ME623: Finite Element Methods in Engineering Mechanics

Instructor: Sumit Basu

Email: [email protected]
Phone office: (0512 259) 7506
Office: NL211A
Phone residence: (0512 259) 8511
Send assignments to
Ramkumar S [email protected] (odd numbered assignments)
and
Deependra Singh Sekhawat [email protected] (even numbered
assignments)
Evaluation
1. Assignments: 20%
2. 2
nd
mid-semester exam: 30%
3. End-semester exam: 30%
4. ANSYS project: 20%
Experiments versus simulations?
Qs. Sophisticated experiments can tell everything. Why do we need the FE method?
1: Experimental results are subject to interpretation. Interpretations are as good as the
competence of the experimenter.
2. Experiments, especially sophisticated ones, can be expensive
3. There are regimes of mechanical material behaviour that experiments cannot probe.
4. Generality of behaviour is often not apparent from experiments.
Experiments and simulations are like two legs of a human being. You need both to walk and it does not
matter which you use first!
1943: Richard Courant, a mathematician described a piecewise polynomial
solution for the torsion problem of a shaft of arbitrary cross section. Even holes.
The early ideas of FEA date back to a 1922 book
by Hurwitz and Courant.
His work was not noticed by engineers
and the procedure was impractical at the time due to the lack of digital computers.
A short history of FEA
1888-1972: b in Lublintz Germany
Student of Hilbert and Minkowski in Gottingen Germany
Ph.D in 1910 under Hilberts supervision.
1934: moved to New York University, founded the Courant Institute
In the 1950s: work in the aircraft industry introduced FE to practicing engineers.
A classic paper described FE work that was prompted by a need to analyze
delta wings, which are too short for beam theory to be reliable.
1960: The name "finite element" was coined by structural engineer Ray Clough
of the University of California
By 1963 the mathematical validity of FE was recognized and the
method was expanded from its structural beginnings to include heat transfer,
groundwater flow, magnetic fields, and other areas.
Large general-purpose FE software began to appear in the 1970s.
By the late 1980s the software was available on microcomputers,
complete with color graphics and pre- and post-processors.
By the mid 1990s roughly 40,000 papers and books about
FE and its applications had been published.
Professor emeritus of Structural Engineering at UC Berkley
Ph.D from MIT
Well known earthquake engineer
Books
Concepts and applications of Finite element analysis: Cook, Malkus and
Plesha, J ohn Wiley and Sons, 2003.
T.R. Chandrupatla and A.D. Belegundu, Introduction to Finite Elements in
Engineering, Second Edition, Prentice-Hall, 1997.
O. C. Zienkiewicz and R. L. Taylor, The Finite element method, vols 1 and
2, Butterworth Heinemann, 2000
Klaus-J urgen Bathe, Finite Element Procedures (Part 1-2), Prentice Hall,
1995.
Solving an engineering problem
Mathematical model: an
equation of motion
Eulers explicit scheme or first order Runge Kutta
scheme
Write a MATLAB code to
integrate the discretised
equations of motion with different
timesteps. Use l=1, g=10, initial
velocity=0, position=45
o.
Compare with the exact solution.
x
x
Alternately, the FEM idea.
A simple example: another look at FEM
A 2-d truss with elements that can
only withstand tension.
x
y
1
2
3
4
5
2
4
Direction of stretch
2 degrees of freedom (dof) per node
2
4
X
Y
2
2/1
4/2
x
y
Local stiffness matrix in local coordinate
system (ndof X nnod, ndof X nnod)
Local force vector (ndof X nnod,1)
Displacement vector (ndof X nnod, 1)
ndof : no. of dofs/node
nnod: no. of
nodes/element
2
4
x
y
Transformation matrix
3/1 4/2 7/3 8/4
3/1
4/2
7/3
8/4
Eg: 2,2 4,4
3,1 7,3
2
4
x
y
1
2
3
4
5
10 4
9 3
4 2
3 1 2
10 4
9 3
2 2
1 1 1
Destination Local dof Elemno
1
2
3
4
5
1
2
3
4
5
6
7
After assembling elements 1 and 2
X+X+X X+X+X X X X X X X
X+X+X X+X+X X X X X X X
X X X+X+X X+X+X X+X X+X X+X X+X
X X X+X+X X+X+X X+X X+X X+X X+X
X X X+X X+X X X
X X X+X X+X X X
X X X X X+X+X X+X+X X X
X X X X X+X+X X+X+X X X
X X X X X+X X+X
X X X X X+X X+X
Stiffness matrix is symmetric, diagonally dominant, sparse and banded.
1
2
3
4
5
1
2
3
4
5
6
7
Sparsity pattern
1
3
5
4
2
1
2
3
4
5
6
7
Half bandwidth
Global force vector Global stiffness matrix
Global displacement vector
Notes:
Global stiffness matrix is singular i.e. it has zero eigenvalues
Hence it cannot be inverted!
Nave approach for imposing displacement boundary conditions
L=a very large number. Also replace the corresponding dofs in the rhs vector by
LXspecified displacement value
The proper way of imposing displacement constraints
Suppose dof k is specified
Traspose negative of the specified value X kth column to the right
Replace k th row and columns in the stiffness matrix to zero
Replace K(k,k) by 1
Set F(k)=specified value
Repeat above steps for all specified dofs.
Solving the equations
K: requires huge storage, largest component of a FE code
Strategies:
Use sparsity: K_sparse=sparse(K);
Matlab command U=inv(K_sparse)*F;
Karl Friedrich Gauss(1777-
1855)
Wilhelm J ordan (1842-1899)
1
st
elimination:
Row2- Row 1*(-6/18)
Row3-Row1*(-6/18)
2
nd
elimination
Row3-Row2*(-2/10)
Row4-Row2*(-6/10)
3
rd
elimination:
Row4-Row3*(-7.2/9.6)
Original problem
Direct methods: Gauss J ordan elimination
Write a code in Matlab to solve a nXn system of linear equations by Gaussian
elimination with pivoting. Check the code on a small system. Take bigger and bigger
systems and plot the time taken by your computer (use the Matlab tic function) against
the system size n. Fit a curve to this plot.
How does the solution procedure work?
A digression: a little bit of linear algebra
Linear independence
Boundary conditions
Force specified: eg. dof 9 and 10 in our example
Displacement specified: eg. dof 1,2,and 6 in our example
Both forces and displacements cannot be specified at the same dof.
Eigenvalues and eigenvectors
Change of basis
Eigenvalues are real if A is real
Eigenvalues are orthogonal
Spectral Decomposition
Norms
A little more about solving systems of linear equations
Iterative method of solving a system of linear equations:
J acobi method
An example
Method of conjugate gradients for solving a system
of linear simultaneous equations
All examples are sourced from An introduction to conjugate gradient method without the
agonizing pain by J R Shewchuk of Carnegie Mellon University. Available on the web.
Examples of +ve definite, -ve
definite, singular and indefinite
systems in 2-d
(from Shewchuk)
Convergence of the example problem starting from (2,-2)
A brief description of the conjugate gradient method
Using coordinate axes as
search directions..
\newcommand{\bm}[1]{\mbox{\boldmath $#1$}}
Note that all previous search
vectors need to be stored. O(N
3
)
operations are needed for the
conjugation
Thus conjugate gradient method works in the following way:
Write a Matlab code to perform the conjugate gradient method on a 2X2 square,
positive definite system of linear simultaneous equations.
The FEM scheme of things
Read in data mainly nodal
coordinates, element connectivity,
force and displacement boundary
conditions and material properties
For each element
Form local stiffness matrix
Assemble into global stiffness matrix & rhs vector
End
Form local rhs vector
Form destination array
Incorporate boundary conditions into stiffness and rhs
Solve
Stop
Write a matlab code that
I.reads in the nodal coordinates and element connectivities of any 2-d truss
structure and displays the structure
2. Reads in the applied forces
3. Reads in the displacement boundary condition information
4. Prepares the destination array
5. Computes the local stiffness in the global coordinate system for each element
and using the destination array puts them into appropriate locations in the global
stiffness
6. Applies boundary conditions
7. Solves the system and displays the deformed truss
Method of steepest gradient in action.
Stiffness matrix from basics!
1
2
L
x
22
12
13
Intuitive but not easy
How do we solve a beam problem with
distributed loads?
wL/4
wL/4
wL/2
wL/8
wL/4
wL/4
wL/4
Introduction to the theory of elasticity
Number of independent elastic constants = 21
Number of elastic constants when there exists one plane of symmetry = 13
Number of elastic constants when there exists three mutually perpendicular planes of
symmetry = 9
Number of independent elastic constants for an Isotropic material = 2
X
1
,X
1
X
2
,X
2
X
3
X
3
Strain displacement relationships
On
L/2
w(x)
2h
y
x
x
y

0
w
H
a
x
y

0
w
H
a
L
a
2h
P
x
x
Plane strain
Plane stress
Axisymmetry
Axisymmetry
Geometry as well as loading should be axisymmetric!
Plane strain
Axisymmetric
C
L
Axisymmetric geometry, Non-
axisymmetric loading
Digression: A little bit of variational calculus
a
b
u(x)
a
b
u(x)
u(x)
y(x)
(x)
Euler equation
Some more interesting problems
Euler equation for several independent variables
This is zero on the boundary
Necessary condition for a function v(x,y) minimising or maximising the
functional J
u(x,y)
v(x,y)
(x,y)
Weak anchoring
What about the other way round?
Linear, differential operator
Example :
Laplace equation
Solving Laplace equation with given boundary conditions is equivalent to minimising
the variational statement
The other way round: strong form to weak form
Example
The route to the variational principle starting from an self-adjoint pde is outlined. The route
for non-self-adjoint cases is not known.
Principle of minimum potential energy
We assume the existence of W
Potential of the loads
Internal energy
Finding a displacement field satisfying the boundary conditions is equivalent to
minimising the above variational statement.
Bar of c/s A, modulus E loaded
by a body force cx where c has
units of force/area
Assignment: This is the Rayleigh-Ritz technique where the problem is reduced to
that of determining a few constant coefficients. Carry out the solution with three
undetermined constants.
Principle of virtual work: weak form of the eqns of equilibrium
Internal virtual work
External virtual work
Virtual work equation
Strong form to weak formanother example (from Reddy,
pp69)
Shear force
Bending moment
A 1-d problem with exact solution
L
x
Piecewise linear approximation: the assumed
displacement method
L
x
Strain displacement matrix
Shape function matrix
V
S
2-d domain, triangular elements
Linear triangle, 3 nodes/element, 2
dofs/node
Shape functions
Strain-displacement matrix
Strains are constant within an element: Constant strain triangle (CST)
x
y
f
S3x
2
3
1
f
S3y
f
S2x
f
S2y
(2,0)
(2,2)
Calculation of nodal forces due to surface tractions:
example
t=thickness of the element
A=surface area of the element
t
A
Assignment 7: Write a Matlab function to generate the plane strain CST
stiffness matrix given the elastic constants and the coordinates of the three
nodes
Example
x
y
El 1
El 2
1
2
3
4
300 psi
1000 lb
3 in
2 in
Thickness (t) = 0.5 in
E= 3010
6
psi
=0.25
(a) Compute the unknown nodal displacements.
(b) Compute the stresses in the two elements.
Realize that this is a plane stress problem and therefore we need to use
psi
E
D
7
2
10
2 . 1 0 0
0 2 . 3 8 . 0
0 8 . 0 2 . 3
2
1
0 0
0 1
0 1
1

Step 1: Node-element connectivity chart

2
4
Node 3
3
3
Area
(sqin)
4
2
Node 2
3 2
1 1
Node 1 ELEMENT
0 0 4
2 0 3
2 3 2
0 3 1
y x Node
Nodal coordinates
Step 3: Compute element stiffness matrices
7
) 1 (
T
) 1 ( ) 1 (
T
) 1 ( ) 1 (
10
2 . 0
0 5333 . 0
0 2 . 0 2 . 1
3 . 0 0 0 45 . 0
2 . 0 2 . 0 2 . 1 3 . 0 4 . 1
3 . 0 5333 . 0 2 . 0 45 . 0 5 . 0 9833 . 0
B D B ) 5 . 0 )( 3 ( B D B

=
= = At k
u
1
u
2
u
4
v
4
v
2
v
1
Step 2: Compute strain-displacement matrices for the elements

=
3 3 2 2 1 1
3 2 1
3 2 1
0 0 0
0 0 0
2
1
b c b c b c
c c c
b b b
A
B
Recall
1 2 3 3 1 2 2 3 1
2 1 3 1 3 2 3 2 1
x x c x x c x x c
y y b y y b y y b
= = =
= = =
with
For Element #1:
1(1)
2(2)
4(3)
(local numbers within brackets)
0 ; 3 ; 3
0 ; 2 ; 0
3 2 1
3 2 1
= = =
= = =
x x x
y y y
Hence
0 3 3
2 0 2
3 2 1
3 2 1
= = =
= = =
c c c
b b b

=
2 0 0 3 2 3
0 0 3 0 3 0
0 2 0 0 0 2
6
1
) 1 (
B
Therefore
For Element #2:

=
2 0 0 3 2 3
0 0 3 0 3 0
0 2 0 0 0 2
6
1
) 2 (
B
7
) 2 (
T
) 2 ( ) 2 (
T
) 2 ( ) 2 (
10
2 . 0
0 5333 . 0
0 2 . 0 2 . 1
3 . 0 0 0 45 . 0
2 . 0 2 . 0 2 . 1 3 . 0 4 . 1
3 . 0 5333 . 0 2 . 0 45 . 0 5 . 0 9833 . 0
B D B ) 5 . 0 )( 3 ( B D B

=
= = At k
u
3
u
4
u
2
v
2
v
4
v
3
Step 4: Assemble the global stiffness matrix corresponding to the nonzero degrees of
freedom
0
1 4 4 3 3
= = = = = v v u v u
Notice that
Hence we need to calculate only a small (3x3) stiffness matrix
7
10
4 . 1 0 2 . 0
0 983 . 0 45 . 0
2 . 0 45 . 0 983 . 0

= K
u
1
u
2
v
2
u
1
u
2
v
2
Step 5: Compute consistent nodal loads

=
y
x
x
f
f
f
f
2
2
1

=
y
f
2
0
0
y
S y
f f
2
1000
2
+ =
The consistent nodal load due to traction on the edge 3-2
lb
x
dx
x
dx N
tdx N f
x
x
x
S
y
225
2
9
50
2
50
3
150
) 5 . 0 )( 300 (
) 300 (
3
0
2
3
0
3
0
2 3
3
3
0
2 3
3
2
=

=
=
=
=

=
=

=

3 2
3
2 3
2
x
N =

lb
f f
y
S y
1225
1000
2
2
=
+ =
Hence
Step 6: Solve the system equations to obtain the unknown nodal loads
f d K =

1225
0
0
4 . 1 0 2 . 0
0 983 . 0 45 . 0
2 . 0 45 . 0 983 . 0
10
2
2
1
7
v
u
u
Solve to get

in
in
in
v
u
u
4
4
4
2
2
1
10 9084 . 0
10 1069 . 0
10 2337 . 0
Step 7: Compute the stresses in the elements
) 1 ( ) 1 ( ) 1 (
d B D =
With
[ ]
[ ] 0 0 10 9084 . 0 10 1069 . 0 0 10 2337 . 0
d
4 4 4
4 4 2 2 1 1
) 1 (

=
= v u v u v u
T
Calculate
psi

=
1 . 76
1 . 1391
1 . 114
) 1 (

In Element #1
) 2 ( ) 2 ( ) 2 (
d B D =
With
[ ]
[ ]
4 4
2 2 4 4 3 3
) 2 (
10 9084 . 0 10 1069 . 0 0 0 0 0
d

=
= v u v u v u
T
Calculate
psi

=
35 . 363
52 . 28
1 . 114
) 2 (

In Element #2
Notice that the stresses are constant in each element
Quadratic traingle (Linear strain triangle)
Can geometry be simplified (plane stress,
strain, axisymm, 3d, ymmetries?)
Material properties? Material? Elastic or
elasto plastic. Model for constitutive
behaviour?
Simplified geometry.
What elements to use?
Mesh. Insert boundary
conditions appropriately.
Solve.
Post process, check the
results.
Solution involves forming
element stiffnesses and loads,
assembling, applying bcs and
solving.
Interpolation and shape functions
The intepolation functions are same as the shape functions we got for the 1-d example.
1
2
3
Quadratic interpolation in 1-d
Bilinear rectangle (through interpolation formulas)
2a
2b
x
y
This is equivalent to assuming a displacement variation with <1 x y xy> terms
An example: Classical Kirchoff Plate
Completeness, compatibility etc.
Requirements of a FE model
Monotonic convergence
Completeness Element must represent rigid body modes + constant strain states
Element must rotate and translate
as a rigid body.
orthonormality
Generalised eigenproblem
Parasitic shear gets worse with (a/b) ratio. This condition is known as shear locking.
For completeness we require:
Rigid body modes to be represented (in 2-d two translations and one rotation)
Constant strain states must be represented (in 2-d two normal and one shear strain)
Interelement continuity of displacements must be satisfied.
Rigid body modes for a plane stress quad
Flexural modes, 0.495, 0.495
Shear mode, 0.769
Stretch mode,
0.769
Uniform
extension mode
1.43
Parasitic shear makes the Q4 element unusable in bending
More on convergence
Starting from a trial function and always adding more terms to it (equivalent to h 0 in
FEM) will always make the potential go towards a lower value.
Decrease in the total potential implies increase in the strain energy. Thus predicted
stiffness is always higher than actual in FE analysis.
Spatially anisotropic elements depend on coordinate
orientations.
Rate of convergence depends on the completeness
of the polynomial used.
Pascals triangle
Isoparametric element formulation
First proposed by: B.M. Irons, Engineering applications of numerical integration in
stiffness methods, AIAA J ournal, v4, n11, 1966, 2035-2037.
The Q4 element revisited
1
2
3 4
r=-1
r=1
r=1/2
s=-1
s=1/2
s=-1/2
2
3
4
s=1
r=-1/2
r
s
x
y
Parent element Physical element
1
1
1
1
Calculation of gradients in 2-d
J
J acobian matrix
Follows from the Q4 element done
earlier. Put a=b=1.
Stiffness matrix needs to be integrated numerically.
Numerical integration: Gauss quadrature-1d
r=1
r=-1
Sampling points between -1 and 1
Weights
r=-1
r=1
The idea behind a quadrature scheme.
0.5555
0.5555
0.8888
0.77459,
-0.77459
0
3
0.34785
0.34785
0.65214
0.65214
0.86114
-0.86114
0.33998
-0.33998
4
1
1
0.57735,
-0.57735
2
2 0 1
weights Sampling points n
Gauss quadrature: sample points and weights
Important rule: n point Gauss quadrature integrates a polynomial of order 2n-1 exactly
GI in 2-d
For the Q4 element, 2X2 quadrature is required
1
r s
r
2
rs s
2
r
3
r
2
s rs
2
s
3
2X2
Write a MATLAB code to formulate the stiffness matrix of a 4 noded quadrilateral
isoparametric element given the nodasl coordinates. Use Gauss quadrature.
Assume material properties as E = 200 Gpa and Poissons ratio=0.3.
Term paper topics
Group 1: Model the problem of a line load on a half space. Compare with the theoretical
solution of stresses near the load point. Refine the mesh to see if you converge to the exact
solution.
Group 2: Model a hard elliptical inclusion in a softer matrix. The matrix is infinitely large
compared to the inclusion and is subjected to uniaxial stresses applied at infinity. Find out
the theoretical solution to this problem and check how the strains inside the inclusion vary
with its ellipticity. Change the Poissons ratio of the matrix and the inclusion and see how the
stress fields in the matrix and inclusion change.
Group 3: Model a sharp crack in a infinite body. Find out the theoretical solution to this
problem. The stresses at the crack tip should be infinite. Compare with the theoretical
solution and see whether with mesh refinement you can get close to the theoretical solution.
Group 4: Model the problem of a series of periodically placed holes in a thin film. This
problem has already been discussed. Discuss further with Dr Ghatak.
Group 5: Model the bending of a functionally graded beam where the gradation is
exponential in the depth direction. Compare with the solution for a homogeneous beam to
see the differences due to the property variation.
Group 6: Model an internally pressurised hollow cylinder in plane strain. Start with a thick
cylinder and using theoretical results show how, as you move towards a thin cylinder the
solution changes.
Group 7: Analyse using you own code a deep beam using a structured mesh composed of
linear strain and constant strain triangles. Check all stresses with theoretical estimates.
Check convergence with mesh refinement.
Group 8: Verify the Euler buckling load of a slender beam using Finite Elements. Your
solutions may break down at the point of buckling. How good are your estimates compared
to the theoretical solutions?
Group 9: Analyse using your own code a internally pressurised plane strain thick cylinder
and check the results with theoretical estimates.
Group 10: Model a infinite wedge with tractions on one of the boundaries. Verify the stress
solutions near the tip of the wedge with theoretical estimates. What happens in the case of
a right angled wedge subject to shear tractions on one edge?
Group 11: A beam is rotating at a fixed angular velocity about its geometric center. Find the
stress distribution in the beam and compare with theoretical results.
Group 12: Solve the stresses for a 90 degree curved beam of inner radius a and outer
radius b subjected to a normal sinusoidal loading Asin( ). Find solutions to the problem as
varies from 0 to 2. Comment on the solution at =1;
Group 13: Solve the problem of a heavy beam on a elastic foundation. Compare your
results with theoretical solutions for different beam aspect ratios. Investigate when the
theoretical solution breaks down.
Group 14: An infinite plate subjected to a remote tensile load contains an elliptical hole.
Using a FEM simulation determine the stress concentration at the tip of the ellipse as a
function of the ellipticity of the hole. Compare with theoretical results.
Step by step guide to formulating a problem in FEA
Eg. Heat conduction in a 2-d domain
Q1. What are the governing equations and boundary conditions for the problem?
Q2. What is the variable to be solved for?
T(x,y) a scalar quantity 1 dof per node.
Q3. Does a variational statement exist?
Check for self-adjointness if self adjoint write variational principle
In this case self adjointness exists and hence
is the variational principle. Thus the correct T(x,y) makes the
above a minimum and also satisfies all the boundary conditions.
If variational principle does not exist, use a weighted residual method (we will learn
about it later)
Q.4 Variational principle exists. Now what?
If we use the Rayleigh Ritz principle, we should start the discretisation right away. Will
deal with that route later..
Q5. What element to choose?
Tricky. Eigenvalue analysis may help. Experience may too.
Let us choose a 4 noded iso-p quad in this case.Thus
What if we used the Rayleigh Ritz technique?
Define geometry
Mesh geometry
Loop over elements
Form local stiffness
Form surface force vector if required
Assemble stiffness
Assemble forces
end
Apply essential boundary conditions
Solve
Post process results
Method of weighted residuals
weights
Example: beam bending
Example 2: Heat transfer in 2-d
Some special elements: singular elements
1
2
3
4
5
6
7
8
9
Include if node I is defined
x
y
Vanishes at x=0 J acobian is
singular at x=0
2
3
5
7
1,4,8
6
Singularity exists all over the element.
Eg. Show that singularity exists along x axis.
What happens if 1,4,8 are given different node
numbers?
Infinite Finite elements
1
1
2
3
a
Errors in FE analysis
Round off errors in ill-conditioned
systems
P
Discretisation error
From Cook, Malkaus
and Plesha (2002)
In a linear element, displacement error is proportional to the square of element size while
strain error is linearly proportional to element size.
Displacements are more accurate near the nodes, strains are accurate inside elements.
Eg. For a problem meshed with 3 noded triangles, p=1
Thus error in energy O(h
2
)
Change to 6 noded triangle: p=2 and energy error O(h
4
)
Reduce element size by half: error reduced by a factor of 4.