You are on page 1of 15

Comput. Methods Appl. Mech. Engrg.

194 (2005) 34383452


www.elsevier.com/locate/cma

Computational methods for form nding and optimization


of shells and membranes
Kai-Uwe Bletzinger *, Roland Wuchner, Ferna Daoud, Natalia Camprub
Lehrstuhl f ur Statik, Technische Universit
at M
unchen, Arcisstr. 21, D-80290 M
unchen, Germany
Received 20 January 2004; received in revised form 23 August 2004; accepted 6 December 2004

Abstract
Free form shells optimized for stiness under given loading and membrane structures act in a pure membrane state
of stresses, either because bending is minimized or not even present by denition. Physical experiments as soap lms and
hanging models have been used since centuries to generate optimal shapes of membranes in tension and shells in compression. The paper presents numerical methods to simulate the physical experiments as well as how they can be merged
among each other and with the most general technology of structural optimization. The combined approach represents
the combined power of each technique. Several examples illustrate the methods and typical applications.
2005 Published by Elsevier B.V.
Keywords: Form nding; Shape optimization; Free form shell; Hanging model; Minimal surface

1. Introduction
Lightweight structures as shells and membranes are dened by the optimal use of material to carry external loads or pre-stress. Material is used optimally within a structural member if the member is subjected to
membrane forces rather than bending. An important objective of a procedure to determine layout and
shape of a lightweight structure is, therefore, to minimize bending or more general, to minimize the strain
energy rather than structural weight as the term lightweight may imply [1].
Two dierent lines of research have developed which deal with the generation of structural shapes: the
elds of form nding and structural optimization, respectively. The methods of form nding are

Corresponding author. Fax: +49 7031 874521.


E-mail address: kub@bv.tum.de (K.-U. Bletzinger).

0045-7825/$ - see front matter 2005 Published by Elsevier B.V.


doi:10.1016/j.cma.2004.12.026

K.-U. Bletzinger et al. / Comput. Methods Appl. Mech. Engrg. 194 (2005) 34383452

3439

usually restricted to tensile structures (cables and membranes) whereas the methods of structural optimization are far more general and can be applied to any kind of structure [24]. The dierences of the two
approaches, however, are not only the level of specialization but also their aims. Form nding methods
are designed to determine structural shape from an inverse formulation of equilibrium. Again, two dierent methods are known: the soap lm analogy for the generation of structures acting in pure tension
which are related to minimal surfaces, and hanging models to generate structures in compression by inversion of tensile structures. These methods are well known from tradition and are established since centuries by physical experiments. Today, these methods can be eectively simulated by nite element
methods as demonstrated in the sequel. In the context of structural optimization other criteria or even a
combination of several criteria can be chosen to dene the shape with respect to the special problem
under consideration. The most general approach of structural optimization, however, has to be paid by
numerical expense and is therefore often restricted to rather few shape degrees of freedom. On the other
hand, numerical simulations of form nding methods are comparatively robust, are suited for complex
shapes but are designed for very special applications. The paper will give an overview on the both
approaches and will show a strategy about how to combine them with respect to eectiveness and mechanical relevance.

2. Form nding methods


2.1. Pre-stressed membranessoap lm analogy
Pre-stressed membranes in tension are optimal in two respects. First, in the case of uniform surface stresses they reect surfaces of minimal area or minimal weight, e.g. minimal surfaces [5], and, secondly, structures of optimal material usage since bending is omitted a priori by denition. Experimentally minimal
surfaces can be realized by soap lms (soap lm analogy) [1]. Numerical simulation results in a complicated inverse problem why there exist many dierent solution approaches, e.g. [615], further references
can be found in [13].
2.2. Virtual work of a surface stress eld
Consider a stress eld r which acts tangential to a surface and which is assumed to be in self-equilibrium
(Fig. 1). The governing equation is the principle of virtual work

z
y

x
Fig. 1. Tangential surface stress eld.

3440

K.-U. Bletzinger et al. / Comput. Methods Appl. Mech. Engrg. 194 (2005) 34383452
reference
configuration
2

deformation
1

G1

G2

g2
P

u( , )

X( 1,2)

Z, z
1

x( , )
Y, y

g1
actual
configuration

X, x
Fig. 2. Deformation of surface.

dw t

Z
r:
a

odu
da t
ox

r : du;x da 0

which is used to nd the unknown shape x of equilibrium. r is the prescribed Cauchy stress tensor which
acts on the surface, du,x is the derivative of the virtual displacement with respect to the geometry x of the
surface in equilibrium. The thickness of the membrane is denoted by t. It is comparatively thin and assumed
to be constant during deformation.
The standard notation of continuum mechanics to describe the deformation of bodies denes the displacement eld u as the dierence between the surface in equilibrium x and any reference surface X,
u = x  X, Fig. 2. Following the chain rule of dierentiation the derivative of the virtual displacement
du with respect to x can be explained by use of the real and virtual deformation gradients F and dF
 
odu odx oX
ox
du;x


d
2
 F 1 dF  F 1 :
ox
oX ox
oX
With (2) the virtual work equation (1) can be written as
Z
Z
Z
T
dw t r : du;x da t
det Fr  F : dFdA t F  S : dF dA 0;
a

S is the second KirchhoPiola stress tensor.


It is related to r by
S det FF 1  r  F T :

The mapping of the area dA from the reference to the equilibrium state da is dened by
da det F dA:

At this point (3) reects alternative formulations of the state of equilibrium. By the special relation (4) of
the stress tensors, r reects the deformation from the reference conguration to the state of equilibrium and
a given tensor S or vice versa. Or, in other terms, if r denes a uniform surface stress then S is non-uniform
unless the reference geometry X coincides with the equilibrium geometry x. Note that all stress components
normal to the surface are zero, i.e.
ri3 r3i S i3 S 3i 0:

K.-U. Bletzinger et al. / Comput. Methods Appl. Mech. Engrg. 194 (2005) 34383452

3441

In the case of a minimal surface the Cauchy stress tensor is a scalar multiple of the unit tensor I, r = sI and
(3) reduces to
Z
Z
det FF T : dF dA 0
7
dw st I : du;x da st
a

which is the governing equation of a surface of minimal area content within a given boundary [5]. The analogy between this geometrical explanation and the initial mechanical statement of equilibrium is the well
known soap lm analogy.
From the geometrical analogy it is obvious, that the variation of shape must have at least a normal component to the surface. Otherwise, the shape maps to itself and the area content does not change. In terms of
a nite element discretization tangential displacement degrees of freedom are irrelevant and lead to singular
structural response.
2.3. Regularization
The conclusion of the previous section implies that a solution of (1) by a nite element technique allows
for only one displacement degree of freedom at each node which remains normal to the surface during the
form nding procedure, Fig. 3.
However, for application to tents the boundaries are exible cables and the surface nodes must move
tangentially also to ensure a proper mesh, Fig. 4.
That means, a generally applicable discretization technique must use three displacement degrees of freedom at each node and, as a consequence, must be enhanced by some regularization methodology to circumvent the above mentioned singularity problems.
There exist several regularization techniques. The method of dynamic relaxation adds articial inertia
terms to the governing equations (3). Other ideas suggest neglecting certain terms after linearization to

diagonal section

move directions of nodes normal to surface

Fig. 3. Form nding of a HP-like surface with rigid boundaries.

initial shape

Nodes must move in all three spatial


directions to guarantee mesh quality

final shape

Fig. 4. Form nding of a HP-like surface with exible boundaries, top view.

3442

K.-U. Bletzinger et al. / Comput. Methods Appl. Mech. Engrg. 194 (2005) 34383452

prevent the stiness matrix from rank deciency. A further alternative is based on an intelligent mixture of
congurations.
Starting from (3) one can introduce a continuation factor k and write
Z
Z
dw kt
det Fr  F T : dF dA 1  kt F  S : dF dA 0:
8
A

Nothing is changed so far.


Next, we assume that instead of applying (4) the second PiolaKircho stresses S are dened identically
as the Cauchy stresses r, e.g. in the case of uniform surface stresses
r S sI:

Now, (8) is modied to be


Z
dwk st k det FF T 1  kF : dF dA 0;

1 < k 6 0:

10

If k is chosen properly the second term stabilizes the original expression and allows for standard nite element discretization and solution. The modication has the convincing property that it disappears at the
solution of a minimal surface. Then the deformation gradient reduces to be the identity tensor. It should
be mentioned that this regularization technique can be generalized for other situations where tangential
movements of nodes must be stabilized.
2.4. The updated reference strategy
The modied and stabilized expression (10) is non-linear with respect to the nal geometry x and must be
solved iteratively applying the NewtonRaphson method. The result deviates from the true solution
depending on the choice for k. The procedure might be repeated with a modied choice of k closer to
one until one gets close enough to the solution [15].
~ of the modied system
Alternatively, k can be chosen as a xed number. Then the resulting geometry x
(10) can be used to update the reference geometry X for the next iterations. By repeatedly updating the reference geometry the procedure safely and robustly converges to the nal solution, Fig. 5. Even k = 0 might
be chosen. Then the problem reduces to be linear and can be solved within one iteration. However, more

Choose

external loop,
update of
reference geometry,
eventually also of
continuation factor

Solve (10)

Inner loop,
Iterative solution of (10)
one linear step if = 0

Fig. 5. Flow scheme of the updated reference strategy.

K.-U. Bletzinger et al. / Comput. Methods Appl. Mech. Engrg. 194 (2005) 34383452

3443

Fig. 6. Updating reference strategy, adaptation of k: Steps 0 to 3; updating geometry: Step 4.

Fig. 7. Membrane structure of constant surface stress (minimal surface).

external loops are necessary to update the reference geometry. The method converges very fast and robust,
Fig. 6. Minimal surfaces can be found in a very reliable manner, Fig. 7.
2.5. Anisotropic pre-stress
There exist several situations where ideal uniform surface stresses are not possible if the structure is realized by anisotropic textile material or is loaded by additional load cases as wind and or snow. Then anisotropic pre-stresses must be considered which result in shapes which dier from the ideal minimal surface.
The updated reference method is readily modied for this new situation if the stress tensors r and S in the
governing equation (8) are formulated accordingly [16]. With respect to the production of membrane structures by the composition of initially at strips we dene the stress eld in terms of an additional, developable surface, denoted as reference strip. For simplicity of presentation we take a plane assuming rather

3444

K.-U. Bletzinger et al. / Comput. Methods Appl. Mech. Engrg. 194 (2005) 34383452
tangent plane

coordinate lines on
reference surface

T 3=G 3

G2

G1
T1

plane of
projection rays
x

f3

f2

F
f1
e3
reference strip

e2
e1

Fig. 8. Projection of pre-stress on membrane surface.

at shapes as solution. Fig. 8 shows part of the reference membrane surface and below the additional reference strip. Vectors f1 and f2 dene the directions of prescribed pre-stress on the strip which are initially
dened as problem input. The direction T1 of pre-stress on the reference surface is dened by the intersection line of tangential and projection plane (Fig. 8). Direction T2 is determined to be orthogonal to T1 and
the surface normal G3.
The following chain of stress transformation is dened:
Pre-stress on the strip:
r
ab f a  f b :
r

11

Pre-stress on reference surface:


ab T a  T b S ab G a  G b ;
Sr

12

where the components of S appear to be


cd G a  T c T d  G b :
S ab r

13

The pre-stress tensor r acting on the equilibrium surface can be determined by an analogous procedure,
dening vectors (t1, t2) dened by the intersection of the projection plane and the tangential plane, now,
at the equilibrium surface. The Cauchy stress components rab with respect to the basis (g1, g2) are dened
as
cd g a  tc td  g b :
rab r

14

The numerical eort of consistent linearization with respect to the discretization parameters for a general
choice of k can be avoided if the URS scheme with continuation factor k = 0 is chosen.
Pre-stress of dierent magnitudes and orientations in adjacent strips are usually not in equilibrium along
the common seam. In these cases the magnitude of pre-stress must be corrected during form nding. This
can be done without altering the general layout of the method.

K.-U. Bletzinger et al. / Comput. Methods Appl. Mech. Engrg. 194 (2005) 34383452

3445

Fig. 9. Snow loaded 5-point tent.

A typical application is shown in Fig. 9 where an anisotropic pre-stress is used to modify the shape of
tents slightly to prevent water ponds under additional snow loading.
2.6. Hanging models
The hanging chain and its inverse is one of the oldest methods which are known to generate the shape of
an arch which is free of bending, subjected only to compressive axial force. The method has been used
intensively during the centuries, e.g. by Antoni Gaud, to mention one well known name among all the
others. Extended to two directions to dene the shape of shells the hanging model concept has been brought
to perfection by Isler [17].
The goal of hanging models is to perform the transition from a bending structure to a membrane structure thus minimizing the bending part of strain energy. The optimal shape generated by hanging models is
the result of a mechanical deformation for one load case. It is dominated by the size of the undeformed
original piece of material (chain, cloth) which has been used for the experiment. In two dimensions for
the generation of shells the choice of the initial shape is also critical with respect to wrinkles and folds which
may develop during the deformation. The choice of material, isotropic or anisotropic, also aects the result.
The implicit interaction of initial and optimal shape through the mechanical deformation yields that the
design variables are not at all obvious. It might be very complex to identify and to vary them. The variety
of possible solutions is almost innite, because a further classication of structural quality besides the absence of bending is not part of the method. And, stability eects cannot be considered by hanging models.
The numerical model is derived from the virtual work equation (3), again allowing for large displacements and considering an external load p

3446

K.-U. Bletzinger et al. / Comput. Methods Appl. Mech. Engrg. 194 (2005) 34383452

Fig. 10. Hanging models and inverted shapes.

dw t

F  S : dF  pdudA t

S : dE  pdudA 0:

15

The isotropic St. VenantKirchho material law is used


S k tr EI 2lE

16

which denes an elastic relation between the GreenLagrange strain tensor E and S where k and l are the
Lame constants. The kinematic assumptions for thin membranes are applied to derive a standard nite element discretization.
The hanging model experiment is done numerically starting from an undeformed reference structure, e.g.
a plane membrane, applying a shape generating load case, usually a distributed surface load like self weight
or snow. In general, the equilibrium shape is found after large displacements and large strains as can be
seen by the following examples. Nevertheless, the St. VenantKirchho law can be used which is restricted
to small strains. The hanging model is used only for shape generation and not for the validation of the nal
structure after inversion and application of additional loading, Fig. 10.

3. Structural optimization
The methods of structural optimization are the most general optimization tools, for an overview see e.g.
[18] They combine highly specialized methods from dierent disciplines as there are computer aided geometrical design, computational mechanics, and non-linear mathematical programming, Fig. 11. Together
they dene a modular tool box which can be applied for the denition of a very general optimization
problem:
f s; u ! min;
s

hi s; u 0; i 1; . . . ; no: of equality constraints;


gk s; u 6 0; k 1; . . . ; no: of inequality constraints:

17

The optimization variables s are the spatial control node positions of the design model which in turn dene
shape and FE mesh, Fig. 12. In terms of Computer Aided Geometrical Design the design model is dened
as tensor product surfaces, e.g. Bezier-spline and B-spline patches. The example shows an initially spherical
shell which is approximated by 6 Bezier patches [3,4].
The analysis model denes type of structure, material and loading conditions, e.g. shell, concrete, self
weight, respectively. The structural behavior is a function of the state variables u. The optimum is dened
by the objective function f and the constraints g and h which are functions of the design s and the state u. As

K.-U. Bletzinger et al. / Comput. Methods Appl. Mech. Engrg. 194 (2005) 34383452

3447

Design
Model

CARAT
Optimization
Model

Analysis
Model

Fig. 11. The models of structural optimization.

Fig. 12. Shape optimization of a shell by means of structural optimization.

objective and constraints all structural properties and response quantities can be used which dene the
considered optimization problem, e.g. weight, mass, strain energy, absolute stresses, stress dierences,

3448

K.-U. Bletzinger et al. / Comput. Methods Appl. Mech. Engrg. 194 (2005) 34383452

Fig. 13. Periodically arranged concrete pillars.

displacements, eigenfrequencies, etc. Finally, the procedure is controlled by the optimizer, where several
strategies can be used, e.g. mathematical programming or optimality criteria methods.
Taking minimum strain energy as objective whilst restricting mass the optimal shape is in principle identical with the related hanging model and acts mainly as a membrane and is almost free of bending.

4. Merging form nding and hanging models


Soap lm analogy and hanging model experiment can be combined to nd structures of maximum stiness in situations which are characterized by a combination of distributed surfaces loads and pre-stress. This
is possible because doubly curved shells are able to carry dierent loads by pure membrane action. Shape
generating and design load cases can be dierent. The governing equations are the sum of (3) and (15)
Z
Z
dw t F  S S pre : dF  pdudA t S S pre : dE  pdudA 0:
18
A

The pre-stresses belonging to the soap lm part are nominated by Spre and S are related to the hanging
model and reect elastic deformations. Depending on the relation between both parts in (18) regularization
by use of the updated reference strategy may be necessary.
The approach was successfully applied for a real application which is shown in Fig. 13. Concrete pillars
had to be designed to carry heavy loads. The original design was motivated by the inversion of a principal
experiment shown at top of the gure. A closed thread is pulled out of a soap lm to form a singular vertex.
The row below shows on the left side the application of this idea to form periodically arranged pillars. This
shape modeled as a shell and loaded by distributed surface load instead of the shape generating load case
can be further improved by simultaneously considering the soap lm analogy and the hanging model experiment. The critical part of the design is the edge of the inner eyes as can be seen by comparison of the two
results. Here the structure reacts very sensitive to modied shape generating loads by bending response.
Doubly curved inner regions are able to carry several load cases by membrane action.
5. Merging hanging models and structural optimization
The general denition of structural optimization is the reason for the power of the approach but also for
its weakness if applied to special tasks. Since the structural geometry is dened in mathematical terms it

K.-U. Bletzinger et al. / Comput. Methods Appl. Mech. Engrg. 194 (2005) 34383452

3449

Fig. 14. Hanging model experiment as shape generator in a structural optimization loop.

cannot reect anything of the mechanical behavior. This is the case for form nding of free form shells
where it is initially known that the optimum shape will be free of bending. The idea of merging hanging
models with structural optimization is to use a numerical hanging model experiment as shape generator
instead of a CAGD design model, Fig. 14 [19,20]. Doing so, we implicitly search for the optimum within
the class of membrane shapes. We add to the hanging model experiment an additional optimization loop
looking for the best available hanging model. Optimization variables can be all parameters which dene a
hanging model experiment, e.g. the intensity and distribution of the shape generating loading case p. In
addition, we can directly test the quality of a hanging model shape with respect to design load cases q
or for stability. Dening the structural shape x(p) as the deection of the hanging model, i.e. as a function
of the load generating load parameters p, the optimization problem states as
f xp; u ! min;
gj xp; u 6 0;

j 1 . . . no: of ineq: constr:;

hj xp; u 0;

j 1 . . . no: of eq: constr:

19

The load parameters p are the optimization variables. For the following examples we took the intensity of a
given load distribution, e.g. the number of optimization variables was reduced to one only. Chargin et al.
reported about a similar method taking also the deformation of statical load cases as shape vectors in the
context of structural optimization [21].
Fig. 15 shows the results of a simple, principal example and compares the mentioned approach with classical shape optimization. The shape of a shell is to be optimized with respect to minimal strain energy. In
the context of classical structural optimization the design model is dened as a patch of four Bezierelements. Considering geometrical continuity and symmetry ve independent shape variables remain which
are the vertical coordinates of design nodes as indicated in the gure. In the middle of the gure the hanging
model of a vertically loaded membrane as alternative design model is shown. The independent shape

3450

K.-U. Bletzinger et al. / Comput. Methods Appl. Mech. Engrg. 194 (2005) 34383452

Fig. 15. Comparison of conventional structural optimization and the merged approach.

variable is the intensity of the shape generating load p. The structural load case q is chosen to be the negative of p. The thickness of the structure is constant and the same for both approaches. The bottom of the
gure gives the optimal results for both alternatives. Both structures act in a membrane state of stresses and
have negative curvature at the free edges. The classical optimum gives the better optimal value of the
objective strain energy p. That reects the fact that the classical approach makes use of the larger number
of degrees of freedom. However, the merged result appears to be more pleasing and natural because the
shape has higher order continuity properties than the alternative. The classical optimum is the result
of minimizing bending whereas the merged optimum reects the best available shape which is generated

conventional result:
load path along edge beam

result of merged method:


load path along diagonal arches

Fig. 16. Load carrying behavior of dierent membrane solutions.

K.-U. Bletzinger et al. / Comput. Methods Appl. Mech. Engrg. 194 (2005) 34383452

3451

Fig. 17. Combining hanging models and structural optimization: two further examples.

by p and loaded by q. A more shallow shape would exhibit higher stresses, a steeper shape would have a
larger volume. Both would give larger values of strain energy. If compared to a plane plate pure bending
strain energy of normalized value 75,084 both alternative results are considerably better solutions.
The optimal structures dier considerably in the load carrying behavior, Fig. 16. Because of the high
curvature along the edges of the conventionally optimized structure the edges act like sti beams. As
a consequence, loads are transferred from the inner domain rst to the edge and then down to the support.
An orthogonal grid of arches is formed. The resulting structure of the merged approach behaves dierent.
The edges remain to be soft. Loads are, therefore, transferred along an internal diagonal pair of arches
down to the supports. Interestingly, both structural alternatives had been realized by engineers. They represent two local optima of shells which are optimized for stiness. Fig. 17 shows two other examples where
the merged methods had been applied.

6. Conclusions
Free form shells and membranes are considered to be optimal with regard to structural behavior if they
act in a pure state of membrane internal forces. This mechanical property is exploited by traditional physical experiments as the hanging model and the soap lm analogy. Very powerful numerical techniques simulate these experiments and allow generating even complex shapes by a simple and robust methodology. If
they are combined or even merged with the most general method of mathematically based structural optimization procedures are dened which combine the power of each single technique. Various examples show
the success. Actual research deals with the generalization of the demonstrated techniques for other applications as shown here.

Acknowledgement
Financial support by the German Ministry for Education and Science is gratefully acknowledged.

References
[1] F. Otto, B. Rasch, Finding Form, Deutscher Werkbund Bayern, Edition A. Menges, 1995.
[2] E. Ramm, Shape nding methods of shells, in: International IASS Symposium on Spatial Structures at the Turn of the Millenium,
Copenhagen, 1991, pp. 5967.
[3] (a) E. Ramm, K.-U. Bletzinger, R. Reitinger, Shape optimization of shell structures, IASS Bull. 34 (1993) 103121;
lements Finis 2 (1993) 377398.
(b) E. Ramm, K.-U. Bletzinger, R. Reitinger, Shape optimization of shell structures, Rev. Eur. E
[4] K.-U. Bletzinger, E. Ramm, Structural optimization and form nding of light weight structures, Comput. Struct. 79 (2001) 2053
2062.
[5] S. Hildebrandt, A. Tromba, Mathematics and Optimal Form, Scientic American Library, 1985.
[6] T. Suzuki, Y. Hangai, Shape analysis of minimal surface by the nite element method, in: International IASS Symposium on
Spatial Structures at the Turn of the Millenium, Copenhagen, 1991, pp. 103110.

3452

K.-U. Bletzinger et al. / Comput. Methods Appl. Mech. Engrg. 194 (2005) 34383452

[7] K. Linkwitz, H.-J. Schek, Einige Bemerkungen zur Berechnung von vorgespannten Seilnetzkonstruktionen, Ingenieur-Archiv.
40 (1971) 145158.
[8] E. Haug, G.H. Powell, Finite element analysis of nonlinear membrane structures, Report UCSESM 72-7, University of California
at Berkeley, 1972.
[9] R. Haber, J. Abel, Initial equilibrium solution methods for cable reinforced membranes, Comput. Methods Appl. Mech. Engrg.
30 (1982) 263284.
[10] M. Barnes, Form and stress engineering of tension structures, Struct. Eng. Rev. 6 (1994) 175202.
[11] W.J. Lewis, Tension Structuresform and Behaviour, Thomas Telford Ltd, London, 2003.
[12] B. Maurin, R. Motro, Surface stress density method as a form nding tool for tensile membranes, Engrg. Struct. 20 (1998) 712
719.
[13] R. Motro (Ed.), Int. J. Space Struct., Tensile structures vol. 14 (1999) (special issue).
[14] K. Ishii, Membrane Structures in Japan, SPS Publ. Co., Tokyo, 1995.
[15] K.-U. Bletzinger, E. Ramm, A general nite element approach to the form nding of tensile structures by the updated reference
strategy, Int. J. Space Struct. 14 (1999) 131146.
[16] K.-U. Bletzinger, R. Wuchner, Form nding of anisotropic pre-stressed membrane structures, in: Wall, Bletzinger, Schweizerhof
(Eds.), Trends in Computational Structural Mechanics, CIMNE, Barcelona, 2001, pp. 595603.
[17] H. Isler, Concrete shells derived from experimental shapes, Struct. Engrg. Int. 3/94 (1994) 142147.
[18] R.T. Haftka, Z. Gurdal, Elements of Structural Optimization, third ed., Kluwer, 1991.
[19] K.-U. Bletzinger, Form nding and optimization of tensile structures, in: International IASS-IACM Conference on
Computational Methods for Shell and Spatial Structures, Chania, Greece, 2000.
[20] L. Schiemann, Formndung und Formoptimierung von Schalen mit numerischen Hangemodellen und mathematischer
Programmierung, Diplomarbeit, Uni Karlsruhe, TU Munchen, 2000.
[21] M.K. Chargin, I. Raasch, R. Bruns, D. Deuermeyer, General shape optimization capability, Finite Elements Anal. Des. 7 (4)
(1991) 343354.

You might also like