AOO-40025
AIAA-2000-4738
2.1 Topologies containing only isotropic solid The number of materials involved may also be in-
or empty finite elements (ISE topologies) dicated: a topology consisting of four possible ma-
terials and void will be denoted 4ISE topology. If
Our discussion is restricted to a given finite set of
all elements must be filled with material (i.e. there
elements affixed shape which may be empty or filled
are no empty elements) then the letter E is omitted
with one of several isotropic materials of given prop-
from the identifier for that topology. For example, a
erties .
topology having solid elements out of four materials
In the terminology of this paper, an element is called
but no empty elements will be called a 4IS topology
• Solid (S), if it is entirely filled with one material; (choice of 4 materials for Isotropic Solid elements).
The simplest subclass of problems in generalized
• Empty (E), if it contains no material; shape optimization is a USE (or "black-and-white")
topology involving the optimal distribution of a sin-
• Porous (P), if it contains one material and void gle material within the design domain. An example
(i.e. empty space); of this type of problems is a perforated plate in which
the plate thickness for any element is restricted to ei-
• Composite (C), if it contains more than one ma- ther zero or a given nonzero value.
terial but no void; and An elementary example of this class of problems is
given in Fig. la, in which we have 2x2 = 4 square
• Composite-Porous (CP), if it contains more than
finite elements in plane stress; two elements (1,2) are
one material and void. supported along their left edge and one element (4)
Using this terminology, ISE topologies contain Iso- is loaded uniformly along its bottom edge. The load
tropic Solid or Empty elements. is to be transmitted to the supports, so that
ones. Their elements are therefore solid, empty or Since most practical problems are associated with
composite. In ISECP topologies, composite-porous ISE and IS topologies at present, we will only discuss
elements are also admitted. in detail methods for these problems. As before, our
investigation is restricted to topologies with a given
2.3.1 Applications of ISEP topologies finite number of elements.
ISEP (ISEC and ISECP) topologies have two very
3.1 The SIMP method
important applications.
In this method, we are using Solid Isotropic Micro-
• They may indicate the exact optimal topology for structures 1 with Penalization for intermediate den-
a given problem, which can be compared with sities.
exact analytical solutions for verification pur- The SIMP method can be easily understood if we
poses. Exact solutions can be used as the ab- consider the example of a perforated plate in plane
solute limit of material economy for a given de- stress, in which the plate thickness is either zero or a
sign problem, providing a basis for assessing the given value (to). In order to explore all possible solu-
relative degree of material economy of practical tions for a large number of elements (e.g. 40000), we
designs. would have to carry out a prohibitively large number
(24oooo ^ 10 i204i) of analyses, and therefore we must
• We may wish to manufacture fibre-reinforced or
resort to iterative methods.
other densely structured composites on the basis
We can, for example, assume that the plate thickness
of "exact" optimal solutions.
t may vary continuously between zero and to (after
The basic features of the above topologies is summa- Rossow and Taylor [15]). The mechanical properties
rized in Table 1. of the plate (e.g. stiffness for in-plane forces, per-
The ISE and ISEP topologies were termed, respec- missible values of in-plane forces etc.) are linearly
tively, SE and SEP topologies in earlier studies (e.g. proportional to its thickness. We can easily mini-
[12, 13]). mize the weight of the above plate by using either
Naturally, we could restrict the material properties an optimality criteria (OC) or a mathematical pro-
and microstructure parameters in ASE and ISEP J
It should be noted that an alternative interpretation for
topologies, respectively, to given discrete values and "M" in SIMP is "Materials" , see e.g. [14]. This author
then we would again finish up with discrete value prefers the term "microstructures" because he uses the word
optimization. "material" for the base material in microstructures, which is
isotropic even in ISEP and ISEC topologies. The limiting (de-
generate) case of a microstructure is indeed a solid element (or
an empty element). As indicated in Section 2.1, multimaterial
3. Basic methods of generalized shape systems are called "composites" (and not "materials") in this
optimization with ISE and IS topologies article.
(c)2000 American Institute of Aeronautics & Astronautics or Published with Permission of Author(s) and/or Author(s)' Sponsoring Organization.
gramming (MP) method. In fact, for a compliance straints and objective function(al). Such microstruc-
constraint, this problem is convex and therefore the tures have been studied completely at present only
(only) optimum is easily and quickly calculated. for "compliance" design, in which the total amount
The catch is that the resulting solution will contain of external work is either minimized or constrained.
all sorts of thicknesses and in ISE topologies we want
only thicknesses of zero or to. Our unwanted result ——— solid isotropic microstructure
can be largely improved if we penalize intermediate ——— rank-2 laminate
thicknesses (with 0 < t < to). —•—- solid microstructure with penalty (SIMP)
This procedure is shown graphically in Fig. 4, in .——— microstructure with sqare holes
which the relation between the normalized plate stiff-
ness s (in plane stress) and the specific cost p (here: Specific Cost
plate thickness) is indicated. For a plate of varying 1.0
thickness without penalty, this relation is a straight
line. We can penalize the intermediate thicknesses
by using the relation (after Bends0e [8])
s= (1)
0.5 -/
where p > I. In our unpenalized procedure (after
[15]), we had p — 1. The relation in (!) is shown
graphically for p = 5 in Fig. 4.
The above penalization will effectively suppress in-
termediate thickness values but the problem becomes
nonconvex even for compliance design and therefore Stiffness
1.0
globality of the optimum cannot be guaranteed. The
results are usually improved if we start at the begin- Figure 4: Stiffness (s) specific cost (g) relation for
ning of the iteration with p — 1 and then increase p various types of microstructures
gradually to a higher value (say p = 5).
For a 2D problem, the optimal (rank2 layered) mi-
3.2 The OMP method
crostructure of an element has three free parame-
In this method Optimal Microstructures with ters (two layer densities and one orientation) and for
Penalization for intermediate densities are used. 3D problems it has five free parameters (three layer
This means that first the solution is optimized using densities and two orientations). For so-called non-
for each finite element an optimal microstructure, de- selfadjoint problems (e.g. Lurie [16]), the rank2 lay-
rived rigorously for the particular type of design con- ered microstructures are not orthogonal and hence
(c)2000 American Institute of Aeronautics & Astronautics or Published with Permission of Author(s) and/or Author(s)' Sponsoring Organization.
the number of free parameters must be increased to In the NOM method, the number of free parame-
three (2D) and six (3D). ters per element may be somewhat lower than in the
However, optimal microstructures do not provide OMP method (for 2D problems with square hole, for
sufficient penalization for ISE (black-and-white) example, two instead of three).
topologies (see Fig. 4). For this reason, some ad- The three main methods used for ISE or IS topologies
ditional penalization is usually introduced. are summarized in Table 2.
A%
20
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
square holes). In this case, the specific material cc
Figure 7: Specific cost of perforated plates taking
2 cost
pm — a(l — a ) —>• a = v/1 — gm/a (2) °^ "sawulg out"mto consideration (see gt)
where a is the material cost per unit plate area and .„„ _ , , , . . , . , , . , . .,
. ,, •„ • , , ,, r ,1 i i TM 20, 21 ) which does not rely on a physical model to
a is the specific side length ol the square holes. 1 he ..... v
justify it.
fabrication cost becomes
Penalization in between given discrete design values (a) Computational efficiency in terms of storage ca-
is a standard method in nonlinear optimization (e.g. pacity and CPU time required, since only one
(c)2000 American Institute of Aeronautics & Astronautics or Published with Permission of Author(s) and/or Author(s)' Sponsoring Organization.
free variable is used per element (OMP requires The OMP method has a potential marginal advan-
up to 6, see Section 3.2). tage, if we wish to calculate both optimal ISE and
ISEP topologies for a given design problem, because
(b) Robustness in the sense that SIMP can be both can be obtained in an extended single operation
readily used for any combination of design con- (by starting with an unpenalized optimal microstruc-
straints. OMP is presently restricted to compli-
ture). This, however, happens rarely in practice.
ance or equivalent designs, because the optimal
microstructure is not known for more compli- 5.3 Disadvantages of the NOM method for
cated design conditions. ISE/IS topologies
(c) Penalization can be adjusted freely, and hence The NOM method may potentially have a smaller
the computationally optimal penalization can be number of variables than the OMP method (e.g.
used, which is not the case with NOM, for ex- square holes), which is a slight advantage. However,
ample. it has the following disadvantages.
(d) Conceptual simplicity, since the algorithm does (a) NOM involves more variables per element than
not require derivations involving higher mathe- SIMP (but possibly less than OMP).
matics (derivations of the optimal microstruc-
ture for OMP relies on advanced mathematical (b) Penalization is fixed and often inadequate, re-
methods). sulting in "grey" regions and/or a nonoptimal
ISE/IS topology.
(e) Since the p-value in (1) is increased progres-
sively, we can start SIMP with a solution for (c) As OMP, NOM is also inherently nonconvex.
p = 1 for which some problems (e.g. compli-
(d) NOM also requires homogenization.
ance) are convex and the solution a global op-
timum. The subsequent gradual incrementation The relative advantages and disadvantages of various
of the p-value is not likely to move the solution methods for ISE and IS topologies are summarized
too far from the global optimum, but this is only in Tables 3 and 4.
an "experimental" finding at this stage.
(f) SIMP does not require homogenization of the 6 Theoretical objections to SIMP and their
microstructure. irrelevance to ISE/IS topologies
The SIMP method has no known disadvantage for
ISE/IS topologies at present. The universal adoption of the computationally supe-
rior SIMP method was delayed by almost a decade
5.2 Disadvantages of the OMP method for owing to certain theoretical objections to this tech-
ISE/IS topologies nique.
The OMP method has no known major advantages 6.1 Lack of physical interpretation of SIMP
for purely ISE/IS topologies but has the following
obvious disadvantages. In first introducing the SIMP method, Bends0e [8]
commented on the efficiency of this method but
(a) It is computationally complicated and relatively noted that "it is impossible to give a physical mean-
inefficient, requiring much more free variables ing for intermediate values" (of density). For this
per element than SIMP. reason, Bends0e [8] refers to an "artificial material"
(b) The optimal microstructures are developed fully with the appropriate density-stiffness relation. At
at present for compliance design only, which the time, Bends0e's comments were absolutely cor-
makes OMP a highly nonrobust method. rect.
As noted already by Rozvany and Zhou [9] at a meet-
(c) Derivation of the exact optimal microstructure ing in 1990, a certain physical interpretation of the
for any new design condition requires a highly SIMP relation in (1) can be achieved by the inclu-
advanced mathematical treatment. sion of manufacturing costs (e.g. the length of the
perimeter of the holes) in the cost function.
(d) The OMP problem is intrinsically nonconvex,
As mentioned in Section 4.1, the above objection was
even for compliance design.
completely removed by Bends0e and Sigmund [14],
(e) OMP requires homogenization of the microstruc- who constructed composites realizing the relation in
ture which is an extra operation in comparison (1) within limits on the p-value. Moreover, these
to SIMP. authors correctly point out:
(c)2000 American Institute of Aeronautics & Astronautics or Published with Permission of Author(s) and/or Author(s)' Sponsoring Organization.
Table 3: Relative advantages of the SIMP, OMP and NOM methods in optimizing ISE and IS topologies
SIMP OMP NOM
(a) computational efficiency additional information about potentially smaller no. of
(one variable per element) the optimal ISEP topology variables per element than
(b) robustness, suitable for (almost) OMP
any design condition
(c) penalization can be adjusted freely
(d) conceptual simplicity, no
higher mathematics required
(e) convexity can be preserved
for the early iterations with p = I
(f) no homogenization necessary
Table 4: Relative disadvantages of the SIMP, OMP and NQM methods in optimizing ISE and IS topologies
SIMP OMP NOM
no known (a) greater computational effort (a) more variables per
disadvantage than SIMP element than SIMP
(b) highly nonrobust (restricted (b) penalization fixed and often insufficient
presently to compliance design) for reaching the correct ISE/IS topology
(c) requires advanced mathematics (c) intrinsically nonconvex
for deriving optimal microstructures (d) requires homogenization
(d) intrinsically nonconvex
(e) requires homogenization
"if a numerical method leads to black-and- Sigmund et al. [23]. In this beam layout problem, a
white designs, one can, in essence, ignore rhombic grillage has simple line supports along two
the physical relevance of 'grey' and in many edges (double lines) and the other two edges (single
situations a better computational scheme lines) are unsupported (Fig. 8), with a point load
can be obtained if one allows the violation P at the unsupported corner. It was shown rig-
of bounds on properties of composites". orously by Prager and Rozvany ([22]) that for the
above problem the optimal layout consists of an infi-
6.2 Mesh-dependence of SIMP results nite number of beams, having the type of beam lay-
out shown in Fig. 8, in which continuous and broken
As another possible disadvantage of the SIMP lines indicate beams under positive and negative mo-
method (then termed "direct approach"), Bends0e ments, respectively. For a finite number (n) of long
[8] correctly noted that the "scheme is very depen- beams, the structural weight has been shown to be
dent on the mesh". [22]
Mesh-dependence can be mostly avoided by con-
straining the length of the internal boundaries or
"perimeter", as demonstrated in an outstanding pa-
per by Haber, Bends0e and Jog [10]. For each con- where k is a constant. The optimal layout and op-
strained perimeter value, the ISE topology remains timal weight for a given n-value is strongly mesh-
stable below a critical mesh size. dependent, with the absolute optimal weight (for
Whilst perimeter control is absolutely necessary if we 77 —> oo) of
want to restrict a practical design to a simple topol- = 7-kPa2 (6)
ogy, mesh-dependence actually becomes beneficial if
we want to demonstrate that ISE topologies tend to In verifying this optimal layout by FE-based meth-
a known exact analytical solution. ods (e.g. [23]), we find that the strongly mesh-
To demonstrate this, we refer (once more) to an op- dependent structural weight clearly converges to the
timal grillage solution by Prager and Rozvany [22], value in (6). Similarly, it is necessary to allow mesh-
which was also confirmed using a FE formulation by dependence if we wish to confirm by ISE topologies
(c)2000 American Institute of Aeronautics & Astronautics or Published with Permission of Author(s) and/or Author(s)' Sponsoring Organization.
(SIMP method) any analytical solutions consisting solution clearly exists and the problem is intrinsi-
of an infinite number of internal boundaries. This cally well-posed, even when N is extremely large (but
procedure would not be possible, if we put an effec- finite). The foregoing conclusion should, in itself,
tive constraint on the perimeter (i.e. total length of disperse any doubt about the validity of the ISE/IS
internal boundaries). topology formulation for any practical problem: with
the rapid improvement of computational capabilities,
in the not so distant future we will be able to em-
ploy the SIMP procedure for ground structures with
millions of elements, and the problem will still be
well-posed.
(b) It was shown that the SIMP methodology has [11] Rodrigues H, Soto C and Taylor JE "A design
several advantages over other techniques (OMP, model to predict optimal two-material composite
NOM). structures", Struct. Optim., 1999, 17, 186-198.
(c) Theoretical objections to the SIMP approach
have been found irrelevant and/or unfounded. [12] Rozvany GIN, Zhou M and Birker T "General-
(d) In the lecture/full paper, computational experi- ized shape optimization without homogenization",
mental evidence and new applications of SIMP will
Struct. Optim., 1992, 4, 250-254.
be outlined. [13] Rozvany GIN, Kirsch U and Bends0e MP "Lay-
out optimization of structures", Appl. Mech. Rev.,
1995, 48, 41-119.
References
[14] Bends0e MP and Sigmund O. "Material inter-
[1] Michell AGM "The limits of economy of material polation schemes in topology optimization". Arch.
in frame structures", Phil. Mag. 8, 1904, 589-597. Appl. Mech., 1999, 69, 635-654.
[2] Rozvany GIN "Grillages of maximum strength [15] Rossow MP and Taylor JE "A finite element
and maximum stiffness", Int. J. Mech. Sci., 1972, method for the optimal design of variable thickness
14,651-666. sheets". AIAA J., 1963, 11, 1566-1569.
[3] Prager W and Rozvany GIN "Optimization of [16] Lurie K "An optimal plate". In: Olhoff N and
structural geometry". In: Bednarek AR and Ce- Rozvany GIN (Eds.) Proceedings of WCSMO-1
sari L (Eds.) Dynamical systems, New York, Aca- (Goslar 1995), pp. 169-176, 1995, Pergamon, Ox-
demic Press, 1977, pp. 265-293. ford.
[4] Rozvany GIN "Optimal layout theory: analyti- [17] Hashin Z and Shtrikrnan S "A variational ap-
cal solutions for elastic structures with several de- proach to the theory of the elastic behaviour of
flection constraints and load conditions", Struct. multipurpose materials". J. Mech. Phys. Solids,
Optim., 1992, 4, 247-249 1963, 11, 127-140.
[18] Sigmund O "Materials with prescribed constitu-
[5] Rozvany GIN and Birker T "On singular topolo-
tive parameters: an inverse homogenization prob-
gies in exact layout optimization", Struct. Optim.,
lem". Int. J. Solids Struct., 1994, 31, 2313-2329.
1994, 8, 228-235. .
[19] Rozvany, GIN, Structural Design via Optimality
[6] Cheng K-T and Olhoff N. "An investigation con- Criteria, Kluwer, Dordtecht, 1989
cerning optimal design of solid elastic plates, Int.
J. Solids Struct., 1981, 17, 305-323. [20] Shin DK, Giirdal Z and Griffin OH jr "A penalty
approach for nonlinear optimization with discrete
[7] Rozvany GIN, Olhoff N, Bends0e MP, Ong TG, design variables", Eng. Opt., 1990, 16, 29-42.
Sandier R and Szeto WT "Least-weight design of
perforated elastic plates I, II", DCAMM Report [21] Bauer J "A survey of methods for discrete op-
306, 1985; Int. J. Solids Struct., 1987, 23, 521- timum structural design", Comp. Assist. Mech.
536, 537-550. Eng. Sci., 1994, 1, 27-38.
[8] Bends0e MP "Optimal shape design as a mate- [22] Prager W and Rozvany GIN "Optimal layout of
rial distribution problem", Struct. Optim., 1989, grillages", J. Struct. Mech., 1977, 5, 1-18.
I, 193-202. [23] Sigmund 0, Zhou M and Rozvany GIN "Lay-
out optimization of large FE systems by optimality
[9] Rozvany GIN and Zhou M "Applications of
criteria methods: applications to beam systems",
COC method in layout optimization". In: Es-
In: Haug EJ (Ed.) Proc. NATO ASI: Concurrent
chenauer H, Mattheck C and Olhoff N (Eds.) Proc.
Engineering Tools and Technologies for Mechan-
Conf. "Eng. Opt. in Design Processes" (Karlsruhe
ical System Design (Iowa 1992), 1992, Springer,
1990), Berlin, Springer-Verlag, 1991, pp. 59-70.
Berlin.
[10] Haber RB, Bends0e MP and Jog CS "A new ap-
proach to variable-topology shape design using a
constraint on the perimeter, Struct. Optim., 1996,
II, 1-12.