NASA Contractor Report 191098

/_ 61

Probabilistic Finite Elements for

Fatigue and Fracture Analysis

Ted Belytschko and Wing Kam Liu

Northwestern University
Evanston, Illinois

April 1993

Prepared for
Lewis Research Center
Under Contract NAG3-822

(NASA-CR-191098) PRQBABILISTI£ N93-23053

FINITE ELEMENTS FOR FATIGUE ANO
FG"ACTU!_E AhALYSI5 Final Report
(_iorthw,'-stern Univ.) ol p Unc I ] s
National Aeronautics and
Space Administration

G3/39 0156315
 PROBABILISTIC FINITE ELEMENTS FOR FATIGUE AND FRACTURE ANALYSIS

TedBelytschko and Wing Kam Liu

Northwestern University
Department of Mechanical Engineering
Evanston, Illinois 60208

SUMMARY

This report presents an overview of the probabilistic finite

element method (PFEM) developed by the authors and their

colleagues in recent years. The primary focus is placed on the

development of PFEM for both structural mechanics problems and

fracture mechanics problems. The perturbation techniques are

used as major tools for the analytical derivation. The report

specifically covers the following: (i) representation and

discretization of random fields; (2) development of PFEM for the

general linear transient problem and nonlinear elasticity using

Hu-Washizu variational principle; (3) computational aspects; (4)

discussions of the application of PFEM to the reliability

analysis of both brittle fracture and fatigue; and (5) a novel

stochastic computational tool based on stochastic boundary

element (SBEM). Results are obtained for the reliability index

and corresponding probability of failure for: (i) fatigue crack

growth; (2) defect geometry; (3) fatigue parameters and; (4)

applied loads. These results show that initial defect is a

critical parameter.

l
 1. INTRODUCTION

It is becoming increasingly evident that traditional and especially deterministic methods will
not be sufficient to properly design advanced structures or structural components.subjected to a
variety of complex, and cyclic loading conditions. Due to uncertainty in loading conditions,
material behavior, geometric configuration, and supports, the stochastic computational mechanics,
which accounts for all these uncertain aspects, has to be applied to provide rational reliability
analysis and to describe the behavior of the structure. The fundamentals of stochastic
computational mechanics and its application to the analysis of uncertain structural systems are
summarized and recapitulated in the book (Liu and Belytschko, 1989).

While the theory of statistics and structural reliability has been used successfully in
modeling the uncertain nature of structures, load environments, and in computing the probability of
failure, its application is only limited in simple structures with linear constitutive behavior. Due to
the complexity in the geometry, external loads, and nonlinear material behavior, more advanced
computational tools such as Finite Element Methods (FEMs) or Boundary Integral Equation
Methods (BIEMs), have to be employed to provide the necessary computational framework for
analyzing structural response. The combination of these advanced computational tools with the
theory of statistics and structural reliability has become a rational way for the safety assessment and
uncertainty characterization of complex structures. It_ this Chapter, attenuation is focused on the
development of Probabilistic Finite Element Method (PFEM), which combines the finite element
method with statistics and reliability methods, and its application to linear, nonlinear structural
mechanics problems and fracture mechanics problems. The novel computational tool based on the
Stochastic Bound agy Element Method is also given for the reliability analysis of a curvilinear
fatigue crack growth.

The existing PFEMs have been applied to solve for two types of problems: (1)
determination of the response uncertainty in terms of the means, variance and correlation
coefficients; (2) determination the probability of failure associated with prescribed limit states.
While the second order statistic moments of a response are not sufficient for a complete reliability
analysis, these moments offer useful statistical information and serve as a measures of reliability.
Furthermore, due to the lack of multivariate distribution function of random variables, a
meaningful risk assessment and failure analysis may not be feasible.

The perturbation method has been used extensively in developing PFEM due to its
simplicity and versatility. Cambou (1975) appears to have been the first to apply the fh-st order
perturbation method for the finite element solution of linear static problems with loading and
system stochasticity. Baecher and Ingra (1981) also used the same techniques for settlements
predictions. The perturbation method in conjunction with finite element method has also been
adopted by Handa and Anderson (1981) for static problems of beam and frame structures, by Ishii
and Suzuki (1987) for slope stability reliability analysis, and by Righetti and Harrop-Williams
(1988) for static stress analysis for soils. The accuracy, convergence and computational efficiency
of the perturbation method have been compared with those from Neumann expansion method and
direct Monte Carlo Simulation (MCS) method (Shinozuka and Yamazaki, 1988; Shinozuka and
Deodatis, 1988). The PFEM based on the second-order perturbation approximation has been
introduced by Hisada and Nakagiri (1981 and 1985) for static problems and for eigenvalue
problems.

Extensive research on the PFEM has been developed by the authors and their colleagues at
Northwestern University in the recent years. The PFEM based on the second--order perturbation has
been developed to estimate the statistic moments of the response for linear static problems (Liu et al.,

2
 1986a),nonlineardynamicproblems(Liu et al., 1986b),andinelasticproblems(Liu et al., 1987).
The.formulation basedon the single-field variational principle has beenextendedby Liu et al.
(1988a)to the three-field Hu-Washizu variational principle formulation, which has far greater
versatility. The numericalinstability resultingfrom the seculartermsin the perturbationhasbeen
removedby Liu et al. (1988b)basedon Fourrier analysis. The perturbationmethodshavebeen
shown to provide efficient and accurateresults for small random fluctuations in the random
parameters.An extensivereview on the applicationof perturbationmethodsin developingPFEM
hasbeengivenby BenaroyaandRehak(1988).

The finite elementmethodcoupledwith the First andSecond-OrderReliability Methods

(FORM and SORM) has been developedby Der Kiureghian and Ke (1985, 1988) for linear
structuralproblemsandLiu andDer Kiureghian(1991)for geometricallynon-linearproblems. The
mostcritical stepin this methodis thedevelopmentof anefficient searchalgorithmfor locatingthe
point at which theresponsesurfaceis to beexpandedin a first or secondorderTaylor series. This
point is obtainedby aniterativeoptimizationalgorithm,which involvesrepeatedcomputationof the
limit statefunction andresponsederivatives. Unlike the methodof direct differentiation (Der
Kiureghianand Ke, 1988;Liu and Der Kiureghian, 1991; Zhang and Der Kiureghian, 1991), the
PFEM based on the perturbation approximation in conjunction with FORM has been developed by
Besterfield et al (1990, 1991) for the reliability analysis of brittle fracture and fatigue. In a slightly
different context, the PFEM has been developed by Faravelli (1986, 1989) that couples response
surface approach with deterministic finite element formulation. The finite element simulation
coupled with the polynomial response surface fitting.has been also proposed by Grigoriu (I982).
Using a deterministic finite element code and finite differences, an advanced algorithm based on the
Fast Probability Integration (FPI) has been developed by Wu et al (1990) to generate the entire part
of the Cumulative Distribution Function (CDF) of the response. The performance of the FPI based
on either advanced mean-value method or advanced mean value first-order method has been
demonstrated by Cruse et al. (1988) through the reliability analysis of turbine blade.

In addition to the PFEM, the Stochastic Boundary Element Method (SBEM) has been
developed and adopted recently by researchers. The SBEM that combines the deterministic
boundary element method with perturbation expansions has been developed by Ettouney et al.
(1989) and Dasgupta (1992) for the determination of the statistic moments of both displacements and
tractions. Most recently, the authors have developed the SBEM, which combines the mixed
boundary integral equation method CLua et al., 1992c) with FORM, for the study of probabilistic
fatigue crack growth (Lua et al., I992d).

This Chapter presents an overview of the PFEM developed by the authors and their
colleagues in the recent years. The primary focus is placed on the development of PFEM for both
structural mechanics problems and fracture mechanics problems_ The perturbation techniques are
used as major tools for the analytical derivation. The remainder of this Chapter is organized as
follows. In Section 2, the representation and discretization of random fields are presented. The
development of PFEM for the general linear transient problem and nonlinear elasticity using Hu-
Washizu variational principle are given in Section 3, and 4, respectively. The computational aspects
are discussed in Section 5. The application of PFEM to the reliability analysis of both brittle fracture
and fatigue is given in Section 6. A novel stochastic computational tool based on SBEM is
presented in Section 7. The final conclusions are drawn in Section 8.

2. RANDOM FIF2_,D DISCRETIZATION

2.1 Background

The randomness of a stochastic system can be described in three forms, random variables,
random process in space, and random process in time. The random process in space is also called

3
randomfield. The aspectsof randomfieldsandits applicationto engineeringproblemsaregiven
by Vanmarcke(1984). Variousmethodshavebeenusedto thenumericalrepresentation of random
processes.The statisticalcharacterizationfor thunderstormwindshasbeengiven by Twisdaleand
Dunn (1983) andTwisdaleandVickery (1991). The spectralrepresentationof randomprocesses
basedon computersimulationhasbeenproposedby Shinozuka(1987).

The spatialvariabilityof mechanicalpropertiesof a systemandtheintensityof a distributed

load canconvenientlyberepresented by meansof randomfields. Due to thediscretenatureof the
finite elementformulation, therandomfield mustalsobediscretizedinto randomvariables.This
processis commonlyknown asrandomfielddiscretization.Variousmethodshavebeendeveloped
in therepresentationof randomfields. They are:themidpointmethod(HisadaandNakagiri,1985;
Der Kiureghian andKe, 1988;Yamazakiet al., 1988),the spatialaveragingmethod(Vanmarcke
andGrigoriu, 1983),andseriesexpansionmethod(Lawrence,1987;SpanosandGhanem,1988).
In this section, the interpolation method(Liu et al., 1986a)is described. In this method,the
randomfield is representedby a setof deterministicshapefunctionsandthe randomnodalvalues
of the field. The sizeof therandomfield elementis controlledby thecorrelationlengthof thefield
andthe stability of the probability transformationusedin the reliability methods(FORM and
SORM). The randomfield meshshouldbeso fine to capturethe fluctuationof the randomfield.
On the other hand,the randomfield meshshouldnot be so small thathigh correlatedstochastic
variablesof adjacentelementscausenumericalinstabilityin the probabilitytransformation,which
is required in the reliability methods(FORM and SORM). As suggestedby Der Kiureghian
(1985), two separate meshes for the finite element and for random fields have to be used in the
numerical implementation.

Since the computational effect in the determination of response derivatives or sensitivities is

proportional to the number of random variables, it is desirable to use as few random variables as
possible to represent a random field. To achieve this goal, the transformation of the original
random variables into a set of uncorrelated random variables has been introduced by Liu et al.
(1986a) through an eigenvalue orthogonalization procedure. Comparison with a Monte-Carlo
simulation demonstrates that a few of these uncorrelated variables with larger eigenvalues are
sufficient for the accurate representation of the random field. This technique along with other
computational aspects is presented in Section 5.

2.2 Interpolation Method

Let b(x) represent the random field. In PFEM, b(x) is approximated by

b(x) = _ Ni(x) b.a (2-1)

i=l

where Ni(x) represent the shape functions and b.1the discretized values of b(x) at x i, i = 1..... q.
It follows from Eq. (2-1) that

db(x) = _ Ni(x) db.1 (2-2)

i=l

and

4
 db2(x) = _ Ni(x) Nj(x) db._db.j (2-3)

ij=l

where

db. = b.-b. (2-4)

1 1 l

and b. represent the mean values of b. (also denoted by the expectation operator E[.]). From Eq.
1 1

(2-1) the expectation and the covadance of b(x) are, by definition,

+OO

E[b(x)]= J" b(x) f(b) db (2-5)

-00

= _ Ni(x) E[b i] (2-6)

i=l
and
.+OO

Cov[b(Xk), b(Xm)]=
-OO
f [b(Xk)- b(Xk) ] [b(Xm) - b(Xm) ] f(b) db (2-7)

_ Ni(Xk) Nj(Xm) Cov(b i , bj) (2-8)

ij=l

respectively, where fib) is the multivariate probability density function; x k and x m are any two
points in the domain of x.

From second-moment analysis, the mean of any function S[b(x),x] at any point x k and the

covariance of the function between any two points x k and Xm can be written as

1 _ a2Sk
E[Sk] = Sk + 5.. _7ff_. Cov(b i , bj) (2-9)
1j=l 1 j

and
 (M/oO
ij=l

where

S k = S[b(x),xk] (2-11)

and the superposed bar implies evaluation atb. The error in Eqs. (2-9) and (2-10) arises from: (i)
the truncation of higher order moments and (ii) the discretization of the random field b(x) by the
finite vector b. If the randomness in b(x) is small, then the first error will be small for a smooth
function and the second-moment analysis is applicable. The error due to discretization in Eqs. (2-
6) and (2-8) has been studied by Liu et al. (1987).

When the random field discretization is coupled with a FEM discretization, as in PFEM, q
need not be equal to the number of finite elements NUMEL and the shape function Ni(x) need not
be the same as the finite element interpolants for the displacement field. As indicated before, two
meshes, one depending on structural topology, the other on correlation length can be employed to
improve the computational efficiency.

3. PROBABILISTIC FINITE ELEMENTS FOR LINEAR PROBLEMS

The probabilistic finite element method (PFEM) is used to study of systems with parametric
uncertainties in both the unknown function and mathematical operators acting on it. The loads can
be either deterministic or random. In this section, the second-order perturbation is employed to
develop PFEM for a general linear transient problem. By applying the second-order perturbation,
the random linear system equations can be replaced by a finite number of identical deterministic
system equations up to second-order. The effective load in each of these equations depends on the
randomness of the system and the solutions of the all the lower order equations.

Due to the space limitation, the review on the deterministic finite element method is not
given here. The state-of-the-art of finite element techniques can be found in the review article by
Noor (1991). Using either the single field variational principle or Galerkin formulation, the
discrefized linear equations of motion are

M a(b, t) + K(b) d(b, t) = f(b, t) (3-1)

where M and K(b) are the (neq × neq) global mass and stiffness matrices, respectively; a(b, t),
d(b, t), and f(b, t) are the (neq x 1) nodal acceleration, displacement, and force vectors,
respectively; neq is the number of equations; and b is a q-dimensional discretized random variable
vector, i.e., b i = b(xi), where x._ is the spatial coordinate vector. The mass is usually assumed to
be deterministic whereas the probabilistic distributions for the stiffness and external force are

represented by a generalized covariance matrix, Cov(bi, bj), i, j = 1 .... , q. It is worth noting that
the stiffness matrix can be expressed in terms of the generalized gradient matrix, B(x), and the
material response matrix, D(b, x). In this formulation, the random vector, b, can represent a
randommaterialproperty(e.g.,Young'smodulus)and/ora randomload.

The application of second-momentanalysisto develop PFEM involves expanding all

randomfunctionsaboutthe meanvalueof the randomvector b, denotedby b, via Taylor series

andretainingonly termsup to second-orderterms. Thatis, for a smallparameter,4, the random

displacementfunctiond(b, t) is expandedaboutb via a second-order

perturbationasfollows:

1 2 1 .1
i=l 1 ij=l i j

where d(t), db (t), and db.b.(t) represent the mean displacement, the first-order variation of
l ij

displacement with respect to b. evaluated at b, and the second-order variation of displacement with
1

respect to b. and b. evaluated at b, respectively and Ab. represents the first-order variation of b.
1 j 1 1

about b.. In a similar manner, K(b), a(b, t), and fib, t) are also expanded about b. via a second-
1 I

order perturbation. Substitution of the second-order perturbations of the random function d(b, t),

K(b), a(b, t) and fib, t) into (3-1) and collecting terms of order 1, {, and 4 2 yields the following
equations:

Zeroth-Order Equations

Ma(t) + Kd(t) = f(t) (3-3)

First-Order Equations (for each Ab i, i= 1..... q)

Mab.(t) + K db.(t) = Fb.(d, t) (3-4)

1 1 1

where

Fb.(d, t) = lb.(t) - Kb. d(t) (3-5)

1 1 1

Second-Order Equations (i and j are summed from 1 to q)

7
 Ma2(t) + Kd2(t) = V2(d, t) (3-6)

where

{ 1
_ -fb.b.(t) _1 Kb.b.
- d(t) - Kb. db (t) } Cov(b i, bj) (3-7)
F2(d, t) = -
ij=l l] 1j x j

a2(t) = 2 _ ab.b.(t)C°v(bi'bj) (3-8)

i j=l 1j

d2(t) = 21- _ d-b.b.(t ) Cov(bi, bj) (3-9)

ij=l 11

The solution process for Eqs. (3-3) through (3-9) c.a:n be performed in parallel since only one
effective stiffness matrix needs to be formed. Therefore, the total solution requires one
factorization of the effective stiffness matrix and q+2 forward reductions and back substitutions of
an (neq × neq) system of linear equations to obtain the zeroth-, f'trst-, and second-order solutions.

To illustrate the performance of PFEM, a simple two degree of freedom spring-mass system
is presented here. The computed results are compared with those obtained using (1) Monte Carlo
Simulation (MCS) and (2) Hermite-Gauss Quadrature (HGQ) schemes. The problem is depicted
in Fig. 7- I. A sinusoidaI vector forcing function is used:

F(t) = I 25.0 x 1060.0 sin 2000 t ] (3-10)

The random spring constants K 1 and K 2 are normally distributed with a coefficient of variation
(i.e. 07#) equal to 0.05. The mean spring constants are 24 x I06 and 12 x I0 6, respectively. The
deterministic masses m I and m 2 are 0.372 and 0.248, respectively. A stiffness-proportional
damping of 3% is included. The probabilistic equations derived earlier are solved by the implicit
Newmark-_ method (Ma, 1986). The mean amplitude dl is depicted in Fig. 7-2 for all the three
numerical methods-PFEM, HGQ and MCS. The PFEM solution compares very well with the
other two methods. For the variance of dl the PFEM solution, plotted in Fig. 7-3, seems to
overshoot the variance at large time. The +3crbounds for the displacement dl is plotted in Fig. 7-
4.

4. PROBABILISTIC FINITE ELEMENTS FOR NON-LINEAR PROBLEMS

The probabilistic finite element method has been developed in the previous section using the
single-filed variational principle. Due to the direct stiffness matrix approach used, it can be only
applied to solve a limited number of problems with uncertainty in loading and material properties.
In orderto consistentlyhandleproblemswith randomnessin theequilibrium equations,domain,
andboundaryconditions,the three-fieldHu-Washizuvariational principle will be employedto
developPFEM. An additionaladvantageof usingthe Hu-Washizuvariationalprinciple involves
the eliminationof the locking phenomena(Belytschkoand Bachrach,1986)and suppressionof
hourglassmodes(Belytschko et al., 1984). Solution of three stationary conditions for the
compatibilityrelation,constitutive law, and equilibrium yield the variations in displacement, strain
and stress. The statistics such as expectation, autocovariance, and correlation of displacement,
strain and stress are then determined.

Using matrix notation, the Hu-Washizu Variational Principle (HWVP) for nonlinear
problems adopted in this Section is (see Liu et al., 1988c)

5_ T(_-cy) df2 - f 8G T(e-Vu) df2

f2 Xq

+ f 5(Vu) T CY dk'2 fT 5u F dff2 5u h dF = 0 (4-1)

if2 f_ af_ h

where E, _, and u are independent random field variab'les representing the nonsymmetric measure
of the strain, first Piola-Kirchhoff stress, and displacement, respectively; _ is a nonlinear function

of the deformation gradient; and a superscript T represents the transpose. In Eq. (4-1): f2, _h'

F, h, and Vu represent the domain, traction boundary, body force vector, prescribed traction
vector, and the nonsymmetric part of the displacement gradient, respectively; 5 represents a virtual
quantity. The surface and volume integrals in Eq. (4-1) can be expressed via parametric
representation:

dF = J dA and d_ = J dR (4-2)
S v

respectively, where J and J represent the surface and volume Jacobians, respectively; and R and
S v
A represent the reference domain and boundary, respectively. Random domains and boundaries
are incorporated into the formulation through randomness in the gradient operator and Jacobians.
The application of second-order perturbation techniques in the HWVP involves the expansion of all

random functions about the mean value of the random field b(x) denoted by b(x) and retaining
only up to second-order terms, i.e., for a given small parameter [{ = the scale of randomness in

b(x)J, the random function _ is expanded about b at a given point x in the reference domain as
follows:

= • 0 + _(V ' + C0e .....

) + _2(V +C_ +C 0 e " ) (4-3)

where the superscripts nought, prime, and double prime represent the random functions evaluated

at b, the first-order variation due to variations in b, and the second-order variation, respectively.
The first elasticity tensor, C in Eq. (4-3) is given by
 (4-4)
Cijkm = 3GijOGk m

where W is the strain energy density function;

and G is the deformation gradient. Similarly, the
T
rest of random functions e, c, F, h, Js, Jv, Vu and _5(Vu) can also be expressed as second
order perturbations (see Liu et al., 1988a). After substituting the second order perturbations of all
these random functions into Eq.(4-1), the following three equations for the zeroth-, first-, and
second-order nonlinear PHWVP axe obtained:

Zeroth-Order Variational Principle

Rf _ T(Xl/0 _ 0) j0v dR - Rf _(_T(E0 - _'0U0) j0v dR

+ 5(V0u) T 0j0dR - 5u dR - 5u T
Rf Rf TF0 J0 _ h0 j0dA =0 (4-5)
v v 3.Rh s

First-Order Variational Principle (_ terms)

T j0 +xl/ j0+ - J ]dR

Rf 8e [(C0e ' -G) ' v ' v ( 0 cO) v'

T . 0 ' ' 0 j0 0u0 )

- SO [(e - X7 u - X7 u ) v + (c0- V Jr] dR
t

+ C , + 5(V , u) T 0] + ,5(V 0u ) T 0 J _} dR
j0
Rf{[_i(V0u) T v

- ,Su (F j0 + J )dR - _u (h j0 + h J )dA = 0 (4-6)

Rf T , F0 , f T , 0 '
v v _ l_ h S S

Second-Order Variational Principle (_2 terms)

Rf 5e T [( C0 E ....-_) j0v+(Ce+_)
.... j0v+( C0E ' +_-_)
' ' J'v + (_/0 - 0)j'_ ]dR

10
 J0+(E-V u-_7uO) J + (a 0 - V0u 0) J_] dR
- G15 [(e -V u -Vu-V u ) v v
T ,, 0 ...... 0 ' O' ' '

+ {[_i( O + G(V u) (5 + _(V I1) ] jO

Rf _70u)T ,, , T , ,, T O0

+ [G(V 0 U) T G , + G(' _7U) T G O] J ' +5( V0u )

v

f Gu T (F " j0v +F ' J v' + F 0 j'_ )dR

R

T " 0+h J +h 0 )dA = 0 (4-7)

- 5U (h Js s
_ J'pT,
h ' , j'_

It should be noted that all random functions with the guperscript C ) or (') in Eqs. (4-6)-(4-7) are,
in general, described through spatial expectation and autocovariance functions. Therefore, in
addition to the usual finite element approximation of the displacement field, the random fields are
also discretized with q shape functions. To be consistent with the finite element approximation and
to maintain the accuracy of the discretized random field [i.e., b(x)], the random functions _1/, C, F,

h, Js' and Jr' which are, in general, functions of b(x) and x, are first discretized with the same q
shape functions as the random fields. For example, the finite element approximation of C is given
by

C = C 0+ _C'+ _2 C " (4-8)

or

c = ¢i(x)( c 0 + _ CI+
' _2 CI" ) (4-9)
I=l

where; 0I(X) are the q shape functions; C 0I denotes the Ith nodal value of C evaluated at b; C I

tt

denotes the first-order variation of C(x I, b) due to variations Abi; and C I denotes the second-order
variation. The last two are then expanded in terms of the random variables bi and given by

(4-10)
CI = (C Ab i
i=l

ll
and

CI - 2 (C Ab.1 Ab.j (4-11)

ij=l

respectively. The factor 1/2 is included in order to be consistent with the second-order Taylor
t II

series expansion. The nodal values (CI) i and (Ci)ij can be obtained by partial differentiation of C
or by a least-square fit to the actual data. Similar definitions can be developed for the rest of
random functions (see Liu et al., 1988a).

Substituting the given approximation of all random functions into the zero-order, first-
order, and second-order PHWVPs (Eqs. (4-5)-(4-7)), and using the three stationary conditions
(strain-displacement, stress-strain, and equilibrium), the zeroth, first and second order equations
can be obtained (see Liu et al., 1988a). The zeroth-order equations require an iterative solution
technique, but the tin'st-order and second-order equation are linear. After determining the zeroth-,
first-, and second- order solutions, the expectations and autocovariance matrices for the
displacements, strains, and stresses can be obtained.
The applicability and effectiveness of the PFI_M for nonlinear problems was demonstrated
by Liu et al. (1988a) through the problem of a cantilever beam subjected to large deflection. The
Saint Venant-Kirchhoff model for nonlinear elasticity with randomness in the external force, beam
heig.ht, and material properties were considered. The probabilistic distribution for displacement,
strata and stress were also presented.

To reduce computational effort, the random variables can be transformed to the uncorrelated
normal form by an eigenvalue problem as shown below.

5. COMPUTATIONAL ASPECTS

5.1. Random Variable Transformation

The mean and covariance can be obtained from the equations in Section 4. However, the
number of derivatives to be evaluated is proportional to q(q+l)/2, where q is the number of
random variables. To reduce computations, an eigenvalue orthogonalization procedure, which is
similar to the modal analysis in structural dynamics, can be employed. The full covariance matrix

Cov(b i , bj) is transformed to a diagonal variance matrix Var (c i , cj) such that

Var(c i , cj) = 0 for i #j (5-1)

and

Var (ci, cj) = Var (ci)

for i = j (5-2)

Therefore, the number of evaluations is proportional to q. The above is achieved through the
eigenproblem:

f2_ = WA (5-3)

12
wherethef2 andA matricesdenoteCov(bi , bj) andVar (ci , cj), respectively;W is a constantq x
q fundamentalmatrix with thefollowing properties:

T T
h° hv = hv hv = I (5-4)

T
A = tp f_tp (5-5)
and

T
b = _Pc or c = q_ b (5-6)

I is the q x q identity matrix and c is the transformed q x 1 vector of random variables. Thus, the
discretized random vector b is transformed to an uncorrelated random vector c with the variance of

c as the eigenvalues of X'-2in Eq. (5-3).

With Eqs. (5-5) and (5-6), the mixed derivatives appearing in Section 5 reduce to second
derivatives and Vat (b i , bj) reduces to Vat (ci). Thus, the mean of any function S[b(x), x] at any
point xk and the coveriance of the function between any two points x k and x m can be written as

1 _ O2Sk (5-7)
E[Sk] = Sk + 2i = 1 7 Var(ci)
1

and

co%, =mj=ij=l
respectively.
It is observed that for one-dimensional random fields, as the correlation length increases
from zero to a large value, the number of largest eigenvalues n, n _<q necessary to evaluate the
mean and covariance in Eqs. (5-7) and (5-8) to a specified accuracy, decreases from q to 1. When
the correlation length is zero the random field is uncorrelated and all q eigenvalues are dominant.
As the field is uncorrelated, all q random variables are necessary to represent the randomness of the
field. As the correlation length increases the number of dominant eigenvalues decreases.
Eventually, for a very large correlation length the random field is closely correlated and there is just
one dominant eigenvalue. As the field is closely correlated, only one random variable,
corresponding to the largest eigenvalue, is sufficient to represent the randomness of the field. This
feature, when present, can easily be exploited to reduce the computations. The value of n can be
chosen based on the distribution of the eigenvalues before solving the PFEM equations. The
eigenvalues here can be interpreted as weighting factors for the corresponding mode shapes
necessary to represent the covariance structure; a large eigenvalue means a dominant mode and vice
versa. Results of the eigenvalue distribution and selection of n, for beam problem and a bar
problem, are discussed in Liu et al. (1986a, 1987).

5.2. Adjoint Method in PFEM

Consider a typical function H(c, d) involving the displacements d and the random variables

__3
c. Chaindifferentiationyields
T I

(5-9)
[rl]c=rl C. +rld dC.
I I I

where the subscript denotes the derivative with respect to c., and
I

T
(5-10)
n d = ( , )
I-ldl l-ld2 ..... rldNEQ

Using the first-order equation of PFEM in the transformed space, i.e.,

t t

Kd. = f. (5-11)
1 1

Eq. (5-9) becomes

TK-1 ,
In] c = nc. + na f. (5-12)
1 1 1

Usually, in the direct method, the above equation is evaluatedfor each random variable c i,

involving 'n' solutions of the linear equation (5-12). In the adjoint method, _. is selected to satisfy

(5-13)
K3. = Fl d

Then, Eq. (5-12) can be rewritten as

T t

= FI +_. f (5-14)
[n]c. c. c.
1 I 1

The adjoint problem, Eq. (5-13), is solved only once in this method. In the direct method, 'n'
solutions of Eq. (5-11) are required. This is the advantage of the adjoint method over the direct
method. Both methods require 'n' inner products with f C. in Eqs. (5-9) and (5-14), respectively.
1
However, it has been shown that when the number of functions is more than the number of
random variables, the computational advantage of the adjoint method is lost (Liu et al., 1988d).
By solving 'q' adjoint problems, the second order sensitivites can also be evaluated. It should be
noted that the adjoint method is applicable to nonlinear problems as well, as the first and second
order equations are still linear.

5.3 Parallel Computing in PFEM

Recent advances in computing hardware and software, have made muhiprocessing in

general and parallel processing in particular a viable and attractive technology. Parallel processing
provides an opportunity to improve computing efficiency by orders of magnitude. Probabilistic
computational mechanics exhibits several inherent levels of both coarse-and finite-grained
parallelism. It is imperative to develop the computational strategies and algorithms to maximize

14
parallel processingefficiency and minimize parallel overhead. The parallelism in both the
probabilisticcomputationsandthe structuralmechanicscomputationshasbeenexploredby Sues
et. al (1991,1992a).The implementationof parallelprocessingto generalprobabilistic structural
analysisproblemshasbeenstudied by Sueset al. (1991, 1992a). The parallel computing for
probabilistic fatigueanalysishasbeenimplementedby Sueset al. (1992b) on both sharedand
distributedmemorymachines.

The parallel implementationof PFEMcan beeasily achievedin the solution of the first-
order equations(sensitivity analysis). As shown from Eqs. (3-3)-(3-9), only one effective
stiffnessmatrix needsto be formulated. Oncethe zeroth-ordersolution is obtained,q equations
(Eq. (3-4)) can be solved in parallel to determinethe responsederivatives. Multiple levels of
,_rallelism can be achieved if the substructuring (Komzsik and Rose, 1991), domain
decomposition (Chan,et al., 1989) and operator splitting (Sues and Chen, 1992c) are also
employedin PFEM.

6. APPLICATION OF PFEM TO THE RELIABILITY ANALYSIS OF BRITTLE FRACTURE

AND FATIGUE

6.1 Introduction

In the previous section, the probabilistic finite element method, which is based on the
second-order perturbation, has been formulated to quantify the statistical moments of the response of
a stochastic system. In this section, the PFEM coupled with the first-order reliability method is
developed to determine the reliability of brittle fracture and fatigue crack growth. The constrained
optimization problem is formulated to calculate the reliability index. A Lagrange multiplier technique
along with gradient projection algorithms is used to solve the constrained optimization problem.

Fracture and fatigue have become important factors in the structural design and safety of
aging structures. The failure of an aging structure is usually resulted from microdefects activation,
propagation, and formation of major cracks. Due to the ramdomness in the configuration of
microdefects and uncertainty in the failure mechanism, the probabilistic fracture mechanics (PFM),
which combines the fracture mechanics with the stochastic methods, provides a useful tool to
address problems with large uncertainty.

The reliability analysis of flawed structures will here be classified into two groups of
problems:

1. brittle material problems, where the material contains flaws with the random location and
orientation. The major question is the reliability of the structure in the presence of these flaws.

2. ductile material problems, where failure is expected to result from the growth of a critical
flaw until it can lead to failure of the structure.

The first category of the problem has been addressed recently by Lua et al. (1992a, 1992b) in
quantifying the inherent statistical distribution of the fracture toughness of a multi-phase brittle
material. The second question is of particular relevance to the safety of aging structures and
nondestructive evaluation techniques. Because the threshold of detection is substantially greater
than flaws sizes which may lead to failure over the course of time, inspection cycles should be set
so that the reliability of an aging structure remains acceptable in these circumstances. Although a
deterministic analysis can obtain an estimate of the fatigue life, the uncertainties in crack growth
rates and the initial crack lengths detract from the usefulness of such solutions.

3.5
 In this section, the application of the PFEM and the first-order reliability analysis for the
brittle fracture and fatigue is presented. A brief review on the reliability methods is given first.
The fusion of the PFEM and reliability analysis for probabilistic fracture and fatigue are then
presented. Performance of the methodology developed is demonstrated on example problems.

6.2 Reliability Analysis

Reliability analysis is used to determine what is the likelihood that a structure subjected to
uncertain loads, material properties and geometry, will satisfy a limit state criterion. Several text
books and monographs on the methods and application of the reliability theory have been written,
i.e. [Ditlevsen (1981), Ang and Tang (1984), Augusti et al. (1984), Madsen et al. (1986) and
Melchers (1987)]. An overview of the first-order and seconder-order reliability methods
(FORM/SORM), as well as various Monte Carlo simulation schemes has been given by Bjerager
(1989). As the PFEM provides a powerful computational tool to determine first- and second-
moment of random parameters, the second-moment reliability methods can be easily combined
with the PFEM to obtain measures of the reliability of the structural systems.

Throughout this section the uncertainties-in load, material properties, component geometry
and crack geometry- are represented by a q-dimensional vector of random variables denoted by b =
T
[b 1, ..., bq] . A random variable reliability problem is described by a performance function, g(b),
which is a continuous measure of the ability of a system to perform as designed. Three states of a
system, namely, the limit-state surface, the failure state, and the safe state, are defined by

g(b)=0 g(b)<0 and g(b)>0 (6-1)

respectively. The probability of failure is given by

f
Pf = ] fB(b) db (6-2)
b)_:0

where fB(b) is the multivariate density function of b. Two difficulties are associated with Eq. (6-
2). First, the domain of integration (g(b) _ 0) is an implicit function of the random vector b.
Second, standard numerical integration of this multiple integral is prohibitively complicated as the
number of random variables becomes large. Two approaches- MCS and failure surface
approximation methods such as the first or second order reliability method (FORM or SORM)-
have been employed extensively to calculate Eq. (6-2). In the first-order reliability method
(FORM), the limit-state surface in the standard normal space is represented by the tangent
hyperplane at the design point. In the second-order reliability method (SORM), the limit-state
surface in the standard normal space is replaced by a quadratic surface, which is tangent at the
design point. While MCS is completely general, it is very expensive and time-consuming for small
probabilities of failure, which is the major concern in reliability engineering. FORM and SORM
are more accurate and efficient for extreme probability of failure (e.g., 0.0001 or 0.9999
cumulative probability), however implementation can be more complex. In the present study, the
FORM is applied to predict the reliability of a flawed component.

In order to make use of the properties of the standard normal space (rotationally symmetric
and exponential decay), a transformation is introduced to map the original random variables b to a
set of standard, uncorrelated normal variables r. Eq. (6-2) in the r-space becomes

16
 (6-3)

where( )t denotesthe transposeof a vector or a matrix, and_(r) (= g(b(r))) is the performance

functionin the transformedr-space. FORM approximatesthecalculationin Eq.(6-3)as follows:
first thepoint r* on thelimit-statesurface(_(r) = 0), which hastheminimumdistanceto theorigin,
is found through an iterative algorithm, then the limit-state surfaceat the design point r* is
replacedwith a tangenthyperplanegivenby

_(r) = --b_(r)(ri - r_) (6-4)

_ri
The resulting first order approximation to Eq. (6-3) is

= _(-[3) (6-5)
= f__ (2/1:) "q/2 exp(- ½ r t r)dr
Pfl g(r_____)
(ri- r_)_<0
_ri

where the reliability index [3 is defined by

13= (r "t r') 1/2 (6-6)

and _( ) is the standard normal cumulative probability. The step to determine the most probable
point (r")on the failure surface is the most critical in the reliability analysis. It generally requires to
form an iteration and optimization scheme to calculate the gradients of the performance function.
In this paper, the reliability index _ is determined by solving the following optimization
problem in r-space, i.e.,

[3 = min (rTr) I/2, subject to g(r) = 0 (6-7)

The optimization can be solved using any general non-linear optimization algorithm such as HL-RF
method (Hasofer and Lind, 1974; Hohenbichler and Rackwitz, 1981; Rackwitz and Fiessler,
1978), gradient projection method (Haug and Arora, 1979) and the modified HL-RF method (Der
Kiureghian and Liu, 1988). A fast convergence rate is essential in selecting an iteration method.

The second order reliability method based on the second order Taylor expansion of the
failure surface is given by Fiessler et al. (1979), Breitung (1984), Der Kiureghian et al. (1987),
and Tvedt (1983).

6.3 Brittle Fracture Reliability Analysis

In order to model the singularity at the crack-tip, Bestfield et al (1990) used enriched
element (Gifford and Hilton, 1978). Other methods such as J-integral approach (Rice, 1968) and
hybrid elements (Akin, 1976; Barsoum, 1976; Henshell and Shaw, 1975; Tong et al., 1973) can
also be used. The enriched element approach has the advantage that mode I and II stress intensity
factors, _: and _: are directly calculated along with the nodal displacement. This simplifies the
! II
development of the sensitivity equations which are needed in f'u'st-order reliability analysis.

3_7
 The discretizedglobal finite elementequationsareobtainedby assemblingthe enriched
elementsthat surroundthe crack-tip and the regular elementsthat model the remainderof the
continuum. The globalsystemof [neq+ 2] equations(i.e.,numberof displacementequationsplus
modeI andI/stress intensityfactors)is

K(b) 5(b) = fib) (6-8)

where the generalized displacement, 8, and external force, f, vectors are

5(b) = _(b) and f(b) = O(b) (6-9)

respectively and the global stiffness matrix, K(b), is given by

T (6-10)
K(b) = I C(b)
R(b) C(b)
E(b) J

In Eqs. (6-8) through (6-10): d and h are the regular displacement and external force vectors,
respectively; R, E, and C are the [neq x neq] regular stiffness matrix, the [2 x 2] stiffness matrix
from the enriched terms, and the [neq x 2] coupled stiffness matrix from the regular and enriched
terms, respectively. The other submatrices in Eq. (6-9) are

K(b) = { l':I(b) t and O(b) = fI(b)

fli(b) } (6-I 1)
_:ii (b)

where the two terms fI and fll are zero if the enriched element is not on a loaded boundary.
Equations (6-8) through (6-11) are solved by condensing out the stress intensity factors (i.e., static
condensation).

For mixed Mode I and Mode II fracture, several kinds of fracture criteria have been
summarized by Wu and Li (1989). Among these criteria, the most widely used are: the maximum
principal stress criterion proposed by Erdogan and Sih (1963) and the minimum strain energy
density criterion, Sih (1974). In the case of mixed mode fatigue, the fatigue laws are generally
based on an equivalent Mode I case to simulate actural mixed mode behavior. In order to be
consistent with the mixed mode fatigue laws, the maximum principal stress criterion (Erdogan and
Sih, 1963) is applied here to determine the equivalent Mode I stress intensity factor. Thus, the
performance function for the mixed mode fracture can be expressed as

g(b) =Kc-r, eq (6-12)

Equation (6-12) implies that fracture occurs when the equivalent Mode I stress intensity factor,

r,eq,.exceed, s the critical value, r,c. The direction of crack growth where the hoop stress becomes
maximum is given by

T
Z(K:,0) = e _¢ = 0 (6-13)

18
where

sin 0
(6-14)
3 cos 0 -

In Eq.(6-14),0 is measuredfrom thecurrentcrackline. Therelationbetweenthe equivalentMode

I stressintensityfactor (_:eq)andstressintensityfactor(_:I,_rr) is given by
T
_c = _ _ (6-15)
eq
where

= cos
2

I IE
el COS

- 3 sin
_-

0
(6-1 6)

and 0 is determined by Eq. (6-13). When only Mode I or Mode II fracture is present, Eq. (6-12)
can be rewritten as

g(b) =nc-hi, i= I, II (6-17)

where r,c is given by

r,.c=_ic (for Mode I) and r-c = 21_--_"Ic (for Mode II) (6-1 8)

In Eq. (6-18), _Ic stands for the fracture toughness. As indicated in Section (6.2), the
determination of the reliability index for calculating the first-order probability of failure in the
FORM is achieved by solving an optimization problem with one constraint (limit-state condition).
In order to incorporate other constraints such as equation of equilibrium, crack direction law (in
fatigue crack growth problem) in the formulation, the method of Lagrange multipliers can be
applied. The statement of the optimization problem for brittle fracture is described in the
following.
The nonlinear programming problem consists of determining the correlated random

variables, b = [b 1, .... b q] T, and the generalized displacements, sT = [d T , _cT ], that minimize the
distance from the origin to the limit-state surface in the independent standard normal space. The
minimizer is termed as the reliability index 13 (Eq. (6-7)). The minimization is subject to the
following equality constraint:

K(b) 8(b) : _b) (6-19)

(i.e., equilibrium) and the following inequality constraint:

19
 g(b) ---0 (6-20)
(i.e., theperformancefunctionbeingon the limit-statesurfaceis a constraintin the optimization
problem).
Equations(6-19),and(6-20)areconvertedto the Kuhn-Tuckerproblem(Arora 1989)by
defininga.Lagrangefunctional,L, of independent variables b, 8, p., k, and a as follows
T 2
L(b,&lt,_.,a) = rTr + _ [f- K 5] + X [g + a ] (6-21)

where la is a Lagrange multiplier for equilibrium, X > 0 is a Lagrange multiplier for the inequality
constraint, and a is a slack variable that is introduced to ensure that g < 0. Depending on the sign
of k, the function to be minimized will increase or decrease with a change in g. In other words, if
_>0, then rTr will decrease (i.e., minimize) while g < 0 (Converse 1970). The Kuhn-Tucker
necessary conditions for the minimization of Eq. (6-2 I) are obtained by setting the derivatives of
the Lagrange function with respect to the independent variables b, & It, k, and o_ to zero, i.e.,

_L - c3[rT
c3br ] + B T {_ [f- K _]} + _.. -_b = 0 (6-22)

_L _ -P r K + _. ag = 0 (6-23)
a5 38

aL
- f - K 6 = 0 (6-24)
ap

aL 2
= g + a - 0 (6-25)
aZ

aL
- 2 _. o_ - 0 (6-26)
ao_

The optimization requires the solutions of Eqs. (6-22)-(6-26) for b, 5, It, _. > 0, and a. Equation
(6-24) is simply equilibrium; and Eqs. (6-25) and (6-26) can be simplified to eliminate the slack
variable, ¢x, such that, _. g = 0 and g < 0 which ensures that _. > 0.

Since 5 and b are independent variables in the Lagrange function (see Eq. (6-21)), the
partial derivative of the second term with respect to b in Eq. (6-22) can be expressed as

d [f_
b--6 Of OK
-5ff a (6-27)

To simplify the right hand side of Eq. (6-27), the first-order probabilistic finite element equation
(Eq. (3-4)) is employed for the present static problem, i.e.,

3fi _f 3K
K _-6- = _--i_- -,b--1_-5 (6-28)

2O
whichyields

(6-29)
d [f-KS]=K _d_i

when Eq. (6-28) is substituted into Eq. (6-27). Now multiplying each side of Eq. (6-29) by 1aT
and using Eq. (6-23) in the right hand side we obtain

dg 25 (6-30)
gT _b [f- KS] = _. _--_ _-_

which can be expressed as

dg _K: (6-31)
p T _b [f- KS] = _. _

since g is only a function of _.

Substituting Eq. (6-31) into Eq. (6-22), the final optimization problem becomes

L [3 + _.Lg = 0 (6-32)

when k >__0 where

Lg = _ + _3_: OK:
3--_ (6-33)

LI3 - _b [rTr] (6-34)

finite difference depending if the

In Eqs. (6-34), z_ [rTr] is computed either explicitly or by
random variables are normal or non-normal. In order to perform the sensitivity analysis on the

stress intensity factors, namely, _-, the probabilistic finite element method described in Sec. 3 can

be applied to accomplish this task. Since we are only interested in the sensitivity of the stress
intensity factors, considerable computational effort can be saved by using the adjoint method as
described in Section 5.2. The iteration algorithm for the brittle fracture reliability is given by
Besterfield et al. (1990).

In order to demonstrate the applicability of this approach to the brittle fracture reliability
analysis, a single edge-cracked beam subjected to a concentrated point load is considered (see Fig.
7-5). The problem constants are given in Table 7-1. Due to symmetry, ten regular 9-node
elements and two enriched 9-node elements are depicted in the left half of the beam as shown in
Fig. 7-5. The applied load is modeled with one random variable with a coefficient of variation of
0.1 and the crack length is also modeled with one random variable and the coefficient of variation
of 0.1. The convergence criterion for the optimization is 0.001. The variance of the Mode I stress
intensity factor with randomness in force, material, crack length and the combination is presented
in Table 7-2 for the adjoint method. Also presented in Table 7-2 are the summaries of the
numerical performance and results of the reliability analysis (e.g., starting point, number of

21
iterations,the failure point, reliability index,andprobability of failure). As shownin Table7-2,
the randomcrack length haslesseffecton the probabilityof failure dueto the smallervariancefor
randomcrack length.
)
6.4 Fatigue Crack Growth Reliability Analysis

Fatigue crack growth is sensitive to many parameters and these parameters can seldom be
determined accurately. Uncertainties in the crack geometry, material properties, crack direction,
crack length, component geometry, and load time history all play a role. Thus, the prediction of
fatigue failure must be treated as a probabilistic problem.

The first order second moment reliability method (FORM) can be applied to this problem as
before by solving a constrained optimization problem. Due to the combined effects of external
loading, unsymmetrical component geometry and crack geometry, cracks rarely grow in a straight
line. Thus, the mixed-mode fatigue crack growth law and crack direction law should be employed.

The most common law for fatigue crack growth is the Paris-Erdogan model (1963), which
gives the fatigue life, T, by

af

da (6-35)

T = f D (A_:eq) n
a
1

where a 1 and af are the initial and final crack lengths, respectively, da is the random crack path; D
and n are primarily material parameters but can also depend on the loading and environmental

effects; and AKeq(a) is the range of equivalent Mode I stress intensity factors, i.e.,

max rain
A_ = _ _: (6-36)
eq eq eq

min max
where n and K: are the minimum and maximum equivalent Mode I stress intensity factors,
eq eq
respectively, associated with the minimum and maximum cyclic applied stresses, respectively. If
the minimum equivalent Mode I stress intensity factor is assumed to be zero, then

max
&_: = K: - _ (6-37)
eq eq eq

The direction of the crack can be considered to be a random function, which will depend on the
material properties and the history of the loading and the crack path. At each step, the statistics of
the crack-tip, as reflected in this random function, in conjunction with the previous length of the
crack and its orientation, will be used to obtain the new configuration. Based on the maximum
hoop stress criterion (Erdogan and Sih, 1963), the crack growth direction Z(_:, 0) given by Eq. (6-
13) is also employed here. Unlike the case of brittle fracture discussed before, the performance
function for fatigue crack growth is given by

g = T - T (6-38)
S

22
where T is the servicelife of the component. In other words, the componentfails when the
s
fatiguelife is lessthanthedesiredservicelife. The performancefunctioncould alsobeexpressed
in termsof acritical cracklength.

Thecalculationof thereliability indexby thefirst-orderprobability theoryis performedin

the sameway as before by solving a constrainedoptimization problem. Before stating the
optimization problem for fatigue crack growth, the crack direction law (Eq. (6-13)) must be
discretizedinto "npts"discretizationpointsalongthecrackpath. At eachcrackpathdiscretization
point,thecrackdirectionis
T
Zk = Ok Kk = 0 k = 1..... npts (6-39)

where_:kandO k represent _ and O evaluated at { = {k' k = 1 ..... npts. Thus, at each crack path
discretization point, the new crack direction is recalculated and the crack is then allowed to grow to
the next discretization point.

The calculation of the reliability index by the first-order probability theory is posed as a
constrained optimization problem. Unlike the previous brittle fracture reliability problem, both
equality and inequality constrains have to be satisfied at each crack path discretization point. Also,
the crack direction law (Eq. (6-39)) has to be included in the Lagrange function. By defining a

Lagrange function, L, of independent variables b,/a 1 .... , I.tnpts, 51 ..... 5npts, tpl ..... tPnpt s,

01 ..... 0npts, k, and ix, we have

L (b, I.ti, _5i, _i' 0i' k, a)

r r + _i If. 1 - K.1 _5.]

1 + tpk Z k + _. T-T S + ¢x (6-40)
i=l k=l

where P'i is a Lagrange multiplier for equilibrium, (Pk is a Lagrange multiplier for the crack

direction law, tx is a slack variable which is introduced to ensure that g < 0 and X _>0 is a Lagrange
multiplier for the inequality constraint.
The necessary conditions of Eq. (6-40) (i.e., derivatives with respect to the independent
variables) then lead to

npts npts

Ob°_L - N°_ [rT r ] + _ If. - K._ _i ] + Cpk (Zk),b +X T, b = 0

i=l k=l
(6-41)

23
 3L
-f. K. _. =0 i = 1.... , npts, no sum on i (6-42)
3_t. 1 i 1
1

35. - - I'ti
wKi + ¢Pk (Zk)'5. + k T,5. = 0
l k=l I 1
i = 1..... npts, no sum on i (6-43)

3L 2
- T-T +a =0 (6-44)
3_. s

3L
-2Xa=O (6-45)

3L
m _ Z,
= 0 i = 1..... npts (6-46)
3<pi 1

0L _ST3Kk _ s '
= " k=l 8k + k=l
(Zk)'0
i
+ XT,0 = 0
1 1

i = 1 ..... npts (6-47)

where: T, b in Eq. (6-41), T,5. in Eq. (6-45), and T,0. in Eq. (6-47) are the derivatives of the
1 1

fatigue life T with respect to b, 5 i, and 0 i, respectively, assuming b, 5 i, and 0.a are independent

variables; and (Zk), b in Eq. (6-41), (Zk),8. in Fq. (6-43), and (Zk),e. in Eq. (6-47) are the
1 I

derivatives of the crack direction law Z k with respect to b, 5.,1 and 0 i, respectively, assuming b,

5, and 0. are independent variables. The optimization requires the solutions of Eqs. (6-41)-(6-47)
1 1

for b, gl ..... lanpts' 51 ..... 5npts' a' X >_0, q_l ..... g_npts and 91 ..... 0npts. Equation (6-42)
is simply equilibrium at each discretization point; Eq. (6-46) is the crack direction law at each
discretization point; and Eqs. (6.44) and (6-45) can be simplified to eliminate the slack variable, a,
such that _. g = 0 and g <_0, which ensures that _. >_0.

Since 5. and b are independent variables in the Lagrange function [see Eq. (6-40)], the
1
partial derivative of equilibrium with respect to the correlated random variables in Eq. (6-41) can be
expressed as

3f. 3K.
rt'i _K 1-
1 5i] = _- -b--ft 5.1 no sum on i, i= 1, ..., npts (6-48)

To obtain an expression for the right hand side of Eq. (6-48), the first-order probabilistic finite

24
elementequation(seeSection3) is employed,i.e.,

_8. Df. DK.

K. 3bl - Obx 1 no sum on i, i = 1 ..... npts (6-49)
Ob 8. 1

which yields

88.
no sum on i, i = 1 ..... npts (6-50)
"ffi - K.1 5 i]. = K.1

are independent
when substituted into Eq. (6-48). Since 51, ..., 8npts, and, 01 ..... 0npts
variables in the Lagrange function, Eqs. (6-43) and (6-47) simplify to:

T
i = 1..... npts (6-51)
" gi K.I + q_i (Zi)'8. + XT,8. = 0 no sum on i,
1 1

and

T c}K.
= 0 no sum oni, i= 1 .... ,npts (6-52)
gi 00. I 8._ + q_i (Zi)'0- + kT,o.
1 1
1

T
respectively. Multiplying each side of Eq. (6-50) by I.ti and substituting in Eqs. (6-51)-(6-52)

yields

OK.

T,8. K-1 18 -T 08.

i 0O. i '0.
1 1 1 1

T _ [fi - K. 8iI = k OK. (Zi),8. + T,8. Ob

1 1
K -1 18.
(Zi)'0. (Zi)'8. i 00. '
1 1 1

no sum on i, i - 1..... npts (6-53)

which can be expressed as

OK.
T,8. K-.1 I 8. - T,O.
I 00. 1
T 0 1 1 1 1

gi b-6 [fi- K., 8il = k< OK. (Zi),_" + T,_:. 0b

1 1
K-.1 i 8.
(Zi)'0- " (Zi)'& _ OO.
1 1 1

no sum on i, i = 1, ..., npts (6-54)

25
sinceT andZ.1areonly functionsof _c i, i = 1,..., npts. After substitutingEqs.(6-52) and(6-54)
into Eq. (6-41),the final form of the optimizationfor reliability for fatiguecrack growthis given
by

L 13+ _. L g = 0 (6-55)

where

L[3 =g6O [rT r ] (6-56)

DK.

T,& K-. I 18. -T

1 D0. 1 '0.
1 1 1 !
L g =n_s < DK. (Zi)'_:. -aft
i=l K-.1 .___..21
8 1 1

(Zi)'0.- (Zi)'8. _ D0. i

1 1 1

• OK.

T,& K-. 1 _8.- T '0.

i O0. 1 I
1 1
+
DK. (Zi)'b + T,b (6-57)
i=1 , K-.1 I 5.
(Zi)'0. (Zi) 5. _ D0.
_. 1 1 1

In Eq. (6-57): T, b, T,K: ., T,0., T,8., (Zi), b, (Zi),_: ., (Zk),0., and (Zi),8. are determined
1 1 1 1 1 1

explicitly in Reference (Besterfield et al., 1991). The sensitivity of the stress intensity factors,
DK.
namely #, is also computed by the PFEM.

In order to demonstrate the performance of the method for reliability analysis against failure
due to fatigue crack growth, a classical Mode I fatigue problem is presented. Figure 7-6 shows a
finite rectangular plate with a single edge crack of length a subjected to a distributed load. The
problem constants and second-moment statistics are given in Table 7-3. Due to symmetry, two
enriched 9-node elements and twenty-three regular 9-node elements are depicted on the upper half
of the plate. The reliability index is plotted versus the service life under the various types of
uncertainties for the reference solution t3 and the solution obtained by PFEM in Figs. 7-7a and 7-
7b, respectively. The same trends as the reference solution with the slight difference in the value
of the reliability index can be observed through comparison of Fig. 7-7a with Fig.7-7b. This
difference is due to the small numerical error in calculating the stress intensity factor by finite
element methods. As shown in Fig.7-7a, for a service life of 4 x 106 cycles, the reliability index
is less for uncertainty in the initial crack length (100 % coefficient of variation) and stress (25 %
coefficient of variation ) than for randomness in the final crack length (10 % coefficient of
variation), fatigue parameter D (30 % coefficient of variation), and fatigue parameter n (2.5 %
variation). When all five of the parameters are treated as random, the combined effect is much

26
greaterthananyoneindividualeffect,asexpected.
7. SBEMFORTHE CURVILINEAR FATIGUECRACK GROWTH RELIABILITY ANALYSIS

7.1 Introduction

The developmentof probabilisticfinite elementmethod(PFEM) and its applicationsto

linear, nonlinear structural mechanicsproblemsand fracture mechanicsproblems have been
discussedin the previoussections.In this section,we presenta novel computationaltool, called
StochasticBoundaryElementMethod(SBEM),for thereliability analysisof a curvilinear fatigue
crackgrowth.
The SBIM basedon the perturbationtechniqueshasbeendevelopedby Ettouney et al.
(1989)andDasgupta(1992)for quantifyingthestatisticmomentsof tractionsanddisplacementsof
a stochasticsystem. A generalmethodology,which combinesthe first orderreliability method
(FORM) with the mixed boundaryintegral equation method (Lua et al. , 1992c), has been
formulatedmostrecentlyby theauthors(Luaet al., 1992d).Thepreformanceandefficiencyof the
developedSBEMhavebeendemonstrated by theproblemof probabilisticfatiguecrackgrowth.
The state-of-the-artof boundary element methods along with various of codes are given in
the Boundary Element Reference Book (Mackerle and Brebbia, 1988). Due to its modelling
efficiency and solution accuracy, BEMs have been used extensively in the field of computational
fracture mechanics (Aliabadi and Rooke, 1991; Cruse, 1988). The application of the BEM to a
curvilinear fatigue crack growth is presented in this section.

The curvilinear fatigue crack path is mainly attributed to the inherent inhomogeneity of the
advance materials such as ceramics, composites or polycrystalline alloys. The existence of a micro-
defect such as a void, a rigid inclusion or a transformation inclusion perturbs the stress field at a
growing crack tip, resulting in a curvilinear crack path. In order to model the singularity at a moving
crack tip, an automatic remeshing in conjunction with the quarter-point singular element (Barsoum,
1976; Henshall and Shaw, 1975 ) has been developed by Saouma (1984) to study the fatigue life of
attachment lugs. A remeshing scheme based on the Arbitrary Lagrangian Eulerian (ALE) together
with enriched finite elements has been developed by Besterfield (1991) in the reliability analysis of a
fatigue crack growth. For problems of multiple fatigue cracks in which elastic interactions of a
fatigue crack with micro-defects are treated, the remeshing scheme will be prohibitively complicated.
The formulation based on the Boundary Integral Equations (BIEs) has several advantages in terms
of solution accuracy and modeling efficiency.

Due to the degeneration of the usual displacement BIE for coplanar crack surfaces, the
traction BIE has to be employed on the crack surface. The traction BIE alone is insufficient to solve
the problem due to the coupling and interaction of the boundary of the component with the growing
crack. Thus, the displacement BIEs have also to be applied. This set of mixed BIEs provide a
unique solution for the boundary value problem. The application of the mixed BIEs to the elastic
interactions of a fatigue crack and a micro-defect can be found in the Reference (Lua et al., 1992c).

By adding a few elements to permit crack extension along the crack growth direction,
remeshing can almost be avoided. Similar to the approach used in enriched finite elements (Gifford
and Hilton, 1987), a special interpolation function which incorporates the stress intensity factors is
employed to model the near tip Crack Opening Displacements (CODs). The mixed BIEs are
presented in Sec.7.2 for a multi-connected region with a fatigue crack. An enriched element which
incorporates the mixed mode stress intensity factors is applied to characterize the singularity at a
moving crack tip. The response gradient, which is key in FORM, is determined in Sec. 7.3 by
direct differentiation. Due to the presence of three random processes in the expression of the
response gradient, namely the mode I and mode II stress intensity factors and the crack direction

27
angle,the first orderresponse-surface
modelis employedto determinethe responsesensitivityof
theserandomprocesses.An iterationschemebasedon the HL-RF method(RackwitzandFiessler,
1978) is employedto find the most probablefailure point (or design point). Due to the high
accuracyof the responsegradientcalculationbasedon thedirectdifferentiation,fastconvergenceis
obtainedin the numerical iteration. The accuracyand efficiency of the presentapproachare
demonstratedin Sec.7.4 through a fatigue crack growth problem with randomnessin the crack
geometry,defectgeometry,fatigueparameters andexternalloads.

7.2 Mixed BIEsfor a Multi-ConnectedRegion

Figure 7-8 showsa finite linear elastic body bounded by outer boundary Fo and inner
boundaries Fi (i = 1, 2 ..... M), containing a finite crack under remote loading t'. A local

Cartesian coordinate system (x',y') with origin at the center of the crack is employed with the y'-
axis normal to the crack surface re. On the displacement boundary Fu, the displacement

components u_ are prescribed; and on the remaining outer boundary I't, the traction components ti
M

are given. The boundary conditions on the inner boundary F1 - _ Fi can be specified based on
i=l
the characteristics of a micro-defect (Lua et al., 1992c).

The usual displacement BIEs (see e.g. Telles, 1983) can be successfully applied on both
inner boundary FI and outer boundary Fo. The resulting BIEs on FIu Fo are

Cij Uj(_) = Uik(_;X) tk(X) dF(x) - tik(_;X) Uk(X) dF(x) - Tin tnk(_ ;X ) AUk(X') dF(x')

.to
.I-o

for _ _ F 1 w F o (7-1)

where the symbol 3("stands for the principal value of the integral in Cauchy's sense, and all the

quantities with the prime (') are defined in the local coordinate system; tk(X) and uk(x) are the

components of traction and displacement, respectively, in the global coordinate system, ui'k(_;x)

and t_k(_;x) represent the displacement and traction, respectively, in the k-th direction at the field
point x corresponding to a unit point force applied at the source point _ in the i-th direction.
Explicit expressions for these free space Green's functions are given in Telles (1983). The
transformation from global to local coordinate system is given by the following transformation
matrix T (or Tin ):

T= I cos00 -sin00] (7-2)

sin00 cos00

The coefficient matrix Cij depends on the smoothness of the boundary; Cij = 21-8ij (for smooth

boundary). The quantity AUm(X') designates the COD on the crack surface Fc in the local
coordinate system defined by

28
 AUm(X')= u+(x') - Um(X') (7-3)

where u_a(x') and u_(x') are the components of the displacement on the upper and lower crack
surface, respectively. The coupling term representing the effect of CODs, which differs from the
usual displacement BIEs, is the 3rd term in Eq. (7-1).

The displacement BIE (7-1) alone is insufficient to solve for all the unknowns, namely, the
unknown displacement and traction on FIu Fo and CODs on F¢. Due to the degeneration of the
displacement BIEs for coplanar crack surfaces, the higher order BIEs based on the traction
boundary data on Fc is employed. Using the prescribed traction boundary condition on Fc, the
resulting traction BIEs become

ti(_) = nj { ui*jk(_;x) tk(X) dF(x) - Vijk(_;x) Uk(X) dF(x) -

.to
.to

- Tin Tjs _tnk(; ;x ) AUk(X') dF(x')} for ; ¢ F_- (7-4)

tW i •

where the symbol (_) stands for the finite-part of a divergent integral, ti are the prescribed traction

components on the upper crack surface F_', and nj=(sin00, -cos00) are the components of the
+ *

normal vector to Fc in the global coordinate system. The free space Green's function Uijk(_;X) and

_i*jk(_; x) are given by Lua et al (1992c).

In order to characterize the crack tip singularity, an enriched element which incorporates
the stress intensity factors (SllZs) is used at the crack tip:

(7-5)
Aul(s) 2(1-v)_ _7_-aKlI
_ , AUz(S) 2(l-v)__ _aKI V2as_s 2

where s is the distance behind the crack tip, a is the semi-crack-length, and KI and KII are mode I
and mode II S_s. In the numerical implementation of the mixed BIEs (Eqs. (7-1) and (7-4)), all

the boundaries, namely FI, Fo and F¢ have to be discretized first. By dividing the boundary
Fi + Fo and the crack surface Fc into NE and NC elements, respectively, the discretized version of
the mixed BIEs can be expressed as

NE NC

cijuj(;)= Z ¢ Z A"L
m
m=l m=l

for _ = x1, X2, ..., XNE (7-6)

[D_ijk(_;Xm) ¢ - S_k(;'Xrn)Uk n] - Z m
Rijk( _ ".,Xm)
' AU
ti(_) = nj(_) {
m=l m=l

for _ = xl , x2 ..... XNC (7-7)

29
wherethecoefficientmatricesG_k, _ , Qrik, m
Dijk, m and
Sij_, R_jk are given by Lua et al (1992c).

7.3 FORM for the Curvilinear Fatigue Crack Growth

As described in Sec. 6.2, for the general correlated and non-normal random variables b,
three steps are required in the first-order reliability analysis (FORM). They are: 1) transformation
of b into the uncorrelated standard normal vector r by Rosenblatt transformation (Rosenblatt,
1952), 2) approximation of the failure surface in the r-space by a flat hyperplane at the most likely
failure point (design point), 3) determination of the reliability index b by computing of the
minimum distance from the origin to the limit state surface. As discussed in Sections 6.3-6.4, the
design point has to be determined by an iterative optimization algorithm, which involves repeated
computation of the limit state function (Eq. (6-1)) and its gradient. In order to ensure rapid
convergence, an accurate determination of the response gradient is required.

Unlike the previous sections (6.3 and 6.4), where the response gradients or sensitivities
have been determined by PFEM, the response gradients are calculated using SBEM. The direct
differentiation coupled with the response-surface method is employed to perform the sensitivity
analysis.

Assuming that the crack geometry (ai, af, x0, Y0,00), fatigue parameters (D, n), external
load (% and defect geometry (Xc, Yc, rc, Pi) (see Fig.8) are modeled by a q-dimensional random
vector b, the performance function for a fatigue problem is given by Eq. (6-38). Since the service
life T s is a deterministic variable, the gradient of the limit state function is given by the response
sensitivity, i.e.,

Og
= _b T(b, K:eq(b)) (7-8)

where T is given by Eq. (6-35). In order to facilitate the response gradient calculation, the line
mapping is applied to map a curvilinear crack path to a local coordinate system, { ( { _ [-1, +1] ).
The mapping function is defined by

a = l[(af - ai)_ +(af+ ai)] (7-9)

Assuming that the crack geometry (ai, at, x0, Y0,00), fatigue parameters (D, n), external load ('r
), and defect geometry (xc, Ye, rc, Pi) (see Fig.8) are modeled by a q-dimensional random vector
b (or r in the transformed space), and using Eqs. (6-37) and (7-9), Eq. (6-35) can be rewritten as

(7-10)

T(b, _c_q(b))
= f(b,_eq(b,_))d_

where the function f in Eq. (7-10) is given by

(7-11)
J(b)
f(b,_eq(b,_) =

The Jacobian J in Eq. (7-11) is defined by

J(b) = ½ (af- ai) (7-12)

3O
Notethat the function neq(b,_) can be described only in an algorithmic form through the mixed
BIEM (Eqs. (7-1) and (7--4)). This is the place where the mixed BIEM interfaces with the FORM
to form the stochastic BIEM. Using Eqs. (7-10), (7-11) and (6-15), the total derivative of the
0T
response (7-) is given by
db

(7-13)
_Kl(b, _) _Kll(b, _) O0(b, _) ]
Ob + +c_,_,, _b + r_,o Ob )] d_
1

where _eq._ , i<eq>:, and _eq.0 are derived from Eq. (6-15) and given by

0 (7-14)
Keq,_ t = COS 3 -_

0 0
(7-15)
_¢q._:u = -3 cos 2 _- sin

= °[c4 (0 o)] (7-16)

Both fb and f,% in Eq. (7-13) can be determined explicitly from Eq. (7-11). The results are

1 0J J OD J ln(_:eq) On (7-17)
fb = n - n - n
D(,:_j _ D_(,__ D(_j _b
n J (Keq) - (n*l) (7-18)
f"_ --" D

Due to both the complicated explicit expressions and implicit functions involved in Eq. (7-
13), numerical integration is required to calculate the response sensitivity (Eq.(7-13)). By dividing
the integration interval [-1, +1] into Npts-1 line elements, which correspond to Npts-1 crack
growth steps, and applying the trapezoidal rule, Eq. (7-13) can be approximated by

_T N_ [f,b + f _:,q( _:_,_:, 0Kr(b, _) 0_ctt(b, _) 30_.__,_). Wn (7-19)

"_-= n=l 0b + _a,_,, 0b + K:eq,0

where _n = -1 + 2(n-1)/(Npts-1), and Wn are the integration weights given by

Wn - 1 (for n=l or n=Npts), Wn - 2 (otherwise) (7-20)

Npts - 1 Npts - 1

Since _Cl,_c_ and 0 are implicit functions of both b and _ in Eq. (7-19), the direct calculation of

3].
 0K:I DK:II _0.
these response sensitivities, namely, Db ' /)b and _-_ is not feasible in this case.
As shown in Eq. (7-19), the key step in the implementation of stochastic BIEM for a
curvilinear fatigue crack growth reliability is to determine three response sensitivities (
D,q D_:u De
_)b ' _b ands-)at each integration point _n (n= 1, 2 ..... Npts). In addition to the implicit

dependence of functions _q(b, 9), _Cn(b, 9) and 0(b, 9) on both b and _, these three functions
represent three random processes due to the presence of the random vector b. Since
_I(_), Kn(_) and 0(_) at a given realization of b can be easily generated by using the mixed BIEM
(see Sec. 7.2), the response-surface approach (Faravelli (1989, 1986)) is employed to determine

the response sensitivity _ Db ' _b and _ at each discretization point _n (n =1, 2 ..... Npts). The
first order response model in b is employed in conjunction with factorial experiments with each
factor at two levels (Myers, 1971). As _I(b, _), _ciI(b, _) and 0(b, _) are independent of the fatigue
parameters, D and n, the dimension of b in _I, _n and 0 is q-2.
With the help of the deterministic solver based on the mixed BIEM, 2q-2 computer
simulations are performed in accordance with predefined factorial simulations. By performing the
least squares fitting process at each crack path discrelization point _,_ (n = 1, 2 ..... Npts), the

D0(b,
¢°)I
history of the response sensitivity of 3b _ _ and -_ j at the L-th iteration, bL
can be determined. Substituting the results
_cI(b, _n) 3K:u(b, _n) _)0(b, _n)
(' Db ' _)b and /)b ; n = 1, 2 ..... Npts) into Eq. (7-19), the sensitivity of the

fatigue life T with respect to the primary random vector b at b E can be determined. For a given
service life Ts, an iterative algorithm to obtain the location of the design point b*, the response
F-,--]
sensitivity at the design point/°tl and the reliability index (or the probability of failure) can be
L bJb"
found in the Reference (Lua et al., 1992d).

7.3 Numerical Results

In order to show the accuracy and efficiency of the stochastic BIEM in a curvilinear fatigue
crack reliability setting, a single edge-cracked plate with a random transformation inclusion is
considered (see Fig. 7-9). The plate geometry (W, L), initial crack location (xo, Y0), initial crack
angle (00), final crack size (af), and material constants (aluminum 7075-T651) are deterministic
parameters given by
L = W = 2.0 in ,x0 = -1.0 in ,Y0 = 0.0 in, 00 = 0.0 (7-21)
af = 0.5 in, _t = 3.866 x 10 6 psi, v = 0.33 (7-22)

where ta is the shear modulus and v is the Poisson's ratio. The crack geometry (ai), external load
('t), fatigue parameters (D, n), the defect geometry (xc, Yc, re), and the internal pressure (Pi)
resulting from the residual strain in the inclusion are assumed to be independent random variables
with specified probability density functions.

The statistical parameters of random input variables (mean, standard deviation and
coefficient of variance (COV)) along with corresponding distribution functions are listed in Table

32
1. As shownin Table4, theinitial cracksizeai hasthe largestdispersion(COV _-60%). For the
initial cracklengthai, a uniform distribution with a tail is employedhere(seeFig. 7-10). The
detection threshold, which is equal to 7.5 x 10 -3 shown in Fig. 7-10, represents the lower limit of
an inspection device to detect the presence of a small crack. Below the detection threshold the
probability density is assumed uniform; above the threshold the probability density decays linearly
to zero, representing false negatives of the inspection technique. For the purpose of verifying the
accuracy of the stochastic BIEM, the Monte Carlo Simulation (MCS) for the sample size
Ns = 2000 is used.

The Cumulative Distribution Function (CDF) of the fatigue life T obtained by the stochastic
BIEM for values of service life Ts are presented in Fig. 7-11. The agreement of MCS and SBIEM
results shown in Fig.7-11 demonstrates the accuracy and efficiency of the stochastic BIEM. As a
rule of thumb (Bjerager, 1989), the sample size necessary for MCS to get a probability estimate
with good confidence is around 100/pf. For small probabilities of failure pf(-- 10 -3 - 10-6), which
are the major interest in reliability engineering, one needs 105 - 108 Monte Carlo simulations to
achieve good confidence. The number of iterations in the stochastic BIEM required to find the
design point b" is only of order 15 to 20 for 13= 3 - 5 (or pf = 0.001 - 0.3 x 10 -6 ). Therefore the
stochastic BIEM based on FORM has an overwhelming advantage over the MCS for small
probabilities of failure in terms of solution accuracy and efficiency.

The reliability index (13)versus the service life (Ts) is shown in Fig. 7-12 along with the
results of no micro-defect. As shown in Fig. 7-12, the presence of a random transformation
inclusion has a detrimental effect on the fatigue life. The comparison of response sensitivities at
Most-Probable-Points (MPPs or design points) versus the probability of failure for both cases is
plotted in Fig. 7-13. As shown in Fig. 7-13, the presence of a random transformation inclusion
changes the response sensitivity of ai significantly. The comparison of the loci of the Most-
Probable-Point (MPP) of crack geometry (ai) is shown in Fig. 7-14. Due to the presence of the
random transformation inclusion, the locus of MPP of ai changes considerably as shown in Fig. 7-
14. When the value of ai increases, the probability of failure pf becomes large (see Fig. 14). This is
the main reason why the routine crack inspection is so important to avoid the large probability of
failure.

8. Conclusions

The Probabilistic Finite Element Method (PFEM) is presented, which is based on the
second-order perturbation. Due to the discrete nature of the finite element formulation, the random
field has to be also discretized. Existing approaches for representation of random fields are outlined.
To the efficient characterization of the random field, the transformation of the original random
variables into a set of uncorrelated random variables is introduced through an eigenvalue
orthogonalization procedure. Both single-field variational principle and three-field Hu-Washizu
variational principle are employed to develop the PFEM for linear, and nonlinear problems,
respectively. The computational aspects in the numerical implementation of the PFEM are also
presented.

The accuracy and efficiency of PFEM in quantifying the statistic moments of a stochastic
system are demonstrated through two examples: 1) a stochastic spring-mass system under sinusoidal
excitation; 2) a cantilever beam subjected to large deflection. The results are in good agreement with
Monte Carlo simulations (MCSs). The computational efficiency of PFEM far exceeds MCS. Since
the PFEM developed essentially involve solution of a set of deterministic problem, it is easily
integrable into any FEM based code.

33
 The PFEMcoupledwith the first-order reliability methodis alsopresentedfor the reliability
analysisof both fracture and fatigue. The methodology consists of calculating the reliability index
via an optimization procedure, which is used to calculate the probability of failure. The PFEM
provides a powerful tool for the sensitivity analysis, which is required in an iterative optimization
algorithm. Performance of the methodology presented is demonstrated on a single edge-cracked
beam with a concentrated load and a classical mode I fatigue crack growth problem.

In addition to the PFEM, the stochastic boundary element method (SBEM), which combines
the mixed boundary integral equation with the first-order reliability method, is also presented for the
curvilinear fatigue crack reliability analysis. Due to the high degree of complexity and nonlinearity
of the response, direct differentiation coupled with the response-surface method is employed to
determine the response gradient. The reliability index and the corresponding probability of failure
are calculated for a fatigue crack growth problem with randomness in the crack geometry, defect
geometry, fatigue parameters and external loads. The response sensitivity of each primary random
variable at the design point is also determined to show its role in the fatigue failure. The results
show that the initial crack length is a critical design parameter. Since crack lengths below the
threshold of an inspection limit are likely to exhibit a large amount of scatter, this makes it imperative
that the life expectancy of a strycture be treated from a stochastic viewpoint.

Probabilistic analysis is becoming increasingly important for the safety and reliability of an
aging structure and for tailoring new advanced materials. Due to the complexity in charactrizing
material behavior, structural response, and failure mechanism, probabilistic mechanics problems are
computationally intensive and strain the resources of currently available computers. Since many
sources of parallelism are inherent in probabilistic mechanics problems, it is evident that the
development of a parallel computing enviroment for probabilistic response analysis in the current
trend in stochastic computational mechanics.

34
 References

1) Akin, J.E.1976. The Generation of Elements with Singularities. Int. ,I. Numer. Methgds
En_.0.._.,10, pp. 1249-1259.

2) Aliabadi, M.H. and Rooke, D.P. 1991. Nomerical Fracture Mechanic_. Computational
Mechanics Publishers, Southampton/Boston.

3) Ang, A.H.S. and Tang, W.H. 1984. prQbability Concepts in Engineering Planning and
Design: Volume II - Decision. Risk. and Reliability. John Wiley & Sons, New York.

4) Arora, J.S. 1989. Introduction to Optimal Design. McGraw-Hill, New York, pp. 122-136.

5) Augusti, G., Baratta, A. and Casciati, F.1984. Probabili_tic M¢_hods in Stru¢t0ral

Engineering. Chapman and Hall, London.

6) Baecher, G. B. and Ingra, T.S. 1981. Stochastic FEM in settlement predictions. Journal of
the Qeotechnical Engineering Division, ASCE, 107 (GT4), pp. 449-63.

7) Barsoum, R.S. 1976. On the Use of Isoparametric Finite Elements in Linear Fracture
Mechanics. Int. J. Numer. Methods Eng.. 10, pp. 25-37.

8) Belytschko, T. et al. 1984. Hourglass Control in Linear and Nonlinear Problems. Comput.
Methods Appl. Mech. Eng. 43, pp. 251-276.

9) Belytschko, T. and Bachrach, W.E. 1986. Simple Quadrilateral with High-Course Mesh
Accuracy. Comput. Methods Appl. Mech. Eng. 54, pp.279-301.

10) Benaroya, H. and Rehak, M. 1988. Finite element methods in probabilistic structural
analysis. Appl. Mech. Rev. 41 (5), 201-213.

11) Besterfield, G.H., Liu, W.K., Lawrence, M.A., and Belytschko, T.B. 1990. Brittle
Fracture Reliability by Probabilistic Finite Elements. ,1. Eng. Mech. ASCE 116 pp. 642-
659.

12) Besterfield, G.H., Liu, W.K., Lawrence, M. and Belytschko, T. 1991. Fatigue Crack
Growth Reliability by Probabilistic Finite Elements. Comput. Methods Appl. Mech. Eng.
86, pp. 297-320.

13) Bjerager, P. 1989. Probability Computation Methods in Structural and Mechanical

Reliability. Computational Mechanic;:s of Probabilistic and Reliability Analysis, Liu, W.K.
and Belytschko, T. eds., Elme Press International, pp. 47-68.

14) Breitung, K. 1984. Asymptotic Approximation for Multinormal Integrals. ,l. Eng. Mech.
ASCE, 110, pp. 357-366.

15) Cambou, B. 1975. Application of first-order uncertainty analysis in the finite element method
in linear elasticity. Proceedings of the 2nd International _Qnference Qn Appti_atiqn of
Statistics and Probability in Soil and Structural Engineering, Aachen, West Germany,
Deutsche Gesellschaft fur Grd-und Grundbau ev, Essen, FRC, pp. 67-87.

16) Chan, T.F., Glowinski, R., Periaux, J., Widlund, O. 1989. Domain decomposition
methods for partial differential equations. SIAM, Philadelphia, PA.

35
17) Converse, A. O. 1970. Qptimizati0n Robert Krieger Publishing Co., New Yprk, pp 243-
248.

18) Cruse, T. A., Wu, Y.-T., Dias, B., and Rajagopal, K. R. 1988a. Probabilistic structural
analysis methods and applications. Comput. Struct. 30, pp. 163-170.

19) Cruse, T.A. 1988b. Bounqtary ElemCnl An_aly_i_ in Computational Fracture Mechanic_.
Kluwer Academic Publishers.

20) Dasgupta, G. 1992. Stochastic finite and boundary element simulations. Proceeding of the
Sixth Specialty Conference, Denver, Colorado, July 8-10, pp. 120-123.

21) Der Kiureghian, A. 1985. Finite Element Methods in Structural Safety Studies. Proc ASCE
Convention, Denver, May.

22) Der Kiureghian, A. and Ke, B.-J. 1985. Finite-element based reliability analysis of frame
structures. Proceedings. Fourth International Conference On Structural Safety and Relial_ility,
Kobe, Japan, I, pp. 395-404.

23) Der Kiureghian, A., Lin, H.- Z., and Hwang, S. J. 1987. Second-Order Reliability
Approximations. ,1. Eng. Mech., ASCE, 113, pp. 1208-1225.

24) Der Kiureghian, A. and Ke, B.-J. 1988a. The Stoizhastic Finite Element Method in Structural
Reliability. Probabilistic Engineering Mechanics 3, pp. 83-91.

25) Der Kiureghian, A. and Liu, P.-L. 1988b. Optimization Algorithms for Structural Reliability.
Computational Probabilistic Mechanics ADM-93, ASME, pp. 185-196.

26) Ditlevsen, O. 1981. l,Jncertain_y Modeling with Applications to Multidimensional Civ.il

Engineering Systems, McGraw-Hill Inc., New York.

27) Erdogan, F. and Sih, G.H. 1963. On the Crack Extension in Plates under Plane Loading
and Transverse Shear. J. Basic Eng., ASME, 85, pp. 519-527.

28) Ettouney, M., Benaroya, H., and Wright, J. 1989. Probabilistic Boundary Element
Methods. (_omputational Mechanic_; of Probabili_tic an.d Reliability Analysis, Liu, W.K. and
Belytschko, T. eds., Elme Press International, pp. 142-165.

29) Fiessler, B., Neumann, H.-J. and Rackwitz, R. 1979. Quadratic Limit States in Structural
Reliability. J. Eng. M_ch. ASCE, 105, pp. 661-676.

30) Fraravelli, L. 1986. A response surface approach for reliability analysis. RILEM Symp on
Stoch. Meth. in Mat. and Struct. Eng..

31) Faravelli, L. 1989. Response surface approach for reliability analysis. J. Eng. Mech. ASME
115, pp. 2763-2781.

32) Gifford, L.N. and Hilton, P.D. 1978. Stress Intensity Factors by Enriched Finite Elements.
Engng Fracture M¢chanic_ 10, pp. 485-496.

33) Grigoriu, M. 1982. Methods for approximate reliability analysis. Structural Safety 1, pp.
155-165.

34) Handa, K. and Andersson, K. 1981. Application of finite element methods in the statistical

36
 analysisof structures.Proceedings of the _rd International Conference on Structural Safety
and Reliability, Elsevier, Amsterdam, pp. 409-417.

35) Hasofer, M. and Lind, N.C. 1974. Exact and Invariant Second-Moment Code Format. J._
Eng. Mech. ASCE, 100, pp.lll-121.

36) Haug, E.J. and Arora, J.S. 1979. Applied Optimal Design: Mechanical and Structural
Systems, ist ed., John Wiley & Sons, New York, pp. 319-328.

37) Henshell, R.D. and Shaw, K.G. 1975. Crack Tip Elements Are Unnecessary. Int. J.
Numer. Methods Eng. 9, pp. 495-507.

38) Hisada, T. and Nakagiri, S. 1981. Stochastic finite element method developed for structural
safety and reliability. Proceedings of the 3rd International Conference on Structural Safety
and Reliability, Elsevier, Amsterdam, pp. 395-408.

39) Hisada, T. and Nakagiri, S. 1985. Role of the Stochastic Finite Element Method in
Structural Safety and Reliability. Proceedings, 4th International Conference on Structural
Safe_ and Reliability, Kobe, Japan, pp. 385-394.

40) Hohenbichler, A.M. and Rackwitz, R. 1981. Non-Normal Dependent Vectors in Structural
Safety. J. Eng. Mech., ASCE, 107, pp. 1227-1238.

41) Komzsik, L. and Rose, T. 1991. Substructuring in MSC/NASTRAN for large scale parallel
applications ..Computer Systems in Engineering. 2 (2/3), pp. 167-173.

42) Lawrence, M.A. 1987. Basis Random Variables in Finite Element Analysis Int. J. Nomer
Method Eng. 24, pp. 1849-1863.

43) Liu, P.-L. and Der Kiureghian, A. 1991. Finite element reliability of geometrically nonlinear
uncertain structures. J. Eng. Mech. ASCE 117 (8), 1806-1825.

44) Liu, W.K., Belytschko, T. and Mani, A. 1986a. Random Field Finite Elements. Int. J.
Numer. Methods Eng. 23, pp. 1831-1845.

45) Liu, W.K., Belytschko, T. and Mani, A. 1986b. Probabilistic Finite Elements for Nonlinear
Structural Dynamics. Comput. Methods Appl. Mech. Eng 56.

46) Liu, W.K., Belytschko, T. and Mani 1987. Applications of Probabilistic Finite Element
Methods in Elastic/Plastic Dynamics. J. Eng. Ind. ASME, 109, pp.2-8.

47) Liu, W.K., Besterfield, G.H. and Belytschko, T. 1988a. Variational Approach to
Probabilistic finite Elements. J. Eng. Mech. ASCE, 114, pp. 2115-2133.

48) Liu, W.K., Besterfield, G.H. and Belytschko, T. 1988b. Transient probabilistic systems.
Compt. Methods Appl. Mech. Eng. 67, pp. 27-54.

49) Liu, W.K., Belytechko, T. and Chen, J.S. 1988c. Nonlinear version of flexurally
superconvergent element. Comput. Methods Appl. Mech. Eng. 71, pp. 241-258.

50) Liu, W.K., mani, A. and Belytschko, T. 1988d. Finite element methods in probabilistic
mechanics. Probabilistic Eng. Mech. 2 (4), pp. 201-213.

51) Liu, W.K. and T. Belytschko 1989. (_omputational Mechanics of Pr0babilistic and

37
 Reliability Analysis Elme Press International.

52) Lua, Y.J, Liu, W.K. and Belytschko, T. 1992a. A Stochastic Damage Model for the
Rupture Prediction of a Multi-Phase Solid: Part I: Parametric Studies. Int. J. Fract7re Mech.
55, pp. 321-340.

53) Lua, Y.J, Liu, W.K. and Belytschko, T. 1992b. A Stochastic Damage Model for the
Rupture Prediction of a Multi-Phase Solid: Part II: Statistical Approach. Int. J. Fracture
Mech. 55, pp. 341-361 .

54) Lua, Y. J., Liu, W.K., and Belytschko, T. 1992c. Elastic interactions of a fatigue crack
with a microdefect by the mixed boundary integral equation method, submitted to Int. J.
Numer. Methods Eng.

55) Lua, Y.J., Liu, W. K., and Belytschko, T. 1992d. Curvilinear fatigue crack reliability
analysis by stochastic boundary element method, submitted to Int. J. Numer. M¢thoO_ Eng.

56) Ma, F. 1986. Approximate Analysis of a Class of Linear Stochastic Systems with Colored
White Noise Parameters. Int. J. Eng. Sci. 24 (1), pp. 19-34.

57) Mackerle, J. and Brebbia, C.A. 1988. The Boundary Element Reference Book
Computational Mechanics Publications, Southampton, Boston.

58) Madsen, H.O., Krenk, S. and Lind, N.C. 1986. Methods of Structural Safety, Prentice-
Hall, Inc., Englewood Cliffs, N.J.

59) Melchers, R.E. 1987. Structural Reliability, Analysis and Predictions, Ellis Horwood Series
in Civil Engineering, Halsted Press, England.

60) Myers, R.H. 1971. Response Surface Methodology. Allyn and Bacon Inc., Boston.

61) Noor, A. K. 1991. Bibligraphy of books and monographs on finite element technology.
Appl Mech Rev 44 (6), pp. 307-317.

62) Paris, P.C. and Erdogan, F. 1963. A Critical Analysis of Crack Propagation Laws. J. Basic
ASME, 85, pp. 528-534.

63) Rackwitz, R. and Fiessler, B. 1978. Structural Reliability under Combined Load Sequences.
Gomput. Struet. 9, pp. 489-494.

64) Rice, J.R. 1968. A Path Independent Integral and the Approximate Analysis of Strain
Concentrations by Notches and Cracks. J. Appl. Mech. ASME, 35, pp. 379-386.

65) Rosenblatt, M. 1952. Remarks on a multivariate transformation. Annals of Mathematical

Statistics. 23, pp. 470-472.

66) Saouma, V.E. 1984. An automated finite element procedure for fatigue crack propagation
analysis. Engng Fracture Mech. 20, pp. 321-333.

67) Shinozuka, M. 1987. Stochastic fields and their digital simulation. Stochastic Methods in
structural Dynamics, Schuellernd, C.I. and Shinozuka, M. eds., Martinus Nijhoff
Publishers, Boston, pp. 92-133.

68) Shinozuka, M. and Deodatis, G. 1988a. Response variability of stochastic finite element

38
 systems. J. Engng. Mech., ASCE, 114 (3), pp. 499-519.

69) Shinozuka, M and Yamazaki, F. 1988b. Stochastic finite element analysis: an introduction.
in Stochastic Structural Dynamics. Pr0_ess in Theory. and Applications, Ariaratnm., S.T.,
Schueller, G.I. and Elishakoff, I. eds., Elsevier Applied Science, pp. 241-291.

70) Sih, G.C. 1974. Strain Energy Density Factor Applied to Mixed Mode Crack Problems. Int_.__
J.. Fracture Mech., 10, pp.305-322.

71) Spanos, P.D. and Ghanem, R. 1988. Stochastic Finite Element Expan_sion for Random
_. Report NCEER-88-0005.

72) Sues, R.H., Chen, H.-C., Twisdale, L.A. I991. Probabilistic Structural Mechanics
Research for Parallel Processing Computers NASA CR-187162.

73) Sues, R.H., Chen, H.-C., Chamis, C.C., Murthy, P.L.N. 1992a. Programming
probabilistic structural analysis for parallel processing computers. AIAA Journal, scheduled
for publication.

74) Sues, R.H. Lua, Y.J., and Smith, M.D. 1992b. Parallel computing for probal_ilistic
response analysis of high temperature composites. NASA CR-26576.

75) Sues, R.H. and Chen, H.-C. 1992c. The stochastic preconditional conjugate gradient
method. Probabilistic Engineering Mechanics, to be published.

76) Telles, J.C.F. 1983. The Boundary_ Element Method Applied to Inelastic Problems, Lecture
Notes in Engineering, 1, Springer-Verlag.

77) Tong, P., Pian, T.H.H. and Lasry, S.J. 1973. A Hybrid-Element Approach to Crack
Problems in Plane Elasticity. Int. J. Numer. Methods En_. 7, pp. 297-308.

78) Tvedt, L. 1983. Two Second Order Approximations to the Failure Probability. Det Norske
.Veritas (Norway) RDIV/20-004-83.

79) Twisdale, L.A. and Dunn, W.L. 1983. Probabilistic analysis of tornado wind risks. J_
Struct. Engng 109 (2).

80) Twisdale, L.A. and Vickery, P.J. 1991. Research on thunderstorm wind design parameters.
submitted for publication in the JOurnal of wind Engineering and Industrial Aerodynamics.

81) Vanmarcke, E. and Grigofiu, M. 1983. Stochastic Finite Element Analysis of Simple Beams.
J. Eng. Mech. 109, pp. 1203-1214.

82) Vanmarcke, E. 1984. Random Fields. Analysis and Synthesis, MIT Press, Second Printing.

83) Wu, X.F. and Li, X.M. 1989. Analysis and Modification of Fracture Criterion for Mixed-
Mode Crack. Engng Fracture Mech. 34, pp. 55-64.

84) Wu, Y.-T., Millwater, H. R. and Cruse, T. A. 1990. Advanced probabilistic structural
analysis method for implicit performance functions. AIAA Journal 28, pp.1663-1669.

85) Yamazaki, F.M., Shinozuka, M., and Dasgupta, G. 1988. Neumann expansion for
stochastic finite element analysis. Journal of Engineering Mechanics ASCE, 114 (8), pp.
1335-1354.

39
86) Zhang, Y. and Der Kiureghian, A. 1991. Dynamic Response Sensitivity of Inelastic
$trvctures. Technical Report UCB/SEMM-91/06, Department of Civil Engineering,
University of California at Berkeley.

4O
Table 7-1. ProblemConstants:SingleEdge-CrackedBeamwith anApplied Load

Parameter Mean Standard Deviation Percent

LengthrE) 10.0 in 0.0 0.0

Width 0,V) 5.0 in 0.0 0.0

Thickness (0 1.0 in 0.0 0.0

Young's Modulus ('E) 30,000.0 ksi 3,000.0 10.0

Poisson's Ratio (v) 0.30.0 0.0

Applied Load (P) 10.0 kip 1.0 kip 10.0

Crack Length (a) 0.1 in 0.01 in 10.0

Stress Intensity Factor !r.i)

33.452 ksi 4_ o.o ksi4E 0.0

Fracnn'e Toughness (_:ic)

43.0 ksi ,/_ o.0 _i 4_ 0.0

43.
Table7-2. NumericalPerformancein Brittle FractureReliabilityAr_,_]ysis

Randomness Variance Starting Number Failure Reliability Probability

in (kip2/in 3) Point Iterations Point Index of Failure

ii

]:orc¢ 12.3248 P=12.5 kip 5 1:'=-12.2

kip 2.173 1.49 %

9.94 %

Young's 10.003087 E-30e3 ksi 0 E---30e3ksi 0.0 %

Modulus 0.18 %
i t

Crack Length 1.8273 a=l.4 in 7. a= 1.29 in 2.911 0.1801%

3.83 %
IHli

Combined 12.5107 P=I2 kip 9 P=I2.1 ki_ 2.079 1.88 %

10.01% E=30e3 ksi E=30¢3 ks5

a-l.1 in a=l.02 in

42
 Table7-3. ProblemConstants:SingleEdge-CrackedPlatewith aDistributedLoad

Parameter Mean Standard Deviation Percent

Length rE) I0.0 in 0.0 0.0

Width (W) 4.0 in 0.0 0.0
Thickness (0 1.0 in 0.0 0.0
Young's Modulus (E) 30,000.0 ksi 0.0 0.0
Poisson's Ratio (v) 0.3 0.0 0.0
Applied Stress (x) 12.0 ksi 3.0 ksi 25.0
Initial Crack Length (al) 0.01 in 0.01 in 100.0

0.1in 0.01 in 10.0

Final crack Length (af)
1
Fatigue Parameter (D) 1.0 x 10 10 3.0 x 10 "1 30.0
Fatigue Parameter (n) 3.25 0.08 2.5

43
 Table 7-4. The Statistical Parameters and Distributions of Input Random Variables of
the Example Problem

Random Parameters Mean Standard deviation COV

ai CUniform with tail) 0.5833 x 10 2 in 0.3584 x 10 "2 in 61.4%

D (Log-normal) 0.3770 × 10 -9 0.1885 x I0 -10 5.0%

n (Log-normal) 3.60 0.18 5.0%

x (Normal) 11.0 ksi 1.1 ksi 10.0%

x_ (Uniform) -0.25 in 0.14433 57.7%

Ye (Uniform) -0.4 in 0.05774 14.4%

rc (Uniform) 0.1375 in 0.03608 26.2%

Pi (Normal) 35.0 ksi 3.5 10.0%

44
 kl m1 k2 m2

F(t)

X2

F(t)=25.0 x 10' Sin(2OOOt)

m1=0.372 .k1=24.0 x 10s

m2=0.248 k2=12.0 x 106

Figure 7-1

45
 DISPLACEMENT (NODE I)
2.000
.......PFEM HGQ ......... MCS

-2.00G ; ; ! ! ! ! ! i ,
.000 .12.5 .250 .375 .500 .625 .750 .87_ 1.000 1.125 1.250

TIME (SEC) (x 10 -2)

Figure 7-2

46
 DJSPLACENIENT(NODE I)
.400.
....... PFEM -- HGQ ......... UCS

._mo'

Z
m

j_ i_.
o.

I__" r i! a i

L_ [l

.100 -

A1 (
i! lj i_ !i
i !i
r_l,

!i
I _

ii i!
:_i,
!iV'

ii
, A
ri

i ! I

.000 ! I I I I
.000 .250 .37S .500 .1125 .750 .87_ 1.000 1.12_ 1.250

TIME (SEC) (x10 -2)

Figure 7-3

47
 DISPLACEMENT BOUNDS AT NODE I(PFEM)

MEAN ..............
UPPER ......... LOWER

..-"... _...
1._W)O•
z - - ,,,.\
/"_. .,

___
d.
.O(X}-
!, /: i : /i / J
,
--°'°%%

X : " ."" : " .'"'".. /

/
.o

,
_ ," ".. : : '.. : -.. /
.o ; • : ; _. •
,'. : .; ;o-"
t o._ # t •
j _ 8 # I m

. ;;
16
,. ,;
,,, '."

-3.000 I I I I I ! I I 1
.1211 .250 .3711 .500 .625 .7_i0 .878 1.000 1.125 1.2.E)O
_.ME(sEc) (xlo -2)

Figure 7-4

48
 Reg I Reg I Reg _1

L ' ' J'l

| Reg I Reg I Enr | -']_-

I ' Reg I Peg I Enl" , a 1

10

Figure 7-5

49
 I
I
R R R R R
I
I
L
I
m

R
4_a R R R R!
_,R' _ _l R

-RTR-I
_-,
_'l R
-4--4-4--I
R

j
L
W
v

Figure 7-6

sO
 5
\
Initial Crack Length
N Fanal Crack Length
4 Fatigue Parameter D
om

om
d_
Fatigue Pararmter n
o_ 3 & Stress
_J Combination

0 ! ! |

0 2000000 4000000 6000000 80O00O0

Service Life (cycles)

Figure 7-7a

51
7

6
__ _. __'ual Crack Length

. FinalCrack Length
= Fatigue Pararr_er D
= Fatigue Parameter n
A Stress
Combination

0 ! ! !

0 2OOOOOO 4OOOOOO 8OOOOOO

Service Life (cycles)

Figure 7-7b

52
 I t_' (on F,)

0 = x

Figure 7-8

53
 _ X

YC
(xo, yo)

l_
I-
w q

Figure 7-9

54
 f_

88.889

{ \o.o15 v

7.5 x 10 .3
ai (in)

Figure 7-10

55
 1.2

[.., 1.0

.< 0.8
o
0.6

[.., 0.4
< Stochastic BIEM
MCS (Sample Size=2000)
0.2

¢,.)
0.0
0 I 2 3 4
SERVICE LIFE (*E+06)

Figure 7-11

56
 4
2:

W!thout a Random Inc!usion

3

0 • ' " I " " " | " ' =

0.00 0.25 0.50 0.75 1.00 1.25

SERVICE LIFE (*E+06)

Figure 7-12

57
•_
30
2.S t

2.0
_---.-o-- Without
With a RandomInclusion
a Random Inclusion I
1.5

_" 1.0

Z 0.5
r_

0.0
0.0 0.2 0.4 0.6 0.8 1.0

PROBABILITY OF FAILURE

Figure 7-13

58
Z
0 -----w--- Without a Random Inclusion
¢I,

Z 8O With a Random Inclusion

6O

<

40

2O

0
0 0.2 0.4 0.6 0.8
0.0

PROBABILITY OF FAILURE

Figure 7-14

59
 Form Approved
REPORT DOCUMENTATION PAGE OMB NO. 0704-0188

Public reporting burden for this collection of inform_ion is estimated to average I hour per response, including the time for revmwing instructions, searching existing data sources.
gathering and maintaining the Oate needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of thn;
collection of information, including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson
Davis Highway, Suite 1204, Arlington, VA 22202-4302° and to the Office of Management and Budget, Paperwork Reduction Project (0704-0188). Washington. DC 20503.

1. AGENCY USE ONLY (Leave blank) 2. REPORT DATE 3. REPORT TYPE AND DATES COVERED

April 1993 Final Contractor Report

4. TITLE AND SUBTITLE 5. FUNDING NUMBERS

Probabilistic Finite Elements for Fatigue and Fracture Analysis

WU-584--03-11
G-NAG3-822
6. AUTHOR(S)

Ted Belytschko and Wing Kam Liu

7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) a. PERFORMING ORGANIZATION

REPORT NUMBER

Northwestern University
Department of Mechanical Engineering
E-7711
Evanston, Illinois 60208

9. SPONSORING/MONITORING AGENCY NAMES(S) AND ADDRESS(ES) 10. SPONSORING/MONITORING

AGENCY REPORT NUMBER

National Aeronautics and Space Administration

Lewis Research Center
NASA CR-191098
Cleveland, Ohio 44135-3191

11. SUPPLEMENTARY NOTES

Project Manager, C. Chamis, Structures Division, (216) 433-3252.

12_ DISTRIBUTION/AVAILABIMTY STATEMENT 12b. DISTRIBUTION CODE

Unclassified - Unlimited
Subject Category 39

13. ABSTRACT (Maximum 200 words)

This report presents an overview of the probabilistic finite element method (PFEM) developed by the authors and their
colleagues in recent years. The primary focus is placed on the development of PFEM for both structural mechanics
problems and fracture mechanics problems. The perturbation techniques are used as major tools for the analytical
derivation. The report specifically covers the following: (1) representation and discretization of random fields;
(2) development of PFEM for the general linear transient problem and nonlinear elasticity using Hu-Washizu variational
principle; (3) computational aspects; (4) discussions of the application of PFEM to the reliability analysis of both brittle
fracture and fatigue; and (5) a novel stochastic computational tool based on stochastic boundary element (SBEM).
Results are obtained for the reliability index and corresponding probability of failure for: (1) fatigue crack growth;
(2) defect geometry; (3) fatigue parameters and; (4) applied loads. These results show that initial defect is a critical
parameter.

14. SUBJECT TERMS 15. NUMBER OF PAGES

6O
Defect size; Defect statistics; Defect growth; Stochastic process; Boundary elements;
16. PRICE CODE
Random variables; Fracture mechanics; Variational principles; Reliability brittle fracture A04
17. SECURITY CLASSIFICATION 18. SECURITY CLASSIFICATION 19. SECURITY CLASSIFICATION 20. UMITATION OF ABSTRACT
OF REPORT OF THIS PAGE OF ABSTRACT

Unclassified Unclassified Unclassified

NSN 7540-01-280-5500 Standard Form 298 (Rev. 2-89)

Prescribed by ANSI Std. Z39-18
298 - 102