You are on page 1of 42

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/232896029

Practical aspects of the finite element method

Article  in  Revue Française de Génie Civil · January 2002


DOI: 10.1080/12795119.2002.9692737

CITATIONS READS

7 323

3 authors:

Manuel Pastor Pablo Mira


Universidad Politécnica de Madrid Centro de Estudios y Experimentación de Obras Públicas
204 PUBLICATIONS   5,648 CITATIONS    52 PUBLICATIONS   792 CITATIONS   

SEE PROFILE SEE PROFILE

Jose Antonio Fernandez-Merodo


Instituto Geológico y Minero de España
102 PUBLICATIONS   1,470 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Project of Preservation and Conservation of Tiwanaku and Akapana Pyramid View project

RISKCOAST "Tools for preventing and managing coastal-georisks in the framework of Global Change" View project

All content following this page was uploaded by Manuel Pastor on 27 May 2014.

The user has requested enhancement of the downloaded file.


Practical Aspects of the
Finite Element Method
Manuel Pastor
Pablo Mira,
José Antonio Fernández Merodo

Centro de Estudios y Experimentación de Obras Públicas


CEDEX, Ministerio de Fomento, SPAIN

E.T.S. Ingenieros de Caminos


Universidad Politécnica de Madrid, SPAIN

Slide 1
1.- Introduction (I)
• FEM has become a powerful engineering analysis tool
• Many engineers use commercial codes : ANSYS, ABAQUS,......
• Beginners face questions such as :

i) Which element ?
ii) How many elements and how big ?
iii) Which material model ?
iv) Is there any saving in reduced integration

• This chapter adresses some of these questions


It does it not in an exhaustive way
Besides there are many other questions
It is just a pocket guide for travellers

2
Slide 2
1.- Introduction (II)
• The questions we will address are:

1) The misteries of bending

2) Volumetric locking
3) The risks of reduced integration

4) Failure loads
5) Why cannot we choose all elements for geotechnical analysis ?

6) Our favorite solver : do we have any?

3
Slide 3
2.- The misteries of bending (I)
• Not all elements perform satisfactorily under bending contitions
• Linear triangles and bilinear quadrilaterals produce a much stiffer
response than they should

• This extra stiffness produces an increase in natural frequencies

• Let us consider a square x∈[-1,+1], y∈[-1,+1]

M M

4
Slide 4
2.- The misteries of bending (II)

Continuum mechanics solution

• Plane stress problem

• Displacement field :

u  − EI
M
xy v  − EI
M
1 − x 2 

• Horizontal displacements at nodes : u 0  M


EI

−u 0 y
• Strain field  x  − EI
M
y  y  0  xy  0  0
0
5
Slide 5
2.- The misteries of bending (III)
Finite Element solution with linear quadrilateral

• Nodal displacements will be :


−u 0 u0 −u 0 u0
û1  û2  û3  û4 
0 0 0 0

• Element displacement vector : ûT  ûT1 ûT2 ûT3 ûT4


y
4 3

• Shape functions: x

N1  1
1 − x − y  xy N2  1
4
1  x − y − xy
4
1 2
N3  1
4
1  x  y  xy N4  1
4
1 − x  y − xy

6
Slide 6
2.- The misteries of bending (IV)

• B matrix :
−1  y 0 1−y 0 1y 0 −1 − y 0
B 0 −1  x 0 −1 − x 0 1x 0 1−x
−1  x −1  y −1 − x 1 − y 1  x 1  y 1 − x −1 − y

• Strain field at element level is therefore:


−u 0 y
  Bû  0
−u 0 x

• Spurious shear strain


7
Slide 7
2.- The misteries of bending (V)
• Computing external and internal work

W ext  2M  2Mu 0

W int    T d    T Dd  2  u 0 yEu 0 yd  2  u 0 xGu 0xd

 2u 0 EIu 0  2u 0 GIu 0  2u 0 E  GIu 0

and equating them we get u0  M


EGI

• Shear locking

8
Slide 8
2.- The misteries of bending (VI)

• With linear triangle T3 spurious shear strain also appears

u0 −u 0
2
1 2 
  0   0
1
u0 −u 0

• Constant strain in each element

• With quadratic elements T6 and Q8 spurious shear strain


does not appear

9
Slide 9
2.- The misteries of bending (VII)
• Poor bending performance for T3 and Q4

• Since they are simple and easy to program attempts have


been made to improve them

• Wilson&Taylor, Simo&Rifai: Q4+Incompatible modes


Quadratic enrichment. Distortion
• Pian-Sumihara: Additional stress field
• Battoz & Dhat: Beam test problem
• Test performed with T3,T6,Q4,Q8,Q4WT and Q4PS

• The most satisfactory Q4WT , Q4PS ,Q8,T6

10
Slide 10
Volumetric locking (I)
• Let us assume a material with a Poisson coeficient close to 0.5
• Bulk modulus: K  31−2
E
→∞
Y F

3 4 6
1
X
1 2 5

• With υ=0.5 and linear triangles none of the nodes move


• Two problems: critical parameter (υ=0.5 ) and element limitations (T3)
• Critical parameter causes ill conditioning of the coeficient matrix 11
Slide 11
Volumetric locking (II)
• Nearly all classical displacement formulations present this type
of limitations under incompressibility, in different degrees

• In 2D continuum mechanics problems, in incompressibility


conditions we have at each point
1) 2 degrees of freedom ndof=2
→ ndof/nrestr = 2
2) 1 restriction (εv=0) nrestr=1

• It would be desirable that in FE problems such ratio be maintained

12
Slide 12
Volumetric locking (III)
• A simple test : incompressible patch test (Nagtegaal,Hughes)
ndof = 2
nrestr ? u  a 1  a 2 x  a 3 y  a 4 xy
v  b 1  b 2 x  b 3 y  b 4 xy

x  a2  a4 y  y  b3  b4 x

 v  a 2  b 3   b 4 x  a 4 y
• Element incompressibility restriction

1 1 x 1 y1 a2 0 Coefficient matrix is rank 3


1 1 x 2 y2 b3 0

1 1 x 2 y2 b4 0
ndof
1 1 x 2 y2 a4 0 r nrestr
 2
3
13
Slide 13
Volumetric locking (IV)

• Ratio values for different classical element types with


full integration, in plane strain and axisymmetric conditions
from Sloan & Randolph (1982)

T3 T6 T10 T15 Q4 Q8 Q12 Q17


4 3 8 2 8
PlaneStrain 1 3 2 5 3
1 1 7
1 2 9 16 2 2 10 16
Axisymmetric 3 3 10 15 5 3 13 19

14
Slide 14
Risks of reduced integration (I)
• Reduced integration consists of using an integration rule of smaller
degree than required to integrate exactly polinomials existing in
the stiffness matrix of undistorted elements

• It has two advantages: a) Reducing computational cost


b) Improving incompressibility performance
(Volumetric locking)
• Two popular reduced integration rules are:
a) One point for bilinear quadrilaterals
b) Two by two for 8 noded quadrilaterals

• It is interesting to study the problem from the point of view of


deformation modes (eigenvectors of stiffness matrix)
• Let us consider the case of the bilinear quadrilateral
15
Slide 15
Risks of reduced integration (II)
• Deformations modes of bilinear quadrilateral

Tv Th Rot

Dil B1 B2

S1 S2

• Description
• Rigid body modes : Zero energy 16
Slide 16
Risks of reduced integration (III)

• Reduced integration: 1 point integration rule

• Stiffness matrix which in general is : K   B T . D. B d


Is computed as : K  B T0 . D. B 0 . W 0
• B is 8x3 and D is 3x3 → K is rank 3

K is 8x8 and rank 3 → 5 zero energy modes

• 5 zero energy modes = 3 rigid body modes +... 2 bending modes


Reduced integration of 8 noded quadrilaterals also causes
Additional zero energy modes

17
Slide 17
Risks of reduced integration (IV)
• Non rigid body zero energy modes might be a problem

• Spurious bending modes


• Ill conditioning → Hourglassing
• Hourglass control
• Selective integration

18
Slide 18
Failure loads (I)
• Elastoplastic problems also present quasi incompressible conditions
• Footing on vertical slope
Material properties
Tipo de material E(Pa)   y (Pa)
Suelo Von Mises 1.0E5 0.35 200.0
Zapata Elásticolineal 1.0E8 0.35

Perfect plasticity : H = 0.00


Softening : H= -0.01E

- Theoretical Solución

Using límit analysis:


- Failure mechanism direction 45º
- P = 1154.7 N
19
Slide 19
Failure loads (II)
• How does mesh orientation affect the finite element solution
of a failure mechanism problem ?

20
Slide 20
Failure loads (III)

t3st t6st, t6bb, t7st, t7bb t4st, t4bb, t4sr

“Right”
Orientation
(r)

“wrong”
Orientation
(w)

21
Slide 21
Failure loads (IV)
• Force-displacement diagrams for H = 0.00

Quadrilaterals Triangles

22
Slide 22
Failure loads (V)

• Failure mechanisms for H=0.00 with quadrilaterals and 3n triangles

23
Slide 23
Failure loads (VI)

• Failure mechanisms for H=0.00 with 6n and 7n triangles

24
Slide 24
5.- Why cannot we use all elements ..........? (I)
• Let us consider the saturated consolidation problem
σy σ 'y p
τ xy τ ' xy

τ xy τ ' xy
p
σx σx = σ 'x + p
σ 'x
τ xy τ ' xy

τ xy τ ' xy p
σ 'y
σy

• Space discretization

S − mp  b  0  T ′
B  d − Qp̄  f u

m Su̇ − ∇ k∇p 
T T ṗ
 ∇ k w b  0
T . .
Q∗ Q u Hp̄  C p̄  f T

• Discretizing in time and solving the non linear problem with NR


.
KT Δt −QΔt dΔ ū n  Gu
. −
−Q T Δt −Δt HΔt  C dΔp̄ n  −Δt G p 25
Slide 25
5.- Why cannot we use all elements ..........? (II)
• Undrained incompressible limit kw→0; Q*→∝, jacobian will be :
.
KT −Q dΔ ū n  Gu
. −
−QT 0 d Δp̄ n  −Gp

• Doesnt look good, determinant might be zero

• A complete mathematical treatment of this problem can be


found in Babuska(1973) and Brezzi(1973)

• A simpler a approach to the problem, although less complete can


be found in Zienkiewicz & Taylor patch test (1986)

26
Slide 26
5.- Why cannot we use all elements ..........? (III)

• Saturated soil column 1 m.


subject to surface load q = 100 exp(-iwt)
• Only uz allowed
tz = 0
• All boundaries impermeable
except top horizontal boundary ux= 0
30 m
∂ pw
Z =0
kw 10 −7 m/s ∂x
n 0.333 X
O
E 7.492 10 8 (Pa)
 ∂ pw
0.2 uz = 0 ux= 0 =0
∂z
s 2. 0  10 3 N/m 3 
w 1. 0  10 3 N/m 3 
27
Slide 27
5.- Why cannot we use all elements ..........? (IV)
• Problem was discretized 1
Pore-pressure distribution Pore-pressure distribution

with a column of 20
0.8

elements 0.8

0.6 0.6

z/L
q4p4

z/L
0.4 0.4

(a) (b)
• Results 0.2 0.2

a) Q* = 104MPa
0. 0.
0. 0.5 1 1.5 0. 0.5 1 1.5 2

•b) Q* = 109MPa p/q p/q

Pore-pressure distribution Pore-pressure distribution


1 1

0.8 0.8

Q8P4 Q8P4

0.6 0.6

q8p8
z/L

z/L
0.4 0.4

(c) (d)
Q* = Water and grain 0.2 0.2

compressibility 0.
0. 0.5 1 1.5 0.
0. 0.5 1 1.5
p/q p/q
28
Slide 28
5.- Why cannot we use all elements ..........? (V)
• Zienkiewicz & Taylor patch test (1986)

KT −Q dΔ ū n  Gu
. −
−QT 0 d Δp̄ n  −Gp

. −1
dΔ ū n   K Q. dΔp̄ n   K G u
−1 u p
.
Q K. QdΔp̄ n   − G p  Q T K −1 G u
T −1

n p × n p = (n p × nu )(nu × nu )(nu × n p ) } nu ≥ np
u p

• This is a necessary but not sufficient condition


Singularity would have to be tested in all cases
29
Slide 29
5.- Why cannot we use all elements ..........? (VI)
• Single element patch q4p4
• Constraints p
u

• nu = 0 < np = 3
Test not satisfied
Precribed displacement
Prescribed Pressure

• Single element patch t6p3

• nu = 0 < np = 2 u p
Test not satisfied

30
Slide 30
5.- Why cannot we use all elements ..........? (VII)

• Six element patch T6P3

• Constraints

• nu = 14 > np = 6
Test satisfied
u p

• Four element patch Q4P4 • nu = 2 < np = 8


Test not satisfied
• Four element patch Q8P4 • nu = 10 > np = 8
Test satisfied
• Conclusions
31
Slide 31
5.- Why cannot we use all elements ..........? (VIII)

KT −Q dΔ ū n  Gu
. −
−Q T
0 d Δp̄ n  −Gp

• If somehow the 0 is avoided there is no problem


• Stabilization methods

- Divergence, Fractional step; methods based in ideas


used in FEM for fluid mechanics
-Special implementation of Simo-Rifai

32
Slide 32
6.- Our favorite solver (Do we have any?)
• One of the main computational tasks in the FEM is linear
Equation system solving

• The issue is especially important with transient, non linear,


....multifield,.....3D analysis

• Which equation solver to use is therefore a crucial question to


answer when thinking of building your own FE code

33
Slide 33
6.- Our favorite solver (Do we have any?)
• Iterative solvers like preconditioned conjugate gradient are
excellent choices : Easy to program
Does not require renumbering

• Iterative methods are specially good for really large problems:


Tens or hundreds of thousands degrees of freedom

• For not so large problems (thousands of dregrees of freedom)


direct methods with special storage and renumbering schemes
are also good choices

• Non symmetric systems : skyline → some changes


conjugate gradient → GMRES
34
Slide 34
6.1 Conjugate gradient methods (I)

• Based on the idea that the solution of Ax = b minimizes


total potential   12 x T Ax − x T b
• It is straightforward to prove the previous idea
• The procedure would therefore be:
Given xk and πk
Find xk+1 such that πk+1 < πk

• Not only we want πk+1 < πk but πk+niter << πk


With niter not too large → we want fast convergence
• In the kth iteration we use a set of linearly independent vectors
p1, p2,....., pk and calculate the minimum potential in the space
spanned by these vectors, thus obtaining xk+1 and also pk+1 in
the process 35
Slide 35
6.1 Conjugate gradient methods (II)
• The algorithm can be summarized as follows:
• Choose the starting iteration vector x¹
• Calculate the residual r¹=b-Ax¹
if r¹=0
quit
else set p¹=r¹
do for k=1,2,..., until ‖rk ‖ ≤ 
T
rk rk
 
k
T
p k Ap k

x k1  x k   k pk
rk1  rk −  k Apk
T
r k1 r k1
 
k
T
r k Ar k
pk1  p k1
  k pk

enddo 36
Slide 36
6.1 Conjugate gradient methods (III)
• Two important orthogonality properties
pTi Apkj  0
PTj rk1  0 where P j  p1 , . . . , pj

• Convergence in at least n iterations.


It should be faster

• Rate of convergence depends on condition number of A


Cond(A)=λn/ λ1 where λn is the largest eigenvalue of A
and λ1 is the smallest eigenvalue of A

• Large condition number → Determinant closer to 0


Small condition number → Closer to identity matrix

37
Slide 37
6.1 Conjugate gradient methods (IV)

• To increase rate of convergence → decrease condition number


−1 −1
• Instead of solvingAx  b, solve A Ax  A b
Where A is the preconditioner, easy to invert and as close as
possible to A-1
−1
A = diag(A) → Jacobi conjugate gradient
Good for diagonal dominant A matrices
−1
• Algorithm with preconditioner introduces: z  A rk
k
T
z k rk
 
k
T
p k Ap k
T
z k1 r k1
 
k
T
z k rk

pk1  zk1   k pk

38
Slide 38
6.2 Skyline storage scheme (I)

• Using a direct solution scheme and storing the hole nxn


coeficient matrix is not an option considered in FE analysis

• To do so would imply :
a) A huge waste of storage space given the sparsity of
The coefficient matrix (many coefficients are 0)
b) A huge waste of computational effort since matrix operations
on null coefficients are trivial
• Because of this it is important to use an efficient storage scheme

• The skyline storage scheme is a good alternative

39
Slide 39
6.2 Skyline storage scheme (II)
• It consists of storing the stiffness matrix in a vector including the
diagonal terms and the off diagonal terms between the diagonal
And the farthest non zero off diagonal term, for each row (column)

• Additionally it will be necessary to use an N component integer


Vector to store the position in the stiffness vector of each diagonal
component

2 −2 0 0 −1 i 1 2 3 4 5 6 7 8 9 10 11 12
−2 3 −2 0 0 ai 2 −2 3 −2 5 −3 10 −1 0 0 4 10
0 −2 5 −3 0
i 1 2 3 4 5
0 0 −3 10 4
jdiagi 1 3 5 7 12
−1 0 0 4 10

40
Slide 40
7.- Concluding remarks
View publication stats

• We have addressed a few practical aspects of FE analysis

• Bending behaviour: Standard low order displacement formulations


Quadratic formulations
Improved low order formulations : mixed.....
• Volumetric locking : Standard low order formulations are bad
Higher order formulations are better but ......

• Reduced integration : It is possible solution for volumetric locking


But careful with it : Hourglassing
Hourglass control or selective integration
• Coupled formulations : Standard u-p formulations → nu > np u8p4
Stabilized : fractional step, Simo-Rifai, div
• Solvers : Iterative: easy to program, very good for large problems
Direct with special storage
Slide 41
and renumbering, also good 41

You might also like