expensive, and time consuming.
• Analytical methods:
• Based on universal physical laws. [ such as energy conservation ].
• Exact or Closed-Form solutions, valid for the given mathematical
model. Remember: The solution is for the model, not for the reality.
• Good for simple cases

5
 2/24/2020

Introduction , Methods of Analysis

• Numerical methods: Approximates the solution to the differential equations
• Integration/Collocation/Partition methods
• Mesh-Based: Mathematical model is imposed on a pre-defined mesh (or connected grid), and certain physical laws are
satisfied on the grid-points.
• Finite Difference Method [FDM]
• Finite Element [FEM] and Extended-Finite Element Methods [X-FEM]
• Finite Volumes Method
• Boundary Element Method
• Mesh-Free: Mathematical model is imposed on un-structured points or overlapping regions
• Smoothed-particle hydrodynamics (SPH) Method
• Element-Free Galerkin Method
• Diffuse Element Method
• Finite Point Method [Petrov-Galerkin Mesh-Free Method]

• Heuristic and Meta-heuristic methods (Artificial Intelligence Learning methods) Solution is

divided to layers (or clouds), with interaction models between layers. Corrective [Learning/Heuristic] criteria are
imposed based on learning from many previously “known” results
• Genetic Algorithm Methods
• Artificial Neural Network
• Swarm Methods
6
 2/24/2020

Fathers of the F.E. Method

Boris Galerkin Ray William Clough Richard Courant Alexander Hrennikoff

7
 2/24/2020

Engineering Method
Change
Physical Problem physical
problem

Mathematical Improve
mathematical
Model model

Increase
Numerical model accuracy

No!
Logical Answer?
Results OK? Refine analysis

YES! Design improvements

Structural optimization
Happy  8
 2/24/2020

Engineering Method , Verification and Validation

How to know that the solution is OK!
Physical Problem
Validation
Mathematical Validation
Model
Verification
Validation
Numerical model

Solver
Verification
Results
9
 2/24/2020

General Procedure of Finite Element Method

1. Create an engineering / mathematical model of the structure. [ Define the Domain, and the Boundaries,
This step is called Modelling, and it requires Engineering Judgement]
2. Discretization [Partition] of the structure [Called Domain] using elements with assumed DOF’s per
element. This is called Meshing.
3. Select an interpolation function [ Also called Shape Function ] that approximates the response
[ displacement, or strain, or forces, or stresses…etc] inside each element.
4. Write energy equation [or differential equation] for each element in terms of DOF’s. Convert the energy
or differential equations into a set of algebraic equations. [ i.e. Derive the behavioral matrices] .
5. Sum the energies for all elements [ This is called Assembly of elements]
6. Impose boundary conditions on the sides of the domain.
7. Solve the resulting system of equations.
8. Plot and draw solution [ This is called Post-processing]
9. Steps 2 to 5 are called Pre-processing
In the overall steps above, the F.E. solution must satisfy the following
a. Equilibrium / Conservation of Energy - On an average sense
b. Compatibility / Kinematic – Depending on continuity of the shape functions and inter-connectivity
c. Constitutive law / Material Behavior – Often Exactly
d. Boundary conditions - Exactly
10
 2/24/2020

For example. Stress-distribution around a hole

Boundary
Domain
Forces

Model Boundary
Assume Plane Real Problem
stress behavior

Mesh:
Using Plane-Stress
triangular 3-noded
element

11
 2/24/2020

Types of Elements

12
 2/24/2020

Common Sources of Error in FEA

• Modelling errors:
• Mathematical model does not account for the physical problem, e.g. cracking, plasticity…etc
• Domain Approximation, the modelling error
• Approximation of the boundary conditions , the modelling error
• Selection of the Element Interpolation Shape functions
• Numerical Integration Errors when integrating the element matrices (Including
Spatial and Time Integration)
• Error due to solution method
• Computer Errors (Round-Off, Floating points, Etc., )

13
 2/24/2020

Assessment of results of FEA

Accuracy:
how close we are to the exact (experimental) solution
Relative Error = |(Exact/Experiment Solution) – (FEM Solution)| / |(Exact/Experiment Solution) |

Convergence:
Rate of improvement of solution towards the exact solution, i.e. how fast the error reduces.
The error can be reduced by:
• Increase Number of Elements [ called mesh refinement] (h-convergence)
• Increase the order of the interpolation function [called Approximation order] (p-convergence)
Generally: Error → 0 as Number of Elements or Approximation Order → ∞

14
 2/24/2020

Simple Element Equation Example

Direct Stiffness Derivation
u1 u2
• The direct stiffness approach is
consider the “birth” idea of the F.E. F1 F2
1 2
• Equilibrium is directly imposed on the k
element and the equations are written Equilibrium at Node 1  F = ku − ku
1 1 2
as a function of the DOF. Equilibrium at Node 2  F = −ku + ku
2 1 2

• This approach is easy for bar and beam

elements, but becomes almost or in Matrix Form

impossible for higher elements.

 k − k u   F 1 1
• We will not follow this approach from  − k k  u  =  F 
     2 2
now on. Stiffness Matrix Nodal Force Vector
[ K ]{u} = {F }
15
 2/24/2020

Bar Example, Approximate Methods

• Consider an elastic bar, that is loaded with a uniformly distributed axial force
q(x) = qo and is fixated at one end. The cross-sectional area of the bar
increases with x as
 =
 1 +  .
• The differential equation for such member is

( ) ( )
1+ 
 + 
 +  = 0


• For the sake of comparison and simplicity, let’s assume 
 =  =  = 1,
then the D.Eq become:
  () ()
1+ + +1=0
  
Since this is a second order D.Eq, then two boundary conditions are required.
Let’s assume those are u(0)=0 “fixed” and the stress at the free end is zero =
()
σ(L) = 0 =  = 0.

16
 2/24/2020

Bar Example, 1D element

• The boundary related to the displacement (the primary variable) are called
essential (or necessary) boundary conditions, while those related to the
derivatives of the displacements are called natural (or non-essential)
boundary conditions.
• Note that D.Eq can be solved exactly, and the “exact” solution is:
  = −  +2ln 1 + 

17
 2/24/2020

Bar Example: The Residual Methods

• The bar example will be solved using many techniques, and assuming single
element.
• All these techniques are based on approximating the solution and forcing the
error to be smallest.
• First, assume an “admissible” function  that satisfies the boundary
conditions.
• Substituting the function  will produce an error. This is called a “Residual
Error”
     
  = 1+ + +1
  
• There are 4 common methods to find the approximate solution:
• Collocation method
• Sub-Domain method
• Least Squares method
• Galerkin Method
18
 2/24/2020

Bar Example: The Residual Methods

• Let’s approximate the solution as a series of trial functions.
"

  ≅   =   !  = $ !$  +  !  + ⋯ &'
#$
• Since the highest derivative in the boundary conditions [For the
D.Eq] is 1, then the trial functions should have a non-zero
derivative of at least of order 1. Thus, the order of the trial
functions must be 2 or higher.
• Also, because we have two boundary conditions, we must at least
have 3 constants ci, to get a logical solution.

19
 2/24/2020

Bar Example, The Residual Methods

• For simplicity, let’s assume the trial solution to be a polynomial . Let’s choose
!$  = 1 , !  = , and !+  =   . The trial solution is:

  = $ +   + +  

• Imposing B.C. it becomes   = + (  − 2)

• The residual can be calculated as:

  = 1 +  2+ + 2+  − 1 + 1 = 4+  + 1

• The total “average” error over the domain is :

$
>    = 0
? 20
 2/24/2020

Bar Example, Collocation Method

• In this method, the residual is forced to become zero at some
selected points x* inside the domain.
• Here since we have only one unknown, we only need one point.
Select x* = 0.5. Hence:
 0.5 = 0.5 4+ + 1 = 0
A$
• Solving gives + = .

$
• Thus the solution is [  ]DEE = − (  −2)


21
 2/24/2020

Bar Example, Sub-Domain Method

• Here we divide the domain into sub-domains, and then assume the total residual in
each domain to be zero, individually, thus creating as many equations as the number of
sub-domains.

• The number of domains equals the number of unknowns in the residual.

• Here we have single unknown c3, thus, one sub-domain is needed.

$ $
>    = > [ 4+  + 1] = 0
? ?

A$
• Solve, gives + =


22
 2/24/2020

Bar Example, Least Squares Method

• Here we minimize the total sum of the square of the residual:
$
• Let S = G?   . To minimize S, we make the derivatives of S with
respect to the unknown parameters ci, to be equal to zero. This gives
HI $
H()
= 0 = 2>    = 0
H ? H
JK JK
The term is called a weight function. N () = , The equation above is
JLM JLM
generated for each value of the weight N
In our example, this gives, since there is only one unknown c3
$
H() $
>    = > 4 4+  + 1  = 0
? H ?
AO
Solving, gives + =
$O
23
 2/24/2020

Bar Example, Galerkin Method

• Here, we take the weight function to be the derivative of the trial solution
P
J
itself., i.e: N () = , which of course gives: N  = !  , thus
JLM
imposing:
$
> !     = 0
?
• In our example, since have only one unknown c3
$ $
> !+     = >   [ 4+  + 1]  = 0
? ?
$
Solving gives: + = −
+

24
 2/24/2020

Bar Example, Comparisons

25
 2/24/2020

Basis of Finite Element, Galerkin Method

• The Genius idea of Galerkin was to assume the weight functions to be the
derivatives of the trial functions themselves.
• From mechanics we know that the equilibrium equations are simply the
derivatives of the total energy Q of the system with respect to the
displacement field.
• Therefore, if we write the displacement as   = ∑  !  , and try to
minimize the total energy of the system, then we have:
HS HS H
= =  !  =0
H H H
• And this is similar to the weighted residual formulation of the Galerkin
method, but it is carried out on an average sense (as integral).
• The integral G !     is called the Weak Formulation of the
differential equation whose residual is R(x).
26
 2/24/2020

Finite Element Method using Energy Method

• Galerkin is purely mathematical method and can be difficult to generalize for
higher elements. Thus, the energy approach will be used.
• Consider a 2-noded element with 1 DOF per node (ui, uj)
Typical Bar Element
dui ui Ε,Α uj du j
Pi = − AE Pj = − AE
dx (i) L (j) dx
(Two Degrees of Freedom)

• The total energy of the system Q is the difference between the internal strain
energy and the external energy Q = Ue – We.
• Strain-Displacement relation (Kinematic, or Compatibility conditions) is

T=

• Constitutive Material behavior (Strain-Stress relation) for elastic is:
U = T
27
 2/24/2020

Finite Element Method using Energy Method

• For the bar member
$ $  
VW = G G UT v = G T
T = G 
.
 
• The external work is XW = G   + ∑ Y ZY
• Assume the solution is   = ∑ Y !Y  , where Y are nodal
displacements at nodes j = 1, 2. !Y  are trial functions (called shape
functions).
• The solution must satisfy the displacements at the nodes. Since we have
two DOF per element, then assume linear function for u, hence   =
[ + \, but we would like to have a,b in terms of the nodal DOF.
• At x=0, u=ui and at x = L, u =uj, solving and re-writing gives:
    
  = 1 −  + Y = 1 −  = ]^
    Y 28
 2/24/2020

Finite Element Interpolation Functions

ui uj
x (local coordinate system)
(1) L (2)

u(x)

x
(1) (2)

Ν1(x) Ν2(x)

x
(1) (2)
Nk(x) – Lagrange Interpolation Functions
29
 2/24/2020

Finite Element Method using Energy Method

• The strain is
 ]
T= = ^ = _^
 
• Substitution:
1  
Π= > 
 − >   −  Y ZY
2  
1
Π = ^a > _ a 
_ ^ − ^a > ]  − ^a b
2
• Differentiation with respect to the vector d and setting equal to zero, gives:

> _ a 
_ ^ = > ]  + b

c d = G _ a 
_ and G ]  is the equivalent nodal forces, and F is the nodal
forces
30
 2/24/2020

Finite Element Method using Energy Method

• Doing the integrations gives:
ef 1 −1 g  1
cd = and G ]  = h
 −1 1  1

• This is the stiffness matrix for the elastic bar element assuming linear
behavior. So, this element is called a linear element [ Because the
interpolation functions are linear].
• We can also assume a quadratic element behavior, but for this we will require
3 nodes per element.
• Quadratic bar element will have a different 3X3 stiffness matrix

31
 2/24/2020

Finite Element – Quadratic Bar Element

[$
 = [$ + [  + [+   = 1    [ = ij
[+ u1 u2 u3
x
Use the boundaries of the element (1) (2) (3)
u1 = a1 $ 1 0 0 [$ L
 = 1 /2  /4 [
2
L L
u2 = a1 + a 2 + a3 u(x)
2 4 + 1   [+
^ = ilj
2
u3 = a1 + a2 L + a3 L
ψ x
From which we get the coefficients vector a: (1) (2) (3)

j = i l Am ^ And then the solution is:

Ν2(x)
Am Am
() = ij = i i l ^ = ]^, nodpd ] = i i l Ν1(x) Ν3(x)

Here N =[N1 N2 N3] 1

3 2  4 4   2 
!$ = 1 − +  , ! = −  , !+ = − + 
      (1) (2) (3)
x
32
 2/24/2020

Finite Element – Quadratic Bar Element

 ]
• The strain becomes: T = = ^ = _^, Here _ = [B1 B2 B3]

3 4 4 8 1 4
r$ = − +  , r = − , r+ =− + 
     

• The Element stiffness matrix c d = G _ a 
_ =

• The equivalent nodal force vector is

1/6
G ]  =   2/3
1/6

33