Computers & Structures Vol. 66, No. 5, pp.

705±709, 1998
# 1998 Elsevier Science Ltd. All rights reserved
Printed in Great Britain
PII: S0045-7949(97)00077-1 0045-7949/98 $19.00 + 0.00

A NEW ARC-LENGTH METHOD FOR HANDLING SHARP

SNAP-BACKS
H. -B. Hellweg$ and M. A. Cris®eld%
$Continental AG, Tyre Research, Postfach 169, 300001 Hanover, Germany
%Department of Aeronautics, Imperial College of Science, Technology & Medicine, London SW7 2BY,
U.K.

(Received 31 August 1995)

AbstractÐVarious constraints for the arc-length method have been successfully employed to pass limit
points. However, at very sharp snap-backs, particularly those occurring in fracture mechanics or
damage mechanics both for failure initiation and crack propagation, standard arc-length methods can
fail to converge and alternative solution procedures are required. A novel algorithm to choose the cor-
rect root for the cylindrical arc-length method is suggested here. It will be demonstrated that this new
arc-length method is robust and can handle very sharp snap-backs. # 1998 Elsevier Science Ltd

INTRODUCTION these circumstances, it might be better to revert to a

During the progressive failure analysis of a double method such as the cylindrical procedure [3], which
cantilever beam (DCB) loaded under pure mode I, automatically includes elements of constraint for all
using interface elements employing a softening type variables.
damage model [1, 2], it was found that the solutions However, as previously discussed, in its present
form the method is not very robust, particularly
often failed to converge beyond a critical displace-
with sharp snap-backs. This paper will show that
ment or load when the conventional cylindrical arc-
the main limitation is related to the choice of roots
length method [3, 4] was used. The convergence pro-
and hence it is possible to simply amend the
blems frequently coincided with the crack initiation,
method while retaining the main elements of the
or started in the initial stages of the crack propa-
existing algorithms.
gation. Convergence problems of some sort per-
sisted for nearly all simulations of DCB specimens.
In some cases they disappeared for particular com- STANDARD ARC-LENGTH METHOD
binations of the ®nite element mesh and the energy
To overcome limit points (Fig. 1), various con-
release rate, but then reappeared for longer crack
straint methods have been proposed (eg normal
lengths. Characteristic of the convergence problems
plane or updated normal plane method, the domi-
were oscillation cycles, where the iterative solutions
nant displacement method, or the spherical or
circled in®nitely between a number of trial solutions
cylindrical arc length method [8±13]). In constraint
and the solution failed to converge. Problems with
methods, a load level parameter l is introduced to
existing arc-length methods have been reported by a
adjust the load level. It is assumed that the problem
number of authors [5±7] and are often associated
involves the monotonic factoring of a ®xed load
with snap-backs (which may be local). One success- vector f so that fn=lnf. The constraint methods are
ful solution [5] involves the use of the crack opening e€ectively Newton Raphson procedures with an ad-
displacement to form the constraint equation. These ditional constraint to de®ne the direction of the
procedures inevitably involve relative displacements iterative path/surface. Many constraint equations
between two (or more) nodes and consequently have been proposed such as the spherical or the
require pre-knowledge of the crack position and its cylindrical arc length method [4], the normal plane
local topology. or updated normal plane method [11] and the domi-
It might be possible to generalize such methods nant displacement method [5, 12].
so as to limit this pre-knowledge, but the coupling In the standard method the iterative displace-
between the resulting constraint relation and the el- ments which account for a ¯exible load level are
ement connectivity would still be inevitable. given by:
Particularly, once multiple cracks are considered, n
very considerable diculties are anticipated. In dpni ˆ dp ni ‡ dlni d^
pi 1†

705
706 H.-B. Hellweg and N. H. Cris®eld

Fig. 1. Limit points during a snap-through and a snap-back response.

is obtained when a cylindrical surface is de®ned,

where the in¯uence of the loads is neglected [13].
t
Dpni Dpni ˆ Dl 2 ˆ const 6†

Substituting Equation (4) in 6 gives a quadratic

equation for the iterative change in load level dl.
The standard way to avoid the doubling back of
the solution (see Fig. 3), is to select the root where
the corresponding incremental displacements encir-
cle the smaller angle [3].
t
€ Dpni Dpniÿ1 † ˆ min 7†

The following numerical results relate to a ®nite el-

Fig. 2. Principle of the arc length method. ement model of a DCB in which the non-linearity
was contained within interface elements which were
given softening properties related to the fracture
where dpni are the iterative displacements{ related to
energy [1]. The main modelling details have no sig-
the residual force vector as in the standard Newton
ni®cant bearing on the current developments apart
Raphson method
from the distinction between ``reversible'' and ``irre-
d p ni ˆ ÿKtiÿ1n gni 2† versible'' materials. The former assumes a comple-
tely reversible ``stress±strain'' law, while the latter,
with the residual gni being computed at the current
on unloading the ``stress±strain curve'' follows a lin-
estimate of the total load parameter lniÿ 1 and dpÃni
ear path back towards the origin. The irreversible
are the iterative displacements corresponding to the
model will also be referred to as an ``elastic damage
total load vector f
model''.
n
pi ˆ Ktiÿ1n f
d^ 3† When we applied the arc length method to simu-
late the crack propagation in a DCB, the conver-
Knowing the iterative displacements, the incremen-
gence problems described in the previous sections
tal displacements can then be updated as
were encountered: the solution oscillated and sub-
n
Dpni ˆ Dpniÿ1 ‡ dpni ˆ Dpniÿ1 ‡ d p ni ‡ dlni d^
pi 4† sequently failed to converge beyond a critical crack
The iterative load level dl is computed from the
constraint equation. In Fig. 2 the principle of
Cris®eld's spherical arc length method is
illustrated [4]. He de®nes a spherical constraint sur-
face by introducing the arc length D l which is kept
®xed during an increment.
t
Dpni Dpni ‡ Dlni f t Dlni f ˆ Dl 2 ˆ const 5†

An equally ecient but simpler constraint condition

{The subscript i denotes the current iteration whereas the

superscript n denotes the increment number. Fig. 3. Illustration of a solution doubling back.
 Handling sharp snap-backs 707

Table 1. Residual at critical increment

Increment Residual
number Root Angle norm
18/1 a 30.588 0.45273E + 04
18/1 b 118.638 0.69244E ÿ 03

NEW ARC-LENGTH METHOD

The convergence problems occurred when we

selected the root corresponding to the smaller angle
between the incremental displacements and the dis-
Fig. 4. Snap-back response during unstable crack growth. placements from the predictor solution using
Equation (7). In Fig. 6 the angles for both roots at
the critical increment are plottedÐthe smaller angle
a results in the doubling back, whereas with the big-
lngth. A restart with a very small arc length in the ger angle b we can immediately follow the unstable
proximity of a critical crack length showed an inter- path, which snaps back sharply. Consequently, an
esting phenomenon. The solution for a reversible alternative criterion for the root selection is
material model doubled back on itself, returned, required.
and then snapped back and continued to sub- Clearly a better technique is required to choose
sequently propagate. In Fig. 4 the converged incre- the appropriate root in the presence of sharp snap-
ments for an extreme case of an unstable crack backs. The proposed method has its origin in Fig. 7,
propagation are shown. It can be seen that before which indicates that the residual following the ap-
the solution followed the unstable path, it doubled plication of the ``wrong solution'' (although smal-
back several times. With an elastic damage material lest angle), is likely to be much bigger than that
model, we were not able to follow the unstable associated with the ``correct solution''. To apply
path, but were trapped in oscillation cycles. Figure this idea, it is necessary to try both alternatives,
5 illustrates the iterative path for the reversible and and update the displacements and compute the
elastic damage model at the critical increment resulting residuals for both cases. The resulting
(increment 18): whilst the reversible model doubles overhead is not very large and is roughly the equiv-
back on itself, the elastic damage model oscillates alent of the cost of one line-search. In relation to
between the predictor and the ®rst trial solution. the previous example, Table 1 shows the data as-
Even though the reversible material model exhibited sociated with the two separate roots at increment
the remote possibility of reaching convergence, the 18 (compare with Fig. 6). To save computer
standard arc-length method is clearly not satisfac- resources this more sophisticated procedure needs
tory and alternative solution procedures are impera- only to be switched on and activated when the user
tive. anticipates sharp snap-backs. In summary, the pro-

Fig. 5. E€ect of the unloading mode on the iterative path during a snap-back.
708 H.-B. Hellweg and N. H. Cris®eld

Fig. 6. E€ect of the choice of roots on the iterative path (reversible un-/loading).

posed new root selection is based on trying both nt

solutions and choosing the one with the minimum pi Dpniÿ1 ‡ d p ni †
c1 ˆ 2d^
residual norm (either Euclidian or maximum). The
steps are outlined in the following algorithm (here t n t t
c0 ˆ Dpniÿ1 D^
piÿ1 ‡ 2Dpniÿ1 p ni ‡ d p ni d p ni ÿ Dl 2
using the maximum norm).
1. standard arc length algorithm up to the compu- store dlni,j j = 1,2 set j = 1
tation of iterative vectors Equation (2 and3) 3. update incremental load level
2. solve quadratic equation for both roots dlni,j Dlni,j=Dlniÿ 1+dlni,j and total load level
Equation (4 and6) ln=ln ÿ 1+Dlni,j
2 4. compute iterative displacements Equation (1)
c2 dlni ‡ c1 dlni ‡ c0 ˆ 0
5. update current total displacements, strains,
nt n
stresses and internal forces
c2 ˆ d^
pi d^
pi 6. compute residual gni,j=lnf + qni,j

Fig. 7. Illustration of the minimum residual criterion.

 Handling sharp snap-backs 709

AcknowledgementsÐThis work was supported by a DTI/

LINK scheme which the authors gratefully acknowledge.

REFERENCES

1. Cris®eld M. A. and Hellweg H.-B., Some recent work

with Reinforced Concrete and other Composite
Materials,, University of Capetown Printing
Department, South Africa, Vol. 1, ed. D. Lanksch.
University of Capetown Printing Department, South
Africa, 1995, pp. 106±122.
2. Hellweg H.-B., Cris®eld M. A. and Davies G. A. O.,
Fig. 8. Snap-back response revisitedÐmodi®ed choice of Failure analysis of composite structures using interface
roots. elements. In Proc NAFEMS Conf. Application of
Finite Elements to Composite Materials, ed. R. E.
Eggington. 1994, pp. 6±9.
7. if the solution has converged, continue with 3. Cris®eld M. A., Non-Linear Finite Element Analysis of
next increment (n + 1) Solids and Structures, Vol. 1. John Wiley & Sons,
8. if the solution has not converged Chichester, 1991.
4. Cris®eld, M. A. A fast incremental/iterative solution
(a) if j = 1 set j = 2 and continue with 3. procedure that handles snap through. Computers &
(b) if j = 2 take dlni,j and dpni,j for min[abs(gni,1); Structures, 1981, 13, 55±62.
abs(gni,2)], update the incremental load level and 5. De Borst, R. Computation of post-bifurcation and
the incremental displacements and continue post-failure behaviour of strain-softening solids.
Computers & Structures, 1987, 25, 823±829.
with next iteration (i + 1). 6. Schellekens J. C. J., Computational strategies for com-
By selecting the root with the minimum residual, posite structures, Ph.D. thesis at Delft University
Netherlands, 1992.
we can follow the load displacement curve continu- 7. Duan Y. and May I. M., A local arc-length procedure
ously for both material unloading modes (Fig. 8). for strain softening. ACME Proc. Computational
Mechanics in UKÐ1994, ed. Geotechnical Engineering
Group, Department of Mechanical Engineering
SUMMARY AND CONCLUSIONS Manchester and UK. 1994, pp. 65±68.
8. Wempner, G. A. Discrete approximations related to
The cylindrical and spherical arc-length methods nonlinear theories of solids. International Journal of
su€er from convergence problems. This applies par- Solids and Structures, 1971, 7, 1581±1599.
ticularly to local instabilities in the form of snap- 9. Bergan, P. G. Solution algorithms for nonlinear struc-
backs. The convergence problems are connected to tural problems. Computers & Structures, 1980, 12,
497±510.
an incorrect choice of roots for the quadratic con- 10. Riks, E. An incremental approach to the solution of
straint equation. In conventional methods, the root snapping and buckling problems. International Journal
which gives the smallest angle between the previous of Solids and Structures, 1979, 15, 524±551.
and the current incremental displacements is 11. Ramm E., Strategies for tracing non-linear responses
selected. It has been shown in the paper that a near limit points. In Non-linear Finite Element
Analysis in Structural Mechanics, ed. W. Wunderlich,
more robust and ecient criterion is the minimum E. Stein and K. J. Bathe. Springer, Berlin, 1981.
residual. This new criterion can easily be added to 12. Rheinboldt, W. C. Numerical analysis of continuation
an existing arc-length method by including one ad- methods for nonlinear structural problems. Computers
ditional loop over the number of roots. With only & Structures, 1981, 13, 103±113.
13. Cris®eld, M. A. An arc-length method including line
the additional overhead of approximately one line- searches and accelerations. International Journal of
search, paths with very sharp snap-backs can now Numerical Methods in Engineering, 1983, 19, 1269±
be followed successfully. 1289.