Sunteți pe pagina 1din 23

Maute, K. et al.

: Structural Optimization

651

ZAMM  Z. Angew. Math. Mech. 79 (1999) 10, 651673

Maute, K.; Schwarz, St.; Ramm, E.

Structural Optimization
The Interaction between Form and Mechanics
Plenary lecture presented at the 76th Annual GAMM Conference, Bremen, April 69, 1998
Usually mechanical laws are applied to determine the structural response, for example deflections and stress state, while
loads, boundary conditions, and geometry of the structure, i.e. the topology and the shape, are given. However, the
mechanical principles can also be used to determine topology and shape of a structure for a prescribed structural response. This inverse method is called structural optimization. Since structural optimization deals in general with nonlinear and implicit functionals, only numerical methods have a chance to solve application-orientated problems in engineering design. Structural optimization can be distinguished into material, shape, and topology optimization depending
on what is varied in the optimization process. The most challenging task is to determine the basic geometrical layout by
topology optimization. In particular, recently the so-called material topology optimization of continuous structures has
gained substantial interest both by mathematicians as well as engineers. The present contribution tries to consolidate
these developments from an engineering point of view. In order to overcome problems of conventional numerical modeling techniques, material topology optimization is extended to geometrically adaptive methods. The adaptive discretization
of the geometrical model in topology optimization turns out to be an appropriate way to obtain reliable results and
simultaneously to reduce the numerical effort. This is verified by topology optimization problems with stress constraints
and considering elastoplastic material behavior. The optimization based on either a linear or a nonlinear structural
response leads to completely different results and shows the relevance of an appropriate mechanical model in the optimization process.
MSC (1991): 73K40, 70-02

1. Concept of structural optimization


Problems of optimal design can be traced back to the origins of structural mechanics. In 1638 Galileo Galilei
already dealt with the optimum shape of cantilever beams in his famous Discorsi (Szabo [75]). Although explicit
theories of beams and elasticity had not been developed at this time, Galileo found some remarkable results based on
his studies of the fracture behavior of beams. He concluded that the weight of a cantilever beam with constant depth
subjected to a point load at its tip is minimal, if the height decreases parabolically from the support to the tip (Fig. 1).
Subsequently, the mathematical characteristics of structural optimization are illustrated by this classical example. The Galileo problem is reformulated as follows: Find a continuous function sx which describes the height of a
cantilever beam such that the structural weight is minimum while the displacement uA at the tip does not exceed a
given limit u:

min rbsx dx ;
sx 2 C 1 `
1
`

with
A

u u  0 ;

12
`

P ` x2
dx :
Ebs3 x

The displacement at the tip uA is determined by the principle of virtual work based on the Bernoulli beam
theory. It is characteristic for problems in structural optimization that the structural response, e.g. the tip displacement uA , depends on the design function sx. The nonlinear constraint optimization problem can be transferred into a

Fig. 1. Structural optimization Galileo problem

652

ZAMM  Z. Angew. Math. Mech. 79 (1999) 10

saddle point problem described by the Lagrangean function Ls; g:


2
3

2
P
`

x
A
Ls; g rbsx dx g4 12
dx u 5 :
Ebs3 x
`

The Lagrangean multiplier of the inequality constraint is denoted by g. The saddle point of the Lagrangean
function is defined by the Euler equations, which are obtained by a variation of Ls; g with respect to the design
function sx and the Lagrangean multiplier g. The Euler equation ds Ls; g 0 leads to the solution s* predicted by
Galileo. The height s0 at the support is controlled by the displacement constraint, which is active at the optimum,
i.e. uA u:
p
4
s*x s0 1 x=` :
However, only few academic optimization problems can be analytically solved. Therefore, today numerical methods are applied approximating the optimal solution. These methods are based on a discretization of the design function
sx and the structural response, e.g. the displacements ux, on the domain W:
with

sh 2 V hs  H 1 W;

s^ 2 Rns ;

^ with
ux  u h x; u

u h 2 V hu  H 1 W;

u^ 2 Rnu :

sx  sh x; s^

5
6
h

The vectors s^ and u^ are the variable parameters of the piecewise smooth approximations s and u . With respect to
this discretization three numerical models can be distinguished (Fig. 2):
(a) D e s i g n m o d e l : Geometry and material properties of the structure are interpolated by numerical schemes, e.g. by
CADG methods (Computer Aided Geometrical Design). Based on a spatial discretization the shape or any structural and material property, e.g. the thickness or the material density, is approximated by local shape functions
 hm et al. [14], Bletzinger et al. [13]). The paramsuch as Lagrange, Bezier, and B-spline interpolations (e.g. Bo
eters of the shape functions, e.g. the position of the control nodes of splines, are the optimization variables s^.
(b) A n a l y s i s m o d e l : In structural optimization most often the finite element method is applied in order to determine the structural response u^ and to evaluate the mechanically orientated design criteria. For the optimizer the
sensitivities, the gradients with respect to s^, are of utmost interest.
(c) O p t i m i z a t i o n m o d e l : Based on the design and the analysis models the originally continuous problem can be
formulated as a parameter optimization problem on an abstract mathematical level:
^ ;
min f^
s; u
s^

with

s^L  s^  s^U ;

^ 0;
h^
s; u

h 2 Rmh ;

7
^  0;
and g^
s; u

g 2 Rmg ;

^ s. The optimization problem is


where u^ is in general an implicit function of the optimization variables: u^ u^
formulated such that the objective f is to be minimized. The range of the optimization variables is in general
restricted by lower and upper bounds s^L and s^U . The feasible design domain is defined by mh equality constraints
h and mg inequality constraints g, where the most essential equality constraint is the structural equilibrium.

Fig. 2. Numerical modelling

The optimization problem is in general nonlinear in s^ and u^ and therefore has to be solved by iterative methods.
However, most often the structural response u^ is already determined within the analysis model, since efficient finite
element procedures can be used. In this case, the state variables u^ are eliminated in the optimization problem (7) and
the optimization only determines the optimization variables s^. Depending on the characteristics of the optimization
problem, like the smoothness of the design criteria with respect to the optimization variables or the number of constraints and optimization variables, diverse mathematical programming (MP) methods or optimality criteria (OC)
methods are applied (e.g. Arora [3], Kirsch [36], Patnaik et al. [61], Vanderplaats [80]).

653

Maute, K. et al.: Structural Optimization

Structural optimization often deals with smooth problems; i.e., the design criteria are continuous smooth functions of the optimization variables. This kind of problems can be efficiently solved by gradient based optimization methods. Therefore, the sensitivities of the design criteria with respect to the optimization variables s^ have to be determined.
The sensitivities linearize the variation of the structural response due to a variation of the optimization variables, e.g. the
variation of a specific displacement due to the variation of the position of a control node, which defines the shape of a
boundary. In structural optimization the structural response is the result of two transformations (Fig. 3): (a) The transformation from the parameter space of the optimization variables into the physical space and (b) the transformation of
the structure from the undeformed into the deformed configuration. Transformation (a) is modeled in the design model,
transformation (b) in the analysis model. Consequently, sensitivity analysis is a task, which ties together the design and
the analysis model. However, since the design transformation is mostly explicit and the mechanical transformation is in
general implicit, the theoretically and computationally challenging part of sensitivity analysis belongs to the analysis
model. For a detailed introduction into sensitivity analysis see Haug et al. [31] and Kleiber et al. [37].

Fig. 3. Design
tion Sensitivity

Deforma-

In order to illustrate these general comments, a standard numerical procedure is applied to the Galileo problem presented above. The procedure essentially consists of three components: a CADG module, a finite element solver
and a gradient based optimization algorithm. The Galileo problem is transferred into a parameter optimization problem according to (7):
min f^
s w^
s=ws* ;
s^

s^L  s^  s^U ;

with
u 1  0:
g^
s u^A =
The objective is defined as the current structural weight w^
s normalized with respect to the weight ws* of the
continuous solution s* according to (4). The displacement constraint is formulated as an inequality constraint g^
s  0.
The lower and upper bounds for the optimization variable are given for the sake of completeness only, but they play
no role in the present example. Firstly, a rather coarse design model with only one free parameter, i.e. optimization
variable s^, is chosen. The change of the height is assumed to be linear (Fig. 4):
sh x s^1 x=` cx=` :
h

9
h

The height s 0 s^ of the beam at the support is variable; the height s ` c at the tip is prescribed.
The iterative optimization procedure starts with an initial guess s^0 . In order to determine a new design s^k1 the
optimization algorithm requires the values for f^
sk and g^
sk as well as their gradients, which are evaluated in the
A
analysis model. The tip displacement u^ is determined by the finite element method assuming a linear structural
response and plane stress conditions. The nodal displacements u^ are obtained solving a system of linear equations.
K u^ P :
10

Fig. 4. Galileo problem design model

654

ZAMM  Z. Angew. Math. Mech. 79 (1999) 10

Fig. 5. Galileo problem optimization model

K is the elastic stiffness matrix and P is the vector of consistent external nodal forces. The derivative of the discretized
^ s:
equilibrium condition (10) with respect to the optimization variable s^ leads to the sensitivities of the displacements du=d^
K

du^ dP dK
^ :

u^ P pseudo u
d^
s
d^
s
d^
s

11

^ s can be interpreted as the structural response due to a fictitious load case. Therefore, the rightThe sensitivities du=d^
hand side of (11) is often called pseudo-load vector P pseudo . Most often, the structural weight and its sensitivity are
also determined based on the finite element discretization.
The applied optimization algorithm, namely a sequential quadratic program (Schittkowski [69]), finds the
optimum of the parameter problem in only few steps. The optimum is the saddle point of the Lagrangean function,
which describes the parameter optimization problem in the ^
s; g-space (Fig. 5). The inequality constraint is active at
optimum s^*, i.e. g^
s* 0. However, the objective function f^
s* at the optimum is considerably larger than one, i.e.,
*
the weight w^
s is larger than the weight of the continuous solution. In order to improve the approximation of the
continuous solution, the geometrical discretization has to be refined either by increasing the number of linear segments
or by increasing the order of the shape function (Fig. 6). In both cases the approximated solution sh ^
s* approaches to
the continuous solution, if the number of the optimization variables becomes infinity, i.e. ns ! 1.

Fig. 6. Galileo problem geometrical refinement

The continuous solution could be improved, if the beam is punched or scratched and afterwards processed
again by shape optimization (Fig. 7). That way, the weight can be reduced to 25% or 10%, respectively. By means of
shape optimization only so-called continuous geometrical transformations can be described keeping the topology of the
structure, i.e., adjacent structural points remain adjacent and non-adjacent structural points remain non-adjacent.
Punching or scratching changes the topology of the structure, which means a discontinuous transformation. Discontinuous transformations can only hardly be described by standard analytical or numerical methods. However, considering the potential of improvements by changing the overall layout of the structure in the optimization process,
topology optimization is a promising approach for engineering design.
Subsequently, a brief survey of different topology optimization methods in structural optimization is given and
the most popular approach, namely material topology optimization, is discussed in detail. Today, material topology
optimization provides a robust tool to determine approximately the conceptual layout for maximum stiffness assuming

655

Maute, K. et al.: Structural Optimization

Fig. 7. Galileo problem


change of topology

a geometrically and materially linear structural response. However, topology optimization methods, which are able to
deal with more complex optimization problems, to consider a non-linear structural response or to determine a detailed
optimum layout, are still a matter of research. In the present contribution, also recent extensions of material topology
optimization to the adaptive discretization of the design function are discussed. The adaptive approach in material
topology optimization provides the possibility to solve sophisticated problems more efficiently and to consider nonlinear structural response. For example this is verified by maximizing the structural stiffness with stress constraints
and maximizing the ductility for an elastoplastic material behavior.
2. Topology optimization
First approaches in topology optimization go back to 1860 when Maxwell [50] found fundamental principles for the
layout of truss structures of minimum weight and prescribed maximum stresses. His results were restricted to structures uniformly stressed either in tension or in compression. In 1904 Michell [52] extended Maxwell's principle to
quasi-continuum truss structures stressed in tension and compression. For one load case Michell determined that the
optimum structure consists of a quasi-continuum mesh of orthogonal curved trusses (Fig. 8). These structures have no
shear stiffness and are only stable for one load case. Michell showed that those structures are simultaneously structures
of maximum stiffness for given structural mass. However, he determined geometrical optimality criteria for the optimum structural layout, but he did not develop a method to generate these structures for given external loads and
boundary conditions. Despite their academic quality Michell's results were practically not relevant from an application
point of view. Michell-structures are statically under-determined and, since even approximated Michell-structures consist of a fine mesh of curved trusses, they are very costly to manufacture.

Fig. 8. Michell-structures

In the late sixties, researchers like Hemp [32] or Prager and Rozvany [62] extended Michell's work to more
complex optimization problems with multiple constraints and load cases as well as to plate problems applying analytical methods. However, these analytical approaches were still restricted to academic problems and not applicable to real
life design problems. Only, when computer power considerably increased, numerical topology optimization methods for
discrete truss and beam structures developed into a design tool (e.g. Kirsch [35]). Most of these methods are based on
the so-called ground-structure approach (Dorn et al. [21]). A given design space is covered by a fine mesh of truss
elements defining the initial structural layout. Based on an integer formulation non-optimal truss elements are successively eliminated during the optimization process. In continuous formulations either the cross-sectional areas or the
forces of the elements are the optimization variables. A truss element is assumed to be non-optimal, i.e. not present in
the optimal structure, if the optimization variable is less than a given threshold value. A detailed survey of the evolution of topology optimization methods of discrete structures is given in Rozvany et al. [67]. However, there are two
fundamental shortcomings of discrete approaches, namely the dependency of the optimization results on the initial
ground structure and the simplified mechanical model. The a-priori assumption that the optimal structure only consists of either truss or beam elements restricts considerably the manifoldness of potential structural layouts. Therefore,
in the last ten years so-called continuous topology optimization methods gained substantial interest.

656

ZAMM  Z. Angew. Math. Mech. 79 (1999) 10

Continuous topology optimization methods can be divided into geometrical and material approaches. In the
geometrical procedure, inserting small holes changes the topology of the structure. Similar to the procedure illustrated
in Fig. 7 the optimal shape of the holes is determined by a subsequent shape optimization step. The fundamental
problem of the geometrical approach is to determine the number and position of holes to be inserted. In heuristic
methods one or more holes are inserted in low stressed areas (e.g. Rosen, Grosse [65]). In contrast, the so-called
bubble-method solves this problem by a local optimization problem (Eschenauer et al. [22], Schumacher [70]).
For a fixed stress state an infinitely small circular or elliptical bubble is inserted and its shape is varied. The hole is
inserted where the gradient of the Lagrangean function with respect to the variation of the shape of the bubble is
minimal. For selected design criteria, like strain energy, the local problem can be analytically solved leading to a
characteristic function. The bubble is inserted at the minimum of the characteristic function. In general however, the
local problem has to be solved numerically by varying the position of the bubble. The latter procedure is numerically
costly and exceeds the ability of most shape optimization programs since the bubble may move through the entire
structure. Moreover, geometric methods lead to a path-dependent optimization procedure, since a bubble of a finite
size is inserted, which means a discontinuous step in the optimization procedure. Due to the fundamental problems of
the geometrical approach most topology optimization methods are based on the material approach which is subsequently discussed in detail.
2.1 Material topology optimization

In material topology optimization the geometry of the structure is described by 01 material distribution in a given
design space Ws (Fig. 9). This leads to an indicator function cx:

1 8 x 2 Wm
;
c 2 Vc  L1 Ws :
12
cx
0 8 x 2 Ws =Wm
The structural body is defined by the set Wm of all material points x with cx 1. In order to describe an arbitrary
structural layout the indicator function cx is an element of L1 Ws , i.e., each point can be void (0) or material
(1) independently of its adjacent points. The topology optimization problem is to find the indicator function cx,
which minimizes an objective subjected to constraints. Thus, the topology optimization problem is transferred into a
material distribution problem. It is obvious, that this formulation includes topology and shape optimization.

Fig. 9. Material topology optimization

The original theory of material topology optimization was derived for the following optimization problem: Find
 is given. For the better undercx such that the strain energy in the design space Ws is minimized while the mass m
standing, the mechanical equilibrium is explicitly given in its weak form considering only forces applied on the surface
Gs of the design space:
t
min 12
e Cc e dW
13
c

with

Ws

Ws

Ws

 0;
rc dW m
deT Cc e dW

rc r0 cx ;

Gs

du T p dG ;

du; u 2 V u Ws  H 1 Ws ;

Cc C 0 cx ;


cx

1 8 x 2 Wm
;
0 8 x 2 Ws =Wm

c 2 Vc  L1 Ws :

The strain tensor is denoted by e, the material tensor by C, and the density by r. Subsequently, a linear elastic
structural response is assumed. Graphically, the optimization problem (13) means to fill a certain amount of mass of
an isotropic homogeneous material with C 0 ; r0 into the design space and to distribute the material in such a way that
the strain energy becomes minimal or, in other words, the stiffness of the emerging structure becomes maximal.
In general, this optimization problem cannot be analytically solved. Applying a standard finite element discreti^ and ch ^
zation, the indicator function and the displacements are approximated by u h u
c:
^ ;
u  u h u
h

c ;
c  c ^

u h 2 V hu Ws  V u Ws  H 1 Ws ;
ch 2 V hc Ws  H 1 Ws :

14

657

Maute, K. et al.: Structural Optimization

Fig. 10. Material topology optimization the ill-posed 01 problem

Additionally to the displacement degrees of freedom each finite element has a material degree of freedom c^i which can
be either 0 or 1. This leads to an integer optimization problem determining the elements with c^ 1 (Fig. 10).
However, integer optimization problems are in general costly to solve. Moreover, if someone compares the results for
different discretization, i.e. different finite element meshes, he would obtain quite different material distributions. The
results of the 01 problem strongly depend on the discretization and numerous local minima exist for each discretization.
2.2 Regularizationrelaxation

These numerical shortcomings can be reduced to a fundamental problem of the 01 material distribution problem.
The 01 formulation is an ill-posed non-convex optimization problem, which has no solution in the chosen subsets
u h 2 H 1 Ws and ch 2 H 1 Ws . Worth noting that this dilemma cannot only be reduced to the integer characteristic of
the indicator function c. Even optimizing continuously the thickness distribution of Kirchhoff plates and restricting the
maximum thickness, Olhoff [57] as well as Armand and Lodier [2] detected a similar numerical behavior. The finer
the discretization is, the more ribs emerge. Cheng and Olhoff [16] concluded that the optimum solution is characterized by an infinite number of ribs. Mathematically orientated studies, for example by Lurie [40], Tartare [76], and
Murat [55], corroborate this assumption. However, this feature cannot be verified by means of numerical procedures.
Finally, Kohn and Strang [38], [39] developed a mathematically rigorous method to solve these kinds of ill-posed
material distribution problems by combining three disciplines, namely structural optimization, relaxation of non-convex
functionals, and homogenization of micro-structured materials. Subsequently, this approach is discussed stressing its
physical interpretation.
Studying the characteristics of the original 01 optimization problem, it turns out that the displacement tends
to oscillate rapidly in the optimum. Physically, this means that the optimum structure consists of an arbitrary number
of arbitrarily small holes. If the characteristic size zc of the holes tends to zero and, accordingly, the frequency of the
oscillations of the displacements becomes infinite, the structural response cannot be described by means of continuum
mechanics. However, recalling the concept of the material point in continuum mechanics this problem can be overcome by introducing porous materials. Thus, rapidly oscillating displacements u z on a microscopic scale, i.e. if zc ! 0,
can be described by smoothly varying displacements u on a macroscopic scale. Introducing porous materials with a
variable density r into the optimization problem (13) can regularize the ill-posed 01 formulation. The modified
optimization problem reads as follows:
t
min 12
e Crx e dW
15
r

with

Ws

Ws

Ws

 0;
rx dW m
deT Crx e dW

Gs

du T p dG ;

du; u 2 V u Ws  H 1 Ws ;

rx fr 2 Vr  L1 Ws j 0 < r  r0 g :
The material tensor Cr describes the macroscopic behavior of the porous material with the density r made by
perforation of the basic material C 0 . Often, this kind of porous materials are visualized as micro-structured material.
The density can vary between zero and the density of the basic material r0 .
This regularization approach is often called relaxation, since the mathematical meaning of the displacements u is
enlarged: the macroscopic displacement u at a point x is a local average of the microscopic oscillations u z . However,
the relaxation has to satisfy certain conditions, which ensure that the modified problem has a solution in u 2 H 1 Ws
and u h 2 H 1 Ws , respectively, and which relate the relaxed formulation to the original optimization problem (Kohn,
Strang [39]):
(a) The modified optimization problem has to be weakly lower semi-continuous in u 2 H 1 Ws , i.e., the problem has a
continuous solution which can be approximated by minimizing sequences fu h g of numerical solutions.
(b) The infima, i.e. the strain energies, of the original 01 and of the relaxed problem are equal.
(c) The (continuous) solution of the relaxed problem converges weakly to the solution of the original problem.

658

ZAMM  Z. Angew. Math. Mech. 79 (1999) 10

The conditions (a) and (b) can be satisfied by using so-called optimal porous materials. The modified optimization
problem (15) has a solution in u 2 H 1 Ws only, if the porous material is constructed that it has a maximum stiffness
for the current density and strain state. The strain energies of optimal porous materials are a quasi-convex hull for the
strain energies of the non-convex 01 problem. If non-optimal porous materials are introduced in (15), the numerical
results do not converge. In this case, the material distribution tries to mimic locally an optimal porous material on a
macroscopic scale. Since the transition from the macroscopic to the microscopic scale is defined as the limit when the
characteristic size of the representative volume element tends to zero, the macroscopic material distributions do not
converge.
In order to prove condition (c) convergence theorems for heterogeneous materials can be transferred onto the
present problem. Consider a heterogeneous composite composed of two non-degenerated materials, i.e. C ; C  0, on
a microscopic scale y x=zc (Fig. 11). The characteristic size of the representative volume element (RVE) is denoted
by zc . On the microscopic scale the state variables u z ; ez ; s z vary rapidly but finite. The convergence theorem proves
that for any external load p 2 L2 W the state variables on the microscopic scale converge weakly in L2 W to the
macroscopic state variables u; e; s, if the characteristic size of the RVE tends to zero (e.g. Dal Maso [17]):
ux lim u z y;
zc !0

ex lim ez y;
zc !0

sx lim s z y :
zc !0

16

Fig. 11. Homogenization

The macroscopic strains and stresses are related by the homogenized material tensor C. Transferring this proof
onto the optimization problem (15) the macroscopic displacement converges weakly to the rapid oscillations on a
microscopic scale, if the material tensor C in (15) is the homogenized material tensor of the porous material. This
holds true for all design criteria, which weakly converge in the displacement u, like strain energy or eigenfrequencies. It
should be noted that this proof holds true for composites made of non-degenerated materials. Therefore, for the sake of
consistency the voids in the porous material have to be replaced by an arbitrary weak material C ! 0. Thus, the
density in the design space is always larger than zero. However, this modification has no relevance.
Due to the fundamental differentiation between macroscopic and microscopic, it is convenient to split the topology optimization problem, more precisely the objective into a macroscopic global part and a microscopic local part
(e.g. Jog et al. [33]):

min 12 minet Cr e dW ;
Cr 2 G r :
17
r

Ws

The optimum density distribution is determined in the global optimization problem. Locally, the strain energy
density is minimized varying the material tensor for given density and strain state. G r is the set of all constitutive
tensors of materials which can be composed by perforating the basic material C 0 with the density r. The complete
description of G r is not known even for a simple isotropic material C 0 . However, in order to solve the local optimization problem it is sufficient to know the convex hull of G r , i.e. the stiffest porous material for any strain state.
2.3 Material models

For an isotropic strain state the convex hull of G r corresponds to the known Hashin-Shtrikman upper bounds for
composites composed of a strong material C and a weak material C ! 0 (Hashin, Shtrikman [30]). In order to
solve the local topology optimization problem (17) also anisotropic strain states have to be considered. For this case,
several abstract approaches were developed to predict the upper bounds of material stiffness of heterogeneous material
without determining the layout of the micro-structure (e.g. Willis [81], Grabovsky [27]). However, in material topology optimization optimum micro-structure models are appreciated since here the macroscopic material properties, like
density or constitutive tensor, can be described as explicit or implicit functions of microscopic parameters. These
parameters are the optimization variables of the relaxed material distribution problem.
Up to now, only one class of micro-structure models is known describing the convex hull of Gr for anisotropic
strain states in 2-D and 3-D, namely the so-called periodically structured rank-n laminates (e.g. Francfort, Murat
[25], Milton [53]). The layering of an orthotropic rank-2 is shown in Fig. 12. On a first microscale z the RVE
consists of a layer q2 of the strong material C and a layer of a rank-1 material. On a second microscale z2 the

659

Maute, K. et al.: Structural Optimization

Fig. 12. Orthotropic rank-2 laminate

Fig. 13. Micro-hole approach

rank-1 material consists again of two layers, one layer q1 of the strong material C and one of the weak material C .
Following this scheme, hierarchically structured material of arbitrary complexity can be built up in 2-D and 3-D varying additionally the angle between the layers. Subsequently, the orthotropic rank-2 material is discussed only. This
material model is known to describe the convex hull of Gr for plane stress conditions considering only one load case in
the optimization problem. The homogenized material tensor can be analytically determined (e.g. Bendse [10]). The
coefficients of the material tensor (plane stress) are explicit functions of the layer thickness q1 and q2 . For C ! 0
and C 0 C the coefficients of the symmetric orthotropic material tensor are
q1 E0
;
q1 q2 1 n20 1 q2
~1111 ;
q2 E0 q2 n0 C

~1111
C

~1111 ;
~1122 q2 n0 C
C

~2222
C

~1212 0 :
C

18

The Young's modulus E0 and the Poisson ratio n0 are the properties of the basic material C 0 . The material
tensor is given with respect to a reference system, which is aligned to the RVE. The orientation of the RVE in the
global reference system is the third variable parameter of the rank-2 material. The density is also an explicit function
of the layer thickness q1 and q2 :
r r0 q1 q2 q1 q2 :

19

The rank-2 material is weak in shear. Thus, a structure made of rank-2 material is only stable for one load case.
This characteristic relates this kind of structures to Michell-structures. The rank-2 model can be considered as the
continuous equivalent to Michell's quasi-continuous truss structures. In the limit r ! 0 both models converge (Allaire,
Kohn [1], Bendse, Haber [8]).
The rank-2 material model provides a convenient tool to solve the material topology optimization problem (15) and
(17), respectively. However, rank-n laminates are becoming quite complex for multiple load cases and, moreover, they
provide a relaxation tool only for the currently discussed optimization problem, i.e. minimizing the strain energy for given
mass or minimizing the mass for given stiffness, respectively. Therefore, Ringertz [64] and Bense et al. [9] propose to
vary directly the coefficients of the material tensor in (17) on a macroscopic scale without any microscopic material model.
Diverse norms of the macroscopic material tensor, like its trace, are used to relate the material stiffness to the density. For
the present optimization problem, i.e. maximizing the stiffness for given mass, the rank-n material models and the macroscopic approach lead to equivalent optimization results relaxing correctly the original ill-posed optimization problem.
In the following, it will be shown for the orthotropic rank-2 material that this approach works perfectly from a
mathematical point of view, but leads to optimum material distributions with a large amount of porous material
0 < r < 1. These results can hardly be transferred into a structure consisting of only one homogeneous material C 0 and
consequently do not satisfy the 01 objective. Therefore, non-optimal material models were introduced into the continuous problem (15). These engineering approaches lead to an almost pure 01 material distribution on the macroscopic
scale, but do not satisfy the conditions for the relaxation of the ill-posed problem. Using non-optimal material models the
continuous problem is still ill-posed, i.e., the optimization results depend on the discretization of the design space Ws .
Non-optimal material models can be split up again into microscopic and macroscopic approaches. Kikuchi and
Bendse [5] introduced the most popular microscopic approach, the so-called micro-hole approach, which means a
simplification of the periodically structured orthotropic rank-2 material (Fig. 13). The dimensions of the rectangular
hole a and b as well as the orientation Q of the RVE in the global reference system are the optimization variables. In
~ has to be determined for any dimensions of the
contrast to the rank-n laminates, the macroscopic material tensor C
micro-hole by means of numerical homogenization methods (e.g. Bensoussan et al. [12], Sanchez-Palencia [68]). In
order to avoid costly homogenization procedures within the optimization process, the coefficients of the homogenized
material tensor are evaluated in advance for a finite set of dimensions of the micro-hole and interpolated by shape
functions (e.g. Guedes, Kikuchi [28]).
A simple macroscopic isotropic model roughly approximates the material stiffness-density relation of porous materials:
E E0 r=r0 b ;

r fr 2 R j 0 < r  r0 g ;

b fb 2 R j b  1g :

20

660

ZAMM  Z. Angew. Math. Mech. 79 (1999) 10

Fig. 14. Stiffness density relation for porous material models

Several authors use this or similar macroscopic isotropic models in order to formulate a continuous material distribution problem (e.g. Bendse [6], Rozvany et al. [66], Mlejnek [54], Swan, Arora [74]). The density r is the only
optimization variable. For plane stress conditions and b 1, the material distribution problem corresponds to a thickness optimization problem. Increasing the exponent b intermediate density values 0 < r < r0 become gradually unfavorable with respect to their stiffness-density relation. Therefore, intermediate densities, i.e. porous material, more
and more vanish in the optimum material distribution and almost pure 01 results are obtained.
In order to reduce the mesh dependency of the isotropic approach, Maute and Ramm [44] extended the isotropic
~ in the reference system aligned to the
model (20) to orthotropy. For plane stress conditions, the material tensor C
principal axis is
2
3
p
E1
n E1 E2 0
p
7
~ 1 6
C
21
E2
0 5;
4 n E1 E2
1 n2
0
0
G
1 n p
i 1; 2 ;
G
r r0 c1 c2 c1 c2 :
Ei E0 cbi ;
E1 E2 ;
2
Similar to the micro-hole approach, the macroscopic orthotropic model has two variables c1 and c2 , which
define the macroscopic density. Additionally, the orientation of the material in the global reference system is the third
variable.
~1111 with
In order to compare the different material models presented above the evolutions of the coefficient C
respect to the density are shown in Fig. 14. The orthotropic rank-2 material marks the upper limit for porous materials. The stiffness of non-optimal porous materials is less. Subsequently, these models are applied to solve a material
distribution problem.
2.4 The Michell problem

The influence of different material models on the optimization results is illustrated by a classic design problem, which
was already addressed by Michell [52]. The design space and boundary conditions are shown in Fig. 15. The material
distribution is determined for the optimization problem (15), i.e. maximizing stiffness for given mass, for one load case
and plane stress conditions. At the beginning of the optimization process the design space consists of equally distributed
porous material with a relative density of r=r0 20%. The design space is discretized by 4480 finite elements. On an
elemental level the displacements are approximated by 8-node serendipity shape functions; the material parameters are
constant. This discretization is fixed during the optimization process.
Since the optimization problem is characterized by a large number of optimization variables and only one equality constraint, a simple but robust and efficient optimality criteria method is applied which is subsequently outlined.
The algorithm is based on the Kuhn-Tucker condition of the resulting parameter optimization problem:
@L=@^
si @f=@^
si h @h=@^
si 0 :

22

Fig. 15. Michell problem

661

Maute, K. et al.: Structural Optimization

Fig. 16. Comparison of material models Optimized material distribution

The optimization variables s^ are the material parameters according to the current material model. The Lagrangean multiplier for the mass constraint is denoted by h. For @h=@^
si 6 0 the Kuhn-Tucker condition (22) motivates the
following updating rule:


@f=@^
si k
k1
k
k
k
k
s^i Yi p ;
Yi mk
;
Yi > 0 ;
p 2 R :
23
s^i
h @h=@^
si
k

The shift-factor mk is introduced in order to satisfy the condition Yi > 0. The exponent p controls the convergence of the optimization process. If p is too large, oscillations may occur; if p is too small, the rate of convergence is
low. Usually p is chosen in the range between 0.5 and 0.8. The Lagrangean multiplier h is determined such that the
linearized equality constraint is satisfied. This algorithm is equivalent to an inverse approximation method solved by
dual techniques (Ma et al. [41]). Since the density of the porous material does not depend on the orientation, i.e.
@h=@^
si 0, the orientation is adjusted in each iteration step of the optimization process by special algorithms (Cheng,
Kikuchi [15], Maute [49]).
In the update rule (23) the sensitivities of the objective, i.e. strain energy, and the equality constraint, i.e. structural mass, are required. Since the strain energy is self-adjoint with respect to material parameters, the sensitivity of
the objective can be determined in a post-processing step (e.g. Bendse, Kikuchi [5]). The evaluation of the gradient
of the mass is also not costly, since the density is an explicit function of the material parameters.
The optimized material distributions for different material models are shown in Fig. 16. Comparing the strain
energies of the optimum material distributions, the orthotropic rank-2 material leads obviously to the stiffest result,
the macroscopic isotropic approach to the highest strain energy. However, the stiffer the porous material is, the more
porous material occurs in the optimal material distribution. This effect is negligible in the case of the micro-hole
approach. If the b-factors in the macroscopic approaches (20) or (21) are sufficiently large, this leads to almost pure
01 material distributions on the macroscopic scale. In other papers, e.g. of Bendse and Kikuchi [5], the microhole approach leads to an almost 01 material distribution. However, these results are not based on the material
data determined directly by homogenization. In order to get an almost 01 material distribution intermediate densities are additionally penalized, i.e., the stiffness of porous material is artificially reduced.
2.5 Stabilization for non-optimal materials

From an engineering point of view, the application of non-optimal material models is favorable, since the optimization
results based on these models can be easily transferred into structures, which consist of only one homogeneous material. However, since the non-optimal models do not correctly regularize the original 01 problem the optimization
results depend more or less on the discretization of the design space. In particular, the material distributions of isotropic models depend considerably on the orientation of the finite element mesh. The results of the optimization
problem (15) based on the macroscopic isotropic material model (20) are shown for two different discretizations A and
B in Fig. 17a. In mesh B the elements are rotated by 45 degree with respect to the elements in mesh A. The optimization problems were solved by the OC algorithm (23). Depending on the orientation of the finite elements the isotropic

Fig. 17. Mesh dependency of isotropic material models a) compared to the orthotropic model b)

662

ZAMM  Z. Angew. Math. Mech. 79 (1999) 10

material model leads to different structural layouts. This shortcoming can be overcome by applying orthotropic material models. In the present example the macroscopic orthotropic material model (21) leads for both discretizations to
topologically equivalent optimization results (Fig. 17b).
However, the results of non-optimal orthotropic material models still depend on the refinement of the mesh.
Therefore, additional stabilization methods, like the perimeter-method by Haber et al. [29] or the filter-method by
Sigmund [72], are necessary in order to get numerically stable optimization results. These stabilization methods are
based on a restriction of the semi-norm of the gradient of the material distribution and introduce a kind of meshindependent smoothing. Formally, the optimization problem is extended by an additional constraint:


@r^
s


 2 R :
R
24
@x dW R  0 ;
L2
Ws

The optimization variables s^ are the material parameters of the finite elements according to the applied material
 means the perimeter of the structure. In the so-called perimetermodel. For an almost 01 material distribution R
method the constraint (24) is explicitly considered solving the modified optimization problem. In the filter-method the
restriction of the semi-norm of the material distribution is indirectly considered by modifying the objective, i.e. optimizing a smoothed strain energy. The stabilized formulation of the optimization problem is
P
i
min
Hi P ej ^
s
25
elements

with

Ws

 0:
rx dW m

P ej denotes the elemental strain energy of the finite element j which borders to the element i. The smoothing operator Hi j is defined by
.P
P
~
~ i j ;
~i j
H
Hi
H
j

~i j
H

rk j 8 rij  r
2 rij =
;
0
8 rij > r

~
~ i j
H

rk jWj j
2 rij =
0

8 rij  r
:
8 rij > r

26

The smoothing has an effect on all elements within a distance smaller than r. The radius rij denotes the distance
i
between the midpoints of the elements i and j. The exponent k scales the elemental strain energy P ej with respect
to the distance. It can be shown that the larger the smoothing is, the smaller the semi-norm of the optimal material
 in (24) is implicitly set by the parameters k and r.
distribution is. The limit R
With the example in Fig. 18 the difference between the original and the stabilized formulations are illustrated.
The optimization problem (15) is solved for different discretizations of the design space, i.e. a coarse mesh with 2  320
elements and a fine one with 2  2880 elements. This problem was already addressed by Sigmund [72] applying the
macroscopic isotropic material model (20), which is known to lead to extremely mesh dependent results. Here, the
stabilization by the filter-method is verified for the macroscopic orthotropic material model (21). The optimal material
distributions were determined by the OC algorithm (23) considering the symmetry of the optimization problem. The
results of the non-stabilized formulation clearly show that the results of non-optimal orthotropic models depend on the
discretization of the design space. A finer discretization leads to a more detailed structure. In contrast to this, the
stabilized formulation converges to one solution refining the discretization of the design space.
2.6 Application of topology optimization a short survey

Based on the material models presented above maximum stiffness problems have been solved in 2-D and 3-D (e.g.
Bendse, Mota Soares [7], Bendse [10]). For plate and shell problems the material models are extended to layered
composites varying the density in one or more layers (e.g. Diaz et al. [20], Maute, Ramm [43]). The extension of
maximum stiffness problem to multiple load cases is straightforward (Diaz, Bendse [19], Maute, Ramm [43]). Moreover, material topology optimization is applied in order to solve a broad range of design problems, e.g. maximizing the
structural flexibility in selected points (Frecker et al. [26]), tuning of eigenfrequencies (e.g. Diaz, Kikuchi [18],
Tenek, Hagiwara [79]) or determining the micro-structural layout of heterogeneous material for desired macroscopic
properties (Sigmund [73]). In the geometrically nonlinear regime the bucking load is maximized based on a linearized
eigenvalue problem (Neves et al. [56], Maute [49]). In the materially nonlinear regime, the structural ductility for
elastoplastic materials is maximized (e.g. Yuge, Kikuchi [83], Maute et al. [48]) and the optimum reinforcement
layout is determined for quasi-brittle materials, like concrete (Ramm et al. [63]).
Even if this list is definitely not complete, it seems that material topology optimization provides a general tool
for structural design. However, up to now the mathematical theory is only proved for maximum stiffness problems. In
particular, for tuning eigenfrequencies and maximizing buckling loads numerical instabilities are reported which can be

Maute, K. et al.: Structural Optimization

663

Fig. 18. Stabilization for non-optimal material models by the filter-method

reduced to an insufficient relaxation of the related 01 problem by the applied material models (Maute [49]). In
order to overcome these problems, fundamental research is required.
The numerical modeling of the material distribution inhibits material topology optimization solving more complex optimization problems with several equality and inequality constraints. The topology optimization problems mentioned above are mainly characterized by one constraint only. Due to the fixed discretization of the design space a fine
finite element mesh and accordingly a large number of optimization variables are necessary in order to identify in
detail the optimum structural layout by the material distribution. Moreover, a fixed discretization leads to discontinuous material distributions with jagged boundaries between void and non-void areas. Along these boundaries artificial
stress singularities arise which may affect the quality of the structural analysis, particularly with regard to a nonlinear
structural response. The problems caused by the rigid numerical modeling motivate to adapt the discretization of the
design space to the current material distribution during the optimization process.
3. Adaptivity in structural optimization
Conventionally, at the beginning of the optimization process the degree of discretizations of both the design function
sh and the displacements u h are fixed and the optimization problem is solved. However, in the course of the optimization process the design often changes considerably and the initial discretization approximates the continuous solutions
s* and u only roughly. This leads to non-optimal designs. Consequently, adaptive procedures are necessary to adapt
the discretization of sh and u h to the current state in the optimization process. Adaptivity in structural optimization
contains both the adaptation of the design function sh to the optimum structure and the adaptation of the state
variables u h to the structural response. The latter aspect, the so-called mechanical adaptivity, is addressed within the
framework of structural optimization e.g. in Bennet, Botkin [11], Kikuchi et al. [34], Banichuk et al. [4], Maute,
Ramm [46]. In the present study the adaptation of the design function sh , the so-called geometrical adaptivity, is
considered only.
The basic idea of geometrical adaptivity has already been illustrated by the Galileo problem (Fig. 6). In order to
improve the approximation of the continuous solution, the discretization of the variable boundary was refined either by
increasing the number of segments in an h-strategy or by increasing the order of the shape functions by a p-strategy. In
this example the refinement has been uniform. However, in order to reduce discretization error as much as possible
with a minimum number of optimization variables, the refinement has to be locally controlled by refinement indicators.
Up to now, only few schemes have been described to adapt the discretization of the design function s. The advantages
of design adaptivity in shape optimization were verified by Mathiak, Schnack [42] and Falk et al. [23] refining the
discretization of variable boundaries. Maute and Ramm [44], [47] presented the benefits of an adaptive approach in

664

ZAMM  Z. Angew. Math. Mech. 79 (1999) 10

material topology optimization. Subsequently, the essential features of design adaptivity are illustrated by an example
in shape optimization and transferred to material topology optimization of plane and shell structures. The adaptive
approach in material topology optimization provides the possibility to solve efficiently problems with local constraints
or elastoplastic material behavior.
3.1 Geometrical adaptivity in shape optimization

The essential features of design adaptivity are illustrated by the so-called fillet problem. In Fig. 19 only the upper half
of the symmetric structure is shown. The initial design is a rectangular membrane clamped on its left edge and subjected to a uniform horizontal load on its right edge. The shape of the marked segment has to be determined such that
the structural weight is minimal. A C 2 -continuity for the design model is required. The maximum allowable equivalent
stress is restricted to sv . The stress resultant is kept constant during the optimization process. The design function
s 2 Vs G  C 2 G describes the variable boundary G. In order to solve numerically the optimization problem the
shape x h of the variable boundary G is approximated by a C 2 -continuous cubic B-spline:
P
x h z; x c
Fi z x ci :
27
The shape functions are denoted by Fi . The parameter z marks the position in a local convective reference system.
The vertical positions of the control nodes x ci are the optimization variables s^. The geometrical discretization spans a
finite-dimension space V hs G, which has to be bounded and closed with respect to V s G. This guarantees that the
numerical solution sh converges to the continuous one s* increasing the refinement of the discretization, i.e. decreasing
the characteristic size h of the subdomains. Without proof this condition is supposed to be satisfied by the chosen
discretization.

Fig. 19. Fillet problem

Adapting the geometric discretization can be interpreted as a generic optimization problem: Determine the discretization such that the continuous solution is approximated as accurate as possible, i.e., the geometrical discretization error is minimum for a given number of optimization variables. The geometrical discretization error kes k can
either be defined by the Lp -norm of the difference between the continuous and the approximated solution,

1=p
jx h ^
s* x *jp dG
;
p fp 2 R j p  1g ;
28
kes kLp
G

s* of the continuous and the approxior by the absolute difference between the objectives f * fs* and f h f^
mated solution,
kes kf jf h f *j :

29

In contrast to error estimates of the finite element method the geometrical discretization error cannot be estimated by local residua or discontinuities of the numerical solution, since in general optimization problems cannot be
reduced to local equilibrium conditions. A robust but costly a-posteriori estimate of the geometrical discretization
error is based on an asymptotic expansion of the numerical solutions f h or kx h ^
s*k increasing gradually the refinement of the discretization h ! 0:
kes kf jf h f *j t1 hk t2 h2k tr h3k HOT :

30

The parameters ti and k depend on the optimization problem and the current discretization. Based on the
results of different discretizations, i.e. different number of optimization variables, these parameters can be determined
solving a system of nonlinear equations. The error estimate determines the quality of the numerical solution. If the
discretization error is larger than a given limit, a refinement indicator is necessary in order to refine locally the discretization.
In general, it is not possible to derive an error distribution, i.e. a local refinement indicator, from the asymptotic
error estimate (30). Up to now, only heuristic refinement indicators have been developed. Mathiak and Schnack [42]
compare the optimized shapes of two different discretizations and increase the refinement where the local Lp -norm of

Maute, K. et al.: Structural Optimization

665

Fig. 20. Adaptive and non-adaptive geometrical refinement

the difference between the continuous and the approximated solution is maximum. Alternatively, the same authors
propose to increase the refinement in splines with large curvature. Falk [24] introduces additional optimization variables s^a . The gradient of the Lagrangean function with respect to these variables serves as refinement indicator. If this
gradient exceeds a given limit, the related optimization variable is added. Based on the refinement indicator the discretization is locally refined and the updated parameter optimization problem is solved. In shape optimization two p- and
h-adaptive strategies can be applied. The r-adaptation, i.e. adapting the position of the control nodes of the splines, is
already included by definition in shape optimization.
In the present example an h-adaptive strategy is applied increasing locally the number of control nodes of the
B-spline, i.e. decreasing the characteristic length of the cubic segments. Starting with a coarse discretization additional
control nodes are added where the local curvature of the shape is maximum. Each parameter optimization problem is
solved by an SQP algorithm (Schittkowski [69]). The structural response is determined by a finite element analysis
with 1152, 8-node plane-stress finite elements. The analytical sensitivities of the design criteria are calculated by a
discrete direct approach. For comparison the discretization is uniformly increased. In Fig. 20 the minimum weight with
respect to the initial weight, i.e. optimized objectives for the different discretizations, over the number of design variables are given for the adaptive and the uniform refinement procedures in a double-logarithmic scale. The optimized
shapes and the related equivalent stress distributions are shown. Controlling the local refinement by a curvature driven
criteria works well in the present example, since the exact optimum shape has an almost discontinuous jump from the
loaded to the supported domain.
In Fig. 21 the exact error and the estimated relative discretization error kes kf are given for the uniform discretization. The exact error is defined by the difference between the current minimum weight and the minimum weight
of a reference solution obtained by an adaptive discretization with 27 optimization variables. The estimated error is
determined by the asymptotic expansion (30) based on the minimum weights of three different successive discretizations. Due to the small number of optimization variables the quality of the error estimates is poor. However, the
adaptive refinement of the geometrical discretization provides the possibility to generate automatically an efficient
design model. This was illustrated by the present shape optimization problem and will be subsequently transferred to
topology optimization.

Fig. 21. Exact and estimated discretization error

666

ZAMM  Z. Angew. Math. Mech. 79 (1999) 10


3.2 Adaptive topology optimization an example

In principle, the same adaptive techniques presented for design adaptivity in shape optimization, such as h-, p-, and
r-adaptive strategies, can also be applied in material topology optimization in order to adapt the discretization of
the material distribution during the optimization process. Analogously to the curvature of the shape the spatial
gradient of the material distribution can be used as refinement indicator. For non-optimal material models this
concept leads to a refinement of the discretization in subdomains with porous materials. However, some specific
features of material topology optimization should be considered:
(a) In contrast to shape optimization the discretizations in the design and in the analysis model are coupled, i.e., the
displacements and the material distribution are approximated on the same finite element mesh.
(b) A p-adaptive strategy requires an adaptation of the order of the shape functions for the displacements, since the
order of the shape functions for the material distribution and the displacements have to satisfy a Babuska-Brezzitype condition (Bendse [10]). Otherwise, so-called checkerboard modes arise in the optimized material distribution.
(c) The optimum material is discontinuously distributed in the design space. The border between void areas and areas
with basic material defines the contour of the optimum structure. In order to determine in detail the shape of
contours, a locally refined mesh has to be generated in this region (Fig. 22). However, this leads to a large number
of elements and optimization variables, respectively. Alternatively, following an r-adaptive strategy the elements
can be aligned along the contours.
(d) Generally, void and non-void areas can already be identified after only few iterations in the optimization process.
Therefore, the numerical efficiency can be considerably increased by neglecting the void areas adapting the discretization in the design space. However, this may not lead to an irreversible process.

Fig. 22. Approximation of a discontinuous material distribution

These specific features are considered by an adaptive topology optimization procedure proposed by Maute and
Ramm [44]. Subsequently, the essential points of this procedure are illustrated by a shell example. In Fig. 23 a segment
of a spherical shell is subjected to a concentrated load in the vertex and vertically supported in the four corners. Due
to symmetry of the problem only one quarter of the shell is analyzed and optimized. The objective is maximum stiffness, i.e. minimum strain energy. The mass is kept constant during the optimization process. The macroscopic isotropic material model (20) is applied. Firstly, the design space is discretized by a rather coarse mesh using 4  36,
eight-node shell elements with constant material distribution. The parameter optimization problem is solved by the
OC-algorithm (23). In the optimum the structural layout can hardly be identified, since in the vertex area porous
material is generated.
Therefore, the discretization is refined in areas with porous material. Additionally, void areas are neglected
approximating the contours of the adapted design space by splines. The adaptation procedure is conveniently carried
out in the parameter space of the shell mid-surface (Fig. 24). Firstly, the material distribution optimized for the coarse
mesh is mapped onto a background mesh in the parameter space. The background mesh serves to store the optimization result of the previous optimization step; it consists of a large number of design patches in order to provide a high
resolution and to avoid mapping errors. The background mesh spans always the original design space during the overall optimization process allowing the adapted design space to shrink and swell. Based on the material distribution in
the background mesh the contours of non-void areas are approximated by splines. The contours are identified by

Fig. 23. Topology optimization of a spherical shell structure

Maute, K. et al.: Structural Optimization

667

Fig. 24. Adaptation of the design model in topology optimization

isolines rc =r0 of the density distribution. Areas with an average density below a threshold value r  rc are supposed
to be void. The non-void areas, i.e. adapted design space, are discretized by a free mesh generator refining locally the
mesh in areas with porous material. The material distribution stored in the background mesh is mapped onto the
adapted design model, which is finally transferred in the physical space. The material parameters of the finite elements
are the optimization variables of the following optimization step.
This procedure is applied to determine the structural layout of the shell problem in detail. In Fig. 25 the adapted
design spaces and the related optimum material distributions are given. Three cycles of optimization runs for the
respective discretizations are necessary until the topology can be clearly identified. Additionally, a smooth approximation of the optimum shape is obtained. During the course of the optimization process the objective, the strain energy,
related to the one of the initial design, is gradually decreased refining the design model. The peaks result from mapping
errors during the mesh changes.

Fig. 25. Adaptive topology optimization

668

ZAMM  Z. Angew. Math. Mech. 79 (1999) 10

Fig. 26. Comparison of adaptive and conventional topology optimization

In order to compare the adaptive method with the conventional procedure the design problem was also solved in
32 iterations using one fixed fine mesh (900 patches per quarter) during the entire optimization process. The final
layouts obtained by the adaptive scheme and the conventional non-adaptive procedure are shown in Fig. 26. Concerning the topology and the shape the optimization results are almost equivalent, but the overall numerical effort can be
reduced by the adaptive method for this example to 16 percent in comparison to the conventional procedure.
However, despite a distinct improvement the optimal shape of the outer and inner boundaries cannot be determined in detail by adaptive material topology optimization alone, since an even finer discretization along the boundaries would be necessary. Therefore, it is appropriate to include also shape optimization in the overall optimization
procedure. The integration of topology and shape optimization in the same process is addressed by Olhoff et al. [58],
Papalambros, Chirehdast [60] and integrated in an adaptive approach by Maute and Ramm [45].
3.3 Adaptive topology optimization case studies

Two advantages of design adaptivity in material topology optimization are verified by the shell example presented
before, i.e. the numerical efficiency and the clearly defined optimization results. Further advantages of this approach
are emphasized by the following two design problems, which can hardly be solved by the conventional procedure.
A fixed descretization leads to discontinuous material distributions with jagged boundaries between void and non-void
areas. Along these boundaries artificial stress singularities arise which may affect the quality of the structural analysis,
particularly with regard to nonlinear structural response. Furthermore, it is difficult to control stress driven design
criteria during the optimization process. These problems can be overcome by the adaptive procedure, which is verified
by the following two design problems: maximizing the structural stiffness with stress constraints and maximizing the
ductility for a nonlinear material behavior.
3.3.1 Material topology optimization with stress constraints

In order to get a reliable design the maximum allowable stress of the material has to be considered. The structural
response of a design generated by topology optimization maximizing the stiffness is characterized by an almost homogeneous stress state. With respect to the topology optimization problems discussed before the overall stress level corresponds to the prescribed mass in the design space. However, local stresses in supported or loaded areas may considerably exceed the overall level. Therefore, it is in general necessary to consider stress constraints.
In topology optimization it is not advisable to consider stress constraints as local criteria. This would lead to
large number of constraints and a costly sensitivity analysis. Therefore, stress constraints are introduced in an integral
penalty formulation (Yang [82]). The modified problem is
8 
91=p

<
=

sv ^
s p
i
 0:
min P ^
s ws
dW
;
r^
s dW m
31

s^
s
v
:
;
Ws
Ws

The maximum allowable equivalent stress is denoted by sv . The penalty term is weighted by ws . An exponent
p > 1 can avoid extreme local violation.
In Fig. 27 the optimization problem (31) is solved for a membrane based on a macroscopic isotropic material
model (20). The sensitivities of the strain energy are self-adjoint, the gradients of the penalty term are determined by a
variational adjoint approach. The filter-method is applied in order to stabilize the optimization results. Clearly defined
structural layouts are generated by the adaptive procedure in three cycles (Fig. 28). Stress singularities are avoided
adapting the contours of the effective design space to the material distribution.
For comparison the corresponding optimization problem is solved neglecting the stress constraints. This leads to
highly stressed areas near the supports. Most often stress constraints result only in an extension of structural elements.
In the present example due to the specific support conditions a different structural layout is generated, if the maximum
allowable stress is considered. It is remarkable, that in contrast to the results obtained for pure stiffness maximization

Maute, K. et al.: Structural Optimization

669

Fig. 27. Topology optimization with


stress constrains

with non-optimal material models large areas with porous material exist in the optimum material distribution if additionally stress constraints are considered. It is still an open question whether this indicates that porous material even
with non-optimal properties is favorable in maximum stiffness design if the maximum stress is restricted or whether
this is only a numerical problem.

Fig. 28. Material topology optimization with and without stress constraints

3.3.2 Topology optimization in elastoplasticity

Material topology optimization is extended to nonlinear material behavior in order to get a save design in the critical
and post-critical range, as it is typical for crash or earthquake loading. The present optimization procedure is based on
a von Mises elastoplasticity model with linear, isotropic work-hardening/softening for small strains in plane stress conditions. The objective of the design problem is to maximize the ductility of the structure for a given range of pre The ductility is defined by the integral of the strain energy over the range u.
 The mass m
 in
scribed displacements u.
the design space Ws is prescribed:
T
s de dW
32
min
s^

with

Ws

Ws e

 0
r dW m

 The ductility characterizes the structural


where e are the total strains according to the prescribed displacements u.
response in the elastic and elastoplastic range.

670

ZAMM  Z. Angew. Math. Mech. 79 (1999) 10

Similar problems have been addressed by Taylor, Washabaugh [78] for truss structures, Yuge, Kikuchi [83]
for beam structures, and Mayer et al. [51] optimizing the crashworthiness of shell structures. Subsequently, the essential points of the optimization algorithm are briefly discussed stressing the specific role of design adaptivity for topology optimization in elastoplasticity; details concerning the overall algorithm, the sensitivity analysis and the numerical
implementation are given in Maute et al. [48].
(a) The classical von Mises material is transferred to a macroscopic isotropic porous material model in order to regularize the topology optimization problem. Analogously to the b-power law (20) nonlinear relations are arranged for
the Young's modulus E, the plastic hardening modulus E h and the initial yield stress sy :
E E0 r=r0 b1 ;

E h E0h r=r0 b2 ;

bi fbi 2 R j bi > 1g ;

sy sy0 r=r0 b3 ;

i 1; 2; 3 :

33

This approach leads to an almost 01 material distribution in the optimum. However, it should be noted that
this material model may not correctly relax the underlying 01 problem.
(b) The structural response is determined by a displacement-controlled Newton-Raphson algorithm. The sensitivities of
ductility are computed by a variational adjoint approach within the path-following algorithm based on the consistent material tangent.
(c) The adaptation of the design model, i.e. the discretization of the material distribution, plays an important role
particularly with respect to reliable results of the structural analysis and numerical efficiency. Jagged boundaries
generated by the conventional procedure with a fixed discretization cause artificial yielding which affects the overall structural response. This defect is avoided by approximating repeatedly the effective design space by smooth
contours. Additionally, reducing the number of elements the numerical effort for the nonlinear structural analysis is
considerably decreased.
The proposed algorithm is verified by an example for plane stress conditions. The deep beam in Fig. 29a is clamped on
both sides and subjected to a vertical load in the center of the upper edge. The objective of the design problem is to
maximize the structural ductility for a range of prescribed vertical displacements uA . The degrees of freedom of the
loaded nodes are coupled. Due to the symmetry of the problem only one half of the structure is analyzed. The optimization problem is solved by the OC-method (23). The discretization of the design space is adapted analogously to the
previous examples.

Fig. 29. Adaptive topology optimization including elastoplastic material behavior

Maute, K. et al.: Structural Optimization

671

The adaptive optimization result considering an elastoplastic material behavior in the optimization process is
shown in Fig. 29b. For comparison, the optimized layout is determined assuming an elastic structural behavior
(Fig. 29c). In the elastic case maximizing the ductility is equivalent to minimizing the strain energy. Afterwards the
structural ductility for the elastic optimum is determined for an elastoplastic material behavior as well and compared
to the ductility of the elastoplastic optimum. The related stress states are given in Fig. 29d. The load-deflection
diagrams show that the load of the elastoplastic optimum is about 22 percent larger in comparison to the one of the
elastic optimum (Fig. 29e). The stiffness of the elastoplastic optimum is slightly smaller in the elastic range, but
yielding begins at a higher load level. The ductility may even be further increased if additional shape optimization
steps are added subsequently, Schwarz et al. [71].
4. Conclusions
Material topology optimization provides a powerful tool in structural optimization reflecting the interaction between
form and mechanics. The optimization problems is transferred into an ill-posed non-convex 01 material distribution
problem. For maximum stiffness problems the ill-posed problem can be regularized introducing so-called optimal porous
material with variable density. However, these material models lead to optimal material distributions with a large
amount of porous material. Thus, the optimum layout of a structure, which consists of only one homogeneous material,
can hardly be identified. Therefore, engineers introduced non-optimal material models leading to almost 01 material distributions on a macroscopic scale. However, non-optimal materials have to be stabilized by imposing an additional constraint on the smoothness of the material distribution, in order to get mesh-independent optimization results.
Today, material topology optimization is tackling a broad range of design problems with diverse design criteria
and considering a nonlinear structural response. This raises two problems: (a) the relaxation of the 01 for other
design criteria and (b) the numerical modelling. Due to a fixed discretization of the design space a large number of
optimization variables is necessary to describe the optimum structural layout. Additionally, the contours of the structure are represented by jagged boundaries, which may affect the quality of the structural analysis, particularly with
regard to a nonlinear structural response. The shortcoming of conventional topology optimization techniques can be
overcome adapting the discretization of the design space to the current material distribution during the optimization
process.
The concept of geometrical adaptivity was illustrated by a shape optimization example and transferred to material topology optimization. In both disciplines geometrical adaptivity provides a promising tool in order to increase the
quality of the optimization results and the numerical efficiency. In particular, design adaptivity plays an important role
in material topology optimization in order to consider more sophisticated design problems. This was verified by maximizing the stiffness with stress constraints and maximizing the ductility for a materially nonlinear structural response.
Moreover, the optimization results of the latter design problem clearly show that it is necessary to include the nonlinear structural response in the optimization problem.
Acknowledgements
This work is part of the DFG research projects Ra 218/16-3 Adaptive Methods in Topology Optimization and Ra 218/11-3 Algorithms, Adaptive methods, Elastoplasticity. The support is gratefully acknowledged. The second author would like to acknowledge
the financial support provided by the Swiss foundation `Besinnung und Ordnung'.

References
1 Allaire, G.; Kohn, R. V.: Explicit optimal bounds on the elastic energy of a two-phase composite in two space dimensions.
Quart. Appl. Math. 60 (1993), 675699.
2 Armand, J.-L.; Lodier, B.: Optimal design of bending element. Internat. J. Numer. Mech. Engin. 13 (1978), 373384.
3 Arora, J. S.: Introduction to optimum design. McGraw-Hill, New York 1989.
4 Banichuk, N. V.; Barthold, F.-J.; Falk, A.; Stein, E.: Mesh refinement for shape optimization. Struct. Optimiz. 9 (1995),
4651.
5 Bendse, M. P.; Kikuchi, N.: Generating optimal topologies in structural design using a homogenization method. Comput. Meth.
Appl. Mech. Engin. 71 (1988), 197224.
6 Bendse, M. P.: Optimal shape design as a material distribution problem. Struct. Optimiz. 1 (1989), 193202.
7 Bendse, M. P.; Mota Soares, C. A. (eds.): Topology design of structures. Kluwer Academic Publishers, Dordrecht 1993.
8 Bendse, M. P.; Haber, R. B.: The Michell layout problem as a low volume fraction limit of the perforated plate topology
optimization problem. Struct. Optimiz. 6 (1993), 263267.
9 Bendse, M. P.; Guedes, J. M.; Haber, R. B.; Pedersen, P.; Taylor, J. E.: An analytical model to predict optimal material
properties in the context of optimal structural design. J. Appl. Mech. 61 (1994), 930937.
10 Bendse, M. P.: Optimization of structural topology, shape, and material. Springer-Verlag, Berlin 1995.
11 Bennet, J. A.; Botkin, M. E.: Structural shape optimization with geometric description and adaptive mesh refinement. AIAA
J. 23 (1985), 458464.

672

ZAMM  Z. Angew. Math. Mech. 79 (1999) 10

12 Bensoussan, A.; Lions, J.-L.; Papanicolaou, G.: Asymptotic analysis for periodic structures. North-Holland, Amsterdam
1978.
13 Bletzinger, K.-U.; Kimmich, S.; Ramm, E.: Efficient modeling in shape optimal design. Comput. Syst. Engin. 2 (1991),
483495.
 hm, W.; Farin, G. E.; Kahmann, J.: A survey of curve and surface methods in CAGD. Computer Aided Design 1 (1984),
14 Bo
160.
15 Cheng, H. C.; Kikuchi, N.; Ma, Z. D.: An improved approach for determining the optimal orientation of orthotropic material.
Struct. Optimiz. 8 (1994), 101112.
16 Cheng, K.T.; Olhoff, N.: An investigation concerning optimal design of solid elastic plates. Internat. J. Solids Struct. 17
(1981), 305323.
17 Dal Maso, G.: An introduction to G-convergence. Birkhauser, Boston 1993.
18 Diaz, A.; Kikuchi, N.: Solutions to shape and topology eigenvalue optimization problems using a homogenization method. Internat. J. Numer. Meth. Engin. 35 (1992), 14871502.
19 Diaz, A.; Bendse, M. P.: Shape optimization of structures for multiple loading conditions using a homogenization method.
Struct. Optimiz. 4 (1992), 1722.
20 Diaz, A.; Lipton, R.; Soto, C.: A new formulation of the problem of optimum reinforcement of Reissner-Mindlin plates. Comput. Meth. Appl. Mech. Engin. 123 (1995), 121139.
21 Dorn, W. S.; Gomory, R. E.; Greenberg, H. J.: Automatic design of optimal structures. J. Mec. 3 (1964), 2552.
22 Eschenauer, H. A.; Kobelev, V. V.; Schumacher, A.: Bubble method for topology and shape optimization of structures.
Struct. Optimiz. 8 (1994), 4251.
23 Falk, A.; Barthold, F.-J.; Stein, E.: Hierarchical modelling in shape optimization. In: [59], pp. 371376.
24 Falk, A.: Adaptive Verfahren f
ur die Formoptimierung von Flachentragwerken unter Ber
ucksichtigung der CAD-FEM-Kopplung.
Ph.D. dissertation, University of Hanover, Hanover, Germany, 1995.
25 Francfort, G. A.; Murat, F.: Homogenization and optimal bounds in linear elasticity. Arch. Rat. Mech. Anal. 94 (1986),
307334.
 z, Z. (eds.): Pro26 Frecker, M.; Kikuchi, N.; Kota, S.: Optimal design of compliance mechanism. In: Gutkowski, W.; Mro
ceedings of the Second World Congress of Structural and Multidisciplinary Optimization, May 2630 1997, Zakopane, Poland,
pp. 345350.
27 Grabovsky, Y.: Bounds and extremal microstructures for two-component composites: A unified treatment based on the translation method. Proc. Roy. Soc. London A 452 (1996), 919944.
28 Guedes, J. M.; Kikuchi, N.: Pre- and post-processing for materials based on the homogenization method with adaptive finite
element methods. Comput. Meth. Appl. Mech. Engin. 83 (1990), 143198.
29 Haber, R. B.; Jog, C. S.; Bendse, M. P.: A new approach to variable-topology shape design using a constraint on perimeter.
Struct. Optimiz. 11 (1996), 112.
30 Hashin, Z.; Shtrikman, S.: A variational approach to the theory of the elastic behavior of multiphase materials. J. Mech. Phys.
Solids 11 (1963), 127140.
31 Haug, E. J.; Choi, K. K.; Komkov, V.: Design sensitivity analysis of structural systems. Academic Press, Orlando 1986.
32 Hemp, W. S.: Optimum structures. Clarendon Press, Oxford 1973.
33 Jog, C. S.; Haber, R. B.; Bendse, M. P.: Topology design with optimized, self-adaptive materials. Internat. J. Numer. Meth.
Engin. 37 (1994), 13231350.
34 Kikuchi, N.; Chung, K. Y.; Torigaki, T.; Taylor, J. E.: Adaptive finite element methods for shape optimization of linearly
elastic structures. Comput. Meth. Appl. Mech. Engin. 57 (1986), 6789.
35 Kirsch, U.: Optimal topology of structures. Appl. Mech. Rev. 42 (1989), 223239.
36 Kirsch, U.: Structural optimization: fundamentals and applications. Springer-Verlag, Berlin 1993.
nez, H.; Hien, T. D.; Kowalczyk, P.: Parameter sensitivity in nonlinear mechanics. Wiley, Chichester
37 Kleiber, M.; Antu
1997.
38 Kohn, R. V.; Strang, G.: Optimal design for torsional rigidity. In: Atluri, S. N.; Gallagher, R. H.; Zienkiewicz, O. C.
(eds.): Hybrid and mixed finite element methods. Wiley, Chichester 1983, pp. 281288.
39 Kohn, R. V.; Strang, G.: Optimal design and relaxation of variational problems. IIII. Comm. Pure Appl. Math. 39 (1986),
113137, 139182, 353377.
40 Lurie, K. A.: On the optimal distribution of the resistivity tensor of the working substance in a magnetohydrodynamic channel.
J. Appl. Math. Mech. 34 (1970), 255274.
41 Ma, Z.-D.; Kikuchi, N.; Cheng, H.-C.; Hagiwara, I.: Topological optimization technique for free vibration problems. ASME
J. Appl. Mech. 62 (1995), 200207.
42 Mathiak, G.; Schnack, E.: An adaptive composition of unknown boundaries for shape optimization of three-dimensional elastic
structures. In: [59], pp. 365370.
43 Maute, K.; Ramm, E.: Topology optimization of plate and shell structures. In: Abel, J. F.; Leonhard, J. W.; Penalba, C. U.
(eds.): Spatial, lattice and tension structures Proceedings of the IASS-ASCE International Symposium, Atlanta, April 1994.
American Society of Civil Engineering, New York 1994, pp. 946955.
44 Maute, K.; Ramm, E.: Adaptive topology optimization. Struct. Optimiz. 10 (1995), 100112.
45 Maute, K.; Ramm, E.: General shape optimization an integrated model for topology and shape optimization. In: [59],
pp. 299306.
46 Maute, K.; Ramm, E.: Adaptivity in structural optimization. In: Topping, B. H. V. (eds.): Advances in structural engineering
optimization. Civil-Comp. Press, Edinburgh 1996, pp. 227237.
47 Maute, K.; Ramm, E.: Adaptive topology optimization of shell structures. AIAA J. 35 (1997), 17671773.
48 Maute, K.; Schwarz, S.; Ramm, E.: Adaptive topology optimization of elastoplastic structures. Struct. Optimiz. 15 (1998),
8191.
49 Maute, K.: Topologie- und Formoptimierung von d
unnwandigen Tragwerken. Ph.D. dissertation. University of Stuttgart, Stuttgart, Germany, 1998.
50 Maxwell, C.: Scientific papers. II. Cambridge Univ. Press, Cambridge 1869.
51 Mayer, R. R.; Kikuchi, N.; Scott, R. A.: Application of topology optimization techniques to structural crashworthiness. Internat. J. Numer. Meth. Engin. 39 (1996), 13831403.
52 Michell, A. G. M.: The limits of economy of material in frame structures. Philos. Mag. Ser. 6, 8 (1904), 589597.

Maute, K. et al.: Structural Optimization

673

53 Milton, G.: Modelling the properties of composites by laminates. In: Erickson, J. L.; Kinderlehrer, D.; Kohn, R. V.;
Lions, J. L. (eds.): Homogenization and effective moduli of materials and media. Springer-Verlag, Berlin 1986, pp. 150175.
54 Mlejnek, H. P.: Some aspects of the genesis of structures. Struct. Optimiz. 5 (1993), 6469.
55 Murat, F.: Contre-examples pour divers problemes ou le contr^ole intervient dans les coefficients. Ann. Mat. Pur. Appl. 112
(1977), 4968.
56 Neves, M. M.; Rodrigues, H.; Guedes, J. M.: Generalized topology design of structures with a buckling load criterion. Struct.
Optimiz. 10 (1995), 7178.
57 Olhoff, N.: Optimal design of vibrating rectangular plates. Internat. J. Solids Struct. 10 (1974), 93109.
58 Olhoff, N.; Bendse, M. P.; Rasmussen, J.: On CAD-integrated structural topology and design optimization. Comput. Meth.
Appl. Mech. Engin. 89 (1991), 259279.
59 Olhoff, N.; Rozvany, G. I. N. (eds.): Proceedings of the first world congress of structural and multidisciplinary optimization,
May 28June 2, 1995, Goslar, Germany. Elsevier, Oxford 1995.
60 Papalambros, P. Y.; Chirehdast, M.: Integrated structural optimization systems. In: [7], pp. 501514.
61 Patnaik, S. N.; Guptill, D. J.; Berke, L.: Merits and limitations of optimality criteria method for structural optimization.
Internat. J. Numer. Meth. Engin. 38 (1995), 30873120.
62 Prager, W.; Rozvany, G. I. N.: Optimization of structural geometry, In: Bednarek, A. R.; Cesari, L. (eds.): Dynamical
systems. Academic Press, New York 1977, pp. 265293.
63 Ramm, E.; Bletzinger, K.-U.; Reitinger, R.; Maute, K.: The challenge of structural optimization. In: Topping, B. H. V.;
Papadrakakis, M. (eds.): Advances in structural optimization. Civil-Comp Press, Edinburgh 1994, pp. 2752.
64 Ringertz, U. T.: On finding the optimal distribution of material properties. Struct. Optimiz. 5 (1993), 265267.
65 Rosen, D. W.; Grosse, I. R.: A feature based shape optimization technique for the configuration and parametric design of flat
plates. Engin. with Comp. 8 (1992), 8191.
66 Rozvany, G. I. N.; Zhou, M.; Birker, T.: Generalized shape optimization without homogenization. Struct. Optimiz. 4 (1992),
250252.
67 Rozvany, G. I. N.; Bendse, M. P.; Kirsch, U.: Layout optimization of structures. Appl. Mech. Rev. 48 (1995), 41119.
68 Sanchez-Palencia, E.: Non-homogeneous media and vibration theory. (Lecture Notes in Physics, 127.) Springer, Berlin 1980.
69 Schittkowski, K.: NLPQL: A FORTRAN subroutine for solving constrained nonlinear programming problems. Anal. Oper.
Res. 5 (1985), 485500.
70 Schumacher, A.: Topologieoptimierung von Bauteilstrukturen unter Verwendung von Lochpositionierungskriterien. Ph.D. dissertation. University of Siegen, Siegen, Germany, 1996.
71 Schwarz, S.; Maute, K.; Ramm, E.: Topology and shape optimization including elastoplastic behavior. Proceedings of the 7-th
AIAA/USAF/NASA ISSMO Symposium on Multidisciplinary Analysis and Optimization, St. Louis, USA, September 1998,
pp. 19111921.
72 Sigmund, O.: Design of material structures using topology optimization. Dissertation. Report S 69. Danish Center for Applied
Mathematics and Mechanics, Technical University of Denmark, 1994.
73 Sigmund, O.: Materials with prescribed constitutive parameters: an inverse homogenization problem. Internat. J. Solids Struct.
31 (1994), 23132329.
74 Swan, C. C.; Arora, J. S.: Topology design of material layout in structured composites of high stiffness and strength. Struct.
Optimiz. 13 (1997), 4559.
75 Szabo, I.: Geschichte der mechanischen Prinzipien. Birkhauser, Basel 1979.
76 Tartar, L.: Problemes de contr^ole des coefficients dans les equations aux derivees partielles. In: Bensoussan, A.; Lions, J. L.
(eds.): Control theory, numerical methods, and computer system modelling. (Lecture notes in economics and mathematical systems, 107.) Springer, New York 1975, pp. 420426.
77 Taylor, J. E.: Truss topology design for elastic/softening materials. In: [7], pp. 451467.
78 Taylor, J. E.; Washabaugh, P. D.: Analysis and design of trussed structures made of elastic/stiffening materials. Struct. Optimiz. 8 (1994), 18.
79 Tenek, L. H.; Hagiwara, I.: Static and vibrational shape and topology optimization using homogenization and mathematical
programming. Comput. Meth. Appl. Mech. Engin. 109 (1993), 143154.
80 Vanderplaats, G. N.: Numerical optimization techniques for engineering design: with applications. McGraw-Hill, New York
1984.
81 Willis, J. R.: Bounds and self-consistent estimates for the overall properties of anisotropic composites. J. Mech. Phys. Solids 25
(1977), 185202.
82 Yang, R. J.; Chen, C. J.: Stress-based topology optimization. Struct. Optimiz. 12 (1996), 98105.
83 Yuge, K.; Kikuchi, N.: Optimization of a frame structure subjected to a plastic deformation. Struct. Optimiz. 10 (1995),
197208.
Received September 8, 1998, accepted January 21, 1999
Address: Dr.-Ing. Kurt Maute, Dipl.-Ing. Stefan Schwarz, Prof. Dr.-Ing. Ekkehard Ramm, Institute of Structural Mechanics,
University of Stuttgart, Pfaffenwaldring 7, D-70550 Stuttgart, Germany

S-ar putea să vă placă și