Engineeringwith Computers(1993) 9:27-35

9 1993 Springer-VerlagLondonLimited Engineering

C~nputers

Form Finding of Shells by Structural Optimization

K.-U. Bletzinger and E. Ramm

Universityof Stuttgart,Institutfiir Baustatik,D-7000 Stuttgart 80, Germany

Abstract. Shell structures are known to be extremely changes of certain parameters, e.g., the reduction of
parameter sensitive; even small changes of the initial design, the buckling load due to only small initial imperfec-
e.g., to the shape of the shell, may drastically change the tions. Since realistic situations of design, such as single
internal stress state. The ideal case for concrete shells is a loads, support conditions, free edges, or shape
pure membrane stress state in compression for all loading incompatibilities, make it very difficult to fulfil the
conditions. Since in many realistic situations the solution for
basic membrane-oriented design rules, a modification
an 'optimal' shape is not obvious, the need for form finding
methods is evident. This paper presents computational methods
of the original design could substantially improve the
of structural optimization as a general tool for the form finding structural behaviour. Additionally, the specific pro-
of shells. The procedure as a synthesis of design modelling, perties of the chosen material have to be considered.
structural analysis and mathematical optimization is discussed It is obvious that the shape of a concrete shell should
with special emphasis on the modelling stage. Several examples be designed to avoid tension to the greatest possible
show the power of the approach and the similarities to extent. The ideal situation would be a pure membrane
experimental solutions. state in compression, hopefully avoiding any buckling.
The problem is to find shapes of that kind.
Keywords. Computer aided geometrical design; Form
finding; Shells; Structural optimization
1.2. Traditional Form Finding Methods

It can be recognized that most shells are of regular

1. Introduction shapes which are often analytically defined. This
reflects interactions of the usual design practice and
1.1. Shells as Structures of Optimal Behaviour classical shell theory which gives closed solutions for
analytical defined geometries only. Typical design
Shells are the most efficient structural alternative for practice is to experiment with standard geometries like
a number of extreme situations such as structures of spheres, cylinders, toil, cones and HP-surfaces. To
long spans, minimum mass, or high resistance [13. At adjust the final shape to the prescribed plan, segments
the same time, they appear very light and graceful, are cut out or different shell types are put together,
and meet aesthetical demands in a natural manner. neglecting the basic conditions of membrane theory
There is no doubt that shells are the epitome of with respect to both equilibrium and compatibility.
structural elegance. The extraordinary behavior of This leads to high bending stresses or large displace-
shell structures is caused by the 'double arch effect' ments, which are usually avoided by additional
which, in contrast to one-dimensional curved struc- stiffening elements. Typical examples are the cylindri-
tures, allows them to carry several different load cal roof shell with edge beams and diaphragm walls
constellations by a pure membrane action. That means at the ends, or the spherical shell over a polygonal
that shells designed to act as membranes are already plan with heavy beams at the free edges.
optimal structures. In some cases they may also show An experimental principle long used to find the
typical characteristics of 'over-optimized' specializa- optimal shape of structures in compression is the
tions with high sensitivity with respect to small hanging model and its inverse. The inverted catenary
was used, for example, in 1748 by Giovanni Poleni to
Correspondence and offprint requests to: K.-U. Bletzinger,Univer- compare the shape with Michaelangelo's design of St.
sitfit Stuttgart, Institut f/Jr Baustatik, Pfaffenwaldring7, Postfach Peter's in Rome, and extensively by Gaudi at the
801140, D-7000 Stuttgart 80, Germany. beginning of this century for many structures around
28 K.-U. Bletzinger a n d E. R a m m

Barcelona (chapel in Colonia Gfiell, Sagrada Familia, which are not necessarily related to mechanical
etc.). Extended to two dimensions, the hanging fabric behaviour, and loads such as body forces and support
can be used to find shapes of shells over almost conditions which change with every modification of
arbitrary plans. In many cases, heavy edge beams can shape and which can only be solved with difficulty by
be avoided by this method, to yield pleasing, naturally experiments, and sometimes not even then.
shaped shells with free edges. This principle was used The bulk of investigations is still devoted towards
very successfully and economically in many practical optimization of cross-sections. In shape optimization,
applications by H. Isler [-2, 3]. most work addresses two-dimensional systems [4].
The basis of the hanging model experiment is that Shell problems are usually restricted to the axi-
one characteristic load case is used to generate the symmetric case; relatively little has been described for
final shape by large deflections of a given membrane. general shells so far [5]. Usually, as in this paper,
So far, the method can easily be simulated by modern non-linear structural response is not taken into
analysis codes which are able to consider geometrically account because of the exponential increase in the
non-linear structural behaviour. The procedure is complexity of the problem. Considerable research is
powerful [1] but has some drawbacks: for example, it currently underway to include these phenomena.
is not possible to find a compromise when different The methods of structural optimization have
load cases are dominant, or to consider criteria for the reached a remarkable level [6], and they have been
genesis of shapes not based on elastic deformations. used as design tools to improve structural quality in
many industrial applications, especially in aircraft [7]
and automotive as well as in other mechanical
2. Structural Optimization - A General industries. Their potential for alternative design in civil
engineering has not yet been fully exploited. This lack
Tool of Shape Design
of practical acceptance may be explained by some
frustrating experiences in the 1960s and early 1970s.
Computational methods which offer a more general
But progress in computational methods and hardware
approach to shell design than the principle of inverting
since those times is also considerable, and this has
hanging models are the procedures of structural
rekindled interest in structural optimization.
optimization. Their foundation is a strict separation
A typical problem of structural optimization is
of structural geometry, mechanical behaviour and the
characterized by an objective f(x) and constraints 9(x)
design objective which is responsible for the generation
and h(x) which are non-linear functions of the
procedure. The idea of form finding by these methods
optimization variables x. It can be stated as:
is therefore very much related to the obvious
engineering approach:
minimize: f(x)
(1) choose a shape;
(2) evaluate the structural behaviour according to the subject to: hj(x) = 0; j = 1,..., me
(1)
given load cases and support conditions, by finite gj(x) ~ 0; j -= m e -t- l . . . . , m
element methods, for example;
(3) check stresses, displacements, buckling load and x L~<x~<xu; x~R"
other safety requirements;
(4) compare the quality of the design with the chosen where xL and x U are lower and upper bounds on the
optimality criteria; variables, respectively.
(5) if necessary, propose a new and better design by It is now possible to consider, for typical objective
means of sensitivity analysis, and repeat the functions which are used in the sense of the design
process. rules above discussed, how to improve the structural
behaviour substantially. For example, strain energy fE
This procedure is absolutely different from the is an appropriate objective to achieve shapes which
hanging model and other related principles where the act in a membrane state of stresses, i.e., tension and
generating rule itself (i.e., mechanical reactions to compression but no bending:
given loads) is already the criterion for optimality.
Nevertheless, the same results can be achieved if
equivalent objectives and load conditions are chosen. fE = ~ ere d V (2)
Because of their general formulation, methods of
structural optimization can tackle problems with
many load conditions, arbitrary design objectives To generate shapes which are able to carry loads
Form Finding of Shells 29

mainly in compression, stress levelling fs can be used [11], which are able to handle all these problems, or
E8]: approximation methods such as the method of moving
asymptotes [-12] (MMA), which are superior in special
cases like weight minimization. Both methods have
fs = fv (~ - a")2 dV (3) been successfully used in shape optimal design of free
form shells [-8, 9, 13].
where cr, is a prescribed goal of stress. There are other As already mentioned, structural optimization is
objective functions, such as construction cost, weight, understood to be a synthesis of various individual
or natural frequencies, which are interesting in the disciplines [10]: (a) design modelling (CAGD),
shell structure design and often dominate the two (b) structural analysis (e.g., FEM), (c) behaviour-
objectives of natural significance already mentioned sensitivity analysis, (d) mathematical programming
above. Applying the methods of multicriterion optimi- and (e) interactive computer graphics as an important
zation, several, even conflicting, objectives can be additional aid. Expertise in all these fields is necessary
considered simultaneously to obtain an 'optimal' to get satisfactory results. The authors developed their
compromise of structural design. This is done, for own program system CARAT (Computer Aided
example, when strain energies of more than one load Research Analysis Tool) [14, 15] which provides
case have to be minimized (see the example in the designing engineer with coordinated program
subsection 4.2). modules. Design modeller, FE-analysis and optimiza-
Inequality constraints g(x) are taken into account tion algorithms are coordinated from the root of
to impose safety and reliability requirements. Typical development and are integrated in a general design
constraints of this type are stress and displacement procedure based on a unique in-core database which
limits. If the stiffness of a structure is to be maximized allows fast data exchanges without any loss of
for a prescribed structural mass, an equality constraint accuracy.
h(x) is introduced. Otherwise, unrealistically massy
solutions could be obtained, as is usually the case
if external loads dominate the self-weight of the 3. Modelling and Modifying Structural
structure. Shapes
Form finding implies optimization of geometry.
Characteristic optimization variables are therefore Design modelling, as an important part of structural
geometric parameters defining the structural shape. optimization, is the backbone of the whole procedure.
The number of variables can be reduced dramatically The general methods of Computer Aided Geometrical
without loss of generality if CAGD concepts are used Design (CAGD) [16, 17] are the basis of modern
[-9, 10]. By these methods, shapes of free form shells pre-processors to design structural geometries in two
can be described by the coordinates of a few so-called and three dimensions. Shapes are approximated
'design-nodes' which can be chosen as variables. In piecewise by 'design patches'. Within each design path,
addition, the thickness variation can be optimized the resulting shape r is parametrized in terms of shape
where discrete thicknesses at design-nodes are taken functions ~i, patch parameters u, v, w, and design
as variables. The use of these methods in modelling nodes rdi which describe the location of the patch in
and modifying surfaces will be described in more detail space:
in the next section.
Form finding of shells results in non-linear optimi- r(u, v, w) = ~ ~i(u, v, w)r~i (4)
i=1
zation problems which exhibit many different proper-
ties of mathematical optimization. Depending on the Many different shape functions are available, e.g.
objective (strain energy, weight, etc.), the constraints Lagrangian interpolation, Coons' transfinite inter-
(equality, non-equality) and their combinations, the polation, B6zier and B-spline approximations. De-
optimization problems vary from totally uncon- pending on their formulation, a huge variety of shapes
strained (often stress levelling as objective) via little can be described without severe restrictions on the
constrained but with equality constraints (strain manifold solutions. In shape design of free form shells,
energy minimization with fixed structural mass) to one-dimensional cubic B4zier and B-splines, and
highly constrained problems like weight minimization, two-dimensional bi-cubic B6zier patches (Fig. 1)
which tends to reduce mass until the limit of material appear to be superior to others. This is because only
resistance is reached. Therefore, only robust and the corner nodes of those design patches interpolate
sophisticated methods can be recommended, such as the resulting shape. The remaining inner nodes only
SQP techniques (sequential quadratic programming) approximate the shape, which yields no differences in
30 K.-U. Bletzingerand E. Ramm

continuity d e s i g n nodes

a) four B6zier patchesdefininga plate

continuity
patches

roo
Fig. 1. B6zierpatch.
b) shift designnodes
the results compared with equivalent Lagrange
interpolation schemes, but allows the construction of
continuity conditions between adjacent design patches
if composite surfaces are to be defined. This is
demonstrated by an example which also reflects the
interactive capabilities of CARAT. A plate is defined
by four 16-noded B6zier patches as shown in Fig. 2.
These elements are connected in such a way that the
shape generated is continuous in slopes across the
common edges of adjacent patches. To obtain this,
corresponding design nodes of the involved patches
have to remain on a common line during all c) generatedcontinousshape
subsequent shape modifications. The same rule holds Fig. 2. Interactive surface modification; continuity patches con-
for the second dimension which leads to linear necting four B6zierpatches.
dependencies between, at most, nine design nodes.
These topological relations are formulated in super- plan of a free form shell described by a total of 16
imposed 'continuity patches' [9]. They are generated B6zier patches, and the generated shape modelled by
automatically and preserved during manual user 8-noded isoparametric shell elements. The generated
interactions and shape optimization (Fig. 2(a), (b), (c)). model of a sea urchin shell is shown in Fig. 4, and
Figure 2 shows different types of continuity patches was the subject of a biomechanical study together with
depending on whether they are connecting two or four biologists [18]. The stiffening effects of the wrinkles
design patches, or if they are defined at an isolated were the main objective of this investigation. The shape
corner. In all cases, four nodes are independent and was interactively adjusted to measured data using
control the locations of the remaining nodes, leading B-splines which were linearly blended in cylindrical
to a reduction of geometrical degree of freedom which coordinates (Coons' interpolation).
is very welcome in structural optimization to stabilize Another important fact which must be considered
the procedure. when 'membrane' shapes are to be determined is the
The idea of continuity patches is very helpful in generation and modification of corresponding support
interactive design of free form shells which can serve conditions. CARAT supplies rules to generate support
as initial shapes for subsequent optimization runs or conditions which are tangential and normal to surfaces
as valuable interactive pre-processor tools for input or edges, respectively, and remain so during the whole
preparation of complex shapes. Figure 3 shows the form finding process.
Form Finding of Shells 31

Fig. 3. Free form shell using

16 B~zier patches. ground plan generated shape

4. Examples
4.1. Bi-parabolie Roof Shell

This example is used to demonstrate the effects of

different objective functions and the variety of shapes
which can be generated by using only two variables.
The structural situation is shown in Fig. 5. A shell of
rectangular plan (b = 6 m , l = 12m) and uniform
constant thickness (t = 0.05m) is supported by
diaphragms at the smaller edges. The shape is
generated by four B6zier patches. The design nodes
are linked (a) to preserve double symmetry and (b) to
describe a bi-parabolic surface which can be controlled
by two vertical coordinates as indicated. In the initial
design, both coordinates are set to sl/2 = 3 m ,
Fig. 4. Finite element model of a sea urchin shell. describing a cylindrical shell. The structure is loaded

linked design node

parabolic shape functions \ diaphragm
variable

~---'--b = 6m-----~ "~t ~-----~1 = 12mr

Fig. 5. Parabolic roof shell:
problem statement. material properties: E = 30,000 MPa, v = 0.2 (concrete)
32 K,-U. Bletzinger and E. Ramm

optimizer: SQP

intial values:
Sl = 3m
s2 = 3m
a) initialshape
1200 optimal values:

Z~ s2* = 3.12m
1080

~ 960

840
e~

720

600
0 1 2 3 4 5
iterations : b) optimal shape, minimization of strain energy

360, optimal values:

Sl* 0.90m
=

~ 340 i s2* = 1-97m

280 k
~ 260, ~ .
240:
012345678
iterations c) optimal shape, stress leveling

36 i t optimal values:
sl* = 1.64m
34 s2* = 1.34m

~ 32
.~ 30

28

26

24
0123456 Fig. 6. Parabolic roof shell; initial design and optimal
iterations d) optimal shape, weight minimi~tion
solutions.
Form Finding of Shells 33

load: load cases: oround olan:

,i::::::::::::::::::::::::
:i i i i ili~i :i~:~i:i:i i ~il, '~
snow: ~!:!i!i~:.i!i!i~:.~!~!iiii~i:i~iiii!~i!iii!:i:~:~:!!i!i~:i!i!ii!i~i~i!!i!!%i|P

p=5~
nl 2
I::iii]i:i~iii~i~i~iii!i]!iii:ii!i!:~i:i]
g
and [i!!i!!!i!i!!ii!i!i!i~-i!i!:ii!ii~iiii!il] ~iiiiiiiiiiiiiiiii~,, lOm
dead load Ii:i:!:i:i:!:~:i:!:~:!:!:i:i:!:]:i:] i

1,9-----------I~,~-------I~
lOm lOm

t o = 8.1cm t* = 8.0cm

,~ lOm =,... lOm =,

Fig. 7. Free form shell subjected

to two load cases. initial shape optimal shape

by uniform vertical load p = 5 kN/m 2 (snow). Support be avoided totally because of the simple shape function
conditions are fixed hinges. Because of the symmetry and the rectangular plan of the structure. It is
of loads and the structure, only one-quarter of the remarkable that the diaphragm - although possible -
shell has to be analyzed. This was done by 72 8-noded does not disappear. If it vanishes, the resulting shape
isoparametric shell elements which are 2 x 2 reduced has a horizontal tangent plane at the corner, leading
integrated. The material properties of concrete are to negative curvature and increased bending.
chosen as E = 3.0 x 107 kN/m 2, v = 0.2. By using 'weight' as ~he objective function, any
In a first optimization run, strain energy was chosen shape between the 'minimal surface' and a plate can
as the objective function without a stress constraint, be determined, which is forced by the additional
assuming the structure is sufficiently reinforced to constraints on stresses and displacements. Figure 6(d)
resist high tension forces. The resulting shape (Fig. shows a result obtained with constraints on v. Mises
6(b)) is an anticlastic surface (HP), very similar to a effective stresses are not allowed to exceed an
minimal surface which acts almost like a membrane arbitrarily chosen value of a m = 400 kN/m 2.
in tension and compression. Since the structural
thickness is fixed, the result is alternatively restricted 4.2. Non-regular Shell with Two Loading Cases
by an upper bound (6 m) on variable s 1.
To get a more suitable design for concrete, the A concrete shell of quadratic plan and uni-
objective 'stress levelling' was used to reduce tension form thickness t ( E = 3.4 x 107kN/m 2, v = 0 . 2 ,
stresses in the lower fibres of the structure, which are 7 = 25 kN/m 3) has been investigated with respect to
caused by interactions of normal forces and bending strain energy minimization (Fig. 7). The total material
moments. A goal stress of o a ~--- - - 1 0 0 kN/m 2 was volume is kept constant during shape optimization;
prescribed. The optimal structure (Fig. 6(c)) is a no further constraints are introduced. Two combina-
synclastic shape (EP) where the area of tension in the tions of dead load and uniform live load p = 5 kN/m 2,
lower fibres is reduced to a minimum. Tension cannot which acts either on the entire shell or on one-half of
34 K.-U. Bletzinger and E. R a m m

side views: A-A . B-B

lore

~r. i'
7.3m 18rn 10m
I,dl L ~
" 141 ~-4
I~
r.,,i
r~ 25.3m El
v l

elan with 23 desien variables s,: continuity patches

IS sl
'" ....... " Ss l 0 variable height
$1 '11
~ : ~~l " fixed hegiht
~ls| X linked height

$23 ~'22 - -$21 -$20

Fig. 8. Geometric model of free
Is form shell.

it, are considered as different load cases. The sum of the optimal shape again shows a clear negative
the non-weighted individual strain energies defines a curvature near the free boundary (Fig. 9) as was
compromise objective. The initial geometry with a demonstrated by Isler in hanging model experiments
thickness of t = 8.1 cm was evaluated in a preliminary and the related beautiful shell structures.
design process. The shape is defined by four Brzier
patches. Design variables are linked and continuity
patches are introduced to preserve symmetry and
continuity of the structure. The optimal shape exhibits 5. C o n c l u s i o n s
a maximal stiffness. Since a pure membrane state is
otherwise not possible, the shell needs a boundary Shape-sensitive structures like shells require high
stiffened by a distinct negative curvature. quality design, analysis and manufacturing. Therefore,
the main objective is a membrane-oriented design,
avoiding as far as possible bending, and also buckling
phenomena.
4.3. Tennis Hall The present paper presents the methods of structural
optimization as general computational tools to find
The shell in Fig. 8 is related to the reinforced concrete the shape of shells subjected to different load cases
tennis hall designed by H. Isler [3]. The material data and certain boundary conditions. The key to the
are E = 3.0 x 107 kN/m 2, v = 0.2, ~ = 25 kN/m 3. One approach is a flexible design modeller which allows
load case of dead load plus uniform live load, the generation and modification of even complex
25 kN/m 2, is considered. One-quarter of the shell is shapes by only a few design parameters. Different
idealized by 4 Brzier elements linked by continuity design objectives can be applied. It was shown that
patches and by 23 vertical nodal coordinates S~ as useful information on optimal shells of preferable
design variables (Fig. 8). A total of 126 reduced stress state can be achieved by the objectives of
integrated 8-node shell elements are used for a linear 'minimal strain energy' and 'stress levelling'. Although
structural analysis. Strain energy is minimized. It can the static analysis is assumed to be linear in this
be recognized that the shape near the free edges is investigation, the entire optimization procedure is
sensitive and might even result in a sharp local highly non-linear, demanding sophisticated algo-
curvature representing a kind of edge beam. If this is rithms and experienced personnel. Together with
not tolerated by prescribed geometrical constraints, progress in computational sciences and hardware,
Form Finding of Shells 35

[3 = .

Fig. 9. (a) Initial and (b)

optimal shape of free form shell.

structural optimization can become a valuable design analyse mit finiten Elementen, PhD Dissertation, Institut f/Jr
aid for shell structures, which will reduce the planning Baustatik, Universit/it Stuttgart
9. Bletzinger, K.-U. (1990) Formoptimierung yon Fl/iehentrag-
time and the experimental expense.
werken, PhD Dissertation, Institut f/Jr Baustatik, Universit/it
Stuttgart
10. Braibant, V.; Fleury, C. (1986) Shape optimal design and CAD
oriented formulation, Engineering with Computers, 1,193-204
Acknowledgements 11. Schittkowski, K. (1981) The nonlinear programming method
of Wilson, Han and Powell with an augmented Lagrangian type
This work is part of the research project SFB 230'Natural Structures
line search function, Numerische Mathematik, 38, 83-114
Light Weight Structures in Architecture and Nature' supported
12. Svanberg, K. (1987) The method of moving asymptotes - a new
-

by the German Research Foundation (DFG) at the University of

method for structural optimization, Int. J. Num. Meth. Engng,
Stuttgart. The support is gratefully acknowledged. The authors also
24, 359-373
would like to thank their former research associate and colleague
13. Ramm, E.; Bletzinger, K.-U.; Kimmich, S. (1991) Strategies in
Stefan Kimmich.
shape optimization of free form shells, Nonlinear Computa-
tional Mechanics - State of the Art (Wriggers, P. and Wagner,
W., Editors), Springer-Verlag, Berlin, 163-192
14. Btetzinger, K.-U.; Kimmieh, S.; Ramm, E. (1991) Efficient
References modeling in shape optimal design, Computing Systems in
Engineering, 2, 483-495
1. Ramm, E.; Mehlhorn, G. (1991) On shape finding methods and 15. Kimmich, S.; Ramm, E. (1989) Structural optimization and
ultimate load analyses of reinforced concrete shells, Engineering analysis with program system CARAT, Proc. GAMM -
Structures, 13, 178-198 Seminar, October 5-7, 1988, Siegen, on Discretization Methods
2. Isler, H. (1990) Elegante Modelle - die moderne Form des and Structural Optimization - Procedures and Applications
Schalenbaus, Deutsche Banzeitung, db 7, 62-65 (Eschenauer, H.A. and Thierauf, G., Editors), Lecture Notes in
3. Ramm, E.; Schunck, E. (1986) Heinz Ister - Sehaten, Kr~mer, Engineering, Springer-Verlag, Berlin, 186-193
Stuttgart 16. Brhm, W.; Farin, G.; Kahmann, J. (1984) A survey of curve
4. Haftka, R.T.; Grandhi, R.V. (1986) Structural shape optimiza- and surface methods in CAGD, Comp. Aided Geom. Des., 1,
tion - a survey, Comp. Meth. Appl. Mech. Engng, 57, 91-106 1-60
5. Kruzelecki, J.; Zyczkowski, M. (1985) Optimal structural design 17. Faux, I.D.; Pratt, M.J. (1979) Computational Geometry for
of shells - a survey, SM Archives, 10, 101-170 Design and Manufacture, Ellis Horwood, Chichester
6. Schmit, L.A. (1981) Structural synthesis - its genesis and 18. Philippi, U.; Nachtigall, W. (1991) Constructional morphology
development, AIAA Journal, 19, 1249-1263 of sea urchin tests, to be presented at the 2nd International
7. Petiau, C. (1991) Structural optimization of aircraft, Thin Symposium on Natural Structures - Principles, Strategies, and
Walled Structures, 11, 43-64 Models in Architecture and Nature, SFB 230, Stuttgart, October
8. Kimmich, S. (1990) Strukturoptimierung und Sensibilitfits- 1-4