FEM II - Lecture 1 Page 1 of 14

_________________________________________________________________________________________________________________________________
FINITE ELEMENT METHOD II
Autumn 2015

Lectures (15h):
1. Accuracy, error estimation and adaptive remeshing
2. Heat flow and thermal stresses in FEM
3. Introduction to structural dynamics, free vibrations
4. Nonlinear problems in mechanics of structures - basic numerical techniques
5. Orthotropic materials and composite structures
6. Parametric modeling and design optimization

Computer lab (15h):

Modeling simple problems of: thermal stresses, contact mechanics, plasticity and residual stresses, free vibrations,
buckling, parametric modeling

References:
[1] Lecture notes from the web site: http://meil.pw.edu.pl/zwmik/ZWMiK/Dla-studentow2/Finite-Element-Method-II
[2] Moaveni S.: Finite element analysis. Theory and applications with ANSYS. Paerson Education, 2015.
[3] Kleiber M. (red.): Komputerowe metody mechaniki ciał stałych, seria Mechanika Techniczna XI, Warszawa PWN 1995.
[4] Xiaolin Chen, Yijun Liuv: Finite Element Modeling and Simulation with ANSYS. Workbench, CRC Press 2014
[5] Huebner K. H., Dewhirst D. L., Smith D.E., Byrom T. G.: The finite element method for engineers, J. Wiley & Sons, Inc.,
2001.
[6] Zienkiewicz O.C., Taylor R.: The Finite Element Method - different publishers and editions
[7] Krzesiński G., Zagrajek T., Marek P., Borkowski P.: MES w mechanice materiałów i konstrukcji. Rozwiązywanie
wybranych zagadnień za pomocą programu ANSYS, Of. Wyd.PW 2015
[8] Bijak-Żochowski M., Jaworski A., Krzesiński G., Zagrajek T.: Mechanika Materiałów i Konstrukcji, Tom 2, Warszawa, Of.
Wyd. PW, 2014
Assessment based on the final test and the results of computer lab work
FEM II - Lecture 1 Page 2 of 14
_________________________________________________________________________________________________________________________________

1. ACCURACY OF FE ANALYSIS.
ERROR ESTIMATION AND ADAPTIVE REMESHING

ε s = ws − wr – modeling error,
Real object
Real result wr
ε d = wd − ws – discretization error
geometry, boundary (initial) conditions
material properties, laws of physics ε n = wn − wd – numerical error
Total error:
Methematical model Exact solution ε c = ε s + ε d + ε n = wn − wr
(continuous) of the mathematical model ws

discretization
The most effective FEM analysis:
εs ~ εd ~ εn
approximation

Discrete Exact solution of the

discrete model wd
model

numerical calculations

Numerical result - wn

Approximate methods – flow chart

(valid also for other methods e.g. Boundary Element Method - BEM and Finite Differences Method - FDM)
FEM II - Lecture 1 Page 3 of 14
_________________________________________________________________________________________________________________________________
Modeling error εs

Depends on the accuracy of the available information about the problem and knowledge of the analyst (1D – 2D – 3D models, linear,
nonlinear, assumed simplifications, reliable information concerning material properties, loads.)

The example – wooden board

Different models for the problem

one dimensional model - beam

a) model belki Mathematical model:
N
q0 •1,2, 3 dimensional
m
•properties of wood: isotropic-orthotropic, moist-
dry, homogenous-inhomogeneous, rate dependent-
rate-independent)
b)two- dimensional
model płyty model - plate •contact interaction model (friction)
N
p0 •loadings
m2

three
c) dimensional
model model – solid
trójwymiarowy bryłyvolume

N
0
m3
FEM II - Lecture 1 Page 4 of 14
_________________________________________________________________________________________________________________________________

Discretization error εd

Depends on the mesh density, types of the elements – shape functions, the shapes of the finite
elements

MX

.186398 50 150 250 400

25 100 200 300

FE mesh and von Mises stress distribution Discrete solution versus exact solution of the continuous problem

There are some mathematical convergence requirements in FEA concerning the mesh, shape functions, and rules of FE model
building.
FEM II - Lecture 1 Page 5 of 14
_________________________________________________________________________________________________________________________________
Numerical error εn

[ K ]{q} = { F } ,
δq  {δ F } [δ K ] 
[ K + δ K ]{q + δ q} = {F + δ F } . → ≤ [K ] [K ]
−1
 + ,
q  {F}
 [K ] 


condition number of the matrix K

cond ( [ K ] ) = [ K ] [K ]
−1
.

Norm of a matrix (vector)– a measure of magnitude

L2- Euclidean norms
1
 2
{q} =  ∑ ( qi )  ,
2
Vector norm
 i 
1
 2 2
Matrix norm [ K ] =  ∑∑ ( kij )  (matrix norm induced by the vector norm)
 j i 

Max norms (L~ norms)

 
{q} = max qi , [K ] = max  ∑ kij  .
i i
 j 

A problem with a lsmall condition number is said to be well-conditioned, while a problem with a high condition number is said to be ill-conditioned.

cond(K) | ≥ 1 cond(K ) ≈ 1 - problem well-conditioned cond(K )» 1 - problem ill-conditioned

Reasons of ill-conditioning of the problems in FE stress analysis – great differences between stiffness of FE elements, unstable boundary conditions
FEM II - Lecture 1 Page 6 of 14
_________________________________________________________________________________________________________________________________

The example (ill conditioned problem):

kA kB
q1 q2 q3

F2 F3
k A −k A   q1   F1 
 k −k  k k     
[ k ]e =  k =  −k k −+kkA
 k−A k+ k B q − k B= F
q2  =  F2 
 −k k  
2× 2  −k B k B   q3   F3  The example of the very small perturbation

k A + kB −k B   q2   F2   F2  0.999 
q1 = 0  −k  =   = 
k B   q3   F3   F3   −1.0 
For changed force vector :
 B

q2   −0.001
kA = 1 k B = 1000  = 
Let’s assume:
 q3  −0.002 
 F2   1 
 =  δ F2   −0.001 δ q2   −0.001
 F3   −1 {δ F } =  =  {δ q} =  = 
The solution: δ
 3  0 
F δ q3   −0.001
1 1 
 q2   k A kA  F 
 =  2
 q3   1 1 1 F
+  3 δF δq
k k A kB  = 0.707 ⋅ 10−3 ⇒ = 1.414
 A F q
The result:
 q2   0 
 =  cond ( K ) ≈ 4i103
 q3   −0.001
FEM II - Lecture 1 Page 7 of 14
_________________________________________________________________________________________________________________________________
First equation
k A + kB F
q3 = ⋅ q2 − 2
kB kB
Second equation
F3
q3 = q2 + q3
kB
System ill-conditioned if
k A + kB
→1
kB
kA α1
→0 α2
kB
k A + kB
tgα 2 = tgα1 = 1
Sensitivity to the slop change kB
KA<<KB system ill conditioned

q2
Round–off error
As a general rule, if the condition number cond(K) = 10k, then you may lose up to maximum k digits of accuracy during the solution of the system of
linear equations. However, the condition number does not give the exact value of the maximum inaccuracy that may occur in the algorithm.

(
r ≥ p − log10 cond ( [ K ] ) )
p – number of significant digits in the computer representation of numbers
r – number of significant digits of the result

In FE models cond(K) reaches 108

FEM II - Lecture 1 Page 8 of 14
_________________________________________________________________________________________________________________________________
A posteriori error approximation techniques
Element and nodal solution in FE program postprocessors (PLESOL, PLNSOL in ANSYS)
FE solution provides the continuous displacement field ( from element to element), and the discontinuous stress field.
To obtain smooth stress distribution, the averaging of the stresses in the nodes is performed ( nodal stresses).
PowerGraphics EFACET=1
EFACET=1 AVRES=Mat
DMX =.086853 DMX =.086853
SMN =-14.333 SMN =-10.372
SMX =370.25
SMX =379.111
-10.372
-14.333 31.919
29.383 74.211
73.099 116.502
116.815 158.793
160.531 201.085
204.247 243.376
247.963 MN
285.667
MN
291.679 327.959
370.25
335.395
379.111

Y Y
Z X MX Z X MX

JUN 11 2008 JUN 11 2008

00:32:13 00:32:37
PLOT NO. 5 PLOT NO. 6
ELEMENT SOLUTION NODAL SOLUTION
STEP=1 STEP=1
SUB =1 SUB =1
TIME=1 TIME=1
SY (NOAVG) SY (AVG)
RSYS=0 RSYS=0
PowerGraphics PowerGraphics
EFACET=1 EFACET=1
DMX =.025111 AVRES=Mat
SMN =116.884 DMX =.025111
SMX =379.111 SMN =116.884
116.884 SMX =370.25
146.02 116.884
175.157 145.036
204.293 173.188
233.429 201.339
262.566 229.491
291.702 257.643
320.838 285.795
349.975 313.946
379.111 342.098
370.25
MX MX

Rectangular plate with a hole under tension. The model of the quarter of the structure. The stress component σy
Discontinuous „element solution” (left) and averaged continuous “ nodal solution” (right). Six-node triangular plane elements
FEM II - Lecture 1 Page 9 of 14
_________________________________________________________________________________________________________________________________
_________________________________________________________________________________________________________________
Basic relations between displacements, strains, stresses and strain energy within finite elements (the relations discussed during FEM I lectures)

Displacement field over the element is interpolated from the nodal displacements:
{u} = [ N ( x, y, z )] {q}e , where {q}e - nodal displacements vector , [N ] - shape functions matrix. u

For example for the simplest triangular element with 3 nodes and 6 DOF the relation is
 u1 
υ  LWE=3 LWE=6 LWE=4 LWE=8
 1
0  w
u ( x, y )   N1 ( x, y ) 0 N 2 ( x, y ) 0 N 3 ( x, y ) u2  ,
 =    elementy trójwymiarowe
υ ( x, y )   0 N1 ( x, y ) 0 N 2 ( x, y ) 0 N 3 ( x, y )  υ2  v
u
u3 
 
υ3 
where Ni are the linear functions
Shape functions Nij are usually polynomials defined in local (element) coordinate systems.
Displacements, strains and stresses within each element are defined as the functions of the nodal displacements
{u} = [ N ] {q}e ,
{ε } = [ R ] {u} = [ R ][ N ] {q}e = [ B ] {q}e , [B] – strain-displacement matrix, [R] -gradient matrix
{σ } = [ D ] {ε } = [ D ][ B ] {q}e .
The strain energy of the element Ωe is:

1
 q  e [ k ]e {q}e .
1
ε  {σ } d Ωe ,
1
U e = ∫  q  e [ B ] [ D ] [ B ] {q}e d Ωe , Ue =
2 Ω∫e
Ue =
T

2 Ωe 2

[ k ]e = ∫ [ B ] [ D ] [ B] d Ωe
T
where is the stiffness matrix of the element (symmetrical, singular, semi-positive defined)
Ωe
FEM II - Lecture 1 Page 10 of 14
_________________________________________________________________________________________________________________________________

Averaged stress vector

at node n
{σ}2
{σ} 1 {σ}nav = Σ {σ}ni / k
( k=6)
2
1

Node n

{σ}i

MX
MX discontinuous continous
stress vector at node n of element i
σ x 
ni

σ 
 y
σ 
{σ }n =  z  {σ }n ≠ {σ }n ≠ {σ }n ≠ ...
i 1 2 3

τ xy 
τ yz 
 
τ xz n
k
∑{σ }n
i

{σ }n =
av i =1
averaged stress vector at node n:
k
FEM II - Lecture 1 Page 11 of 14
_________________________________________________________________________________________________________________________________

{∆σ }n = {σ }n − {σ }n
i i av
Stress error vector at node n of element i
vector {∆σ } within the element ’ i’ may be determined by standard approximation using the stress error
i
The stress error
vectors at nodes of element i {∆σ }n .
i
Strain energy of the element i
1
U i = ∫ σ  {ε } d Ωi
i i

2Ω i

{ε } = [ D ] {σ } , D] – stress-strain matrix
−1
Then for each element so called energy error can be estimated
1
ei = ∫  ∆σ  [ D ] {∆σ } d Ωi
−1
1
σ  [ D ] {σ } d Ωi
i i
(ETABLE-SERR) −1
Ui =
i
∫
i
2Ω i 2Ω i
[
JUN 11 2008 JUN 11 2008
10:49:02 10:49:26
PLOT NO. 1 PLOT NO. 2
ELEMENT SOLUTION ELEMENT SOLUTION
STEP=1 STEP=1
SUB =1 SUB =1
TIME=1 TIME=1
SERR (NOAVG) SDSG (NOAVG)
DMX =.086853 DMX =.086853
SMN =.320E-08 SMN =.00959
SMX =.962E-03 SMX =12.223
.320E-08 .00959
.107E-03 1.367
.214E-03 2.724
.321E-03 4.081
.427E-03 5.438
.534E-03 6.795
.641E-03 8.152
.748E-03 9.509
.855E-03 10.866
.962E-03 12.223

MX
Y MX Y
Z X Z X

Initial mesh. SERR and SDSG error distribution

SDSG - ∆σi = maximum absolute value of any component of {∆σ }n = {σ }n − {σ }n for all nodes connected to element
i i av
FEM II - Lecture 1 Page 12 of 14
_________________________________________________________________________________________________________________________________

The energy error over the model

l .el .
e = ∑ ei
i =1
The energy error can be normalized against the strain energy
1
 e 2
SEPC = 100  
U + e 
U – total strain energy over the entire model

SEPC – percentage error in energy norm

The ei values can be used for adaptive mesh refinement. It has been shown that if ei is equal for all elements, then the model using
the given number of elements is the most efficient one.

This concept is also referred to as “error equilibration” ( ei = const , SEPC< S0).

FEM II - Lecture 1 Page 13 of 14
_________________________________________________________________________________________________________________________________

Adaptive Meshing Techniques

- Automatic refinement of FE meshes until converged results are obtained

- User’s responsibility reduced to generation a good initial mesh

TIME=1
START SY (AVG)
RSYS=0
DMX =.086851
SMN =-11.174
SMNB=-14.773
Initial mesh SMX =374.04
SMXB=375.198
-11.174
31.628
Mesh refinement 74.429
(different mesh 117.231
IF
generation techniques) 160.033
(ei – e/N )/(e/N )<eps1 202.834
SEPC<eps2 NO
245.636
288.437
MN
331.239
YES 374.04
Y Y

END Z X Z X MX

Final FE mesh and the results - Sy distribution (percentage error in

energy norm SEPC=0.811%, uniform error distribution ei= const)
FEM II - Lecture 1 Page 14 of 14
_________________________________________________________________________________________________________________________________

Selective adaptive meshing

If mesh discretization error (measured as a percentage) is relatively unimportant in some regions of the model, the procedure may be speed up by
excluding such regions from the adaptive meshing operations. Also - near singularities caused by concentrated loads and at boundaries between
different materials.

Selective adaptive meshing

in ANSYS

Types of refinement in adaptive meshing :

h-refinement: reduction of the size of the element (“h” refers to the typical size of the elements)
p-refinement: increase of the order of the polynomials on an element (shape functions from linear to quadratic, etc.)
r-refinement: re-arrangement of the nodes in the mesh
hp-refinement: combination of the h- and p-refinements.