You are on page 1of 34

The Stochastic Finite Element Method: Theory and

Applications

George Stefanou

Institute of Structural Analysis and Seismic Research


National Technical University of Athens, Greece

Clermont-Ferrand
June 2009

4/6/2009

IFMA Seminar, France

Outline of the presentation


Stochastic processes and fields
Simulation of Gaussian stochastic processes/fields
Simulation of non-Gaussian stochastic processes/fields
The stochastic finite element method (SFEM)
SFE (static) analysis of shells with random material and geometric properties response variability calculation
SFE-based stability analysis of shells with random imperfections
Nonlinear dynamic analysis of frames with random material properties under
seismic loading

4/6/2009

IFMA Seminar, France

Simulation of Gaussian stochastic fields


The spectral representation method (Rice 1954, Shinozuka & Deodatis 1991)

Fundamental theorem:

f ( x) = [ cos( x)du ( ) + sin( x)dv( ) ]


0

N 1

f ( x ) = An cos( n x + n )

Spectral representation of field f ( x ) :

n=0

An = 2 S ff ( n )

n = n

n = 0,1, 2,..., N 1

Truncated form (finite number of terms)

u
N

9 f ( x ) is asymptotically a Gaussian stochastic field as N due to the central


limit theorem

4/6/2009

IFMA Seminar, France

9 Its mean value and autocorrelation function are identical to the corresponding
targets as N
9 f ( x ) is periodic with period 0:

The Fast Fourier Transform (FFT):

(i )

Bn = 2 An ein

n = 0,1,..., M 1
p = 0,1,..., M 1

T0 =

M 1 i np 2
f (i ) ( px ) = Re B e M
n
n =0

M is the number of points on which


f is generated: M = 2

9 FFT-based sample functions are always generated on a domain equal to one


period 0
9 The reduction of step x (stochastic mesh refinement) leads to a larger number
of terms M in the FFT series
4/6/2009

IFMA Seminar, France

The Karhunen-Love expansion method (Love 1977, Ghanem & Spanos

1991)

f ( x, ) = ( x) + n n ( x) n ( )

f ( x, ) = ( x) + n n ( x) n ( )

n =1

n =1

Exact form (infinite terms)

Truncated form (finite number of terms)

n , n ( x) eigenpairs of the covariance function C ff ( x1 , x2 ) obtained from the solution of


the following eigenvalue problem:

ff

( x1 , x 2 ) n ( x1 )dx1 = n n ( x 2 ) (Fredholm integral equation of the second kind)

9 Exact solution feasible only for simple geometries and special forms of C ff ( x1 , x2 )
9 Numerical solution with the conventional Galerkin approach
very costly to compute and invert
4/6/2009

Dense matrices

IFMA Seminar, France

Some remarks on K-L expansion


9 Few K-L terms are needed for the simulation of strongly correlated (narrowbanded) stochastic fields
9 Few K-L terms are needed for the simulation of stochastic fields with smooth
(differentiable) target covariance function
9 More efficient simulation when the exact (analytical) solution of the eigenvalue
problem is feasible (Huang et al. 2001)
9 The variance of the stochastic field is underestimated and the simulated field f ( x, )
is in general not homogeneous since its variance is a function of x (Field & Grigoriu
2004):
N

n =1

n =1

Var[ f ( x, )] = [ n n ( x )] 2 [ n n ( x )] 2 = Var[ f ( x, )]

4/6/2009

IFMA Seminar, France

Non-ergodic characteristics of K-L sample functions (Stefanou & Papadra-

kakis 2007)

- Ergodicity in the mean: Pr lim f ( x)


L

= E [ f ( x)] = 1

Proof:
Without any loss of generality, we consider that E [ f ( x)] = f ( x) = 0

f ( i ) ( x)

lim f (i ) ( x)
L

L
1 L (i )
1 L N
1 N
(i )
(
)
(
)
f
x
dx
=

x
dx
=
n n(i ) n ( x)dx

n
n
n

0
L 0
L 0 n =1
L n =1

L
1 N

= lim n n(i ) n ( x)dx


0
L L
n =1

N
1 L

= n n(i ) lim n ( x)dx


L L 0

n =1

9 Existence of limit
9 Right hand side of the expression:
random variable
9 K-L sample functions: in general
not ergodic in the mean

=?

4/6/2009

IFMA Seminar, France

- Ergodicity in autocorrelation: R (ffi ) ( ) = RTff ( ) f ( i ) ( x)


Proof:

R (ffi ) ( ) = f ( i ) ( x + ) f ( i ) ( x )

1 L (i )
f ( x + ) f ( i ) ( x ) dx

0
L

L
1 LN N
1 N N
(i ) (i )
(
)
(
)

x
+

x
dx
=
n mn(i)m(i) n (x + )m (x)dx

n
m
n
m
n
m

0
L 0 n=1 m=1
L n=1 m=1

lim f ( i ) ( x + ) f ( i ) ( x)

L
1 N N

= lim n m n( i ) m( i ) n ( x + )m ( x)dx
0
L L
n =1 m =1

N N
1 L

= n mn(i)m(i) lim n (x + )m (x)dx


0
L L

n=1 m=1
N

RTff ( ) = nn ( x + )n ( x) even in the case n=m


n =1

9 K-L sample functions: in general not ergodic in autocorrelation

4/6/2009

IFMA Seminar, France

Solution of Fredholm integral equation


9 Exact solution feasible only for simple geometries (line, square, circle) and special
forms of C ff ( x1 , x2 )
9 Two main categories of numerical methods are available: integration formulaebased methods (e.g. quadrature method) and expansion methods (e.g. Galerkin)
9 Numerical solution with the conventional Galerkin approach
matrices very costly to compute and invert

Dense

9 An efficient numerical solution of the Fredholm integral equation is indispensable


especially when higher order eigenpairs are needed for an accurate representation
of the stochastic field

4/6/2009

IFMA Seminar, France

The wavelet-Galerkin method


9 Wavelet basis functions enhance the performance of the Galerkin method in the
solution of integral equations (Phoon et al. 2002)
9 If the kernel of the integral equation C ff ( x1 , x2 ) is a rapidly decreasing function, the
application of the wavelet-Galerkin method leads to sparse matrices (Beylkin,
Coifman & Rokhlin 1991)
Basic steps of the wavelet-Galerkin approach:
1. Selection of a set of M wavelet basis functions 1 ( x), 2 ( x),..., M ( x)
Usually Haar wavelets are used (the simplest form of
Daubechies wavelets)
1 0 x < 0.5
Haar mother wavelet function: ( x) = 1 0.5 x < 1
0 otherwise

4/6/2009

IFMA Seminar, France

10

2. Approximation of each eigenfunction of the covariance kernel by a linear combination of Haar wavelet basis functions:
M 1

n ( x) = di(n) i ( x) = ( x)D(n)
i =0

di(n): wavelet coefficients


M = 2m, m: wavelet level

Remark 1. Number of terms in the truncated K-L expansion: N M


3. Solution of the generalized eigenvalue problem: ( x)D = ( x) AHD
Remark 2. A crucial difference of this approach with the conventional Galerkin
method: the computation of matrices A and H does not require numerical
integration
Matrix A is the 2D wavelet transform of C ff ( x1 , x2 ) and H a diagonal MxM matrix
Remark 3. Matrix A is sparse by nature and can be made further sparse by
ignoring elements below a threshold value
This may lead to computational
instabilities in the numerical procedure

4/6/2009

IFMA Seminar, France

11

Numerical examples (Stefanou & Papadrakakis 2007)


1. First order Markov stochastic field with exponential covariance function
Case 1: C ff ( ) = R ff ( ) = 2e

/b

S ff ( ) =

2
b
- zero mean, unit variance
(1 + b2 2 ) - selected threshold values in step 3: 10-3, 10-7

2. Stochastic field with square exponential covariance function

Case 2: Cff ( ) = Rff ( ) = 2e

/ b2

S ff ( ) =

b2 2 - zero mean, unit variance


2b
exp

-12
2
4 - selected threshold value in step 3: 10

9 A smaller threshold value is selected for case 2 because the square exponential
kernel is sparser by its nature
9 Parameters: Wavelet level m=7 (M=27=128 eigenpairs can be computed at
most), number of K-L terms-points in the discretization of the wave number domain
N=16, number of terms in the FFT series M=128

4/6/2009

IFMA Seminar, France

12

Ensemble variance of
truncated K-L expansion:

0.95

Variance

Var[ f ( x)] = nn2 ( x)

0.9

n =1

0.85

- Wavelet level m = 7

thres.10^-3
thres.10^-7

- Threshold values in matrix A:


10-3, 10-7

0.8
0

32

64

96

128

x [-1,1]

Mean variance: 0.912


Good approximation for 5 K-L
terms

The variance is fluctuating w.r.t. x


9 The error at the boundaries is larger compared
to the middle region
9

Threshold value 10-3: substantial oscillations observed throughout the range [-1,1]
due to the numerical instabilities in the calculation of eigenpairs
9 Threshold value 10-7: a stable solution is obtained
9

4/6/2009

IFMA Seminar, France

1.2

1.4
=0.2

1.2

=0.4
=1.0

0.8

=0.2
=0.4

eigenvalue

eigenvalue

13

=2.0
0.6
0.4
0.2

=1.0

0.8

=2.0

0.6
0.4
0.2

0
0

16

24

32

index i

16

24

32

index i

Case 1: = 2b

Case 2: = b

Eigenvalue decay for =0.2, 0.4, 1.0 2.0 (: Vanmarckes scale of fluctuation)
The most rapid eigenvalue decay is observed in case 2 for =2.0 (strongly correlated
stochastic field with smooth autocovariance function)
9

Few K-L terms are needed for an accurate simulation of random fields of this kind

4/6/2009

IFMA Seminar, France

14

1.2

Ensemble variance

Ensemble variance

1.2

0.8
0.6
K-L

0.4

Spectral
0.2

0.8
0.6
K-L

0.4

Spectral
0.2
0

0
0

12

16

20

24

Case 1: =0.2

16

20

24

Case 2: =0.2
1.2

1.2

Ensemble variance

Ensemble variance

12

Number of terms N

Number of terms N

0.8
0.6
K-L

0.4

Spectral
0.2

0.8
0.6
K-L

0.4

Spectral
0.2

0
0

12

16

20

24

Number of terms N

Case 1: =2.0

12

16

20

24

Number of terms N

Convergence of each
approach to the target
variance

4/6/2009

Case 2: =2.0

IFMA Seminar, France

Samples generated with the K-L expansion


(N=16, =0.2)
Worst

Best

Average

15

Samples generated with the K-L expansion


(N=16, =0.2)

Target

Worst

Best

Average

Min. value

-3.3076

-2.0820

-1.7178

Min. value

-1.5411

-1.7148

-1.5988

Max. value

1.1902

1.7706

1.5981

Max. value

1.9598

1.8932

1.8455

Target

Mean

-0.3434

-0.1370

-0.0115

0.000

Mean

0.5846

0.0813

0.0872

0.000

Variance

0.8954

0.7032

0.6791

1.000

Variance

0.5237

0.8803

0.7592

1.000

Skewness

-1.1962

0.0153

-0.0761

0.000

Skewness

-0.5614

0.0166

0.0595

0.000

Samples generated with the spectral


representation (N=16, =0.2)
Worst

Best

Average

Samples generated with the spectral


representation (N=16, =0.2)
Target

Worst

Best

Average

Min. value

-1.3331

-1.9034

-1.9070

Min. value

-1.2252

-2.3388

-1.9791

Max. value

3.2233

1.8684

2.0226

Max. value

3.4019

2.1147

2.0328

Target

Mean

0.0000

0.0000

0.0000

0.000

Mean

0.0000

0.0000

0.0000

0.000

Variance

0.7848

0.7848

0.7848

1.000

Variance

0.9069

0.9069

0.9069

1.000

Skewness

1.6556

-0.0103

0.0711

0.000

Skewness

1.7822

-0.0004

0.0678

0.000

Case 1

Case 2

9 Spectral representation produces ergodic sample functions in a sample-by-sample


sense i.e. every sample function has the target mean and SDF (see below SDF plots)

4/6/2009

IFMA Seminar, France

16

1.2

1.2

Sample

Sample

Target

Target
0.8

0.8

CDF(f)

Ensemble CDF(f)

0.6

0.6

0.4

0.4

0.2

0.2
0

0
-3

-2

-1

-3

-2

-1

K-L expansion: ensemble average

Case 2
=0.2

1.2
Sample

Sample

Target

Target

0.8

0.8

CDF(f)

Ensemble CDF(f)

K-L expansion: Case 1, =0.2, m=9

1.2

0.6

0.6

0.4

0.4

0.2

0.2
0

0
-3

-2

-1

-3

-2

-1

Spectral representation: ensemble


average
4/6/2009

Spectral representation: Case 1,


=0.2, m=9

IFMA Seminar, France

0.03

Target
Sample

0.025

17

Ergodicity of spectral representation-based


samples with regard to SDF: a perfect
matching of the target SDF is observed
(cases 1-2, =0.2)

0.035

SDF

0.02
0.015
0.01
0.005
0
0

10

20

30

40

50

60

wave number

1
S ( ) =
2 T0
(i )
ff

T0

f ( i ) ( x) exp(i x )dx

Case 1
0.035
0.03

Target
Sample

SDF

0.025
0.02

Wavelet level, power in FFT series (m)

K-L expansion (1 sample)

2.5

60.0

Spectral representation (1 sample)

3x10-4

10-3

0.015
0.01
0.005
0
0

10

20

30

40

50

60

Computational performance of K-L


expansion and spectral representation

wave number

Case 2
4/6/2009

IFMA Seminar, France

18

Simulation of non-Gaussian stochastic fields


Theoretically all the joint multi-dimensional density functions are needed to fully
characterize a non-Gaussian stochastic field
Much of the existing research has focused on a more realistic way of defining a
non-Gaussian sample function e.g. as a simple transformation of some underlying Gaussian field with known second-order statistics:

f ( x) = F 1 [ g ( x)]

Translation field theory (Grigoriu 1984, 1998):


- Spectral distortion:

S ff ( ) S Tff ( )

- Compatibility of F - R ff ( ) : RTff ( ) = E [ f ( x) f ( x + ) ] = E h [ g ( x)] h [ g ( x + )]


T

F 1

RTff ( ) =

F 1

[ ( g1 )]F 1[( g 2 )] [ g1 ,g 2 ; Rgg ( )]dg1dg 2

4/6/2009

IFMA Seminar, France

19

where g1 = g ( x), g 2 = g ( x + )
and [ g 1 , g 2 ; R gg ( )] the joint density of {g 1 , g 2 }

Characteristics of translation fields


9 The relationship between the two autocorrelation functions can have a closed
form only in few cases
9 Strictly speaking, if the target F R Tff ( ) are proven to be incompatible, there is
no translation field with the prescribed characteristics.
9 The problem of incompatibility becomes even greater for highly skewed narrowbanded stochastic fields (Grigoriu 1998)
9 Analytical expressions of crossing rates are available for translation fields

4/6/2009

IFMA Seminar, France

20

- In order to address the problem of spectral distortion


Yamazaki-Shinozuka 1988, Deodatis-Micaletti 2001)

Iterative methods (e.g.

Repeated updating of S gg ( ) in every iteration:

S Tff ( ) ( j )
( j +1)
S gg ( ) = ( j )
S gg ( )
S ff ( )

Unwanted correlations between the terms of the spectral representation series


After the first iteration, the underlying Gaussian field is no more Gaussian and
homogeneous. Finally, the generated non-Gaussian sample functions will not have
the prescribed marginal PDF
Use of extended non-Gaussian to non-Gaussian mapping: f ( x) = F 1 F [ g ( x)]
T
- To address the issue of possible incompatibility between F and R ff ( ) : Spectral
preconditioning
Generated non-Gaussian field: approximately translation
field

4/6/2009

IFMA Seminar, France

21

- Correlations between the terms of the spectral representation series (proof):


N 1

( j +1)
g 0( j +1) ( x) = 2 S gg
( n ) cos( n x + n )

in iteration j+1

n =0

1
S ( ) =
2 T
( j)
ff

f
0

( j)
0

1
F 1[ g 0( j ) ( x)]exp(i x)dx
( x) exp(i x)dx =
2 T 0

1
N 1 ( j )

1
F
=

2 S gg ( n ) cos( n x + n ) exp(i x)dx

2 T 0
n=0

9 S ff ( ) depends on the random phase angles n


( j)

( j +1)
gg

S Tff ( ) ( j )
( ) = ( j )
S gg ( )
S ff ( )

9 S gg ( ) also depends on the random phase angles n


( j +1)
marginal PDF of g 0 ( x) due to c.l.t.
( j +1)

4/6/2009

IFMA Seminar, France

alteration of the

22

Proposed enhanced hybrid method (Lagaros, Stefanou & Papadrakakis 2005)


Basic idea: replace the updating scheme of the Gaussian spectrum S gg ( ) (source
of all important difficulties in the simulation) by a NN-based regression model
9 The unwanted correlations between the terms of the spectral representation
series become negligible
9 Use of the extended mapping:

f ( x) = F 1 F [ g ( x)]

9 Spectral preconditioning is not required


an algorithm covering a wider
range of non-Gaussian fields is obtained (not only translation fields).
9 A very small number (<50) of iterations is required until convergence
reduction of the computational effort required for simulation

4/6/2009

IFMA Seminar, France

23

E(w) =

1 N
[S ff ( j ) S Tff ( j )]2
2 j =1

0.6

0.8
Target PDF

0.6

Target PDF

Sam ple PDF

PDF(f)

Gaussian PDF(f)

drastic

0.4
0.2
0
-4

-2

Sam ple PDF

0.4

0.2

0
-4

Exact and sample Gaussian PDF at


first iteration
4/6/2009

-2

Exact and sample beta PDF at final


iteration (translation field case)

IFMA Seminar, France

24

Numerical example: a highly skewed narrow-banded stochastic field


Characteristics of target non-Gaussian field:
- Lognormal distribution defined in the range [-1.30, 10.0]
skewness = 2.763, kurtosis = 19.085
- Target correlation structure: S Tff ( ) =

1 2 3 2
b exp[ b ]
4

Sample function

10

Sample function

10

6
4
2
0
-2

= 1 b = 5

6
4
2
0
-2

64

128

192

256

64

128

x (m)

Sample function generated using the DM algorithm: [-1.23, 7.87]


4/6/2009

IFMA Seminar, France

25

0.8
Target PDF

Target PDF

0.6

0.6

Sample PDF

PDF(f)

PDF(f)

256

Sample function generated using the


proposed EHM: [-1.25, 8.89]

0.8

0.4
0.2

Sample PDF

0.4
0.2

0
-2

10

-2

10

Marginal PDF of sample function


generated using the proposed EHM
versus target lognormal PDF
0.8

Spectral density

0.8
Starget

0.6

Sng

0.4
0.2

Starget

0.6

Sng

0.4
0.2
0

0
0

wave number (rad/m)

wave number (rad/m)

SDF of sample function generated using


the D-M algorithm versus target SDF
4/6/2009

Marginal PDF of sample function


generated using the D-M algorithm
versus target lognormal PDF

Spectral density

192

x (m)

SDF of sample function generated using


the proposed EHM versus target SDF

IFMA Seminar, France

26

9 Perfect matching of the target PDF due to the (exact) extended mapping
9 Perfect matching of the target SDF achieved by the proposed (use of Rprop
training, Riedmiller & Brown 1993)

Computational performance of D-M and proposed


algorithm (EHM)
Method

Iterations

Time (sec)

D-M*

29279

146

EHM-SD

82

2.0

EHM-CG

25

0.6

EHM-Quickprop

45

1.2

EHM-Rprop

32

0.8

9 Substantially smaller cost of


proposed EHM (~ 2 orders of
magnitude)

9 Reduction of cost for the


stochastic analysis of realistic
structures with uncertain nonGaussian properties

* : without spectral preconditioning


4/6/2009

IFMA Seminar, France

27

Statistical comparison of D-M and EHM algorithms

D-M*
Mean value: 38242
St. dev.: 28701
Max: 94494
Min: 4414

D-M* (Iterations)

EHM-Rprop (Iterations)

29279

32

46843

38

4414

28

24335

29

5935

31

94494

35

53295

33

10628

29

61618

28

51582

29

EHM-Rprop
Mean value: 31.2
St. dev.: 3.5
Max: 38
Min: 28

* : without spectral preconditioning

4/6/2009

IFMA Seminar, France

28

Identification of random shapes from images based on polynomial chaos


expansion (Stefanou, Nouy & Clement 2009)

Material layout of a quadratic microstructure

9 Physical problems with random geometry:


disordered systems and random media,
fluctuating domain boundaries, shells with cutouts, heterogeneous materials with random
distribution of inclusions
9 Usually a limited number of samples of the
geometry is available in practice: its complete
probabilistic characterization (in terms of joint
densities) is infeasible
9 Shape recovery from simple images allows
obtaining many samples at a low cost + it is
relatively precise

Cylindrical shell with rectangular cut-out

4/6/2009

IFMA Seminar, France

29

Short description of the procedure


- Shape recovery with the level-set technique: construction of a collection of discretized level-set functions corresponding to a collection of images
- Reduction of information through empirical Karhunen-Love expansion:

(k )

+ U i X i( k )
i =1

- Probabilistic identification of random vector X = ( X 1 ,..., X m ) from samples X ( k )


decomposition on a polynomial chaos basis of degree p in dimension m
(Wiener 1938):

m , p

X H ( ( ))

- Identification of chaos coefficients:


1. Without independence hypothesis: maximum likelihood estimation for random
vector X
2. With independence hypothesis: maximum likelihood estimation for each random
variable or projection method based on empirical CDF
4/6/2009

IFMA Seminar, France

30

The stochastic finite element method


Solution of stochastic elliptic boundary value problems
[ (x, )u (x, ) ] = f (x, ), x D
u (x, ) = 0,
x D

(, , P )
u : D

Well posed problem (existence and uniqueness of solution):


1
9 C ( D)
9 0 < min < max Pr ( x, ) [ min , max ], x D = 1

A(u , v) = B (v ) v X = H 01 ( D )

Weak form:

deterministic

A(u , v ) = u vdx, B (v ) = fvdx u , v X


D

A(u , v ) = B (v )

v X L ()
2
P

stochastic

Finite element approximation


selection of suitable function spaces
h
e.g. X = span{1 , 2 ,..., N x } X and W h = span{ 1 ( ), 2 ( ),..., N ( )} W

= (1 ,..., M )

4/6/2009

IFMA Seminar, France

31

3D elasticity problem:
Formulation of the stochastic stiffness matrix
D( x, y, z ) = D0 [1 + f ( x, y, z )]

- Case of material randomness:

k ( e) =

( e )T

D0( e) B( e) dV ( e) +

e
V( )

( e )T

D0( e) B( e) f ( e) ( x, y, z ) dV ( e)

e
V( )

k (e ) = k 0(e ) + k (e )

- Discretization of the stochastic fields


Midpoint method
(e)
(e)
(e)
Local average method: k = k0 (1 + a )
NW

Weighted integral method: k ( e ) = k0( e ) + X l( e ) kl( e )


l =1

4/6/2009

IFMA Seminar, France

32

Response variability calculation


Direct Monte Carlo Simulation (e.g. Rubinstein 1981)

E (u i ) =

1
NSIM

NSIM

ui ( j)

2 (u i ) =

j =1

NSIM 2

1
2
u i ( j ) NSIM E (u i )
NSIM 1 j =1

9 Solution of SIM deterministic problems


substantial computational cost
especially for high dimensional problems and for large number of simulations
9 It must be combined with efficient discretization methods (e.g. local average
method)
9 Particularly suitable for parallel computing environment (embarrassingly
parallel)
9 Assisted by the spectacular growth of computing power, the only available
method for the solution of large-scale realistic problems

4/6/2009

IFMA Seminar, France

33

The perturbation method (Kleiber & Hien 1992)


9 Taylor series expansion of the stochastic finite element matrix and of the
resulting response vector:
N

K = K 0 + KiI ai +
i =1

u = u0 + uiI ai +
i =1

1 N N II
Kij ai a j + ...
2 i =1 j =1

KiI =

2K
K
, K ijII =
ai a j a =0
ai a=0

1 N N II
uij ai a j + ...
2 i =1 j =1

9 Calculation of the stochastic displacement vector:

u0 = K 01 P0
uiI = K 01 ( Pi I K iIu0 )
uijII = K 01 ( PijII K iIu Ij K Ij uiI K ijIIu0 )

9 Satisfactory results only for small coefficients of variation of the uncertain input
parameters
9 Improvement in accuracy obtained using higher order approximations: small
compared to the disproportional increase of computational effort

4/6/2009

IFMA Seminar, France

34

The spectral stochastic finite element method (Ghanem & Spanos 1991)
9 Karhunen-Love expansion of the stiffness matrix + polynomial chaos expansion
(PCE) of the displacement vector:
M

k ( e ) ( ) = k (0e ) + k l( e )i ( ),
l =1

k i( e ) = i i ( x)B T D 0 Bd e
e

P 1

U( ) = U j j ( ),
j =0

P +1 =

( M + p)!
M ! p!

9 Application (mainly) to linear problems with small variability + Gaussian


assumption for the uncertain input parameters (Sudret & der Kiureghian 2002)
9 Prohibitive computational cost for the solution of problems with large stochastic
dimension (increase of the order of PCE)
K 00 K 0, P 1 U 0 F0
K


10 K 1, P 1 U 1 = F1


K P 1,0 K P 1, P 1 U P 1 FP 1
NP NP
4/6/2009

IFMA Seminar, France

35

- Spectral SFEM: recent advances and extensions


9 Cases of non-smooth solutions (non-linearities, discontinuities): use of other basis
functions e.g. wavelets (Le Maitre et al. 2004) or adaptive sparse generalized PCE
(Wan & Karniadakis 2005, Blatman & Sudret 2008).
9 Stochastic reduced basis methods (Nair et al. 2002-2008): limited to random linear
systems
9 Non-intrusive, stochastic response surface approaches (Baroth et al. 2006, Berveiller
et al. 2006): can take advantage of powerful deterministic FE codes
9 Spectral SFEM in a multi-scale setting (Xu 2007): stochastic variational approach +
scale-bridging multi-scale shape functions
9 X-SFEM (Nouy et al. 2007-8): implicit representation of complex geometries using
random level-set functions

4/6/2009

IFMA Seminar, France

36

Stochastic finite element analysis of shells


The multi-layered triangular shell element TRIC (Argyris et al. 1997)
9 TRIC is a triangular, shear-deformable facet shell element suitable for the analysis
of thin and moderately thick isotropic as well as composite plate and shell structures
9 The element formulation is based on the natural mode finite element method
(Argyris et al. 1979)
9 The treatment of the element kinematics eliminates automatically shear locking
phenomena (Argyris et al. 2000)
9 No need to perform numerical integration for the computation of the deterministic
stiffness matrix which is carried out in closed form
derivation of a cost-effective
stochastic stiffness matrix
9 Computational efficiency in large-scale stochastic finite element computations

4/6/2009

IFMA Seminar, France

37

9 The natural stiffness of the


TRIC element is based only on
the 12 independent natural
straining modes:

k qc
( 66 )
k =
(1212 )

qh
( 33)

k az
( 33)

kqc : axial and symmetric bending stiffness terms


kqh : anti-symmetric bending and shear stiffness terms
kaz : stiffness terms due to in-plane rotations (azimuth stiffness terms)
4/6/2009

IFMA Seminar, France

38

Derivation of the stochastic stiffness matrix


- Random variation of Young modulus (Argyris, Papadrakakis & Stefanou 2002)

E ( x, y ) = E 0 [1 + f ( x, y )]
k

1.

qc
( 66 )

= BtcT ct Btc dV
V

2a.
2.

{[ ] + [k ] }

b
k qh = k qh

(3 3 )

s
qh

= (1 + a )( k qc ) 0

a=

1
f ( x , y )d

qc
( 66 )

9 It is proved that the axial and


symmetric bending stiffness
terms have a local average form

b
b
kqh
= ( kqh
) + X i k i

b
k qh
= BthT ct Bth dV

1 1

2b.

i =1

9 6 weighted integrals i

k qhs = BshT X s Bsh dV

k qhs = (1 + a )(k qhs ) 0

9 Local average form


4/6/2009

IFMA Seminar, France

39

e.g. weighted integrals 1, 4 : X 1 = 2a f a , , d

a , ,

X 4 = a f a , , d

the natural (triangular) coordinates

Matrix k1 of the fluctuating part of the anti-symmetric bending stiffness terms:

k1 =

symm.

h/ 2

0
6 2
9
z k aa + 2 z k a + 2 z 2 k
4
la
la
l

l 2

0
l l 2
3l 2
3l 2
9 2
z k
4 z k aa 2 z k a 2 z k a
l l
la
la l
l a l

l2 2
6 2
9 2

z
k
z
k
z
k
+
+
aa
a

l a4
l a2
l2

1 N
z kij = z ij dz = ijk ( zk3 zk3+1 ) i , j = , ,
3 k =1
h / 2
2

4/6/2009

- ij the entries of the constitutive


matrix ct
- k = 1,2,,N the number of
layers

IFMA Seminar, France

40

kzaa kza kza


1 0.5 0.5


kaz =
kz kz = kz 0.5 1 0.5
(33)
symm.
0.5 0.5 1
kz

3.

kz the maximum of the three edge bending stiffness values:


1
k z = max 2
l a

h/2

h / 2

z 2 dz ,

1
l 2

h/2

h / 2

z 2 dz ,

k z(i ) =

1
l2

h/2

h / 2

z 2 dz

h/2 2
z ii dz
l i2 h / 2

k z = max (1 + a ) k z( a ) , (1 + a ) k z( ) , (1 + a ) k z( )

a=

1
f ( x , y )d

9 The azimuth stiffness terms have also an inherent local


average form

4/6/2009

IFMA Seminar, France

41

- Combined random variation of Young modulus and Poisson ratio (Stefanou &
Papadrakakis 2004)
9 The entries of the elasticity matrix are nonlinear functions of Poisson ratio
consideration of random variation of Lam constants and
0
1
E
12 =
1 0
1 2
0 0 1

( x, y ) = 0 [1 + f ( x, y )] +
9
9
9
9

12

+ 2

=
+ 2
0
0

0
0
2

(x, y ) = 0 [1 + f ( x, y )]

E
1 2

=G=

E
2(1 + )

Axial and symmetric bending stiffness terms: local average form


Anti-symmetric bending stiffness terms: 12 weighted integrals
Anti-symmetric shear stiffness terms: local average form
Azimuth stiffness terms: local average form

4/6/2009

IFMA Seminar, France

42

- Random variation of the thickness


9 The thickness h appears into the integrals expressing the stiffness matrix of the
shell element
9 The application of the weighted integral method is not straightforward
The local average approach is applied leading to greater computational efficiency
h( x, y ) = h0 (1 + a h )

4/6/2009

ah =

1
f h ( x , y ) d

IFMA Seminar, France

L
z

Numerical examples
(Stefanou & Papadrakakis 2004)

P
h

rigid diaphragm

y
x

43

C
R

rigid diaphragm

R = 300 mm
L = 600 mm
h = 3.0 mm
E = 3000 N/mm2
= 0.3

1. Pinched cylinder

= 0.3

2. Scordelis-Lo shell
4/6/2009

IFMA Seminar, France

44

R1
R1 = 4800 mm
R2 = 8000 mm
L = 20000 mm

h = 40 mm

E = 21000 N/mm2
= 0.25
y

R2
3. Hyperboloid shell

4/6/2009

IFMA Seminar, France

45

9 Use of 2D homogeneous Gaussian stochastic fields for the representation of the


uncertain material and geometric properties
9 Sample functions of the stochastic fields generated by the spectral representation
method
9 Selection of two different correlation structures:

S ff ( 1 , 2 ) =

2f
b1b2 (b12 12 + b22 22 ) exp[ (b12 12 + b22 22 )]

S ff ( 1 , 2 ) =

2
f

b1 1 2 b 2 2 2
b1 b 2
exp


4
2
2

9 Calculation of COV of displacement at the characteristic node C:


COV (u i ) =

(u i )
E (u i )

9 Investigation of the effect of various parameters of the stochastic fields


(f, Sff, b1, b2) on the response variability
4/6/2009

IFMA Seminar, France

46

4/6/2009

IFMA Seminar, France

COV of w at node C

47

Scordelis-Lo shell (f = 0.1):


Random variation of Young modulus
and thickness (local average method)

0.25
E
h
E,h

0.2

f = 0.2 b1 = b2 = 2.6

0.15

9 Important effect of thickness


variation

0.1
0.05
0
0.065

0.65

1.3

2.6

26

65

130

Correlation length parameter b

Hyperboloid shell (f = 0.1):


Random variation of Young modulus
and thickness (local average method)

0,3

COV of w at node C

0,25
E
0,2

h
E,h

0,15

9 Important effect of thickness


variation

0,1
0,05
0
0,2

80

200

Correlation length parameter b1

4/6/2009

IFMA Seminar, France

48

COV of w at node C

0.15

Scordelis-Lo shell (f = 0.1):


Random variation of Young modulus
and Poisson ratio

0.1

0.05

0
0.065

0.65

1.3

2.6

26

65

130

9 Negligible effect of Poisson ratio


variation

Correlation length parameter b


E

E,v (Local average)

E,v (Weighted integral)

Hyperboloid shell (f = 0.1):


Random variation of Young modulus
and Poisson ratio

COV of w at node C

0,1

0,05

9 Negligible effect of Poisson ratio


variation
0
0,2

80

200

Correlation length parameter b1


E

E (Weighted integral)

E,v (Local average)

E,v (Weighted integral)

4/6/2009

IFMA Seminar, France

49

0,3

COV of w at node C

0,25
0,2
0,15
h
h(P.s.2)

0,1
0,05
0
0,2

80

200

Correlation length parameter b1

Investigation of the effect of


correlation structure (power spectrum)

0,35
0,3

COV of w at node C

Hyperboloid shell (f = 0.1):


Random variation of Young modulus
and thickness (local average method)

E,h

0,25

E,h(P.s.2)

0,2

Var ( wC ) S ff ( )

0,15
0,1
0,05
0
0,2

80

200

Correlation length parameter b1

4/6/2009

IFMA Seminar, France

50

Conclusions
9 The stochastic stiffness matrix of the TRIC shell element is formulated in terms of
a minimum number of random variables
9 Good agreement between weighted integral and local average methods
9 Significant effect of correlation length parameter b on the response variability
9 Slight effect of correlation structure (expressed via the form of the power
spectrum) on the response variability
9 Random variation in the shell thickness has significant effect on the response
variability compared to the effect of random Young modulus
9 Small effect of random variation of Poisson ratio on the results

4/6/2009

IFMA Seminar, France

51

Stability analysis of shells with random imperfections


9 Large scatter in the measured buckling loads
9 Big discrepancy between the deterministic predictions of buckling loads and the
corresponding experimental results

9 Presence of initial imperfections which occur during the manufacturing and


construction stages
9 Realistic description of imperfections in the framework of a robust stochastic finite
element formulation
9 Importance of available data banks for the realistic simulation of imperfections
(estimation of probability distribution, correlation structure), e.g. Elishakoff & Arbocz
(1982, 1985)

4/6/2009

IFMA Seminar, France

52

Methods based on random variables


9 Use of 2D Fourier series with random Fourier coefficients (e.g. Chryssanthopoulos
& Poggi 1995, Noirfalise et al. 2007) for the representation of geometric imperfections
9 This idea is related to the fact that an analytical buckling analysis of cylindrical
shells yields a 2D Fourier series representation of the critical modes

Methods based on stochastic fields


9 Simulation of the imperfections using stochastic fields combined with the finite
element method for the solution of the stability problem
9 Use of direct Monte Carlo simulation for the computation of buckling load
variability

4/6/2009

IFMA Seminar, France

53

Stability analysis of cylindrical shells with random geometric imperfections


(Schenk & Schueller 2003)
9 Use of the data bank by Arbocz & Abramovich (1979) for the estimation of the
probability distribution and correlation structure of the imperfections
9 Selection of 2D non-homogeneous Gaussian stochastic fields for the description of
the imperfections
9 250 direct Monte Carlo simulations for the computation of buckling load variability
(finite element model with geometric non-linearity)

9 The experimentally determined


scatter in the limit load can be
predicted numerically
9 A more accurate prediction
requires the incorporation of other
kinds of imperfections e.g. material,
thickness, boundary conditions

4/6/2009

IFMA Seminar, France

54

Stability analysis of cylindrical shells with random geometric, material,


thickness and boundary imperfections (Papadopoulos, Stefanou & Papadrakakis
2009)
9 Use of TRIC shell element with geometric and material non-linearity
9 Selection of 2D non-homogeneous Gaussian stochastic fields for the description
of geometric imperfections (data bank by Arbocz & Abramovich, 1979)
9 Effect of non-Gaussian assumption for E, t on buckling load variability
X

Randomly varying quantities:

r ( x , y ) = R + a 0 ( x , y ) + g1 ( x , y )

Loading P

E ( x, y ) = E0 [1 + f1 ( x, y )]
R

t ( x, y ) = t0 [1 + f 2 ( x, y )]
P ( x ) = P0 [1 + g 2 ( x ) ]

E=104410N/mm2
=0.3
L=202.3mm
R=101.6mm
t=0.11597mm

f1, f2: homogeneous non-Gaussian


translation fields

4/6/2009

Initial geometric imperfections:

IFMA Seminar, France

55

r ( x , y ) = R + a 0 ( x , y ) + g1 ( x , y )

Measured initial unfolded shape of shell A7

Shell

R(mm)

t (mm)

L (mm)

E (N/mm2)

P (N)

A-7

101.6

0.1140

203.20

104110

3036.4

A-8

101.6

0.1179

203.20

104800

3673.8

A-9

101.6

0.1153

203.20

101350

3724.8

A-10

101.6

0.1204

203.20

102730

3196.9

A-12

101.6

0.1204

209.55

104800

3853.0

A-13

101.6

0.1128

196.85

104110

3108.8

A-14

101.6

0.1110

196.85

108940

3442.9

Geometry, material properties and experimental


buckling loads of A-shells (Arbocz & Abramovich,
1979)

a0(x,y)

9 The ensemble average a0(x,y) as well as the


evolutionary power spectrum of stochastic field
g1(x,y) are derived from a statistical analysis of
experimentally measured imperfections of A-shells

4/6/2009

IFMA Seminar, France

56

Material and thickness imperfections: use of non-Gaussian translation fields


Lower bound
Lognormal

-1

Upper bound
+

Shape parameters
-

Beta - Case 1

-0.5

0.5

p=12

q=12

U-beta - Case 2

-0.16

0.16

p=0.8

q=0.8

L-beta - Case 3

-0.13

0.26

p=0.8

q=1.6

E ( x, y ) = E0 [1 + f1 ( x, y )]

t ( x, y ) = t0 [1 + f 2 ( x, y )]
S gg (1 , 2 ) = 2f

b 2 b 2
b1b2
exp 1 1 2 2
4
2 2

Correlation structure of the imperfections

In-plane edge (boundary) imperfections: non-uniform axial loading


9 It is assumed that boundary imperfections are produced by a non-uniform random
axial load distribution on the upper edge of the cylinder, modeled as a 1D homogeneous Gaussian stochastic field
loading surface

P ( x ) = P0 [1 + g 2 ( x ) ]
S gg ( ) =

cylinder

4/6/2009

g2

bg exp (bg2 2 )
4
4

IFMA Seminar, France

Numerical results

57

NSIM=100

9 bg = 100 mm with g = 5% are selected for the description of g2


9 b1 = b2 = 50 mm with f = 10% are selected for the description of Young modulus
and thickness
these values are responsible for the minimum mean(Pu) and it is
likely to lead to worst case scenarios w.r.t. lowest Pu

9 The choice of PDF has a significant effect on first and second order properties of Pu
distribution
9 Max Cov(=0.16) for a Gaussian PDF, Min Cov(=0.06) for an L-beta PDF

4/6/2009

IFMA Seminar, France

58

Pu(perfect) = 5350N

Lognormal E, t

Gaussian E, t

U-beta E, t

Beta E, t

4/6/2009

IFMA Seminar, France

59

Conclusions
9 The choice of marginal PDF of E, t is crucial: it
affects significantly the shape as well as the
extreme values of Pu distribution
9 A large magnification of uncertainty has been
observed in the Gaussian case: Cov(Pu) ~ 1.6f
L-beta E, t

9 The lognormal and beta non-Gaussian


assumptions led to estimates of the scatter of Pu
closer to the experimental measurements
9 The tri-modal shape of buckling loads observed
in the experiments has been reproduced by the
corresponding numerical simulations
9 The lowest Pu has been found to represent only
the 28-60% of the Pu of the perfect shell

Experimental

4/6/2009

9 Towards a robust design of imperfect shell


structures

IFMA Seminar, France

60

Nonlinear stochastic dynamic analysis of frames


(Stefanou & Fragiadakis 2009)

A 3-storey steel moment-resisting


frame designed for a Los Angeles
site (SAC/FEMA program)
Fundamental mode period: T1=1.02
sec
5 integration sections defined in
every beam-column element
Geometrical nonlinearities not
considered in the analysis
Material law: bilinear with pure kinematic hardening
Loading: 3 sets of 5 strong ground motion records corresponding to 3 levels of
increasing hazard: low, medium and high + spectrum-compatible artificial
accelerograms

4/6/2009

IFMA Seminar, France

61

The ground motion records used


ID
( level)

Earthquake

Station

Mw2

R3

Soil4

PGA

1 (1/1)

Imperial Valley, 1979

Plaster City

045

6.5

31.7

C,D

0.042

2 (2/1)

Imperial Valley, 1979

Plaster City

135

6.5

31.7

C,D

0.057

3 (3/1)

Imperial Valley, 1979

Westmoreland Fire Station

180

6.5

15.1

C,D

0.11

4 (4/1)

Imperial Valley, 1979

Westmoreland Fire Station

090

6.5

15.1

C,D

0.074

5 (5/1)

Imperial Valley, 1979

Compuertas

285

6.5

32.6

C,D

0.147

6 (1/2)

Northridge, 1994

LA, Baldwin Hills

090

6.7

31.3

B,B

0.239

7 (2/2)

Imperial Valley, 1979

Plaster City

090

6.5

31.7

C,D

0.057

8 (3/2)

Loma Prieta, 1989

Sunnyvale Colton Ave

270

6.9

28.8

C,D

0.207

9 (4/2)

Superstition Hills, 1987

Wildlife Liquefaction Array

090

6.7

24.4

C,D

0.18

10 (5/2)

Loma Prieta, 1989

Sunnyvale Colton Ave

360

6.9

28.8

C,D

0.209

11 (1/3)

Superstition Hills, 1987

Wildlife Liquefaction Array

360

6.7

24.4

C,D

0.2

12 (2/3)

Northridge, 1994

LA, Hollywood Storage FF

360

6.7

25.5

C,D

0.358

13 (3/3)

Loma Prieta, 1989

Hollister South & Pine

000

6.9

28.8

,D

0.371

14 (4/3)

Loma Prieta, 1989

WAHO

000

6.9

16.9

,D

0.370

15 (5/3)

Loma Prieta, 1989

WAHO

090

6.9

16.9

,D

0.638

1
4

4/6/2009

Component 2 Moment magnitude


USGS, Geomatrix soil class

Closest distance to fault rupture

IFMA Seminar, France

62

Upper bound
+
0.26

Shape parameters
p=0.8
q=1.6

1D stochastic variation of Young


modulus and yield stress described
by zero-mean lognormal and L-beta
translation fields with f =10%
The variability of the maximum
inter-storey drift max is examined
using 1000 Monte Carlo simulations:

0.3
0.2

Sample function

Lower bound
-1
-0.13

Lognormal
L-beta

b=0.2
b=1

0.1

b=2
0

b=10
b=20

-0.1

b=100
-0.2
-0.3

( max )
COV ( max ) =
E ( max )

-1

-0.5

0.5

normalised length (m)

A set of sample functions of lognormal stochastic


fields characterizing the spatial variation of Young
modulus E in a beam

Investigation of the sensitivity of


max w.r.t. the correlation length
parameter b

4/6/2009

IFMA Seminar, France

63

Mean value, COV of max for different values of


correlation length parameter b and the ground motion
records of slide 61 (lognormal distribution of E, y)
0.025

mean of

max

0.02
b=0.2
b=1.0
b=2.0
b=10
b=20
b=100

0.015

0.01

9 The effect of b on COV is


important in many cases

0.005

1/1

2/1

3/1

4/1)

5/1

1/2

2/2

3/2

4/2

5/2

1/3

2/3

3/3

4/3

5/3

record
0.2

)=/

0.18
0.16
0.14
b=0.2
b=1.0
b=2.0
b=10
b=20
b=100

0.12

max

COV(

9 The different characteristics of the


seismic records are transferred to the
response statistics
significant
record-to-record variability

0.1
0.08
0.06

9 A large magnification of
uncertainty is observed in some
cases
max COV=18% (~1.8f)
for record 4/2
9 The mean value is practically not
affected by b

0.04
0.02
0

1/1

2/1

3/1

4/1)

5/1

1/2

2/2

3/2

4/2

5/2

1/3

2/3

3/3

4/3

5/3

record

4/6/2009

IFMA Seminar, France

64

Mean value, COV of max for different values of


correlation length parameter b and artificial
accelerograms compatible with records 1/1, 4/2, 1/3
(lognormal distribution of E, y)

mean of max

0.015

0.01
b=0.2
b=1.0
b=2.0
b=10
b=20
b=100

9 The effect of b on COV is


important in some cases

0.005

1/1

4/2

1/3

record
0.25

0.2

COV( max ) = /

9 The different characteristics of the


seismic records are transferred to the
response statistics
significant
record-to-record variability

b=0.2
b=1.0
b=2.0
b=10
b=20
b=100

0.15

0.1

9 The magnification of uncertainty is


even more pronounced in this case:
max COV=23% (~2.3f) for record
1/1
9 The mean value is practically not
affected by b

0.05

1/1

4/2

1/3

record

4/6/2009

IFMA Seminar, France

65

Skewness of max for different values of correlation


length parameter b and the ground motion
records of slide 61 (lognormal, L-beta distribution
of E, y)
2

Input skewness values:


- Lognormal distribution: 0.3
- L-beta distribution: 0.59

1.5
1

b=0.2
b=1.0
b=2.0
b=10
b=20
b=100

0
0.5
1
1.5
2
2.5
3
3.5

1/1

2/1

3/1

4/1)

5/1

1/2

2/2

3/2

4/2

5/2

1/3

2/3

3/3

4/3

5/3

record
2
1.5
1
0.5

Skewness

Skewness

0.5

0
b=0.2
b=2.0
b=100

0.5
1

9 Important record-to-record
variability
9 Significant influence of b in many
cases
9 Values of skewness substantially
different from those of the system
properties due to the strong nonlinearity of the problem

1.5
2
2.5
3
3.5

1/1

5/1

1/2

2/2

2/3

5/3

record

4/6/2009

IFMA Seminar, France

66

Conclusions
9 A stochastic response history analysis of a steel frame having uncertain non-Gaussian
material properties and subjected to seismic loading has been performed.
9 The Young modulus and the yield stress were described by uncorrelated homogeneous
non-Gaussian translation fields.
9 The effect of the probability distribution of the input parameters on the response
variability was negligible due to the small input COV(=10%).
9 The significant influence of the scale of correlation of the stochastic fields and of the
different seismic records on the response variability has been revealed.
9 A large magnification of uncertainty has been observed in some cases.
9 Importance of a realistic uncertainty quantification and propagation in nonlinear
dynamic analysis of engineering systems.

4/6/2009

IFMA Seminar, France

67

Thank you for your attention!

4/6/2009

IFMA Seminar, France

68

You might also like