INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING, VOL.

23, 83 1-846 (1 986)

EFFECTIVENESS OF LINEAR BIFURCATION

ANALYSIS FOR PREDICTING THE NONLINEAR
STABILITY LIMITS OF STRUCTURES

S.-C. CHANG' AND J.-J. CHEN:

Department of Civil Engineering, National Taiwan University, Taipei, Taiwan, Republic of China

SUMMARY
As an effort to predict effectively the actual collapse load of a structure, a series of numerical studies on
the stability of shell structures are made. The difference in formulation between the two types of linear
buckling loads, the classical and the fully linearized, is first demonstrated. Their correlations with respect
to the actual stability limit of the structure are compared, and finally the two types of critical load
approximations are obtained at various stages of a nonlinear analysis to study the pattern of convergence
to the actual collapse load. It is found that the fully linearized buckling analysis, when combined with
nonlinear analysis, can serve as a useful tool for prediction of the stability limit of a structure. While for
most types of structures the approximation is within engineering accuracy, the rate of convergence of the
extrapolated critical load also gives some insight to the accuracy of the approximation.

INTRODUCTION
It is well known that the geometrically nonlinear behaviour of structures in finite element analysis
can be obtained through a number of approaches. Unfortunately, due to their high computational
expense, none of these approaches appeal to the engineering profession, and it is thus carried
out only for very special structures.
Among the engineering structures of most concern are those that exhibit softening or the
snap-through phenomenon. For these types of structures it is the load level that initiates the
snap-through or that causes considerable deformation that is being sought for. For this purpose
a nonlinear load-deflection curve, as depicted in Figure 1, needs to be traced. A simplification
of the elaborated nonlinear analysis procedure is the linear buckling analysis, in which the
stiffness matrix relating to quadratic terms of the generalized displacement is neglected. This
gives rise to a generalized eigenvalue problem to solve for the critical load. It is known that an
approximation of this type is effective only in special cases and has limited applications.
Other simplifications beyond the scope of finite element analysis have been tried. These
approaches deal mainly with cylindrical shells, since stability is of concern mostly for thin shell
structures and the cylinder is a commonly used shape. Batdorf' first idealized the cylindrical
shell structure by neglecting the pre-buckling rotation as well as the hoop stress and thus arrived
at the critical stress for axially loaded cylinders. Brush2 then considered the effect of pre-buckling
rotation and improved Batdorf's approximation. Crol13estimated the lower bound for the critical

'Associate Professor.
;i M. S. Candidate.

0029-598 1/86/08083 1 -l6$08.OO Received 29 November 1984

0 1986 by John Witey & Sons, Ltd. Revised 14 August 1985
832 S.-C. CHANG AND J.-J. CHEN

IP

1/“ Figure 1. Nonlinear load-deflection curve

loads of cylindrical shells by neglecting the strain energy due to membrane action of the structure.
Critical loads so found can be used safely in the design work of cylindrical shells.
In the realm of finite element analysis, there have been few studies into the effectiveness of
the two types of linear buckling analyses, the so-called ‘classical’ and ‘fully linearized’ buckling
analyses, as a means for critical load prediction. It is only known that fhese two types of buckling
analyses give better stability limit predictions when pre-buckling rotation in the structure is not
prominent, and that the ‘fully linearized‘ buckling analysis would degenerate to that of the
‘classical’ if the pre-buckling rotation vanishes. Few shell structures encountered in practice
today conform to this requirement, hence linear buckling analysis alone is not sufficient to
satisfactorily predict the nonlinear behaviour of the shell structure. One goal of the current study
is to explore, beyond the well-known statement above, the existence of an observable pattern
associated with the two types of linear buckling predictions with respect to the actual stability
limit of the structure.
An improved scheme for critical load prediction has been devised by Brendel and Ramm?
where additional linear buckling analyses are carried out at a number of intermediate load levels
before instability takes place. These supplementary linear buckling analyses would in general
give eigenvalue functions corresponding to better stability limit predictions over that based on
the unstressed structure. It was suggested that such predictions can serve as a useful tool to
judge extent of nonlinearity of the structural behaviour, and failure load can be approximated
through proper extrapolation procedures. Based on this methodology, it is the intention of this
paper to demonstrate, through a number of examples, that by properly combining linear buckling
as well as a minimal number of nonlinear analyses, the critical loads of shell structures can be
predicted within engineering accuracy simply by linear extrapolation. The prediction can either
be used in the practical design of structures where instability is a concern, or serve as a guideline of
structural behaviour in tracing the load-deflection curve.
The elements used in the numerical examples contained herein are the conventional beam
element and a triangular thin-shell element, the latter being developed by Wus and the first
author.6

LINEAR BUCKLING ANALYSIS

The complete nonlinear behaviour of a structure is governed by the equilibrium equation
[KO + Nl(u)/2 + N2(u2)/3](uf = ( P } (1)
which for a selected degree-of-freedom may exhibit a typical load-deflection curve as shown in
Figure 1. In equation (l), [ K O ]is the linear stiffness matrix, and [ N l ] and “21 are the stiffness
 EFFECTIVENESS OF LINEAR BIFURCATION ANALYSIS 833

matrices linearly and quadratically dependent on the generalized displacement vector ( u } ,

respectively. { P ) is the load vector applied and is assumed to be always of a linear scale to
a certain reference load vector (Pref).Equation (1) can be cast into an incremental form of
[KO+ Nl(u) + N2(u2)](Au)= (AP) (2)
The tangent stiffness matrix [ K O+ Nl(u) + N2(u2)]( = [ K T ] )is singular at the critical load level,
indicating that an infinitesimal increment in load vector would cause the displacement vector
to increase toward infinity. Hence we are seeking a configuration of the structure described by
a vector ( a ) that renders the stiffness matrix singular, under which circumstance the
corresponding load vector { P } is the critical load (Per).
Assume that the reference load vector (Pref)is first applied, and the structure deforms into
the corresponding state described by the displacement vector (ure,). A basic assumption of
linear buckling analysis is that the structure behaves linearly before the critical load is reached.
Therefore, for a positive constant ‘c’, the structural deformation is described as c(uref) for
the applied load vector c(Pref).A scale factor c,, corresponding to that of the critical load
is defined as
(PCJ = C c r ~ P ~ e z l
and is found through solving the equation
+ + ( 0~ ~ ~ ~ ) I
det [ K O cc,Nl(uref) C ~ , N ~ =
or
I.( = ( 0 )
+ CcrNl(uref)+ c,Z,N~(U,Z~)I (3)
where resulting ( u ) would represent the buckled mode shape. Solution of the above equation
requires a quadratic eigenvalueextraction procedure. No stable and efficientalgorithm is presently
available, however. A simplification is to neglect the “23 matrix and one arrives at
[KO + ccrNl(urefI1( u > = (01 (4)
Procedures for generalized linear eigenvalue extraction can be adopted to obtain the load factor
c,, as well as the mode shape { u ] . Since [Nl] contains all terms linearly dependent on
displacement (u), this analysis procedure is termed the ‘fully linearized’ buckling analysis.
Note that the result found through the fully linearized analysis is not the same as that found
through the commonlyknown buckling analysis, or more precisely the ‘classicalbuckling’ analysis.
In face, if the incremental strain-displacement relation ((Ae) - (Au)) of the structure
described by deformation ( a ) is written as
= I[B(u)l(Au) (5)
where
“41 = CBt(u)l+ LBO1 (6)
then [Nl] is expressed as

In the above, [El is the elasticity matrix, and {s) is the stress vector of the structure. The
834 S.-C. CHANG AND J.-J. CHEN

Table I. [Nl] matrix of a beam element

(Row-Col) Expression (factored by L2/EA)

0
-6 ( ~-
2 ~ 5 ) /-
5 L(u3 + ~,)/10

Table 11. [ K c ] matrix of a beam element

(Row-Col) Expression (factored by LIP)

0
0
6/5
0
1/10
2L2/15
0
0
0
0
0
- 6/5
- L/lO
0
6/5
0
1/10
- L2/30
0
- L/10
2L2/15
 EFFECTIVENESS O F LINEAR BIFURCATION ANALYSIS 835

Table 111. Comparison of classical and linearized buckling loads

Test case EKcl “11 Non 1inear
I +
0.520( 419)’ 0.173( + 73)’ 0.100
2 0.477( + 414) +
0.158( 71) 0.093
3 2S.35( + 4.3) 20.58( + 0.5) 20.47
4 16.36(+ 184) 3.427( - 40) 5.75
5 60.61(+ 124) 22.30( - 17) 27.0
6 4006( + 33.5) 125.0( - 58.3) 300.0
7 +
39.94( 247) +
13.05( 13) 11.50
8 10.05( + 327) 0.484( - 79) 2.35
9 2.612( + 49) 0.288(- 83) 1.75
10 7.214( - 20) 0.100(- 99) 9.00
11 0.361 (N/A) 0.285(N/A) stable
12 1.378(+ 2.6) 0.844( - 37) 1.344

‘Deviation percentage from nonlinear stability limit.

derivation of equation (7) is not elaborated here as it is now deemed a fairly straightforward
nonlinear finite element procedure. In particular, the first two terms of equation (7) correspond
to the ‘initial displacement matrix’ [ K , ] and the last term is the ‘initial stress matrix’ [K,] used
in classical buckling analysis.
The difference in [ K G ]and [ N l ] arises from the linear part of the strain-displacement matrix
[ B J . As an example, the “11 and [ K c ] matrices for a beam are given in Tables I and TI. In
these tables, the degrees-of-freedom u1-u6 are given in the orders of axial, traverse displacements
and rotations at the two nodes of the element. It is obvious that the difference is significant if
there is prominent traverse displacements and rotations prior to buckling. While the utilization
of [ N l ] and [ K c ] would give identical critical load for an ideal straight column, for most
engineering structures different critical loads will result. Although [ N l ] has more terms included
and thus is mathematically more complete, the critical loads derived from the fully linearized
buckling analysis is not necessarily more accurate. This is seen in Table 111, where 12 test cases
with critical loads from these two types of linear buckling analyses as well as from the full
nonlinear analysis are given for comparison. Although no significant pattern of the critical loads
with respect to the actual collapseload can be followed for either type of analyses, two observations
on these examples should be highlighted: first, the buckling loads from the classical approach
are always higher than those of the fully linearized, and secondly, the fully linearized buckling
loads are not always conservative with respect to the actual stability limits.
Figure 2 gives the details of the tested examples. In all cases the loadings are assumed to be
applied along their original directions throughout the nonlinear analysis, i.e. no follower-load
effect is included. In addition, all pressure loads are assumed to be initially normal to the shell
surface. In particular, two structures (cases 9 and 12) are subjected to non-uniform pressure
loads dictated by the expression of Fourier series: p(8) = p(u, cos (no)). With the proper value of
internal suction included, the Fourier coefficients for case 9 are: a, = 0.22, a, = 0.338, u2 = 0.533,
u3 = 0.471, a4 = 0.166, a s = - 0.066, u6 = - 0.055, while for case 12 the followings are in effect:
U, = 0.329, U1 = 0278, U2 = 0589, U3 =10*47,Uq = 0.063, U s = - 0.12, ff6 = - 0.027, a7 =: 0.044,
a8 = 0.002, a, = - 0.02. For the critical loads presented, the maximum windward pressure
corresponding to the critical state are implied. It should be mentioned that test case 9 is adopted
from Brendel and Ramm? but that a different shell element is used.
836 S.-C. CHANG AND J.-J. CHEN

I+----- 400 ---W

i \
w- 400 ------+

1=2000 E=2000 A=391 I = 2000 E = 2000 A= 391

test c a e 1 test case 2

10-deg. 35

1=2000 E=2000 A=391 1=2000 E = 2 0 0 0 A=391

test caw 3 test case 4

1=2000 E=2000 A = 3 % I=2000 E = 2000 A=391

test cme 5 test case 6

 EFFECTIVENESS OF LINEAR BIFURCATION ANALYSIS 837

all boondaries hinged

t =3.9154 t = 0.01
E=10000 v = 0 . 3 R=100 E=IOOOOOOO v = 0 . 3 R = 2 . 5
tesi case 7: spherical cap test case 8: cyliridrical panel

clamped

L=120 t =0.1064 R=100 t = l 1 = 1

E = 30000OOO v = 0.3 R = 40 € = l o 0 0 0 v = O A=12
test case 9: wind-loaded cylinder test case fO

“i
free

pressure

& 11.46
z -&g.q
ti1
R ~~

clamped
H1=11.918 H2 = 3.644
all boundaries clamped R1= 6.223 R2=4 R3=4.257
t = 0.125 E = 140oooOO v=0.3 t =0.007
E=450000 v = 0 . 3 R = 1 0 0 test case 12: wind-loaded hyberbdii
test case 11 coding tower

Figure 2. Summary of tested examples

838 S.-C. CHANG AND J.-J. CHEN

COMBINED ANALYSIS
It is seen from the previous discussions that the linear buckling analysis alone cannot give any
useful prediction about the actual critical load of the structure. A logical step to improve the
approximation is to perform either of the linear buckling analyses based on a stressed structure
under a certain load level (Phase), where
CPbase) = Cbase(Pref) (8)
The linear buckling load so obtained should, in most cases, be closer to the actual stability
limit than that obtained from an unstressed structure. To carry out such analysis, the selected
load level (Phase) is first applied either in increments or in a single step, and the conventional
nonlinear analysiswith Newton-Raphson equilibrium iteration is performed. The tangent stiffness
matrix of the structure [ K T ] corresponding to the equilibrium state can then be formulated. As
the succeeding step, a trial load increment is applied, and the [ K c ] or [Nl] matrix due to the
stress and displacement increment of this trial load step is found. Based on the assumption of
linear buckling analysis, we seek to find the proper load level beyond this fundamental state
(Phase) that one is to apply, so that the stiffness [KT] is completely degraded by increments
of [K,] or “11.
To find this additional load level, a generalized eigen-analysis procedure must be resorted to.
If the eigenvalue extraction based on this fundamental load level gives a critical load factor
Ac,,, then the predicted critical load factor cX, as shown in Figure 3, is
c,*, = Chase + Accr (9)
,c: differs from the c, in that the former is a predicted value from linear buckling analysis,
while the latter corresponds to the actual nonlinear stability limit. At various base load levels,
different linear critical load predictions may be calculated, and if the stability limit does exist
for the structure, the curve formed by connecting these predicted critical load levels should
intersect the nonlinear load-deflection curve at the point of instability. The curve is termed the
‘eigenvalue function’ by Brendel and Ramm4 and is called the ‘predicted-load’ curve here for
easy reference (see Figure 4).
Note that since only a particular selected degree-of-freedomis used to plot the load-deflection
curve, the shape of the predicted-load curve would depend upon the degree-of-freedom chosen.
The shape of the curve can be made unique if the predicted critical load factors are plotted
against the base load factors upon which the eigenvalue extractions are performed, Chase, as

Figure 3. Linear buckling load as obtained upon a stressed structure

 EFFECTIVENESS OF LINEAR BIFURCATION ANALYSIS 839

P
1

Figure 4. Predicted load curve in P-U system

Figure 5. Predicted load curve in load-factor system

shown in Figure 5. The predicted-load curve will always terminate at the point of co-ordinates
(ccr,ccr).For any point on the curve, the vertical distance to the 45-degree line is the load
factor Accr obtained from the eigenvalue analysis, and the height below the 45-degree line is
chase upon which Ac,, is calculated.

PREDICTION OF CRITICAL LOADS

The combined nonlinear and linear buckling analysis outlined above can be applied with either
the fully linearized or the classical buckling analysis. The group of examples previously listed
are all analysed with such approaches, and the resulting predicted-load curves are given in
Figure 6. In all cases, the curves of the classical buckling loads vary noticeably in shape, while
those of the fully linearized buckling loads give significantly less variation. The fully linearized
buckling analysis is therefore seen to be an appropriate tool for critical load prediction.
A simple prediction procedure using linear extrapolation is suggested as follows: first, a fully
linearized buckling analysis on the unstressed structure is performed, and an approximated critical
load level (c,*,)~ is obtained; then, a suitable load level of factor c2 (usually taken as ( c , * , ) ~
or 1) to the structure is applied and iterations are carried out until equilibrium. This configuration
of structure is taken as the base state upon which another fully linearized buckling analysis is
performed to obtain a new prediction ( c , * , ) ~Finally,
. on a c , * ~ - graph,
c ~ ~ ~ extrapolation
~
from points (0,(c:,)~) and (C,,(CX)~) is made to find its intersection with the 45-degree
840 S.-C. CHANG AND J.-J. CHEN

0
2-
/
/
/ I 0
%r /
/

/
/ /
1 /
/ /
/ /
/ /
/ /
/
/
/ ' 0 ' CLASSICAL 0: CLASSICAL
/
A : FULLY-LWARIZED A : FULLY - LINEARIZED
I I I I f I
4 00 6 00 000 2.00 4.00 6.00

I test case 1
base
I test case 2
base

rF /
/
/
/

/
/
/
/
/
/
/

/
/ /
/ /

0:
CLASSICAL

0 A : FULLY - LINEARIZED / A FULLY- LINEARIZED

0
0 I I I I 1
000 050 1 .oo 1.50 000 1.00 2 .oo 3.00
base chase
test case 3 test case 4
0 0
0 - 0 -
M / @J /
/ /
- / /
/ /
/ /
/ /
/ c
/
/ /
L -
/ b -
u" /
0

/ O'CLASSICAL / 0 CLASSICAL
/ /
/ A FULLY -LINEARIZED /
A FULLY -LINEARIZED
0
6.
0 8
000
' L
100
,
2 00
t

3 00
I o /
"600 0&7 1 33
I 8
2 00
I

base 'base
test case 5 test case 6

Figure 6. Summary of critical load extrapolations (load factors normalized with respect to ccr)
 0
8 0.00 0.67
Ei

\ P
\ 0
\ U
In
b 0 \\
\
\
\
\
\
\
\

Ccr
000 0.50 100 1.50
I 1

\
\
\ \
\ \
\ \
z
m
r \ \
\ \
w - 2
O R
c)
842 S.-C. CHANG AND J.-J. CHEN

Table IV. Buckling load predictions

Test case CKG1 "11 Nonlinear

1 0.100( + 2.7)+ +
0.100( 2.1)' 0.100
2 0.097( + 4.3) 0.097( + 4.3) 0.093
3 19.19(- 2.8) 20.38( + 3.7) 20.47
4 +
7.978( 38.7) +
5.656( 7.9) 5.75
5 46.4 (+ 71.9) 265 1 (- 1.8) 27.0
6 diverged 315.2( + 5.1) 300.0
7 8.568(- 25.5) 10,63(- 7.6) 11.50
8 2.556( + 8.8) 5.296(+ 125.3) 2.35
9 2.07( + 18.3) +
1.828( 4.4) 1.75
10 8.949( - 0.5) 456( - 49.3) 9.00
11 0.243(N/A) 0.22(N/A) stable
12 +
1.383( 2.9) 1683(+ 25.2) 1.344

Deviation percentage from nonlinear stability limit. .

line. The load level of the intersection point is then the extrapolated prediction on the critical
load factor ccr.More precisely, the extrapolated load factor is
(ccr)e = (CX)1 cz,"cZ + (c,*,)I- (cX)Zl (10)
which is subjected to the constraint (c& > 0. For values less than zero, the structure is actually
stiffening and therefore will not buckle. The philosophy behind this approach is fairly
straightforward: If one takes the critical load as the trial load level, then the eigenvalue extraction
would give a value of zero and leave the predicted critical load level to be the same as the trial
load level. The advantage of using the [Nl] matrix for predicting the critical load is largely due
to the linearity of the corresponding predicted-load curve, hence different base load levels will
give more or less the same prediction. This statement on the other hand cannot be made about
the classical buckling analysis. Table IV gives the predicted values versus those found through
elaborated full nonlinear analysis, and it is observed that the predicted values from the fully
linearized approach are mostly within 10 per cent of the actual critical load. The only few
exceptions are discussed in detail in the following section.
The described approach for critical load prediction requires significantly less computation
than that for the conventional full geometrically nonlinear analysis. For cases where the
load-deflection curve needs to be traced, the predicted critical load would serve as a valuable
guide for applying load increments on a later step-by-step nonlinear analysis.
The prediction can be somewhat improved if the order of extrapolation is raised to that of
quadratic, namely, if the prediction is obtained from the intersection of the 45-degree Iine with
the parabola defined by three linear buckling samplings. In this case, an additional step of
nonlinear analysis as well as an eigenvalue extraction would be required. However, the
improvement is noticeable only for the case 12 presented, where the critical load predicted changes
to 1.366 from the original value of 1.683. The corresponding relative error improves from 25-2
per cent to 1.6 per cent. The improvements for all other cases turn out to be negligible, and the
extra computational effort can hardly be justified. Therefore, quadratic extrapolation remains
more of a remedy rather than the suggested common practice.

EFFECTIVENESS O F PREDICTION
Although the extrapolation of linear buckling results using the [ N 11 matrix gives satisfactory
predictions for most cases tested, a criterion needs to be established so that the effectivenessof
 EFFECTIVENESS O F LINEAR BIFURCATION ANALYSIS 843

Figure 7. Deflected shape of point-loaded deep circular arch

the extrapolated critical loads can be assessed. Since the prediction is obtained as the intersection
point of the predicted-load curve with the 45-degree line, it is natural to expect valid prediction
only when the solution of the intersection is well-conditioned. Indeed, a close examination of
Table IV and Figure 6 reveals that the two cases where the predictions are off (cases 8 and 10)
are both associated with predicted-load curves of slopes very close to 1. It is logical to conclude
that one should not be too confident about the prediction on stability limit if the trial linear
buckling analyses give a line that is nearly parallel to the 45-degree line.
In addition, the distance between the two lines can give a useful indication as to how close
the structure is to the verge of buckling. If these two lines intersect only after being close to
each other for a considerable length, then the structure would be associated with fairly pronounced
deformation. This is the case dictating the behaviour for test case 10 of the point-loaded deep
arch. Figure 7 gives the deflected shape of the arch when the point load applied is 8-5, which is
more than 90 per cent of the critical load (9.0). The crown has deflected a magnitude larger than
its original radius, hence it would be unreasonable to expect the linear buckling predictions to
behave accurately within this cited range of deformation.
There are also cases where seemingly accurate critical load approximation is obtained while
the actual behaviour of the structure never reaches the unstable state. A well-known example
is the Euler column. The load-deflection curve of a fixed-free column under axial load is given
in Figure 8.7 Shown in the same figure are those of a column subjected to horizontal perturbing

-1 ,PERFECT
9

@v
0
yO.lo/o - IMPERFECT

' 1%- IMPERFECT

0

000 020 040 060 080 100

TIP ROTATION (in radian)

Figure 8. Load-deflection curves of Euler columns
844

.-
u
ln
a
P,
I
-
Pcr- predicted
S.-C. CHANG AND J.-J. CHEN

base

Figure 9. Predicted-load curve of 0.1 per cent imperfect column

- predicted

Figure 10. Predicted-load curve of 1 per cent imperfect column

CENTRE DEFLECTION (in)

Figure 11. Load-deflection curve of cylindrical panel under normal pressure
 EFFECTIVENESS O F LINEAR BIFURCATION ANALYSIS 845

loads with magnitudes of 0 1 per cent and 1 per cent of the vertical load. Figures 9 and 10 give
the corresponding predicted-load curves. The perturbations are assumed to grow linearly with
the applied axial load during the nonlinear analysis that gives these figures. For Figure 9, despite
the fact that the curve is nearly linear and horizontal (slope = 0), the curve turns sharply near
the critical load and becomes parallel to the 45-degree line thereafter, indicating that no instability
would occur. This phenomenon is less pronounced in Figure 10 where the perturbing load is
increased to 1 per cent of the vertical load. While mathematically the proposed approximation
algorithm fails, the load level beyond the predicted buckling load can hardly be realized in
practice because the deflection becomes unacceptably large.
The predicted load will nevertheless serve as a useful design guide. One of the previously
tested examples, that of case 11, is found to observe the above description. Figure 11 gives the
load-deflection curve for the pressure loaded cylindrical panel. It can be seen that although the
structure is stable throughout the loading history, the curve does exhibit a plateau region where
the structure is very flexible. The stiffness is regained only after the crown has deflected over
several times of its own thickness. Both the classical and the fully linearized buckling analyses
give fairly ‘linear’ and well-conditioned predicted-load curves in terms of the flexible region
(Table IV) where the corresponding pressures are between 0.22 psi and 0 3 psi. Accordingly,
the predicted critical load is, as stated, good for practical design purposes.

CONCLUSIONS
The validity of independent linear buckling analyses based on unstressed structure, for both the
classical and the fully linearized, is studied, and a methodology for predicting the critical buckling
loads of structures is presented. The fully linearized buckling analysis is first performed on the
unstressed structure. A proper load level is then imposed on the structure, after which a second
buckling analysis is conducted, The resulting critical loads are used to extrapolate the stability
limit of the structure. The fully linearized analysis is adopted as the tool of prediction since it
gives results that exhibit a better pattern of convergence over the deformation path than those
from the classical buckling analysis. Significant improvement of the stability limit prediction
from the extrapolation can be observed.
The credibility of the prediction is judged by the slope of the extrapolation curve, which is a
straight line in most cases. The effectiveness of the prediction cannot be assured if the slope of
the line is very close to 1, This case needs to be studied further and is usually associated with
very flexible structures with pronounced deformation. Extrapolation of higher orders may be
employed for cases where the predicted-load curve deviates considerably from straight lines.
The proposed prediction algorithm is applied to a number of test cases, and has given results
of significant practical value. It is expected that the proposed method can be used in place of
full nonlinear analysis for the majority of structures encountered.

ACKNOWLEDGEMENT

The authors are grateful to the Department of Civil Engineering at National Taiwan University,
Taipei, Taiwan, for the support and easy access in computing resources.

REFERENCES
1. S. B. Batdorf, ‘A simplified method of elastic stability analysis for thin cylindrical shells’, N A C A TR-No. 874,
Washington D.C. (1947).
846 S.-C. CHANG AND J.-J. CHEN

2. D. 0. Brush, ‘Prebuckiing rotations and cylinder buckling analysis’, Proc. ASCE, 106(EM2), (April 1980).
3. J. G. A. Croll, ‘Towards simple estimates of shell buckling load’, Der Stahlbau (Sept. 1975).
4. B. Brendel and E. Ramm, ‘Linear and nonlinear stability analysis of cylindrical shells’, Comp. Struct., 12, 549-558
(1980).
5. S.-C. Wu, ‘An integrated system for finite element shell analysis-surface representation and curved shell element’,
Dept. of Structural Eng. Re., Report No. 80-7, Cornell Univ., New York.
6. S.-C. Chang, ‘An integrated finite element nonlinear shell analysis system with interactive computer graphics’, Dept of
Structural Eng. Res., Report No. 82-4, Cornell Univ., New York.
7. S. P. Timoshenko and J. M. Gere, Theory of Elastic Stability, McGraw-Hill, New York, 1961.