Download as pdf or txt
Download as pdf or txt
You are on page 1of 16

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/250692192

Nonlinear analysis using a reduced number of variables

Article in Computer Methods in Applied Mechanics and Engineering · September 1985


DOI: 10.1016/0045-7825(85)90020-9

CITATIONS READS

26 30

2 authors, including:

Kuo Mo Hsiao
National Chiao Tung University
63 PUBLICATIONS 1,425 CITATIONS

SEE PROFILE

All content following this page was uploaded by Kuo Mo Hsiao on 06 April 2020.

The user has requested enhancement of the downloaded file.


COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING 52 (1985) 899-913
NORTH-HOLLAND

NONLINEAR ANALYSIS USING A REDUCED NUMBER OF VARIABLES

A.S.L. CHAN and KM. HSIAO


Department of Aeronautics, Imperial College of Science and Technology, London SW7 ZBY, U.K.

Received 15 October 1984


Revised manuscript received 6 March 1985

1. Introduction

A nonlinear structural analysis usually proceeds incrementally and requires a series of


repetitive solutions of a changing stiffness matrix which is time consuming. Considerable
efforts have therefore been made in various directions at finding ways of reducing the
computational cost. Most of these aim at perfecting the numerical solution scheme to improve
on the basic Newton-Raphson iterative algorithm, giving rise to the derivative forms such as
the modified Newton, quasi-Newton, conjugate Newton, and secant-Newton methods, etc.,
each one more (or less) suitable for certain types of problems. A further improvement may be
achieved by using a smaller number of generalized variables to represent the usually large
number of degrees of freedoms, thereby reducing the size of the stiffness matrix involved in
the calculation. The outcome will then depend on how well the actual structural behavior is
reproduced by the transformation. The orthodox modal analysis method [l] and the use of the
so-called path-derivatives [2] have been proved successful.
This paper investigates the possibility of taking a normalized set of correction vectors,
derived during the normal course of a nonlinear analysis step using an appropriate iterative
scheme (the modified Newton-Raphson method was adopted here), to be the basis vectors of
this transformation. These vectors correct a linear trial solution to the nonlinear one, therefore
must have contained the information concerning the local characteristics of the nonlinear
behavior. Indeed they can be shown to be related to the more rigorously derived path-
derivatives of the generalized load-deffection curve. They can be obtained just as easily in the
presence of material nonlinearity, as well as for the purely geometrical nonlinear problems,
which is perhaps the greatest advantage of this simple device over the more theoretical
approaches.

2. Formulation of the noniinear structural auaiysis

A brief resume of the formulation is here included for completeness. The structure is
idealized by the usual finite element method and either a total Lagrangian or an updated
Langrangian description may be used. The strain increment ds can be expressed in terms of the
displacement increment dr as
dc=(&+&)dr=Bdr, (1)

00457825/85/$3.30 @ 1985, Elsevier Science Publishers B.V. (North-Holland)


900 A.S.L. Ghan, KM. Hsiao, Nonlinear analysis using a reduced number of variables

where BL arises from the linear terms of the strain/displacement compatibility relationship and
BN from the nonlinear terms which is a linear function of the displacements 1. Application of
the principle of virtual displacement then resolves the stresses S into internal nodal forces F in
the form of

(2)

so that the increments

dF=$-dr= (B’g+$k)drdV=Kdr. (3)

The tangent stiffness

(4)

is a nonlinear function of r not only because of the geometry dependence of B but also
implicitly through the stresses S when material nonlinearity is present. The internal forces F
should of course be in equilibrium with the equivalent external nodal force vector P,
representing the actual forces applied on the body. However, for small finite increments AP,
the solution can only be obtained iteratively until the residual force

AR =AP-AF (5)

diminishes below a certain prescribed tolerance. In this paper the modified Newton-Raphson
method has been used, but other predictor-corrector methods might serve the purpose equally
well. For each incremental step with load AP, a trial solution

ArO= K-‘AP (6)

is obtained, from which the internal force AF can be calculated. This will in general not be the
same as the applied load, and a residue AR = AP - AF is found and used for correcting the
trial solution. The process is repeated until no further correction is necessary. The jth
correction is given by

Ari = K-’ ARj. (7)

3. Step-size control

When the loading is conservative, the finite loading increment AP can be written as AAP
where P is a normalized reference loading vector and Ah a parameter governing the
A.S.L. Chan, K.M. Hsiao, Nonlinear analysis using a reduced number of variables 901

magnitude of the incremental step. Alternatively, one prominant displacement in the vector Ar
may be chosen as the step-size controlling parameter. These will work for simple nonlinear
problems where no instability is involved. However, the computation will fail whenever the
problem exhibits multi-valued behavior such as snap-through or snap-back. Many methods
[3-71 have been devised in recent years to overcome these type of difficulties, amongst which
the use of a non-decreasing arc-length control seems to be the most versatile [8-10]. This
paper adopts the method proposed by Crisfield, in which the generalized arc-length increment
is defined to be the Euclidian norm of Ar, or

AL2 = Ar’ Ar . (8)

In one version [8], this is kept constant for each step, then the load-size parameter Ah can be
adjusted simultaneously with the displacement corrections during the iterations by satisfying
the constraint equation (8). In another version, the secant method [6,7] is used, but the
load-size parameter AA is regulated at each iteration such that the residual force is minimized
(see Fig. 1) giving

Ah = P’AF. (9)

The former method seems to be marginally more efficient in practice.

Fig. 1. Minimizing the residual force.

The iterative solution is terminated for each step when the convergence criterion, defined
here to be the weighted Euclidian norm of the residual force I]RII/NA becomes sufficiently
small, where N is the total number of degrees of freedom and A the loading parameter. The
tolerance is taken to be between 10e3 and 10e4 in the numerical studies.

4. Solution with a reduced number of variables

The idea is to reduce the size of the problem by using a small number of generalized
variables q to represent the large number of displacement freedoms r in the form of a
transformation

r= Tq. (IO)

The columns of the matrix T are called the basis vectors. The principle of virtual displace-
ment then transforms (3) to
902 A.S.L. Chan, K.M. Hsiao, Nonlinear analysis using a reduced number of variables

dfl==Kdq (loa)

and (5) to

Ag=AQ-AF (11)
where
AR= T’AR, (114
AQ = T' AP = AAT’P. W)
AP = T’AF, (Ilc)
K= T’KT (114

are the residual force, the equivalent applied force, the internal force and the tangent stiffness
matrix of the reduced system respectively. The iterative solution procedure can be the same as
that for the original system. However, the internal force AP cannot be evaluated in terms of
the variables Aq directly, but must return to the full system of variables to find AF element by
element and then transform according to (11~). The convergence criterion for the reduced
system is also taken to be the same one as defined before for the full system in this paper, in
order to obtain a true comparison between the two analyses, although a simpler convergence
criterion could be defined directly in terms of AR. In spite of these inconveniences, it is
reasonable to expect the solution in the reduced system to be more economical since the size
of the problem is so much smaller.

5. Choice of the basis vectors

Obviously, the successfulness of the reduced analysis will depend on the ability of the basis
vectors in the transformation matrix T to reproduce correctly the structural behavior,
manifested as displacements r but represented only by the generalized displacements q. Following
the long-standing practice of modal synthesis, one would expect the first few eigenmodes of
the initial buckling analysis to be the most obvious candidates to fulfill this role since they
dominate the initial stages of the nonlinear behavior, and have indeed been utilized as such in
[l]. However, care must be taken to include the critical modes in the basis near the limit or
bifurcation points for the post-buckling behavior to be correctly traced, and the choice of such
modes is not always obvious. Noor et al. [2,11-141 have successfully used a displacement
vector and its derivatives with respect to the control parameter (the so-called path-derivatives)
as the basis vectors. These vectors were essentially calculated by differentiation of the
nonlinear stiffness matrix and must have contained all the information governing the change of
the local tangent, although their computation looks formidable. The use of the solution vectors
at successive incremental steps, which may be considered as different combinations of the first
vector and its derivatives, has also been tried and proved acceptable [15]. The first two
methods outlined above are generally applicable for elastic systems only when the usual
displacement method of formulation is used; while the last one, although not presented as
such, can be adopted for inelastic structures as well. Recently path-derivatives have been
A.S.L. Chart, K.M. Hsiao, Nonlinear analysis using a reduced number of variables 903

obtained for pin-ended frames using a mixed formulation in dealing with material nonlinearity
[16], but it will be difficult, to say the least, to extend it for general usage.
This paper investigates the feasibility of choosing the solution vectors and the correction
vectors generated during the iterative procedure to be the basis vectors. The solution of the
ith step may be expressed in terms of the prediction vector Ar,,, (6) and the correction vectors
Ari, (7) as

Vi = ri-l+ Ar = ri-1+ 2 Aq, (12)

which may be written into the form of (10) with

T= [ri_l, Are, Ar,,Arz, . . .] and q’= [l, 1, 1, 1, . . .] . (13)

This clearly means that the vectors in T are the basis vectors for ri, and may be used to obtain
further solutions of the subsequent steps.
The relationship between these vectors with the more rigorously derived path-derivatives
can be shown by considering ri to be a function of the generalized arc length L. A
Taylor-series expansion about the previous solution point ri_l gives

where dr”/dL” is the nth-order path-derivative and AL the arc-length increment of the step.
Comparison between (12) and (14) enables the basis vectors of the present choice to be related
to the path-derivatives approximately. If the series on the right-hand side of the two equations
are truncated to only one term as a first approximation, then clearly

Ar =d’AL
’ dL ’
W)

By progressively taking more terms in both series, the general relationship

1 1 d”r .
Al) =--AL" where n= j+l (15’4
n! dL”

can be established. However, these are not exact identifications since they are based on
successive approximations.
If the path-derivatives are approximated by the finite difference formula in terms of the
solution vectors at successive steps, then the use of these vectors as basis can also be seen to
be related.
The basis is found to be most effective, amongst other possibilities, if it is chosen, after
suitable normalization and orthogonalization (the Gram-Schmidt procedure was used in the
examples investigated), in the following order of priority:
904 A.S.L. Chan, K.M. Hsiao, Nonlinear analysis using a reduced number of variables

T = [ri-l, ri, Are, Arl, Ar2, . . .] , (16)

where rj is the nonlinear solution vector at the ith step, calculated from the previous solution
ri_l, the trial solution Ar,, and the correction vectors A5 (i = 1,2, . . .) (see Fig. 2). If i = 1, rO is
taken to be the linear solution and ArO is omitted. Usually, between 4 and 8 vectors are
sufficient. The inclusion of ri compensates for the omission of the higher-order corrections.
The advantage of the present choice is twofold: Firstly, the basis vectors are generated
numerically during the normal course of one incremental step of the nonlinear analysis. They
are approximately equivalent to the use of the path-derivatives, but require no special
calculations for their generation. In this sense, this method is more convenient than the first
two methods (the use of eigenmodes or path-derivatives); and it does not have to perform as
many full-system solutions in order to obtain enough basis vectors as in third. Secondly, its
application is not confined to elastic problems as in the first two methods, when using the
conventional displacement finite element procedure, but can deal with nonlinear material
properties just as easily. In addition, the method also has the flexibility such that the number
of basis vectors may be changed, and new vectors may be substituted in place of old ones. All
these can be accomplished with relatively little computing cost and enable the range of
convergence to be extended for a given tangent stiffness of the full system.

ri-I

Fig. 2. Iterative solutions.


A.S.L. Chan, K.M. Hsiao, Nonlinear analysis using a reduced number of variables 905

An analysis using this technique would therefore assemble and decompose the linear elastic
stiffness matrix for the full system, and generate the necessary basis vectors during the first
nonlinear iterative step. The solutions for the subsequent steps are obtained with the reduced
system. When convergence becomes difficult (i.e. fails to converge within certain pre-specified
number of iterations), different strategies may be adopted. The first simply adds more
correction vectors into the basis, which expands the stiffness matrix E of the reduced system
without changing what is already in it; and when the number of vectors reaches a specified
maximum, changes the basis vectors rip ri-1, etc. to the most up-to-date values. All this can be
done without altering the stiffness K of the full system. Only when this simple strategy fails,
the calculation needs to go back to the full system to form an updated tangent stiffness and
decompose again for the generation of new basis vectors. This is usually indicative that the
characteristic of the structural behavior has been so radically changed by the deformation that
its representation by the original stiffness becomes totally inadequate. The updated full system
is then kept and the analysis proceeds in the reduced system as before, and so on. Apart from
these, there are also other possibilities which has not yet been explored.

6. Examples

A few simple problems of nonlinear structure have been analyzed by the reduced basis
method as described. In all the load-deflection curves here presented, the individual points are
the results obtained by the technique discussed above, and superimposed on the results

lOP, lOP,

EA = 9000
8 A = 1.

_I-

Fig. 3. Truss tower.


906 A.S.L. Chart, K.M. Hsiao, Nonlinear analysis using a reduced number of variables

obtained from the full-system analysis represented by the continuous line. The first strategy
refers to the updating of the basis matrix T alone, based on results of the reduced calculation,
while the second strategy refers to the updating of the full-system tangent stiffness and the
generation of new basis based on the new decomposition.

EXAMPLE 6.1. Truss tower. The geometry and loading for this simple structure is shown in
Fig. 3. The material property is first taken to be linearly elastic, and then assumed to be
nonlinear with a stress/strain relationship of the Ramberg-Osgood type giving a plastic strain

(17)

where m = 8. Results for the two cases are given in Figs. 4 and 5 respectively. The elastic
analysis uses 7 basis vectors throughout while the analysis including material nonlinearity has a
changing number of basis vectors as indicated. The tangent stiffness has to be updated twice in
both cases.

EXAMPLE 6.2. Cylindrical shell. Only symmetrical deformations are considered for
this example. Fig. 6 gives the details of a fairly thin shell under uniform pressure. A

5 .I

4 .!
- FULL SYSTEM ( 26 DOF)

0 REDUCED SYSTEM f 7 BASIS Vt!CtOI’S)

4 .I

3 .!

3.1

a”

*f 2.:

B
0
-I 7.c

I .5

Basis generated
I .a
in the full system l first strategy used

0.S )( second strategy used

0.0
I 2 3 4 5 6 7 6 9 IO II 12 13

Displacement A
Fig. 4. Load-deflection curve for the truss tower in the elastic case.
A.S.L. Chart, KM. Hsiao, Nonlinear analysis using a reduced number of variables 907

2.0 -

2.1 -

a”
” 2.0-
+
0
: l.6-

I .2 - ,, _-Number of basis vectars

Iksis gencrotcd
in the full system 0 first strategy used

second strotegy used

I...l...l...l...l...l...l...l...l...,...

0 .2 0 .4 0.6 0 .6 I.0 I .2 I.4 1.6 I .6 2.0 2.2

DISPLACEMENT A

Fig. 5. Load-deflection curve for the truss tower in the elastic-plastic case.

quarter of the shell is divided into an 8 x 8 grid. A 4-noded degenerated shell element (SH22)
[17,18] with 5 degrees of freedom per node is used. The total non-zero degrees of
freedom is 288. The calculation was based on a reduced integration with 1 x 1 x 2 Gaussian
points. Fig. 7 gives the results showing first softening and then hardening but no limit point.
Fig. 8 shows a thicker shell geometry, but the shell is supported differently at the boundaries
and subjected to a point-load at the centre. A quarter of the structure is now modeled by a
2 x 2 grid of 9-noded degenerated shell elements (SH33) of the same stable as before,
requiring a 2 x 2 x 2 set of integration points per element but giving only 87 non-zero degrees
of freedom. The load-deflection curve shown in Fig. 9 is typical for the snap-through
phenomenon. Nevertheless, both calculations required the tangent stiffness to be updated
three times.

EXAMPLE 6.3. Shallow arch. The details are shown in Fig. 10. Only symmetrical defor-
mation has been considered. A quarter of the arch is idealized by a 4 x 1 mesh of SH33
elements. The numerical integration is based on 3 x 2 x 2 points when the material is
considered to be purely elastic and 3 x 2 x 3 points when it is taken to be elastic/perfectly
plastic. The results for both cases are shown in Fig. 11. For the elastic analysis, only the first
strategy was necessary, whereas the plastic analysis required the tangent stiffness of the full
system to be updated four times because of the plastic deformation.
908 A.S.L. Chan, KM Hsiao, Nonlinear analysis using a reduced number of variables

R = 2.54~1

t = 3.175x la-3a

6 = 0.1 rad
E = 3.10275 x 106kN/m2

v = 0.3

All edges are clamped

Fig. 6. Clamped cylindrical shell.


A.S.L. Chan, K.M. Hsiao, Nonlinear analysis using a reduced number of variables 909

Number of basis vectors

Basis enerated
in the full system . first strategy used
0.5 x second strategy used

A-.---_., .~ -_~-_~L-
0.5 1.0 1.5 2.0 2.5 3.0 3.5

Central displacement ratio W,/t

Fig. 7. Load-deflection curve for the clamped cylindrical shell,

7. Conclusions

From these examples, it can be fairly concluded that this method of choosing the basis
vectors for the reduced system is capable of producing results indistinguishable from the
analyses using the full set of variables, for both geometrical and material nonlinear problems.
However, for the examples, in which the number of degrees of freedom are small, the saving
in computing cost has not been as great as hoped for, mainly because of the necessity of going
back to the full system of variables at every iteration for the calculation of the residual force,
and the stringent error norm imposed on the iterations of the reduced system. A rough
estimate based on the operation count of some key features of the procedures projects the
break-even point at 3000-5000 unknowns, depending on the element used. Above this number
the reduced basis technique will be of definite advantage. This is based on the experience of
these examples which are mainly concerned with proving the capability of so choosing the
basis vectors, rather than with the efficiency of the procedure. No doubt, a more efficient
error-sensing procedure, a simpler convergence criterion based on the reduced variables alone,
and perhaps greater coding skill may bring down the threshold a great deal. The objective of
the present investigation is to obtain a strict comparison between the results of the full and the
910 A.S.L. Chart, K.M. Hsiao, Nonlinear analysis using a reduced number of variables

R = 2.54~1
t = 0.016m
0 = 0.1 rad
E = 3.10275x 106kN/m2
v = 0.3
Strait edges are hinged
Curved edges are free

Fig. 8. Hinged cylindrical shell.


A.S.L. Chart, K.M. Hsiao, Nonlinear analysis using a reduced number of variables 911

8 .I D-

7. Z-
- FULL SYSTEM ( 87 DOF 1

0 REDUCED SYSTEM
6. 4-

5 .I 6-

z 4.1 3- i
V

a?
4 .c l-

._c”

B 3.; , _
s

2.4 1 -

I .6
I Basis generated
first strategy used

0.8
in the full system second strategy used

i
,..I.., ..,.,4..,,.,. ,

3 6 9 12 15 18 21 24 27 30 33

Central Displacement W,(mm)

Fig. 9. Load-deflection curve for the hinged cylindrical shell.

I * TOP
_____;____--__-___ view
IC
,

L = 663.6mm
Z, = 27.69mm
R = 3361.lmm
b = 25.4mm
t = 14.99mm
E = 6.695xlO’MPa
” = 0.
(I * = 137.9MPo

Fig. 10. Clamped shallow arch.


912 A.S.L. Chan, K.M. Hsiao, Nonlinear analysis using a reduced number of variables

4.01 1
t

3.6- A
Q f?muCED SYSTEM
‘0 ‘a _- -+- _----__ELASTIC
6$I___---
5 0 . --*----_____.
- FULL SYSTEM
.w-- 0
5 @/_’ 0
3.2 ---- (se DOF)

2.8 -

9 2.4-
Y
‘=
a 2.0-
Number of bosis VeCtWS
F

P
‘5
4-
o 1.6-
s
Basis generated
1.2 -
in the full System

Central Displacement W,(mm)

Fig. 11. Load-deflection curves for the clamped shallow arch.

reduced analyses so as to establish the feasibility of the proposed method of choosing the basis
vectors, and the findings appear to confirm the validity of this simple technique.

References

[II D.A. Nagy and M. Konig, Geometrically nonlinear finite element behavior using buckling mode superposition,
Comput. Meths. Appl. Mech. Engrg. 19 (1979) 447-484.
PI A.K. Noor and J.M. Peters, Reduced basis technique for nonlinear analysis of structures, AIAA J. 18 (4)
(1980) 455-462.
[31 P.G. Bergan, Automated incremental-iterative solution schemes, in: C. Taylor, EHinton and D.J.R. Gwen,
eds., Proc. Internat. Conf. on Numerical Methods for Nonlinear Problems, Swansea (1980) 291-305.
[41 P.G. Bergan, G. Horrigmoe, B. Krakeland and T.H. Soreide, Solution techniques for nonlinear finite element
problems, Intemat. J. Numer. Meths. Engrg. 12 (1978) 1677-1696.
PI P.G. Bergan, Solution algorithms for nonlinear structural problems, Comput. & Structures 12 (1980) 479-509.
[61 M.A. Crisfield, A fast modified Newton-Raphson iteration, Comput. Meths. Appl. Mech. Engrg. 20 (1979)
267-278.
[71M.A. Crisfield, Incremental/iterative solution procedures for nonlinear structural analysis, in: C. Taylor, E.
Hinton and D.J.R. Owen, eds., Proc. Internat. Conf. on Numerical Methods for Nonlinear Problems, Swansea
(1980) 261-290.
PI M.A. Crisfield, A fast incremental/iterative solution procedure that handles ‘snap-through’, Comput. &
Structures 13 (1981) 55-62.
A.S.L. Chan, K.M. Hsiao, Nonlinear analysis using a reduced number of variables 913

[9] E. Ramm, Strategies for tracing the nonlinear response near limit points, in: W. Wunderlich, E. Stein and K.J.
Bathe, eds., Nonlinear Finite Element Analysis in Structural Mechanics. (Springer, New York, 1981) 63-89.
[lo] E. Riks, An incremental approach to the solution of snapping and buckling problems, Internat. J. Solids
Structures 15 (1979) 529451.
[ll] A.K. Noor and J.M. Peters, Tracing post-limit point paths with reduced basis technique, Comput. Meths.
Appl. Mech. Engrg. 28 (1981) 217-240.
[12] A.K. Noor and J.M. Peters, Bifurcation and post-buckling analysis of laminated composite plates via reduced
basis technique, Comput. Meths. Appl. Mech. Engrg. 29 (1981) 271-295.
[13] A.K. Noor, C.M. Andersen and J.M. Peters, Reduced basis technique for collapse analysis of shells, AIAA J.
19 (3) (1981) 393-397.
[14] A.K. Noor, Recent advances in reduction methods for nonlinear problems, Comput. & Structures 13 (1981)
31-44.
[15] B.O. Ahnroth, F.A. Brogan and P. Stern, Automatic choice of global shape functions in structural analysis,
AIAA J. 16 (5) (1978) 525-528.
[16] A.K. Noor and J.M. Peters, Instability analysis of space trusses, Comput. Meths. Appl. Mech. Engrg. 40 (1983)
199-218.
[17] K.J. Bathe and A.P. Bolourchi, A geometric and material nonlinear plate and shell element, Comput. &
Structures 11 (1980) 2348.
[18] E. Ramm, A plate/shell element for large defiections and rotations, in: K.J. Bathe, J.T. Oden and W.
Wunderlich, eds., Formulations and Computational Algorithms in F.E. Analysis. U.S.-Germany Symposium
(MIT, Cambridge, MA, 1977) 264-293.

View publication stats

You might also like