Vous êtes sur la page 1sur 15

Multilevel optimization in aircraft structural design evaluation

Lars U. Hansen
*
, Peter Horst
Institute of Aircraft Design and Lightweight Structures (IFL), TU Braunschweig, Hermann-Blenk-Str. 35, 38108 Braunschweig, Germany
Received 21 December 2006; accepted 2 May 2007
Available online 22 June 2007
Abstract
Changes of the structural design are one of the essentials in the investigation and optimization of alternative and innovative aircraft
concepts. In this paper the choice of an appropriate design is seen as an optimization problem and solved by the application of a mul-
tilevel optimization procedure based on detailed Finite Element models of certain structural parts. The design is variable by its principle
layout, its material and its dimensions.
At the top level an Evolution Strategy drives the topology parameters. The second level of the optimization procedure is based on the
deterministic, gradient-based optimization method of MSC.Nastran

Sol200 that is used to optimize thicknesses and cross-sections of the


model with respect to dierent design constraints. The model generation is based on Patran PCL-routines. A parallel evaluation is used
to increase calculation speed.
Two examples are presented in this context. The rst one shows advantages of the multilevel approach in simultaneous sizing and
topology optimization of a generic framework structure. The second example is an application to a structural design optimization of
a Blended Wing Body aircraft fuselage structure. This example shows the full scope of the method by consideration of metal and com-
posite designs in single, double and sandwich layout under multiple load and constraint conditions.
2007 Elsevier Ltd. All rights reserved.
Keywords: Multilevel optimization; Structural optimization; Evolution Strategy; Evolutionary algorithms; MSC.Nastran

Sol200; Blended Wing Body


(BWB); Aircraft
1. Introduction
The eciency and environmental friendliness are of high
importance for future aircraft designs. In particular, the
structural weight has a large inuence on the overall per-
formance of an aircraft which has in turn a major impact
on emissions and fuel consumption.
Especially in aircraft design studies of unconventional
congurations a weight estimation is extremely dicult
but essential to rate a concepts potential and quality (cp.
with work on ying wings by [9]). The simplest weight esti-
mation methods often rely on empirical relations and are,
as a consequence, based on existing aircraft. An improve-
ment of such methods can be achieved by adding a physical
representation of the aircraft and its loads to the method,
either by creating relatively simple physics-based models
or by more detailed representations with global Finite Ele-
ment models, which again dier in their delity and quality
by the way loads are comprised (cp. with work presented in
[3,18,19], and especially [15]). In either case the denition
of principal properties of the structure relies on the know-
ledge related to the detailed layout of the structure. Often
this knowledge is obtained from prior aircraft generations
and their structural layouts.
In order to evaluate unconventional aircraft congura-
tions (Blended Wing Body, Oblique Flying Wing, etc.),
where it is likely that structural designs will be dierent
from todays standard, as investigated by Hansen et al.
[7,8] and O

sterheld et al. [14], an optimization task should


be performed to nd a suitable and light structural layout,
respectively design, and to link its properties (e.g. sti-
nesses, specic weights) to the global weight prediction
modules.
0045-7949/$ - see front matter 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compstruc.2007.05.021
*
Corresponding author. Tel.: +49 5313919919; fax: +49 5313919904.
E-mail address: L.Hansen@tu-bs.de (L.U. Hansen).
www.elsevier.com/locate/compstruc
Available online at www.sciencedirect.com
Computers and Structures 86 (2008) 104118
To cope with this situation an approach has been chosen
where loads and geometry data are extracted manually
from representative parts of a global Finite Element model
(cp. Fig. 1).
This information provides the basis for a detailed, para-
metric model that is capable of covering a large range of
possible structural designs (dierent layouts, materials,
proportions, etc.). Many combinations of design parame-
ters can be found that fulll its design requirements.
The need for an optimization approach is obvious and
due to the lack of information about the shape of the
design space an algorithm should be chosen that is capable
of nding global optima. An increase of eciency is achiev-
able by separating this task into two optimization levels:
the rst level uses an Evolutionary Algorithm due to the
capability of nding global optima.
The second level has been realized by implementing a
deterministic optimization method, driven by gradient
information, which is readily provided by the Solution
200 of the commercial Finite Element Code MSC.Nas-
tran

(v2005). The set of design variables is divided into


a group of variables describing the principal concept layout
and a second group of variables that aect the proportions
of the model. This partition leads to a topology design
optimization task performed by the Evolution Strategy
and a second level optimization task that determines an
ideal and feasible combination of thicknesses and cross-sec-
tions to fulll design requirements and constraints. The
result of the second level is handed to the rst optimization
level through the denition of a tness function.
The choice of an Evolutionary Algorithms for the rst
optimization level is based on dierent reasons: one advan-
tage of most Evolutionary Algorithms is their capability of
nding global optima in ragged design spaces and handling
many design variables without a signicant increase in
computation time. The advantage of dealing directly with
real valued design variables without encoding was decisive
for applying an Evolution Strategy. Furthermore the imple-
mentation of a lifespan into the optimization allows to
enforce somehow robust design solutions: slightly distorted
design variables need to create reproducible good results in
order to remain longer than a single lifespan in the popula-
tion. Consequently optimum solutions, that resulted only
from numerical eects, will be removed without perma-
nently misleading the optimization.
This paper is structured as follows. In Sections 2 and 3
the optimization approach and the parametric models are
presented. Section 4 and 5 present dierent applications
of the method. Section 4 places the emphasis on comparing
the multilevel with a single-level approach. In Section 5 the
multilevel approach is examined in application to a model
with high complexity and large changes in shape and topol-
ogy during design optimization. The evaluation ends with a
conclusion and summary in Section 6.
Nomenclature
r stress
s shear stress
q density
l population size
m Poissons ratio
j life span
k ospring per generation
d
i
mapping function
g inequality constraint function
h equality constraint function
n number of design variables
n
h
/n
j
number equality/inequality constraints
n
s
number of strategy parameters
n
x
, n
y
force ux in x-, y-direction
m mass
p pressure
w weighting factor
A cross-section area
C, D shape constants
E Youngs modulus
F
i
failure index
G shear modulus
F tness function
K stiness matrix
u displacement vector
P load vector
X design variable vector

registered trademark
Indices
c compression
t tension
11, 22 principal direction 1, 2
vMises equivalent von Mises stress
allow allowable
Fig. 1. Global FEM model of a Blended Wing Body aircraft [8].
L.U. Hansen, P. Horst / Computers and Structures 86 (2008) 104118 105
2. Optimization approach
The objective of the optimization task is to nd a struc-
ture that has a minimum weight while fullling other
design requirements and constraints. The scope of the
design space covers dierent model topologies and topog-
raphies, each one composed of parts in dierent dimensions
and materials.
Consequently, the representation of such structural lay-
outs requires:
Design variables that describe the model topology and
topographie (e.g. number of stringers, existence of addi-
tional stieners, curvature radii).
Design variables that inuence the model proportions
and, by changing the stiness distribution, the force ow
in the structure (e.g. skin thicknesses, cross-section
areas, 2D-beam heights).
The challenge in this optimization task is the correlation
between some of the design variables. For instance, a simul-
taneous variation of design variables has a dierent inu-
ence than the separate variations and their superposition.
This dependency between variables is especially obvious
when it comes to variables that are more or less parameters
inuencing the model layout: e.g. a parameter adding addi-
tional tension struts to the structure will change the force
ow within the structure in such a signicant way, that gra-
dient information of preceding calculations will become
useless (the sensitivity of the objective and the constraint
functions will be completely dierent). Consequently, all
methods purely relying on gradient information, like a
response surface method (RSM), might lead to a very inef-
cient or even diverging optimization process.
A dierent approach has been taken by Raq and Wil-
liams [16] by interfacing a trained neural network and a
Genetic Algorithm in order to increase the eciency of
the optimization approach. Dierent, representative points
in the search space are chosen for evaluation by Finite Ele-
ment analysis and used to train the neural network. All fur-
ther tness evaluations can then be based on the neural
network without the need of performing time consuming
Finite Element analyses. The avoidance of detailed analy-
ses is advantageous as long as the answers of the neural
network are reliable: strong interference of dierent design
variables might be challenging for such an approach and
increase the number of relevant training cycles of the neu-
ral network.
This paper presents another ecient way to solve the
structural optimization task by dividing the complete task
into two separate optimization problems: a sizing task
and a layout (or topology) task (see Fig. 2). This multilevel
optimization (or maybe better referred to as hybrid strat-
egy) has a separate set of design variables for each level.
This leads to a certain type of optimization strategy, where
the rst level optimizes the principle layout by creation of
multiple sub-tasks that are optimized by a second level.
This separation in two levels is, due to its treatment of
every structural layout as a separate sizing optimization
task, by some means comparable to the mechanisms of a
multi-population approach in an Evolutionary Algorithm.
The lower optimization level the sizing task can be
solved very eectively by making use of gradient informa-
tion, that can be determined eciently by calculating the
sensitivities of the system stiness matrices. Such a process
is included in the optimizer of Nastran.
2.1. Evolution Strategy (level 1)
Evolution Strategies (ES) and Genetic Algorithms (GA)
belong to the group of Evolutionary Algorithms (EA) that
all have a common base Darwins theory of evolution: a
group of solutions, each one dened by their genes (geno-
type) and referred to as individuals, is put under selection
pressure and either reduced by the weakest or selected for
recombination by its ttest. This idea of survival of the t-
test is simulated over multiple generations in order to
determine the best solution.
The simulation requires parameters that describe the
model and that characterize each individual. These param-
eters are coded in genes that describe each individuals
behavior. In the case of the GA these genes are represented
by binary encodings of the parameters and represent the
genotype. A mapping function translates the genotype to
the phenotype (e.g. real-value model parameters). In ES
the genotype is normally equal to the phenotype which
means that every real-valued parameter of the optimization
model can be, without mapping functions, part of the gen-
ome. In addition to the object variables each individual
possesses in ES additional strategy parameters, dening
the standard deviations (variances) and rotation angles
(covariances). The object variables and the strategy param-
eters are changed by mutation and recombination and
hence are part of the optimization problem.
All EA possess certain and specic ways to recombine,
mutate and select individuals in order to simulate multiple
Fig. 2. Multilevel optimization task.
106 L.U. Hansen, P. Horst / Computers and Structures 86 (2008) 104118
generations of evolution. Two major branches of Evolu-
tionary Algorithms were developed independently from
each other:
The Evolution Strategies were established in 1964 ini-
tially to support experimental work on aerodynamic
panel models (cp. [20]). The strategy was used to inte-
grate each obtained result in a population and to nd
new, and better combinations of parameters for the suc-
ceeding experiments. The strategy was able do deal
(without encoding or mapping) directly with the param-
eters of the model. This rst approach showed the
potential of the strategy although it lacked details of
todays ES. Many improvements to the method were
contributed by Schwefel [21].
The Genetic Algorithms were introduced in the 1970s by
Holland (see [6]). Their binary encoding is an abstract
representation of the real-valued parameters. The choice
of an adequate encoding strategy is an additional chal-
lenge and might contribute to the methods performance.
Due to the various dierences in the EA, this paper deals
only with one implementation of an ES with a l, j, k strat-
egy where the parameters l denote the size of the parent
population, k the number of ospring and j the lifespan
of an individual. This representation includes the following
selection mechanisms (comprehensive overviews are given
in [2,22]):
The l + k ES selects out of the ospring and parent
population. This elitist selection allows parents to survive
until they are superseded by better performing ospring.
In this case it might be possible that well performing
individuals survive forever.
The l, k ES selects only out of the ospring population.
The so called non-elitist strategy replaces in every gener-
ation all parents by their ospring.
The l, j, k ES includes both mechanisms by introduc-
ing the lifespan as an additional parameter. With a life-
span of j = 1 the method is equal to the l, k-strategy
and for j = 1 it equals the l + k-strategy. Intermediate
strategies are possible for 1 < j < 1.
The selection mechanisms dier in the selection pressure
exerted on the population. Depending on the problem type,
dierent arguments for the choice of an adequate life span
can be given:
j = 1: Ref. [1] describes the l, k ES as the state-of-the-
art approach since this mechanism does not necessarily
require to keep the best individual, which means that
the selection is non-elitist. Temporarily deteriorations
become possible that might be necessary to leave the
local optimum to search for a region of similar or higher
attraction.
j = 1: If the feasibility, the objective function or the
object variables are subject to noise an elitist selection
strategy can get stuck at an outlier. The survival of an
outlier over all generations reduces the possibilities to
explore regions of higher attraction. This is especially
a problem in experimental settings (cp. [17]). On the
other hand, the elitist selection of the l + k ES is
advantageous for individuals with small numbers in
their strategy parameters (e.g. small variances), due to
their reduced capability to introduce large distortions
to the objective variables.
In the examples of this investigation the individuals t-
ness is always based on results obtained by the FEM. In
order to avoid improvements due to numerical eects, it
was decided to treat the results like an experimental result
where outliers and noise have to be expected. Therefore a
non-elitist l, j, k ES with j = 5 has been chosen as an
adequate approach for the structural problems presented
in this context.
The principal operations of an ES are, as shown in
Fig. 3, performed in multiple steps:
(1) denition and creation of an initial population with l
individuals;
(2) creation of k ospring;
random selection of parents out of an initial
population,
recombination and mutation to create a new indi-
vidual (ospring),
evaluation of individuals tness (e.g. FEM based
structural analysis, experiment),
integration of individual in population,
(3) reduction of the population size by age (life span j);
(4) selection of the l ttest individuals;
(5) convergence check (either restart in step (ii) or
nished).

o
, +)
a
o
o
o
n
Fig. 3. General setup of an Evolution Strategy.
L.U. Hansen, P. Horst / Computers and Structures 86 (2008) 104118 107
In order to give a short introduction to the Evolution
Strategy, the principal elements and operators of the ES
are explained in brief.
2.1.1. Ospring population
The ospring population is created by recombination
and mutation. The number of individuals in the ospring
population is k. Each ospring is characterized by its gen-
ome, age and tness.
2.1.2. Parent population
The parent population is formed by the best l individu-
als out of the total number of ospring and the previous
parent population (depending on the chosen selection
mechanism).
2.1.3. Convergence check
The convergence check is performed by either evaluating
the improvement of the tness or by evaluating the changes
in the design variables.
2.1.4. Recombination operator
The genetic information of two (or more) parent indi-
viduals is used to produce one ospring by recombination
(the successor is hence called recombinant). The recombina-
tion (in GA as well referred to as crossover) can be either of
discrete (a dominant recombination, where each variable is
selected randomly from either parent) or intermediate type
(a recombination, where the variables are created by the
arithmetic mean of the parents variables). Another princi-
pal distinction can be made between global or local recom-
bination: the local recombination performs a choice of
parameters after the parents have been chosen while the
global recombination chooses parents for every gene sepa-
rately. In addition, several other recombination operators
exist (for details see [21]). In the following only local
recombination will be considered.
2.1.5. Mutation operator
The mutation operator adds distortion to the individ-
uals design variables. The amount of distortion, the muta-
tion strength, is chosen randomly with a probability given
by a Gaussian distribution. The standard deviation of the
normal distribution and correlation information are saved
separately in so-called strategy parameters. The correlation
between the variables is saved in the form of correlation
angles. The total number of strategy parameters n
s
for n
design variables is
n
s
n
variance
n
covariance
1
with
n
variance
n 2
n
covariance

nn 1
2
: 3
2.1.6. Selection operator
The selection pressure can be applied in multiple phases
of the process. The ES normally selects in each generation
l individuals by their tness. In addition to the selection by
tness, a life span enables to change between dierent strat-
egies (see above the description of the l, j, k strategy).
2.1.7. Fitness function
The tness of an individual is the measure for its quality.
Depending on the optimization setup, the tness function
can contain the objective and the responses of the system.
Dierent penalization approaches exist which include an
appropriate design constraint balance between intermedi-
ate violations and the selection pressure relief.
2.2. Gradient-based sizing (level 2)
The second level of the optimization task is performed
by MSC.Nastran Sol200. The optimization approach
implemented in Nastran Sol200 is based on gradient infor-
mation, that is obtained via an internal approximation of
the structural responses. The optimization problem is
dened by the search for optimal design variables X that
minimize the design objective function F(X)
min F X 4
subject to n
g
inequality and n
h
equality constraints
g
j
X 6 0 j 1; . . . ; n
g
n
g
inequality constraints 5
h
k
X 0 k 1; . . . ; n
h
n
h
equality constraints 6
and n
si
side constraints
x
lower
i
6 x
i
6 x
upper
i
i 1; . . . ; n
s
7
where the design variables x
i
are dened by X
X fx
1
; x
2
; . . . ; x
n
g: 8
The optimizer implementation in Nastran, as described in
[12], is based on the DOT optimization algorithms from
Vanderplaats Research & Development, Inc. The complete
optimization process is performed in multiple steps:
(1) Starting with an initial design a Finite Element Anal-
ysis (FEA) of the model is performed and all
requested design responses are calculated (e.g. mass,
displacements, strains, stresses).
(2) Constraint screening is performed in order to distin-
guish between active and inactive constraints. A
two stage process rst sorts out all inactive con-
straints by applying a threshold level to the normal-
ized constraints. The second stage tries to localize
violated regions within the model in order to sort
out local exceedings. The reduction of the number
of constraints enhances the performance of the sensi-
tivity analysis.
(3) Sensitivities are the rate of change of the responses
with respect to changes of chosen variables or param-
eters. Nastran Sol200 includes a design sensitivity
108 L.U. Hansen, P. Horst / Computers and Structures 86 (2008) 104118
analysis that evaluates sensitivities for all retained
responses, regardless of the future response use (e.g.
as a design objective or design constraint). Dierent
approaches are available to calculate sensitivities for
static and dynamic responses.
(4) Based on the sensitivity information an approximate
model is created that provides information for the
optimizer.
(5) The optimizer performs the numerical optimization
task and determines the search direction and the step
size. Dierent algorithms, modied method of feasible
directions (MMFD), sequential linear programming
(SLP), sequential quadratic programming (SQP),
are available but their detailed description goes
beyond the scope of this description. In the following
the MMFD is used as the default method.
(6) Convergence is checked based on multiple criteria.
The optimizer is able to distinguish between dierent
types of convergence: unachieved convergence (e.g.
due to a limitation on the number of design cycles),
non-achievable convergence (impossible design), soft
convergence (no signicant change in the variables
or properties of two consecutive approximations) or
hard convergence (convergence in the objective func-
tion calculated by exact analyses).
Especially the performance of the approximation can be
crucial for the design optimization. The basic idea behind
Nastrans implementation is the approximation of the
response by an approximation with Taylor series. The gradi-
ents required to build the series are determined in most cases
by (central or) forward nite dierences without the need to
invert the stiness matrix more than once during each design
iteration: Writing the general form of the system equations
Ku P 9
the derivative of Eq. (9) with respect to design variables X
gives
K
ou
ox
i

oP
ox
i

oK
ox
i
u: 10
Eq. (10) shows, that the rate of change of the displacements
u with respect to changes of the ith design variable x
i
can be
calculated by solving only the right-hand side. The stiness
matrix K is already available in decomposed form as a re-
sult of the usual inversion of the system equation (Eq. (9))
and hence can be reused. This requires only the calculation
of the right-hand side derivatives (also known as pseudo-
loads). In many cases the load vector P is invariant with re-
spect to the design variables (e.g. material properties,
dimensions if structural inertia loads are neglected)
oP
ox
i
0: 11
In this case the right-hand side is simplied and only the
sensitivity of the stiness matrix
oK
ox
i
u 6 0 12
remains unknown. The left-hand side of Eq. (12) can be
approximated with a nite dierence approach.
3. Simulation model
3.1. Parametric model
The simulation models presented in this paper are based
on Finite Element models. These models are created with
the commercial pre- and postprocessor MSC.Patran

by
a routine written in PCL (Patran Command Language).
The user can access the routine via a graphical user inter-
face (GUI) or via command line and ASCII-les.
To increase the model generation speed, Patran has been
run in command line mode (batch mode). In this case the
parameters are provided by an ASCII-le input; sup-
pressed graphics processing also increases the model gener-
ation speed.
3.2. FE-model
The Finite Element models are created in Patran and
translated to MSC.Nastran

format. In this case the inter-


nal optimizer of Nastran provides an ecient way to imple-
ment a gradient-driven optimization in the multilevel
approach (Sol200). Depending on the optimization task
dierent Nastran solution sequences (e.g. linear static
(Sol101), linear buckling (Sol105), etc.) are chosen.
3.3. Evolution Strategy: EStruct
The Evolution Strategy used for this investigation is an
in-house tool called EStruct fully written in Python and
based on methods proposed by Schwefel [21]. The program
makes use of the object orientated features of Python by
handling each ospring as a separate object of a common
class. This basic idea leads to a transparent code that is
easy to adapt. The current version features possibilities to
restart from an earlier run, to distribute the jobs on multi-
ple machines (nodes) and to calculate the tnesses exter-
nally (e.g. Nastran). The main procedure is described in
the following pseudo-code:
> start process via MPI-Python on multiple
machines (e.g. cluster)
> chose master node to steer the job
> put slave nodes to waiting condition
> master node: restart or initialize individuals
of the rst generation
> do while generation < max. generations
> do while (individuals in generation) < k
> initialize individual: birth date (integer),
parameter names (string array), parameter types
(integer array), parameter values (real array),
deviation values (real array), correlation angles
(real array)
L.U. Hansen, P. Horst / Computers and Structures 86 (2008) 104118 109
> master node: distribute individuals via
MPI-Python to slave nodes
> slave nodes: receive individuals and
start calculation of their tness (e.g.
Finite Element Analysis)
> slave nodes: return of individuals status (fail-
ure/success) and fitness to master node
> slave nodes: wait for new job
<end do
> master node: integrate all k new individuals in
population
> master node: perform selection (l, j, k)
> master node: create individuals for the next
generation by choosing parents, recombination and
mutation
< end do
4. Optimization example 1: generic framework structure
4.1. Analysis model and optimization setup
This section introduces the analysis models for the
design test case of a framework optimization. The struc-
tural model is chosen for explanation and verication pur-
poses due to its simplicity (low number of elements, simple
loading condition).
Two dierent approaches to optimize the location of the
nodes and the cross-sections of the framework structure
will be shown. The goal of the optimization is in both
cases the lightest structure. As a design constraint the
members of the framework are not allowed to exceed a
maximum stress level of 100 MPa in compression or ten-
sion (r
i
6 j100 MPaj for i = 1, 2, . . ., n
elements
).
In the following, the rst optimization approach is
called case A: ES+Sol200. It is based on the two level
strategy, as described before, that uses at bottom level Nas-
trans sizing capabilities (gradient-based method) and at
the top level the Evolution Strategy EStruct.
The second approach, herein after referred to as case
B: ES, which is similar to work presented by Gieger
and Ermanni [6], uses only an Evolution Strategy (EStruct)
in a single optimization level. Again, the goal is a suitable
combination of parameters that meet the design require-
ments and improve the design.
4.2. Parameterization and design variables
The Finite Element model of the framework (cp. Fig. 4)
consists of 13 nodes and 22 struts. Displacement boundary
conditions are applied to the two nodes on the left and a
vertical force acts on the right end of the framework.
Each nodal position is dened by two design variables
that are related to the two coordinates (x, y). The nodes
related to boundary conditions and all nodes on the lower
side of the framework are not subject to changes and
remain unchanged during optimization. Hence only the ve
positions of the upper nodes (as shown in Fig. 4) are
included in the parametric model and dened by the design
variables 110.
Each strut of the model is idealized by a Nastran
CROD-Element and characterized by a stiness related
to the area of its cross-section. In order to meet the design
constraints the cross-section areas are dened as design
variables and subject to change during the optimization.
Depending on the chosen approach the design variables,
constraints and the tness function are dierent. Details
are given in Table 1, Sections 4.3 and 4.4.
For comparability purposes the model generation is,
despite the simplicity of the model, performed with Patran
in batch mode. In order to use Nastrans internal optimiza-
tion sequence Sol200, additional job parameters have been
dened: they include the design responses, the design objec-
tive and the design constraints. The 32 design variables x
i
are summarized by X
X fx
1
; x
2
; . . . ; x
5
; y
1
; y
2
; . . . ; y
5
; A
1
; A
2
; . . . ; A
22
g
with
x
i
x-coordinate of node i,
y
i
y-coordinate of node i,
A
j
cross-section area of strut j.
Fig. 4. Framework optimization test case.
Table 1
Dierences in design variable denition depending on optimization approach
Approach Design variables Fitness function Design constraints
Case A: ES+Sol200 10 (ES) + 22 (Sol200) d(m) Fullled by Sol200
Case B: ES 32 (ES)

d(m) + d
i
(r) Incl. in tness function
110 L.U. Hansen, P. Horst / Computers and Structures 86 (2008) 104118
4.3. Design constraints
The task of the design is to minimize the weight subject
to n
g
inequality constraints (g = 1, 2, . . ., 22 struts)
g
j
X 6 0 13
with
g
j
X
F
j
A
j

100 MPa jr
j
j 100 MPa 6 0
for j = 1, . . ., n
struts
with n
struts
= 22.
The (axial) stress in each strut results from tension or
compression. For demonstration purposes the same stress
level has been set for tension and compression.
4.3.1. Case A: ES+Sol200
The gradient-based optimization, included in the
approach of case A: ES+Sol200, accounts for the stress
constraints by introducing Nastran DRESP1 and DCONSTR
cards in the Sol200 bulk deck. Internally theses constraints
become part of the objective (Lagrange) function and con-
sequently of the sensitivity approximation process.
1
4.3.2. Case B: ES
In the optimization setup of case B: ES the stress con-
straints have been included in the tness function by the
use of mapping relations (as shown by Gieger and Ermanni
[6] and Ko nig [11]). They allow to include all design
responses in a common tness function. This leads to a
slightly dierent optimization objective where the tness
is not only dened by the mass but also by the confor-
mance with design constraints. Consequently, the optimi-
zation aims on minimizing the function f(X)
f X

n
i1
w
i
d
i
dmass 14
with w
i
being weighting factors (here set to 1) and d repre-
senting the results of dierent design responses. These re-
sponses are translated by mapping functions in a way
that a value of zero corresponds to the desired condition.
Unwanted conditions (e.g. stress violations, excessive
weight, large displacements) create a response of d
i
> 0.
The mapping functions are normally dened in the range
of 0 6 d
i
6 1 to avoid unequal treatment or unwanted
weighting of the dierent responses.
Mapping functions, presented in [6], have been adapted
to the given problem as follows.
4.3.2.1. Mass response mapping function. The mass map-
ping function links the structural mass to an articial t-
ness value that is attractively low for desired mass range
(m
d
) and increases exponentially for masses higher than
the upper, initial mass (m
i
).
dmass a mass b
a
15
The factors a, b are calculated prior to the optimization by
using an estimated desired (m
d
) and undesired (m
u
) mass to
create answers of the mapping function of zero and one.
The exponent a has been set to a = 5 (cp. [6]).
A function has been chosen that penalizes masses above
a maximum level exponentially by translating the mass
response in tness values >1. In this case the tness result-
ing from the mass response is very likely exceeding all other
addends to the overall tness (see Eq. (14)). Consequently
the mass is desired to have a large impact on the quality of
the design.
4.3.2.2. Stress response mapping function. The stress
response is translated by a dierentiable step function to
a tness response d. The following relations have been
used:
d
i
r 1 e
k
r
r
allow
D
_ _
_ _1
16
for i = 1, . . ., n
struts
with n
struts
= 22 and
k
1
C
feas
ln
1
D
limit
1
_ _
ln
1
D
feas
1
_ _ _ _
17
and
D
1
k
ln
1
D
limit
1
_ _
1: 18
The shape of the stepping function is inuenced by shape
constants D
feas
, C
feas
, C
limit
and D
limit
. These values have
been set to D
feas
= 0.5, C
feas
= 0.1, C
limit
= 1.0 and D
limit
=
0.01 (cp. [6]).
The stress response mapping function creates tness val-
ues in the range between 0 6 d
i
6 1. The penalization of the
structure is thus limited to two prescriptive limits: fulll-
ment or violation of the constraint. Since all members of
the structure contribute to the overall tness the inuence
of every member is bounded by the denition of the map-
ping relation.
4.4. Design objective/tness
The objective of the design is to minimize the weight of
the structure. Due to the dierent approaches the objective
is treated dierently. The multilevel optimization with Nas-
tran Sol200 fullls the design constraints implicitly in the
second level so that the tness value of the each individual
is only dened by its mass.
4.5. Results
The results of the dierent approaches are presented in
Table 2. Both approaches have been used to run an
1
This design constraint proved to be challenging: the gradient driven
optimizer tries to fully utilize both allowable stress levels (in tension and in
compression). Hence gradients might be calculated that point in opposite
directions. The proper adaption of parameters inuencing Nastrans
constraint screening capabilities was necessary to create a feasible
optimization problem.
L.U. Hansen, P. Horst / Computers and Structures 86 (2008) 104118 111
estimated number of 150,000 Finite Element Analyses
(FEA). Both methods are able to create results that meet
the design constraints. But the multilevel approach leads
to a much lower nal weight (nal framework shown in
Fig. 5).
In order to understand this dierence in performance the
convergence behavior is depicted in Figs. 6 and 7. Both dia-
grams show the tness value vs. the number of function
evaluations (FEA). The multilevel approach shows from
the beginning a much better capability to nd a suitable
combination of parameters and hence a lower structural
weight.
The reason is given by the dierent sizing behavior of
the dierent approaches: keeping in mind that the rst 10
design variables change the topology of the structure one
can imagine, that an ideal combination of the nodal coor-
dinates will not necessarily lead to an overall low tness if
the design variables 1132 (cross-section areas) are improp-
erly chosen. The multilevel approach has a principle advan-
tage over the single-level Evolutionary Strategy since the
lower level (Sol200) is able to determine the ideal combina-
tion of cross-section variables. The tness value that is
passed back to the rst level optimizer (ES) clearly rates
the quality of the nodal positions. Even after performing
over 1500 generations with 100 ospring per generation
no signicant improvement of the nal result of case B:
ES was achieved. Although changes to the optimization
parameters might help to improve this performance it
becomes obvious that the multilevel approach of case A:
ES+Sol200 has signicant advantages over the single-
level approach of case B: ES.
The multilevel approach requires a much lower number
of generations for the convergence to be achieved. For fur-
ther investigations a number of 25 generations each with
100 ospring has been chosen.
5. Optimization example 2: BWB fuselage structure
5.1. Analysis model
The investigation of dierent model parameters requires
a common tness indicator in order to rate dierent con-
cepts. The following investigation uses the structural mass,
resulting from the structures density multiplied by its vol-
ume, directly as a tness value. Taking into account
the results from the rst investigated example of the
Table 2
Dierences in design results depending on optimization approach
Approach Max. stress
(MPa)
Final
mass (kg)
Exact FEM analyses
ES 100 94.5 kg 1500 gen 100 ospring
(=150,000 FEA)
ES+Sol200 100 62.5 kg 75 gen 100 ospring SOL200
(150,000 FEA)
Fig. 5. (Top) Final design after sizing task with xed nodal positions;
(bottom) nal framework design after approach of case A: ES+Sol200.
Fig. 6. Convergence behavior: Evolution Strategy (model topology
optimization)+Nastran Sol200 (sizing optimization).
Fig. 7. Convergence behavior: Evolution Strategy (model topology and
sizing optimization).
112 L.U. Hansen, P. Horst / Computers and Structures 86 (2008) 104118
framework optimization case, a two level strategy is cho-
sen: the investigation of dierent model parameters is han-
dled by an Evolution Strategy. The tness of each of these
models is determined by sizing with the optimizer of Nas-
tran Sol200. This procedure enables a fast convergence
(cp. Fig. 7).
The investigated models represent a typical part of the
fuselage structure of a Blended Wing Body aircraft, which
oers a variety of general solutions, but is not yet solved.
The fuselage body, as shown in Fig. 8, encloses a wide
cabin that is partitioned by vertical walls reaching from
the lower shell to the upper shell of the fuselage. In order
to be able to operate in high altitudes, pressurization of
the cabin is necessary. The pressurization of the at cabin
creates an unfavorable loading condition that is dominated
by large bending loads in the upper and lower skin of the
fuselage and large non-linear eects, if no special measures
are taken. The advantages of a BWB aircraft strongly rely
on the fuselage weight. In order to rate detailed structural
designs and to create specic weights (as well for further
investigations of such aircraft) here only two dominant,
pressurization load cases are used for demonstration pur-
poses of the method.
The analysis models represent a pressurized structure of
the fuselage upper shell. By limitation to symmetric mod-
els, a reduction of calculation time (investigation of half-
models) is possible. These models are created with a PCL
routine and consist of 10,00015,000 elements, depending
on the model topology (single skin, double skin, sandwich,
etc.). A single skin design for multiple bays is depicted in
Fig. 9. Only the part marked with dashed lines, the half-
model of the structure, is considered during optimization.
Symmetry conditions are applied to the edge in the symme-
try plane. The major components of the model are labeled.
Models of a double skin or a sandwich are similar and only
dier in the presence of an inner skin, which is stiened by
stringers and attached to the inner frame anges. The sand-
wich structure has in addition solid elements between inner
and outer skin (core height: four solid elements) while
stringers on the inner or outer skin are omitted for this
case.
Two dierent load cases will be considered for this opti-
mization task. Based on FAR $25.365 [5] a fatigue load
case, with recurring loads at limit load level, and a failure
case at ultimate load level have been chosen. The limit load
case assumes a loading condition with a normal, opera-
tional cabin pressurization of 1Dp. The failure case is based
on $25.365(d) with
1:5 Dp
limit load
1:33 2Dp: 19
The calculation of the pressure dierential
2
results from the
cabin pressurization and from the dynamic pressure of
aerodynamic loads.
During operational ight conditions the cabin static
atmospheric pressure (cabin altitude a
cabin
= 2.5 km) is
equal to
p
cabin
746:83 mbar:
At the maximum operating altitude (a
max
= 13 km) of the
aircraft the static atmospheric pressure equals
p
atmos
165:10 mbar:
Considering an additional pressure reduction of
p
aero
50:00 mbar
due to the lift acting on the upper surface of the aircraft,
the static pressure on the outside is reduced to
p
static
p
atmos
p
aero
p
static
165:10 mbar 50 mbar 115:20 mbar:
The resulting dierential pressure on the pressurized skins
of a single or double skin concept is dierent due to the
assumption of the pressure condition between the two
skins:
The loading condition of a single skin concept results
from the dierence between inside and outside pressure
Fig. 8. Location of an exemplary upper shell structure in a global BWB
model.
Fig. 9. Finite Element model topology.
2
All pressure values are taken from tables of the Standard Atmosphere
[4].
L.U. Hansen, P. Horst / Computers and Structures 86 (2008) 104118 113
Dp p
cabin
p
static
20
Dp 746:83 mbar 115:20 mbar 632 mbar:
The loading condition of a double skin concept can be cal-
culated assuming a static atmospheric pressure between in-
ner and outer skin
Dp
inner skin 1
p
cabin
p
atmos
582 mbar
Dp
outer skin 1
p
atmos
p
static
50 mbar
Dp
sum
Dp
inner skin
Dp
outer skin
632 mbar:
21
Due to the non-circular shape of the fuselage, additional
forces are applied to edges of the model. These loads have
been extracted for one panel location in the fuselage from a
global nite-element model (load case: 1Dp).
In longitudinal direction a load of n
x
= 222 N/mm and
in circumferential direction of n
y
= 192 N/mm have been
applied.
For the 2Dp case (ultimate load case) the loads of the
1Dp case are scaled.
The parametric model allows as well to change the mate-
rial models and material parameters for the pressurized
skins. Two dierent, linear elastic materials are chosen
for this investigation: isotropic, homogeneous material
(Aluminum, e.g. 2524 T3) and alternatively anisotropic,
layered composite material (CFRP). Both materials can
be combined with an optional honeycomb core material
(cp. Table 6). The material properties and sizing allowables
are summarized in Tables 3 and 4. In order to obtain a
symmetric laminate and to keep the number of design vari-
ables low, the laminate is dened with a constant stacking
sequence and layer orientation. Due to symmetry, only the
layers thicknesses on one side of the midplane are variable
(cp. Table 5).
5.2. Design variables
The design variables of the model can be subdivided into
a group of parameters inuencing the model topology
(Table 7) and sizing variables that aect the model dimen-
sions (Table 8). The total number of design variables is
dependent on the chosen model topology and material.
The herein before mentioned multilevel optimization
approach is based on the repartition of topology and sizing
variables (cp. Tables 7 and 8). Consequently, the topology
variables are handled by the Evolution Strategy and the siz-
ing variables by the gradient based optimization of Nastran
Sol200.
Unlike the rst example the design variables of this
second example do not directly inuence any nodal
coordinates. Instead variables are dened that change the
parametric design at certain control points (e.g. frame
height, skin curvature, Table 7). Ragged surfaces are con-
sequently already avoided by the denition of the paramet-
ric model without a need for further penalization of
geometrically unpractical or unwanted solutions.
5.3. Design constraints
The design variables are constrained by side-constraints
given in Tables 7 and 8 (the range is marked in columns
with min and max). These side-constraints are inde-
pendent of the load case.
Buckling is prevented by additional design constraints
on the lowest Eigenvalues. This constraint enforces rst
buckling above the corresponding load level. The buckling
constraint is only active for the limit load case.
Additional design constraints are used for the sizing of
the structure. The sizing of the structure distinguishes
between fatigue (recurring pressurization at limit load)
and ultimate (maximum pressure) load: as a consequence
the allowable stress level for all metal parts is dierent
for the fatigue case and for the ultimate case (cp. Table 9).
Fiber reinforced material oers in many cases a good
fatigue behavior which permits a dierent design approach
for CFRP parts of the structure. Hence only the ultimate
load case is used for the sizing with the Tsai-Hill failure cri-
terion. It is applied as a design constraint which takes the
residual strengths of each layer and the multi-axial stress
condition into consideration. The limit load case is again
used for design constraints on the critical buckling load.
Table 4
Unidirectional carbon ber prepreg material: anisotropic material properties and strengths
E
11
(GPa) E
22
(GPa) m
12
G
12
(GPa) G
23
(GPa) G
13
(GPa) q (kg/m
3
)
192 10.6 0.31 6.1 6.1 3.7 1.8e6
r
t11
(MPa) r
t22
(MPa) r
c11
(MPa) r
c22
(MPa) s
allow
(MPa)
2715 56 1400 250 101
Table 5
Stacking sequence of generic CFRP material
Layer no. Orient. angle () Initial thickness (mm) Design variable
1 0 0.125 t
layer1
2 +45 0.125 t
layer2
3 45 0.125 t
layer3
4 90 0.125 t
layer4
58 Symmetry
Table 3
Aluminum material: isotropic material properties and allowable fatigue
and residual stresses
E (GPa) m
12
q (kg/m
3
) r
vMises,fatigue
(MPa) r
vMises,fracture
(MPa)
72 0.3 2.8e6 110 >270
114 L.U. Hansen, P. Horst / Computers and Structures 86 (2008) 104118
The denition of the failure index is dened in Eq. (22).
The calculation of the index is performed for each element
separately. Sensitivities with respect to the failure index are
being calculated internally in Nastran Sol200 and become
part of the objective function (cp. [13]). The material allow-
ables applied for the example are given in Table 4.
1
r
t11

1
r
c11
_ _
r
1

1
r
t22

1
r
c22
_ _
r
2

r
2
1
r
t11
r
c11

r
2
2
r
t22
r
c22

s
2
12
s
2
allow
F
i
22
The existence of constraints based on element stresses and
on Eigenvalues requires dierent analyses. Within Sol200 a
linear static analysis (Sol101) and a buckling analysis
(Sol105) is performed.
5.4. Design objective/tness
The objective of the optimization is minimum weight.
Design constraints and side constraints (cp. Section 5.3)
have to be fullled.
5.5. Results
The Evolutionary l, j, k-Strategy EStruct has been used
with the settings l = 15 (population size), j = 5 (life span)
and k = 100 (ospring per generation). The optimization
was stopped after 25 generations: In total 25 100 =
2500 individuals have been modeled and analyzed in the
rst level of the optimization. In the second level each of
theses models was sized and optimized in multiple Sol200
Table 6
Sandwich core: material properties (similar to Aramid Fibre Honeycomb HexWeb

HRH-10-1/4-3.1, [10])
E
11
(MPa) E
22
(MPa) E
33
(MPa) G
12
(MPa) G
23
(MPa) G
31
(MPa) m
12
q (kg/m
3
)
1.0 1.0 145.0 44.8 20.7 1.0 0.4 4.96e10
Table 7
Design variables: topology variables
Topology variable Type Minmax Number of design variables
Inner skin Curvature r 0 mm1 1
Frame Height h 0 mm1 1
Inner/outer skin Material AluminumCFRP 1
Inner skin Existence Single skindouble skin 1
Sandwich design Existence With corew/o core 1
Tension struts Existence Yesno 1
Pole attachment Topology Y or W design 1
Rib arrangement Topology Withw/o add. ribs as web stieners 1
Inner skin curvature Shape type Cylindricspheric 1
Inner skin curvature Shape type Normalinverse 1
Table 8
Design variables: sizing variables
Sizing variable Dimension Minmax No. of sizing variables
Inner skin (CFRP) Layer thickness 0.125100 mm 4 layers (sym.) 7 regions
Inner skin (alum.) Thickness 1.2100 mm 7 regions (if single skin design chosen)
Outer skin (CFRP) Layer thickness 0.125100 mm 4 layers (sym.) 7 regions
Outer skin (alum.) Thickness 1.6100 mm 7 regions
Frame web Thickness 1.260 mm 7 regions
Frame web stiener Height 4200 mm 1 region
Rib Thickness 1.4100 mm 1 region
Inner frame ange Flange width 60450 mm 7 regions
Outer frame ange Flange width 20400 mm 7 regions
Inner skin stringers Height 55125 mm 7 (if core exists)
Outer skin stringers Height 55125 mm 7 (if core exists)
Poles Cross-section 24200 mm
2
1 region
Tension struts Area 5005000 mm
2
1 (if struts desired)
Table 9
Design constraints depending on material choice
Isotropic, metallic material (cp. Table 3)
Buckling constraint: First Eigenvalue > 1 at limit load
Stress constraints: v.Mises-stress in shells, beams and
struts <r
vMises,fatigue
at limit load
Anisotropic, composite material (cp. Table 4)
Buckling constraint: First Eigenvalue > 1 at limit load
Failure constraints: Tsai-Hill component failure index
F
i
< 1.0 at ultimate load
L.U. Hansen, P. Horst / Computers and Structures 86 (2008) 104118 115
design cycles until convergence was achieved or the maxi-
mum number of design cycles was reached (max. 35 cycles).
The second level required 20 min on an AMD
Opteron64, (2.2 GHz, 2 GB RAM) Linux system to evalu-
ate an individuals tness. By parallel evaluation on 5
CPUs 6.5 h/generation were required (<1 week for 25
generations).
Due to the total number of 2500 individuals not all of
the results can be presented in detail. As an example for
the structural sizing behavior of the second level optimiza-
tion (Sol200) a single skin CFRP model is depicted in Figs.
10 and 11. The sizing history of the models optimization
and the corresponding thicknesses of the laminates layers
are shown. All layers started the optimization with an ini-
tial thickness of 0.125 mm (cp. Table 5). In the rst design
cycle the failure criterion (Tsai-Hill) was already fullled
but the rst Eigenvalue was much too low (0.038). The
optimizer, as in Fig. 10, required 25 design cycles to meet
the design requirements, in this case the Eigenvalue of 1.0
(with a tolerance of 3%).
The nal laminate, depicted in Fig. 11 shows an
increased thickness in the circumferential bre layers (90
layer), a moderate increase in the diagonal bre layers
(45) and no signicant increase in longitudinal bre ori-
entation (0). Besides the skins all other parts of the struc-
ture have been sized as well.
The convergence behavior is depicted in Fig. 12. The t-
ness is plotted over the number of generations. Each circle
in the diagram represents the tness of an ospring (100
ospring/generation). The three curves in Fig. 12 show
the improvement of the tness over the number of genera-
tions. The tness of the best and the worst individual of
each generations parent population and the corresponding
average tness is plotted. Although the tness of the best
individual still improves slightly after 25 generations, the
average tness of the parent population remains nearly
constant. This indicates that future generations might be
subject to small changes and large tness improvements
are not expected.
Fig. 10. CFRP single skin example: Sol200 sizing history (frame height
700 mm, inner frame curvature 10 m, 2 load cases, Tsai-Hill & Eigenvalue
constraints).
Fig. 11. CFRP single skin example: thicknesses of UD-layer 14 of symmetric laminate after sizing (frame height 700 mm, inner frame curvature 10 m, 2
load cases, Tsai-Hill & Eigenvalue constraints).
116 L.U. Hansen, P. Horst / Computers and Structures 86 (2008) 104118
Overall best result. The overall best result was deter-
mined for a single skin design with a composite skin
(23.6 kg/m
2
). The optimized frame has a height of
1400 mm at the panel edges and of 300 mm in the mid-
dle of the panel. The inner ange is curved with a radius of
4 m. Additional ribs in longitudinal directions help
avoiding stability problems of the large frames.
The relevance of the material is depicted in Fig. 13. Each
sphere in the diagram represents one individual out of the
parents populations. The vertical position of the spheres
denes their tness. The color of the spheres depicts the
material of the pressurized skins of the models. All of the
best performing individuals of each generation have a com-
mon material type: the pressurized skin(s) are made of
composite materials which dier in every layer, in every
design region and for every model due to the separate siz-
ing of each layers thickness.
3
The principal evolution of the shell design, dened by
single or double skin arrangements, is shown in Fig. 14.
Again, the vertical position of the spheres represents their
tness while the color marks the design principle: single
or double skin. A trend cannot be detected. All generations
show a mixture of single and double skin designs in their
best 15 individuals. These 15 individuals are used to create
the parent population for the next generation.
All investigated sandwich designs resulted in a higher
weight. The result might be dierent in the case of an addi-
tional constraint on the maximum frame height (e.g limited
by space requirements of the cabin).
6. Conclusion
Evaluation of structural alternatives requires investiga-
tion of multiple sensitivities. In order to rate and judge cer-
tain design alternatives a method is necessary that is able to
optimize parameters and topologies while taking certain
design constraints or design requirements into account.
This task can be fullled by an Evolution Strategy that
tries to nd suitable combinations by selection and repro-
duction of the best combinations. As shown for a simple
design case of a framework structure the pure reliance on
the ES is computationally inecient, especially since ideal
topologies and poor dimension of the members of a struc-
ture lead to the same treatment in a tness function as poor
topologies and ideal dimensions. This mixture of design
responses in one tness function can be inuenced by
weighting factors. Nevertheless, the mixture of design and
Fig. 12. Convergence behavior of BWB fuselage pressurized panels.
Fig. 13. Fitness improvement vs. parent population of each generation:
colored spheres mark material of pressurized skins.
3
Here it has to be taken into account, that the material properties and
allowables strongly inuence which design performs best. For demon-
stration purposes of the method the choice of properties is done
exemplarily and may not be suitable for a nal investigation.
Fig. 14. Fitness improvement vs. parent population of each generation:
colored spheres mark single or double skin models.
L.U. Hansen, P. Horst / Computers and Structures 86 (2008) 104118 117
sizing variables in one optimization task is less ecient
than dividing both tasks into two optimization levels.
The multilevel approach presented in this paper seems to
nd ecient combinations of topology parameters by an
automated procedure. As shown in Figs. 6 and 12 a fast
convergence is achieved by considering only (previously)
sized models for the population. Such a sizing task is per-
formed by a second optimization level that very eectively
uses gradient information to nd the best search directions.
However such investigations are limited by the required
amount of calculation time and by the need for automated
mesh generators.
Future work is planed to improve the interface between
the local, detailed and the global Finite Element models. A
more or less automated transfer of loads and boundary
condition to the local level and a way of returning these
optimized shell properties to the global level is necessary.
Acknowledgement
Part of the work presented was funded by the 5th frame-
work program of the European Union during the project
VELA (Very Ecient Large Aircraft).
References
[1] Back T, Schwefel H-P. Evolutionary computation: an overview. In:
Proc. of the 1996 IEEE intl. conf. on evolutionary computation
(IECC96), Nagoya, Japan. NY: IEEE Press; 1996. p. 209.
[2] Back T, Hammel U, Schwefel H-P. Evolutionary computation:
comments on the history and current state. IEEE Trans Evolut
Comput 1997;1:317.
[3] Bradley K. A sizing methodology for the conceptual design of
Blended-Wing-Body transports, NASA/CR-2004-213016; September
2004.
[4] DIN. Norm-atmosphere, DIN-Norm 5450, Stand 1968.
[5] FAR. Federal Aviation Authorities, Part 25 Airworthiness Stan-
dard: Transport Category Airplanes; October 2006.
[6] Gieger M, Ermanni P. Development of CFRP racing motorcycle rims
using a heuristic evolutionary approach. Struct Multidiscip Optimizat
2005;30:5465.
[7] Hansen L, Horst P. Strukturkonzepte in konventioneller und
zweischaliger Bauweise fu r ache nichtzylindrische Rumpfkong-
urationen, DGLR Jahrestagung, Dresden, Jahrbuch Band III;
2004.
[8] Hansen L, Heinze W, Horst P. Representation of structural solutions
in Blended Wing Body preliminary design. In: Proc. of the 25th
international congress on aeronautical sciences ICAS2006, Hamburg;
2006.
[9] Heinze W. Ein Beitrag zur quantitativen Analyse der technischen und
wirtschaftlichen Auslegungsgrenzen verschiedener Flugzeugkonzepte
fu r den Transport grosser Nutzlasten. PhD thesis, TU Braunschweig,
ZLR Forschungsbericht 94-01; 1994.
[10] Anon. HexWeb HRH-10 Aramid bre/phenolic honeycomb, Product
Data, Hexcel Corporation, Dublin, CA; 2005.
[11] Ko nig O. Evolutionary design optimization: tools and applications.
Dissertation, Swiss Federal Institute of Technology, Zurich, ETH
No. 15486; 2004.
[12] MSC, Design sensitivity and optimization, MSC.Nastran 2004,
Users Guide; 2004.
[13] MSC, Reference manual, MSC.Nastran 2004, Users Guide; 2004.
[14] O

sterheld C, Heinze W, Horst P. Preliminary design of a Blended


Wing Body conguration using the design tool PrADO. In:
Proceedings of the CEAS Conference on Multidisciplinary Aircraft
Design and Optimisation, Cologne; 25/26 June 2001.
[15] O

sterheld C. Physikalisch begru ndete Analyseverfahren im integrier-


ten multidisziplinaren Flugzeugvorentwurf. PhD thesis, TU Braun-
schweig, ZLR Forschungsbericht 2003-06; 2003.
[16] Raq MY, Williams C. An investigation into the integration of
neural networks with the structured genetic algorithm to aid
conceptual design. Lecture notes in articial intelligence, vol.
1454. Springer Publication; 1998. p. 295307.
[17] Homeister F, Back T. Genetic algorithms and Evolution Strategies:
similarities and dierences. Technical Report SYS 1/92, Systems
Analysis Research Group, Department of Computer Science, Uni-
versity of Dortmund, Dortmund, Germany.
[18] Mukhopadhyay V. Sobieszczanski-Sobieski J. Analysis design and
optimisation of non-cylindrical fuselage for Blended-Wing-Body
(BWB) Vehicle. In: 9th AIAA/ISSMO symposium on multidisciplin-
ary analysis and optimisation, Atlanta, AIAA 2002-5664; 2002.
[19] Mukhopadhyay V. Blended-Wing-Body (BWB) fuselage structural
design for weight reduction. In: 46th AIAA/ASME/ASCE/AHS/
ASC structures, structural dynamics and materials conference,
Austin, AIAA 2005-2349; 2005.
[20] Rechenberg I. Optimierung technischer Systeme nach Prinzipien der
biologischen Evolution. Frommann Verlag; 1973.
[21] Schwefel H-P. Evolution and optimum seeking. New York: Wiley-
Interscience/John Wiley & Sons; 1995.
[22] Schwefel H-P, Rudolph G, Back T. Contemporary Evolution
Strategies. In: European conference on articial life; 1995. p. 893
907.
118 L.U. Hansen, P. Horst / Computers and Structures 86 (2008) 104118