American Institute of Aeronautics and Astronautics
1
State Transition Approach to Reliability
Based Design of Composite Structures
Varun Sakalkar* and Prabhat Hajela
†
Rensselaer Polytechnic Institute, Troy, NY, 12180
The paper presents a new approach to the optimal design of composite structural systems in
the presence of uncertainties. A progressive damage propagation model for composites,
based on transformation field analysis (TFA), is used to develop a design for reliability that
includes multiple failure modes typical in laminated composite structural systems. This
analysis involves multiscale computations, with complex interactions between response
parameters at different length scales. Not all failure modes are equally catastrophic and may
only result in degraded structural performance. A design problem formulation based on a
state transition approach is introduced, and allows for the handling of multiple failure
modes in a rational manner. This methodology termed as a system effectiveness approach,
models designer preference as to acceptability of degraded performance, and is used to
develop optimal designs. A comparison of these designs against those obtained from a more
widely used competing risk methodology provides insight into the advantages of the new
approach.
I. Introduction
Analytical and computational modeling is at the very core of designing high performance composite structural
systems for aerospace applications. The behavioral response available from such models is used to provide the ‘what
if’ input to the design process. There is concern that inherent modeling errors, if not quantified appropriately, can
introduce significant deviations in the behavior of the physical system from predicted values. Composite structures,
with distinctly different failure modes and patterns from homologous material systems, are particularly vulnerable to
this deficiency. Failure in composite structures is typically obtained from mechanistic models of damage that span
length scales ranging from the material constituent level to component or macro structural level. Each level of
modeling has uncertainties which influence the behavioral response used to guide design decisions. The proposed
paper will focus on a novel state transition [1] based design tool that is especially applicable to developing damage
tolerant composite structures, with an explicit focus on risk related to operating degraded (by progressive damage)
structural systems. Model development and analysis for the composite systems is based on a transformation field
analysis (TFA) framework.
Methods to characterize progressive damage in a composite have been studied extensively in literature. A number of
models for fiber debonding, matrix cracking, delaminations at crack intersections and free edges, fiber breaks, and
final ply failure have been developed. Models for interfacial debonding and sliding were developed initially by
solving a solitary inclusion problem with appropriate interfacial discontinuities [2-7]. A shear lag model for
debonding with sliding resistance was developed by Hutchinson and Jensen [8]. More recently, Dvorak and
co-workers [9-11] have used the transformation field analysis with equivalent eigenstrain to model the local fields
due to interfacial de-cohesion. For both micromechanical and intermediate scale models, however, it is imperative to
include a non-deterministic treatment so as to model the uncertainty associated with their predictive capability. At a
fundamental level, the relationship between damage/failure modes and the microstructure of composite should
define the manner in which these structures should be designed to resist the long-term influence of mechanical,
hygrothermal and chemical loading. This can be facilitated by adoption of suitable yield/failure criteria based on
measurable mechanical and strength properties of the constituent phases and their interfaces [12-20]. Given
significant variations in these properties, it would be necessary to include stochastic aspects of specific damage
mechanisms, such as Weibull distributions for interface and ply strengths. Some preliminary efforts in this regard
*Corresponding author, PhD Candidate, Department of Mechanical, Aerospace, and Nuclear Engineering, AIAA
Student Member. †Professor, Department of Mechanical, Aerospace, and Nuclear Engineering, AIAA Fellow.
50th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference <br>17th4 - 7 May 2009, Palm Springs, California
AIAA 2009-2179
Copyright © 2009 by the American Institute of Aeronautics and Astronautics, Inc. All rights reserved.
American Institute of Aeronautics and Astronautics
2
have been implemented in modeling of composites with progressive reinforcement debonding [9,11,21]. However,
most widely used analytical or numerical mechanistic models do not explicitly model uncertainty in the prediction
of damage onset or progression. Several recent publications have focused on using probabilistic design approaches
for the design of composite structures. Chamis and Abumeri [22] present a Monte Carlo approach along with FORM
for the dynamic buckling analysis of composite shell structures. Liu and Mahadevan [23] use a Monte Carlo
approach for the progressive fatigue damage analysis of composite laminates. Abumeri and Chamis [24] use a
double-loop strategy (probabilistic constraints within an optimization problem) for the design of composite shell
structures, while Soliman and Kapania [25] use FORM for reliability assessment of first ply failure analysis of
composite cylinders. Arbocz and Hilburger [26] present a probabilistic analysis for the buckling of composite shells
using the Monte Carlo and the first-order second-moment methods. Current work attempts to provide system level
reliability estimation by utilizing the progressive failure characteristics of the micro TFA model.
A critical step towards estimating system reliability is the identification of all possible modes of failure that
contribute significantly to the overall failure probability. For simple systems, or systems that are statically
determinate, it may be possible to identify these limit states by inspection. However, for a complex structure with
high redundancy, this step may prove to be the more difficult [27]. A failure mode approach can be used for this
purpose [28] which identifies all possible failure sequences – this process that may be impractical for large-scale
indeterminate structures [29,30]. However, it is unlikely that all failure modes contribute equally to the overall
system failure. Therefore, much of the recent work focuses on efficient methods to identify only the significant
failure modes of a complex structure. Mahadevan and Raghothamachar [31] present an approach based on a
combination of Monte Carlo sampling and analytical methods to identify significant failure modes of a complex
structure. Mehr and Tumer [32] present a detailed approach to handling reliability issues for large-scale aerospace
systems at NASA. They assert that most generic system reliability techniques, such as the failure mode approach,
may not be adequate for large-scale aerospace systems, where multiple engineering teams work concurrently on
different aspects of the design. Instead, a dedicated approach that caters to their specific requirements may need to
be developed. Vittal [1] describes a state transition approach to evaluate system reliability, where probabilities of a
system transitioning from one failure to another are considered. This approach has been adapted for the design of
laminated composite structures in the current paper.
Computation of the system level reliability involves characterization of uncertainty in damage prediction models as
the first step towards in developing effective tools for designing damage tolerant composite structures. In this
context, probabilistic approaches have shown significant promise as design tools. These approaches are based on
combining nondeterministic analysis methods with design optimization algorithms. Central to this methodology are
methods by which to compute the reliability indices that are used as performance metrics in the design optimization.
Probabilistic design algorithms such as FORM/SORM have been proposed and are widely used in design practice
[28]. The numerical efficiency of these approaches is directly dependent upon the methods used in the deterministic
analysis of composites for damage initiation and progression. The following section summarizes the transformation
field analysis framework that has been shown to be effective in this context, and can be adapted to include
uncertainties in geometric and material properties of the composite structure.
American Institute of Aeronautics and Astronautics
3
Applied Laminate
Stress
Stiffness matrix +
Coordinate
transformation matrix
Ply Stresses
Mori-Tanaka Averaging
+
Concentration Factors
Phase Stresses
Failed
Phases
Eigen Stresses
Corresponding to
Failure
Transformation Matrix
+
Influence Functions
UPDATED Phase
stresses
Laminate Strains
Zero Eigen
Stresses
Levin’s Equation
+
Influence Functions
Ply Eigen Stresses
Compliance Matrix +
Coordinate
Transformation Matrix
Eigen Strains
+
UPDATED
Laminate Strains
-Phase material properties
- Laminate geometry
- Volume fraction
- Phase material properties
- Volume fraction
- Failure Modes
-Ultimate Strength Values
- Phase material properties
- Volume fraction
- Phase material properties
- Laminate Geometry
- Volume fraction
1
2
3
4
5
NO
YES
Figure 1. Micro-TFA model
II. Progressive Failure of Laminated Composites
This section summarizes the main features of the micro-TFA model. Fig. 1 shows the flow of computations in
the transformation field analysis based analysis of composite structures. In this figure, the entities in the oval are
intermediate response quantities. The entities in the rectangular boxes are coupling variables and are numbered 1
through 5. Each coupling variable transforms one input response to the corresponding output response. For example,
the stiffness matrix and co-ordinate transformation matrix listed in box 1 transform applied laminate stress into ply
stress. The curly braces indicate the quantities needed to derive each coupling variable. The figure clearly
demonstrates the coupled nature of the analysis.
The fibrous laminate under consideration consists
of n fully bonded thin elastic plies, Fig. 2. The ply
thickness is denoted as ih , i = 1, n, such that ih h=∑ ,
the total thickness of the laminate. Hence, /i ic h h= is
the volume fraction of ply i, such that 1ic =∑ . Two
coordinate systems are defined as shown in Fig. 2, one
is overall ( jx , 1,2,3j = ), and one is local
( kx , 1,2,3k = ). Fiber orientation of the i-th ply is
given by angle iϕ between the local 1x and the overall
1x axes. In the present work, only laminates consisting
of identical plies and symmetric layups are considered.
Membrane mechanical loads confined to the mid-plane
1 2x x of the laminate are considered. For symmetric
laminates, they cause uniform stresses 11σ , 22σ and
12σ as shown in Fig. 2.
Properties of the unidirectionally reinforced plies are given in terms of elastic moduli of the fiber (f) and matrix (m)
materials, and their volume fractions, ,f mv v where 1f mv v+ = . The matrix is assumed to be isotropic with Young’s
modulus mE and Poisson’s ratio mν . The fiber is transversely isotropic with f
LE and f
TE as the longitudinal and
transverse Young’s moduli, fLν and
fTν the associated Poisson’s ratios, and
fLG the longitudinal shear modulus.
Consequently, the unidirectional plies are transversely isotropic with overall elastic properties analogous to those of
the fiber ( LE , TE , Lν , Tν , LG ). The overall properties of each ply can be either measured or computed from the
local properties using a micromechanical model [33].
Figure 2. Geometry and loading of a fibrous laminate.
American Institute of Aeronautics and Astronautics
4
Assuming that loading includes both membrane forces, [ , , ]1 2 12=NNNN N N N , and bending moments,
[ , , ]1 2 12=MMMM M M M , the classical laminate analysis (CLA) approach allows for the following load-response relations
for an undamaged laminate.
oε ′ ′ ′= + +A B fN MN MN MN M κ ′ ′ ′= + +C D gN MN MN MN M (1)
Where,
2 2 3 3
1 1 1
1 1 1
1 1( ) , ( ) , ( )
2 3
n n n
i i i i i i i i i
i i i
h h h h h h− − −
= = =
= − = − = −∑ ∑ ∑A L B L D L (2a)
2 2
1 1
1 1
1( ) , ( )
2
n n
i i i i i i
i i
h h h hλ λ− −
= =
= − = −∑ ∑f g (2b)
1( ) −′ ′= −A I B B A , 1−′ ′= −B A BD , 1−′ ′= −C D BA , 1 1[ ]− −′ = −D D BA B (3a)
′ ′ ′= − −f B g A f , ′ ′ ′= − −g C f D g (3b)
and the arrays ′f and ′g are overall eigenstrains and curvature. The eigenstrains may be caused by a local event in
any of the constituent materials, including damage or material degradation. The transformation field analysis
approach centers on finding an auxiliary transformation stress field in the plies and/or their fiber and matrix phases
such that the magnitude of the corresponding net stress components, which violate the underlying failure criteria, are
zero. The fiber and matrix stresses for an undamaged laminate subjected to a known overall load are first
determined. Next, the relevant failure criteria are examined for all plies, and the violating stress components are
identified. These stress components are then set to zero and the eigenstresses necessary to satisfy equilibrium are
then computed. These newly computed eigenstresses alter the stress fields in the undamaged constituents and these
have to be re-examined for additional failure. The process is repeated until no further damage is found under the
applied overall loading before incrementing the load to the next level. This coupled multiscale analysis is depicted in
Figure 3.
Constituent
level (micro)
- Volume fraction
- Microstructure assumptions
- Materia l strength values
Lamina level
(macro)
- Ply thickness
- Ply orientation ply
stress
Laminate
load
constituent
stress
Damage criterion +
TFA system of
equationsphase
eigenstress
Laminate level
analysis
Eigenstress calculations
ply
eigenstress
laminate
stress
laminate
eigenstrain
Σ
Total laminate
response Repeat until no further damage at
current load-level
Figure 3. Multiscale TFA data flow
In summary, there are computations at two distinct length scales. At the macro ply and laminate level, the relevant
equations are as follows.
American Institute of Aeronautics and Astronautics
5
1
1
1
1
1
, ,
,
{ ) { )
N
i i
i
N
i i i i i i i i
i
N
i i i ij j
j
L L h
L L
P N Q M U
σ ε λ ε σ µ λ λ
σ ε λ ε σ µ
σ λ
−
=
−
=
=
= + = + =
= + = +
= + +
∑
∑
∑
(4)
Here σ and ε are the total laminate stress and strain, respectively; L is the overall stiffness matrix; λ and µ are
the laminate eigenstresses and eigenstrains and subscripted quantities refer to these responses in the ith
ply. At the
material or phase level, a corresponding set of constitutive relations can be written as follows.
1
.
,
, r r r r r r r r
T Ti f f f m m m
r r i r s s
s f m
L L
c A c A
B F
σ ε λ ε σ µ
λ λ λ
σ σ λ
−
=
= + = +
= +
= + ∑ (5)
Here rσ and rε are the local stress and strain fields respectively; rL is the local stiffness matrix; rλ and rµ are the
local eigenstresses and eigenstrains, all for phase “r”; rA and rB represent strain and stress concentration factors for
the fiber and matrix; rsF are transformation influence functions which depend on mutual constraints of the phases
and their elastic moduli. It is clear from the two sets of equations that to complete the ply and laminate stress
computations, the eigenstresses are required. These eigenstresses in turn must be completed by the microscale
computation that takes as an input the ply stresses computed at the macroscale. This coupled analysis is both
complex and computationally cumbersome, requiring the computation of stress distribution functions, stress
concentration factors, and stress and strain transformation influence functions; these quantities are obtained from
geometrical considerations and from the phase material properties. A more complete description of the TFA
approach is available in [34].
A representative stress strain response obtained from a multiscale TFA analysis is shown Fig. 4. It shows the
comparison of the results predicted by the micro-TFA model against available experimental data for a carbon-epoxy
[0, +/-45, 90]s laminate. Fig. 4 shows stress-strain characteristics ε22 vs. σ22, when the laminate is subjected to a
biaxial loading σ22:σ11 = 2:1, the solid blue line represents the results obtained from the micro-TFA model, the black
stars represent the experimental data from Ref. [35,36], and the dotted red line represents the results obtained from
the ply discount model, which has been widely used in the literature [34]. Note that the ply discount model considers
a single ply as the basic element instead of material phases. Additionally, when a failure occurs in a ply, its stiffness
is not taken into account for further analysis.
A key feature of the Micro-TFA damage model developed as part of this research is its ability to reveal
various stages of laminate damage as the load is increased (stress control method). In Fig. 4, the letters m and f
denote matrix and fiber failure, for a particular ply, respectively. Subscripts 11, 22, and 12 denote the mode of
failure in the normal, transverse, and shear directions, respectively. From Fig. 4, we can also observe that the micro-
TFA model predicts results that closely match the experimental data. The first and second failure strains are denoted
by 1 2,
f fε ε respectively.
American Institute of Aeronautics and Astronautics
6
Ply discount
1fε 2fε
Figure 4. Stress strain response for a [0, +/-45, 90]s laminate with progressive damage
2.1 Response Approximation using Radial Basis Neural Network
The computational approach described in the previous section is integrated into a multi-parameter reliability
analysis of a composite structural system. This approach is computationally demanding, and must be repeated
several times to evaluate the performance functions required for reliability analysis. Further integration of the
approach into an iterative design environment would be computationally prohibitive, and surrogate models of
analysis provide an alternative approach. The present work adopts a surrogate model based on a Radial Basis
Function (RBF) neural network, and this approximate model is used in lieu of the actual micro TFA evaluation.
Details of the RBF model functions are not included in this paper but are available in Ref. [37-39].
Building a RBF model requires that training sets or samples be first created. Several exact computations of the
function ( )g X were performed so as to cover the range of values of x that span the design space. The number of
sampling points required to accurately model the performance function is dependent on the number of random
variables, the nonlinearity of the problem considered and the assigned computational accuracy. Guidelines on the
“design” of sampling points can be found in various statistical textbooks and neural network monographs [40-42].
These sampling data points were then used as the training and testing data in the RBF training so as to
approximately represent the performance function ( )g X . A sample set of results showing the training and validation
mean squared error for the approximation of the first and second failure strain against exact TFA computations is
shown in Fig. 5. The micro TFA model was setup for a [0, +/-φ , 90] s laminate with φ and fiber volume fraction
( rc ) as the independent variables. The radial basis network modeled the TFA analysis and yielded a surrogate
approximation between input variable φ and rc , and the output quantities1 2,
f fε ε . The fiber and matrix material
properties assumed for the TFA calculations are shown in Tables 1a and 1b. A total number of 225 training points
were selected and 40 validation points were used in testing the approximate model. A small training and validation
error at convergence (given in Fig. 5) shows that the simulated RBF network output compares well with the actual
TFA solution.
American Institute of Aeronautics and Astronautics
7
Table 1a Mechanical properties of fiber [34]
Material LE (GPa) TE (GPa) LG (GPa) TG (GPa) Lν uσ (GPa) fρ (kg/m
3)
AS4 Carbon 26.2 262.6 15 7 0.2 3.35 24
Table 1b Mechanical properties of matrix [34]
Material E (GPa) ν uTσ (MPa) uCσ
(MPa) uτ (MPa) Lη mρ (kg/m
3)
3501-6 epoxy 4.2 0.34 69
250 50 0.3
12
0.0
0.2
0.4
0.6
0.8
1.0
1.2
0 50 100 150 200 250
No
rma
lize
d M
ean
sq
ua
re e
rro
r (%
)
Epoch
Epsilon_f1 (Strain at first failure)
Epsilon_f2 (Strain at second failure)
0.0
0.2
0.4
0.6
0.8
1.0
1.2
0 5 10 15 20 25
Norm
alize
d M
ean
Sq
ua
re E
rro
r
(%)
Epoch
Epsilon_f1 (Strain at First Failure)
Epsilon_f2 (Strain at Second Failure)
Figure 5a. Training Error of the Radial Basis network Figure 5b. Validation Error of the Radial Basis
network
III. Reliability based design optimization
The design optimization problem involving nondeterministic design variables and constraints requires a
modified problem formulation. A system objective J is to be minimized as shown in Eq. (6), and is a function of the
design variables x that may be nondeterministic in nature. In such cases, the statistical characteristics of the function
J would be considered as the design objectives. The constraint vector g is also a function of these design variables,
and in contrast to a deterministic constraint, the optimization problem formulation considers bounds on probability
of constraint violation. In the following problem statement, fP is the probability of failure and is bounded by an
upper limitallowfP . Other measures of reliability can also be used in lieu of the probability of failure.
min ( )
. . ( ( ) 0)
x
allowf f
J x
s t P g x P≤ ≤ (6)
The computation of the probability of failure and other estimates of reliability has been researched extensively.
A widely used approach is the first-order reliability method that is both efficient and effective. A first-order second-
moment approximation of nonlinear limit states using normal variables (or equivalent normal variates of non-normal
variables) is typically required in these approaches. The failure probability is estimated by formulating and solving
an optimization problem. In this approach, all the variables are transformed to a standard normal space as
' ( ) /x xx x µ σ= − (7)
where x are the original normal variables and 'x are the standard normal variates. In this reduced space, the
reliability index is defined as
American Institute of Aeronautics and Astronautics
8
( '*) ( '*)Tx xβ = (8)
Where '*x is the minimum distance point (most probable point of failure or MPP) from the origin of the reduced
space to the limit state boundary. An optimization problem [28] can be formulated to obtain this reliability index as
'min ( ') ( ')
. . ( , ') 0
T
xx x
s t g x x
β =
=
(9)
Finally, the failure probability is approximately calculated using the cumulative distribution function (CDF) of the
standard normal distribution as
( )fp β= Φ − (10)
It is not sufficient, however, to only compute probabilities of failure. The subject of estimating system reliability is
more involved, and probability of failure is one step in that estimation.
3.1 System Reliability Assessment
The component reliability estimation described in the previous section must be extended to include situations where
a system of multiple interacting components is to be considered. A structural system composed of more than one
structural element may also include multiple modes of failure, such as those resulting from large deflections, or from
buckling, shear, and corrosion, among others. Evaluation of the reliability of such a system involves multiple limit
state functions, and obtaining independent failure probability estimates for each constraint. Component-level
reliabilities can be estimated using the techniques described above. The final system reliability is likely to be a
function of the probabilities for the individual modes of failure, and any correlation between the different modes.
To obtain estimates of system failure probabilities, the traditional approach is to model the system as either a series
or a parallel system [43,28]. Series systems are those where the failure of the system occurs when any of the
individual limit states is violated. Such systems are also called weakest link systems: a statically determinate
structure is a typical example. In this case, the probability of failure is the union of the individual failure
probabilities. If iEF is the failure event corresponding to the i-th failure mode, the system failure probability is
given as [28,29]
( )series iP P EF= ∪ , (11)
The probability of the union is not trivial to calculate because of the presence of joint probabilities of one or more
events. Alternatively, system failure might occur only if some or all the elements fail (parallel system), and for
which case the system failure probability is the probability of intersection of the individual failure events as follows.
( )parallel iP P EF= ∩ , (12)
Statically indeterminate or redundant systems can be termed parallel systems. In practice, a structural system might
need to be decomposed into a set of series or parallel systems, where the final system reliability (or failure
probability) evaluation requires repeated applications of the above relations. Several approximations to the above
series and parallel reliabilities are in use. Conservative first-order series bounds are given as [27,28]
[ ] [ ]max ( ) 1 1 ( )i series iP EF P P EF≤ ≤ − ∏ − , (13)
Narrower second-order bounds that are widely used can be found in Haldar and Mahadevan [28]. For a parallel
system, first-order bounds are given as
American Institute of Aeronautics and Astronautics
9
[ ] [ ]1
max 1 1 ( ) , 0 min ( )
n
i parallel i
i
P EF P P EF
=
− − ≤ ≤
∑ , (14)
where the upper bound is exact if the events are perfectly correlated. It is important to note that if the individual
failure probabilities in the above equations are obtained using a first or second order approximation to the limit state
(FORM/SORM), then the accuracy of the above bounds may be affected further.
In some applications it is important to distinguish between the different failures modes. For purposes of improving
reliability, it is essential to identify the cause of failure down to the component level and, in many applications,
down to the actual physical cause of failure.
The failure time of a system with two or more failure modes can be modeled with a series-system or competing
risk model. Each risk is like a component in a series system. When one component fails, the system (i.e., product)
fails. Each unit has a potential failure time associated with each failure mode. The observed failure time is the
minimum of these individual potential failure times. Most generic system reliability techniques, such as the failure
mode approach, may not be adequate for large-scale aerospace systems, where multiple engineering teams work
concurrently on different aspects of the design. Instead, a dedicated approach that caters to their specific needs may
need to be developed. Vittal [1] describes a state transition approach to evaluate system reliability or effectiveness,
and where probabilities of a system transitioning from one failure to another are taken into consideration.
3.2 State Transition Method for System Effectiveness
The state transition method is an alternative approach for estimating risk in a system with multiple failure modes
(limit states). The concept is loosely patterned on ideas developed in the military operations research community in
the field of system effectiveness analysis. The approach is particularly significant in that it can compute the
probability of a system existing in different states (failure modes) as a function of system operation history and
failure mode relationship. This makes it possible to develop accurate risk models and “manage” the life of the
system in a manner that maximizes usage and minimizes risk. In the context of composite design using TFA, each
failure mode corresponds to progressive damage in the various constituent plies.
One of the early attempts to quantify system effectiveness is available as a report published by the Weapon
Effectiveness Industry Advisory Committee (WSEIAC) in 1965 [44]. Effectiveness is defined as a product of
availability (A), the dependability (D), and the capability (C) of the system under consideration, and expressed
mathematically as the following matrix product.
E A D C= × × (15)
In expanded form, this product is obtained as follows.
{ }
11 12 1 1
21 22 2 2
1 2 3
1 2
j
j
i
ji i ij
D D D C
D D D CE A A A A
CD D D
=
�
�…
�� � � �
�
(16)
In Eq. 16, ‘E’ is the System effectiveness and is a measure of the extent to which a system will perform its mission;
it is function of the probability to survive various states of damage, as well as the consequence of arriving in a
particular state (condition). ‘A’ is the availability, a measure of the system condition at the start of a mission. For the
purpose of the system degradation, it is the probability that a system is put into service in various conditions ranging
from new (full life), to a fully degraded state with no life. Ai is the probability of a system starting its mission state.
In context of composites it can be the probability of a particular ply to be available at any point of time during its
service life. ‘D’ is the dependability indicated by a transition matrix, where elements of the matrix dij represent the
probability of a system starting in state ‘i’ and ending in state ‘j’. ‘C’ is the capability, a measure of the system’s
ability to achieve a mission objective. For system-lifing applications, a system with full life will complete its desired
objective that may involve supporting given loads or surviving applied temperature gradients.
American Institute of Aeronautics and Astronautics
10
This model is extremely useful in that it provides a snapshot of the system or part in terms of various operational
states or part conditions. In many problems of probabilistic design, the probability of interest is the reliability of a
system, and the algorithm (as discussed in section II) usually produces a static number. By formulating the limit
states for each failure mode, a given design can be evaluated in terms of its probability of surviving these failure
modes. This clearly recognizes the fact that all failure modes are not as severe and risk to the system depends on the
type of failure as well as its combination with other failure modes.
In the context of a composite design problem, each of the failure modes like the ones described in Fig. 4 can be
included in the design problem. The feasibility of this approach can be illustrated using a composite design problem
involving two failure modes as follows.
Mode I: The failure strain at the point of first failure ( 1fε in Fig. 1) does not exceed an allowable value (0.5%).
Mode II: The failure strain at the point of second failure ( 2fε in Fig. 1) does not exceed an allowable value
(0.75%).
The failure mode II is assumed to be more critical. Let PI, PII be the probability of failure due to failure mode I, II
respectively. The following state transition tree can be developed to indicate the nature of the progressive damage.
Composite
Laminate
Fail in
mode II
Fail in
mode I
State 3
State 2
State 1
Yes
PII
No
1-PII
No
1- PI
Yes
PI
Figure 5. State Transition Tree
The various states are as follows.
State 1 – The undamaged laminate.
State 2 – A benign failure where the laminate has excessive preliminary failure strain.
State 3 – A complete failure where the laminate has excessive secondary failure strain.
The transition rules diagram can be expressed as shown in Fig. 6. It shows how the system can either stay in its
present state or transition to one of the others with a certain probability.
State
1
2
3
Figure 6. State Transition Diagram
The state transition matrix then takes the following form, representing the probabilities of transitioning to one of the
other states.
American Institute of Aeronautics and Astronautics
11
11 12 13
22 23
33
0
0 0
d d d
D d d
d
=
(17)
In the above, the components of D can be expressed as:
11 12 13
22 23 33
(1 )(1 ); (1 ) ;
(1 ); ; 1.0
I II II I II
II II
d P P d P P d P
d P d P d
= − − = − =
= − = = (18)
Note that the probabilities of modes I and II defined here as PI and PII are computed using a typical FORM
approach. The starting vector A can be assumed to be [ ]1.0 0 0 , i.e., indicating that the laminate starts with no
damage at all. The capability vector (C) can be assumed of the form [ ]1.0 0.5 0 which means that if the laminate
is in state 1 than it is 100% capable of accomplishing the mission; in state 2 it has 50% capability, and in state 3, it is
incapable of performing the desired mission. Here, the second entry of the capability vector ( 2c ) can take any value
from 0 to 1 and the effect of this choice on the overall results will also be shown in a successive section.
Using classical system reliability methods, reliability of a system with two or more failure modes can be modeled
with a series-system or competing risk model. For the simple case of two failure modes as in this problem, the
system reliability would be written as follows.
(1 )(1 )system I IIR P P= − − (19)
The proposed system effectiveness technique can be extended to handle multiple failure modes; it provides a
convenient way to handle the interaction between the modes that is not as obvious in probabilistic reliability based
approach. For the design of composite systems where damage in various modes may occur progressively, correlating
the individual failure modes needs to be given relative importance. Not all damage modes will be equally
catastrophic for system performance, and more meaningful composite structural designs are possible by constraining
the system effectiveness in contrast to designs based strictly on system reliability.
The calculation of individual failure mode probabilities is done by considering the limit state corresponding to each
failure mode and solving them using a FORM based approach explained in the previous section. For example, the
probability of failure corresponding to the first failure mode limit state is shown below in Eq. 20.
1 1
( )f allow
I fP P ε ε= ≤� � (20)
Here, the ~ overscript denotes a random variable and 1fε� and 1allowε� are the random variables corresponding to first
failure strain and the allowable strain at first failure in the laminate. Pf is the probability of failure of this limit state
obtained using FORM.
3.3 Composite System Reliability and Effectiveness Design Formulations
In context of composite design optimization, a wide array of design problem formulations can be deployed. The
optimization can be single or multi-objective, and can be a discrete or continuous programming problem. In the
presence of uncertainty in design variables and problem parameters, the system reliability or system effectiveness
constraints can be incorporated as part of the design statement.
The suitability of system effectiveness approach in probabilistic optimization is illustrated through a comparative
study of deterministic and nondeterministic design optimization solutions as follows.
Case A: Deterministic optimization formulation - the composite design problem is first formulated as a
deterministic multiobjective optimization problem that seeks to minimize the weight of the laminate and maximize
American Institute of Aeronautics and Astronautics
12
the modulus of elasticity ( 11E ) in the direction of unidirectional in-plane loading (11-direction from Fig. 1). The
design variables are chosen as the volume fraction of the fiber material ( rc ) and the ply angle φ in a [0, +/-φ , 90]s
laminate. The design problem statement is explicitly stated as follows.
11
1 1
2 2
:
:
. .
0.2 0.9
30 60,
f allow
f allow
r
Min Weight
Max E
s t
c
Integer
ε ε
ε ε
φ
≤
≤
≤ ≤
≤ ≤
(21)
The ply angles φ are deterministic integer design variables bounded between 30 deg and 60 deg. This is a design
problem with both continuous and integer design variables. A genetic algorithm (GA) based optimization solution
was implemented in this problem given the mixed nature of the design variables and expected multimodalities in the
design space. The GA based approach required a very large number of function evaluations and the use of the RBF
approximation was critical to this implementation. The RBF model describing the relationship between rc andφ to
the outputs 1fε and 2fε was used in the optimization. A flowchart showing the process of creating the RBF based
surrogate model and its subsequent integration into the GA based search is shown in Fig. 7. The multiobjective
problem is solved using a weighted objective function approach described in [46]. This formulation gives the Pareto
optimal set of solutions shown in Fig. 8, and which will be used for comparison with nondeterministic cases.
Design variables (x)
E.g. ,rc φK samples
using Space
filling Latin
Hypercube
design
ith iteration
Micro-TFA
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 130
35
40
45
50
55
60
cr
φ
(i+1)th iteration
ith Response (y)
E.g. First and
second failure
strains
Generating Training and validation data set for Neural Network
Generate
approximation
to y=
Using a Radial
Basis Function
Network
( )f x
GA based
optimizer for
mixed integer
programming
Figure 7. High fidelity Global optimization using a Radial basis function
American Institute of Aeronautics and Astronautics
13
3 4 5 6 7 8 9 10 110
1
2
3
4
5x 10
10
Weight (kg)
E11 (
lam
ina
te)
(N/m
2)
Pareto Frontier for a deterministic optimization formulation
Figure 8. Pareto optimal solutions for the optimization formulation in Eq. 21
Case B: Nondeterministic optimization formulation (reliability constraint) - the composite optimization problem is
formulated as a non deterministic multiobjective optimization problem seeking to minimize the mean weight
(Weightµ ) of the laminate while also maximizing the mean axial modulus of elasticity ( 11Eµ ) of the laminate. The
design variables are the volume fraction of the fiber material and the ply angle φ in a [0, +/-φ , 90]s laminate. The
design problem statement is explicitly stated as follows.
11
1 1 2 2min
:
:
. .
; (1 )(1 ) (1 ( ))(1 ( ))
0.2 0.9
30 60,
f allow f allowsystem system I II f f
r
Min Weight
Max E
s t
R R where R P P P P
c
Integer
µ
µ
µ
ε ε ε ε
φ
≥ = − − = − ≤ − ≤
≤ ≤
≤ ≤
� � � �
�
(22)
Here, rc� is a normal random variable ( ,0.01)r rc N cµ=� � with its mean value ( rcµ� ) as the design variable in the
optimization problem. The allowable failure strains follow a normal distribution, with values
as 1 2(0.5,0.05); (0.75,0.075)allow allowN Nε ε= =� � . The system reliability (using Eq. 19) and system effectiveness
(using Eq. 15) are computed for each of the solutions on the Pareto front. The results obtained are presented in Table
2 and also shown on Fig. 9.
American Institute of Aeronautics and Astronautics
14
Table 2 Pareto solution for Case B formulation
Volume Fraction Ply angle Weight Laminate Axial Stiffness Rsystem Esystem
0.898 57 10.780 4.979E+10 1.000 1.000
0.898 56 10.775 4.949E+10 1.000 1.000
0.870 57 10.445 4.114E+10 1.000 1.000
0.789 57 9.465 2.766E+10 1.000 1.000
0.757 56 9.081 2.456E+10 0.997 0.999
0.734 56 8.811 2.278E+10 0.982 0.991
0.730 55 8.763 2.248E+10 0.975 0.986
0.727 57 8.720 2.224E+10 0.969 0.983
0.717 57 8.604 2.158E+10 0.943 0.970
0.714 55 8.563 2.135E+10 0.924 0.958
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1.00
1.01
7.0 7.5 8.0 8.5 9.0 9.5 10.0 10.5 11.0
E_sy
stem
, R
_sy
stem
Weight (Kg)
R_system
E_system
Figure 9. Pareto solution for Case B formulation
Case C: Nondeterministic optimization formulation (effectiveness constraint) – this case involves the solution of a
multiobjective optimization problem as before, with the addition of system effectiveness based constraints. The
problem formulation is as follows.
11
min
:
:
. .
0.2 0.9
30 60,
system
r
Min Weight
Max E
s t
E E
c
Integer
µ
µ
µ
φ
≥
≤ ≤
≤ ≤
�
(23)
American Institute of Aeronautics and Astronautics
15
This problem formulation is the same as in case (B) except that reliability constraint is replaced by effectiveness
constraint. The results for this problem formulation are summarized in Table 3 and depicted graphically in Fig. 10.
Table 3 Pareto solution for Case C formulation
Volume Fraction Ply angle Weight Laminate Axial Stiffness Rsystem Esystem
0.899 56 10.784 4.977E+10 1.000 1.000
0.896 56 10.755 4.886E+10 1.000 1.000
0.880 55 10.563 4.358E+10 1.000 1.000
0.863 56 10.354 3.925E+10 1.000 1.000
0.777 57 9.330 2.648E+10 0.999 0.999
0.730 57 8.763 2.250E+10 0.975 0.987
0.728 57 8.741 2.237E+10 0.972 0.985
0.715 57 8.584 2.147E+10 0.938 0.967
0.710 56 8.517 2.111E+10 0.915 0.956
0.695 57 8.337 2.021E+10 0.825 0.909
0.80
0.85
0.90
0.95
1.00
1.05
7.0 7.5 8.0 8.5 9.0 9.5 10.0 10.5 11.0
Esy
stem
, R
syst
em
Weight (Kg)
R_system
E_system
Figure 10. Pareto solution for Case C formulation
3.4 Discussion
In Tables 2 and 3, it is evident that the reliability is bounded between 0.825 and 1.0, and effectiveness is bounded
between 0.909 and 1.0. For the same design on the Pareto front, the reliability is always lower than effectiveness.
This is due to the fact that competing risk model emphasizes both failure modes equally whereas the state transition
approach places less emphasis on the first failure mode. Thus, the effectiveness metric provides a more realistic
estimate of the true system capability and is based on the acceptance by the designer that the residual load carrying
capacity of the structure continues to be significant. This is also demonstrated in the second formulation with the
effectiveness constraint. The designs generated in Tables 2 and 3 are in general lighter than the designs generated by
the reliability based approach. In this case study, first failure strain was considered a benign failure, i.e., from
operational perspective a laminate with initial strain failure has still completed a part of its mission. This “risk
reduction” is captured by the effectiveness based formulation and is finally reflected in the optimal designs
produced.
American Institute of Aeronautics and Astronautics
16
Another aspect of effectiveness approach is to understand the impact of the terms in the capability vector “C” (only
2c can be chosen independently in the case of 2 failure modes). A detailed sensitivity study was done by varying 2c
from 0.2 to 0.9 for both Case B and C formulations. The resulting values for systemE and systemR was plotted against
each other as shown in Fig. 11 and Fig. 12, for Case B and C formulations, respectively. This clearly shows that as
2 0c → the systemE values approach systemR . In other words, for a Pareto design point the corresponding systemR is the
lower bound to systemE for all values of 2 0c > . Thus, the value of 2c can be a designer preference to penalize or relax
a particular failure mode based on prior knowledge of the system, expert input or system configuration. In general,
for N failure modes there would be 2N-1 states. So, the column vector (2 1,1)NC − will have 2N-3 values which can be
independently chosen by the designer as per their preference.
0.9 0.92 0.94 0.96 0.98 10.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
Esystem
Rsyste
m
Comparison of Esystem
and Rsystem
for formulation "B"
c2=0.2
c2=0.3
c2=0.4
c2=0.5
c2=0.6
c2=0.7
c2=0.8
c2=0.9
c2=0.2
c2=0.5
c2=0.9
Figure 11. Sensitivity of solutions to designer definitions of system capability (Case B)
American Institute of Aeronautics and Astronautics
17
0.9 0.92 0.94 0.96 0.98 1
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Esystem
Rsyste
m
Comparison of Esystem
and Rsystem
for formulation "C"
c2=0.2
c2=0.3
c2=0.4
c2=0.5
c2=0.6
c2=0.7
c2=0.8
c2=0.2
c2=0.5
c2=0.8
Figure 12. Sensitivity of solutions to designer definitions of system capability (Case C)
IV. Closing Remarks
The paper describes a new optimization problem formulation for laminated composite structures that are subject to
progressive failure. The TFA approach for analysis is shown to be effective in predicting the failure loads, and also
accounting for residual load carrying capacity of damaged composites. The method can be extended to provide a
probabilistic analysis tool that takes into consideration uncertainty in material and geometric parameters of the
laminate. However, as shown in this paper, the approach is computationally cumbersome due to coupling across the
length scales. In the presence of uncertainties, the issue of computational costs is further exacerbated, and defies
integration into an optimization framework without recourse to the use of function approximations. In the present
paper, a radial basis neural network is used to construct a surrogate model of the TFA approach that was coupled to
a genetic algorithm based optimization methodology. The genetic search was preferred over traditional
mathematical programming due to the mixed design variable space involving both continuous and discrete variables,
and also because of the expected multimodal behavior of the design space.
The optimization problem formulation is based on a concept of system effectiveness that models the degraded
behavior of a structure undergoing progressive damage during service. System effectiveness reflects the ability of
the structure to perform a required mission, and is a convenient tool to model designer preference in the optimization
problem formulation. The method uses routine uncertainty modeling tools to build an effectiveness measure that can
be used as an objective criterion in the optimization formulation. The paper compares numerical results obtained in
this new approach against those derived from more traditional system reliability formulations. The latter are known
to yield more conservative designs, and not as amenable to including designer preferences. Additional work on
implementing the proposed numerical scheme to larger scale design problems is currently in progress, and will be
reported in another publication.
V. Acknowledgements
The support received for this work under a grant from the AFOSR (Award no: FA 9550-05-1-0140) is gratefully
acknowledged.
American Institute of Aeronautics and Astronautics
18
References
[1] S. Vittal, 2002, “Modeling Risk and Uncertainty in Mechanical Design and optimization.” Ph.D. Thesis,
Rensselaer Polytechnic Institute, NY.
[2] M. Toya, 1974, “A Crack Along the Interface of a Circular Inclusion Embedded in an Infinite Solid,” J.
Mech. Phys. Solids, Vol. 22, pp. 325 348.
[3] IB.L. Karihaloo and K. Vistivanathan, 1985, “Elastic Field of a Partially Debonded Elliptical
Inhomogeneity With Debonding Over an Arc (Antiplane Strain)”, J. Appl. Mech., Vol. 52, pp. 91 97.
[4] T. Mura and R. Furuhashi, 1984, “The Elastic Inclusion with a Sliding Interface.” J. Appl. Mech., Vol.
51, pp. 308 310.
[5] T.Mura, I. Jasiuk and B. Tsuchida, 1985, The Stress Field of a Sliding Inclusion,” Int. J. Solids
Structures, Vol. 21, p. 1165.
[6] S.S. Steif and A. Dollar, 1988, “Longitudinal Shearing of a Weakly Bonded Fiber Composite,” J. Appl.
Mech., Vol. 55. pp. 618 623.
[7] R. Furuhashi, J.H. Huang and T. Mura, 1992, “Sliding Inclusions and Inhomogeneities with Frictional
Interfaces,” J. Appl. Mech., Vol. 59, pp. 783 788.
[8] J.W. Hutchinson and H.M. Jensen, 1990, “Models of Fiber Debonding and Pullout in Brittle Composites
with Friction,” Mechanics of Materials, Vol. 9, pp. 139-163
[9] G. J. Dvorak and J. Zhang, 2001, “Transformation field analysis of damage evolution in composite
materials,” J. Mechanics and Physics of Solids, 49, 2001, 2517-2541.
[10] M. Sejnoha, 1996, “Micromechanical Modeling of Unidirectional Fibrous Composite Plies and
Laminates,” Ph.D. Thesis, Rensselaer Polytechnic Institute, Troy, New York.
[11] M.V. Srinivas, 1997, “Micromechanical Modeling of Composite Plies and Laminates”, Ph.D. Thesis,
Rensselaer Polytechnic Institute, Troy, New York.
[12] S.W. Tsai and E.M. Wu., 1971, “A General Theory of Strength for Anisotropic Materials”, J. Comp.
Mat., Vol. 5, pp. 58-80.
[13] Z. Hashin, 1980, “Failure Criteria for Unidirectional Fiber Composites,” J. Appl. Mech., Vol. 47, pp. 329-
334.
[14] T.V. Parry and A.S. Wronski, 1985, “Kinking and Compressive Failure in Uniaxially Aligned Carbon
Fiber Composite Tested Under Superimposed Hydrostatic Pressure,” J. Mat. Sci., Vol. 17, pp. 893-900.
[15] R.M. Christensen, 1988, “Tensor Transformations and Failure Criteria for the Analysis of Fiber
Composite Materials,” J. Comp. Mat., Vol. 22, pp. 874-897.
[16] R.M. Christensen, 1990, “A Critical Evaluation for a Class of Micromechanics Models,” J. Mech. Phys.
Solids, Vol. 38, pp. 379-404.
[17] R.M. Christensen, 1996, “Stress Based Yield/Failure Criteria for Fiber Composites,” Int. J. Solids
Structures, Vol. 34, pp. 529-543.
[18] W.W. Feng, 1991, “A Failure Criterion for Composite Materials,” J. Comp. Mat., Vol. 25, pp. 88-100.
[19] H.T. Hahn and M.N. Kallas, “Failure Criteria for Thick Composites,” Ballistic Research Laboratory
Report, BRL CR 691, Aberdeen, MD.
[20] L.J. Hart Smith, 1993, “Should Fibrous Composite Failure Modes be Interacted or Superimposed,” Vol.
24, pp. 53 55.
[21] G.J. Dvorak, 2000, “Damage evolution and prevention in composite materials,” in Mechanics for a New
Millenium, ed. by H. Aref and J. W. Phillips (Proceedings ICTAM 2000, the 20th Intl. Congress of
Theoretical and Apllied Mechanics), Kluwer Academic Publishers, 2000, pp. 197-210.
[22] C.C. Chamis and G.H. Abumeri, 2005, “Probabilistic Dynamic Buckling of Composite Shell Structures,”
Composites: Part A, Vol. 36, pp. 1368–1380.
[23] Y. Liu and S. Mahadevan, 2005, “Probabilistic Fatigue Life Prediction of Multidirectional Composite
Laminates,” Composite Structures, Vol. 69, pp. 11-19.
[24] G.H. Abumeri and C.C. Chamis, 2004, “Non-Deterministic Optimization of Composite Structures
Reliability,” AIAA 2004-1896, 45th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics &
Materials Conference, Palm Springs, CA.
[25] H.E. Soliman and R.K. Kapania, 2005, “Probability of Failure of Composite Cylinders Subjected to
Axisymmetric Loading,” AIAA Journal, Vol. 43, No. 6, pp. 1342-1348.
[26] J. Arbocz, and M.W. Hilburger, 2005, “Toward a Probabilistic Preliminary Design Criterion for Buckling
Critical Composite Shells,” AIAA Journal, Vol. 43, No. 8, pp. 1823–1827.
[27] R.E. Melchers, 1999, “Structural Reliability Analysis and Prediction,” 2nd edition, John Wiley and Sons.
American Institute of Aeronautics and Astronautics
19
[28] A. Haldar and S. Mahadevan, 2000, “Probability, Reliability, and Statistical Methods in Engineering
Design”, John Wiley & Sons, New York.
[29] S. Park, S. Choi, C. Sikorsky, and N. Stubbs, 2004, “Efficient Method for Calculation of System
Reliability of a Complex Structure,” International Journal of Solids and Structures, 41, 5035-5050.
[30] E.S. Gharaibeh, D.M. Frangopol, and T. Onoufriou, 2002, “Reliability-based Importance Assessment of
Structural Members with Applications to Complex Structures,” Computers and Structures, 80, 1113-1131.
[31] S. Mahadevan and P. Raghothamachar, 2000, “Adaptive Simulation for System Reliability Analysis of
Large Structures,” Computers and Structures, 77, 725-744.
[32] A. F. Mehr and I. Y. Tumer, 2005, “A New Approach to Probabilistic Risk Analysis in Concurrent and
Distributed Design of Aerospace Systems,” DETC2005/DAC-85056, Proceedings of Design Engineering
Technical Conferences and Computers and Information in Engineering Conferences, Long Beach, CA.
[33] G.J. Dvorak and Y.A. Bahei-El-Din, 1995, “Transformation Analysis of Inelastic Laminates,” In: Pyrz,
R. (Ed.), IUTAM Symposium on Microstructure-Property Interactions in Composite Materials. Kluwer
Academic Publishers, Netherlands, pp. 89–100.
[34] R. Khire, P. Hajela and Y. Bahei-El-Din, 2007, “Handling Uncertainty Propagation in Laminated
Composites Through Multiscale Modeling of Progressive Failure,” AIAA-2007-1913, Proceedings of the
48th AIAA Structures, Structural Dynamics, and Materials Conference, Honolulu, Hawaii.
[35] P.D. Soden, M.J. Hinton, A.S. Kaddour, 2002, “Biaxial test results for strength and deformation of a
range of E-glass and carbon fibre reinforced composite laminates: failure exercise benchmark data - Part I”,
Composites Science and Technology, Vol. 62, No. 12.
[36] P.D. Soden, M.J. Hinton, A.S. Kaddour, 2002, “Biaxial test results for strength and deformation of a
range of E-glass and carbon fibre reinforced composite laminates: failure exercise benchmark data - Part
II”, Composites Science and Technology, Vol. 62, No. 12.
[37] J. Moody, C.J. Darken, 1989, “Fast learning in networks of locally tuned processing units”. Neural
Computation 1, 281–294.
[38] C.M. Bishop, 1995, “Neural Networks for Pattern Recognition”, Oxford University Press.
[39] J. Deng, 2006, “Structural reliability analysis for implicit performance function using radial basis
function network”, International Journal of Solids and Structures, 43, 3255–3291.
[40] J. Lawson and J. Erjavee, 2001, “Modern Statistics for Engineering and Quality Improvement”
Thompson Learning, Pacific Grove, CA, USA.
[41] R. Hecht-Nielsen, 1989, “Neurocomputing”, Addison-Wesley Publishing Company, San Diego.
[42] S. Haykin, 1999 “Neural Networks: A Comprehensive Foundation”, second ed. Prentice-Hall, New
Jersey.
[43] O. Ditlevsen and H.O. Madsen, “Structural Reliability Methods”, John Wiley and Sons, London, 1996.
[44] MIL-HDBK-338B, 1998, “Military Handbook – Electronic Design Reliability Handbook”.
[45] A. Messac, J.G. Sundararaj, R.V. Tappeta, J.E. Renaud, 2000, “Ability of Objective Functions to
Generate Points on Non-Convex Pareto Frontiers,” AIAA Journal, Vol. 38, No. 6, pp. 1084-1091.