Beruflich Dokumente
Kultur Dokumente
_
(1.1)
The problem in (1.1) is solved iteratively by gradually changing the design
variables, a, according to the gradients computed in the sensitivity analysis until
a lower value of the objective, f, cannot be found. This methodology is referred
to as gradient based design optimization and is generic to the three commonly
used classes of methods for doing structural design optimization: topology, shape
and size. To distinguish between the three let us take the example of the 2D
structure in Fig. 1.2. Topology optimization (left) can be used to gure out where
to distribute a limited amount of material and as such, it can introduce internal
holes in the structure. This usually provides a coarse outline of the structure so
shape optimization (middle) can be used to rene the boundaries. Finally, size
optimization (right) can be used to nd the optimal thickness distribution over
the structure, indicated in the edge view in Fig. 1.2. The applicability of the
three methods is far greater than indicated here, but the example illustrates the
fundamental dierences.
In the following the present study will be mapped out in more detail with reference
to related studies.
1
In the present study we employ the Nested ANalysis and Design (NAND) approach in which
the equilibrium equations will be assumed to be satised prior to solving the optimization problem
itself.
4 1.2. Background of work
Figure 1.2: Side and edge view of some 2D structure subjected to topology optimization
(left), shape optimization (middle) and size optimization (right).
1.2 Background of work
Over the last two decades the strive for lighter and stronger structures has resulted
in an increasing use of composite materials. In particular, ber reinforced polymers
(FRP) have gained an ever increasing popularity due to their very high strength
to weight ratio. In structural applications ber reinforced polymers are usually
stacked in a number of layers, each consisting of strong bers bonded together by a
resin, to form a laminate. The bers may be uniformly oriented in typically one or
two directions or they may be oriented in no particular ordered fashion. The best
use of the material is achieved when ordering the bers in specic directions to
obtain high stiness in the loading directions and lower stiness in other directions.
Exploiting this directionality of the material is at the core of ecient design with
laminated composites. Furthermore, to obtain an optimal design the engineer must
choose where to put material and which materials to use both in general and in
individual layers. However, proper choice of material layout, materials, stacking
sequence and ber orientation is a far from simple task since laminates can often
consist of as many as 500 or more dierent layers.
This brings forward the need for ecient and reliable numerical design tools
in particular when dealing with large scale structures involving complicated
geometries, multiple layers, multiple materials and multiple load cases. Eorts for
developing such tools are already well under way and this work naturally draws
on results from previous studies in the elds of both nite element analysis and
gradient based optimization. With the vast amount of literature available on these
topics it is well beyond the scope of this brief review to give an exhaustive account
of the works preceding this. In the following particular emphasis will therefore be
placed on topics directly related to the present work.
1.2.1 Finite element analysis of laminated composites
The use of composite materials, especially in the automotive and aerospace
industries, has naturally fueled the eort for developing nite element methods
suitable for such materials. Laminated composite shell structures are used
extensively in these industries and have therefore received an increasing amount of
attention over the last decade, see Noor et al. (1996), MacNeal (1998), Yang et al.
(2000) and Mackerle (2002). This progress has been supported by the development
of robust and reliable shell nite elements, which do not suer from deciencies
Chapter 1. Introduction 5
such as locking and hour-glass modes usually associated with shell nite elements.
The nite element method has, of course, also benetted tremendously from the
exponential growth in computer power, which has made possible the analysis of
realistic and industry relevant models (Bathe et al., 1997; Noor, 1999).
Several classes of nite elements have been developed and the choice depends
on the level of detail required for design purposes. The simplest and most
popular class combines an Equivalent Single Layer (ESL) laminate description
with nite elements based on First order Shear Deformation Theory (FSDT)
to obtain highly ecient and reliable elements for global response analysis. In
this group the most popular choice are linear elements, rather than higher-order
elements, as it compensates for the increased computational cost of computing the
laminate stiness. To avoid locking problems when using lower-order elements the
Assumed Natural Strain (ANS) technique developed by e.g. Hughes and Tezduyar
(1981), MacNeal (1982), Dvorkin and Bathe (1984) and Bletzinger et al. (2000) is
employed, resulting in very robust elements. This is also the strategy chosen in this
work (as described in Chapter 3) and other examples from the literature include
Barut et al. (2000), Alfano et al. (2001) and Wagner and Gruttmann (2002). The
ANS elements have also appeared in major commercial codes such as ANSYS,
ADINA and MSC.NASTRAN/MARC while more advanced methods such as layer-
wise elements e.g. Reddy (1993), Brank and Carrera (2000) and To and Liu (2001)
or solid shell elements e.g. Klinkel et al. (1999) and Sze et al. (2002), accounting
for varying degrees of local behavior, are still reserved for research codes. This is
mainly due to their higher complexity and computational cost. For an extensive
review of layer-wise methods in nite element applications see Ochoa and Reddy
(1992), Carrera (2003) and Reddy (2004).
1.2.2 Topology optimization
Topology optimization was introduced some 30 years ago and was an important
contribution to structural optimization in that it provided engineers with the
ability to optimize not only the shape of existing topologies but also the topology
itself. Since then the eld has been the subject of extensive research and is
probably the most active eld in optimization at present. The basis for topology
optimization as it is today was laid out by Bendse and Kikuchi (1988) who used
the homogenization technique and Bendse (1989) who introduced the SIMP
2
methodology, which was derived independently and extensively implemented
by Zhou and Rozvany (1991) and Rozvany et al. (1992). Ever since, numerous
extensions have been made to the method both in terms of capabilities and its
range of applicability to industrial problems. For extensive reviews of the method
the reader is referred to Eschenauer and Olho (2001), Bendse and Sigmund
(2003) or Mackerle (2003).
The extension of topology optimization to shells (e.g. Maute and Ramm, 1997) has
2
Solid Isotropic Material with Penalization
6 1.2. Background of work
not received the same amount of attention as other elds in topology optimization.
This is partly due to the fact that the essence of classical topology optimization
is the ability to introduce holes in a structure, thereby reducing material usage and
in turn reducing weight. However, the ability to introduce through-the-thickness
holes in shell structures has little practical relevance since holes deteriorate the
membrane load carrying ability of the shell. Furthermore, for the majority of
engineering shell structures (fuselages, wings, ship hulls, turbine blades, pressure
vessels etc.) it is not viable to introduce holes since the boundary is usually
prescribed by other factors (e.g. aerodynamical considerations). Thus, topology
optimization of shell structures only has real engineering applications with the use
of multilayered shell nite elements. In that context topology optimization can
be used to add or remove material in specic layers rather than through-the-
thickness. This opens the door to stiener and core layout design, which has been
treated for plates by e.g. Diaz et al. (1995) and Krog and Olho (1999) and for
shells by e.g. Lee et al. (2000), Belblidia et al. (2001) and Belblidia and Bulman
(2002).
The above mentioned works are all concerned with linear problems. However,
nonlinearities play an important role in the failure of large composite structures,
such as wind turbine blades, and should therefore be considered as well. Some
of the earliest work involving geometrical nonlinearities in stiness design of
continuum structures was that of Jog (1996) and Yuge et al. (1999). Recent
developments of importance to the present work include Buhl et al. (2000),
Gea and Luo (2001) and Bruns et al. (2002) who used topology optimization on
2D structures with geometrical nonlinearities. To the best of our knowledge no
work prior to this addresses the inuence of nonlinearities on the stiening topology
of laminated composite shell structures as treated in Chapter 4.
One of the exiting new developments in topology optimization is the extension
beyond two phases (solid and void) to include multiple phases. This work has
been pioneered by Sigmund and co-workers, e.g. Sigmund and Torquato (1997) or
Sigmund (2001), who used multiphase topology optimization for material design
and design of 2D continua as well as compliant mechanisms. The same idea was
used recently by Wang and Wang (2004) in a level-set framework for solving similar
problems. These ideas lay out the ground for Chapter 5 and will be discussed in
detail there.
1.2.3 Optimization with orthotropic materials
Topology optimization can solve the material distribution problem but the ber
orientation has at present not been solved by any of the optimization branches
mentioned above. For this another branch of optimization has emerged, dedicated
to nding the optimal orientation layout for orthotropic materials (such as ber
reinforced polymers). This work has been established in large part by Pedersen
(Pedersen, 1989, 1991) and the key aspects of the method are summarized in
Pedersen (2004), which also provides a number of examples. Other authors
Chapter 1. Introduction 7
have contributed as well, e.g. Luo and Gea (1998) who used a very similar
approach to that of Pedersen for plates, and Thomsen and Olho (1990) who used
optimality criteria combined with mathematical programming to solve for ber
orientation and density in discs and plates. Other approaches have been taken by
e.g. Miki and Sugiyama (1993) and Foldager et al. (1998) who used lamination
parameters to overcome the inherent diculties with local minima in this type
of problems. The prevailing method seems to be that of Bruyneel and Fleury
(2002) and Moita et al. (2000) who use customized mathematical programming
techniques and such methods have been implemented in the commercial software
packages BOSS QUATTRO from SAMTECH and OPTISTRUCT from ALTAIR.
In the present work the methods above are not extended further but used as
reference in Chapter 5.
1.3 Objectives of work
The general objective of this work is to develop nite element based optimization
techniques for laminated composite shell structures. Furthermore, these methods
should be applicable to practical problems of engineering interest in Danish
industry. The key aspects chosen for investigation in this work are:
Robust and ecient analysis methods for laminated composites
Solution of the material distribution problem
Solution of the ber angle optimization problem
Investigation of large displacement eects on optimal design
The last point springs from our collaboration with the Danish wind turbine
industry who wants to develop still longer and lighter wind turbine blades, which
are subject to very large displacements under running conditions. In the context of
such large laminated composite structures it has been chosen to focus on stiness
maximization (compliance minimization) as the optimization objective.
1.3.1 The MUltidisciplinary Synthesis Tool MUST
The MUltidisciplinary Synthesis Tool (MUST) is a nite element based analysis
and optimization code developed by the Computer Aided Engineering Design
Group at the Institute of Mechanical Engineering, Aalborg University (MUST,
2004). The system has been developed in Fortran 90/95 by Associate Professor
Erik Lund and co-workers over the last six years and the development is set to
continue in the future. Some of the major contributors are Henrik Mller (Mller,
2002) and Lars Jakobsen (Jakobsen, 2002) who have both been associated with the
research project Interdisciplinary Analysis and Design Optimization of Systems
with Fluid-Structure Interaction that spawned MUST.
8 1.3. Objectives of work
MUST
Inputmodule
(ANSYS, COSMOS, ODESSY)
Finiteelementlibrary
(Beam, Solid/fluid 2D/3D, shell)
Sensitivityanalysismodules
(Finitedifference, analytical)
Optimizer library
(MMA, simplex)
Optimizationmodules
(Shape, size, topology, fiber)
Databasemodule
(Resultgeneration)
Postprocessor
(FEPlot)
Solver library
(Linear, nonlinear)
Analysismodules
(Solid, fluid, thermal, FSI)
Figure 1.3: Major components of the MUltidisciplinary Synthesis Tool (MUST). The
gray components are those changed in the present work.
The major reason for choosing MUST as platform of implementation is that the
need for developing a nite element framework from scratch is circumvented.
Furthermore, the use of MUST allows the implemented methods to be used by
colleagues and students an opportunity that has already been exploited in
several graduate studies. This supports the philosophy behind MUST which is
to support both research and education by providing understandable program
code and allowing easy implementation of new features.
MUST is a stand-alone application but relies on external software for meshing
and general preprocessing. It reads a modied input le from ANSYS, COSMOS
or ODESSY
3
, solves the problem, and generates a database for visualization in
FEPlot, which is an in-house postprocessor continuously developed by Erik Lund
and Henrik Mller. The major features of the MUST system are depicted in
Fig. 1.3 where the parts aected by this work are marked by gray.
1.3.2 Outline of thesis
The thesis is organized in four main chapters.
Chapter 2 introduces the basic concepts of structural analysis and optimization.
This includes brief treatments of governing equations, equation solving, sensitivity
analysis and the optimizer.
Chapter 3 is dedicated to nite element analysis of laminated composite shells
3
The Optimum DESign SYstem. In many ways a predecessor to MUST, today maintained
and used largely by Associate Professors Erik Lund and John Rasmussen.
Chapter 1. Introduction 9
and describes the implemented shell elements in MUST in terms of assumptions,
implementation and performance.
Chapter 4 treats the inuence of large displacements (geometrical nonlinearities)
on multilayered shell topology optimization problems, and a number of benchmark
examples demonstrate the dierence in optimal topologies between linear and
nonlinear solutions.
Chapter 5 is devoted to material layout and ber angle optimization and introduces
the concept of discrete material optimization (DMO), which is tested for various
2D and 3D examples.
Chapter 6 summarizes the conclusions drawn from the present work.
10 1.3. Objectives of work
2
Analysis and optimization
T
his chapter lays out the analysis and optimization tools necessary
for solving the generic optimization problem stated in (1.1). Solving the
optimization problem is a process involving several steps as illustrated in Fig. 2.1.
Starting from an initial design (with dened design variables) a design analysis is
performed (solution of the physical problem) and subsequently, analysis and design
improvement is performed consecutively to gradually obtain the nal, optimal
design.
The optimization process is thus a chain of events, each leading (hopefully) towards
the optimal design. As usual, the chain is only as strong as its weakest link and so,
a considerable amount of time has in this work been invested in implementation
of reliable nite element technology. The methodology for doing so is introduced
in Section 2.1 and later in Chapter 3, the particular implemented elements will be
described in greater detail. Another important aspect of solving the optimization
problem is the design improvement step, Fig. 2.1. In this work a shortcut has
Designanalysis
(finiteelement)
Sensitivity
analysis
Improvedesign
(optimizer)
Newdesign
variables
Optimize?
Preprocessing
(Initialdesign)
Yes
Postprocessing
(Finaldesign)
No
Figure 2.1: Flow chart of the solution process for the generic optimization problem in
(1.1). The gray boxes indicate topics treated in this chapter.
12 2.1. Analyzing the design
been taken in this step since an o the shelf optimizer has been implemented in
MUST, only requiring an appropriate interface to be programmed. A thorough
treatment of the theory behind optimizers will be left the established literature
but the fundamentals will be presented in Section 2.2 with special emphasis on
the applicability of such methods in this work.
2.1 Analyzing the design
The problem statement so far has been generic for any class of problems but
in the following, focus will be on structural problems. The governing equations
for the physical problem are Newtons laws of motion and for the nite element
method these are recast as energy conservation equations. This is standard and
for the sake of brevity a derivation of the governing equations will not be given
here, as it may be found in numerous textbooks such as Cook et al. (1989),
Zienkiewicz and Taylor (1991), Bathe (1996), Bonet and Wood (1997), Hughes
(2000) or Belytschko et al. (2000).
In this work both linear and nonlinear problems are considered and as the
formulation of the latter encompasses the former, emphasis will be on deriving
the nonlinear expressions. The starting point is the governing equations for the
static structural problem, which are stated as an axiom:
_
V
s
T
dV
. .
Internal work
_
V
p
b
u dV +
_
A
p
s
u dA
. .
External work
= 0 (2.1)
Here u u
i
is a displacement increment and s s
ij
is the second Piola-
Kircho stress, which is work conjugate with the Green-Lagrange strain increment,
ij
. The external forces are divided into body forces, p
b
, and surface forces,
p
s
. All quantities in (2.1) are tensors but have been expressed using matrix
notation, which is convenient for deriving the element matrices. In that context
the strain and stress vectors are dened as =
11
,
22
,
33
, 2
12
, 2
23
, 2
13
T
and
s = s
11
, s
22
, s
33
, s
12
, s
23
, s
13
T
, respectively.
In nite element analysis the governing equations (2.1) are recast in vector form
by introducing the strain-displacement matrix, B, which is dened from = Bu
where u indicates nodal values. Alternatively, the strain-displacement matrix can
be expressed from the variation of the strain =
u
u as:
B =
u
(2.2)
The form in (2.2) is convenient for deriving the strain-displacement matrix as
will be shown in Section 3.4. Introducing B and rewriting the strain variation in
(2.1) to a displacement variation, an internal element nodal force vector, r, can be
Chapter 2. Analysis and optimization 13
derived from the internal work term as:
r =
_
V
B
T
s dV (2.3)
This must be balanced by the external element nodal force vector, p, which can
be derived in the standard way by computing equivalent nodal forces from the
distributed loads p
b
and p
s
in the external work term. Thus, the residual vector
1
,
R, is dened over all elements, N
e
, as:
R(u) =
N
e
k=1
_
r
k
p
k
_
(2.4)
For the system to be in static equilibrium the internal and external forces must
balance each other, i.e.
R(u) = 0 (2.5)
which represents the governing system of equations for linear and nonlinear static
problems. Thus, the generic statement physical laws in (1.1) can be replaced by
R = 0 to form the basic form of the optimization problem solved in this work:
Objective : min
a
f(a)
Subject to : g(a) G
a
min
a a
max
R(u, a) = 0
_
_
(2.6)
in which u is a xed point indicating that R(u, a) = 0 has been solved prior to
solving the optimization problem, which constitutes a Nested ANalysis and Design
(NAND) approach.
2.1.1 Solving the equations
Obtaining the stationary solution to the physical problem (2.5) requires an
iterative methodology since the internal force vector, r, is a function of the
displacements due to the nonlinear terms in the Green-Lagrange strain tensor
(3.6). We employ Newton-type solvers, which use the following linearization of
the governing equations, (2.4):
R(u + u) R(u) +
R(u)
u
u = 0 (2.7)
To solve the system in (2.7) the tangent stiness matrix, K
T
, is dened as:
K
T
=
R(u)
u
(2.8)
1
This notation is somewhat inconsistent since capital letters usually denote matrices but this
is standard in the literature
14 2.2. Improving the design
and used to nd the displacement increment u by solving the linear system:
K
T
u = R(u) (2.9)
In MUST (2.7) is solved continuously until the ratio |R|/|p| is below some
specied value, typically around 10
8
10
6
. At present, MUST encompasses a
number of nonlinear solvers due to Mller (2002), including a full Newton-Raphson
solver, a modied Newton-Raphson solver and a quasi-Newton BFGS solver all
of which can be used with or without line search. Line search gives better global
convergence properties than full N-R and modied N-R directly and has been used
with full N-R when solving for a single equilibrium point. For tracking equilibrium
paths an arc-length solver has been used with modied N-R for the sub-iterations.
Derivation of the tangent stiness matrix is an essential step in implementing
nite elements and to proceed we assume design independent loads and use the
denition of the internal nodal force vector (2.3) to write (2.8) as:
K
T
=
N
e
k=1
_
V
(B
T
k
s
k
)
u
k
dV
=
N
e
k=1
_
V
_
B
T
k
u
k
s
k
+B
T
k
s
k
u
k
_
dV
(2.10)
If the problem under consideration is geometrically linear the need for (2.4)(2.10)
can be circumvented and the static problem simplied. Using that s = CBu a
linear stiness matrix can be dened from the internal work term in (2.1) as:
K =
_
V
B
T
CB dV (2.11)
whereby the static equilibrium for the linear problem becomes:
Ku = p (2.12)
which is an algebraic linear system of equations that can be solved directly for
u using any linear solver. In MUST we employ the direct sparse solver from the
Compaq eXtented Math Library (CXML), which has proven extremely ecient
and far superior the direct prole solver used previously. Furthermore, a number
of iterative linear solvers are available but these have not been employed in the
present work.
The equations (2.3) and (2.10) above form the basis for proceeding with the nite
element formulation and implementation as described in Chapter 3.
2.2 Improving the design
Having successfully analyzed the design the next step towards the optimal design
is the design improvement phase as depicted in Fig. 2.1. To improve the design it
Chapter 2. Analysis and optimization 15
must rst be established how the performance of the design changes with changes
in the design variables. This is achieved through a design sensitivity analysis
(DSA) in which the gradients of the characteristic properties of the optimization
problem (2.6) are found.
2.2.1 Design sensitivity analysis
Sensitivity analysis is an important part of the gradient based optimization method
and particularly, ecient and accurate computation of gradients is essential for the
success of the method. In this work we strictly use analytical sensitivities, which
can be formulated explicitly and implemented in a general way.
In this work the objective of the optimization is to maximize stiness, as mentioned
in Section 1.3, which can be recast to a minimization problem by introducing the
compliance, C(a):
C(a) = p
T
u(a) (2.13)
which now becomes the objective, i.e. f(a) = C(a). However, this does not
in any way imply that the methods developed later are restricted to compliance
minimization. We will assume that the external load is independent of the design
variables and write the compliance sensitivity for the ith design variable as:
dC(a)
da
i
= p
T
du(a)
da
i
(2.14)
This gradient of the objective will indicate whether a change in a particular design
variable will increase or decrease the performance of the design. In the same way
the constraint gradients, dg(a)/da, indicate how the design variables should be
changed to keep the design within the constraint bounds. In the following we
assume that a design sensitivity analysis has already been performed. Details
concerning computation of the gradients for the specic optimization problems
considered in this work will be treated more extensively in Chapters 4 and 5.
Solving the problem of changing the design variables based on the gradient
information such that the performance is improved (minimization of f) while
observing the constraints falls to an optimizer.
2.2.2 The optimizer
Several options are available for the choice of optimizer but the most popular in
structural topology optimization is the family of convex approximation methods
such as CONvex LINearization (CONLIN) (Fleury and Braibant, 1986) or the
Method of Moving Asymptotes (MMA) (Svanberg, 1987). Both these methods
solve the optimization problem by generating convex approximations as illustrated
in Fig. 2.2 and solving the approximated problem using a dual formulation.
Since their introduction these optimizers have proven very ecient in numerous
applications in various elds of structural optimization. The convex approximation
16 2.2. Improving the design
a
a
1
2
g a G ( )
Originaldesignspace
Convexapproximation
g a G ( )
~
aopt
a
~
opt
f ( ) a
Figure 2.2: Design space and convex approximation.
methods are particularly ecient for problems involving many design variables and
few constraints, which makes them ideally suited for topology optimization.
In this study the Method of Moving Asymptotes has been implemented in
MUST using FORTRAN77 source code kindly made available by Professor Krister
Svanberg, Royal Institute of Technology, Stockholm, Sweden.
MMA solves the optimization problem by creating a convex monotonic approxi-
mation from a rst order Taylor series expansion around the design point, a. The
approximation is made in mixed variables of linear and reciprocal terms, i.e. a
and 1/a. This provides a set of approximation functions for both objective,
f(a),
and constraints, g(a), and thus, the approximated optimization problem to solve
is stated from (2.6) as:
Objective : min
a
f(a)
Subject to : g(a) G
a
min
a a
max
R(u, a) = 0
_
_
(2.15)
The solution of (2.6) is then achieved by successive solution of increasingly good
(hopefully) approximations (2.15) until convergence is reached. The quality of
the approximation is controlled using the lower and upper asymptotes, L and
U, respectively as shown in Fig. 2.3. Depending on the sign of the gradient in
the design point, df(a
3
.
but not necessarily normal. This also implies that the transverse shear strains will
be continuous across the thickness, which is important in relation to multilayered
structures as discussed in Section 3.2. The last assumption of zero transverse stress
(s
33
0) will be enforced through the constitutive relation as shown later.
The node directors are very important in that they are used to dene the nodal
displacements and rotations. This is done by setting up, in each node, a Cartesian
base called the director coordinate system having the node director as z-axis. The
node director will be denoted v
3
and dened as:
v
3
=
g
1
g
2
|g
1
g
2
|
(3.3)
This might seem inconsistent with Fig. 3.2 since the ghost nodes need not lie
on a surface normal but in practice (3.3) is the only viable approach since the
coordinates of the ghost nodes are non-existent in the model. However, using
(3.3) directly can lead to discontinuities in the displacements since the surface
normal may change from element to element depending on the mesh generator and
element type. This will in particular be the case when using four-node elements to
model curved geometries. Consequently, a simple algorithm that averages adjacent
node directors for all elements is applied in MUST to ensure continuity. The two
additional base vectors in the director coordinate system are dened from v
3
and
the auxiliary vector a as:
v
1
=
a v
3
|a v
3
|
; v
2
= v
3
v
1
(3.4)
where a = j if j v
3
,= 0 and a = k if j v
3
= 0. Here, the vectors j and k
are global unit vectors along the y- and z-axis, respectively. The key role of the
director coordinate system is to dene the local rotations, and , of the node
director, v
3
. This introduces two additional degrees of freedom per node and
is consistent with Reissner-Mindlin shell theory. These rotations are dened as
shown in Fig. 3.3.
24 3.1. Geometry, kinematics and material
As indicated in Fig. 3.3 a rotation, , of v
3
about v
1
causes a linear displacement
of magnitude sin in the direction of v
2
. In the same way the director vector
tip is displaced in the direction of v
1
a distance sin . Assuming small rotations
this amounts to a local linear displacement of u
a
=
a
,
a
, 0 which can be
transformed to a global displacement using a simple tensor transformation rule (see
e.g. Sokolniko (1956)). This results in a global displacement expressed in terms
of the local rotations as u
a
=
a
v
a
1
a
v
a
2
. This expression is essential in shell
analysis since it describes the displacement of any point in the thickness direction,
relative to the reference surface. Now, the total displacement of any point in the
shell may be written in terms of the in-plane part of the shape functions, i.e.
N(r, s, 0), and the local rotations as:
u =
A
a=1
N
a
(r, s, 0)
_
u +
t
2
h
_
v
1
v
2
_
_
a
(3.5)
where a is the nodal number, A is the number of nodes in the element and h
is the shell thickness. Note that all quantities must be evaluated at node a.
The expression in (3.5) completely denes the kinematics of the shell element in
terms of ve degrees of freedom, (u
i
, , ), while observing the Reissner-Mindlin
assumptions. The expression (3.5) may be expressed in terms of a modied
shape function operator,
N, dened as
N =
N(N(r, s, 0), t) such that u =
N
u
where
u is dened as a vector containing all the nodal degrees of freedom, i.e.
u = [u
a
x
, u
a
y
, u
a
z
,
a
,
a
[
T
. This rather tedious notation will be abandoned
and from this point and on, N and u will always imply modied shell quantities
as dened by (3.5).
The strain components follow naturally from (3.5) and may be expressed in terms
of the covariant base vectors as:
ij
=
1
2
_
t
g
i
t
g
j
0
g
i
0
g
j
_
=
1
2
_
u
x
j
g
i
+
u
x
i
g
j
+
u
x
j
u
x
i
_
(3.6)
where
ij
indicates local strain since (3.6) is stated in the covariant base vectors.
The subscripts t and 0 indicate deformed and initial state, respectively, i.e.
t
g
i
= g
i
(x +u). The strain may also be obtained from terms of the deformation
gradient, F =
0
t
x, which is perhaps a more direct approach:
ij
=
1
2
_
F
T
F I
_
(3.7)
This can be eciently evaluated in terms of the displacement gradients, which are
readily available in a nite element framework such that F =
0
t
u + I. Which
of (3.6) and (3.7) is most convenient depends on the type of nite element being
implemented.
Chapter 3. Multilayered shell nite elements 25
3.1.3 Material
As briey mentioned above the third of the Reissner-Mindlin assumptions (s
33
= 0)
is enforced through the linear constitutive relation s = C, where s is the second
Piola-Kircho stress. This will be addressed in the following.
First, we dene the constitute matrix of an orthotropic material, which is the most
general case for the present work:
C =
_
_
C
11
C
12
C
13
0 0 0
C
22
C
23
0 0 0
C
33
0 0 0
C
44
0 0
Sym. C
55
0
C
66
_
_
(3.8)
Enforcing s
33
= 0 simply involves deleting the third row and column in (3.8) and
the most general form of C is then:
C =
_
C
11
C
12
0 0 0 0
C
22
0 0 0 0
0 0 0 0
C
44
0 0
Sym. C
55
0
C
66
_
_
(3.9)
where the coecients are given in terms of the engineering constants as:
C
11
=
E
1
1
12
21
C
22
=
E
2
1
12
21
C
12
=
21
E
1
1
12
21
C
44
= G
12
C
55
= G
23
C
66
= G
13
(3.10)
where the modied terms
C
11
,
C
12
and
C
12
can be derived from the general
expressions by setting the transverse Poisson ratios equal to zero (see e.g. Reddy
(2004)). In (3.10) the factor is the shear correction factor taken to be 5/6 for both
multi- and single-layered structures. This is somewhat crude but the predictive
capabilities of the formulation are quite good so no further steps have been taken
towards improving this along the lines of e.g. Pai (1995) or Auricchio and Sacco
(1999). For sandwich structures, however, MUST supports an alternative scheme
in which is set to 1.0 for the core layer while the transverse shear stinesses
are set to zero for the skin layers. This corresponds to assuming that transverse
shear is supported only by the sandwich core and provides very good predictive
capabilities for sandwich structures (Jensen et al., 2001).
From (3.10) it is apparent that only in-plane normal material properties and out-of-
plane shearing properties are taken into account, which is consistent with the shell
assumptions. Each of the expressions above are stated in the orthotropic principal
26 3.1. Geometry, kinematics and material
a
b
v
3
a
v
3
b
Fiberdirection
c
v
3
c
Figure 3.4: Conceptual dierence between director vector (dotted line) and material
coordinate system (solid lines), which changes through the thickness.
directions, which are of no particular interest. Therefore the need arises for a
coordinate system that uniquely denes the orientation of the orthotropic material
at any point in any layer of the element. This coordinate system is very ttingly
called the material coordinate system and the major dierence between this and
the director coordinate system is that the material coordinate system changes
from layer to layer as illustrated in Fig. 3.4 for a curved nine-node element. For
at (three- and four-node) elements the coincidence of all normal vectors could
be exploited to bypass evaluation of some of these coordinate systems but as
the implemented element routines are generic the same approach is used for all
elements.
The material coordinate system is spanned by the base vectors m
i
and for the
constitutive properties to be meaningful the system must span the tangent plane
in any given point. As such it is natural to dene m
3
to be normal as:
m
3
=
g
1
g
2
|g
1
g
2
|
(3.11)
which is the same denition as used for the director vector, v
3
, the dierence
being that (3.11) is updated in all Gauss-points and layers (Fig. 3.4). The in-
plane material base vectors may in principle point in any direction as long as they
are orthogonal to m
3
but for practical modeling purposes it proves convenient to
introduce a projection of a dened (or default) global material system onto the
element plane. This procedure has been adopted from the commercial software
packages ANSYS and COSMOS and can be stated as follows.
A global material system is dened in terms of three vectors, d
i
, which are either
taken to be the global Cartesian base vectors (default) or, if dened, read from
user input. These vectors, d
i
, will be projected onto the element plane as follows.
If the angle between d
1
and m
3
is smaller than 45
, d
1
is projected as m
1
. Else
d
2
is projected as m
1
. The procedure is outlined in Algorithm 3.1.
The denition in Algorithm 3.1 provides a material coordinate system that by
Chapter 3. Multilayered shell nite elements 27
Algorithm 3.1: Pseudo code for setting up material coordinate system
Get covariant base vectors g
1
and g
2
Compute shell normal, m
3
, from (3.11)
if (Global material coordinate system dened) then
Read dened system into vectors, d
1
, d
2
and d
3
else Default
Use global Cartesian, i.e. d
1
= i, d
2
= j and d
3
= k
end if
if
_
abs(d
1
m
3
) < 45
_
then Project using either d
1
or d
2
Compute a = d
1
m
3
else
Compute a = d
2
m
3
end if
Compute m
1
= m
3
a
Compute m
2
= m
3
m
1
default is oriented identically for all elements and has an intuitive feel when
handling materials in modeling situations. The user also has the option of dening
global material coordinate systems, which allows the user to control the material
direction explicitly on various parts of the model. The procedure outlined above
provides a well-dened framework for handling orthotropic materials using a single
angle, . This angle is dened relative to the local, projected material coordinate
system in any point, as shown in Fig. 3.5.
The constitutive behavior can now be described by (3.9) and an in-plane rotation of
Csuch that C = T
T
()CT() where T() is a standard transformation matrix (see
e.g. Cook et al. (1989)). Depending on the element formulation the constitutive
matrix must be transformed once more since the material coordinate system is
usually not the preferred frame for setting up the element stiness matrix. In
the present work the element stiness is either expressed in the global Cartesian
frame (i, j, k) or the covariant frame (g
1
, g
2
, g
3
) depending on what provides
fewest element level computations. The transformations employed are stated
in Cook et al. (1989) for transformation to (i, j, k) and in Heinbockel (2001) for
transformation to (g
1
, g
2
, g
3
).
3.2 Laminate description
The topics discussed in the previous section are generic for single- and multilayered
structures but provide no framework for handling multilayered elements. This
will be addressed in the following by introducing the Equivalent Single Layer
(ESL) laminate description, which has been adopted in this work. In an ESL
description the layers of the laminate are assumed to be perfectly bonded together
and thus, displacements will be continuous across the thickness. Due to the
kinematic assumptions this implies that in-plane strains (
11
,
22
and
12
) are
28 3.2. Laminate description
m
1
m
2
m
3
q
1
2
Figure 3.5: Orthotropic principal directions (1, 2) with reference to the material
coordinate system base vectors m
i
.
continuous across the thickness as well and furthermore, that transverse strains
(
13
and
23
) are constant through the thickness. Consequently, a single node in
the thickness direction is sucient for describing the kinematics of the laminate
and consequently, all quantities derived in the previous sections apply directly
to ESL shell models. The downside of using ESL is that interlaminar eects
such as delamination are dicult to predict due to the absence of normal stress
and strain components in the thickness direction. This can be remedied to some
extent by introducing stress recovery as discussed by e.g. Cho and Choi (2001)
in which the equilibrium equations are integrated a posteriori to obtain a better
result for the stress components. Another approach is to introduce a modied
kinematic assumption, allowing for piecewise continuous displacements. This is
known as the zig-zag approach and has become increasingly popular over the last
years, see e.g. Carrera (2003) for a review. However, the unmodied laminate
description employed here is still the most widely used since it provides a good
approximation of the structural stiness and is computationally less expensive.
Both these properties are desirable in the present context and thus, the method
chosen provides a sucient and well-established basis for doing global analysis and
optimization.
The notation associated with laminates is shown in Fig. 3.6 where the number of
layers is designated N
l
and the layer thickness is h
l
. Each layer is also associated
with a material, C
l
, as well as an angle,
l
, for orthotropic materials. When mixing
element- and layer-wise quantities superscripts e and l will be used, respectively,
to relate each quantity.
Each layer, l, in the laminate of Fig. 3.6 is described by the constitutive relation:
s
l
= C
l
(
l
)
l
(3.12)
and consequently, the stresses will be layer-wise continuous since strains are
continuous and C
l
changes from layer to layer. The constitutive behavior of
the entire laminate is obtained by integration through the thickness. In Classical
Laminate Theory (CLT) this is achieved explicitly by forming the extension matrix,
Chapter 3. Multilayered shell nite elements 29
Layer 1
Layer 2
Layer 3
Layer N -1
Layer N
h
h
l
l
l
Figure 3.6: Laminated composite shell structure.
A, the bending-extension matrix, B, and the bending matrix, D, which together
describe the laminate behavior (see e.g. Jones (1998)). In the present study
numerical through the thickness integration has been employed as a more general
scheme for obtaining the response of any laminate while maintaining the 3D
continuum mechanical formulation. This is also the prevailing approach in the
literature and commercial nite element software packages.
3.2.1 Numerical integration
Evaluation of the integrals in (2.8) is done using full Gauss quadrature and for
laminated structures this can be extended to a layer-wise integration scheme.
From laminate theory it is well-known that the laminate behavior is dependent
on the thickness coordinate to a power of three, i.e. we may write tentatively
C = C(t, t
2
, t
3
). Using a two-point Gauss quadrature in the thickness direction is
therefore sucient as long as the constitutive relation is linear, which will always
be the case in this study. The integration is made layer-wise by introducing an
additional mapping such that the thickness coordinate, t, is expressed from the
layer thickness coordinates, t
l
, as:
t = 1 +
2
h
N
l
i=1
_
h
i
h
l
(1 t
l
)
_
(3.13)
where t
l
=
1
3
for two-point quadrature. Now, the global tangent stiness matrix
becomes:
K
T
=
N
e
k=1
_
_
N
l
q=1
___ _
B
T
q
u
q
s
q
+B
T
q
s
q
u
q
_
h
l
h
drdsdt
l
_
_
k
(3.14)
30 3.3. Unlocking Assumed Natural Strain
where the ratio h
l
/h arises from (3.13) since dt = (h
l
/h)dt
l
. Similarly, the internal
force vector is obtained as:
r =
N
e
k=1
_
_
N
l
q=1
___
_
B
T
q
s
q
_
h
l
h
drdsdt
l
_
_
k
(3.15)
In the following the layer-wise form of volume integration in (3.14) and (3.15) will
be implied whenever stating the stiness matrix or internal force vector in integral
form.
Using (3.14) eectively subdivides each element into N
l
sub-elements in the
thickness direction and so, the computational cost will increase linearly with the
number of layers. For large models with many layers this can be an impediment,
particularly for higher-order elements. Alternatively, explicit thickness integration
can be used by making simplifying assumptions regarding the thickness variation of
the inverse Jacobian, which allows the element matrices to be decomposed and inte-
grated analytically. This is discussed for the AG-method of Zienkiewicz and Taylor
(1991) by Kumar and Palaninathan (1997, 1999) who used it for geometrically
linear and nonlinear structures, respectively. Such methods are not widely used
and cannot be implemented in the existing element routines in MUST without rst
making considerable changes. Consequently, no steps have been taken towards
implementing explicit integration but this could be considered in the future to
obtain higher computational eciency.
3.3 Unlocking Assumed Natural Strain
An important aspect of shell nite elements is the problem of locking, which
has historically haunted shell elements. The problem arises in elements derived
directly from the kinematics presented above, i.e. by inserting the displacement
interpolation (3.5) in the strain denition (3.6) to obtain the strain-displacement
matrix. This approach results in deciencies in the element formulation in that
the element will exhibit overly sti behavior in some situations the element mesh
is said to lock.
3.3.1 Locking
Locking is usually characterized as either shear locking, membrane locking, volume
locking or thickness locking and the dierent types are encountered depending on
the element type and conditions of the model. This brief discussion is limited
to the elements in MUST and an exhaustive account is left the literature, e.g.
Cook et al. (1989) or Bischo (2004).
In short, locking arises due to an elements inability to properly represent the
deformation it is supposed to model. Such problems are more profound in some
element types than others and as a rule of thumb the problem decreases as the
Chapter 3. Multilayered shell nite elements 31
element order gets higher. Indeed, the SHELL16 element in MUST has proven
very robust and only exhibits locking behavior under extreme conditions such
as when elements get highly curved and distorted, which can cause membrane
locking. The SHELL16 element can therefore serve as reference for other elements
although it is not ecient for optimization problems in general due to its high
computational cost. The SHELL9 element in MUST is less expensive and has
been used extensively with good results. However, the element may exhibit shear
locking when subjected to non-constant bending moments and can also suer from
membrane locking when curved. The diculty in using the SHELL9 and SHELL16
elements lies in the fact that the occurrence of locking is hard to predict and can
be dicult to detect unless it is severe. The only viable approach with these
elements is therefore to recognize the problem and perform comparative solutions
to validate the results. For this and other reasons (Section 3.5) the linear MITC
elements remain the choice of preference.
3.3.2 Linear MITC elements
The problem of locking is particularly profound in linear three- and four-node
elements, which are known to suer from shear locking when the element
thickness becomes small compared to the element edge length. This is due to
parasitic transverse shear strains, which result in an gross overestimation of the
element stiness. Still, the linear elements remain very popular due to their low
computational cost and consequently, considerable eort has gone into improving
these elements and eliminating the locking behavior. Historically, this was rst
achieved using a reduced order of integration in the Gauss quadrature but this gives
rise to problems with spurious energy modes. In recent years, the preferred cure
for locking has instead been the Assumed Natural Strain (ANS) methods, which
aim at xing the deciencies in the element formulation by introducing modied
(assumed) strain expressions in the natural element (r, s, t)-space. Elements
derived using assumed strain expressions will be called stabilized elements and
denoted as MITCn while elements derived directly from the kinematics will be
referred to as non-stabilized elements and denoted SHELLn (n indicates the
number of element nodes).
In the present work the ANS method introduced by Dvorkin and Bathe (1984)
as Mixed Interpolation of Tensorial Components (MITC) has been applied
successfully to four- and three-node elements, eectively eliminating problems
with shear locking. The basic idea is to use a new set of points, p = 1 . . . N
p
,
for strain evaluation instead of the Gauss points, which would be used in non-
stabilized elements. These new points are called tying points and are chosen such
that the strains evaluated in these points are free from parasitic strains. Having
obtained a set of correct strains these are then interpolated across the element,
thus eliminating problems with parasitic strains altogether. The strains in the
tying points, (r
p
, s
p
, t
p
), are evaluated directly from the strain denition in (3.6),
i.e.
ij
[
p
ij
(r
p
, s
p
, t
p
). Introducing a set of interpolation functions for each
32 3.3. Unlocking Assumed Natural Strain
r
s
A
C
B D
1 2
4 3
s = -1
r = 1 r = -1
s = 1
1 2
3
B D
C
A
r
s
Figure 3.7: Three- and four-node elements with MITC tying points (AD) used for
transverse shear strain evaluation.
strain component, N
ij
p
, the assumed strains (AS) are expressed as:
AS
ij
=
N
p
p=1
N
ij
p
(r, s)
ij
[
p
(3.16)
where N
p
is the number of tying points, p, for the ijth strain component. The
MITC interpolation functions, N
ij
p
, must naturally fulll the relation:
N
ij
p
[
q
=
kq
, q = 1 . . . N
p
(3.17)
so that the pth interpolation function assumes the value 1 in the pth tying point
and the value 0 in all other tying points. It is therefore natural to choose Lagrange
polynomials of an order appropriate to the number of tying point, i.e. rst order
Lagrange polynomials for two tying points etc.
The two expressions (3.16) and (3.17) above are general for all MITC elements,
which include linear, quadratic and cubic elements. The location and number of
the tying points will vary depending on the order of the element but the basic idea is
the same (see Bucalem and Bathe (1993) for details). For the three- and four-node
MITC elements there are four tying points, AD, used for transverse shear strain
evaluation (two tying points for each strain component). Both transverse shear
strains are interpolated linearly but
13
is interpolated along s between points A
and C and
23
is interpolated along r between points B and D. The four points are
located at the element mid-sides, i.e. A = (0, 1, 0), B = (1, 0, 0), C = (0, 1, 0),
D = (1, 0, 0), as shown in Fig. 3.7. The reason for choosing these locations for the
tying points is that the element mid-sides are the only points in which the strain
denition (3.6) is able to evaluate zero transverse shear strain in pure bending.
This is because the mid-sides are the only points in which the deformation gure
is correct, i.e. the normal remains normal whereby the transverse shear strains are
zero, Fig. 3.8. However, the choice of tying point for the MITC3 element involves
a problem since the transverse strains become dependent on the node numbering.
Chapter 3. Multilayered shell nite elements 33
(a) (b)
Figure 3.8: Linear shell element subjected to constant bending moment (a) and
corresponding actual deformed shape (b) with correct deformed shape (dotted line).
This problem is also encountered in other formulations, e.g. Bletzinger et al.
(2000), but recently Lee and Bathe (2004) proposed a MITC3 element with an
alternative interpolation, which renders the element insensitive to node numbering.
This approach has not been adopted yet but tests have revealed that the results
from the MITC3 element in MUST vary by no more than 3% when permutating
the node numbering
2
. This is satisfactory considering that the MITC3 element is
mainly complementary to the MITC4 element in mixed meshes and as such rarely
used exclusively.
The assumed transverse shear strains for both the three- and four-node elements
are now found by evaluating the tying point strains,
ij
[
p
, and using (3.16) for
N
p
= 2 with rst order Lagrange polynomials:
13
=
1
2
(1 s)
13
[
A
+
1
2
(1 + s)
13
[
C
23
=
1
2
(1 r)
23
[
D
+
1
2
(1 + r)
23
[
B
(3.18)
which provides linearly varying transverse shear stresses in the r, s-plane and
constant stresses in the t-direction. The assumed strains (3.18) are used for
determining the strain-displacement matrix, and as such the stiness matrix, and
allow the elements to show proper behavior for thin elements.
3.4 Implementation
The shell element library in MUST originally consisted of the MITC4, SHELL3,
SHELL4, SHELL6 and SHELL9 elements with linear and nonlinear capabilities.
These were implemented by Jensen et al. (2001) and later, a linear MITC3 element
was added and tested by Moser (2002) and Poulsen et al. (2003) adapted MUST
to accommodate a SHELL16 element and performed extensive testing. In those
implementations the invariant properties of the energy terms in (2.1) were not
exploited and consequently, a number of essentially redundant transformations
were used. Removing these transformations and formulating the stiness matrix
in the covariant base vectors instead of the global Cartesian system lowered the
computational cost of forming the global tangent stiness matrix by approximately
2
Tests performed by Moser (2002) in connection with the original implementation of the linear
MITC3 element.
34 3.4. Implementation
15% and rendered the elements very ecient when comparing to commercial codes.
However, a number of trivial errors existed in the original implementation of the
tangent stiness matrix of the MITC4 elements. This did not cause noticeable
problems for moderate nonlinear analysis but was quite clear when comparing the
stiness matrix to a central nite dierence approximation. The implementation
strategy was (and still is) to formulate the tangent stiness matrix analytically,
which results in some rather involved expressions and it is natural to suspect
insucient dierentiation in the derivation of the linearized terms or other errors
in the derivations. Unfortunately, even after spending considerable time on the
problem, a proper solution has not been found. Instead, the MITC3 and MITC4
element routines have been completely re-implemented and furthermore, the
SHELLn element routines have been optimized for higher computational eciency.
3.4.1 The MITC elements
As in the original implementation we proceed from (2.10) by splitting the strain
into a linear and nonlinear part, i.e. = e + , whereby the strain-displacement
matrix is also split such that B = B
e
+B
k=1
_
V
_
(B
)
T
k
u
k
s
k
+ (B
e
+B
)
T
k
C
k
(B
e
+B
)
k
_
dV (3.19)
The idea in the modied implementation is to create a general framework for
computing (3.19) using dierent routines depending on the element. In other
words it is a modular approach in which routines are needed for determining B
e
,
B
and (B
T
u
_
g
T
i
N
r
j
u +g
T
j
N
r
i
u
_
=
1
2
_
g
T
i
N
r
j
+g
T
j
N
r
i
_ (3.20)
Chapter 3. Multilayered shell nite elements 35
where i and j are set as appropriate for the given strain component (row in B
e
),
i.e. i = j = 1 for e
1
= e
11
and so forth. This leads to the usual denition of the
linear strain-displacement matrix, i.e. e = B
e
u and from the strain denition
(3.6) it follows that e = B
e
u. It is implied in (3.20) that the covariant vectors are
computed in the appropriate node depending on the degree of freedom being used
from u. The nonlinear contribution is derived as:
B
=
u
=
1
2
u
_
u
T
N
T
r
j
N
r
i
u
_
= u
T
N
T
r
j
N
r
i
(3.21)
and thus it follows that = B
, can be
determined as shown in Algorithm 3.3. Due to symmetry it is only necessary to
determine the upper or lower triangle of the matrix, which saves quite a number
of lines in the code and improves the computational eciency. In both Algorithm
3.2 and 3.3 the variable ndof is the number of degrees of freedom for the element,
which is the number of nodes multiplied by the number of degrees of freedom per
node, e.g. ndof = 4 5 = 20 for four-node elements.
The code generated by MAPLE is ecient but the practical down-side is that it
consists of many lines of code, rendering debugging somewhat dicult. Also, the
element routines are rather cumbersome to work with in the editor but still, from
a computational and modeling point of view the result is very satisfactory. To
obtain a more editor-friendly implementation the element routines could be stored
in a pre-compiled library, but this has not been implemented.
3.4.2 The SHELLn elements
The SHELL3, SHELL4, SHELL6, SHELL9 and SHELL16 elements are imple-
mented using the so-called AG-method suggested by Zienkiewicz in which the
nonlinear element matrices are expressed in terms of three auxiliary matrices A,
Gand Hsuch that B
= AGand K
=
_
G
T
HGdV (see Zienkiewicz and Taylor
(1991) for details). The current implementation of the AG-method in MUST is
not computationally ecient but very convenient since the element routines can be
generalized for any number of nodes (as opposed to the MAPLE routines, which
must be derived for each individual element). Therefore, the AG-method is still
used in the current implementation although the element routines have been code
optimized resulting in about 10% performance increase compared to the original
routines implemented by Jensen et al. (2001).
36 3.5. Numerical verication
Algorithm 3.2: Pseudo code for deriving B
e
and B
in MAPLE
Compute e
Compute
for i = 1 to 6 do
for j = 1 to ndof do
Compute (B
ij
)
e
= e
i
/u
j
Compute (B
ij
)
=
i
/u
j
end for
end for
Algorithm 3.3: Pseudo code for deriving (B
T
/u)s in MAPLE
Compute B
T
Compute s
for i = 1 to ndof do
for j = 1 to i do
Compute (K
ij
)
6
k=1
((B
T
ik
)
/u
j
)s
k
end for
end for
3.5 Numerical verication
The numerical verication of the shell elements consists of two parts: patch testing
and general verication by result comparison. The elements chosen for evaluation
are the MITC3, MITC4 and SHELL9 elements, which are the three elements used
most frequently for analysis and optimization with MUST. Extensive testing has
also been done to determine if the MITC elements suer from shear locking when
the element thickness becomes small compared to the element edge length. These
tests have all shown that this is not the case but for the sake of brevity no such
results are presented here.
3.5.1 Patch testing
Patch testing is done in order to verify that the elements exhibit proper
convergence properties by checking element consistency and stability as discussed
by e.g. Razzaque (1986) and Felippa (2003). For the element types implemented
in MUST, which are fully integrated and employ standard shape functions, these
requirements are expected to be met and consequently, the patch test can be
construed as a check for correct implementation.
As patch test strategy for choice of mesh and boundary conditions we adopt the
approach of Dvorkin and Bathe (1984) and Lee and Bathe (2004) for the MITC4
and MITC3 elements, respectively. The patch test is successfully passed if the
element can display constant stress when subjected to appropriate boundary
conditions in a non-regular mesh. If the element passes it implies that the
Chapter 3. Multilayered shell nite elements 37
2.000E+005
2.000E+005
2.000E+005
2.000E+005
2.000E+005
2.000E+005
Name:MITC3patchtest s13 -
2.000E+005
2.000E+005
2.000E+005
2.000E+005
2.000E+005
2.000E+005
Name:MITC4patchtest s11 -
Sxx Sxz
Figure 3.9: Randomly selected patch test results for MITC4 elements (left) and MITC3
elements (right). The same uniform stress results are obtained for all other components.
kinematics are correct since stresses are computed from strains, which are in turn
computed from displacements. As expected both MITC elements pass all patch
tests for the ve stress components s
11
, s
22
, s
12
, s
23
and s
13
. Not all results are
reproduced here but the results for s
11
and s
13
obtained with the MITC4 and
MITC3 elements, respectively, are shown in Fig. 3.9.
The same results have been obtained for the SHELLn family of elements, but these
have been left out for brevity.
3.5.2 Nonlinear comparative test
To demonstrate the nonlinear performance of the implemented elements in MUST
they are tested in a single- and multilayer conguration and the results are
compared to those obtained with commercial software packages. The problems are
solved using the arc-length method. The geometry chosen for the tests is shown
in Fig. 3.10 for the multilayered case. The cap spans a base of 1000 1000 [mm]
and the apex rises 100 mm above the base (detailed geometric information is
available in Section 4.3). For the multilayered case the total thickness of the shell
is 20 mm distributed as 1/1/16/1/1 [mm] with the thin sheets being glass/epoxy
composite with E
x
= 36 GPa, E
y
= E
z
= 6.4 GPa, G
xy
= G
yz
= G
xz
= 2.4
GPa,
xy
=
xz
= 0.27 and
yz
= 0.33 while the thick middle layer is taken to
be isotropic polymeric foam with E = 125 MPa and = 0.3. For the single-layer
case the total thickness is 25 mm of aluminum with E = 70 GPa and = 0.3.
For the single-layer case three reference solutions have been obtained with ANSYS
and ADINA. The ANSYS SHELL181 element is a four-node element formulated
with MITC as the MITC4 element in MUST and is therefore a natural choice.
Furthermore, two quadratic elements have been used the SHELL93 element in
ANSYS and the MITC9 element in ADINA, which is widely recognized as one of
the most robust elements available. As can be seen from Fig. 3.11 the results from
MUST and the commercial codes correlate very well.
38 3.5. Numerical verication
Glass/epoxycomposite(45)
Polymericfoam
Glass/epoxycomposite(-45)
Glass/epoxycomposite(-45)
Glass/epoxycomposite(45)
Figure 3.10: Geometry of spherical cap used for performance comparison.
For the multilayer test three reference solutions have been obtained using ANSYS,
ADINA and COSMOS. The element of choice was again the SHELL181 element
in ANSYS but unfortunately the element was unable to converge. ANSYS
has recognized the problem and their Research and Development department is
currently investigating the cause of this problem. Instead the SHELL4L element
in COSMOS has been used as well as quadratic elements in ADINA and ANSYS.
The results are compared in Fig. 3.12 and again, the correlation is very good.
Equally good correlation has been found with solutions obtained using the
quadratic TYPE75 element in MSC.MARC
3
and the MITC4 element in ADINA,
but these have been left out for the sake of brevity.
3.5.3 General remarks
During the course of this work a large number of other examples have been tested,
both by the author and students working on MUST, and the performance of the
shell elements in MUST has consistently been found to be very good. Furthermore,
with the modications made in this work to the element routines the computational
eciency of MUST for the tested cases is actually higher than or equal to that
achieved with commercial codes. In the multilayered example above MUST
uses approximately 4 minutes with the MITC4 element for generating the entire
equilibrium path, only matched by COSMOS, while ADINA uses approximately
10 minutes. For the same example MUST uses 14 minutes with the SHELL9
element while ANSYS solves the same problem in 18 minutes using the SHELL91
(eight-node) element and again, ADINA takes last place by using 25 minutes with
the MITC9 element
4
.
3
Thanks to my colleague, Ph.D. student Lars Christian Terndrup Overgaard, who actually
pushed the buttons in PATRAN.
4
The poorer performance of ADINA is probably due to its force-control solver, which seems
to be less ecient than the arc-length solvers employed by the other packages.
Chapter 3. Multilayered shell nite elements 39
0,0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1,0
0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0
Normalized center point displacement
L
o
a
d
f
a
c
t
o
r
,
MUST (MITC4)
MUST (MITC3)
MUST (SHELL9)
ANSYS (SHELL181)
ANSYS (SHELL93)
ADINA (MITC9)
k
Figure 3.11: Nonlinear performance of MUST elements compared to commercial codes,
single-layer test with isotropic material. All results are normalized to MUST (MITC4)
results (solid line).
k
0,0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1,0
0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0
Normalized center point displacement
L
o
a
d
f
a
c
t
o
r
,
MUST (MITC4)
MUST (MITC3)
MUST (SHELL9)
ANSYS (SHELL91)
COSMOS (SHELL4L)
ADINA (MITC9)
Figure 3.12: Nonlinear performance of MUST elements compared to commercial codes,
multilayer test with iso- and orthotropic materials. All results are normalized to MUST
(MITC4) results (solid line).
40 3.6. Summary and conclusions
These comparisons are interesting in that computational eciency is very im-
portant for the eciency of the optimization where multiple analyses are run
successively. The results render the MITC3 and MITC4 elements particularly
interesting since they have demonstrated to oer computational eciency while
maintaining good predictive capabilities.
Another interesting observation regarding element choice is that higher-order
elements are often overkill in optimization since the benet of using fewer elements
to obtain the solution contradicts the need for many elements to accurately
parametrize the design. This is especially true in topology optimization where
a ne mesh is needed to represent the boundaries of the optimal design. Indeed,
when reviewing the literature, linear elements are found to dominate the arena in
this eld, which just emphasizes the fact that the MITC3 and MITC4 elements
represent an important tool.
3.6 Summary and conclusions
This chapter provided a birds eye view on the theory supporting the shell
nite elements in MUST. The elements are formulated using the assumptions
from First order Shear Deformation Theory (FSDT) and composite laminate
behavior is encompassed using an Equivalent Single Layer (ESL) description. The
preferred elements are the stabilized MITC3 and MITC4 elements, which use
Mixed Interpolation of Tensorial Components (MITC) to avoid problems with
shear locking. These elements have been implemented using a modular approach
based on FORTRAN code generated directly from MAPLE, which has resulted
in computationally ecient elements. A family of non-stabilized elements has
also been implemented and of these particularly the SHELL9 element has proven
ecient. The MITC3, MITC4 and SHELL9 elements were tested and showed good
predictive capabilities as well as high computational eciency.
4
Nonlinear topology optimization
T
his chapter shifts focus from analysis to optimization by addressing
the topic of material distribution in multilayered structures. Specically, the
multilayer shell formulation developed in Chapter 3 is deployed in a topology
optimization framework. This allows material to be added/removed in individual
layers, simulating stiening of the shell structure. The purpose of the chapter is
to investigate the inuence of nonlinear eects from large displacements on the
optimal topology in order to determine whether nonlinearities can be disregarded
when designing laminated structures using structural optimization. Additionally,
we want to get experience with nonlinear optimization of laminated composite
shell structures on familiar territory before proceeding to other optimization
elds. The bulk of the topics covered in this chapter are also documented in
Stegmann and Lund (2005b).
The testing ground for this investigation is topology optimization of multilayered
structures, which is fundamentally identical to classical topology optimization
of single-layered structures. The dierence is that instead of adding/removing
material over the entire thickness we add/remove material in specic layers. This
will be explained in greater detail in Section 4.1.
The remaining part of the chapter is organized as follows. In Section 4.1 the design
parametrization is discussed and the optimization problem is briey dened. In
Section 4.2 the evaluation of sensitivities is explained and nally, four numerical
examples for plates and shells are presented in Section 4.3.
4.1 Design parametrization
For multilayered topology optimization we distinct between voided and solid layers
of the laminate as shown in Fig. 4.1. The solid layers are those that remain
unchanged during optimization and the voided layers are those in which material is
added/removed. The strategy for adding/removing material is the SIMP method,
see e.g. Bendse (1989) or Bendse and Sigmund (2003) for details.
In this methodology the material stiness is scaled by introducing a density
42 4.1. Design parametrization
Solid
Solid
Voided
Voided
Voided
Solid
(b) (c) (a) (d)
Voided
Solid1
Solid2
Solid1
Voided
Voided
Figure 4.1: Typical lamina sequences for plate/shell topology optimization: (a)
classical topology optimization, (b) symmetric lay-up with voided core, (c) symmetric
lay-up used with voided outer layers and (d) generic scheme.
parameter,
e
, in the constitutive relation as:
s
l
=
e
C
l
l
(4.1)
This is standard in topology optimization and the only dierence for multilayered
structures is that the scaling is employed on specic layers (as indicated by
superscript l in (4.1)) instead of on the entire element. The solid layers are
not scaled and the stiness of these is calculated using the standard constitutive
relation, i.e. s = C. As in topology optimization the element densities,
e
,
constitute the design variables and only a single design variable is assigned to any
element. It should therefore be noted that even though several voided layers may
be specied, all voided layers within the same element will be scaled in the same
way.
Choosing in which layers to apply the scaling in order to perform various types of
optimization is the next step and several optimization schemes are available.
4.1.1 Optimization schemes
In principle there are no restrictions on how the scaling should be applied
but a few simple schemes have become standard in shell and plate topology
optimization of laminated structures. The available options are shown in
Fig. 4.1 but in the literature schemes (a) and (c) constitute the most employed
ones, see e.g. Soto and Diaz (1993), Lee et al. (2000), Belblidia et al. (2001) or
Belblidia and Bulman (2002).
The single-layer scheme in Fig. 4.1(a) represents classical topology optimization
and allows formation of through-the-thickness holes. Using this involves numerical
problems since the stiness matrix becomes singular when the element stiness
tends to zero. Remedies for this problem have been treated by e.g. Buhl et al.
(2000) and Bruns and Tortorelli (2001) who introduced various numerical tricks
to bypass the problem. As the topic of interest has not been this type of
Chapter 4. Nonlinear topology optimization 43
K 0 =
K 0
K 0
d
d
K 0
y
x
y
x
Figure 4.2: Laminate optimized with scheme (b) to zero stiness (K = 0) of the core,
which is not equivalent to removing the layer.
optimization no such measures have been implemented, and scheme (a) has not
been utilized.
The scheme in Fig. 4.1(b) seems interesting since core layout design is of
practical importance in many applications involving stiening with sandwich cores.
However, the scheme lags in that the removal of core material by scaling
e
to
zero (or near-zero) does not eliminate its stiness contribution entirely since the
presence of the solid layer in the numerical integration (3.13) still contributes to
the total stiness of the laminate. This can be illustrated by considering the
moment of inertia, I
zz
, of the structure illustrated in Fig. 4.2. Scaling the stiness
to zero eliminates the contribution to I
zz
from the core layer (white) but due to
the parallel axis theorem there will still be a contribution to the total moment of
inertia, proportional to d
2
. Consequently, scheme (b) does not provide a viable
method for solving the core layout problem and should be replaced by some update
scheme, which removes/adds both the stiness of layers and their contribution to
(3.13). There is no immediate solution to this problem and no solution has been
sought out in this work.
The two remaining schemes, however, have both been employed in the present
study. The rib stiener scheme in Fig. 4.1(c) allows formation of symmetric ribs
on both sides of a xed surface, which has many practical applications and suer
from none of the deciencies outlined above. Scheme (c) has been used by e.g.
Lee et al. (2000) and will be used in the rst numerical example in Section 4.3.
The general scheme in Fig. 4.1(d) illustrates the point that any lay-up is treatable
in the present formulation and indeed, two of the three numerical examples in (4.3)
employ an asymmetric lay-up to form ribs on one side of a shell structure. However,
care must be taken to obtain a physically sound model when using scheme (d)
and furthermore, introduction of internal voided layers should be avoided unless
specically required by the physical model being optimized.
4.1.2 Problem formulation
Having established a design parametrization the optimization problem may now
be stated. As previously mentioned the objective function is compliance and
44 4.2. Objective function sensitivities
introducing a volume constraint, V
c
, the problem is stated as:
Objective : min
C() = p
T
u()
Subject to : V V
c
0 <
min
e
1
R(u, ) = 0
_
_
(4.2)
The constraint acts on the eective volume, V , dened as the sum of the scaled
volume of all voided layers, N
vl
, in all elements, N
e
, i.e.:
V =
N
e
e=1
N
vl
n=1
e
V
n
e
(4.3)
The volume constraint is included to ensure that the optimizer does not increase
the stiness (reduce the compliance) by pushing all design variables to 1, which
would be a rather useless result. To reduce the occurrence of intermediate densities
the SIMP methodology is applied to the stiness scaling whereby (4.1) is actually
computed for each layer as:
s
l
=
p
e
C
l
l
(4.4)
where the SIMP power p 1 penalizes densities between 0 and 1 by making
them uneconomical. This helps push the optimal topology towards a black/white
(0/1) design with no or few intermediate values of
e
as described by e.g.
Bendse and Sigmund (2003). The power is often chosen statically to p 3 but
in this work an incremental scheme is also used in which the power, p, is increased
from a starting value to a nal value in steps of some chosen size. In this way the
penalization is increased as the design converges, which in our experience results
in more distinct topologies. As the power increases, however, the term
p
e
tends to
zero quite rapidly, which may cause numerical problems. To avoid this problem
entirely, a conservative lower bound of
min
= 1 10
3
has been used for all design
variables, as indicated in (4.2).
Solution of the problem in (4.2) is done as described in Section 2.2 using analytical
design sensitivities and the Method of Moving Asymptotes (MMA). This approach
has proven ecient in other works on nonlinear topology optimization (see e.g.
Bruns et al. (2002); Buhl et al. (2000)).
4.2 Objective function sensitivities
In determining the sensitivities of the objective function we assume design
independent loads and write for a particular design variable,
e
:
dC
d
e
= p
T
du
d
e
(4.5)
Chapter 4. Nonlinear topology optimization 45
The sensitivities du/d
e
are not readily available and hence, it is not possible to
determine the sensitivities of the objective function directly from (4.5). To circum-
vent this problem a few paths are available: (1) nite dierence approximation,
(2) direct dierentiation method and (3) the adjoint variable method. For the
nite dierence to be accurate a central nite dierence approximation should be
used but this involves two additional analyses and is therefore highly impractical
from a computational eciency point of view. The direct dierentiation method
is only favorable if the number of constraints is larger than the number of design
variables, which is not the case here. The adjoint variable method is fairly simple
to implement and has proven ecient for large numbers of design variables and
has therefore been chosen for this work. The procedure for forming an adjoint
problem (see e.g. Haug et al., 1986) is described in the following.
4.2.1 Adjoint sensitivity analysis
To establish an expression for the displacement derivative, du/d
e
, we use the
chain rule to obtain the derivative of the residual, R(u(), ) = 0, dened in
(2.4):
dR
d
e
=
R
u
du
d
e
+
R
e
= 0 (4.6)
Assuming that the structural problem has been solved, i.e. R 0, (4.6) may be
rewritten to yield:
du
d
e
=
_
R
u
_
1
R
e
(4.7)
where R/u is the global tangent stiness matrix dened in (2.8). Inserting the
above into (4.5) eliminates the unknown derivative du/d
e
:
dC
d
e
= p
T
_
R
u
_
1
R
e
(4.8)
In principle (4.8) is sucient for performing the sensitivity analysis. However,
doing so requires nding the inverse of the global tangent stiness matrix and
as such, that would be a very impractical approach. Alternatively, the following
equation could be solved for each design variable:
K
T
du
d
e
=
R
e
(4.9)
and the compliance sensitivities can then be obtained by substituting this solution
into (4.5) for all design variables. This constitutes the direct dierentiation
approach but instead we proceed from (4.8) by dening the adjoint variable vector,
, as follows:
T
= p
T
_
R
u
_
1
(4.10)
46 4.2. Objective function sensitivities
Finding the adjoint variables is very simple and fast since it just requires solving
the linear system:
K
T
= p (4.11)
in which the factored stiness matrix is readily available from solving the analysis
and the force vector, p, is known. Now, having solved for the objective function
sensitivities may be found directly as:
dC
d
e
=
T
R
e
(4.12)
Finding the derivative of the residual in (4.12) with respect to
e
can be done
by nding the derivative of the internal force vector dened in (2.4) since
displacements have been assumed to be independent of the design variables, .
The internal element force vector is stated again for reference:
r =
_
V
B
T
s dV (4.13)
Finding derivative of (4.13) eectively only involves nding the derivative of the
stress vector since the strain-displacement matrix, B, is not dependent on the
design variables. Thus, the residual derivative can be found by computing the
derivative of (4.4) as:
s
l
e
= p
(p1)
e
C
l
l
(4.14)
And so, the vector product in (4.12) only yields contributions for the given element,
e, which renders the sensitivity calculation quite ecient.
It is important to note that the analysis must be solved accurately in order to
obtain accurate sensitivities from (4.12) since one of the basic assumptions is
R = 0. This is because of the NAND approach taken in this work where the
analysis is solved prior to computing the sensitivities. In our experience the error
tolerance on the residual in the nonlinear solution should be no more than around
10
10
to 10
8
on the nal iteration to avoid problems with numerical noise.
4.2.2 Multiple load cases
When solving the nonlinear topology optimization problem for a number of distinct
loads it is likely that the individual designs, although optimal at a single load, are
not optimal at other loads. To obtain a good compromise that works well at
both light, medium and heavy loading we solve the problem for the same load
case but for dierent load magnitudes. The objective, C(), from (4.2) is then
expressed as a weighted sum:
C() =
N
q
q=1
w
q
p
T
q
u
q
() (4.15)
Chapter 4. Nonlinear topology optimization 47
0
200
400
600
800
1000
0
20
40 60 80 100
p
u
a
b
c
e
d
Figure 4.3: Fictitious load-displacement space with selected load points a = (1, 10),
b = (15, 200), c = (25, 350), d = (80, 550) and e = (100, 1000).
where q is the load case number and w
q
is a weight factor for load case q. The
weights can be chosen arbitrarily but the natural approach would be to choose
them such that the contribution from each load case is approximately equal. This
ensures that no single load is favored over others and should therefore yield the
design that performs best at all loads. In the present work two strategies have
been tested:
I : w
q
=
|p
1
|
|p
q
|
II : w
q
=
(
p)
T
1
u
1
(
p)
T
q
u
q
(4.16)
where (
p)
T
q
u
q
in scheme II is the estimated compliance at load number q.
Obtaining these estimates either requires qualied guessing or additional analysis.
The approach used most frequently in this work is to run an initial nonlinear
solution using the arc-length solver to obtain the response over the entire load-
displacement space. This initial analysis, although not optimized, can then be
used to estimate the compliance at dierent load levels. This requires quite a bit
of manual processing and so, scheme I is by far the easiest to handle and could be
preferred for this reason. However, using scheme I can lead to favoring of the high
load cases since the displacements in this scheme (4.16) are disregarded resulting
in a factor of 10 1000 on the weight. To illustrate the dierence between the
two schemes a ctitious load-displacement space has been created and ve distinct
loads, ae, have been selected, Fig. 4.3.
Now, using (4.16) the weights for the compliance in (4.15) can be found from
Fig. 4.3 by inserting. The results are listed in Table 4.1. As indicated, the
dierence between the two schemes is signicant but for large numbers of q this
can be leveled out to some extent by choosing more loads in the lower part of the
spectrum. In general, however, it is more fool proof to choose scheme II since
it is hard to predict how the weighting aects the results. There is no right
approach to choosing the weighting scheme and a bit of tweaking of the weighting
factors must probably be endured before obtaining satisfactory results.
48 4.3. Numerical examples
Table 4.1: Weights computed with scheme I and II in (4.16) for ctitious load-
displacement space in Fig. 4.3. Values are rounded o to six digits.
Scheme I Scheme II
Weight: w
i
w
i
C
i
w
i
w
i
C
i
Load a 1.00000 10.0000 1.00000 10.0000
Load b 0.05000 150.000 0.00333 10.0000
Load c 0.02857 250.000 0.00114 10.0000
Load d 0.01818 800.000 0.00023 10.0000
Load e 0.01000 1000.00 0.00010 10.0000
The weighting factors naturally enter into the sensitivity expression as well
whereby (4.12) for multiple load cases is written as:
dC
d
e
=
N
q
q=1
w
q
T
q
R
q
e
(4.17)
The major disadvantage of solving the multiple load case problem is the increased
computational time and memory consumption caused by the need to store multiple
tangent stiness matrices. The latter issue is not a generic problem but exists in the
current implementation in MUST. For solving problems of practical engineering
interest this currently poses a severe limitation on MUST as memory usage may
easily approach the 4 GB limit on standard PC systems. This will, however, be
remedied in a later version of the software.
Following a general discussion four examples will now be presented to demonstrate
the capabilities and limitations of the methodology used.
4.3 Numerical examples
When presenting the results, load-displacement plots showing the evolution of the
topology with increasing load factor, , will be used. These curves are the result of
multiple nonlinear optimizations, solved at various loads, and thus do not represent
the load-displacement curve from a single nonlinear analysis. Consequently, each
marker (in Figs. 4.5, 4.9 and 4.16) represents the end result of a nonlinear
optimization at that particular load (a dotted line marks the response predicted
with linear theory.). The load-displacement curve in these plots therefore marks
the boundary for optimal designs, i.e. all non-optimal designs must lie beneath
that curve.
The optimal boundary curve may in practice be obtained in a number of ways. One
is to start all analyses from the same conditions, e.g. evenly distributed material
Chapter 4. Nonlinear topology optimization 49
(uniform gray) and u = 0. This strategy may lead to local optimum solutions
so at each load, alternatives should be sought out to see if a better solution can
be found. This constitutes solving for the same load several times with dierent
initial conditions, e.g. restarting from the previous or subsequent load with state
variables alone or both state variables and topology (this is illustrated in Fig. 4.11).
As such, the generation of the optimality boundary in the load-displacement space
may require a substantial amount of work. Even then, it is not guaranteed that
no better solution can be found at a particular load. This could be the case
for problems exhibiting non monotonic behavior in the load-displacement space,
e.g. snap-through behavior, since the Newton-type solvers are not well suited
for these types of problems. To provide a better understanding of the nonlinear
behavior over the entire load spectrum, the arc-length method has been used on
some optimal designs as illustrated in Figs. 4.12 and 4.14.
The rst example is a plate example, which serves as a validation example for the
present formulation since the linear topology solution is well known. The second
and third examples are for a doubly-curved shell structure and are included to
demonstrate the eects of the nonlinearities on a more general structure for single
and multiple load cases, respectively. Finally, a multilayered, curved panel of
orthotropic lamina is included.
Note that in the chapter the examples have been solved exclusively using the
SHELL9 element since the MITC4 element had not been fully reimplemented
when these results were generated. Consequently, checker-boarding does not
pose a problem but a sensitivity lter has still been employed to ensure mesh-
independency (Bendse and Sigmund, 2003).
4.3.1 Simply supported 3-layer square plate
The square plate supported at the four edges and subjected to a central point
load is a well-described problem in topology optimization. It has been included
mainly to validate the current formulation by comparing linear results to those
obtained by others. For the sake of comparison the model data have been adopted
from Lee et al. (2000). The dimensions of the plate are 2000 2000 [mm] with a
total thickness of 150 mm distributed over three layers as 37.5/75.0/37.5 [mm] all
consisting of isotropic material with E = 21 MPa and = 0.30. The skin layers
are voided and the core layer is maintained constant (scheme (c) in Fig. 4.1).
A quarter of the plate is modeled using a 20 20 mesh of SHELL9 elements. The
model is simply supported on two edges and has symmetry boundary conditions
on the opposing edges. This approach is valid in this example since the plate
does not exhibit snap-through and therefore has well-dened nonlinear behavior.
The SIMP power is initially 3.0 and is increased by 1.0 every 10 iterations until
it reaches 6.0. This causes some minor uctuations in the objective function as
shown in Fig. 4.4 but is without consequence to the results. The volume constraint
is V
c
= 0.5.
50 4.3. Numerical examples
0,0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1,0
0 10 20 30 40 50
It erat ions
N
o
r
m
a
l
i
z
e
d
o
b
j
e
c
t
i
v
e
3
4
5
6
P
o
w
e
r
Figure 4.4: History of objective function (left axis, continuous curve) and SIMP power
(right axis, step curve) for linear optimization of square plate example. Load factor is
0.027.
Figure 4.5 shows the evolution of the topology with increasing load factor, , and
two distinct shifts in topology can be observed (the gray areas). As expected
the topologies at small loads are the same for the linear and nonlinear cases, and
correspond well to those obtained by e.g. Lee et al. (2000).
With an increasing load two eects can be noted in Fig. 4.5 due to increasing
stiening from in-plane stresses; 1) the four corner reinforcements become wider
and extend further into the plate and 2) the center reinforcement is reduced in size
and instead ribs extend towards the boundaries until at maximum load, the corner
and center reinforcements become a single reinforcement. This example both
validates the formulation and gives a rst indication that the nonlinear eects are
of increasing importance as the load increases. This corresponds well to the usual
progressive nature of nonlinear eects. At light loading the optimal topologies
are identical, see Fig. 4.5, but show increasing deviation as the load increases.
However, the increase in performance in very small as indicated in Fig. 4.6, where
the topology obtained with a linear solution ( = 0.001) has been compared to the
topology obtained with a nonlinear solution at maximum load ( = 1.000). This is
achieved by performing a nonlinear analysis of the optimal topologies and tracing
the equilibrium path over the entire load spectrum using an arc-length solver.
As indicated in Fig. 4.6 the linear solution performs slightly better than the
nonlinear solution in the lower part of the spectrum and vice versa as the load
increases. This is to be expected but the dierence in performance is only
approximately 0.45%, which amounts to nothing for practical purposes. This was
not particularly encouraging so we continue to investigate the same phenomena
for shell structures.
C
h
a
p
t
e
r
4
.
N
o
n
l
i
n
e
a
r
t
o
p
o
l
o
g
y
o
p
t
i
m
i
z
a
t
i
o
n
5
1
0,0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1,0
0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1
Normalized cent er point displacement
L
o
a
d
f
a
c
t
o
r
,
k
Figure 4.5: Topological evolution of square plate with increasing center point load. The dotted line is the linear load-displacement
curve, the solid line is the equivalent nonlinear response. Both curves are for the optimized structure.
52 4.3. Numerical examples
0,0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1,0
0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0
Normalized cent er point displacement
L
o
a
d
f
a
c
t
o
r
,
Boundary of opt imal designs
= 0.001
= 1.000
k
k
k
Figure 4.6: Nonlinear response of optimized topologies for nonlinear optimization of
plate example. The curves represent the response of the optimal topologies obtained for
= 0.001 and = 1.000, respectively. The solid line indicates the boundary of optimal
designs as obtained from Fig. 4.5.
4.3.2 Hinged 4-layer spherical cap single load case
The doubly-curved shell has been chosen because its behavior is fundamentally
dierent from that of the plate. Instead of stiening with increasing load the
spherical cap becomes more compliant and eventually snaps a behavior often
encountered in shell structures. The base of the shell structure is spanned by a
1000 1000 [mm] square and the center point rises 100 mm above the base. The
four edges are prescribed by parabolas and the surface is generated by dragging
one parabola along an identical parabola. In mathematical terms the surface may
be described as:
z(x, y) = h
2h
l
2
_
x
2
+ y
2
_
(4.18)
where h is the height of the center point and l is the edge-length, here 100 [mm]
and 1000 [mm], respectively. The thickness of the shell is 24 mm divided over four
layers as 1/20/1/2 [mm] where the voided layer is the 2 mm skin at the center
of curvature side. The three skins consist of aluminum with E = 70 GPa and
= 0.3 and the core is an isotropic foam with E = 125 MPa and = 0.3. The full
geometry is shown in Fig. 4.7 with actual thickness and distribution of layers. In
this conguration the stiening constitutes a relatively small portion of the total
load carrying capacity.
Chapter 4. Nonlinear topology optimization 53
Solid: Aluminum
Solid:Foam
Solid: Aluminum
Voided: Aluminum
Figure 4.7: Geometry of spherical cap example with actual thickness and distribution
of layers. The model is subjected to a center point load and hinged along the four edges.
0,0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1,0
0 10 20 30 40 50
It erat ions
N
o
r
m
a
l
i
z
e
d
o
b
j
e
c
t
i
v
e
3
4
5
6
P
o
w
e
r
Nonlinear,
Linear
= 1 k
Figure 4.8: History of objective function (left axis, continuous curve) and SIMP power
(right axis, step curve) for nonlinear optimization of spherical cap example. Normalized
to initial compliance.
The full shell geometry is modeled using a 40 40 mesh of SHELL9 elements and
the model is hinged (u
i
= 0) on the four edge curves. The volume constraint is
V
c
= 0.5 and the SIMP power is increased from 3.0 to 6.0 in steps of 1.0 over the
rst 30 iterations as shown in Fig. 4.8. As indicated in Fig. 4.8 the optimal design
only has a moderately improved performance over the initial design (uniform grey),
which is due to the moderate thickness of the stiening layer as mentioned above.
Figure 4.9 shows the evolution of the topology with increasing load factor.
5
4
4
.
3
.
N
u
m
e
r
i
c
a
l
e
x
a
m
p
l
e
s
0,0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1,0
0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0
Normalized cent er point displacement
L
o
a
d
f
a
c
t
o
r
,
k
Figure 4.9: Topological evolution of spherical cap with increasing center point load. The dotted line is the linear load-displacement
curve, the solid line is the equivalent nonlinear response. Both curves are for the optimized structure. The gray areas mark the
transitions between topologies.
Chapter 4. Nonlinear topology optimization 55
(a) (b) (c)
Figure 4.10: Actual deformed shapes of the three basic topologies of a quarter of the
spherical cap geometry: (a) moderately nonlinear, = 0.10, (b) intermediate topology,
= 0.37, (c) nal conguration, = 1.00.
As the load increases the optimal stiening topology changes ve times as marked
by the gray zones in Fig. 4.9. The topology changes over these zones due to
changes in the load carrying mechanisms of the structure as shown for the three
major topologies in Fig. 4.10. The rst transition (from (a) to (b) in Fig. 4.10)
occurs because material is needed to reinforce the corners, which are subjected to
local bending. At the same time the center area is highly aected by the load and
must be reinforced as well. The change (from (b) to (c) in Fig. 4.10) is caused by
increasing membrane eects due to the near-at surfaces extending from the load
towards the hinged boundaries and along the symmetry planes.
It is interesting to note the major jump in topology that occurs when the structure
snaps through (the horizontal gray area in Fig. 4.9). The sudden change in
topology can be seen clearly when starting an optimization from a high load
factor (the post-snap regime, > 0.35) and solving for a smaller load factor.
This is illustrated in Fig. 4.11 where the state variables obtained at = 0.36 have
been reused as initial guess for solving for = 0.35. As can be seen, the optimizer
eectively pushes the design to the pre-snap conguration between iteration 19 and
20. This also indicates that the curve in Fig. 4.9 may be discontinuous between
these two points or, at least, have near-zero slope. This makes good sense since
the pre-snap conguration is by far more attractive in terms of compliance than
the post-snap conguration.
The increased performance of the nonlinear design over the linear solution is
modest since the stiening layer only contributes moderately to the total stiness.
However, the structure exhibits noticeably dierent nonlinear response for the
design obtained with the linear method compared to some of those obtained with
the nonlinear method. Consider the three load-displacement curves in Fig. 4.12,
generated using an arc-length solver for small, medium and heavy load ( = 0.01,
= 0.42 and = 1.00, respectively). As expected all three curves stay below
the boundary of optimal designs (obtained from Fig. 4.9) but the response of the
three designs is not the same over the entire load spectrum. At low load factors
the curves are almost indistinguishable but as the load increases the dierent
topologies show their strengths and weaknesses.
56 4.3. Numerical examples
0,0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1,0
0 25 50 75 100 125 150
It erat ions
N
o
r
m
a
l
i
z
e
d
o
b
j
e
c
t
i
v
e
3
4
5
6
P
o
w
e
r
Figure 4.11: History of objective function (left axis, continuous curve) and SIMP power
(right axis, step curve) for nonlinear optimization of spherical cap example, restarted from
state variables at = 0.36 and solved for = 0.35.
0,0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1,0
0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0
Normalized cent er point displacement
L
o
a
d
f
a
c
t
o
r
,
Boundary of opt imal designs
= 0.01
= 0.42
= 1.00
k
k
k
k
Figure 4.12: Nonlinear response of optimized topologies for nonlinear optimization
of spherical cap example. The curves represent the response of the optimal topologies
obtained for = 0.01, = 0.42 and = 1.00, respectively. The solid line indicates the
boundary of optimal designs as obtained from Fig. 4.9.
Chapter 4. Nonlinear topology optimization 57
Solid: Aluminum
Solid:Foam
Solid: Aluminum
Voided: Aluminum
Figure 4.13: Geometry of spherical cap (multiple load cases) example with actual
thickness and distribution of layers. The model is subjected to a center point load and
hinged along the four edges.
In Fig. 4.12 it is noteworthy that the result for = 0.42, although not optimal in
all congurations, has a superior ability to sustain load during snap-through. This
is another interesting aspect of utilizing nonlinear optimization for such problems.
As demonstrated, the spherical cap example shows signicant inuence of nonlin-
ear eects, particularly for very large displacements. The changes in topology can
be explained by the change in load carrying properties, which is not accounted for
with small displacement analysis. The characteristic property of the topologies for
large displacements is that they seem to be more global since the deformation in
these cases involve the majority of the structure. The result is a simpler and more
continuous reinforcement pattern, which is desirable from a manufacturing point
of view since that type of pattern will be much easier to manufacture than the
discontinuous reenforcement pattern suggested by the optimal topology obtained
with a linear solution. This will be discussed further in the last example.
4.3.3 Hinged 4-layer spherical cap multiple load cases
This example is identical to the previous one in geometry and materials but the
thickness distribution is now 1/10/1/4 [mm], where the 4 mm layer is voided as
shown in Fig. 4.13. With a thickness share of 25% the stiening layer constitutes
the majority of the total structural stiness and consequently, the distribution
of material in the stiening layer plays a decisive role for the load carrying
mechanisms of the structure. This reduces the robustness of the present method in
the sense that individual topologies will potentially exhibit very dierent nonlinear
behavior over the load spectrum. This is not an acceptable behavior for engineering
structures and choosing a single topology is thus not a viable approach. Instead,
we solve the problem for 5 and 10 dierent loads simultaneously, as described in
Section 4.2.2.
58 4.3. Numerical examples
The multiple load case optimizations are solved at values of 0.01, 0.17, 0.33,
0.67, 1.0 (5 load cases) and 0.03, 0.10, 0.17, 0.33, 0.40, 0.47, 0.53, 0.60, 0.67, 0.83
(10 load cases). For the 10 load case optimization the weights are chosen as a
scaling of the load (scheme I in (4.16)) while the 5 load case optimization is scaled
using the estimated compliance values at each load (scheme II in (4.16)). The
estimated values are obtained by running an initial analysis of the non-optimized
structure. Consequently, the weighting is not entirely uniform as indicated in
Table 4.1, page 48, but uniform to an order of magnitude, which is sucient to
avoid severe favoring of individual load cases. The choice of load factors, , is the
result of an iterative process and is used primarily to seek out dierent topologies,
testing their nonlinear response, and deciding which ones perform satisfactorily.
This procedure is somewhat unscientic and illustrates the tweaking involved in
using the weighted sum formulation.
Using the approach described above should result in the topology (or topologies)
that constitutes a good compromise between the very dierent topologies ob-
tained at distinct loads, given the criteria dened in the weighted sum formulation.
In Fig. 4.14 the obtained topologies for 5 and 10 load cases are shown together with
reference topologies obtained using a linear and nonlinear solution, respectively.
An arc-length solver has been used to determine the nonlinear response of both the
topology obtained from a linear solution, the topology obtained from a nonlinear
solution at = 1 and the topologies obtained with 5 and 10 load cases.
Note that the curves in Fig. 4.14 represent transverse center point displacement
in the thickness direction. The step in the curves for the nonlinear solution
( = 1.00) and the 5 load case solution indicates that the geometry is deforming
with components perpendicular to the transverse direction, which is a well-known
behavior of such structures. The reason that this is observed here and not in
the previous example is the reduced thickness, which reduces the stiness of the
shell in a way that allows it to snap in other directions before actually snapping
through.
Considering the graphical representation of the optimal topology obtained for 10
load cases (Fig. 4.14) it is apparent that it represents a compromise between the
linear and nonlinear topologies. Furthermore, the nonlinear response of the 10 load
case design represents a good compromise at all loads. This is a very promising
result, which indicates that the weighted sum formulation may indeed be used to
generate structures that perform well at very dierent load levels.
C
h
a
p
t
e
r
4
.
N
o
n
l
i
n
e
a
r
t
o
p
o
l
o
g
y
o
p
t
i
m
i
z
a
t
i
o
n
5
9
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0
Normalized cent er point displacement
L
o
a
d
f
a
c
t
o
r
,
k
5LC 10LC Linear k =1.00
Figure 4.14: Nonlinear response of optimized topologies for nonlinear optimization with multiple loads of spherical cap example. The
curves represent the response of the optimal topologies obtained with the linear solution and the nonlinear solution for = 1.00 as
well as the results of solving with 5 and 10 load cases (LC).
60 4.3. Numerical examples
Voided:Glass/epoxybiax(+45/-45)
Solid:Foam
Solid:Glass/epoxyUD(0)
Voided:Glass/epoxybiax(+45/-45)
Solid:Glass/epoxyUD(0)
Figure 4.15: Geometry of cylindrical shell example shown with actual thickness and
distribution of layers. Fiber angles are taken with respect to the straight edges of the
shell. The model is subjected to a center point load and hinged at the four corners.
4.3.4 Corner hinged 5-layer cylindrical shell
This example demonstrates the use of orthotropic materials in the present
formulation. The geometry is a cylindrical shell segment as shown in Fig. 4.15
where all edge lengths are 1000 mm and the curved edge spans = 1/0.86 rad
of a circle with radius 860 mm. The shell rises 141.3 mm above the base, which
measures 1000 944.6 [mm]. This is a standard geometry for shell examples
and has been used by numerous authors. The lay-up of the shell is a 5-layer
symmetric laminate [45/0/0/0/45] with a total thickness of 14 mm distributed
as 1/1/10/1/1 [mm]. The four skin layers consist of an orthotropic glass ber
reinforced epoxy composite with the following material properties: E
x
= 36 GPa,
E
y
= E
z
= 6.4 GPa, G
xy
= G
yz
= G
xz
= 2.4 GPa,
xy
=
xz
= 0.27 and
yz
= 0.33. The core is taken to be the same isotropic foam as used above with
E = 125 MPa and = 0.3. For the 45 biax layer we use averaged material
properties, which discards the bending-twist coupling that would occur if a +45
layer was stacked upon a 45 layer. The properties used are E
x
= E
y
= 6.6
GPa, E
z
= 5.8 GPa, G
xy
= 5.2 GPa, G
yz
= G
xz
= 2.2 GPa,
xy
= 0.50 and
yz
=
xz
= 0.30.
The shell is modeled using a mesh of 50 50 SHELL9 elements and the model
is supported in three directions (u
i
= 0) at the four corner nodes. This renders
the model very exible and snap-through occurs at a relatively low load factor as
shown in Fig. 4.16, which also shows the change in topology with increasing load.
In Fig. 4.16 the cross-shaped topology obtained for large displacements coincide
with the primary material directions of the biax reinforcement layer and as such
correspond well to the anticipated design.
C
h
a
p
t
e
r
4
.
N
o
n
l
i
n
e
a
r
t
o
p
o
l
o
g
y
o
p
t
i
m
i
z
a
t
i
o
n
6
1
0,0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1,0
0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0
Normalized cent er point displacement
L
o
a
d
f
a
c
t
o
r
,
k
Figure 4.16: Topological evolution of cylindrical shell with increasing center point load. The dotted line is the linear load-displacement
curve, the solid line is the equivalent nonlinear response. Both curves are for the optimized structure. The gray areas mark the
transitions between topologies.
62 4.4. Summary and conclusions
Figure 4.17: Optimal stiening topology for cylindrical shell, solved for 10 load factors
(0.01, 0.04, 0.08, 0.12, 0.20, 0.40, 0.56, 0.64, 0.80, 1.00).
As observed in the previous example the large displacement topologies are
continuous over the geometry and consequently have a better manufacturability.
This is particularly the case for laminated composite structures since bres are
most often laid out in mats covering larger areas. Making the local patterns
obtained with a linear solution involves a number of problems.
First, manufacturing is more dicult because several smaller and geometrically
more complex patches of bre mat must be cut out and handled. Second, localized
reinforcement would lead to an increased number of local stress concentrations
and incidently, an increased number of locations from which a potential crack
might spring. These aspects are very important when designing reenforcements for
laminated composite structures and should be taken into account when choosing
the optimal topology to go from.
As in the previous example the problem has also been solved for multiple load
factors to obtain a compromise design that will perform well throughout the load
spectrum. The 10 load factors considered are 0.01, 0.04, 0.08, 0.12, 0.20, 0.40,
0.56, 0.64, 0.80, 1.00, which yields the design shown in Fig. 4.17 using scheme I
from (4.16). Again, the multiple load factor design is a compromise between the
dierent topologies, which is evident when comparing Fig. 4.16 to Fig. 4.17.
4.4 Summary and conclusions
In this chapter the eect of including the full Green-Lagrange strain measure
in the topology optimization formulation for layered shell structures has been
investigated. The static problem is solved iteratively using the Newton-Raphson
method and the minimum compliance optimization problem is solved using the
adjoint variable method and the Method of Moving Asymptotes (MMA). The
design parametrization allows multiple voided layers in each element to be scaled
by a single design variable while any additional layers remain xed (solid). This
circumvents problems of near-zero terms in the stiness matrix and renders the
Chapter 4. Nonlinear topology optimization 63
problem solvable without the need for any additional numerical tricks. Multiple
load cases are handled in a weighted sum formulation and allows us to solve for
best compromise designs in a wide load spectrum.
Four numerical examples demonstrate the capabilities of the method and show
the eects of the nonlinear terms. The performance of the nonlinear designs over
the linear solutions is better in all examples (lower compliance) but the dierence
in compliance may be relatively small. This is the case in the rst spherical cap
example where the optimal distribution of an isotropic reinforcement material is
found for a single and relatively thin layer. If the thickness of the reinforcement
layer is increased substantially the dierence in compliance also increases as
demonstrated in the second spherical cap example. Here, a major dierence in
the nonlinear response for topologies obtained at low and high loading can also
be observed. Furthermore, the example shows the best compromise response
obtained when solving for 10 dierent load factors simultaneously. This indicates
that a multiple load case approach should be taken for structures operating at both
low and high loads. Finally, a cylindrical shell example shows the reinforcement
pattern of two opposite orthotropic biax layers. The result correspond well to the
anticipated result in that the reinforcement coincides with the principal material
directions of the biax.
In all the examples a signicant change in topology is observed for increasing loads.
Also, the results are consistent with the progressive nature of nonlinear eects in
that a gradual increase of loads yields a gradual change in topology (starting from
the linear and nonlinear topologies being identical and tending towards increasing
deviation of the nonlinear topologies). A common characteristic of the large
displacement topologies is that they are global in nature where the linear results
tend to involve local reinforcements. This is to be expected since an increasing
portion of the geometry becomes aected when the load increases. An interesting
observation is that the manufacturability of the large displacement results seems to
be higher since they are continuous over the geometry. For laminated composites
this is essential since application of smaller reinforcement patches would be a highly
impractical approach.
64 4.4. Summary and conclusions
5
Discrete material optimization
T
his chapter is dedicated to presenting a novel approach to material
layout and ber angle optimization, called Discrete Material Optimization
(DMO). For laminated composites the choice of material and orientation can
be dicult but it is an important topic since ecient design with laminated
composites is dependent on the engineer being able to exploit the directionality of
the material. This has historically been associated with great diculty but it is
believed that DMO can be used to solve the general material optimization problem
eciently for laminated composite shell structures. This objective is three-fold
since it requires simultaneous solution for material, orientation and stacking. As
will be shown, this can be encompassed in the same parametrization and solved
within a nite element framework. The main points presented in this chapter are
also published in Stegmann and Lund (2005a).
Displacements will in this chapter be assumed small and material behavior linear
but the proposed parametrization could easily be used for nonlinear problems as
well. For the linear case the element stiness matrix, K
e
, is dened as in (2.11)
but stated here again for reference:
K
e
=
_
V
B
T
CB dV (5.1)
which is summed over all elements and used for solving the linear static equilibrium
equations:
Ku = p (5.2)
Again, the compliance, C, is used as objective function and may for the linear case
be stated as:
C = u
T
p = u
T
Ku = 2U (5.3)
where U is the total strain energy. These equations form the basis for implementing
the proposed parametrization and will be addressed in more detail in Section 5.2.
The rest of the chapter is organized as follows. Section 5.1 discusses orientational
optimization methods to place DMO in a context, Section 5.2 introduces the
DMO parametrization, Section 5.3 takes a closer look at the element level
66 5.1. Orientation optimization with orthotropic materials
q
1
q
m
q
N
e
Figure 5.1: Illustration of the classical concept of orientation optimization in a nite
element framework. The arrows represent the 1st principal material direction (denoted
by
m
) of the orthotropic material.
parametrization and Section 5.4 states the optimization problem. In Section 5.5 a
number of 2D and 3D examples are presented to illustrate the capabilities of the
DMO method.
5.1 Orientation optimization with orthotropic materials
The classical approach to solve for optimal orientation of orthotropic materials and
minimum compliance has been to use the local orientation (denoted by local ber
angles, ) as design variables. In a nite element framework this can be depicted as
shown in Fig. 5.1, where each arrow represents the 1st principal material direction
and is uniquely dened in each element (or layer) by the angle,
m
, relative to
some xed frame. The design variables are then the continuous parameters,
m
,
and thus the method is often referred to as Continuous Fiber Angle Optimization
(CFAO). The optimization problem may be stated as:
Objective : min
C() = p
T
u
Subject to : Ku = p
min
max
_
_
(5.4)
where
min
and
max
contain the lower and upper bounds on the design variables,
respectively (typically 90
and +90
m
u
e
m
(5.5)
This expression is valid if
m
only pertains to a single element. Later we will
introduce so-called patch variables which aect several elements and in such cases
Chapter 5. Discrete material optimization 67
x
y
y
x
1
(a) (b)
P
P
P
P
q
2
q
2
5
10
Figure 5.2: Example geometry and boundary conditions (a) and optimum solution (b).
the sensitivity dC/d
m
must be evaluated as a sum over all elements aected
by the variable. The expression in (5.5) is common to both CFAO and topology
optimization and is in practice quite ecient to evaluate either analytically or more
commonly for CFAO by a nite dierence approximation for general geometries.
The major diculty faced when using this type of formulation is that the global
design space becomes non-convex. To illustrate this the simple problem in Fig. 5.2
has been constructed and solved using a continuous formulation and the Method
of Moving Asymptotes (MMA) by Svanberg (1987). The result is the design space
shown in Fig. 5.3(a), where four extrema can be found one global and three local
(see also Table 5.1). To obtain the global optimum solution the initial guess must
be within the non-shaded part of the design space (the safe domain) as illustrated
in Fig. 5.3(b). In a simple case like this it is fairly easy to determine how to start
the optimization but for complicated geometries with multiple layers it becomes
almost impossible to state the problem a priori within the safe domain. It follows
that for generic problems the results obtained in this manner with CFAO will often
be sub-optimal albeit better than the initial design.
Table 5.1: Extremum values for the design space in Fig. 5.3.
Optimum: Global Local #1 Local #2 Local #3
Element 1 24.2
29.3
90.0
90.0
Element 2 41.6
90.0
90.0
47.3
This is, of course, not a new realization and several methods have already been
proposed to circumvent the problem of local optimum solutions. These methods
roughly fall within one of four categories:
1. Analytical methods, which rely on the closed-form formulation of an
optimality criterion as described by Prager (1970). This constitutes a
68 5.1. Orientation optimization with orthotropic materials
-90 -60 -30 0 30 60 90
-90
-60
-30
0
30
60
90
Element 2
E
l
e
m
e
n
t
1
1,72-1,80
1,64-1,72
1,56-1,64
1,48-1,56
1,40-1,48
1,32-1,40
1,24-1,32
1,16-1,24
1,08-1,16
1,00-1,08
-90 -60 -30 0 30 60 90
-90
-60
-30
0
30
60
90
Element 2
E
l
e
m
e
n
t
1
1,72-1,80
1,64-1,72
1,56-1,64
1,48-1,56
1,40-1,48
1,32-1,40
1,24-1,32
1,16-1,24
1,08-1,16
1,00-1,08
-90
-60
-30
0
30
60
90
-90
-60
-30
0
30
60
90
1,00
1,08
1,16
1,24
1,32
1,40
1,48
1,56
1,64
1,72
1,80
Normalizedocjective
(compliance)
Element 2
Element 1
(a) (b)
Optimum
Figure 5.3: Normalized objective function for CFAO test example. The red dots mark
the local extrema and the shaded area marks values that will lead to local optimum
solution.
limitation in the applicability of the methods but for simpler geometries the
methods have been applied very successfully and often serve as benchmark
solutions for purely numerical methods. The major contributor to optimality
criterion methods in orientation design is Pedersen (1991) who has used
the method to optimize 2D continua, beams and plates. Luo and Gea
(1998) used a similar approach to solve for optimal orientation of plate bead
stieners.
2. Mathematical programming techniques are purely numerical and are con-
cerned with tuning the optimizer itself rather than reformulating the problem
or change the parametrization. Among others Bruyneel and Fleury (2002)
and Moita et al. (2000) have applied such methodology with success. Still,
these methods do not ensure convergence to the globally optimum solution.
3. Parametrization methods start from the realization that the problem of
local optimum solutions is inherent in the CFAO method. The goal of
the methods is to change the parametrization and obtain a convex design
space. Such methods include the Lamination Parameter Method introduced
by Tsai and Pagano (1968) and used by e.g. Miki and Sugiyama (1993) for
orientation optimization of plates and Foldager et al. (2001) for plates and
cylindrical shells. The method requires closed-form analytical formulation of
the feasible domain of the lamination parameters which has so far only been
achieved for relatively simple geometries.
4. Evolutionary techniques, attributed to Holland (1975), are fundamentally
search methods but employ ideas derived from genetics and Darwins sur-
vival of the ttest principle to more eectively select the best solution from
Chapter 5. Discrete material optimization 69
a population of solutions. Genetic Algorithms (GAs) are particularly useful
for discrete problems and problems where the sensitivities are impossible
or extremely dicult to compute. If the population is suciently large
the method reduces the risk of obtaining a local optimum solution but still
cannot guarantee convergence to the global optimum. The major problem
faced with GAs is the amount of computational eort involved when solving
problems with a large number of design variables, particularly if the analysis
itself is computationally expensive. GAs have been successfully applied to
e.g. stacking sequence optimization by Riche and Haftka (1993), Adali et al.
(1995) and others.
The work-horse of engineering design in industry today is the mathematical
programming techniques, which have been integrated in e.g. the commercial codes
BOSS QUATTRO from SAMTECH and OPTISCTRUCT from ALTAIR. Such
software cannot guarantee that the obtained solution is the global optimum but
is still a valuable design tool since it can provide engineers with a signicant
performance increase. The goal of the present work is to provide a novel method
that obtains a close approximation to the global optimum solution while being
applicable to problems of industrial interest. To this end a new parametrization
is proposed.
5.2 The discrete material optimization method
The basic idea in the discrete material optimization (DMO) parametrization is
essentially an extension of the ideas used in structural topology optimization
but instead of choosing between solid and void we want to choose between
any distinct number of materials. This methodology can be stated as: for all
elements in the structure nd one distinct material from a set of pre-dened
candidate materials such that the objective function is minimized. Metaphorically
this is like painting a picture from a palette of colors the optimal picture is
the one having the best combination of colors over the canvas (although some
artists might object to this crude portrayal of their trade). The idea of using
topology optimization for material selection in this way was rst introduced for
two and three candidate materials (phases) as multiphase topology optimization
by Sigmund and co-workers. The rst applications of the three-phase method
was that of Sigmund and Torquato (1997) who used it for designing materials
with extreme thermal expansion properties and later by Gibiansky and Sigmund
(2000) for designing materials with extreme bulk modules. The method has also
been extended to actuator design, Sigmund (2001), which is closely related to
compliant mechanism design, Sigmund (1997). Recently Wang and Wang (2004)
introduced a level-set method based on the same ideas as an alternative to the
approach of Sigmund and co-workers. The level-set method is applied with up to
four phases to 2D structural problems and shows promising results.
Common to the parametrization used in the works stated above is that materials
70 5.2. The discrete material optimization method
are assumed to be isotropic, the elements are single layered and the maximum
number of phases involved is four three distinct materials and void. Furthermore,
the structures considered have been either 2D continua or plates. The DMO
formulation extends the scope of application by considering multiple phases and
3D structures and furthermore by allowing the structures to be multilayered and
of orthotropic materials. However, the method is still limited to operate on a xed
design domain, i.e. thicknesses and shape are not changed during optimization.
In the context of orientation optimization, dierent materials simply mean the
same material oriented at various angles in space (ber angles) but in general,
it might as well mean Carbon or Glass Fiber Reinforced Plastic (CFRP/GFRP),
polymer foam, steel, aluminum or any other material at any orientation. As such,
the proposed formulation is very versatile and can be used to optimize the material
constitution of structures in general and composite structures in particular.
5.2.1 The methodology
As in topology optimization the parametrization of the DMO formulation is
invoked at the nite element level. The element constitutive matrix, C
e
, is
expressed as a weighted sum of candidate materials, each characterized by a
constitutive matrix, C
i
. In general, this may be expressed as a sum over the
element number of candidate materials, n
e
:
C
e
=
n
e
i=1
w
i
C
i
= w
1
C
1
+ w
2
C
2
+ + w
n
eC
n
e, 0 w
i
1 (5.6)
It follows that the number of candidate materials is also the number of element
design variables and if N
e
is the number of elements, the total number of design
variables, N
dv
, for single layered structures is N
dv
=
N
e
i=1
n
e
i
. Note that
the classical topology optimization formulation having one design variable per
element is obtained by setting n
e
= 1 in (5.6).
The weights, w
i
, in (5.6) must have values between 0 and 1 as no matrix can
contribute more than the physical material properties and a negative contribution
is physically meaningless. In this way, as in classical topology optimization,
the weights on the constitutive matrices become switches that turn on and o
stiness contributions such that the objective is minimized and a distinct choice of
candidate material is made. This underlines that the DMO method relies heavily
on the ability of the optimizer to push all weights to the limit values. Any element
having intermediate values of the weights must be regarded as undened since
the constitutive properties are non-physical. For the same reason any element
having more than a single weight of value 1 must be considered undened as well.
Consequently, the single most important requirement for the DMO method is that
every element must have one single weight of value 1 and all other weights of value
0. Failing to comply with this essentially renders the results meaningless. It follows
that the choice of weighting functions, w
i
, is very important for the performance of
Chapter 5. Discrete material optimization 71
-90 -60 -30 0 30 60 90
-90
-60
-30
0
30
60
90
Element 2
E
l
e
m
e
n
t
1
1,72-1,80
1,64-1,72
1,56-1,64
1,48-1,56
1,40-1,48
1,32-1,40
1,24-1,32
1,16-1,24
1,08-1,16
1,00-1,08
y
x
-41.6
(b) (a)
24.2
q
i
q
i
Figure 5.4: Normalized objective function for DMO test example. (a) The candidate
materials at 12 angles,
i
. (b) The yellow dots mark possible combinations of candidate
materials and the large and small white dots mark optimum solutions obtained with
DMO and CFAO, respectively.
the DMO method and several formulations have been developed and evaluated (as
described shortly). The initial values of the design variables, x
i
, may in principle
be any set of numbers between 0 and 1 but in general the values should be chosen
such that the initial weighting is uniform, i.e. w
i
= w
j
for all i, j = 1 . . . n
e
. This
provides the most fair starting guess since no materials are favored a priori.
To illustrate the methodology for ber angle optimization we solve the example
in Fig. 5.2 using DMO
1
with the same orthotropic material oriented at 12
dierent angles (0
, 15
, 30
, 45
, 60
, 75
, 90
/45
is marked by the
large white dot. This solution is the best t to the global optimum solution
24.2
/41.6
i=1
(x
e
i
)
p
. .
w
i
C
i
= (x
e
1
)
p
C
1
+(x
e
2
)
p
C
2
+ +(x
e
n
e)
p
C
n
e, 0 x
i
1 (5.7)
In this formulation each design variable scales only one constitutive matrix and has
no inuence on any of the other matrices. To push the design variables towards
0 and 1 the SIMP method has been adopted by introducing the power, p, as a
penalization of intermediate values of x
i
(see e.g. Bendse and Sigmund (2003)
for details). The method in (5.7) is not very ecient as it fails to push the design
variables to the limit values for all the cases tested.
5.3.2 DMO scheme 2
The problems with scheme 1 were also realized by Sigmund and co-workers
(Gibiansky and Sigmund, 2000; Sigmund and Torquato, 1997) who instead used
the following formulation for two distinct materials (three phases), which has been
adopted in MUST:
C
e
= (x
e
0
)
p
_
[1 (x
e
1
)
p
]C
1
+ (x
e
1
)
p
C
2
_
, 0 x
i
1 (5.8)
This formulation is fundamentally dierent from scheme 1 in that x
0
scales the
entire contribution to C
e
while x
1
slides between C
1
and C
2
. As such, the
formulation encompasses simultaneous topology optimization (through x
0
) and
multiple material optimization (through x
1
). The major dierence between this
formulation and (5.7) is the term (1 x
e
i
), which links a single design variable to
more than one material. The benet of that is that adding weight in one place
automatically reduces weight in others, thus helping to push the weights towards
0 and 1. For pure material selection design (such as ber angle optimization) the
x
0
variable can be left out.
As can be seen from (5.8) the SIMP methodology is still encompassed in this
formulation. However, several implementations of the power, p, have been
Chapter 5. Discrete material optimization 73
p = 1 p = 3 p = 15
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
x1
x2
x1
x2
x1
x2
x2 x2 x2
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
x1 x1 x1
w
2
w
2
w
2
w
1
w
1
w
1
Figure 5.5: Weight functions, w
1
and w
2
, for two materials, computed with DMO
scheme 3. Top and bottom row represent w
1
and w
2
, respectively.
suggested but here we follow the strategy suggested by Gibiansky and Sigmund
(2000). The formulation in (5.8) is not very general since it can only handle two
materials, so the scheme has been adapted to multiple materials in MUST.
5.3.3 DMO scheme 3
This scheme is a simple extension of scheme 2 that can encompass any number of
materials by adding terms to (5.8), e.g. for three distinct materials (four phases):
C
e
= (x
e
0
)
p
_
[1 (x
e
1
)
p
]C
1
+ (x
e
1
)
p
_
[1 (x
e
2
)
p
]C
2
+ (x
e
2
)
p
C
3
__
= (x
e
0
)
p
_
[1 (x
e
1
)
p
]
. .
w
1
C
1
+ (x
e
1
)
p
[1 (x
e
2
)
p
]
. .
w
2
C
2
+ (x
e
1
)
p
(x
e
2
)
p
. .
w
3
C
3
_
(5.9)
where the limits 0 x
i
1 have been excluded for brevity. The expression in (5.9)
becomes tedious to write out for larger number of variables but can be generalized
for any number of candidate materials:
C
e
= (x
e
0
)
p
n
e
1
i=1
_
[1 (x
e
i
)]
p
i1
j=1
(x
e
j
)
p
_
. .
w
i
C
i
+
n
e
1
j=1
(x
e
j
)
p
. .
w
n
e
C
n
e (5.10)
where we have that
w
i
= 1. To accommodate ber angle optimization the
scheme can in MUST be used either with or without the topology optimization
variable, x
0
. DMO scheme 3 has proven ecient for up to three phases but when
the number of phases is greater the formulation tends to get stuck in local optima.
This is particularly the case when the remaining choice is between two almost
74 5.3. Element level parametrization
p q = 3 = 3 , p q = 3 = 9 , p q = 3 = 15 ,
x1 x1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
x1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
x2 x2 x2
x2 x2 x2
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
x1 x1 x1
w
2
w
2
w
2
w
2
w
2
w
2
p q = 1 = 1 , p q = 1 = 3 , p q = 1 = 15 ,
Figure 5.6: Weight function, w
2
, for two materials, computed with DMO scheme 3. Top
and bottom row represent w
2
with q-penalization and w
2
with combined pq-penalization,
respectively.
p = 1 = 3 ,q p q = 3 = 3 ,
x1
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
x2
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
x2
x1
Figure 5.7: Sum of weight functions, w
1
and w
2
, for two materials, computed with
DMO scheme 3.
equally good materials. This can be explained by considering a plot of the weights
for two materials as shown in Fig. 5.5. Since the two weights add up to 1 any spot
in the design space is equally good and the optimizer cannot gure out where to
go since nothing forces it to choose. This behavior has been observed for elements
in pure shear where the same orthotropic material at 45
is equally good. To
counter this a modied SIMP scheme has been tried in which a second power, q, is
introduced on the weights as [1 (x
e
i
)
p
]
q
and (x
e
j
)
pq
. This changes the behavior
of the weight as illustrated for the second weight, w
2
, in Fig. 5.6 while the rst
weight, w
1
, is unaected. Using the pq-power scheme the weights no longer add
up to one and the middle region of the space becomes unfavorable as shown in
Fig. 5.7. This improves the capabilities of the method in some cases but requires
some tweaking of the power q to converge to the optimum solution. Consequently,
interpolation scheme 3 in (5.10) is not generally employed.
Chapter 5. Discrete material optimization 75
x1 x1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
x1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
x1 x1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
x2 x2 x2
x2 x2 x2
x1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
p = 1 p = 3 p = 15
w
2
w
2
w
2
w
1
w
1
w
1
Figure 5.8: Weight functions, w
1
and w
2
, for two materials, computed with DMO
scheme 4. Top and bottom row represent w
1
and w
2
, respectively.
5.3.4 DMO scheme 4
As an alternative to scheme 3 above the following scheme has been developed,
which is essentially an extension of scheme 1 in (5.7). The dierence is that each
design variable now aects all weights as:
C
e
=
n
e
i=1
_
(x
e
i
)
p
n
e
j=1;j=i
_
1 (x
e
j
)
p
_
. .
w
i
C
i
(5.11)
The dierence from (5.7) is the term (1x
e
j
), which is included so that an increase
in x
i
automatically involves a decrease in all other weights. The behavior of the
interpolation for two materials is illustrated in Fig. 5.8.
The introduction of the (1 x
e
j
) term helps drive the design towards 0/1 and the
method has proven quite eective for the problems tested. The dierence between
(5.10) and (5.11) becomes clearer when writing out the expression for e.g three
materials (phases) and comparing to (5.9):
C
e
= (x
e
1
)
p
[1 (x
e
2
)
p
][1 (x
e
3
)
p
]
. .
w
1
C
1
+ (x
e
2
)
p
[1 (x
e
1
)
p
][1 (x
e
3
)
p
]
. .
w
2
C
2
+ (x
e
3
)
p
[1 (x
e
1
)
p
][1 (x
e
2
)
p
]
. .
w
3
C
3
(5.12)
The interpolation is quite simple to implement in a general way from (5.11) and
the weight factors can be computed eciently for both analysis and sensitivity
analysis as illustrated in Algorithm 5.1.
76 5.3. Element level parametrization
Algorithm 5.1: Pseudo code for computing weight factors with DMO scheme 4.
w
i
= 1 for all i = 1 . . . n
e
Initialization
if (analysis) then
for i = 1 to n
e
do
w
i
= x
p
i
for j = 1 to n
e
do
if (i ,= j) w
j
= w
j
(1 x
p
i
)
end for
end for
else if (sensitivity analysis) then
for i = 1 to n
e
do
if (k = i) then k is the active variable, i.e. w
i
/x
k
w
i
= p x
p1
k
From x
p
k
term
else
w
i
= p x
p1
k
From (1 x
p
k
) term
w
i
= w
i
x
p
i
Constant term
end if
for j = 1 to n
e
do
if ((i ,= j) and (i ,= k)) w
j
= w
j
(1 x
p
i
)
end for
end for
end if
The structure of Algorithm 5.1 is identical for the other interpolation schemes in
MUST and consequently, pseudo codes for these have not been included.
The disadvantage of DMO scheme 4 in (5.11) is that the weighting functions in
general do not add up to unity, i.e.
w
i
,= 1, which is illustrated in Fig. 5.9.
Consequently, the element stiness will be unrealistically low initially and only
slowly (due to move limits on x
i
) be pushed towards the physical stiness. When
the problem has converged, and the design variables have been pushed to 0 and 1,
the sum of the weights is 1 as indicated in Fig. 5.9.
If the convergence in compliance of the problem considered is monotonic then the
initially low stiness poses no problem but increases the number of iterations to
reach optimum and thus increases the computational cost. For problems relying
on a realistic stiness, however, the low initial stiness presents a major problem
and can lead to erroneous results. This has been observed when optimizing for
frequency response where the low stiness regions give rise to very large variations
in the ratio between stiness and mass (k/m), which ultimately leads to numerical
problems and renders the analysis unsolvable. Furthermore, evaluating the mass
constraint or other physical constraints from (5.11) is meaningless due to the
articially small scaling. To circumvent these problems a scaled version of (5.11)
has been formulated.
Chapter 5. Discrete material optimization 77
p = 1 p = 3
x1
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
x2
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
x2
x1
0
0.2
0.4
0.6
0.8
1
Figure 5.9: Sum of weight function, w
1
and w
2
, for two materials, computed with DMO
scheme 4.
5.3.5 DMO scheme 5
DMO scheme 5 is an extension of scheme 4 and introduces a common denominator
on all weights, equal to the sum of the weights. This ensures that the sum of the
weights is always one, i.e.
w
i
= 1, and for the general case scheme 5 is written
as:
C
e
=
n
e
i=1
w
i
n
e
k=1
w
k
. .
w
i
C
i
where w
i
= (x
e
i
)
p
n
e
j=1;j=i
_
1 (x
e
j
)
p
(5.13)
This formulation gives faster convergence to a near optimum compliance value
but it cannot converge fully since it is less eective in driving the weights to
0/1. The reason is that the scaling to unity alters the eect of penalizing the
design variables as shown in Fig. 5.10 for two materials. Essentially the scaling
increases the number of favorable combinations of design variables and thus makes
the interpolation less distinct. Incidently, increasing the penalization will not
help much as this only increases the size of the at triangular plateau in the
design variable space, as shown in Fig. 5.10. However, scheme 5 is still very useful
for two reasons. First, it can be used to compute the physical constraints in a
combined interpolation scheme where scheme 4 is used for stiness interpolation.
This strategy has been applied successfully to a number of examples. Second,
scheme 5 can be used as is for problems requiring so, e.g. frequency optimization,
and the inability of the scheme to fully push all the design variables to their limit
values can be accepted as a necessary trade-o. In fact, this issue seems to pose
less of a problem than expected. For most tested examples scheme 5 is able to
push the design variables to their limit values in 65 75% of all elements. In the
remaining elements the solutions gets stuck on the triangular plateau in Fig. 5.10
but often a material has been selected with e.g. w
1
= 0.7 and w
2
= 0.3, which
gives strong indication that material 1 should be chosen. Consequently, a manual
selection subroutine has been implemented in MUST, which can a posteriori push
the largest weight to 1 and the remaining weights to 0 to complete the material
selection. This has not been employed on any of the examples presented in this
work but could be used (and have been tested) for practical applications.
78 5.3. Element level parametrization
x1
x1
x1
x1
x1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
x1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
x2
x2
x2
x2
x2
x2
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
p = 1 p = 3 p = 15
w
2
w
2
w
2
w
1
w
1
w
1
Figure 5.10: Weight functions, w
1
and w
2
, for two materials, computed with DMO
scheme 5. Top and bottom row represent w
1
and w
2
, respectively.
5.3.6 DMO schemes 6 and 7
The two last interpolation schemes, 6 and 7, are identical to schemes 4 and 5,
respectively, with the dierence that they accommodate combined material and
topology optimization by reintroducing the variable x
0
as:
C
e
= x
0
n
e
i=1
_
(x
e
i
)
p
n
e
j=1;j=i
_
1 (x
e
j
)
p
_
. .
w
i
C
i
(5.14)
from scheme 4 (5.11) and complementary to this we have the scaled version:
C
e
= x
0
n
e
i=1
w
i
n
e
k=1
w
k
. .
w
i
C
i
where w
i
= (x
e
i
)
p
n
e
j=1;j=i
_
1 (x
e
j
)
p
(5.15)
from scheme 5 (5.13). These two schemes have been implemented recently and
therefore not tested extensively yet, but this will be part of our future work.
5.3.7 Multi layered structures
For multi layered structures the interpolation method described in the previous
section can be applied directly. The only dierence is that the interpolation must
be invoked layer-wise instead of element-wise, i.e. for all layers in all elements.
Consequently, the interpolation scheme is written by layer, her illustrated for DMO
Chapter 5. Discrete material optimization 79
1
(a) (b)
1
1 1
2 2
2 2
3 3
3 3
4 4
4 4
1
1
2
2
3
3
4
4
Figure 5.11: Four patch variables (14) used to reduce the total number of design
variables by collecting elements over the structure (a) or layers in an element (b).
scheme 4 from (5.11), as:
C
l
=
n
l
i=1
_
(x
l
i
)
p
n
l
j=1;j=i
_
1 (x
l
j
)
p
_
. .
w
i
C
i
(5.16)
where l again denotes layer and thus, n
l
is the number of candidate materials
for the layer.
The number of element design variables, n
e
, for multi layered elements is then
the sum of the number of design variables per layer, n
l
, over all layers, N
l
, i.e.
n
e
=
k=1
N
l
n
l
k
. As before, the total number of design variables in the problem is
N
dv
=
N
e
i=1
n
e
i
, which for multilayered structures implies a signicant increase in
the total number of design variables. To counter this, patches of design variables
are introduced.
5.3.8 Patch design variables
Collecting design variables in patches reduces the number of total design variables
by merging several design variables from dierent layers and elements into a single
variable a patch design variable. The idea springs from the manufacturing process
of laminated composites where ber mats covering larger areas are often used. A
single variable could then govern the orientation of the ber mat even though it
covers several elements, Fig. 5.11(a). Of course, the layout of the patches is left up
to the engineer a priori and consequently, the nal result will also be dependent
on the initial patch layout. Patches of design variables may also be used to enforce
laminate symmetry by assigning the same design variable to opposite laminae, see
Fig. 5.11(b). This is a very convenient feature for many practical applications
where laminate coupling eects may be unwanted.
The total number of design variables, N
dv
, when using patch design variables
80 5.4. The optimization problem
is N
dv
=
N
p
i=1
n
p
i
where n
p
is the number of candidate materials for the patch
variable and N
p
is the number of patches. This can provide a signicant reduction
in the number of design variables but at the same time requires some extra eort
and insight on the part of the engineer.
5.4 The optimization problem
The discrete material optimization problem can be stated in the same way as a
topology optimization problem, subject to an optional constraint on total mass:
Objective : min
x
C(x) = p
T
u
Subject to : (m m
c
)
0 x
min
x 1
Ku = p
_
_
(5.17)
where m is the mass of the structure and m
c
is the allowable mass, analogous to
the constraint on volume in topology optimization, (4.2). The mass constraint is
compared to the weighted mass of all, n
l
, candidate materials in all layers, N
l
,
over all elements, N
e
, i.e.:
m =
N
e
e=1
N
l
n=1
n
l
i=1
(w
i
i
V
i
)
l
e
(5.18)
This constraint is not active when doing pure ber angle optimization since a
change in ber angle involves no change in mass and consequently, the constraint
may be left out entirely (which in MMA is achieved by setting the constraint
bound very high). When doing multi material optimization the mass constraint is
important since it eectively determines the amount of light material in the nal
structure. Note that in (5.18) the density,
i
, is the physical material density in
[kg/m
3
] of material i and not a scaling as was the case in Chapter 4.
5.4.1 Design sensitivity analysis
Obtaining the gradients of the objective and constraints follows the methodology
of Section 4.2. The compliance sensitivity is again written for the ith design
variable as:
dC
dx
i
= p
T
du
dx
i
=
d
dx
i
_
u
T
Ku
_
(5.19)
and rearranging the derivative of (5.2) it can be shown as in (5.5) that:
dC
dx
i
= (u
e
m
)
T
K
e
m
x
i
u
e
m
(5.20)
which again indicates that the sensitivity can be computed on the element level,
thus reducing computational time. However, the expression in (5.20) can be
Chapter 5. Discrete material optimization 81
computed more eciently by employing the residual (2.4) as:
dC
dx
i
= (u
e
m
)
T
R
e
m
x
i
(5.21)
This expression has been implemented and provides an increased performance of
more than 300%, depending on the number of stiness matrix evaluations saved.
For patch design variables the same expressions apply by introducing a summation
over the number of elements in the patch, N
ep
:
dC
dx
i
=
N
ep
j=1
(u
e
j
)
T
R
e
j
x
i
(5.22)
Obtaining the derivative of the residual (or the stiness matrix) consists of
computing the derivative of the internal force vector (2.3) since the forces are
assumed independent of the design variables. In turn, the gradient is the derivative
of the weighting functions introduced earlier in Section 5.3. These are relatively
simple polynomial expressions that allow the sensitivity analysis to be implemented
analytically in a general and simple way. The weighting function derivative is also
used to compute the constraint function sensitivities directly from (5.18).
5.4.2 DMO convergence
To determine whether the optimization has converged to a satisfactory result, i.e.
a single candidate material has been chosen in all elements and all other materials
have been discarded, a DMO convergence measure is dened. For each element
the following inequality is evaluated for all weight factors, w
i
:
w
i
_
w
2
1
+ w
2
2
+ + w
2
n
e (5.23)
where is a tolerance level, typically 95-99.5%. If the inequality (5.23) is satised
for any w
i
in the element it is agged as converged. The DMO convergence, h
, is
then the ratio of converged elements to the total number of elements:
h
=
N
e
c
N
e
(5.24)
The DMO convergence is denoted h
99.5
if the tolerance level is 99.5% (and so forth)
and full convergence, i.e. h
99.5
= 1, simply means that all elements have a single
weight contributing more than 99.5% to the Euclidian norm of the weight factors.
This provides a good measure of the convergence although it is not a rigorous
mathematical denition. For multilayered structures the DMO convergence is
simply computed layer-by-layer instead of element-by-element, i.e. h
= N
l
c
/N
l
.
5.4.3 Explicit penalization
As discussed in Section 5.3, some of the DMO parametrization schemes (3 and 5 in
particular) often have diculty pushing the design variables to their limit values.
82 5.5. Numerical examples
In an attempt to improve the performance of these schemes a penalty function
method has been implemented, which explicitly penalizes the objective as:
C(x) = p
T
u + w
N
dv
i=1
x
p
i
(1 x
p
i
) (5.25)
where w is a scaling parameter introduced to ensure that the penalty term does not
dominate the problem. The value of w can be set in various ways e.g. according to
the compliance value, the largest sensitivity or some other measure. The penalty
function is designed to automatically decrease as the design variables tend towards
0/1 by encompassing the term (1 x
p
i
).
An elaborate framework has been implemented in which the penalty function can
be introduced either depending on the iteration number or the level of convergence
reached, i.e. when h
, 75
, 60
, 45
, 30
, 15
, 0
].
Chapter 5. Discrete material optimization 83
Figure 5.12: Optimal ber angle distribution in cantilever beam with uniformly
distributed top load.
Figure 5.13: Optimal ber angle distribution in cantilever beam with uniformly
distributed top load. Solved using 48 patches of 4 4 elements.
The optimization converges monotonically to full DMO convergence (h
99.5
= 1.0)
in 157 iterations taking just under 7 minutes on a desktop PC. The optimal ber
angle distribution determined is shown in Fig. 5.12 and agrees very well with the
results obtained by e.g. Pedersen (1991).
To illustrate the patch variable methodology the problem has been solved using
48 patches of 4 4 elements, which reduces the number of design variables to 576
and reduces the runtime by approximately 16%. The resulting optimal ber angle
distribution is shown in Fig. 5.13.
5.5.2 Beam subjected to four-point bending
This example demonstrates the ability of the DMO method to simultaneously
choose material type and material orientation. The domain is as dened for the
previous example and a mesh of 768 SHELL9 elements is used (again, all out-
of-plane displacements are xed). As candidate materials the same orthotropic
glass/epoxy as above is taken at [90
, 45
, 0
, 45
, 0
, 45
, and 90
.
Furthermore, only the interior layers are allowed to consist of foam material since
foam is not a realistic choice for the skin layers. Consequently, the distribution
of design variables for the 16 layers is [4, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 4], which
brings the number of design variables per element to 78 and thus the total number
of design variables is 6006 for the patch variable model and 748800 for the full
Chapter 5. Discrete material optimization 91
0
10
20
30
40
50
60
0 10 20 30 40 50
Iterations
N
o
r
m
a
l
i
z
e
d
o
b
j
e
c
t
i
v
e
0,0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1,0
D
M
O
c
o
n
v
e
r
g
e
n
c
e
(
9
9
.
5
%
)
Object ive (pat ch)
Object ive (full)
DMO convergence (pat ch)
DMO convergence (full)
Figure 5.24: Convergence of objective function (left axis) and DMO convergence ratio
(right axis) for wind turbine main spar example with patched and full variables.
variable model.
Both models were solved on a desktop PC. The patch variable model used 131
iterations to reach h
99.5
= 0.975 taking approximately 33 hours of computational
time while the full variable model used just 81 iterations to reach h
99.5
= 0.997
in 21 hours. The vast majority of time, around 98%, is spent on the design
sensitivity analysis and increasing this performance is going to be addressed in
future work. Again the MMA algorithm proves very ecient and solves the
approximate subproblem of the full variable model in just under 4 seconds. The
convergence is shown for the rst 50 iterations in Fig. 5.24 where the subsequent
iterations have been left out for clarity. The objective of the full variable model is
approximately 5% lower than for the patch variable model.
The result of the optimization using patch variables is shown for all 16 layers in
Figs. 5.26, 5.27, 5.28 and 5.29 4 layers at a time. Similarly, the result of the
optimization using full variables is shown in Figs. 5.30, 5.31, 5.32 and 5.33. The
optimized material directions for the glass ber are shown for all 16 layers, and
if no material direction is indicated it is implied that soft core material has been
chosen for that particular layer of a given patch/element. Note that layer 1 is the
inner (bottom) layer and layer 16 the outer (top) layer.
The two models yield very similar results and, as expected, most of the foam
core material has been put in the webs in the internal layers 215 close to the
root of the main spar in order to increase the bending stiness where the bending
moment is strong. Looking at the two models in detail it is apparent that the
full variable model chooses to use foam material in all internal layer of the web,
distributed continuously over approximately the rst third of the main spar. In
contrast the patch variable model only places foam in layers 214, and distributes
92 5.5. Numerical examples
(a) (b)
Figure 5.25: Optimal material directions for the full variable main spar model. Shown
at the tip for layer 1 (a) and layer 8 (b).
it in two distinct regions rather than continuously. This illustrates that choice of
material and orientation for each patch is a compromise between all elements in
the patch while in full variable model, the properties of each element can be set
individually. The orientation of the glass/epoxy material also corresponds well to
what was expected from the basic considerations made earlier regarding the load
carrying mechanisms of the spar. In the anges all layers are dominated by 0
orientation, i.e. along the length of the spar, to account for bending while the
webs are dominated by 45