Compurm & Strurrures Vol. 62, No. 4.

pp, 737-745, 1997

Copynght 0 1996 Elsevier Science Ltd
Printed in Great Britain. All rights reserved
PII: SO0457949(96)00234-9 004s7949/97 117.00 + 0.00

ASTROS FOR RELIABILITY-BASED MULTIDISCIPLINARY

STRUCTURAL ANALYSIS AND OPTIMIZATION
Xiaodong Luo and R. V. Grandhi
Department of Mechanical and Materials Engineering, Wright State University, Dayton, OH 45435,
U.S.A.

(Received 10 May 1995)

Abstract-k1 conducting multidisciplinary structural optimization, there is uncertainty and randomness

in the input information of finite element models, material properties and boundary conditions. Because
of the stringent performance requirements on the missions of modern aircraft, there is a need for reducing
the structural failure probability due to uncertainties by conducting stochastic analysis instead of a
deterministic approach. This paper presents a strategy for conducting such an analysis in a widely used
structural ar,alysis and optimization program ASTROS. The execution control sequence of ASTROS is
modified and reliability analysis modules are incorporated for analysis and multidisciplinary optimization.
The implementation and modification of ASTROS for incorporating the random nature of input
information is demonstrated through truss and wing structure examples. Copyright 0 1996 Elsevier
Science Ltd.

INTRODUCTION ASTROS combines mathematical optimization

algorithms with traditional structural analysis disci-
Aerospace structures are usually operated in extreme plines such as static forces, normal modes, static
and uncertain environments, and subjected to the aeroelasticity and dynamic aeroelasticity (flutter), all
requirements of minimum weight, high performance in a finite element context, to perform automated
and reliability. The: probabilistic or reliability-based preliminary design of an aircraft structure. The
approach has been widely used in structural design to comprehensive capability and executive control
deal with uncertain parameters, and significant sequence of ASTROS have made it a powerful and
advancements have taken place during the last decade widely used design tool in both general purpose
in both the theoretical and computation techniques of preliminary structural design and specific structural
structural reliability analysis [14]. As the extensive design aspects. The idea of this paper is to integrate
development and wide application of stochastic an efficient reliability analysis procedure with
analysis techniques continue, it is appropriate and ASTROS for probabilistic analysis and optimization.
necessary to combine the reliability-based analysis The ASTROS analyzer is used to generate structural
procedure with multidisciplinary optimization tech- information such as active constraints and function
niques, in which the structural parameters and design sensitivities. This information is extracted from the
variables are considered as random, to obtain a database in evaluating the reliability analysis
powerful approach to the practical optimal design. procedure to evaluate the safety index and sensi-
Work in this aspect has been reported [5-81, and it tivities, and they are fed back to the ASTROS
has been shown analytically and experimentally that optimizer for safety index constrained optimization.
the reliability-base:d optimization can be more The ASTROS standard executive sequence is
effective than using a factor of safety in deterministic modified to include new constraints and constraint
optimization for incorporating randomness of input sensitivities for reliability considerations, and to reset
information [S]. Most of the reported work developed the required ASTROS modules.
either a general purpose finite element analysis for ASTROS is an executive controlled software
structural reliability or developed limited reliability procedure. One of the functions of the ASTROS
optimization capabilities in research work. executive system is to determine the sequence in
The purpose of this paper is to build reliability which the modules of the procedure are invoked.
analysis and optimization capabilities in the widely ASTROS uses Matrix Analysis Problem Oriented
used structural design software ASTROS (Auto- Language (MAPOL) to perform set procedures in
mated STRuctural Optimization System). ASTROS structural analysis and design. It provides the user
is a computer program for the multidisciplinary with an opportunity to install new codes and
analysis and design of aerospace structures, functional capabilities into the system, and interacts
developed by Wright Patterson Air Force Base, Ohio. with the software. MAPOL allows the user to

CAS 6214-E
137
738 Xiaodong Luo and R. V. Grandhi

manipulate the software system in many ways to with a hyperplane or a quadratic approximation
tailor the available capabilities to perform particular surface. Then the first-order approximation to Pr is
tasks and, at a higher level, add modules to the given by the expression
system or replace modules that already exist. In this
paper, the standard ASTROS MAPOL sequence is
Pr=Q(-B) (3)
modified to include new constraints and constraint
sensitivities for structural safety indices, and manip-
where @(.) is the standard normal cumulative
ulate the required ASTROS modules in analysis,
probability, and p is known as the safety index, i.e.
sensitivity and optimization phase, and finally add
the distance from the origin to the limit-state surface.
the new reliability analysis module to compute the
Since the probability density in the standard normal
structural safety indices and the sensitivities.
space decays exponentially with distance from the
The reliability analysis is an iterative process where
origin, the point with a maximum failure probability
the calculation of the safety index for each limit state
on the limit-state surface for approximating the
function needs more than one structural analysis
surface is the point of minimum distance to the
and sensitivity calculation with respect to random
origin. To find this point, a constrained optimization
variables. Hence the design procedure is a nested
problem can be formulated as part of reliability
scheme with an inner loop on safety index
analysis:
convergence. ASTROS is used as an analyzer even in
the inner loop. In this paper, an efficient reliability
analysis developed by Wang and Grandhi [4] is Minimize: F(Y) = YrY (4a)
implemented, in which the intervening variables and
nonlinear approximation methods are used for safety Subject to: G(Y) = 0. (4b)
index calculation, and an analytical approach is used
for probabilistic sensitivity evaluation. A reliability Mathematical optimization schemes or iterative
analysis program IS compiled and defined as a algorithms can be used to solve this problem. In this
functional module of the ASTROS, and then linked paper, an efficient safety index calculation approach
to the modified ASTROS sequence to generate a new developed by Wang and Grandhi [4] is implemented,
ASTROS execution system. in which the limit state function is expanded in terms
The effectiveness of the proposed approach is of intervening random variables instead of in basic
demonstrated through three sample cases. Ten-bar random variables.
truss, ACOSS-II model and the intermediate The intervening variables, S = (s,, s2, . , s~)~,are
complexity wing (ICW) are considered with displace- defined as
ment and frequency constraints.

s, = x: i= 1,2 , >N, (9
STRUCTURAL RELIABILITY SYNTHESIS

1. Safety index calculation where x,(i = 1, 2, . , N) are the physical variables,

In the analysis of structural reliability, the and r represents the nonlinearity index, which varies
“performance function” or “limit state function” from one iteration to another. This nonlinearity index
g(X) is defined in terms of basic random variables is obtained from the following feedback formula:
X = (xl, x2, , x,)~ such that g(x) = 0 separates
the design space into “failure” and “safe” regions. g(%I)
The probability of structural failure in the specified
mode is

(1)
-
[
g(x,)++-‘)~ (x:. 4 _ , - x: h)
I
= 0

wherefx(X) is the joint probability density function (6)

of X, and R is the failure region (g(X) > 0). The
solution of this multiple integral is extremely where X, is the current design point and X, _, 1sthe
complex. A widely used alternative method is the previous design point. x,,~ is the ith variable of the kth
safety index approach [4, 51, in which the basic design point.
random variables X are transformed into standard, First, a linear approximation of the performance
uncorrelated and normal variates Y = T(X), where function is generated by using the first-order Taylor’s
T(X) is the transformation matrix. In the standard series expansion about the mean values, and the most
normal space, first- and second-order approximation probable failure point XI and safety index pi are
of the failure probability P, can be obtained by computed using the Hasofer-Lind [9] algorithm.
replacing the limit-state surface For the second iteration onwards the nonlinearity
index r is determined by solving eqn (6) based on
g(T-‘( Y)) = G(Y) = 0, (2) the information of the current and previous points,
 ASTROS 139

and a nonlinear approximation of the performance problem (9). The perturbations for Y* must satisfy
function is constructed as the following condition:

6G=VhG6b+VyG6Y=O, (IO)
Z(x) = g(X) -I- ; 5 x$-” q (X: - X&)_ (7)
where the derivatives are obtained at Y* and bD.
Kuhn-Tucker necessary conditions must be satisfied
This approximate equation is used repeatedly to at b”, since Y* and b@are optimal solutions for the
reach the limit state function boundary of a minimum minimization problem (8b). This can be given as
distance point. The function and gradient infor- follows by introducing the Lagrangian multiplier 1:
mation is updated with ASTROS exact analysis; also
the nonlinearity index r is computed based on the VyF + lTVyG = 0 (I la)
most recent information. This procedure is repeated
until the safety index fi is converged. Typically the G(bO, Y*) = 0. (1 lb)
safety index calculation took 2-3 exact analyses using
ASTROS. The approximation given in eqn (7) From eqns (9)-(11a), the following expression is
represents the limit function more accurately with the obtained as
nonlinearity index r, so it reduced the number of
times exact analysis is conducted. This has become an !$=ITVbG( Y*, b”) (12)
inner optimization using the ASTROS analysis
module. The details are presented in the ASTROS
executive sequence modifications section (Fig. 2). = patay
ay ab
2. Sqfety index serrsitivity analysis
The Lagrangian multiplier 1 of the above equation
An important task in reliability based optima1
can be obtained from Kuhn-Tucker conditions (1 l),
design is to obtain sensitivities, which are used to
and the expression is given as follows by considering
study the effect of parametric variations, calculate the
eqn (8b) and its derivatives:
search directions for finding an optimum design, and
construct function approximations. An analytical
approach for the sensitivity analysis of the safety
index presented in Ref. [lo] based on an advanced
’ = ,,VyG(;*, b”),,’

first-order second-moment (AFOSM) method is used From eqns (12) and (13), the exact derivatives of
in this paper. When an AFOSM method is used to safety index p with respect to design variable b can
evaluate the failure probability, the single mode of be obtained when Y* is the real optimal solution of
failure probability constraint can be written as the minimization problem (8b), that is, if the fi is
exact.
\!?- f&‘(P) > 0 @a) In many practical problems, the design variables b
are usually modeled with uncertain means and
1 = yei; (P Y)“’ (8b) uncertain (or fixed) standard deviations. For this
case, aY/ab in eqn (12) can be derived easily since Y
A ==(ylG(b, Y) = 0) (84 is a function of means and standard deviation, and
the 1 and aG/a Y in eqn (12) are the by-products of
where Y denotes a vector of standard, uncorrelated the reliability analysis. Thus no additional analysis is
and normal variables, b stands for a vector of needed for sensitivity calculation after conducting the
optimization variables, I$-‘( .) represents the inverse traditional structural analysis using ASTROS.
cumulative standard normal distribution function,
and P is the required probability of survival for
RELIABILITY-BASED OF’TIMIZATlON
failure mode G(b, Y).
To obtain @ in terms of 66, the minimization A design problem for achieving a robust structural
problem in eqn (8b) is solved for design variables b”. system subject to performance, weight, cost and
Then, the perturbations for the most probable failure reliability constraints with design variables which are
state Y* are stated as follows: random can be generally defined as

s/3 = nl” {VYF(Y)BY} (9) Minimize F(b) (144

where F(Y) = (P Y)“*, VvF( Y) represents the deriva- subject to P[Gi(b, x) > 0] 2 P,
tives vector of F(Y) with respect to normalized
random variables Y. Assuming that the minimum j=l,2,...,J (l4b)
point Y* for the problem (8b) is unique, the variation
can be taken in t’he context of the minimization b:<b,<bj’, i=l,2 ,..., N, (l4c)
740 Xraodong Luo and R. V. Grandhi

where X is a vector of random variables, b is a vector Initialization (PRBPACE) Segment

of design variables, P[ .] denotes the probability of the START OPTIMIZATION LOOP
event [.I, and P, is the required probability of survival
for the jth constraint or failure mode, G,(b, X). ANALYSiS PHASE
In this paper, all design variables are assumed to For Each Boundary Condition Do
be random, and their mean values are treated as the Discipline 1
Subcase 1, Constraints
design variables. Generally, it is almost impossible to Subcase 2, constraints
calculate the probability of eqn (14b) by a multiple . ..
Discipline 2
integration, thus the above probabilistic optimization .. . .
problem is converted into the following equivalent End Do

deterministic one by using a first-order second-mo-

SENSITIVlTY PHASE
ment method:
Select Active Coastmints
For Each Active Bouodary Constraint Do
Active Discipline 1
Minimize F(b) (Isa) Active Subcase 1, Constraiat Sensitivities
Active Subcase 2, Constraiat Sensitivities
. .....
Subject to g,(b) = @-‘(P,) - /?, < 0 Active Discipline 2
. ..
End Do
j = 1,2, . )J (15b)
[ OPTIMIZATION PHASE 1
b:<b,<b:, i=1,2 ,..., N. (15c) Redesign Based on Current Active Constraints
sad Constraint Sensitivities
END LOOP
The constraint (15b) ensures that the probability of
survival for the limit function j is greater than the For Each Boundary Condition Do
desired value. Discipline 1
Subcase 1
This equivalent deterministic optimization prob- Subcase 2
....
lem is solved using ASTROS to build the reliability- Discipline 2
based optimization capabilities in ASTROS. . .. ...
End Do

Fig. 1. Structure of the standard MAPOL sequence

MODIFICATIONS OF ASTROS EXECUTIVE SEQUENCE

1. Standard ASTROS executive sequence

A standard executive sequence, or MAPOL auxiliary computations not available or execute only
procedure, is available in ASTROS and it supports a portion of the solution. Special MAPOL commands
multiple engineering disciplines and optimization and EDIT declarations in the ASTROS input stream
features. The standard MAPOL sequence consists of allow these modifications to be performed on any
two major components: the variable declarations and matrices, database entities, functions and procedures
the solution algorithm. The solution algorithm can be available in ASTROS, but if the user wishes to
further divided into preface modules, an optimization include an external procedure or function, then the
segment and a final analysis segment. The declaration ASTROS system generation program needs to be
segment declares all variables used in the MAPOL modified and executed to include new utrlity and
sequence, including all integer and real scalar application modules, link the libraries and assemble
variables as well as high order variables such as the new executive Image of ASTROS.
relations, matrices and unstructured database enti-
2. Modifications of‘ ASTROS executive sequence
ties. Within the solution algorithm, the preface
modules comprise a group of engineering modules As described previously, there are three phases m
exercised prior to the boundary condition loops to ASTROS optimization segment: analysis, sensitivity
perform a number of system initializatron tasks. and optimization. In reliability based optimization,
Separate optimization and analysis segments consist these three phases need to be manipulated to include
of a loop on the number of (optimization or analysis) external procedures, new constraints and sensitivities.
boundary conditions in the current execution. In the The idea of modifying the ASTROS executive
optimization segment, a second boundary condition sequence can be illustrated through the algorithm
loop is performed to obtain the sensitivities of active flow-chart shown in Fig. 2.
constraints in preparation for the optimization task. After extracting the structural information from
Figure 1 provides the standard algorithm structure the input stream, ASTROS starts optimizatron
showing how the multidisciplinary optimization is iterations. To conduct a reliability analysis, ASTROS
performed in ASTROS [ 1I, 121. takes the structural model information and sends
In some cases, the user may wish to modify the it to an external procedure. This procedure combines
standard ASTROS solution in order to perform some the structural information from ASTROS and the
 ASTROS 741

1 ASTROS I 1 ReliaMitv I

Fig. 2. Flow-chart of the algorithm using ASTROS.

probabilistic information, such as coefficients of outer loop will be continued until the optimization is
variation and distribution type, to generate a finished.
stochastic model in which the design variables are For the tasks described above, several major
assumed to be random about their mean values. The modifications of the ASTROS executive sequence
reliability analysis procedure then starts the safety were made as discussed below.
index /3 iterations, which need limit state function and (1) Two ASTROS processes are required in the
their sensitivities. So it calls another ASTROS algorithm. The first one acts as a controller to
process only for analysis and sensitivity phases, which conduct the reliability based optimization, and the
provides the active constraints (limit state functions) second one supplies the structural constraint and
and their sensitivities with respect to the random sensitivity information to the safety index algorithm.
design variables. Using this information, an adaptive Both of them need MAPOL modifications to
nonlinear approximation is developed for the limit resequence the required ASTROS modules related to
state function. The iteration for reaching the shortest analysis, sensitivity and optimization phases, as well
distance of the function boundary is done using the as the external procedure, to realize the computations
Hasofer-Lind algorithm just on the approximation shown in Fig. 2. In the controlling process, the
function. The function value and the sensitivity analysis and sensitivity phases in optimization
information is updated using ASTROS, and the iterations are skipped, and the flow is shifted to the
iterations are repeated until /? convergence is realized. external procedure and then returned back to the
This finishes the inner loop shown in Figure 2. redesign phase. In the second ASTROS process, the
Next, the reliability constraints and their sensi- analysis and sensitivity phases are required while the
tivities with respect: to random design variables are redesign phase is skipped. All of these are done by
constructed using a chain rule discussed in earlier adding control statements and procedure invoking
sections, and fed back to the ASTROS optimization statements in the standard MAPOL sequence or in
phase, in which the mean values are considered as the ASTROS input stream.
design variables. With new constraints and sensi- (2) An external procedure is added as a functional
tivities, the ASTROS optimization process conducts component of the ASTROS library to conduct
a redesign task and checks for the convergence. The reliability analysis. ASTROS is capable of including
142 Xiaodong Luo and R. V. Grandhi

standard FORTRAN 77 subroutines and functions.

This is accomplished by compiling the procedure,
defining a new module in the module definition
database, and executing the ASTROS system
generation program to relink the ASTROS system.
The procedure calls the second ASTROS process
repeatedly to supply constraint and sensitivity
information durmg the safety index fi iterations. This
ASTROS process is called using the operating system
commands, since ASTROS is not designed to be
called as a subroutine.
(3) New constraints and constraint sensitivities are El
introduced in the controlling ASTROS process. In
ASTROS, the optimization is conducted by calling I P
the redesign module which appears as: Fig. 3. Ten-bar truss.

CALL DESIGN (. . . CONST, [AMAT], .)

cross section areas. The optimization gives the final
where CONST is a RELATION entity with 20 weight of 4885.28 lb with one displacement constraint
attributes for multidisciplinary constraints infor- after 14 iterations. In reliability optimization, the
mation, and [AMAT] is a MATRIX entity for safety index fl constraint is applied to ensure that the
sensitivities information. A new RELATION entity reliability is not less than 0.999 (/I 2 3.09). All the
BCONST is defined for reliability constraints. At cross section areas b, are considered as random design
present, the BCONST entity is defined in place of variables, which is denoted as case 1. In case 2, the
displacement constraints, which makes ASTROS Young’s modulus E is also treated as a random
consider handling reliability constraints just like parameter, and in case 3, both Young’s modulus E
displacement constraints. To introduce new sensi- and load P are taken into account as random
tivities, a new MATRIX entity [BMAT] is defined, in parameters. All the variables are assumed to have
which each column contains sensitivities of an active normal distributions. In all three cases, the initial
reliability constraint with respect to design variables. design starts from the point realized after IO
Thus the above calling statement is replaced by: iterations of the deterministic optimization [ 131.
Table 1 lists the mean values of the initial design and
CALL DESIGN (. BCONST, [BMAT], .) the corresponding coefficients of variation (values in
parenthesis are the reliability optimization results of
With this new definition the ASTROS control case 1, which will be discussed later).
sequence conducts the structural optimization for The iterations histories of the three optimization
minimizing the weight subject to the reliability cases, the safety Indices and the corresponding
constraints. reliability are listed in Table 2, in which the numbers
in parentheses represent the number of iterations
EXAMPLES AND DISCUSSIONS
needed for computing fl. In case 1, only cross section
areas (design variables) are random. The initial
The effectiveness of the proposed approach is
demonstrated through three sample cases. A IO-bar
truss, ACOSS-II model and ICW model are
examined with displacement, frequency and stress
constraints. Member sizes, material properties and
external loads are considered as random parameters,
the mean values of member sizes are treated as design
variables, and the structural weight is minimized by
placing the constraints on the reliability indices of
displacements, natural frequencies or stresses.

Example 1: IO-bar truss

The IO-bar truss structural model, shown in Fig. 3,
has six nodes and 10 truss elements made of
aluminum with a Young’s modulus of E = 10’ psi.
The deterministic design problem minimizes the
structural weight while limiting the transverse
displacement at node 2 to 2.0 in under the load of
P = 10s lb. The design variables are the truss element Fig. 4. ACOSS-II finite element model.
 ASTROS 143

Table I. Statistical data of IO-bar truss increased and the other six are decreased, resulting in
Random Mean Coefficients a decreased structural weight. This is consistent with
variables values of variation the results presented in Ref. [8]. In this case, the
bl 29.536 (33.955) 0.05 number of fi iterations is always two, indicating the
b2 2.7327 (0.7092) 0.05 efficiency of the B algorithm.
bz 29.456 (23.436) 0.05 In case 2, the random variables include cross
bs 13.589 (16.066) 0.05
bs 0.1000 (0.1000) 0.05
section areas and Young’s modulus of the material.
bo 2.1663 (0.6626) 0.05 The reliability of the initial structure is 0.6436
b, 14.045 (4.9484) 0.05 (/I = 0.3680). After 10 iterations, the structure is
b# 19.184(24.101) 0.05 optimized with the increased reliability of 0.999
bs 19.236 (23.748) 0.05
(b = 3.0900) along with a higher weight of
b,o 3.0002 (0.5333) 0.05
E lb7 0.05 5955.89 lb. In case 3, random variables cover all the
P 105 0.05 design variables, Young’s modulus and loading P.
With these uncertain parameters, the reliability of the
initial design is as low as 0.6077 (/l = 0.2735). The
reliability is increased to 0.999 (/? = 3.0893) with the
design has a structural weight of 5616.71 lb and
optimum weight being 6241.69 lb after 10 iterations.
reliability of 0.829 (/3 = 0.9509). This indicates that
In these two cases, the number of !3 iterations is two
the initial design needs to be redesigned to meet the
or three in the inner loop.
reliability goal of 0.999. The optimization process
Table 2 shows that with the increase in the number
minimizes the structural weight to 5412. 58 lb with
of random variables, the optimum weight increases
the reliability of 0.999 (p = 3.0881). The optimization
quite significantly, from the deterministic optimum
decreases the weight while increasing the reliability by
weight of 4885.28 to 5412.58 lb in case I, 5955.89 lb
redistributing the element cross section sizes. This
in case 2 and 6241.69 lb in case 3. In this example
redistribution is shown in Table I in which the values
where the displacement constraint is considered, the
in parentheses represent the optimum design vari-
variation of Young’s modulus E of the material seems
ables. It can be seen that four design variables are
to have the most significant effect on reliability based
optimization.
Table 2. Iteration history of IO-bar truss The move limits selection plays an important role
in the reliability optimization. The default move
Iteration
number Weight (lb) Safety index fl Reliability P limits of ASTROS were modified to get a stable
convergence rate. The first six iterations had
Case I
MOVLIM of 2.0 and from the seventh iteration
5616.71 0.9509 (2) 0.8292
5774.04 3.3036 (2) 0.9995 onwards it was 1.4 for cases 1 and 2, and 1.3 for case
5691.10 3.2307 (2) 0.9994 3. These values were selected after several trials.
5607.89 3.1809 (2) 0.9993
5539.04 3.1605 (2) 0.9992 Example 2: ACOSS-II model
5477.36 3.1142 (2) 0.9991
5433.27 3.1036 (2) 0.9990
The structural model of this example is the
8 5420.83 3.0871 (2) 0.9990 modified ACOSS-II (Active Control of Space
9 5412.58 3.0881 (2) 0.9990 Structures-Model 2) presented by Grandhi and
Venkayya [14]. The finite element model, shown in
Case 2
Fig. 4, has 33 nodes and 113 truss elements made of
5616.71 0.3680 (2) 0.6436
2 6479.97 3.2681 (2j 0.9995 a graphite epoxy material. The deterministic design
3 6335.82 3.1625 (2) 0.9992 problem minimizes the weight of the structure while
4 6236.05 3.1356 (2j 0.999 l imposing a lower bound frequency constraint of
5 6141.44 3.1151 (2) 0.9991
2.0 Hz on the first mode. The design variables are the
6 6074.50 3.1045 (3) 0.9990
6011.30 3.0886 (3) 0.9990 113 truss element cross sectional areas with a lower
8 5989.03 3.0914 (2) 0.9990 bound of 0.10 in*. A converged solution after 13
9 5971.43 3.0893 (2) 0.9990 redesign cycles gives the objective function value of
IO 5955.89 3.0900 (2) 0.9990 26.4525.
In reliability optimization, all the truss element
Case 3
5616.71 0.2735 (3) 0.6077 cross sectional areas are considered as random design
2 6952.46 3.5224 (3) 0.9998 variables, with a safety index p constraint (with
3 6657.99 3.1606(3) 0.9992 respect to lower bound frequency constraint of 2.0 Hz
4 6546.71 3.1269 (3) 0.999 1 on the first mode) not less than 3.09, indicating that
5 6447.68 3.1083 (3) 0.999 1
6 6372.81 3.0981 (3) 0.9990 the reliability is greater than 0.999. The final design
7 6323.21 3.0937 (3) 0.9990 of deterministic optimization is taken as the initial
8 6289.29 3.0915 (2) 0.9990 point of the reliability optimization. Table 3 shows
9 6264.14 3.0901 (2) 0.9990 the design iteration history for the reliability
10 6241.69 3.0893 (2) 0.9990
optimization process.
744 Xlaodong Luo and R. V. Grandhi

Table 3. Iteration history of ACOSS-II Table 4. Iteration history of ICW model

-
Iteration Objective Safety index Reliability Number
number function B P Iteration Weight Safety index of p Iterations
I 26.4525 number (lb) 4m>n (averaaeiconstrainth
-0.0218 (3) 0.4913 _ ‘I

2 27.1448 3.0451 (3) 0.9988 I 76.3483 -0.01 IO 102 (3.4)

3 27.0946 3.0171 (3) 0.9987 2 84.5604 2.0640 82 (2.7)
4 27.0940 3.1014 (3) 0.9990 3 84 5260 3.0921 82 (2.7)
5 27 0741 3.0653 (3) 0.9989 4 84.5291 3.0899 82 (2.7)
6 27.0647 3.0989 (3) 0 9990
7 27.0367 3.0825 (2) 0.9990
8 27.0268 3.0880 (3) 0.9990
each of 30 active stress constraints. The reliability
optimization starts from the final design of determin-
It is noted from Table 3 that the reliability of the istic optimization. The design iteration history is
deterministic optimum design (the initial design in shown in Table 4. The results show that the reliability
reliability optimization) is 0.4913 (1 = - 0.0218). of 0.4957 (corresponding to the most active stress
After the optimization, the reliability reaches the constraint) at the deterministic optimum design
design goal of 0.999 (fl = 3.09) with a very small violates the constraint. Hence the structure has to be
increase of as low as 2% in the objective function. redesigned to reach the reliability goal of 0.999. After
This is attributed to the redistribution of design four design cycles, the minimum reliability is
variables during the reliability optimization. Mean- improved to 0.999 (p = 3.0899), with a slight weight
while, the number of fi iterations never exceeds three increase (less than 11%) to 84.5291 lb. The numbers
for this example with 113 design variables. In this in the parentheses are the average fl iterations needed
example, ASTROS default move limit 2.0 is used. for each optimization iteration. In the first optimiz-
ation iteration, for instance, fl calculation for each
Example 3: intermediate complexity wing stress constraint needs average 3.4 iterations, so it
This example demonstrates the reliability based needs a total of 102 p iterations, because there are 30
optimization with multiple constraints. The struc- active /I constraints, which means 102 structural
tural model of the intermediate complexity wing analyses. It 1s noted in Table 4 that after the first
(ICW), shown in Fig. 5, uses 62 quadrilateral and two iteration, the structural weight changes very slightly,
triangular membrane elements to model the wing while the minimum value of safety index is increased
skins, 55 shear panels to model the substructure, from 2.0640 to the required value of 3.09, by
and 39 rod elements as posts to complete the redistributing the design variables.
interconnection of the upper and lower surfaces. The Figure 6 shows the material distributions of ICW
deterministic design problem minimizes the structural structure after determmistlc and probabilistic optim-
weight under static loads representing subsonic air izations. It can be seen that the probabilistic
loads, subject to the von Mises stress constraints on optimization does not increase the design variables
all the 158 elements, with a tensile stress limit of uniformly from the deterministic optimization. Some
67 ksi, compressive stress limit of 57 ksi and shear member thicknesses are increased, and some are
stress limit of 39 ksi. Quadrilateral and triangular decreased. This redistribution may increase the
membrane elements representing the corresponding structural weight slightly, but the reliability increased
upper and lower skins are linked, resulting in 32 significantly.
global design variables. The shear panels representing
each spar and rib are linked, giving 11 global design SUMMARY REMARKS
variables. All rods are linked as one additional
variable for a total of 44 global design variables. 30 In this paper an attempt has been made to
active constraints are selected among 158 stress incorporate the recently developed adaptive approxi-
constraints. After six redesign cycles, the optimiz- mations based stochastic analysis and optimization
ation phase gives the minimum weight of 76.35 lb. capabilities in ASTROS. This feature improves
In reliability optimization, all 44 global design the reliability of optimal designs subjected to the
variables are considered as random variables, with uncertain nature of the input information. The
the reliability requirement of 0.999 (fl > 3.09) for calculation of structural safety index is itself an
iterative process within the loop of structural
optimization. The ASTROS executive sequence has
been modified to include an external procedure which
repeatedly uses ASTROS for the safety index and
its gradient calculation with respect to random
parameters. Also, ASTROS provides flexibility for
defining new constraints on structural failure
probability. This paper accomplished the transfer
Fig. 5. Intermediate complexity wmg. of technology from the basic research work of
 ASTROS 745

Deterministic Optimization

Fig. 6. Thickness distributions of ICW.

developing efficient reliability analysis and optimiz- 5. E. Nikolaidis and J. Strout, Reliability-based optimiz-
ation: a proposed analyticalexperimental study. In:
ation to a general purpose multidisciplinary program
AIAA-94-1446-CP, 35th AIAA/ASME/ASCE/AHS/
such as ASTROS. F’reliminary results showed that for ASC Structures, Structural Dynamics, and Materials
a small percentage increase in structural weight from Conf., Hilton Head, SC, 18-20 April, Part 2,
the deterministic designs, the reliability can be pp. 1105-11 I4 (1994).
improved significantly. Future works needs to 6. 1. Enevoldsen and J. D. Sorensen, Reliability-based
optimization in structural engineering. J. struct. Safety
address the implementation of the system reliability 119, 1069-1084 (1993).
issues. 7. M. V. Reddy, R. V. Grandhi and D. A. Hopkins,
Reliability based structural optimization: a simplified
Acknowledgements-bdr Ray Kolonay and Dr V. B. safety index approach. J. Comput. Struct. 53,
Venkayya of Wright Patterson Air Force Base, Dr Robert 1407-1418 (1994).
Canfield of Air Force Institute of Technology and MS 8. L. P. Wang, R. V. Grandhi and D. A. Hopkins,
Geetha Bharatram and Dr Liping Wang of Wright State Structural reliability optimization using an efficient
University greatly assisted in implementing the reliability safety index calculation procedure. Int. J. numer. Meth.
analysis capabilities in ASTROS. This research work was Engng 38, 1721-1738 (1995).
sponsored by Wright Patterson Air Force Base through the 9. R. Rackwitz and B. Fiessler, Structural reliability under
Contract F33615-94-C-321 I. combined load sequences. Comput. Struct. 9, 4899494
(1978).
IO. B. M. Kwak and T. W. Lee, Sensitivity analysis for
REFERENCES reliability-based optimization using an AFOSM
method. Comput. Struct. 27, 399406 (1987).
1.H. 0. Madsen, S. Krenk and N. C. Lind, Methods of II. D. J. Neil1 and D. L. Herendeen, ASTROS Enhance-
Structural Safety. Prentice-Hall, Englewood Cliffs, NJ ments In: ASTROS User’s Manual, Vol. I, WL-TR-93-
(1986). 3025 (1993).
2. C. Chamis and C. Shiao, IPACS-integrated probabilis- 12. D. J. Neill, D. L. Herendeen and R. L. Hoeslv.
tic assessment of composite structures: code develop- ASTROS Enhancement. In: ASTROS Programmer’s
ment and application. In: Third NASA Advanced Manual, Vol. II, WL-TR-93-3038 (1993).
Composites Technology Conf., Long Beach, CA 13. T. Y. Torng and R. J. Yang, An advanced reliability
(1992). based optimization method for robust structural system
3. Y. T. Wu, H. R. Millwater and T. A. Cruse, Advanced design. In: AIAA-93-1443. 34th AIAAIASMEIASCEI
probabilistic structural analysis method for implicit AHSIASC Structures, Structural Dynamic.s and Ma:
performance functions. AIAA J. 28, 1663-1669 (1990). terials Conf., La Jolla, CA. pp. 1198-1206 (1993).
4. L. P. Wang and R. V. Grandhi, Efficient safety index 14. R. V. Grandhi and V. B. Venkayya, Structural
calculation for structural reliability analysis. J. Comput. optimization with frequency constraints. AIAA J. 26
Struct. 52, 103-I I I (1994). (7), 858-866 (1988).