COMMUNICATIONS IN NUMERICAL METHODS IN ENGINEERING, Vol.

14, 51±58 (1998)

A USER-CONTROLLED ARC-LENGTH METHOD

FOR CONVERGENCE TO PREDEFINED
DEFORMATION STATES

J. G. TENG* AND Y. F. LUO

Department of Civil and Structural Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong

SUMMARY
The arc-length method has been widely used in non-linear analysis of structures. While it can handle
complex load-de¯ection paths e€ectively, the analyst has no control over the load incrementation scheme to
achieve convergence to speci®c locations along the load-de¯ection path. There are a number of situations in
which such control is necessary, for example when convergence to prede®ned load levels, displacement levels
or stress levels is required. This paper describes an improved arc-length method in which this de®ciency is
overcome. # 1998 John Wiley & Sons, Ltd.
Commun. Numer. Meth. Engng, 14, 51±58 (1998)

KEY WORDS arc-length method; user control; accumulated arc-length control; non-linear analysis; prede®ned
deformation states; convergence

INTRODUCTION
Among the many existing incremental iterative non-linear solution methods, the arc-length
method1±4 as given by Cris®eld4 for application in ®nite element analysis appears to be the most
popular.5 The arc-length and other non-linear solution methods with similar capabilities
generally rely on iterating for a constant increment in a control parameter which is neither a pure
load nor a pure displacement parameter. Instead, the control parameter captures the e€ects from
both load change and displacement change, so both load limit points (snap-through points) and
displacement limit points (snap-back points) can be handled e€ectively. While the arc-length
method (or other methods with similar capabilities) can handle complex load-de¯ection paths
e€ectively, the analyst has no control over the load incrementation scheme to achieve converg-
ence to speci®c locations along the load-de¯ection path. There are a number of situations in
which such control is required.
When non-linear structural analysis programs are used in structural design, both the ultimate
load carrying capacity and the deformations under service loads need to be found. The arc-length
method is suitable for the former task, but not the latter. Precisely for this reason, Chan6 argued
for the inclusion of a number of solution procedures in such programs. The analyst needs to
conduct two separate analyses if he needs to know both the ultimate load and the state of the
structure at a particular load level. This is inconvenient for both programming and program

* Correspondence to: Prof. J. G. Teng, Department of Civil and Structural Engineering, The Hong Kong Polytechnic
University, Hung Hom, Kowloon, Hong Kong.

CCC 1069±8299/98/010051±08$17.50 Received 31 January 1997

# 1998 John Wiley & Sons, Ltd. Accepted 25 July 1997
52 J. G. TENG AND Y. F. LUO

application. Other situations in which convergence to prede®ned deformation states is required

include the assessment of dynamic characteristics of structures at a given level of loading and ®rst
yield design of structures which exhibit strong geometric non-linearity as demonstrated later in
the paper. This paper thus presents an improved form of the arc-length method of Cris®eld4 so
that it can achieve convergence to prede®ned load levels, or prede®ned values of any other
parameters which are a function of load.

THE CONVENTIONAL ARC-LENGTH METHOD AND THE CONCEPT

OF ACCUMULATED ARC-LENGTH
In the solution of a system of non-linear equations for a ®nite element model of a structure using
the arc-length method,4 the accumulated displacements of the ith loading step after the jth
iteration are given by

j j ÿ1 j j
fDda gi ˆ fDda gi ‡ Dli fdI gi ‡ fDdR gi …1†

in which Dlji is the change of the load factor for the jth iteration within the ith loading step, {dI}i is
the tangent displacement vector under an arbitrary reference load level, and {DdR}ji is the vector
of displacement increments due to residual forces. The change of load factor Dl is constrained by
the arc-length increment li for the ith loading step through the following equation:

T
fDda gij fDda gij ˆ l 2i …2†

The arc-length increment li for the ith loading step is determined by the following procedure:3

p
l i ˆ l iÿ1 …J d =J iÿ1 † …3†

in which li71 is the arc-length increment of the previous loading step, Jd is the desired number of
iterations for the ith loading step before convergence and Ji71 is the number of iterations required
to converge in the (i 7 1)th loading step.
The concept of an accumulated arc-length was ®rst proposed by Teng and Luo7 for use with
the conventional arc-length method in non-linear bifurcation analysis to achieve convergence to
an anticipated bifurcation point. The accumulated arc-length at the ith loading step Li is de®ned
as the sum of all arc-length increments up to and including the ith loading step

X
i
Li ˆ lk …4†
kˆ1

in which lk is the arc-length increment of the kth loading step. The parameter L represents the
current state of the structure and depends not only on the characteristics of the structure and its
loading but also on the process of load incrementation during the analysis.

COMMUN. NUMER. METH. ENGNG, VOL. 14, 51±58 (1998) # 1998 John Wiley & Sons, Ltd.
 USER-CONTROLLED ARC-LENGTH METHOD 53

ACCUMULATED ARC-LENGTH CONTROL FOR CONVERGENCE

TO PREDEFINED LOAD LEVELS
General
For simplicity of description, it is assumed here that convergence is required at a single
prede®ned load level where the load factor is ld , although a number of prede®ned load levels can
be found without extra diculty. Initially the conventional arc-length method as described above
is used to trace the load-de¯ection path, during which continuous monitoring is carried out to see
if the converged load level li is near the pre-de®ned load level ld . The accumulated arc-length
process proposed below is started once the load factor li at the convergence of the ith loading step
is close to ld by a rational criterion.

Quadratic approximation to the desired arc-length increment

To facilitate the formulation, it is necessary to de®ne a new parameter c as

c…Li † ˆ li ÿ ld …5†

The desired arc-length increment for the (i ‡ 1)th loading step ld is to make the accumulated
arc-length Ld satisfy the following equation:

c…Ld † ˆ 0 …6†

where

Ld ˆ Li‡1 ˆ Li ‡ l i‡1 ˆ Li ‡ l d …7†

Taylor's expansion of equation (6) leads to

dc…Li † 1 d2 c…Li † 2
c…Ld † ˆ c…Li ‡ l d † ˆ c…Li † ‡ ld ‡ l ‡


ˆ 0 …8†
dL 2 dL2 d

Omitting third and higher order terms as ld is small compared to Li results in the following
quadratic equation:

dc…Li † 1 d2 c…Li † 2
c…Li † ‡ ld ‡ l ˆ0 …9†
dL 2 dL2 d

The ®rst derivatives of c with regard to L at loading steps i and i 7 1 are approximated by the
following backward ®nite di€erence formulae:

dc…Li † c…Li † ÿ c…Liÿ1 † l ÿ liÿ1

ˆ ˆ i …10†
dL Li ÿ Liÿ1 Li ÿ Liÿ1
dc…Liÿ1 † c…Liÿ1 † ÿ c…Liÿ2 † l ÿ liÿ2
ˆ ˆ iÿ1 …11†
dL Liÿ1 ÿ Liÿ2 Liÿ1 ÿ Liÿ2

# 1998 John Wiley & Sons, Ltd. COMMUN. NUMER. METH. ENGNG, VOL. 14, 51±58 (1998)
54 J. G. TENG AND Y. F. LUO

and the second derivative is again evaluated based on the ®nite di€erence approximation

dc2 …Li † dc…Li †=dL ÿ dc…Liÿ1 †=dL

ˆ …12†
dL Li ÿ Liÿ1

Introducing (10)±(12) into (9), the new arc length increment for the (i ‡ 1)th loading step to
reach the prede®ned load level can be found by solving the above quadratic equation. The
appropriate root is the one which is closer to the linear solution. By analysing the roots of (9), it
can be demonstrated that this root is in general much smaller than the other root in terms of
absolute values (the sign of the appropriate root may be positive or negative depending on the
location of the desired point). Theoretically speaking, imaginary roots are possible in some
special cases when the value of dC(Li)/dL approaches 0, but this very rarely occurs in numerical
calculations. In all the numerical calculations presented later, the problem of imaginary roots
did not appear. If the magnitude of ld given by (9) is larger than that given by (3), then the value
from (3) is used instead and a subsequent re-evaluation of ld will be carried out for the (i ‡ 2)th
loading step. Generally, the analysis converges to the prede®ned load level with a small number
of repeated applications of (9).

Linear approximation to the desired arc-length increment

Omitting the second-order term in (9) results in the following equation:
c…Li †
ld ˆ ÿ …13†
dc…Li †=dL

Introducing (10) into (13), the new arc-length increment ld for the (i ‡ 1)th loading step to reach
the desired load level is found as
ld ÿ l i l ÿ li
ld ˆ …L ÿ Liÿ1 † ˆ d l …14†
li ÿ liÿ1 i li ÿ liÿ1 i

Alternative quadratic approximation to the desired arc-length increment

An alternative quadratic approximation for the desired arc-length increment is formulated as
follows. This alternative formulation has the advantage that the solution of a quadratic equation
is not necessary and hence the issue of choosing the appropriate root and the possibility of
imaginary roots are removed. De®ning the accumulated arc-length L as a function of the load
factor
L ˆ L…l† …15†

after convergence has been achieved at a load factor of li at the end of the ith loading step, the
accumulated arc-length is equal to Li . Expanding the desired value of the accumulated arc-length
Ld corresponding to a load factor of ld about Li leads to

dLi 1 d2 Li 2
Ld ˆ L…ld † ˆ L…li ‡ Dld † ˆ Li ‡ Dld ‡ Dld ‡


…16†
dl 2 dl2

COMMUN. NUMER. METH. ENGNG, VOL. 14, 51±58 (1998) # 1998 John Wiley & Sons, Ltd.
 USER-CONTROLLED ARC-LENGTH METHOD 55

where

Dld ˆ ld ÿ li …17†

Retaining only the ®rst- and second-order terms in (16) leads to the expression for the desired
arc-length increment

dLi 1 d2 Li 2
l d ˆ Ld ÿ Li ˆ Dld ‡ Dld …18†
dl 2 dl2

in which the derivatives of L with regard to l are approximated by the following ®nite di€erence
expressions:

dLi L ÿ Liÿ1
ˆ i …19†
dl li ÿ liÿ1
d2 Li dLi =dl ÿ dLiÿ1 =dl
ˆ …20†
dl2 li ÿ liÿ1

If only the linear term in (18) is retained, the resulting formula for the desired arc-length
increment is the same as that of (14).

GENERALIZATION FOR OTHER PARAMETERS

The procedure described above can be generalized for use to achieve convergence to prede®ned
values of other parameters such as displacement or stress. Assuming that the parameter to be
controlled is Z and convergence should be achieved to a prede®ned value Zd , then in all the above
equations l should be replaced by Z and ld by Zd .

NUMERICAL EXAMPLES
The method proposed here has been coded into the latest version of the NEPAS program,8,9
which can perform elastic±plastic large de¯ection analysis of shells of revolution subjected to
combined axisymmetric loading including torsion. Two numerical examples are presented below
to demonstrate the validity and capability of the proposed method.
A clamped spherical cap under a reference load equal to the classical buckling pressure p0 for a
spherical shell is considered ®rst (Figure 1). The NEPAS program was instructed to trace the
load-de¯ection path for a cap with a shallowness parameter l (Figure 1) of 5.5 and in the
meantime to converge to all points with a load factor of 0.6. The load-de¯ection path determined
by the conventional arc-length method (Figure 2) shows that no converged point lies close to a
load factor of 0.6. Figure 3 shows the load-de¯ection path traced with convergence to a load
factor of 0.6 using the linear version of accumulated arc-length control. The converged points
with the load factor equal to 0.6 are shown as large crosses and other converged points as circles.
The method is seen to be able to locate all three points corresponding to the prede®ned load
factor of 0.6 along the whole load-de¯ection path precisely. A comparison between the results
from the linear version and those from the two quadratic versions shows that the simple
linear version works best for this problem, requiring only three trials each time; the alternative

# 1998 John Wiley & Sons, Ltd. COMMUN. NUMER. METH. ENGNG, VOL. 14, 51±58 (1998)
56 J. G. TENG AND Y. F. LUO

Figure 1. Spherical cap under uniform external pressure

Figure 2. Load-de¯ection path tracing by conventional arc-length method

quadratic version is the worst (3, 3 and 5 trials for the three points, respectively) and the quadratic
version is in between (3, 3 and 4 trials). The linear version appears to o€er a simple and e€ective
approach, but a general conclusion cannot be drawn until further evidence becomes available.
The arc-length solution procedure was carried out with the following values for the control
parameters: the tolerance for convergence of the non-linear iterative process for each loading
step ˆ 0.001 based on the criterion speci®ed in Reference 8, the desired number of iterations for
each loading step ˆ 4, and the starting load factor for the ®rst loading step ˆ 0.2. In the analyses,
the prede®ned load level was judged to have been reached if the di€erence between the converged
and prede®ned load factor is less than 1075, which is tighter than may be needed in many
practical situations.

COMMUN. NUMER. METH. ENGNG, VOL. 14, 51±58 (1998) # 1998 John Wiley & Sons, Ltd.
 USER-CONTROLLED ARC-LENGTH METHOD 57

Figure 3. Load-de¯ection path tracing with convergence to a prede®ned load level

Figure 4. Load-de¯ection path tracing with convergence to multiple prede®ned stress levels

The proposed approach is equally e€ective for convergence to a prede®ned value of any other
parameter and to multiple values of a parameter. As an example, an elastic circular plate with a
completely ®xed edge was analysed for convergence to multiple stress levels. The plate has the
following properties: radius ˆ 100 mm, thickness ˆ 1 mm, Young's modulus ˆ 2.0  105 MPa
and Poisson's ratio ˆ 0.3. The reference load of the plate is 1 MPa. The NEPAS program was
instructed to ®nd the load levels at which the maximum von Mises equivalent stress is equal to
100 MPa, 200 MPa and 250 MPa, respectively (Figure 4). A prede®ned stress level was judged to
have been reached if the di€erence between the converged and prede®ned stress level is less than
0.1 per cent. This kind of analysis is useful in design application of non-linear analysis where
®rst yield constitutes a limit state as it can precisely determine the load level corresponding
to ®rst yield. The starting load factor for the ®rst loading step ˆ 0.001 and the values of the
other two control parameters in the arc-length solution procedure are the same as for the ®rst
example.

# 1998 John Wiley & Sons, Ltd. COMMUN. NUMER. METH. ENGNG, VOL. 14, 51±58 (1998)
58 J. G. TENG AND Y. F. LUO

CONCLUSIONS
This paper has presented an improved arc-length method which can converge to prede®ned load
levels, or prede®ned values of other parameters. This user control is achieved by coupling the
accumulated arc-length control procedure of Teng and Luo7 with the existing arc-length method
of Cris®eld.4 The method is easy to implement in both existing and new computer programs for
non-linear analysis, and can be useful in many practical applications. A particularly useful
application lies in the use of non-linear analysis in structural design as demonstrated by the
numerical examples. The study also shows the way for augmenting other non-linear solution
procedures to achieve convergence to prede®ned load levels, or prede®ned values of other
parameters. Numerical examples have demonstrated the e€ectiveness of the proposed method.

ACKNOWLEDGEMENT

This research was supported by a research grant from the Hong Kong Polytechnic University,
for which the authors are most grateful.

REFERENCES

1. G. A. Wempner, `Discrete approximations related to nonlinear theories of solids', Int. J. Solids Struct., 7,
1581±1599 (1971).
2. E. Riks, `An incremental approach to the solution of snapping and buckling problems', Int. J. Solids
Struct., 15, 529±551 (1979).
3. E. Ramm, `Strategies for tracing nonlinear response near limit points', in Nonlinear Finite Element
Analysis in Structural Mechanics, W. Wunderlich, E. Stein and K. J. Bathe (Eds.), Springer-Verlag,
New York, 1981, pp. 68±89.
4. M. A. Cris®eld, `A fast incremental/iterative solution procedure that handles snap-through', Comput.
Struct., 13, 55±62 (1981).
5. M. J. Clarke, and G. J. Hancock, `A study of incremental-iterative strategies for nonlinear analysis',
Int. J. Numer. Methods Eng., 29, 1365±1391 (1990).
6. S. L. Chan, `A generalized numerical procedure for nonlinear analysis of frames exhibiting a limit or a
bifurcation point', Int. J. Space Struct., 6(2), 99±114 (1991).
7. J. G. Teng and Y. F. Luo, `Post-collapse bifurcation analysis of shells of revolution by the accumulated
arc-length method', Int. J. Numer. Methods Eng., 40(13), 2369±2383 (1997).
8. J. G. Teng and J. M. Rotter, `Elastic-plastic large de¯ection analysis of axisymmetric shells',
Comput. Struct., 31(2), 211±235 (1989).
9. J. G. Teng and J. M. Rotter, `Non-symmetric bifurcation of geometrically non-linear elastic-plastic
axisymmetric shells subject to combined loads including torsion', Comput. Struct., 32(2), 453±477 (1989).

COMMUN. NUMER. METH. ENGNG, VOL. 14, 51±58 (1998) # 1998 John Wiley & Sons, Ltd.