Shape optimization of shell structures

Ekkehard Ramm* - Kai-Uwe Bletzinger** -Reiner Reitinger***

*Professor, Dr. Ing., Institutfii.r Baustatik,

University of Stuttgart, Germany
**Dr. Ing. Senior Research Associate, Institutfii.r Baustatik,
University of Stuttgart, Germany
***Dip/. Jng., Research Assistant, Institut fur Baustatik,
University of Stuttgart, Germany

'The paper was originally published in the bulletin of the International Association for Shell and Spatial
Structures, Vol. 34 (1993) n° 2"

ABSTRACF. Shells are known to be optimal in many ways, provided certain basic shell oriented design
rules are followed. The shape, thickness and material distribution play a dominant role. Minimum
material, a specific frequency response, maximum load carrying capacity, a pure membrane stress
state are typical design objectives. In the present contribution the form finding and thickness
variation are embedded in the concept of structural optimization which combines design modelling,
structural and sensitivity analyses and mathematical optimization schemes to a general design tool.
The structural response may be based on linear elastic, eigenvalue and geometrically nonlinear
analyses. In particular, the imperfection sensitivity with respect to buckling is discussed. A few
selected examples demonstrate the versatility of optimization schemes in shell design, among these
are the tuning of a bell and the form finding of a classical reinforced concrete dome shell.
RESUME. Les coques sont reconnues comme etant des structures optimales sous differents aspects
dans la mesure ou certaines regles de base de conception sont respectees. La forme, l'epaisseur etla
repartition du materiau constituant Ia coque jouent un role preponderant. Les criteres d'optimisation
typiques sont la minimisation du volume de materiau utilise, une reponse frequentielle specifique,
une capacite portante maximum et un comportement en membrane pure. Cet article concerne
/'optimisation de forme et d'epaisseur en utilisant un outil general de conception qui combine
/'analyse des structures, le calcul des sensibilites et les algorithmes d'optimisation. L'etude de Ia
structure peut etre baste sur des analyses liniaires elastiques, non liniaires geometriques ou pour
les vibrations libres. En particulier, nous discutons des effets des imperfections sur Ia stabilite.
Quelques exemples demontrent le caractere general des schimas d'optimisation pour Ia conception
des coques. Parmi ces exemples, nous presentons l'itude d'une cloche et d'un dome classique en
bitonarmi.
KEYWORDS : shells, shape optimization,form finding, sensitivity analysis.
MOTS-cLts : coques, optimisation de forme, analyse de sensibilite, analyses par ilements finis,
paramitrisation.

Revue europeenne des elements finis. Volume 2- n° 3/1993, pages 377 a398
378 Revue europcenne des elements finis. Vol. 2- n° 3/1993

1. INTRODUCTION and versatile technique as design tool. This

statement is based on the fact that each de-
Shells belong to the most common and most
sign follows essentially an optimization pro-
efficient structural elements in nature and
cess. Specifically with respect to the form-
technology. They are used whenever high re-
finding of a shell structure the objective can
sistance, large spans and minimum material
be stated as follows [22]:
are required or a shel(l)ter and containment
function is needed. In this respect they also Find the shape and thickness distribution of
may be termed optimal structures showing a shell, so that
excellent structural performance and in many
cases also architectural beauty. However, as • the boundary conditions and all pos-
sible load cases are considered,
symptomatic of optimized systems shells can
be extremely sensitive with respect to their • material properties are taken into ac-
mechanical behavior as well as their aesthet- count (e.g. no tension for masonry or
ics [20). Any design should consider this sen- concrete),
sitivity to become finally successful.
• stresses and displacements are limited
Besides the thickness distribution the overall to certain values,
shape of the shell is directly related to this as- • an almost uniform membrane stress
pect. It is well known that an extremely thin- state results,
walled shell heavily relies on a load carrying
principle based on a membrane stress state • buckling, excessive creep and nega-
avoiding bending as far as possible. Further- tive environmental effects are avoided,
more, the stress state has to reflect the char- • a reasonable life time is guaranteed
acteristics of the chosen material: a fabric
and hopefully
membrane needs enough prestress avoiding
wrinkling under compression, a reinforced • manufacturing and service costs are
concrete shell ought to be mainly in compres- justified and the design is aesthetically
sion. This ideal situation can of course rarely pleasing.
be achieved, if a so-called geometrical, These requirements interact with each other
mostly analytically defined shape is adopted. and are in some cases even contradictory so
In this case extra structural elements like that a compromise has to be made. For exam-
additional reinforcement, prestress, stiffen- ple, a free form shell may lead to expensive
ers, edge beams etc. ("prostheses") are formwork and an efficient concrete shell
needed to put the shell into the desired posi- might look rather bulky.
tion. In contrast natural or structural shapes
try to avoid most of these extra stiffening com- Structural optimization is currently under-
ponents. They are obtained by an inverse ap- stood as a synthesis of several individual dis-
proach in which the objective of a desired ciplines like structural and sensitivity analy-
structural response is prescribed and the ini- ses, computer aided geometrical design
tial design, e.g. the shape and thickness dis- (CAGD), mathematical optimization, interac-
tribution, is looked for. The interrelationship tive graphics etc. Apparently, it is a computa-
between shape and structural response has tional method, consequently only those re-
been intensively discussed by the authors quirements described above may be part of
in [21). [22]. It has been mentioned that one the process which can be cast into a mathe-
of the described methods, namely structural matical formulation. Unfortunately, also the
optimization, seems to be the most general term 'optimization' is misleading since it sug-
 Shape optimization of shell structures 379

gests that there is only the one optimal solu- mathematical programming schemes and -
tion. Firstly, only parts of the complete task most important - a general concept to incor-
can be included up to now so that always a porate all kinds of objectives, constraints and
model problem is investigated. Secondly, design variables. The formulation is currently
even for this restricted model a local optimum extended to geometrically nonlinear struc-
is reached. In other words, creativity of the de- tures including buckling and its related
sign is still kept as part of the game; fortu- imperfection sensitivity (24].
nately, the process is - in this rather general
perspective - extremely parameter sensitive 2. OPTIMIZATION MODEL
and allows a lot of design freedom. Structural
optimization is nothing else then an addi- 2.1 General Procedure for Shape
tional design aid. Its applications can be clas- Optimization of Shells
sified into: Hanging fabric models or their numerical sim-
• homogenization problems (uniform ulation are excellent techniques in the form-
finding process for a membrane oriented
stress state etc.),
shell design [17]. They are simple and ideally
• optimal use of material (trimming, maxi- suited in the initial phase since they always
mum load carrying capacity), give a rough picture of a potential shape.
• optimal structural response (tuning, However, their application has certain limits.
e.g. frequencies).
Usually, the material of the membrane used in
In this paper we concentrate on shape and
the experiment or analysis is not related to
thickness optimization. The tuning of axisym-
that of the real shell. Wrinkling of the hanging
metric shells to certain frequencies and buck-
fabric in general cannot be avoided and it is
ling loads by shape modification is described
not clear, how different load cases can be in-
in [16]. Minimum weight and cost design of
corporated. Furthermore, experimental data.
rotational shells is addressed in [23]. Shape
have to be processed anyway for a subse-
optimization schemes are used in (15] to elim-
quent structural analysis.
inate bending and minimize membrane
stresses of arbitrary shells with constant
Therefore, a more general approach to shell
thickness. Shape optimization of prismatic
design is advocated here following the prin-
curved shells and axisymmetric shells is in-
ciples of structural optimization. They ideally
vestigated in [9], (1 0] where also further refer-
reflect the individual design stages every en-
ences are- given. The literature for optimiza-
gineer usually goes through, namely
tion of nonlinear shell structures is rather
limited, e.g. [25]. In this context a different 1. choose a reasonable initial shape,
shape sensitivity with respect to the influence
of small geometrical imperfections on the 2. evaluate the structural response for
buckling and failure load of optimized struc- several load cases,
tures has to be mentioned.

The present paper summarizes a general 3. check stresses, displacements, buck-

scheme for shape and thickness optimization ling load and other safety requirements
of free-form shells developed in the authors' and serviceability conditions, costs,
research group [3], (7], (11], (12], (19]. This
includes the geometrical parameterization, 4. compare the quality of the design with
sensitivity analysis, the application of certain any chosen optimality criteria,
380 Revue europeenne des elements finis. Vol. 2- no 3!1993

5. if necessary, evaluate a new trend ap- 2.2 Definition of Optimization Problem

plying sensitivity analyses and propose
an improved design, The rather abstract mathematical statement
of a non-linear optimization problem
6. repeat the process until all criteria and minimize the objective f(x)
constraints are satisfied. subject to
equality
To allow some flexibility and generality constraints: g1(x) = 0 ; j = 1, ... , m.
the process has to be structured in optimiza- (1)
inequality
tion, i.e. geometry definition, mechanical be- constraints: g 1(x) s 0 ; j = m. + 1, ... ,m
havior, design objectives and constraints and
bounds for
mathematical optimization module are optimization variables: .!h s x s x.
strictly separated (Fig. 1).

This procedure is absolutely different to the has to be redefined in mechanical terms

hanging model or other principles where the (Fig. 2).
shape generating rule itself (i.e. mechanical
For shells coordinates and thicknesses of cer-
response to given loads) is already the crite-
tain selected design or structural nodes are
rion for optimality. Nevertheless, the same re-
selected as optimization variables. In order to
sults can be achieved if equivalent objectives,
allow a smooth and ·efficient solution, the
material data, boundary and load conditions
number of design variables should be kept as
are chosen. But beyond this the methods of
small as possible but still allowing enough
structural optimization can handle problems
freedom for a general shape. Besides the
with many load cases, arbitrary design objec-
most common objective "weight" there are
tives and constraints not necessarily related
other functions of natural significance like
only to mechanical behavior, loads like body
strain energy minimization which is equiva-
forces and support conditions which change
lent to maximizing the stiff-ness. This means
with every modification of shape.
that the bending strains in a shell are mini-

CAGD

mathematical linear
programming eigenvalue
approximation nonlinear
methods dynamics
optimality
criteria

objective(s) variational SA
constraints discrete SA
derivatives

Figure 1 - Design = optimization loop

 Shape optimization of shell structures 381

mized so that a membrane stress state is • Design variables x

achieved. Stress levelling with a prescribed coordinates of selected nodes
target of stress oavg can be applied to gener-
thickness of selected nodes
ate a shell mainly in compression. Tuning to a
certain response, for example a single fre- other cross sectional parameters A
quency or a desired spectrum or maximizing. • Objectives f(x)
the failure load are classical objectives in en- weight or volume
gineering. In all cases we have to keep in mind
that optimized structures may result in ex- strain energy
treme parameter sensitivity.
stress leveling
Inequality constraints are taken into account
to check the safety and reliability require-
tuning function
ments which have to be satisfied. Typical
constraints of this type are stress and dis-
fundamental frequency
placement limits. If the stiffness or the critical
load factor are to be maximized a prescribed critical load factor -A.
structural mass is introduced via an equality
constraint. This constraint prohibits an accu- • Constraints
mulation of mass which would otherwise pro- weight or volume gw = .:!L - 1 =0
Wall
duce unrealistic massive structures.
For multi-objective optimization the problem
displacements gu = U~u - 1 S 0
has to be generalized (Fig. 3) allowing only a
compromise. Here, either several weighted
stresses g.=~l -1 so
objectives are combined to one compromise
or one dominant function f 1 (x) is chosen as frequencies
leading objective whereas the other functions
are introduced as constraints. Alternatively, load factor
Pareto optimal solutions are located on the
so-called functional- efficient curve A- B Figure2 - Typicalvariab/esandfunctionsin
in the criteria space with C as the min - max optimal shape design of shells
solution. Also different load cases can be han-

• general objective weighting factors

k number of individual
• leading objective objective functions

as constraints
I
• Pareto optimization f, I
f--A Pareto optimal solutions

>z
"--V______ f2
Figure 3 - Multi-objective optimization
 382 Revue europeenne des elements finis. Vol. 2- n° 3/1993

died in a similar way if for example the values

for one objective of several load cases are
added to one general function.

3. GEOMETRIC REPRESENTATION (3]

3.1 Design Model

Shape finding means optimization of geome-
try. Characteristic optimization variables are
therefore geometric parameters defining the
structural shape. The number of variables
can be reduced dramatically without loss of Figure 4 - Bezier patch
generality if CAGD concepts are used. By
these methods shapes of free formed shells There are many different shape functions
can be described by the coordinates of a few available, e.g. Lagrangian interpolation,
so-called "design nodes" which can be cho- Coons transfinite interpolation, Bezier and
sen as variables. Additionally, the continuous B-spline approximations. In shape design of
thickness variation can be optimized where free form shells one dimensional cubic Bezier
thicknesses at design nodes are taken as dis- .and B-splines and two dimensional bi-cu-
crete variables. bic Bezier patches (Fig. 4) appear to be supe-
rior to others.
The general methods of CAGD are the basis
of modern pre-processors to design struo- Continuity conditions between adjacent
tural geometries in two and three dimensions. patches of composite surfaces can be formu-
Shapes are approximated piecewise by "de- lated in superimposed "continuity patches".
sign patches". Within each design patch the They are generated automatically and pre-
resulting shape r8 is parameterized In terms served during manual user interactions and
·of shape functions Hio patch parameters u, shape optimization. Fig. 5 shows different
v, w and design nodes rdi which describe types of continuity patches depending on
the location ofthe patch in space: whether they are connecting two or four de-
sign patches or are defined at an isolated cor-
ner. In either case four nodes are independent
( 2)
and control the locations of the remaining

shift design nodes generated shape

Figure 5 - Interactive surface modification, continuity patches connecting four Bezier patches
 Shape optimization of shell structures 383

generated shape

ground plan

Figure 6 - Free formed shell (16 Bezier elements)

nodes leading to a reduction .of geometrical variables x of the optimization model Is de-
degrees of freedom which is very welcome in fined as:
structural optimization to stabilize the proce-
dure.

The idea of continuity patches is very helpful with: (4)

in interactive design of free formed shells be-
cause they can serve as initial shapes for sub- In these relations r~ and r3 denote coordi-
sequent optimization runs or as valuable in- nates of analysis and design models, respec-
teractive pre-processor tools for input tively, which remain constant during the opti-
preparation of complex shapes. Fig. 6 mization process. Linking matrices Lax and
shows the plan of a free form shell de- Ldx describe linear relations between opti-
scribed by a total of 16 Bezier patches and the mization variables x and variable coordi-
generated shape modeled by 8-noded iso- nates of analysis and design model, respec-
parametric shell elements. tively. Had denote nonlinear relations
between design and analysis model and are
identified as shape functions. In addition,
3.2 Linking
specific linking rules can be introduced to
The concept of linking is a necessary tech- prescribe move directions of nodes, linear
nique to introduce certain geometrical combinations of nodal variables, symmetry
constraints. One application has already conditions, projection rules etc. (3].
been mentioned above in the context of conti- The concept of linking can drastically reduce
nuity requirements. Another one is the inter- the number of optimization variables. It also
action of the analysis model (e.g. finite ele- allows the user to guide the solution Interac-
ment model) and the design model tively into a reasonable design.
(geometrical macro element model). A com-
monly used rule which links variables r 8 of
the analysis model via the design model with
384 Revue europeenne des elements finis. Vol. 2- no 3/1993

4. ANALYSIS The discretized nonlinear 'equilibrium equa-

tions G(u, A.) = 0 are usually linearized, lead-
4.1 Structural Analysis
ing to the tangent stiffness matrix KT, and it-
Once the design model is defined a finite ele- eratively solved by a Newton type of iteration,
ment mesh can easily be interpolated within in general combined with a path-following
each design element (eqn. (2)). This concept scheme, e.g. the arc- length -method
cannot totally avoid mesh distortions during (24]. The maximum load, which is either a bi-
shape optimization. However, by choosing furcation or a limit point, is controlled by the
reasonable patches and prescribed move stability criterion (6 a). Usually the criterion is
directions the effect can at least be dimin- utilized during the loading process to check
ished. In most cases the structural response whether or not the structure is in a stable equi-
for shape optimization is based on linear elas- librium position or a bifurcation into another
tic analyses or eigenvalue formulations for equilibrium path occurs.
buckling or vibration. The equilibrium condi-
In order to pinpoint a critical load exactly a
tion of the discretized system G = 0 leads to
so-called extended system can be utilized
the usual stiffness expression. Its homoge-
[30].
neous counterpart defines an eigenvalue for-
mulation for the load parameter >... The ho-
mogeneous equation of motion yields the free G(u, A.) } linearization
vibration problem: K1 (U, A.) ell = 0 ~
{
equilibrium G(u,A.) = 0 (5 a)
e(ell) (8 a) V
=t> Ku=R (5 b)

[(K~:) (K~ :).l][::]= -[K~J

stability K1 ell 0 (6 a)

J
(6 b) •• : 1
=t> (K-A.K 0)eii=O
free vibration Ku+Mii=O (7a) or e.. o M e (8 b)
=t> (K - fil M) ell = 0 (7b)
Here displacements u, buckling mode ell
and the critical load parameter >.. are the un-
K, K9 , M and R are the elastic stiffness ma- knowns. e(~) is an additional constraint on
trix, geometric stiffness matrix, mass matrix the length of the buckling mode ~. Again the
and external load vector, respectively. KT is linearized equation (8 b) is the basis of a New-
the sum of K and Kg. u denotes the vector
ton type iteration. ( ), denotes the corre-
of the nodal displacement parameters. sponding partial derivatives. A solution of the
Shells are known to be extremely sensitive unsymmetrical set of equations (8 b) Is by-
with respect to small deviations of their ideal. passed by a partitioning method. This, in turn,
shape. The buckling load of the real, requires the factorization of the tangent stiff-
imperfect shell may be drastically lower than ness matrix KT which is singular at critical
that of the perfect structure. This is in particu- points. In (24], [30] different schemes are
lar true for optimized structures. It brings up proposed to augment (8 b) and circumvent
the question, how nonlinear effects and this difficulty.
imperfection sensitivity can be included in the Although the extended system can be started
optimization procedure. This problem is dis- already at the undeformed configuration, the
cussed in detail in (24] for maximizing the crit- direct solution for the critical point may not
ical load of geometrically nonlinear strue> converge. Thus a few steps are computed by
tures and is briefly outlined in the following.
 Shape optimization of shell structures 385

the path following algorithm coming close to ture. This gives the new buckling mode used
the critical point before the extended system as imperfection. And again the imperfect
is turned on. structure is investigated in the same manner
by the extended system. The procedure is re-
Once the critical point of the perfect structure
peated until convergence is reached, i.e. no
including the buckling mode is determined,
further increase of the critical load can be ob-
the original structure is perturbed by a frac-
served. For details see [24].
tion of this mode and the resulting Imperfect
structure is investigated again by the ex-
4.2 Sensitivity Analysis [11]
tended system. Its critical load - usually a
limit point - is the basis for the maximization The sensitivity analysis supplies gradient in-
of the load factor by the optimization proce- formations on objective and constraints with
dure (Fig. 7). A new design is obtained. The respect to optimization variables. In general,
critical point of the new perfect structure is any function t (objective or constraint) de-
computed starting with the extended system pends on optimization variables x and state
at the critical stage of the old perfect struo- variables u, e.g. for displacements. Thus,

Path Following (PF) and Extended System method (ES)

perfect structure imperfect structure

geometry perturbed
~
by fraction of A.
buckling mode c

Uc

Sensitivity analysis
Mathematical optimization
0
Inew design I
0
perfect structure imperfect structure
geometry perturbed
~ Ac
by fract1on of
buckling mode

Uc

~
I Convergence check I
Figure 7 - Optimization including imperfection sensitivity
386 Revue europeenne des elements finis. Vol. 2- n° 3/1993

the total derivative of t with respect to x is than the numerical forward difference
given as: scheme if only parts of the structure are af-
fected by shape variations. It definitely de-
pends on the optimization problem which
.Q!_ =.ill_ du + .ill_ (9)
dx au dx ax method is to be preferred. Therefore, all of
them should be provided.

where the determination of the response sen- For the optimization of the critical load (sec-
sitivity du/dx is part of the job. It can be car- tion 4.1) the total derivative of the critical load
ried out by several different techniques. They factor dA.c/dx has to be determined which in
can be divided into variational or discrete turn depends on the derivative ducfdx of the
methods, depending on whether the gradi- related displacement field. For details see
ents are obtained before or after discretiza- (24].
tion [1], [8]. Nevertheless, the same results
will be obtained if variation, discretization and
derivation are done consistently [11]. Within 5. MATHEMATICAL OPTIMIZATION
the present approach all three variants of dis- The optimization problem eq. (1) is in most
crete sensitivity analysis (DSA) are adopted: cases definitely nonlinear since all functions
numerical, semi-analytical and analytical. (objective f(x) and constraints g(x)) are
nonlinear functions of the optimization vari-
The analytical derivation ofthe state variables
ables x. A local solution is characterized as
is given as:
a stationary point of the corresponding La-
grangian function:
(10)
L(x, u) = f(x) + vr g(x) (12)

where R is the load vector and K is the sys- v are the Lagrange multipliers or dual vari-
tem stiffness matrix. The major concern is the ables. The necessary condition ·:for the sta-
calculation of the pseudo load vector: tionarity of L or the corresponding con-
strained minimum of f(x) is defined by the
Kuhn-Tucker conditions. They give a set of
R = dR _ dK u (11) nonlinear equations to determine the optimal
dx dx
solution x*, v*:
which results in analytical derivations of R
and K.

Since the element stiffness matrix is decom-

posed into several individual matrices (e.g.
aL *
-~- = g 1• = 0 ; j = 1, ... m. (13)
kinematic operator, constitutive matrix, de- .,vi
terminant of Jacobian) the analytical deriva-
tion has to be carried out for all individual ma- *
aL =
- gl• VI=
• 0 }
trix elements. av1 j = m. + 1, ... m
vj ~ 0
However, the analytical version of DSA ap-
pears to be very reliable for all kinds of ap-
where f*, g* and L* are the function values
plications at the cost of a higher programming
at the optimal solution. Without loss of practi-
effort. Additionally, it is even more efficient
 Shape optimization of shell structures 387

cal relevance for the presented range of ap- second derivatives of the Lagrangian with re-
plications the problem functions are stated to spect to x.
be continuous in gradients and curvature.
SOP-methods have been used rather infre-
The methods of non-linear programming quently within structural optimization so far
can be divided into (i) primal methods (e.g. but with the increasing complexity of prob-
method of feasible directions), (ii) penalty lems like shape optimization these methods
and barrier methods (e.g. sequential uncon- get more and more accepted (27]. The au-
.strained minimization technique), (iii) dual thors can report very good experiences with
methods and (iv) Lagrange methods (14]. all kinds of structural optimization problems.
They are distinguished by the type and num-
The performance of the iterative design loop
ber of independent variables they use.
can be improved in certain applications if in
Lagrange methods can be stated to be the every iteration step objective function and
most sophisticated numerical optimization constraints are replaced by proper approxi-
techniques and they are applicable for all mations. Usually, approximations are derived
kinds of constrained problems. They are de- from first order linearizations with respect to
signed to solve the Kuhn-Tucker conditions problem oriented, generalized variables. The
(13) directly and are operating In the full idea came up first for sizing of statically deter-
space of primal and dual variables. Iterative minate trusses where cross sectional areas
solution of (13) by subsequent linearization as the design variables enter the stress
leads to a natural extension of the classical constraints in the denominator. Conse-
Newton-Raphson procedure which became quently, if areas are substituted by their recip-
known e.g. as SOP-method (sequential qua- rocal value the correct solution emerges. The
dratic programming [26]). Since in the con- idea has been generalized and applied also in
text of structural optimization the evaluation shape optimization. Depending on the kind of
of second derivatives with respect to opti- approximation these techniques became
mization variables is far too expensive quasi- known as hybrid approximation, convex li-
Newton variants are used. In the k-th itera- nearization or method of moving asymptotes
tion step the corresponding quadratic (MMA). for a review see (2], (3]. The special
subproblem states as: advantages of approximation methods are
convex and separable sub-problems which
minimize: can be solved efficiently by specialized solu-
tion schemes, e.g. dual optimizers. In (6] an
extended version of the method of the moving
asymptotes (EMMA) is proposed which es-
sentially demonstrates that the method can
subject to:
be embedded in the SOP-formulation as
agl(xk) dk + ( k) = 0 ; j = 1, ... m.
special case.
ax gl x

agl(xk) k 6. EXAMPLES
-ax- d + g1(x") s 0 ;j = m. + 1, ... m
6.1 Bl- parabolic Roof Shell [5)
with: dr s ~"- xr This example is used to demonstrate the ef-
fects of different objective functions and the
variety of shapes which can be generated by
where Bk is the current approximation of the only two variables. The structural situation is
388 Revue curopeenne des elements finis. Vol. 2- no 3/1993

shown in Fig. B. A shell of rectangular plan (b duce tension stresses in the lower fibres ofthe
= 6 m, I = 12 m) and uniform constant thick- structure which are caused by interactions of
ness (t = 0.05 m) is supported by dia- normal forces and bending moments. A tar-
phragms at the smaller edges. The shape is get stress of o8 = - 100 kN/m 2 was pre-
generated by four Bezier patches. The design scribed.
nodes are linked (i) to preserve double sym-
The optimal structure (Fig. 9 b) is a synclastic
metry and (ii) to describe a bi-parabolic sur-
shape (EP) where the area of tension in lower
face which can be controlled by two vertical
fibres is reduced to a minimum. Tension can-
coordinates as indicated. In the initial design
not be avoided totally because of the simple
both coordinates are set to x1, 2 = 3 m de-
scribing a cylindrical shell. The structure is
optimal values: initial values:
loaded by a uniform vertical load p = 5 kNfm2
s1* = 6m s1 =3m
(snow). Support conditions are fixed hinges. s2* = 3,12m s2 =3m
Due to symmetry of loads and structure only
one quarter of the shell has to be analysed.
This was done by 72 eight - noded isopara-
metric shell elements which are 2x2 reduced
integrated.
In a first optimization run strain energy was
chosen as objective function without stress
constraint assuming the structure is suffi-
ciently reinforced to resist also high tension
a) optimal shape, minimization of strain energy
forces. The resulting shape (Fig. 9 a) is an an-
ticlastic surface (HP), very similar to a minimal optimal values: initial values:
surface which acts almost like a membrane in s 1* = 0.90m s1 =3m
tension and compression. Since the struc- s2* = 1.97m s2 =3m
tural thickness is fixed, the result is alterna-
tively restricted by an upper bound (6 m) on
variable x 1.
To get a more suitable design for concrete the
objective "stress leveling" was used to re-
b) optimal shape, stress leveling

optimal values: initial values:

s1* = 1.64m s1 =3m
s2* = 1.34m s2 =3m

E =30,000 MPa,
c) optimal shape, weight minimization
v =0.2 (concrete)
Figure 8 - Parabolic roof shell: problem Figure 9 - Parabolic roof shell:
statement initialshapesndoptimizationresults
 Shape optimization of shell structures 389

shape function and the rectangular plan of the are only important for the convergence of the
structure. It is remarkable that the dia- algorithm.
phragms.- although possible- do not disap-
pear. If they vanish, the resulting shape has a
a) Design-Model
horizontal tangent plane at the corner leading
to negative curvature and Increased bending.
If "weight" is used as objective fu.nction any
shape between the "minimal surface" and a
plate can be determined which is forced by
additional constraints on stresses and dis-
placements. Fig. 9 c shows one result ob-
tained with constraints on v. Mlses effective
stresses which are not allowed to exceed an
thickness geometry
arbitrarily chosen value of om = 400 kNJm2. distribution description
6.2 Tuning of a Bell [4], [12]
b) Optimization Model
The major design aspect for a bell is to pre-
serve high tonal quality. For this reason the
tuning of the basic frequencies denoting the
partial tones of the bell is introduced as objec-
tive of the optimization problem without any
other constraints. The frequency require-
ments of a minor- and a major-third bell are 0.70 x1
very much influenced by the number n of 0.82 X
goal frequencies A.; 0 and the individual 0.87 X
0.92 1
weighting factors w; used in the objective f>,. variables for
sizing shape

f =f (1.1 - A. 10)
2
w Figure 1o - Definition of the optimization
1. 1=1 A~ I modal of the bell

The harmony of the lowest five partial tones

The variables for the optimization model are
leads to the following frequency require-
selected with special care to obtain a well
ments [Hz]:
posed optimization problem. As shown in Fig;
major-third minor-third 10 eight sizing and ten shape values in the de-
bell bell sign model are used as relevant optimization
hum 512 512 variables. Some restrictions for design coor-
fundamental 1024 1024 dinates and nodal thickness values are
third 1290 1218 introduced to obtain a useable optimization
fifth 1534 1534 result. The number of variables can be further
octave 2048 2048 reduced by using linking schemes for sizing
and shape variables, like the vertical shape
This tuning problem Is a multi-criterion opti-
coordinates in Fig. 10 which altogether are
mization problem. As optimization strategy a
linked to the variable x1 •
SOP-method is used. Since all frequency re-
quirements are met exactly at the optimal From the above mentioned 18 relevant opti-
solution, the individual weighting factors w; mization variables 15 are used as indepen-
390 Revue europeenne des elements finis. Vol. 2- n° 3/1993

dent variables to improve the minor-third ally nonlinear analyses,

bell. Seven of. them are sizing variables.
The resulting shape (Fig. 11) shows only initial minor major
slight modifications of the initial shape.

To obtain the major-third bell, the height of

the bell is introduced as additional variable.
The shape of this bell has an increased height
and a moderate bump at half the height of the
bell, also described by Schoofs [28).
In both optimization procedures the objective
function becomes zero within a given toler-
ance bound. This means that the frequency
requirements are fulfilled exactly (see final
frequencies in the table of Fig. 11 ). The itera-
tion history of the objective function and the
lower three eigen-frequencies are shown in
Fig. 11 for both, the minor- and major-
third bell. The difference is pointed out by
the third eigen-frequency which is in-
creased for the major-third bell.

6.3 Kresge-Auditorium at MIT [22)

6.3.1 Structure, Objective 18 11 211 2Z

Saarinen's famous Kresge auditorium (1955)

at the MIT (Fig. 12) belongs to the class of
shells with a geometrical form. The three
point supported dome is a 1/8 segment of a
sphere with a radius of 34.29 m (Fig. 13), a z....... ,.•••••.,
side length of about 48.5 m and an elevation
atthe vertex of 14.5 m and atthe crown ofthe
arches of 8.23 m. The shell has a thickness
of 8.9 em which increases to 12.7 em atthe 2 4 B a ~ n u ~ ~ 211
edge beam and to 49.5 em ·at the supports.
The edge beam is 25.4 em wide and varies
from 50.8 em depth at the crown to 91.4 em
at the support. A steel casting is added which
is pinned to the abutment. The dead load is
3.98 kNfm2, a uniform live load, representing
- .....
+-+ •••••
roughly also the wind, is 1.44 kNJm2.

to demonstrate that rigid geometrical

shapes do not lead to an appropriate
shell-like behavior. This is not only ver-
ified by linear elastic analyses, It can be
Figure 11 - Optlmlzatlonresultsandltera-
confirmed by geometrically and materi-
tion history of minor- and major-third bell
 Shape optimization of shell structures 391

Figure 12 - Kresge auditorium at MIT. Cambridge, MA

Figure 13 - 1/8 - segment of sphere

to answer the question how a natural 6.3.2 Scope of Study

shape of a dome of this size and kind
The initial analyses are based on linear elastic
should look like.
material properties with a Young's modulus
In optimization the total strain energy is used of 3 . 104 N/mm2 and Poisson's ratio of
as objective thus minimizing the bending in 0.2. This holds also for the shape optimiza-
1
the she 11. In the first instance it looks Uke that tion. Afterwards the original and the opti-
this design objective may cause a conflict mized shells are investigated by geometri-
since a membrane oriented stress state cally nonlinear analyses. Finally, the
mainly in compression turns out to be more nonlinear behavior of reinforced concrete is
sensitive with respect to buckling and geo- added. On'y dead load is applied based on a
metrical imperfections (see results for geo- concrete weight 25 kNfm3. This load which
metrically nonlinear analysis). The objective leads to 2.23 kNJm2 for a thickness of
has to be judged in the context of the chosen 8.9 em is augmented in the entire shell area
rrtaterial, i.e. reinforced concrete with a low by a uniform load of 1.75 kNJm2 for the extra
tension capacity. coverage. No live load is considered. In order
392 Revue europeenne des elements finis. Vol. 2- n° 3/1993

to avoid high stress concentrations the shell are 1.7 cfn and 5.9 em, respectively. Con-
corner is slightly cut off at the supports (at a siderable bending occurs not only at the sup-
distance of 1.20 m from the ideal corner port but also beside the edge beams, leading
point). to tension stresses of up to 17 N/mm2. Of
course if no edge beams are added unaccept-
The finite element analyses are based on
able values for the displacements (maximum
8-node isoparametric shell elements with re-
.29 em) and stresses(> 25 N/mm2) result.
duced integration. 42 elements are used for
1/6 of the shell. Initial stability and geometri- • Shell with "optimized" shape
cally nonlinear analyses of the perfect and The above described techniques for shape
imperfect structure are performed to under- optimization were applied to the three-point
stand the buckling sensitivity of the shell. In supported shell having the same plan dimen-
this case the entire structure is modelled to al- sions, vertex elevation and construction vol-
low for unsymmetrical failure modes. The ume (amount of concrete). For simplicity only
imperfections are based on the fundamental two design elements (bicubic Bezier ele-
eigenmode scaled to a fraction of the thick- ments) are introduced, defined only by five
ness. geometric variables. in addition, extra thick-
ness parameters are chosen allowing an "op-
The material nonlinear analyses are based on timal" thickness variation. Two studies were
an incremental inelastic, orthotropic material performed: In the first one an edge beam with
model with a smeared crack approach for the a varying thickness can develop, in the other
concrete (13). [18]. The layer-wise model
one the shell is supposed to have a free edge.
utilizes a 20 failure curve, a nonlinear stress In this case the elements of the boundary are
- strain curve in compression including soft- prevented from geometrically bending down-
ening and tension stiffening. The chosen ma- wards indirectly developing into an edge
terial properties are: initial modulus of elastic- beam. As side constraints the minimum and
ity E0 = 3 . 104 N/mm2, uniaxial strength in maximum thickness is chosen as 8.9 em
compression fc = 30 N/mm 2 and ten- (thickness of MIT -shell) and 100 em, re-
sion f1 = 2.2 N/mm 2, strain at maximum spectively. No other constraints are used.
compression Ec = 0.004, maximum strain of
tension stiffening Eut = 0.0025. Each steel Figure 14 c shows the shape of the optimized
layer has multi-linear 1D properties. Double shell with a free edge. It can be seen that it
layered reinforcement with 2.57 cm2Jm in differs substantially from the original
both directions is used. An elastic, linear sphere. it is more parabolic at the edge and
hardening model with E = 2.1 . 1os N/mm2 has a more pronounced curvature perpendic-
and Eh = 0.2. 105 N/mm 2 is applied. The ular to the boundary. The niain difference is
yield limit fy is 240 N/mm2. the large increase of the elevation of the
crown which comes along with a negative
6.3.3 Shape Optimization
Gaussean curvature. The thickness varies
• Analysis of original dome from 8.9 em to 42.8 em. The displacements
For simplicity, the thickness is varied only in at the center and at the crown of the arch are
the direction of the supports with a maximum reduced to 0.2 em and 0.3 em, respectively.
value of 0.35 m. The edge beams have a The stresses are clearly smaller (max.
width of 0.25 m, their thickness varies from v. Mises stress 6.4 N/mm 2), bending almost
0.51 to 0.91 m. in the analysis it is simu- vanishes. Although there is tension it is al-
lated by a thicker shell element. The displace- most negligible (about 1/10 compared to the
ments at the vertex and crown of the arches previous solution).
 Shape optimization of shell structures 393

a. original shell

b. optimized shell
with edge beam

c. optimized shell
w/o edge beam

d. inverted membrane

Figure 14 - Shapes of shells

The shell exhibiting an edge beam has a this investigation could be improved H geo-
slightly different shape (Fig. 14 b).The thick- metrical as well as structural models are re-
ness of the edge beams varies from 8.9 em at fined,
the crown to 100 em at the corners. The shell
• Shell as Inverted membrane
itself is also thinner in this region. The wavy
A uniform dead load is applied to an extremely
character above the support disappears be-
thin membrane with very low bending stiff-
cause now the shell carries the loads more via
ness having the same plan dimensions. The
the edge beams to the foundation. Although
deformation of the membrane with free edges
the displacements are further reduced the
is monitored by a geometrically nonlinear
overall stress state is similar. The results of
analysis until the center deflection reaches
394 Revue europeenne des clements finis. Vol. 2- no 3/1993

0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8
characteristic displacements [m]

Figure 15a - Load - deflection diagram: initial buckling and geometrically nonlinear analyses

6.---------------------------------------------------,
optimized shell
5 • w. imperfections

' -- _ \ - - 0

original shell
• w. imperfections

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - ~- - - - - 0

o~~------~~--.-.--r~,-~-.-.,-.--.-.--T-~-.--r-4
0 0,05 0,1 0 0,15 0,20 0,25 0,30 0,35 0,40 0,45 0,50
characteristic displacements [m]
Figure 15b - Load - deflection diagram: geometrically and materially nonlinear analyses

-------

optimized shell
inverted membrane
original shell

Figure 16 - Comparison of shapes

 Shape optimization of shell structures 395

the value of the total height of the original This is different for the optimized shell. Be-
sphere. This shape is inverted (Fig. 14 d). A cause of the free edge the buckling load is
thickness variation similar to the original slightly lower in this case (A.= 11.3 in an initial
structure is assumed and a linear elastic buckling analysis,>..= 16.4 in a large deflec-
analysis is performed. tion analysis, Fig. 15 a). As expected the opti-
mized structure is more sensitive with respect
The final shape differs from the optimized to initial imperfections. For a maximum
shape primarily near the free edge. Oppo- imperfection amplitude of 1.5 . thickness the
site to this it exhibits a positive Gaussian nonlinear elastic buckling load drops to >.. =
curvature (which was artificially pre- 8.6.
vented in the optimization model). Further- The real collapse behavior is very much in-
more, slight membrane wrinkles can be visu- fluenced by material failure and therefore
alized at the support. But the structural yields to lower failure loads compared to elas-
analysis leads to a similar positive response tic analyses. The structural response already
compared to the optimized shell: small dis- depicted in the left lower corner of diagram 15
placements, low stresses, little tension, less a is zoomed in Fig. 15 b. Now the anticipated
bending. result is obtained. Although the optimized
The three shapes are compared In Fig. 16. shell with a free edge is still imperfection sen-
v. Mises and principal stresses are plotted In sitive to a certain extent its safety margin is
(22). still sufficient. The original shell with edge
6.3.4 Nonlinear Analyses beams but without mullions results in an
extremely poor behavior with a maximum
The original shell with edge beam and the op-
load multiplier >.. slightly above 1.0.
timized shell with free edges are further inves-
tigated first by a geometrically nonlinear but It needs to be mentioned that the above de-
elastic analysis (Fig. 15 a). Afterwards the ma- scribed study is still of academic nature and
terially nonlinear formulation is added (Fig. is intended only to demonstrate some charac-
15 b) simulating the collapse of both struc- teristic features of shells. For a real design
tures. much more investigations (environmental ef-
The Initial buckling analysis of the original fects, long time behavior, non-uniform loads
shell leads to a load multiplier of >.. = 17.6 with etc.) have to be undertaken to verify the feasi-
a symmetrical failure mode which Is 29 per- bility of the design.
cent of the Zoelly load for a perfect spherical 6.3.5 Conclusion for MIT -Shell
shell under pressure The load multiplier >.. is The comparison of all four shapes (Fig. 16)
defined with respect to the variable weight of gives a good insight what a "natural" form
the concrete and the uniform load of the cov-
means. In the sphere the stress flow is di-
erage. A large deflection analysis indicates
rected towards the cut off free edge and does
distinct nonlinearities but leads almost to the
not find a stiffening element unless a heavy
same buckling load (A.= 16.2), see the load
edge beam or a support is added. Contrary to
deflection diagram Fig. 15 a. The same holds
this the other forms develop an arch with a suf-
for the imperfect structure with a maximum
ficient stiffness by increasing the curvature
imperfection amplitude of 1.5. thickness (A.=
perpendicular to the free edge. This causes a
15.3). This means that the "amputated" shell
considerable reduction in displacements and
itself anticipates the enormous imperfection
stresses. Whether this curvature is positive
sensitivity usually present in complete spheri-
(inverted membrane solution) or negative
cal shells and spherical caps.
(enforced by constraints in the optimized
396 Revue europeenne des elements finis. Vol. 2- n° 3/1993

solution) does not play an important role. It is REFERENCES

interesting to note that the three-point sup-
[1] Arora, J.S., Haug, E.J.: Methods of De-
ported shell, Investigated in [18] in which
sign Sensitivity Analysis in Structural
three kinks between the corners and the
Optimization. AIM-Journal, 17 (1979),
center were admitted, carries the loads
970-974.
mainly in the direction of these lines.
[2] Barthelemy, J.-F. M., Haftka, R.T.:
7. CONCLUSIONS Approximation Concepts for Optimum
Structural Design - a Review. Structural
The methods of structural optimization have
Optimization, 5 (1993), 129-144.
been presented as general computational
tools to find the shape of shells subjected to [3) Bletzinger, K.-U.: Shape Optimization
many different combinations of objectives of Surface Structures (in German). Ph.D.
and constraints. This Includes of course the dissertation, lnstitut fur Baustatik, Univ.
form finding of membranes. The correspond- of Stuttgart, 1990.
ing objectives (e.g. prestress) and
[4] Bletzinger, K.-U., Reitinger, R.: Shape
constraints (e.g. a desired design space) can
Optimization of Shells. Proceedings Int.
be formulated as an optimization problem as
Symposium on "Natural Structures -
well, for example applying the least-square
Principles, Strategies and Models in Ar-
principle or the minimum surface solution.
chitecture and Nature", Sonderfors-
In general, shape optimization Is based on lin- chungsbereich 230, University of Stutt-
ear elastic structural response including at gart, 1991.
best linear buckling or eigen-frequency
(5] Bletzinger, K.-U., Ramm, E.: Structural
analyses. In this paper the formulation is ex-
Optimization as Tool for Shape Design.
tended also to geometrically nonlinearities in-
Proceedings First European Conference
cluding instability phenomena. A key point of on Numerical Methods In Engineering
this approach Is that it allows to include the (ECCOMAS), Brussels, Sept. 1992.
imperfection sensitivity with respect to buck-
ling. Further studies are currently under way [6] Bletzinger, K.-U.: Extended Method of
to verify the method for shape optimization of Moving Asymptotes Based on Second
nonlinear shell structures. The effects of ma- Order Information. Structural Optimiza-
terial non-linearities as well as time depen- tion, 5 (1993), 175-183.
dent influences have not been considered but [7] Bletzinger, K.-U., Ramm, E.: Form Find-
are challenging tasks for further studies. ing of Shells by Structural Optimization.
J. Engineering with Computers, 1993.
ACKNOWLEDGEMENT (8) Haftka, R.T., Adelman, H.M.: Recent De-
velopments in Structural Sensitivity
This work is part of the research project SFB
Analysis. Structural Optimization, 1
230 "Natural Structures - Light Weight Struc-
(1989),137-151.
tures in Architecture and Nature" supported
by the German Research Foundation (DFG) at (9] Hinton, E., Rao, N.V.R., Slenz, J.: Finite
the University of Stuttgart. The support is Element Structural Shape and Thick-
gratefully acknowledged. ness Optimization of Axisymmetric
Shells. Engineering Computation, 9
(1992), 499-527.
 Shape optimization of shell structures 397

[10] Hinton, E., Rao, N.V.R.: Analysis and (20] Ramm, E.: On Structural Efficiency and
Shape Optimization of Variable Thick- Elegance. Proceedings Int. Symposium
ness Prismatic Folded Plates and Shells. on "Natural Structures - Principles,
Submitted to J. Thin-Walled Structures. Strategies and Models in Architecture
and Nature", Sonderfor- schungsber-
[12] Kimmich, S., Reitinger, R., Ramm, E.: In-
eich 230, University of Stuttgart, 1991,
tegration of Different Numerical Tech-
Part Ill, 69-78.
niques in Shape Optimization. Structural
Optimization, 4 (1992), 149-155. (21] Ramm, E.: Shape Finding Methods of
[13] Kompfner, T.A.: Ein finites Elementmo- Shells. Bulletin of the International
dell fur die geometrisch und physika- Association for Shell and Spatial Struc-
lisch nichtlineare Berechnung von Stahl- tures (lASS), 33 (1992), n. 2, 89-99.
betonschalen. Ph.D. thesis, report No.2, (22] Ramm, E., Reitinger, R.: Force Follows
lnstitut fur Baustatik, Universitiit Stutt- Form in Shell Design. Proceedings
gart, 1983.
IAss·-cscE Int. Congress on "Innova-
[14] Luenberger, D.G.: Linear and Nonlinear tive Large Span Structures", Toronto,
Programming, Addison-Wesley, Read- July 1992.
ing, 1984.
(23] Reinschmidt, K.F., Narayanon, R.: The
[15] Mohr, C.A.: Design of Shell Shape Using Optimum Shape of Cooling Towers.
Finite Elements. Computers & Struc- Computers & Structures, 5 (1975),
tures, 10 (1979), 745-749. 321-325.
[16] Plaut, R.H., Johnson, L.W., Parbery, R.: (24] Reitinger, R., Ramm, E.: Optimization of
Optimal Forms of Shallow Shells with Geometrically Nonlinear Buckling Sensi-
Circular Boundary; Part 1: Maximum tive Structures. Proceedings OPTI 93
Fundamental Frequency; Part 2: Maxi- "Computer Aided Optimum Design of
mum Buckling Load; Part 3: Maximum Structures", July 1993, Zaragoza.
Enclosed Volume. J. Applied Mechan-
(25] Ringerts, U.T.: Optimal Design of Nonlin-
ics, 51 (1984), 526-539.
ear Shell Structures. Report FFA TN
[17] Ramm, E., Schunck, E.: Heinz Isler 1991-18, The Aeronautical Research
Schalen. Karl Kramer Verlag, Stuttgart, Institute of Sweden, Structures Depart-
1986. ment, 1991.
(26] Schittkowski, K.: The Nonlinear Pro-
[18] Ramm, E., Mehlhorn, G.: On Shape Fin-
gramming Method of Wilson, Han
ding Methods and Ultimate Load Analy-
and Powell with an Augmented Lagran-
ses of Reinforced Concrete Shells. Engi-
gian Type Line Search Function. Numer-
neering Structures, 3 (1991), 178-198.
ische Mathematik, 38 (1981), 83-114.
[19] Ramm, E., Bletzinger, K.-U., Kimmich, (27] Thanadar, P.B., Arora, J.S., Li, G.Y., Lin,
S.: Strategies in Shape Optimization of T.C.: Robustness, Generality and Effi-
Free Form Shells. In "Nonlinear Com- ciency of Optimization Algorithms for
putational Mechanics - a State-of- Practical Applications. Structural Opti-
the-Art" ('eds. P. Wriggers, w. Wagner), mization, 2 (1990), 203-212.
Springer, Berlin, 1991.
398 Revue europeenne des elements finis. Vol. 2- n° 3/1993

[28] Schoofs, B., van Asperen, F., Maas, P.,

Lehr, A.: Experimental Design Theory
and Structural Optimization Design for a
Major-third Church Bell. CAMP '85, C
4.2, 1985.
[29] Werner, H.R.: Kresge Auditorium. A Nu-
merical Analysis and Criticism. M.Sc.
thesis, Princeton Univ., 1971.
[30] Wriggers, P., Simo, J.C.: A General Pro-
cedure for the Direct Computation of
Turning and Bifurcation Points. Int. J. Nu-
merical Methods in Engineering, 30
(1990),155-176.