You are on page 1of 11

6th World Congresses of Structural and Multidisciplinary Optimization

Rio de Janeiro, 30 May - 03 June 2005, Brazil

Fast Reanalysis of Geometrically Nonlinear Problems After Shape


Modifications
Mirko Zeoli, Fred van Keulen and Matthijs Langelaar
Structural Optimization and Computational Mechanics Group,
Department of Precision and Microsystems Engineering,
Delft University of Technology,
Mekelweg 2, 2628 CD, Delft, The Netherlands
email: F.vanKeulen@wbmt.tudelft.nl

Abstract

Fast Reanalysis for geometrically nonlinear problems using either Newton iterations or Modified Newton
iterations starting from a previously obtained solution can be efficient. However, the first iteration within
these iteration processes resembles the semi-analytical design sensitivity. The latter expose severe inaccuracies when applied to shape design variables. The present paper demonstrates that the efficiency of
the geometrically nonlinear reanalysis schemes is less for shape design variables. In addition, techniques
developed for shape design sensitivities are used to improve the efficiency of the geometrically nonlinear
reanalysis schemes. The basic idea is to use accurate sensitivities, e.g. based on a refined semi-analytical
formulation, to improve the first step of the iteration process used in the reanalysis. The standard and
modified reanalysis schemes are studied for a thin shell problem.

Keywords:

Reanalysis, geometrically nonlinear, semi-analytical, design sensitivities, rigid body modes

Introduction

Fast reanalysis is very attractive in situations requiring computational evaluations of many configurations with little or minor differences. Characteristic examples can be found in (structural) optimization
and reliability-based analysis. Typically, configurations must be analyzed that differ in terms of loading
conditions, material properties, dimensions and/or shape. However, the difference between the configurations is often relatively small. The fact that hundreds or even thousands of model evaluations may
be required, prompts the need for methods that capitalize on the fact that the solution for a slightly
different configuration is already known. Particularly for path-independent problems, fast reanalysis
techniques can be formulated relatively easily.
In the context of optimization, fast reanalysis may be crucial if design sensitivities based on global
finite differences are being evaluated. Clearly, the demand for efficient reanalysis techniques becomes
very urgent if the number of design variables at hand increases.
In the finite element context, fast reanalysis techniques have been formulated for linear statics,
eigenvalue problems (natural eigenfrequencies, linearized buckling), geometrically nonlinear transient
analysis, see [18], among many others. In particular for the geometrically nonlinear setting, Kirschs
Combined Approach [1] is mentioned. In the present paper, attention will be on fast reanalysis for
geometrically nonlinear problems. The focus will be put on a single point of the geometrically nonlinear
transient analysis, e.g. the solution for a specified load level. This in contrast to settings requiring the
solution of an entire transient analysis. In addition, the paper is restricted to regular solutions, i.e.
stability points are not included in the present study.
For geometrically nonlinear problems, the easiest approach to fast reanalysis is to use an available
solution as the starting point for the iterative solution process to solve for a slightly different configuration
(or perturbed configuration). Potentially, this has two obvious advantages. Firstly, if the original (or

nominal) configuration required an expensive incremental-iterative solution procedure, the solution of


the perturbed configuration may be restricted to a single solution point. Hence, only a single set of
iterations is required. Secondly, the number of iterations required after a minor configuration change
may be small. That is, the solution for the nominal configuration may be relatively close to the solution
for the perturbed configuration. Consequently, the latter solution may require only few iterations.
Summarizing, in the geometrically nonlinear setting, the reanalysis may benefit from (i) the fact that
only a single solution point is to be regenerated, while (ii) only few iterations are to be expected.
It is noted, that the solution for the perturbed configuration may be obtained via a Newton process,
or alternatively, a modified Newton process may be employed. Another solution has been proposed by
Kirsch with his Combined Approach, see [1, 2, 4, 9, 10].
In case of a very small configuration change, one may anticipate that minor computational efforts
are required in the reanalysis. In the limit, Haftka [11] demonstrated that calculating design sensitivities
via a Global Finite Differences (GFD) scheme with a reanalysis techniques as sketched above, becomes
identical to a semi-analytical approach to design sensitivities. Notice, that for (semi-analytical) sensitivities, the sensitivity equations are linear. In the global finite difference scheme this equivalence is nearly
reached as soon as only a single iteration is needed. (There may still be a minor difference, due to an
updating of the tangent operator, provided a full Newton process is used.) Hence, for very small design
perturbations, reanalysis and semi-analytical design sensitivities become very similar.
Semi-analytical design sensitivities have a very bad reputation if applied to shape design variables.
Depending on the nature of the structure and the loading pattern at hand, severe accuracy problems
emerge, see, e.g. [12, 13]. These accuracy problems have been extensively studied for linear statics,
see [12, 1417]. Remedies include the use of alternative and higher-order finite difference schemes. The
drawbacks of these methods are the additional computational costs and the fact that accuracy problems are only reduced and not eliminated. Rigorous improvements can be obtained using the exact
formulation [18] or differentiation of rigid body modes. The latter crystalized in the so-called refined
semi-analytical method, which has been documented for linear statics, eigenvalue problems, limit points
and geometrically nonlinear problems, see [19, 20].
Recently, it has been demonstrated that errors similar to the ones spoiling the accuracy of standard
semi-analytical derivatives may also deteriorate the efficiency of fast reanalysis techniques. Sofar, the
focus was on fast reanalysis for linear statics and the complications were only studied in the context of
Kirschs Combined Approach (CA). Apart from identifying these complications, also a remedy has
been proposed, see [2].
Prompted by the problems observed for the CA in a linear context and the similarity between semianalytical derivatives and GFD derivatives based on fast reanalysis, fast reanalysis for geometrically
nonlinear cases will be investigated for similar complications. The present study involves both reanalysis
using full Newton as well as Modified Newton iteration schemes. In addition, the use of derivatives to
improve reanalysis will be investigated.
The present paper is organized as follows. In Section 4, governing finite element equations, rigid
body modes, design sensitivities and reanalysis are discussed. Section 5 provides a brief comparison
between reanalysis and semi-analytical design sensitivities. This discussion is entirely restricted to the
geometrically nonlinear setting. A possibility for improvement is proposed in Section 6. Section 7 will
include numerical examples. Conclusions are provided in Section 8.
In the present paper, a formal notation will be used which entirely avoids the precise details of the
finite elements at hand. If quantities refer to the element level a superscript (e) will be used. For
quantities related to the system level, these superscripts are omitted.
Throughout the present paper, matrices will be denoted using bold capitals, e.g. D,S and J, whereas
for vectors bold lower-case characters, e.g. p and q, will be used. Partial derivatives will be shortly
denoted using a comma. To simplify the discussion, only a simple design variable (s) will be considered.

Preliminaries

This section is organized as follows. In Section 4.1, the governing general FE equations will be summarized. The associated rigid body modes are presented in Section 4.2. Discrete design sensitivities are

summarized in Section 4.3. Finally, a fast reanalysis using the full or standard Newton method and a
modified Newton method are described in Section 4.4.

4.1

Governing finite element equations

For every element in a finite element model, a set of nodal degrees of freedom ue can be identified.
Notice that the superscript indicates element level. In addition, for each element a set of generalized
deformations can be defined, which will be denoted e . The energetically conjugated generalized stresses
are e .
At element level, the generalized deformations e may be expressed as functions of the nodal degrees
of freedom ue and the design variable s, that is
e = e (ue , s).

(1)

This expression is typically nonlinear in terms of both the nodal degrees of freedom (ue ) and the design
variable (s). Notice, that the nodal degrees of freedom are implicit functions of the design variable.
Assuming linear static material behaviour, the generalized stresses are related to the generalized
deformation via
e = Se (s)e ,

(2)

where the matrix S depends on the material properties, the design variable and the actual element at
hand.
Using Eq.(1), the virtual work of deformation may be written as
X
X
X
Wint =
Weint =
eT e =
(3)
eT De ue = qT u,
e

where the components of De are defined as


e
(ue , s) =
Dij

ei
.
uej

(4)

The load q(u, s) is often referred to as the internal load vector and is calculated on the basis of the
individual element contributions
qe = DeT e .
(5)
The applied, or external, load is denoted by p and it is often scaled with a load parameter , thus
p(s) = (s)f (s).

(6)

Wext = pT (, s)u.

(7)

The corresponding virtual work reads

For simplicity it has been assumed that the internal load is independent of the displacements. With the
principle of virtual work, it follows from Eq.(3) and Eq.(7) that
q(u, s) = p(, s).

(8)

The corresponding rate equations become


J

du
dp
=
,
d
d

(9)

where the tangent operator J is assembled from individual element contributions Je . The latter are
given by
(10)
Je = Ke + Ge ,

where the physical stiffness follows from


Ke = DeT Se De

(11)

and the components of the geometrical stiffness matrix are determined by


Geij =

ke

2 ek
.
uei uej

(12)

Both the physical stiffness matrix and the geometrical stiffness matrix are configuration dependent.

4.2

Rigid body modes

For the present paper, it is important to define for each element a basis of rigid body modes. In the actual
configuration, a rigid body mode e is characterized by the fact that the corresponding deformation rates
are zero. Thus, from Eqs.(1) and (4) follows
De e = 0.

(13)

Clearly the space of rigid body modes corresponds to the kernel of the linear mapping represented by the
matrix De . An orthogonal basis rek (ue , s) is introduced for the kernel of this mapping. The dimension
of the space spanned by this basis equals the number of independent rigid body modes to be described
by the element under consideration. Hence, for 2-D elements this space is typically three dimensional,
whereas for 3-D elements this space becomes six dimensional.
The basis vectors rek satisfy Eq.(13) for any configuration, i.e. for any combination of ue and s.
Differentiating the equation De rek = 0 with respect to the design variable gives the consistency conditions
De,s rek + De rek,s = 0.

(14)

The analytical derivation of the partial derivatives of De is tedious, whereas the partial derivatives rek,s
can be easily evaluated analytically without information on the precise formulation of the element at
hand. This observation has been used to improve semi-analytical design sensitivities in [19]. For this
purpose, the basis vectors rek will be used to decompose pseudo-loads into a self-equilibrating and a
non-self-equilibrating component. Therefore, it is convenient to introduce a matrices He and Fe defined
as
He =

X re reT
k k

e
reT
k rk

(15)

Fe = I H e .

(16)

These matrices can be used to extract the self-equilibrating load components at element level. Note that
both He and Fe are symmetric.
The partial derivatives of He and Fe with respect to the design variable s are given by
Fe,s = He,s = F0e + F0eT + 2

e
reT
k,s rk

rek reT
k
.
e 2
(reT
k rk )

(17)

Here the matrix F0 refers to a pseudo-gradient, which is defined as


e

F0 =

X rek,s reT
k
k

e
reT
k rk

(18)

Note that F0 is non-symmetric. With definitions (15) and (18) and consistency equation (14), we may
verify that
De,s He = De F0e .

(19)

This consistency equation can be used to replace partial derivatives of De by partial derivatives of rek .
4

4.3

Design Sensitivities

A straightforward differentiation of the equilibrium equation with respect to the design variable s gives
J

p,
= p,s q,s .
s
s

(20)

Provided the load parameter () is independent of the design variable s, the above equation simplifies
to
u
J
= p,s q,s .
(21)
s
The right-hand side of this equation is referred to as the pseudo-load. From Eq.(21) we can directly
obtain the displacement sensitivities or system derivatives for the geometrically nonlinear case
u
= J1 (p,s q,s ).
s

(22)

Clearly, the regularity of the Jacobian matrix is assumed, i.e. stability points are not included in this
study.
To solve Eq.(22) we need to evaluate partial derivatives. The Semi-analytical (SA) formulation utilizes
finite differences for this purpose. It was shown, that severe accuracy problems may result from replacing
q,s by its finite difference approximation for shape design variables, see, e.g., [12,13]. These inaccuracies
for shape design variables emerge as soon as individual elements undergo rotations. Moreover, mesh
refinement leads to even more severe accuracy problems. Particularly, slender structures may expose
relatively large rotations. Hence, the accuracy problems become most evident in such structures.
A possible solution has been given by Van Keulen and De Boer [19]. To achieve more accurate design
sensitivities, the following term is investigated in more detail:
e
eT e
qe,s = DeT
,s + D ,s .

(23)

In this equation the first term gives a non self-equilibrating contribution. As a consequence, an error as
a result of a finite difference approximation to DeT
,s tends to distribute over the entire model. On the
other hand, the second term leads to a self-equilibrated contribution, which immediately follow from
eT e
reT
k D ,s = 0,

(24)

where rek are the rigid body modes. As a consequence a finite difference error in the differentiation of
e,s tends to damp out.
To solve the accuracy problems caused by the first term, we decompose it in a self-equilibrating
component and one which is not. In this way Eq. (23) becomes
eT e
e
e eT e
qe,s = He DeT
,s + F D,s + D ,s .

(25)

Here, only the first term gives a contribution which is not self-equilibrating and using the consistency
equations, see Eq. (19), it may be replaced by one that only involves partial derivatives of rek
e

e
eT e
qe,s = F0 DeT e + Fe DeT
,s + D ,s .

(26)

In the Refined Semi-Analytical design sensitivities the derivatives of e and De are evaluated using finite
differences. Although this leads to certain errors, they do not spread over the entire domain. The real
advantage is the analytical differentiation of the rigid body modes rek used in the matrix F0 . Using
the refined semi-analytical derivatives eliminates the severe accuracy problems, without affecting the
advantages of a semi-analytical formulation. Moreover, the derivatives of the rigid body modes are easy
to calculate and are even zero for rigid body translations.

4.4

Fast Reanalysis

Starting point of a reanalysis is an available solution obtained by a previous analysis of an unperturbed or


nominal configuration. Before focussing on the reanalysis, we first summarize the incremental-iterative
solution procedure for the unperturbed configuration. This configuration corresponds to the setting s
for the single design variable. Starting point is the set of equilibrium equations
q(u, s) p(, s) = 0.

(27)

Assuming the load parameter is independent from s, the rate equations become
J(u, s)

u(s)
= f (s),

(28)

which follows immediately from (27). Load and displacement increments are now related by the approximate relation
J(u, s)u f (s)
(29)
and an incremental-iterative solution procedure relies on
J(ui , s)ui = f (s) q(ui , s)

(30)

ui+1 = ui + ui .

(31)

If the load level () is kept constant, then Eqs. (30) and (31) represent a standard Newton process. The
solution of the nominal configuration is typically found using an incremental-iterative solution technique
based on the equations above. The iteration processes will be aborted as soon as an acceptable unbalance
has been achieved, i.e. (27) is satisfied within a certain error tolerance.
Assuming an accurate solution has been found on the basis of (30) and (31), then the design configuration will be perturbed from s to s + s. The solution found for the nominal configuration will
be denoted u0 and is the starting point for reanalysis using a standard or full Newton process. In that
case, the reanalysis is governed by
J(ui , s + s)ui = f (s + s) q(ui , s + s)

(32)

ui+1 = ui + ui .

(33)

For the first iterate, i.e. i = 0, the displacement vector equals the solution of the nominal configuration
(u0 ). A small modification is to use the tangent operator of the unperturbed configuration, i.e. the
reanalysis is entirely based on a modified Newton process. Then the reanalysis scheme becomes
J(u0 , s)ui = f (s + s) q(ui , s + s)

(34)

ui+1 = ui + ui .

(35)

Naturally the number of iterations required for the reanalysis increases, however only the internal forces
are re-evaluated at each iteration.

Comparison Reanalysis versus Semi-analytical Design Sensitivities

In many previous studies, it has been demonstrated that Semi-Analytical design sensitivities applied to
shape design variables may expose severe accuracy problems. As mentioned, these inaccuracies become
more pronounced when we consider slender structures subject to rotation and when fine meshes are used.
The severe accuracy errors are due to the numerical differentiation of the internal force vector.
The starting point for a reanalysis is the solution u0 (s), and this solution will be used to search the
solution u(s + s). The first iterate of the fast reanalysis using the full Newton method, see (32), follows
from
(36)
J(u0 , s + s)ui = f (s + s) q(u0 , s + s).
6

In case a forward finite difference scheme is used in a semi-analytical formulation, then the sensitivity
equation becomes
J(u0 , s)

u
((f (s + s) f (s)) (q(u0 , s + s) q(u0 , s)))
=
,
s
s

(37)

u
(f (s + s) q(u0 , s + s))
=
.
s
s

(38)

which may be simplified to


J(u0 , s)

We note that Eq. (36) and Eq. (38) are very similar except for the fact that in the sensitivity equation
1
J(u0 , s) and the factor s
are used.
This similarity is even more pronounced when a modified Newton scheme is considered. In that case,
the first iterate of the reanalysis is determined by
J(u0 , s)ui = f (s + s) q(u0 , s + s).

(39)

1
In this case the only difference is given by the factor s
.
The foregoing comparison indicates that the first increments used in the reanalysis schemes (nearly)
resemble semi-analytical design sensitivities. This observation puts forward the burning question whether
the first iterate is effective in situations for which semi-analytical derivatives would lead to accuracy
problems. In fact, numerical examples demonstrate that indeed the efficiency of reanalysis schemes
deteriorates for cases exhibiting rotations of the elements. This deterioration is observed by a relatively
large number or iterations used in the reanalysis or even no convergence.

Improved reanalysis

In the previous section it has been explained why fast reanalysis may become ineffective for shape design
variables. In the present section, we shall device a simple remedy. The basis idea is to use accurate
design sensitivities, e.g. based on the Refined Semi-Analytical method, as a predictor step in the Newton
process used for the reanalysis. This may be done by multiplying the design sensitivities with the applied
design perturbation to obtain an high-quality estimate for the first increment. Thus, the first iterate
becomes
u1 = u0 + u,s s.
(40)
After this modified first step in the reanalysis, a standard or, alternatively, a modified Newton process is
again applied to obtain the converged solution for the perturbed configuration. As will be demonstrated
in the next section, this correction decisively improves the convergence of the reanalysis schemes.

Example

A simple rectangular strip as shown in Figure 1 is considered. The length L of the long sides is chosen as
design variable. The short side at the left-hand side is clamped and at the right-hand side a distributed
bending moment is applied, whereas for the long sides symmetry boundary conditions are applied. Both
a coarse (Figure 1) and a fine (Figure 2) mesh based on triangular finite rotation shell elements [21]
are used. Different load levels have been studied. Figure 2 provides an indication of the deformed
configuration at load level = 1000.
In a first study, either the load parameter or the design perturbation is varied. The results are
depicted in Figure 3 and Figure 4. The results labelled Full Newton refer to a standard reanalysis,
SA-Newton refers to the modified scheme but using standard semi-analytical derivatives and RSANewton corresponds to the modified scheme using refined semi-analytical derivatives. For all cases,
standard Newton iterations are used. Similar results are obtained using the fine mesh, as shown in
Figure 5 and Figure 6.

1
0
1
0
10
20
30
40
50
60
70
80
10
90

5
100

Figure 1: Unperturbed strip configuration. Coarse mesh.

10

15

20
25
30
35
15

10

10
0

5
10
15

Figure 2: Deformed strip configuration with a tip moment applied ( = 1000). Fine mesh.

8
+

Iterations

7
6

600

700

800

900

1000

3
2

+
+

5
4+

RSA-Newton
SA-Newton
Full Newton

200

300

400

1
0
100

500

Figure 3: Difference between Full Newton, SA-Newton and RSA-Newton for a relative design perturbation of 0.1% and a coarse mesh.

35

Iterations

30

RSA-Newton
SA-Newton
+
Full Newton

25
20
15

10

+
5

0
0.1

0.5

1
5
Length Perturbation

10

20

Figure 4: Difference between Full Newton, SA-Newton and RSA-Newton for =500 and a coarse mesh.
Horizontally the relative design perturbation has been given.
8

Iterations

600

700

800

900

1000

6
5+

3
2

RSA-Newton
SA-Newton +
Full Newton
+
+

200

300

400

1
0
100

500

Figure 5: Difference between Full Newton, SA-Newton and RSA-Newton for a relative design perturbation of 0.1% and a fine mesh.
The same example has also been tested using different line loads at the tip and using a modified
Newton scheme. (These results will be included in a full-length journal paper.) All results are consistent with observations found in studies on semi-analytical derivatives. Moreover, all results indicate a
significant improvement of the performance if a corrected predictor is used in the reanalysis. The most
spectacular results can be found in combination with a modified Newton scheme.

Conclusions

There are two main conclusions from the present study. Firstly, accuracy errors in the first increment used
in geometrically nonlinear reanalysis schemes may affect efficiency severely. In particular for modified
Newton processes the effect may be very large. Thus, configurations leading to accuracy problems for
standard semi-analytical design sensitivities, are likely to cause difficulties in the reanalysis. Secondly,
using accurate design sensitivities to construct an accurate first increment in the reanalysis is an effective
method to avoid the observed complications and to conserve the efficiency of the reanalysis.

60

RSA-Newton
SA-Newton
Full Newton
+

Iterations

50
40

30
+
20
10

+
+

0
0.1

0.5

1
5
Length Perturbation

10

20

Figure 6: Difference between Full Newton, SA-Newton and RSA-Newton for =500 and a fine mesh.
Horizontally the relative design perturbation has been given.

References
[1] U. Kirsch. A unified reanalysis approach for structural analysis, design and optimization. Struct.
Multidisc. Optim., 25:6785, 2003.
[2] F. Van Keulen, K. Vervenne, C. Cerulli, and U. Kirsch. Improved combined approach for shape modifications. 10th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, Albany,
Ny, Aug. 30 - Sep. 1, 2004.
[3] K. Vervenne, J. Fatemi, and F. van Keulen. Efficiency improvement of response surface building
using fast reanalysis methods. AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics
and Materials Conference, Palm Springs April 19-22, 2004.
[4] U. Kirsch. Design-oriented analysis of structures - a unified approach. Dordrecht: Kluwer Academic
Publisher, 2002.
[5] J. Gullickson and R.C. Averill. Finite element solution refinement using combined approximations
reanalysis. 42nd AIAA/ASME/ASCE/AHS/ASC Structures,Structural Dynamics, and Materials Conference and Exhibit 16-19 April, 2001.
[6] U. Kirsch and M. Bogomolni. Procedures for approximate eigenproblem reanalysis of structures.
Int. J. Num. Meth. Engng., 60:19691986, 2004.
[7] U. Kirsch and Y. Papalambros. Exact and accurate solutions in the approximate reanalysis of
structures. AIAA Journal, 39(11):21982205, 2001.
[8] H.R. Bae, R.V. Grandhi, and R. Canfield. An efficient successive reanalysis technique for engineering
structures. 45th AIAA/ASME/ASCE/AHS/ASC Structures,Structural Dynamics adn Materials
Conference,Palm Springs,Apr. 19-22, 2004.
[9] U. Kirsch and M. Bogomolni. Refined consistent formulation of a curved triangular finite rotation
shell element. 7th International Conference on Computational Structures technology, Lisbon,
Sep. 7-9, 2004.
[10] U. Kirsch. Efficient, accurate reanalysis for structural optimization. AIAA Journal, 37(12):1663
1669, 1999.
[11] R.T. Haftka. Semi-analytical static nonlinear structural sensitivity analysis. AIAA Journal, 31:1307
1312, 1993.
[12] B. Barthelemy and R.T. Haftka. Accuracy analysis of the semi-analytical method for shape sensitivity calculation. AIAA paper 88-2284, 1988.

10

[13] B. Barthelemy, C.T. Chon, and R.T. Haftka. Accuracy problems associated with semi-analytical
derivatives of static response. Finite Elem. Anal. and Design, 4:249265, 1988.
[14] G. Cheng, Y. Gu, and Y. Zhou. Accuracy of semi-analytic sensitivity analysis. Finite Elem. Anal.
and Design, 6:113128, 1989.
[15] P. Pedersen, G. Cheng, and J. Rasmussen. On accuracy problems for semi-analytical sensitivity
analyses. Mech. Struct. Mach., 17:373384, 1989.
[16] P. Fenyes and R.V. Lust. Error analysis of semianalitic displacement derivatives for shape and sizing
variables. AIAA Journal, 29:271279, 1991.
[17] N. Olhoff and J. Rasmussen. Study of inaccuracy in semi-analytical sensitivity analysis - a model
problem. Struct. Optimiz., 3:203213, 1991.
[18] N. Olhoff, J. Rasmussen, and E. Lund. A method of exact numerical differentiation for error elimination in finite-element-based semi-analitycal shape sensitivity analyses. Mech. Struct.
Mach., 21:166, 1993.
[19] F. Van Keulen and H. de Boer. Refined semi-analytical design sensitivities. Int. J. Solids and
Structures, 37:69616980, 2000.
[20] E. Parente and L.E. Vaz. On evaluation of shape sensitivities of non-linear critical loads. Int. J.
Numer. Meth. in Engng., 56:809846, 2003.
[21] F. Van Keulen and J. Booij. Refined consistent formulation of a curved triangular finite rotation
shell element. Int. J. Num. Meth. Engng., 39:28032830, 1996.

11

You might also like