jse03_form-finding

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Form Finding of Sparse Structures with Continuum

Topology Optimization
Salam Rahmatalla1 and Colby C. Swan, M.ASCE2

Abstract: A continuum topology optimization methodology suitable for finding optimal forms of large-scale sparse structures is pre-
sented. Since the need to avoid long compressive spans can be critical in determining the optimal form of such structures, a formulation
is used wherein the structure is modeled as a linear elastic continuum subjected to design loads, and optimized in form to maximize the
minimum critical buckling load. Numerical issues pertinent to accurate solution of the linearized buckling eigenvalue problem and
accurate design sensitivity analysis are discussed. The performance of the proposed design formulation is demonstrated on a few problems
designed to find optimal forms of a canyon bridge, long-span bridges, and an electrical transmission tower. In all cases, very credible
structural forms are obtained with the proposed design formulation. The results of the design examples solved are typically superior
structural forms with regard to buckling stability than those obtained to minimize the mean structural compliance.
DOI: 10.1061/共ASCE兲0733-9445共2003兲129:12共1707兲
CE Database subject headings: Optimization; Buckling; Sensitivity analysis; Eigenvalues.

Introduction they are now quite standard. In the recent past, such methods have
been investigated quite extensively for civil engineering type
Designing sparse large-scale structures that are not susceptible to structures with a good deal of attention paid to achievement of
buckling instabilities is a longstanding challenge in structural de- designs that are stable with respect to design loads 共Oberndorfer
sign optimization. Many structural design optimization methods et al. 1996; Rozvany 1996; Zou 1996; Achtziger 1999; Bojczuk
are in use today and include cross-section selection, or size and and Mróz 1999; Kočvara 2002兲. In such frameworks, an attempt
shape optimization of structural cross sections. These types of is typically made to achieve stable designs by enforcing local
methods are applicable toward the final stages of designing a Euler buckling constraints. This can be challenging, due to diffi-
structure, but they tend not to be very helpful when one is trying culty in identifying buckling lengths that can be considerably
to find the optimal forms of a structure based on its size 共or span兲, larger than individual member lengths when several collinear and
support conditions, and the loads that it is expected to carry. compressive members form an isolated sequence.
While the designer’s intuition and awareness of design precedents While discrete structural optimization methods date back at
will always play a vital role, structural topology optimization least one century to the work of Michell, continuum structural
methods are objective tools that might also be very useful in topology optimization methods have more recent origins, devel-
identifying suitable, if not optimal, structural forms. Among struc- oped in the 1980s as extensions of shape and size optimization
tural topology optimization frameworks are there are essentially techniques 共Cheng and Olhoff 1981; Bendsøe and Kikuchi 1988兲.
two classes of methods 共see Ohsaki and Swan 2002兲: 共1兲 discrete In these methods, the structure is modeled as a continuum and the
ground-structure topology optimization methods and 共2兲 con- form of the structural system is optimized using a system of dis-
tinuum structural topology optimization methods. tributed continuous design parameters 关see the recent review by
One of the pioneering works in the field of structural topology Eschenauer and Olhoff 共2001兲 for a survey of the numerous con-
optimization per se involved presentation of optimal discrete truss tinuum topology optimization formulations developed over the
structures now called Michell structures 共Michell 1904兲. More past 2 decades兴. Since these methods are not in any way restricted
recently, discrete ground structure truss topology optimization to truss-like structures, they have been investigated for a consid-
methods have been employed 共Dorn et al. 1964; Dobbs and Fel- erable range of applications including: the design of composite
ton 1969; Hemp 1973; Rozvany 1976; Save and Prager 1990兲 and material microstructures, 共see Swan and Arora 1997 for one ex-
ample兲; compliant mechanisms in microelectromechanical sys-
1
Dept. of Civil and Environmental Engineering, Center for Computer- tems 共Yin and Ananthasuresh 2002兲 intermediate scale plate and
Aided Design, The Univ. of Iowa, Iowa City, IA 52242. shell structures 共Swan and Kosaka 1997兲; and civil engineering
2
Dept. of Civil and Environmental Engineering, Center for Computer- type structures 共Mijar et al. 1998; Swan et al. 1998兲.
Aided Design, The Univ. of Iowa, Iowa City, IA 52242 共corresponding In spite of its arguable successes in design of small and inter-
author兲. E-mail: colby-swan@uiowa.edu mediate scale mechanical systems, there have been some chal-
Note. Associate Editor: Shahram Pezeshk. Discussion open until May lenges in applying continuum topology optimization to the con-
1, 2004. Separate discussions must be submitted for individual papers. To
ceptual design of large-scale civil engineering type structures for
extend the closing date by one month, a written request must be filed with
the ASCE Managing Editor. The manuscript for this paper was submitted which economy of material is typically a vital issue. Large-scale
for review and possible publication on July 18, 2002; approved on No- civil engineering structures such as bridges and transmission tow-
vember 19, 2002. This paper is part of the Journal of Structural Engi- ers are characteristically very sparse with the volume of material
neering, Vol. 129, No. 12, December 1, 2003. ©ASCE, ISSN 0733-9445/ that comprises the structural system constituting a very small
2003/12-1707–1716/$18.00. fraction 共⭐5%兲 of the structure’s total envelope volume. Con-

JOURNAL OF STRUCTURAL ENGINEERING © ASCE / DECEMBER 2003 / 1707


tinuum topology optimization methods, unlike discrete truss op- to satisfactory concept designs of large-scale sparse structures.
timization methods, have some difficulty modeling this sparsity In this paper, basic elements of continuum structural topology
since it requires extremely refined models with very high degree optimization are presented including material distribution design
of freedom counts. If the design method cannot capture this spar- parameters, problem statements, and design sensitivity analysis
sity, then it will have difficulty replicating and addressing the expressions. Finally, practical examples including concept de-
vibrational and stability characteristics of the structure. As a re- signs of large structures are demonstrated followed by discussion
sult, when continuum topology methods were applied in preced- and concluding remarks.
ing efforts to design the form of structures for both optimal stiff-
ness and vibrational characteristics, the designed forms of the
structure were realistic, but the computed vibrational characteris- Problem Formulations
tics, such as eigenfrequencies, were not 共Ma et al. 1995; Mijar
et al. 1998; Swan et al. 1998; Min et al. 2000兲. Structural Model and Material Layout Description
For large-scale sparse structures, stability considerations can
often be the controlling factor in determining the overall form of The objective of continuum structural optimization is to find a
the structure. For example, in the design of long-span bridges, layout of a structural material of specified properties in a defined
tension structures are typically optimal since they preclude poten- spatial region that provides optimum structural performance. In
tial buckling under the traffic loading and self-weight loading. order that the widest possible class of structural layouts can be
Previous applications of continuum topology optimization to find- considered, the methods in question must accommodate such gen-
ing optimal forms of bridges based on minimization of the linear erality. In this work, the spatial region that the candidate struc-
elastic structural compliance under traffic loading, or maximiza- tural models can occupy is denoted ⍀ s . To facilitate both descrip-
tion of the fundamental vibrational eigenfrequencies of the struc- tion of the structural material layout in ⍀ s and analysis of the
ture, or combinations of both functionals using multicriterion op- performance associated with each layout considered, the domain
timization approaches 共Swan et al. 1998兲 have not been is discretized into a relatively fine mesh of nodes and finite ele-
satisfactory. Neither linear elastic compliance, nor vibrational ments.
eigenfrequencies of a structural system are directly related to It is desired that at the end of the form-finding process, the
structural region ⍀ s will be decomposed into a collection of re-
structural stability in sparse structures. For this reason, the result-
gions cumulatively denoted ⍀ A that contain the structural mate-
ing concept designs of structures from these preceding works just
rial in question, and the remaining regions ⍀ B ⫽⍀ s /⍀ A that are
cited often feature long slender compression members that would
devoid of structural material. Since solution of the form-finding
be highly problematic if the design were pursued from the con-
problem in this way is ill-posed, an alternative relaxed approach
ceptual stage through the detailed design stage.
is usually employed, wherein it is assumed that an amorphous
In continuum structural topology optimization formulations,
‘‘mixture’’ of structural material A and a void material B exists
achievement of designs that are inherently stable in relation to the
throughout the structural region ⍀ s . In each region of ⍀ s , the
design loads can be very challenging. For example, it can be
nature of the mixture is characterized by a local volumetric den-
difficult to preclude local buckling behaviors with Euler-type con-
sity ␾ A of structural material A. By permitting mixtures, the
straints, since it is very difficult to identify discrete structural
structural material A and a fictitious void material B are allowed
members, their geometrical properties, and their end support con-
to simultaneously occupy an infinitesimal neighborhood about
ditions from the vector of design variables. One promising ap-
each Lagrangian point X苸⍀ s . The volumetric density of struc-
proach to addressing global geometrical instabilities in continuum
tural material A at a fixed Lagrangian point X苸⍀ s is denoted by
structural topology optimization 共Guedes and Rodrigues 1995;
␾ A (X) and represents the fraction of an infinitesimal region sur-
Neves et al. 1995兲 and that has been employed in design of brac-
rounding point X occupied by material A. Natural constraints
ing systems for portal frames is to model the structure as a lin-
upon the volumetric densities are
early elastic system and to use the minimum critical buckling load
computed via eigenvalue analysis either in the objective function 0⭐␾ A 共 X 兲 ⭐1; 0⭐␾ B 共 X 兲 ⭐1; ␾ A 共 X 兲 ⫹␾ B 共 X 兲 ⫽1 (1)
or as a design constraint. While such an approach is very straight-
Clearly, when ␾ A (X)⫽1 the point X contains solid structural
forward in principle, a complicating factor is that when the buck-
material, and when ␾ A (X)⫽0 the point X is devoid of structural
ling eigenvalues of the structure are nonsimple 共repeated兲, their
material. The last physical constraint of Eq. 共1兲 states that the
design derivatives are discontinuous. Accordingly, an alternative
material volume fractions at X are not independent and so one
approach to designing stable, sparse structures in a continuum
need only be concerned with the layout of structural material A.
topology framework is to model the structure as an elastic con-
The design of a structure is here considered to be the spatial
tinuum taking into account finite deformation effects and the as-
distribution of the structural material A in ⍀ s .
sociated instabilities 共Buhl et al. 2000; Bruns and Tortorelli 2001;
To describe the distribution of material A throughout ⍀ s using
Gea and Luo 2001; Rahmatalla and Swan 2003兲. Nevertheless,
a finite number of design parameters, the volumetric density at
the advantage of the first approach over the latter is its substan-
each of the NUMNP nodal point forms a set of NUMNP design
tially lower computational cost.
variables. These are then interpolated over the space of all inter-
In the body of this paper, a linear elastic continuum topology
mediate points in the structure using the nodal shape functions
formulation is introduced for minimization of mean structural
NUM N P
compliance, and for maximization of minimum critical buckling
values computed by linearized buckling analysis. One of the in- ␾ 共 X兲 ⫽ 兺
i⫽1
bi N i 共 X 兲 ᭙X苸⍀ s (2)
tents is to demonstrate the advantages of designing to maximize
minimum critical buckling loads as opposed to designing to mini- where bi ⫽nodal volumetric density values associated with the
mize linear elastic structural compliance. This point is salient as structural material; and N i (X)⫽nodal shape functions. This ap-
designing for compliance minimization has been frequently ex- proach yields a C 0 continuous design variable field that is not
plored in the research literature, and yet it will generally not lead susceptible to ‘‘checkerboarding’’ instabilities.

1708 / JOURNAL OF STRUCTURAL ENGINEERING © ASCE / DECEMBER 2003


Given the finite element model of the structural region ⍀ s , the
structural loads and restraints 共or supports兲 on this region are
specified as the set of design loads. For each set of design load-
f ext⫽ 冕
⌫s
Nhd⌫ s ⫹ 冕 ⍀s
N␳gd⍀ s

ings, and for each realization of the design vector b In all of the above, N denotes the nodal shape functions and B
⫽ 兵 b 1 ,b 2 ,...,b NUM N P 其 , the response performance of the structure denotes the standard strain–displacement matrices 共cf. Bathe
will be analyzed as a boundary value problem. From the com- 1996兲. The structural stiffness matrix K is positive definite due to
puted response of the structure, the performance of the structure the characteristics of the effective elasticity tensor C*, and this
will be quantified, as will be the sensitivity of the performance to guarantees a unique solution to the structural analysis problem for
variations in the design variables. each realization of the design b.
Once the equilibrium solution to the problem of Eq. 共5兲 is
Constitutive Mixing Rules obtained, then the linearized geometrical stiffness matrix G can
be computed based on the stress field ␴ in the structure


In the proposed design framework, each finite element comprising
the spatial domain ⍀ s of the structure will generally contain a
jk ⫽
G LM ␴ mn␦ jk d⍀ s
L M
N ,m N ,n (7)
spatially varying mixture of the structural and void materials. It is ⍀s
necessary to prescribe the stiffness 共or elastic moduli兲 of such
It is worth noting that G is not necessarily positive definite but
mixtures in terms of the stiffness characteristics of the solid ma-
rather depends heavily upon the nature of the stress field in the
terial Csolid , those of the fictitious void material Cvoid , and the
structure. A purely tensile stress field clearly makes G positive
local volumetric density of the structural material ␾共X兲. Here, the
definite, although for any compressive stresses, G will not be
well-known powerlaw formula 共Bendsøe and Sigmund 1999兲 is
positive definite.
used to accomplish this task, providing the local effective stiff-
ness of the mixture C* as
C* ⫽␾ p Csolid⫹ 共 1⫺␾ p 兲 Cvoid (3)
Structural Performance Measures

where typically the mixing rule parameter p苸 关 1,4兴 . With p⫽1,


Overview
the Voigt rule of mixtures is obtained which does not penalize
mixtures, but which does yield a convex formulation for compli- As noted previously, structural topology design problems can be
ance minimization problems 共Swan and Kosaka 1997兲 so that formulated in a number of alternative ways through utilization of
only one solution exists for the design problem. With p⫽4, mix- assorted objective and constraint functions. Generally, the objec-
tures are penalized in the final design, so that regions of ⍀ s tend tive function measures the performance of the structure, and the
to be either solid or void, but the optimization problem is not constraint function limits the amount of structural material that
convex, and will admit a number of solutions that satisfy the first can be used, although the roles can be reversed equally well. The
order optimality conditions. significant aspects of using CSTO to design large-scale sparse
structures can be demonstrated here using the linear elastic struc-
Structural Analysis tural compliance performance measure and the critical load buck-
ling factor.
For each design, a structural analysis problem is solved on the
continuum domain ⍀ s . In general terms, the structural analysis
problem solved for each realization of the design vector b is the Linear Elastic Structural Compliance
following: Find the displacement field u共X兲 ⍀ s →R3 such that the If a structure features a linear elastic response behavior, the re-
variational equilibrium problem is solved sulting displacement field u in response to a set of applied exter-

冕⍀s
␴:␦⑀d⍀ s ⫽ 冕
⌫s
h•␦ud⌫ s⫹ 冕 ⍀s
␳g•␦ud⍀ s (4)
nal loads f ext will be simply u⫽K⫺1 "f ext, where K represents the
stiffness matrix of the structure. For a given set of loads, the
compliance ⌸共b兲 of the structure is simply
where ␴共X兲⫽local stress field in the structure; h⫽traction vector
consistent with the design loads being applied to the structure; ⌸ 共 b兲 ⫽ 21 f ext"u (8)
␳共X兲⫽local mass density of the structural material; Structural concept designs b that are stiff with respect to the ap-
g⫽gravitational body force vector; ␦u⫽kinematically admissible plied loads will have small compliance ⌸共b兲, whereas structures
variational displacement field; and ␦⑀⫽corresponding variational that are not stiff with respect to the applied loads will have large
strain field. In the structural model, the material features linear compliance. To facilitate usage of gradient-based optimization so-
elastic behavior such that ␴⫽C*:⑀, where the effective elasticity lution techniques, it is necessary to compute the design deriva-
tensor is design dependent and prescribed in accordance with Eq. tives of the compliance function. It can be shown that the design
共3兲. The matrix problem associated with variational equilibrium gradient of structural compliance is provided by the following
of the discrete finite element structural model, for which u(X) expression:

冉 冊
⫽ 兺 i N i (X)ui is
d⌸ 1 ⳵K ⳵f ext
0⫽K"u⫺f ext⫽f int⫺f ext (5) ⫽⫺ u• •u⫺ (9)
db 2 ⳵b ⳵b
where

jk ⫽
K LM 冕 ⍀s
B Lm j C * M
mn B nk d⍀ s
Linearized Bucking Performance Measure
Linearized buckling eigenvalue analysis proceeds as follows: A


prescribed force loading f ext is applied to the structure with its
f int⫽K"u⫽ BT ␴d⍀ s (6) magnitude necessarily being less than that required to induce geo-
⍀s metric instability in the structure. Once the resulting linear, elas-

JOURNAL OF STRUCTURAL ENGINEERING © ASCE / DECEMBER 2003 / 1709


tostatic displacement solution u⫽ 兵 u i 其 苸RN in response to the ap-
plied loading f ext is obtained (K"u⫽f ext), where K is the
linearized stiffness matrix, then the following eigenvalue problem
is solved:
关 K共 b兲 ⫹␭G共 u,b兲兴 •␺⫽0 (10)
In the preceding, b⫽ 兵 be 其 苸RM is again the vector of design vari-
ables; K⫽tangent stiffness operator; G共u,b兲⫽linearized geomet-
ric stiffness matrix; ␭⫽⫺共␺"K"␺/␺"G"␺兲⫽eigenvalue denoting
the magnitude by which f ext must be scaled to create instability in
the structure; and ␺⫽normalized eigenvector satisfying
␺"K"␺⫽1. To avoid numerical difficulties in the solution of Eq.
共10兲 stemming from the indefinite characteristics of G, it is com-
mon 共Bathe 1996兲 to solve a modified eigenvalue problem that
Fig. 1. Undeformed and deformed configurations of coarse beam
deals with two positive definite matrices
model with first buckling mode shown
关共 K⫹G兲 ⫺␥K兴 •␺⫽0 (11)
where
librium constraint 共r⫽0兲, and the second term can be made to
␭⫺1 1 vanish by selecting the adjoint displacement vector to solve the
␥⫽ ⇔␭⫽ (12) following linear adjoint equality statement
␭ 1⫺␥
In Eq. 共11兲, the matrix K is positive definite irrespective of the ⳵G
K"ua ⫽␺• •␺ (18)
loading applied to the structure, whereas the matrix 共K⫹G兲 will ⳵u
only be positive definite when the magnitude of the loading ap- Since it can be shown that d⌶/db⫽d f E /db, it follows that the
plied to the structural model is less than the critical magnitude design gradient expression for the objective function is

冉 冊 冉 冊
that creates instability in accordance with linearized buckling
theory. dfE ⳵G 1 ⳵K ⳵K ⳵f ext
⫽⫺␺• ⫹ •␺⫹ua • •u⫺ (19)
The design problem is formulated to maximize the calculated db ⳵b ␭ ⳵b ⳵b ⳵b
minimum-buckling load factor 共␭兲, and accordingly the objective
The preceding expression is valid only when the minimum
function f E to be minimized for this problem would simply be the
eigenvalue is a simple, or nonrepeated, eigenvalue. When the
reciprocal of the lowest eigenvalue ␭ as follows:
minimum eigenvalue is nonsimple, or repeated, the variation of
1 the eigenvalue in design space is nonsmooth, and direct usage of
f E 共 u,b兲 ⫽ (13) the expression in Eq. 共19兲 is technically incorrect 共Choi et al.
min共 ␭ 兲
1983; Seyranian et al. 1994兲. Resolution of this issue is non-
The optimization problem is thus stated to minimize the recipro- trivial, although it can be ameliorated somewhat by using small
cal of the first 共or minimum兲 critical buckling load as follows: and variable move limits in the design optimization process. De-

min f E ⫽min
b,u b,u
冉冊 冉
1

⫽min ⫺ max
b,u 储 ␺ 储 ⫽0
␺"G"␺
␺"K"␺ 冊 (14)
spite this challenge, designs that successfully maximize the buck-
ling stability of a structural system can nevertheless be obtained.

subject to the normal bound constraints on the design variables


Eq. 共1兲, the linear structural equilibrium state Eq. 共5兲, and a con- Numerical Issues
straint on material resources.
The design gradient of the objective function can be expressed Beyond the nonsmooth nature of repeated eigenvalues, when they
as occur, there are additional numerical issues associated with reli-
able solution of the buckling eigenvalue problem and the atten-
d f E ⳵ f E ⳵ f E ⳵u dant effect on accurate sensitivity of buckling eigenvalues with
⫽ ⫹ • (15)
db ⳵b ⳵u ⳵b respect to design changes. As noted previously, due to indefinite
characteristics of the linearized G, the modified buckling eigen-
To avoid explicit computation of the term ⳵u/⳵b, adjoint design value problem of Eq. 共11兲 is usually solved rather than that of Eq.
sensitivity analysis is employed by augmenting the objective 共10兲. However, when the magnitude of loading applied to the
function f E with the equilibrium state equation as follows: structure is greatly exceeded by that of f crit, then ␭→⬁ and ac-
⌶⫽ f E ⫹ua "r (16) cordingly ␥→1. As this happens, it is increasingly difficult to
accurately compute both ␭ and d␭/db. This is demonstrated by
where u ⫽adjoint displacement vector which functions as a ma-
a
considering the simple beam model shown in Fig. 1 for which a
trix of Lagrange multipliers and determined by the solution of a load of magnitude 储 f ext储 ⫽1 MN is applied, and for which f crit
linear adjoint problem. The design derivative of the augmented ⫽␭ * 储 f ext储 ⫽5.1855 MN, giving ␭⫽5.1855. One would expect
Lagrangian is then written as follows: that as the magnitude of the loading applied to the structural
d⌶
db

⳵b冉
⳵fE
⫹ua •
⳵r
⳵b

⳵u ⳵u 冊冋
⳵ f E ⳵r a ⳵u
⫹ •u • ⫹ r•
⳵b
⳵ua
⳵b 册 冉 冊 model changed, the computed value of ␭ would change in accor-
dance with ␭⫽ f crit/ 储 f ext储 . The numerical results shown in Table
1 indicate otherwise, however, and that as the inaccuracy in com-
(17)
puted values of ␭ increases, the inaccuracy in the design sensi-
The last term of Eq. 共17兲 vanishes due to satisfaction of the equi- tivities increases dramatically. Accordingly, when computing the

1710 / JOURNAL OF STRUCTURAL ENGINEERING © ASCE / DECEMBER 2003


Table 1. Buckling Factors Computed with Standard Methods for results presented in Table 2. As can be seen from this table, the
Model Shown in Fig. 1. computed buckling eigenvalues now scale in inverse proportion
Finite difference to the magnitude of the applied loads f ref, and the computed
f ext Analytical DSA DSA design sensitivity results are now in close agreement with those
共N兲 ␥ ␭ Node 14 Node 14 computed by converged finite difference analysis.
10 2
0.999981 5.29781⫻104 ⫺6.7096⫻10⫺5 ⫺6.7491⫻10⫺5
104 0.9981 5.24000⫻103 ⫺6.8476⫻10⫺3 ⫺7.0158⫻10⫺3 Demonstrative Examples
106 0.8072 5.1855⫻100 ⫺7.1877⫻10⫺1 ⫺7.1877⫻10⫺1
Note: Inaccuracies exist in computed eigenvalues and also in both ana- Material Properties and Mixing Rules
lytical and finite difference design sensitivity results.
In the following examples, the CSTO formulation to maximize
minimum critical buckling loads computed by linearized eigen-
buckling eigenvalue of the structural model, greater accuracy is value analysis is tested and compared with the more commonly
achieved by using a scaled load magnitude 储 f ext储 such that used compliance minimization formulation. In all cases, the initial
储 f ext储 / f crit苸(0.1,1). The ratio cannot exceed unity however since starting designs feature a completely solid structural domain; the
in that case 共K⫹G兲 loses its positive definiteness. To avoid the solid structural material in all problems is steel with a Young’s
problems associated with application of a load to the structural modulus of 206 GPa and shear modulus of 79.2 GPa and a mass
model that is either far too small 共i.e. 储 f ext储 / f critⰆ1), or too large density of 7,800 kg m⫺3. The powerlaw mixing rule with p⫽4 is
共i.e. 储 f ext储 / f crit⭓1) an algorithm such as that shown in Fig. 2 is used in all computations to achieve material layouts that, in the
necessary. end, are more or less discrete and interpretable. In addition, the
The essential idea behind the proposed algorithm in Fig. 2 is nodal design variable formulation of the ‘‘Problem Formulations’’
that while a fixed set of structural loads, here denoted by f ref are Section is employed without any spatial filtering of design vari-
applied to the structural model for all realizations of the design ables and without any perimeter control. All design optimization
vector b, the computed value of ␭ ref⫽ f crit(b)/ 储 f ref储 for each re- problems were solved using fairly standard sequential linear pro-
alization of the design b could either be excessively large 共i.e., gramming techniques with variable move limits. Optimization
␭ refⰇ1) and thus contain a large amount of error, or alternatively problems were terminated when the designs satisfied the Kuhn–
too small 共i.e., ␭ ref⭐1) making it very difficult to compute the Tucker first order optimality condition.
minimum eigenvalue associated with Eq. 共10兲. An accurate com-
putation of the buckling eigenvalue with respect to the loads f ref
can be achieved, however, by iterative scaling of the actual sys- Canyon Bridge Problem
tem of loads applied to the structure f ext⫽k * f ref until ␭苸共1,10兲, In this case, we consider seeking optimal forms of a bridge to
where here ␭ is the minimum eigenvalue with respect to the carry self-weight and traffic loads across a span of 1,000 m. The
scaled loads. Since f crit(b)⫽␭ ref* 储 f ref储 ⫽(␭ ref/k)(k * 储 f ref储 ) bridge is designed in two dimensions and the design traffic load
⫽␭ * 储 f ext储 , it follows that ␭ ref⫽k * ␭. To demonstrate the effec- applied uniformly to the deck level of the bridge is 10 kPa. The
tiveness of the proposed algorithm, the computed buckling eigen- candidate spatial region that the bridge superstructure can poten-
values and their design gradients associated with the test problem tially occupy is shown in Fig. 3共a兲. Since it is desired that the
of Fig. 1 were computed using the algorithm of Fig. 2, and the structural form obtained be sparse, the volume of structural ma-
terial used is constrained to be less than or equal to 12.5% of the
bridge envelope volume. Even with this material usage constraint,
the gross weight of the bridge structure greatly exceeds the mag-
nitude of the design traffic load. Accordingly, the bridge form is
designed considering only the traffic loading, although once the
designs are obtained, the performance of the structure under both
traffic and self-weight loading are considered. If the self-weight
loading of the bridge were considered during the optimization
process, they would be dominant and unsatisfactory concept de-
sign would be obtained 共see Swan et al. 1998 for both an example
and discussion兲.
The concept design solution obtained by minimizing the linear
elastic structural compliance under the design loading, without
consideration of potential buckling instabilities, is shown in Fig.
3共b兲. The primary structural system is a long compression arch
that spans the full canyon width. The deck is very slender, and
supported in the first and last thirds of the span by a system of
very slender compression members, and in the central third by
slender tension members. The first buckling mode associated with
this material layout is shown in Fig. 3共c兲 and indicates buckling in
two of the slender compression members.
Two alternative structural concepts were obtained by solving
slight variations of the optimization problem to maximize the
minimum critical buckling factor subject to material usage con-
Fig. 2. Algorithm for accurate computation of buckling load factor
straints. In the first design 关Fig. 3共d兲兴, the entire structural domain
eigenvalue ␭ ref with respect to fixed set of design loads f ref
⍀ s was treated as designable, while in the second 关Fig. 3共f兲兴

JOURNAL OF STRUCTURAL ENGINEERING © ASCE / DECEMBER 2003 / 1711


Table 2. Buckling Factors Computed Using Algorithm of Box 2, Along with Accurate Design Sensitivity Analysis Results Confirmed with
Converged Finite Difference DSA
Analytical Finite difference
f ref DSA DSA
共N兲 k ␥ ␭ ref Node 4 Node 4
1 1⫻106 0.8072 5.1855⫻106 ⫺7.1877⫻10⫺7 ⫺7.1877⫻10⫺7
102 1⫻104 0.8072 5.1855⫻104 ⫺7.1877⫻10⫺7 ⫺7.1877⫻10⫺7
104 1⫻102 0.8072 5.1855⫻102 ⫺7.1877⫻10⫺3 ⫺7.1877⫻10⫺3
106 1⫻100 0.8072 5.1855⫻100 ⫺7.1877⫻10⫺1 ⫺7.1877⫻10⫺1
108 3.125⫻10⫺2 0.3974 5.1855⫻10⫺2 ⫺7.1877⫻101 ⫺7.1877⫻101
1010 3.0517⫻10⫺4 0.4115 5.1855⫻10⫺4 ⫺7.1877⫻103 ⫺7.1877⫻103
1012 3.8147⫻10⫺6 0.2644 5.1855⫻10⫺6 ⫺7.1877⫻105 ⫺7.1877⫻105

design variable values of unity were imposed on all nodes at the buckling mode for this material layout is also highly localized and
deck level. The two structural concepts obtained are markedly not visible on the global scale 关Fig. 3共g兲兴.
different. The first layout resembles the compliance minimizing The computed performance characteristics associated with the
design 关Fig. 3共b兲兴, in that it primary feature is stout compression three layout designs of Fig. 3 are provided in Table 3. Not sur-
arch that crosses the span. The deck is substantially thicker that of prisingly, the compliance minimizing material layout has the
the compliance minimizing design, however, reducing the need smallest compliance of all three designs under the traffic loading.
for the system of slender compression members that transfer deck It also shows a computed buckling eigenvalue slightly greater
loads to the arch. For this reason the secondary compression than that of the first buckling-resistant design 关Fig. 3共d兲兴, although
members appear to be substantially more stout than those in the substantially smaller than that of the second buckling-resistant
compliance minimizing design. The computed fundamental buck- design 关Fig. 3共f兲兴. Under loadings that also include the self-weight
ling mode associated with this material layout 关Fig. 3共e兲兴 is not of the structural material, the same trends in compliance and
visible on the global scale and is thus a highly localized mode. In buckling stability persist. The computed performance characteris-
the second buckling stability design, the primary structural feature tics of the layout designs results should be viewed cautiously,
is a long deep tension member crossing the span, with a system of however, since in taking these concepts to more detailed final
secondary compression members to support the deck. Since the design stages, the performance characteristics could change con-
deck would appear to be predominantly in compression along the siderably.
span direction, the proposed design method has stabilized it with
an irregular system of reinforcing members. The fundamental
Long-Span Concept Designs
For main span lengths greater than 1,000 m, suspension bridges
that use primarily tension to carry both the design loads and their
own weight are generally optimal in that the primary structural
elements are not subject to buckling. Here the conceptual layout
optimization of a very long span 共3,000 m兲 bridge is considered in
which the candidate structural region is selected to lie at or above
the traffic deck level as shown in Fig. 4共a兲. Again, the design
traffic loading on the bridge is 10 kPa uniformly distributed on
the deck level. The structural material usage is limited to 12.5%
of the envelope volume. The design domain is meshed with
10,000 bilinear continuum finite elements, and the problem is
solved first to maximize the minimum critical buckling load 关Fig.
4共b兲兴 and then to minimize the structural compliance of the struc-
ture under the traffic loading 关Figs. 4共c and d兲兴. Since the
compliance-minimizing design shown in Fig. 4共c兲 is somewhat
difficult to interpret, the problem was re solved at a substantially
higher mesh resolution, and is shown in Fig. 4共d兲.
In both of the compliance-minimizing designs, the proposed
methodology yields designs that use flexure of a flying beam-like
structure whose supports are cantilevered out into the span. While
these designs are in many ways quite plausible and realistic, par-
Fig. 3. 共a兲 Design domain and boundary/loading conditions; 共b兲 re- ticularly regarding the distributed support of the deck by systems
sulting material layout to minimize mean compliance of structure of cables suspended from the compression cord of the beam, the
with lowest buckling mode 共c兲; 共d兲 layout to maximize minimum very serious problem with these designs is that the top chord of
buckling eigenvalue with lowest buckling mode 共e兲; 共f兲 layout to the beam-like structure is very long, slender, and in compression
maximize minimum buckling eigenvalue, by considering nondesign- under the design loading. If these compliance-minimizing concept
able layer along bridge deck where traction forces are applied with
designs were to be taken into a secondary more detailed design
lowest buckling mode 共g兲; computed performance characteristics are
stage, the long compression cord members would need to be sized
provided in Table 3
very large to avoid buckling, and the resulting design would be

1712 / JOURNAL OF STRUCTURAL ENGINEERING © ASCE / DECEMBER 2003


Table 3. Computed Performance Characteristics Associated with Three Canyon Bridge Designs of Fig. 3
Buckling factor
Buckling factor under traffic load
Structural compliance Buckling factor under traffic load and reduced self
Performance measure under traffic loading 共N m兲 under traffic load only and full self weight weighta
Compliance design 1.70⫻103 1.74⫻104 65.9 4.84⫻103
Buckling design No. 1 2.49⫻103 1.47⫻104 45.4 3.88⫻103
Buckling design No. 2 1.00⫻104 4.46⫻104 83.3 6.98⫻103
a
Reduced self-weight assumes that bridge weight is only approximately 1% of the weight it has in the model. This approximates the weight of a bridge
that occupies only 0.125% of the envelope volume ⍀ s .

excessively heavy and inefficient. It is worth noting that by uti- regions 关Fig. 5共a兲兴 is solved. The layout design obtained by maxi-
lizing this linear elastic compliance minimizing formulation, mizing the buckling stability 关Fig. 5共b兲兴 again appears superior to
similar topologies will result regardless the length of the span or the compliance-minimizing design solution 关Fig. 5共c兲兴 in terms of
the magnitude of the external loads. On the other hand, the layout global stability.
design of Fig. 4共b兲 shows a conceptual design that maximizes the For the long-span bridge designs, the weight of the structural
linearized critical buckling load factor. As can be seen, the sus- material used 共12.5% of envelope volume multiplied by the unit
pension concept design solution uses compression in the rela- weight of the structural material兲 again dwarfed the magnitude of
tively stout ‘‘towers’’ that elevate the cable, tension in the long the design traffic loading. Consequently, the weight of the struc-
suspension cable that extends across the span, and tension in the tural material was neglected in the design optimization problem.
relatively short hanger system that suspends the deck from the The compliances and the critical buckling load factors for all of
suspension cable. That the proposed formulation produces a sus- the designs in Figs. 4 and 5 are provided 共Tables 4 and 5兲, both
pension type concept design 关Fig. 4共b兲兴 resembling actual long under the design traffic loading, and under the combined traffic
span bridges in usage today is an encouraging development. It is loading and the weight of the structural material. Based on the
worth noting here that a concept design similar to that in Fig. 4共b兲 computed performance characteristics shown in these tables, it
was obtained by Oberndorfer et al. 共1996兲 with a discrete ground appears that the layout designs obtained by maximizing the lin-
structure topology optimization method that considered only local earized buckling stability buckling factors 关Figs. 4共b兲 and 5共b兲兴
buckling instabilities. are indeed superior performance to the compliance-minimizing
To investigate the ability of the linearized buckling approach designs 关Figs. 4共c兲, 4共d兲, and 5共c兲兴 in terms of stability. As noted
to obtain practical long-span designs under different support con- above, however, the only computed design performance charac-
ditions, a similar problem but with two spans and three support teristics that are truly meaningful are those based on the final
detailed structural design. For long-span suspension bridges, it is
actually more realistic that structural volume will lie between 0.1
and 1% of the envelope volume of the bridge. Hence, in proceed-
ing from concept designs to realistic final detailed designs, the
structural models would need to undergo substantial refinement
that would significantly further reduce overall weights.

Fig. 4. 共a兲 Design domain with loading and support conditions; 共b兲
layout design obtained by maximizing minimum buckling eigen-
value; 共c兲 layout design obtained by minimizing structural compli- Fig. 5. 共a兲 Design domain and boundary conditions for two-span
ance; 共d兲 compliance minimizing layout obtained with more refined three-support problem with all dimensions in meters; 共b兲 resulting
model. Computed performance characteristics are provided in material layout to maximize minimal buckling eigenvalue; and 共c兲
Table 4 layout to minimize mean compliance of structure

JOURNAL OF STRUCTURAL ENGINEERING © ASCE / DECEMBER 2003 / 1713


Table 4. Computed Performance Characteristics Associated with Long-Span Bridge Layouts Shown in Fig. 4
Structural compliance Buckling factor Buckling factor
under traffic loading Buckling factor under traffic load under traffic load
Performance measure 共N m兲 under traffic load only and total self weight and reduced self weighta
Buckling design 3.16⫻106 1.25⫻103 5.82 412
Compliance design 2.92⫻105 1.68⫻102 0.882 58.3
a
Reduced self-weight assumes that bridge weight is only approximately 1% of the weight it has in the model. This approximates the weight of a bridge
that occupies only 0.125% of the envelope volume ⍀ s .

Transmission Tower Design to maximize the minimum critical buckling load of the structure
has been demonstrated here on the concept design of bridge struc-
Failure of electrical power transmission towers by buckling dur-
tures and a transmission tower. Based on the example problems
ing ice storms is a potential problem in the power industry. In this
solved, designing to maximize the minimum critical buckling
example, the proposed design formulation is used to obtain opti-
mal forms in two dimensions for transmission tower that can load appears to be more effective at consistently achieving stable
carry the static, vertical loads associated with six cables. The structural forms than does minimizing the generalized compliance
static loading and boundary conditions on the spatial design do- of the structural system. The proposed methods are somewhat
main are shown in Fig. 6共a兲 and are consistent with those used by promising in that they yield structural concept designs in some
Kocer and Arora 共2002兲 in a work entailing detailed optimal de- cases that are known to be optimal for certain design problems.
sign of transmission towers with respect to cross-section selection For example, in the design of very long-span bridges, the method
of individual members. In the current layout optimization prob- yields suspension bridge type designs. If the design tools can be
lem with continuum topology design, the spatial domain of Fig. confirmed on a number of such design applications, then they can
6共a兲 is discretized with 12,000 bilinear continuum finite elements be used with greater confidence in new classes of design prob-
and concentrated loads of 10 kN are applied at the six cable lems for which there is not necessarily any preceding experience
support positions. As in the preceding problems the structural to guide the designer.
material usage is constrained to 12.5% of the design domain’s A number of investigators have recently demonstrated that lay-
envelope volume. The design problem was first solved to mini- out optimization of sparse structures can also be achieved to
mize the mean structural compliance under the design loads, with maximize critical buckling loads computed via geometrically
the resulting topology shown in Fig. 6共b兲. While this design has a nonlinear structural buckling analysis. For large-scale sparse civil
low linear elastic compliance, it is relatively unstable with regard structures, the nonlinear analysis within the models can be quite
to buckling because it utilizes many long slender unbraced com- computationally expensive, and the proposed formulation pre-
pression members. As in the preceding bridge design problems, sented here based on linearized buckling analysis can achieve
the major weakness of the compliance minimizing formulation is similar results much more efficiently 共see Rahmatalla and Swan
its inability to detect potential buckling instabilities and to arrange 2003兲.
the structure in a way that minimizes the likelihood of their actual The emphasis here has been on problem formulations and the
occurrence. Fig. 6共c兲 shows a layout design solution obtained by resulting structural design solutions obtained. Numerical and
maximizing the fundamental linearized buckling eigenvalue. This computational issues have also been addressed to facilitate accu-
solution more closely resembles the form of existing transmission rate calculation of buckling eigenvalues in accordance with the
towers, than the solution of Fig. 6共b兲. The comparative perfor- linearized theory, and accurate design sensitivity analysis of these
mance characteristics of both designs in terms of compliance and buckling eigenvalues. Nonsimple 共repeated兲 eigenvalues can and
buckling stability are provided in Table 6. often do occur when optimizing the layout of a structure to maxi-
mize the fundamental buckling eigenvalue. While there remain
unresolved issues associated with the design sensitivity of such
Discussion and Conclusions non-simple eigenvalues, when they occur, it is worth noting here
that even when the DSA expressions for simple eigenvalues are
The objective of this paper has been to present and apply a con- employed in such cases, the optimization problem still tends to
tinuum structural topology optimization formulation that can be converge, although not monotonically, to designs that achieve op-
used to detect and avoid buckling instabilities in the conceptual timized buckling stability. The computational cost of performing
design stage of large sparse structural systems. The motivation is the design examples of two-dimensional structures presented
to develop design tools that will produce conceptual structural herein has been quite modest, being on the order of a cpu hour
forms that are optimal and that will be less problematic in the each on an HP J-class single processor workstation. Beyond com-
ensuing detailed design stages. Usage of the proposed formulation putational resource issues, which are significant, there is no con-

Table 5. Computed Performance Characteristics of Two-Span, Three-Support, Long-Span Bridge Layouts Shown in Fig. 5
Structural compliance Buckling factor Buckling factor
for traffic loading Buckling factor for traffic load for traffic load and
Performance measure 共N m兲 for traffic load only and total self weight reduced self weighta
Buckling design 1.52⫻105 4.84⫻103 63.3 5.15⫻103
Compliance design 2.06⫻104 1.94⫻103 7.18 5.29⫻102
a
Reduced self-weight assumes that bridge weight is only approximately 1% of the weight it has in the model. This approximates the weight of a bridge
that occupies only 0.125% of the envelope volume ⍀ s .

1714 / JOURNAL OF STRUCTURAL ENGINEERING © ASCE / DECEMBER 2003


Bendsøe, M. P., and Sigmund, O. 共1999兲. ‘‘Material interpolations in
topology optimization.’’ Arch. Appl. Mech., 69, 635– 654.
Bojczuk, D., and Mróz, Z. 共1999兲. ‘‘Optimal topology and configuration
design of trusses with stress and buckling constraints.’’ Struct. Optim.,
17, 25–35.
Bruns, T. E., and Tortorelli, D. A. 共2001兲. ‘‘Topology optimization of
nonlinear elastic structures and compliant mechanisms.’’ Comput.
Methods Appl. Mech. Eng., 190, 3443–3459.
Buhl, T., Pedersen, W., and Sigmund, O. 共2000兲. ‘‘Stiffness design of
geometrically nonlinear structures using topology optimization.’’
Struct. Multidisc. Optim., 19, 93–104.
Cheng, K. T., and Olhoff, N. 共1981兲. ‘‘An investigation concerning opti-
mal design of solid elastic plates.’’ Int. J. Solids Struct., 17, 305–323.
Choi, K. K., Haug, E. J., and Seong, H. G. 共1983兲. ‘‘An iterative method
for finite dimensional structural optimization problems with repeated
eigenvalues.’’ Int. J. Numer. Methods Eng., 19, 93–112.
Dobbs, W., and Felton, L. P. 共1969兲. ‘‘Optimization of truss geometry.’’
Proc. ASCE, 95共ST10兲, 2105–2519.
Fig. 6. 共a兲 The design domain and boundary/loading conditions for
Dorn, W., Gomory, R., and Greenberg, H. 共1964兲. ‘‘Automatic design of
the transmission tower; 共b兲 resulting layout for compliance minimi-
optimal structures.’’ J. Mech., 3, 25–52.
zation; 共c兲 resulting layout to optimize global stability. Computed
Eschenauer, H. A., and Olhoff, N. 共2001兲. ‘‘Topology optimization of
performance characteristics are in Table 6
continuum structures: A review.’’ Appl. Mech. Rev., 54, 331–390.
Gea, H. C., and Luo, J. 共2001兲. ‘‘Topology optimization of structures with
Table 6. Computed Performance Characteristics Associated with geometrical nonlinearities.’’ Comput. Struct., 79, 1977–1985.
Transmission Tower Layout Designs Shown in Fig. 6 Guedes, J. M., and Rodrigues, H. C. 共1995兲. ‘‘Necessary conditions for
the optimal design of structures with nonsmooth eigenvalue based
Performance measure Compliance 共N m兲 Buckling eigenvalue
criterion.’’ Struct. Optim., 9, 52–56.
Buckling design 25.0 2.44⫻103 Hemp, W. S. 共1973兲. Optimum structures, Clarendon, Oxford, U.K.
Compliance design 0.782 3.37⫻102 Kocer, F. Y., and Arora, J. S. 共2002兲. ‘‘Optimal design of latticed towers
subjected to earthquake loading.’’ J. Struct. Eng., 128共2兲, 197–204.
Kočvara, M. 共2002兲. ‘‘On the modeling and solving of the truss design
ceptual difficulty in extending the proposed design methods to problem with global stability constraints.’’ Struct. Multidisc. Optim.,
form-finding of structures in three dimensions. 23, 189–203.
In the example problems solved in ‘‘Demonstrative Ex- Ma, Z. D., Kikuchi, N., and Cheng, H. C. 共1995兲. ‘‘Topological design for
amples,’’ the structure was permitted to occupy up to 12.5% of vibrating structures.’’ Comput. Methods Appl. Mech. Eng., 121, 259–
the structure’s envelope volume. In reality, suspension bridges 280.
and transmission towers typically feature a higher degree of spar- Michell, A. G. M. 共1904兲. ‘‘The limits of economy in frame structures.’’
sity, occupying on the order of 1% or less of the structure’s en- Philos. Mag., 8共47兲, 589–597.
Mijar, A. R., Swan, C. C., Arora, J. S., and Kosaka, I. 共1998兲. ‘‘Con-
velope volume. Within the current continuum topology optimiza-
tinuum topology optimization for concept design of frame bracing
tion framework, modeling structures with such a high degree of
systems.’’ J. Struct. Eng., 124共5兲, 541–550.
sparsity would require finite element models of much greater
Min, S., Nishiwaki, S., and Kikuchi, N. 共2000兲. ‘‘Unified topology design
resolution than the models used herein with 104 – 105 elements. of static and vibrating structures using multiobjective optimization.’’
Nevertheless, even though the current stability optimization Comput. Struct., 75, 93–116.
framework is limited in its ability to capture the true sparsity Neves, M. M., Rodrigues, H., and Guedes, J. M. 共1995兲. ‘‘General topol-
characteristics of structures, it does yield material layouts that are ogy design of structures with a buckling load criterion.’’ Struct.
quite realistic. Optim., 10, 71–78.
Oberndorfer, J. M., Achtziger, W., and Hörnlein, H. R. E. M. 共1996兲.
‘‘Two approaches for truss topology optimization: a comparison for
Acknowledgments practical use.’’ Struct. Optim., 11, 137–144.
Ohsaki, M., and Swan, C. C. 共2002兲. ‘‘Topology and geometry optimiza-
This research was funded in part by a grant from the University of tion of trusses and frames.’’ ASCE/SEI State-of-Art-Rep. on Structural
Iowa CIFRE Program, and in part by Grant No. NSF DMS- Optimization, S. A. Burns, ed., New York.
9874015. Rahmatalla, S. F., and Swan, C. C. 共2003兲. ‘‘Continuum topology opti-
mization of buckling-sensitive structures.’’ AIAA J., 41共6兲, 1180–
1189.
References Rozvany, G. I. N. 共1976兲. Optimal design of flexural systems, Pergamon,
Oxford, U.K.
Achtziger, W. 共1999兲. ‘‘Local stability of trusses in the context of topol- Rozvany, G. I. N. 共1996兲. ‘‘Difficulties in truss topology optimization
ogy optimization Part II: A numerical approach.’’ Struct. Optim., 17, with stress, local buckling and system stability constraints.’’ Struct.
247–258. Optim., 11, 213–217.
Bathe, K. J. 共1996兲. Finite element procedures, Prentice-Hall, Englewood Save, M., and Prager, W. 共1990兲. Structural optimization, Plenum, New
Cliffs, N.J. York.
Bendsøe, M. P., and Kikuchi, N. 共1988兲. ‘‘Generating optimal topology in Seyranian, A. P., Lund, E., and Olhoff, N. 共1994兲. ‘‘Multiple eigenvalues
structural design using a homogenization method.’’ Comput. Methods in structural optimization problems.’’ Struct. Optim., 8, 207–227.
Appl. Mech. Eng., 71, 197–224. Swan, C. C., and Arora, J. S. 共1997兲. ‘‘Topology design of material layout

JOURNAL OF STRUCTURAL ENGINEERING © ASCE / DECEMBER 2003 / 1715


in structured composites of high stiffness and strength.’’ Struct. for structures with nonlinear material behaviors.’’ Int. J. Numer. Meth-
Optim., 13共1兲, 45–59. ods Eng., 40, 3785–3814.
Swan, C. C., Arora, J. S., Kosaka, I., and Mijar, A. R. 共1998兲. ‘‘Concept Yin, L., and Ananthasuresh, G. K. 共2002兲. ‘‘A novel topology design
design of bridge structures for stiffness and vibrations using con- scheme for multiphysics problems of electro-thermally actuated com-
tinuum topology optimization.’’ Structural engineering worldwide, N. pliant micromechanisms.’’ Sens. Actuators A, 3226, 1–12.
K. Srivastava, ed., ASCE, New York. Zou, M. 共1996兲. ‘‘Difficulties in truss topology optimization with stress
Swan, C. C., and Kosaka, I. 共1997兲. ‘‘Voigt–Reuss topology optimization and local buckling constraints.’’ Struct. Optim., 11, 134 –136.

1716 / JOURNAL OF STRUCTURAL ENGINEERING © ASCE / DECEMBER 2003

You might also like