Sunteți pe pagina 1din 17

A R C - L E N G T H M E T H O D FOR PASSING L I M I T POINTS

IN STRUCTURAL CALCULATION
By W. F. Lam 1 and C. T. Morley 2

ABSTRACT: A new solution procedure is presented, based on the arc-length method,


Downloaded from ascelibrary.org by QUEEN'S UNIVERSITY on 04/22/13. Copyright ASCE. For personal use only; all rights reserved.

for passing limit points (load or displacement peaks) in nonlinear finite element
analysis of structures. In addition to the usual equilibrium equations, a quadratic
arc-length constraint equation is specified so that the nonlinear solution is sought
on a small ellipsoidal surface in load-deflection space. The main new feature of
the proposed procedure is the resolution of the out-of-balance loads into a parallel
and an orthogonal component with respect to the vector of applied external loads.
Methods for avoiding complex roots to the quadratic constraint equation, and for
selecting the appropriate root, are presented. Applications of the proposed pro-
cedure to two reinforced concrete problems give satisfactory results. The method
is effective and versatile in handling both snap-through and snap-back problems.

INTRODUCTION

In finite element structural calculations, with an incremental/iterative


method of allowing for nonlinearities due to large geometry change and/or
material behavior, it is important to incorporate an effective solution pro-
cedure for finding and proceeding beyond peaks (limit points) on the load-
deflection curve. For example, in reinforced concrete structures there may
be several load peaks under service loads, due to successive cracking, before
the ultimate maximum load is reached. An effective algorithm to allow an
analysis to proceed past such peaks is clearly beneficial, and is often essential
if ultimate collapse is to be properly defined, to reveal the collapse mode
as well as the maximum load, and throw light on the ductility or deformation
capacity of the structure and on the nature of the failure generally.
In most iterative schemes, the aim is to reduce to less than a specified
acceptable limit the out-of-balance loads, which are the difference between
(1) the nodal loads calculated by the computer as corresponding to the
current assumed displacements and material properties; and (2) the applied
nodal loads at that stage. Once convergence has been achieved, some struc-
tural variable is incremented a small amount further, and renewed iterations
are undertaken to convergence.
Attempts to apply positive load increments will naturally encounter dif-
ficulties near load peaks, and methods of arranging positive stiffness by
additional artificial springs (Wright and Gaylord 1968; Ramm 1981) or of
allowing load reductions within load control (Cope and Rao 1981; Bergan
and Holand 1979; Crisfield 1982b; Phillips and Zienkiewicz 1976) require
great care and are not always satisfactory. Snap-through (but not snap-back)
problems can be tackled by displacement control, treating the load intensity
as unknown (Batoz and Dhatt 1979), but proper choice of the pattern(s) of
displacement to control can be difficult. Two methods proposed by Bergan
'Engr., Howard Humphreys and Partners Ltd., 21a-37a South St., Dorking, Surrey
RH42
2J2, England.
Lect., Univ. Engrg. Dept., Trumpington St., Cambridge CB2 1PZ, England.
Note. Discussion open until June 1, 1992. To extend the closing date one month,
a written request must be filed with the ASCE Manager of Journals. The manuscript
for this paper was submitted for review and possible, publication on April 10, 1990.
This paper is part of the Journal of Structural Engineering, Vol. 118, No. 1, January,
1992. ASCE, ISSN 0733-9445/92/0001-0169/$1.00 + $.15 per page. Paper No.
26518.
169

J. Struct. Eng. 1992.118:169-185.


have given satisfactory results: the current stiffness parameter method of
identifying limit points (Bergan et al. 1978), and the unbalanced force min-
imization technique of adjusting external load levels (Bergan 1980). How-
ever, in a survey of available methods for passing load peaks, Ramm (1981)
concludes that the most versatile is a modified form of the constant arc-
length method of Riks (1972, 1979) and Wempner (1971). This method has
Downloaded from ascelibrary.org by QUEEN'S UNIVERSITY on 04/22/13. Copyright ASCE. For personal use only; all rights reserved.

been successfully applied by Ramm (1981) and Crisfield (1981,1982a, 1983,


1984) to a wide range of problems.
In this paper, some difficulties with the usual form of the constant arc-
length method are identified. A reformulation of the method is then pre-
sented, incorporating modifications to overcome these difficulties, and ap-
plications of the proposed procedure to some reinforced concrete problems
are described.

CONSTANT ARC-LENGTH METHOD

As in displacement control, the arc-length method treats the applied load


intensity X as an additional variable during equilibrium iterations, so that
curves of load against displacement(s) p that decline or fold back can be
explored. The load increment is governed by an extra constraint equation
with the general form
ApT Ap,- + A0 A\2tqTq = A/2 (1)
where Ap, = the trial incremental displacement vector at the rth iteration;
AQ = a scalar parameter that governs the relative contributions of the
displacement and load increments; AX., = the predicted incremental load
factor; q = the specified external load vector; and A/ = the specified arc
length for the current increment. The intention is that the solution should
not stray too far, in load-displacement space, from the last converged point
in fact, the solution is sought on a small ellipsoidal surface surrounding the
last converged point, of size determined by the prescribed arc length A/.
In practice, at the rth iteration, displacement increments Ap, satisfying
(1) are assumed, and structural calculation is carried out to find the cor-
responding out-of-balance loads e,, the difference between the desired load-
ing Xq and the loading calculated to correspond to the total displacements
p,. If the out-of-balance is sufficiently small, the solution has converged; if
not, a new set of displacement increments must be chosen for the next
iteration (/ + 1), hopefully by making estimates from the results of the
i-th iteration.
Riks and Wempner adopted a linearized version of constraint equation
(1) for the (i + l)th iteration, represented by
Sf Ap,, + A0 8X,: AX.,- q r q = 0 (2)
where 8, and 8X, = the required corrections to Ap, and AX,-, respectively.
This results in the so-called fixed normal plane constraint, with the solution
iterates all lying on a hyperplane normal to the tangent at the last converged
point in load-displacement space.
Originally, constraint equation (2) was solved directly with the n equilib-
rium equations, making a total of (n + 1) simultaneous equations. This
procedure reduces the solution efficiency, since the symmetric banded na-
ture of the system stiffness matrix can no longer be exploited. Ramm (1981)
and Crisfield (1981, 1982b, 1983, 1984) rectified this by introducing an
170

J. Struct. Eng. 1992.118:169-185.


indirect solution scheme for the constraint equation (2). They related 8, to
8\, so that
8,- = K- 1 ^. + 8\, K *q (3a)
8,- = 8e, + 8X<8, (35)
_1
Downloaded from ascelibrary.org by QUEEN'S UNIVERSITY on 04/22/13. Copyright ASCE. For personal use only; all rights reserved.

where 8e, = K e,- = the displacement vector resulting from eliminating


e, the out-of-town balance load; and 8 9 = K~xq = the parallel displacement
vector computed and stored after every update of the stiffness matrix K.
The next trial incremental displacement can then be expressed as
Ap,- + x = Ap, + 8e, + 8\,8, (4)
Substituting Ap, + 1 in the constraint equation (1) with A0 = 0 (as frequently
adopted) results in a quadratic equation in 8\, so that
afiXf + o28X., + a3 = 0 (5a)
where
i = 8|B, (5b)
a2 = 28J (Ap,- + 8e/) (5c)
a3 = (Ap,- + 8 e ,) r (Ap,- + 8e;) - AP (5d)
The coefficients in (5) can all be computed from the results of the iih
iteration, so that the equation may be solved and a root chosen for 8X.,-. Eq.
(4) is then used to compute the next trial displacement increments Ap, +1 .
The procedure is known as the indirect spherical arc-length method. Cris-
field applied the technique to concrete structures where numerical difficul-
ties are particularly severe because of the sudden localization of strain on
cracking. Other proposals for the constraint equation can also be found in
the literature, e.g., Bathe and Dvorkin (1983) chose to use a fixed increment
of external work.
Despite the sophistication of the arc-length procedure of Ramm and Cris-
field certain difficulties remain in its application. The first concerns the
choice of root for the quadratic equation (5a). Crisfield proposes the eval-
uation of the scalar product (Ap r , +1 Ap,) for each root and then chooses
the root that gives the largest product, and hence the smallest angle between
Ap, and Ap, +1 . In a multidimensional problem with nonlinear physical or
geometrical behavior and/or unloading, there is some doubt whether the
use of Ap,- as the reference vector for choosing Ap, +1 is always the best
strategy. Secondly, complex roots of the quadratic equation (5a) sometimes
occur; the reasons for this are yet unclear, and the appropriate corrective
action uncertain. One can simply reduce the arc length AZ and try again
but this, by no means, always works.
The purpose of the next section is to present a reformulation of the arc-
length method, with a simplified graphical representation, in the hope of
making the method easier to comprehend and pointing the way to resolving
these difficulties.

MODIFIED ARC-LENGTH SOLUTION PROCEDURE

Basic Formulation
A static finite element problem is considered, with n unknown nodal
displacements p. Nodal loads q, of dimension n, are applied with load factor
A., and for each new value of X starting from zero, the solution p is sought.
171

J. Struct. Eng. 1992.118:169-185.


Typically (Fig. 1), a converged point (X-oq,Po) n a s been found, and at the
ith iteration thereafter, estimates AA.,q and Ap, have been computed of the
changes required to find the next converged point. The aim is to use com-
putations after the ith iteration to choose the next trial values A\, + 1 q and
Ap, +1 so that convergence can be achieved even beyond a limit point.
At the ith iteration, the estimated applied load E, is (X0 + AX,)q. From
Downloaded from ascelibrary.org by QUEEN'S UNIVERSITY on 04/22/13. Copyright ASCE. For personal use only; all rights reserved.

the displacement (p0 + Ap,), the program with nonlinear properties com-
putes the corresponding nodal loads F in equilibrium with the internal
forces, but in general not in agreement with E, nor in the specified load
direction q (see the left half of Fig. 1).
The out-of-balance loads e, = E,- - F, may be resolved into a component
g,-q in the specified applied load direction, where
e?q
gt (6)
qrq
and a component h, orthogonal to q, where
h, = e, &q (7)
If the out-of-balance loads e, or the orthogonal component h, satisfy the
convergence criterion, the solution is accepted, otherwise further iteration
is required.
To achieve convergence, elimination of at least the orthogonal component
h, must be attempted by amending the trial displacements in the next it-
eration, but the term g,q is not so important since the load intensity has not
been specified and may be recomputed. In the proposed arc-length method,
further load h, is applied, plus x,q in the load direction, where xt is a scalar
multiplier to arrange that the point C in Fig. 1 is always at a chosen fixed
arc length A/ from the last converged point. In this way, it is hoped that
difficulties due to snap-through or snap-back limit points will be overcome.
Supposing that the current stiffness matrix K, or an estimate of it, is
known or has been formulated, the next displacement increment is chosen
with

Load in specified
'direction'
Xq Xq
,
--.
X
Ef N C

Aj^
/-V
-th iteration \
AXiq Ay | * " " - i - t h iteration
1 "\

F 4 -,
\ X 0 ql Ap,
I

\ ^~ ast converged point


Loads orthogonal\
to q
0 i Displacement p
Po
la] 'Load' plane (b) Load-displacement graph

FIG. 1. Sketch of Typical Iterative Step


172

J. Struct. Eng. 1992.118:169-185.


Ap,+i = Ap, + K- 1 (x,q + Ti,h,) (8a)
Ap, +1 = Ap,- + xfi + TI,S,. (8b)
-1
where 8q = K~*q; S,. = K h,-; and T|,- = a scalar factor, for the time being
taken to be unity, whose use will be explained later.
The estimated applied load for the next iterative step is expressed as
Downloaded from ascelibrary.org by QUEEN'S UNIVERSITY on 04/22/13. Copyright ASCE. For personal use only; all rights reserved.

E / + 1 = F, + Xiq + h, (9a)
E / + 1 = E,- - g,q + x,q (9b)
i.e.
AX,-+1 = AX; - g, + x, (9c)
where xt is to be governed by the arc-length constraint equation
A/2 = Apr,.+ 1 Ap ;+1 + A0 AX?+1 (10)
The scalar A0 converts everything to the same dimensions and order of
magnitude. Many analysts set A0 = 0, but we take

A0 = ^ (11)

except at the origin {p0 = X0 = 0, [Fig. 1(b)]}.


Now, substituting (8b) and (9c) into (10) gives
ax] + 2bxi + c = 0 (12a)
where
a = A0 + SJB, (126)
b=A0 (AX,- - gl) + 8J (Ap, + -n,.8.) (12c)
2
c = >10(AX,. - g,) - AF + (AP/ + ^Mr (Ap; + -0,8,,) (12d)
All the terms in (12b)-(12d) can be computed at the rth iteration, xt
comes from (12a), and the next trial displacement values from (8). Usually
(12a) gives real roots, and we choose the one that enables solution advance
rather than retreat back down the original equilibrium path towards the
origin. Supposing the previous changes AXpr and Appr between the last
converged solutions are recorded, a scalar product Ap,^x Appr can be eval-
uated for each root, and the one that gives the largest product (hence, the
smallest angle between Ap'"' and Ap,+1) is chosen.

Complex Roots: First Method


Thus far, though presented differently, the approach could lead to the
same new estimates AX,+1 and Ap, +1 as those of Crisfield (1981, 1984) for
the same A0. However, by resolving the out-of-balance loads e, into parallel
and orthogonal components g,q and h,-, it is perhaps easier to see what steps
should be taken when (12a) gives complex roots. Now, complex roots imply
that it is not possible to find a real xt to give the chosen arc length one
reason may be that the orthogonal error h, is too large. One possible cor-
rective action is therefore to attempt to eliminate h,-, by setting
Ap- = Ap,- + 8W (13)
173

J. Struct. Eng. 1992.118:169-185.


(superscript cr = value related to the complex roots). From the displacement
(p0 + Apcr) the corresponding internal nodal load F cr is calculated, which
has a component X"q in the specified direction, where

X" = ^ (14)
Downloaded from ascelibrary.org by QUEEN'S UNIVERSITY on 04/22/13. Copyright ASCE. For personal use only; all rights reserved.

Using changes Apcr and AXcr, where AXcr = Xcr XQ, the actual arc length
is given by

A/Cr = V(Ap c O r (Ap") + A0 (AX")2 ^ 15 )

To revert back to the specified arc length A/, a scaling factor

|x = - ^ - (16)

is used to give the new estimates


Ap /+1 = jiAp" i. (17A)
AX,+1 = jiAX" . (176)

First Iteration from Converged Point


In each increment, the first estimates AXxq and Apt of the changes from
the last converged point (other than the origin) can conveniently be chosen
by projecting the solution forward from the last two converged points, i.e.,
projecting in the direction it is currently advancing. A scalar u.x is used,
where
A/
u, = (18)
KtWY (Ap"r) + A0 (A\"02
so that
Apx = |x1 ApPr (19a)
AXj = p.! AX"' (196)
Iteration continues until the error loads e, are sufficiently small in com-
parison with the current applied loads E,. Convergence has also occurred
if loads h, rather than e,- satisfy the convergence criterion, in which case we
disregard the arc-length requirement and compute X from F, using the method
of (14).

Convergence Criterion
In this work, the convergence criterion is expressed as a ratio of the norm
of the out-of-balance force e; to the norm of the total applied load X,q, in
each case expressed in consistent units. So the solution is taken to have
converged if

r= =S 8 (20)
x,-M,
174

J. Struct. Eng. 1992.118:169-185.


where X, = t n e c u r r e n t load factor. Typically the specified tolerance e is
1/500 or 1/1,000.

Solution Formulation at Origin


To define the first arc length at the start of an analysis, an initial load
increment A\ t is specified. However, at the origin both (11) and (10) are
Downloaded from ascelibrary.org by QUEEN'S UNIVERSITY on 04/22/13. Copyright ASCE. For personal use only; all rights reserved.

undefined. Suppose we consider a converged point close to the origin, so


that both X0q and p 0 are small, we have
Po = XoK-iq = X08? (21)
Substituting p 0 into (11) gives
A0 = 8J8, (22)
The first trial displacement from the origin is
A Pl = A X ^ (23)
So (10) becomes
A/2 = ApfAPl + A0 AX2 = 2AX2 8J8 (24)
The arc length Al from (24) is used for iterations near the origin, and is
later increased or decreased if convergence has been fast or slow, aiming
for about four iterations to each converged point. The adjustment used by
Ramm is adopted, so that

A/
= J-ZTr Al"r (25)
where Ide = the desired number of iterations to each converged point; Ipr
and bdpr = the actual number of iterations required and the arc length used
in the previous increment, respectively.
Normally, the tangent stiffness matrix K or an estimate of it is computed
every five iterations. Also, if convergence is not achieved in, say 20 itera-
tions, the practice here is to return to (X0q, p 0 ), recompute K, and proceed
again with a smaller arc length.

PREFERRED PROCEDURE FOR AVOIDANCE OF COMPLEX ROOTS

When applying the arc-length method proposed thus far, results showed
that, in many cases, the steps of (13)(17) do enable the analysis to proceed
and subsequently to find further converged points, in the event of complex
roots to the quadratic equation (12a). Yet occasionally, particularly in some
reinforced concrete analysis problems when there is severe strain localization
and unloading of adjacent elements on cracking, difficulties in solution
convergence still persist, while the procedure repeatedly gives complex roots.
This was also the experience of Crisfield (1983, 1984) who obtained con-
siderable improvement when he introduced line searches, by multiplying
the second and third terms on the right of (4) by a factor y\ + 1.0 and
determining T| by minimization of potential energy.
Because of the difficulties encountered, further work was necessary to
develop a solution procedure to tackle the problem of complex roots. Here,
a pseudo line-search technique is introduced, with a similar factor r\ that
is, however, computed without introducing energy considerations.
175

J. Struct. Eng. 1992.118:169-185.


Recalling the quadratic equation (12a),
2
ax + 2bxt + c = 0
where a, b, and c are defined in (12b)-(12d). For real roots, we must have
b2 - ac> 0, i.e., from (12fe)-(12d)
Downloaded from ascelibrary.org by QUEEN'S UNIVERSITY on 04/22/13. Copyright ASCE. For personal use only; all rights reserved.

[AQ (AX, - g) + &l (Ap, + i)$M)? ~ (A0 + 8JB) x


[A0 (AX, - giy - AP + (Ap, + u-8 w ) r (Ap, + T|,8;,)] > 0 . . . . . . (26a)
or

(i) T|,<K)

~<1-0<T1 2

b'
(Hi) I|,<1-0 <--^

(iv) 1-0 <T),

FIG. 2. Selection of Line-Search Factor t),

176

J. Struct. Eng. 1992.118:169-185.


A>[2(A\, - ft)8,r(Ap,- + -vfi,,,) - (Ap, + in,-8w)T(Ap,- + TI,8W)
2
- (AX,. - ft) 8JBJ - (8JB,)(Ap, + i i ^ ^ A p , + ^,8,,)
+ {[8J(Ap, + T,,.8;.)]2 + A0A/2 + 8J8,AZ2} > 0 (266)
Now, if A0 and A/ are to be kept unaltered within the current increment,
Downloaded from ascelibrary.org by QUEEN'S UNIVERSITY on 04/22/13. Copyright ASCE. For personal use only; all rights reserved.

a logical way to satisfy relation (26) is to introduce a nonunity factor in, on


8W. By expanding and regrouping terms, relation (26) becomes
V = a'rtf + 2b' T\I + d < 0 (27a)
where
a' = {A, + 8JB,)(88W) - (8JBW)2 (276)
6' = (A0 + 8?r89) (Apf8,.) - [A0(AX,. - ft) + SJ Ap,] (8|BW) . . . . (27c)
c' = (A0 + 8JB,) [(ApfAp,) - AP] - [2A0(A\, - ft) + (8JAP/)] x
[8|-ApJ + A0 (AX,- - ft)2 (SJB,) (27d)
All the terms here can be computed at the rth iteration (many are already
determined when formulating the quadratic equation (12a) in x,. The range
of values of T|,- for which the x,s are real then comes from relations (27).
Usually relation (27a) gives real roots (when equality holds), say % and T|2
(where r\2 s T|I). For convenience and simplicity, a suitable -n; is chosen
close to 1.0, say at 5% of |T)2 %( within the valid range for real xts, that
is [Fig. 2, which covers cases where in = 1.0, gives complex roots to (12)]
TI, = Ti2 - i if ife < 1.0 (28a)

Tl, = Tfe + ? if - -- < 1.0 < ih (286)

b
%= % - if TU < 1.0 < - -, (28c)

-n,, = T|J + i if 1.0 < -Hi (28d)


where
S = 0.05 |-n2 " "nil (28c)
In most cases, no line search is required to give real roots to the basic
constraint equation (12a). So the practice here is first to solve (12a) as
described previously with -r\t = 1.0. If complex roots are obtained, relations
(27) and (28) for the factor in, are then formulated and solved. The chosen
value of T|, is fed back to (12a), giving real roots, and the analysis can then
proceed as before. In the rare situation where complex roots are also ob-
tained for relation (27a), the previous strategy represented by (13)-(17) is
adopted.
The introduction of this pseudoline-search technique was found to give
great improvement in the convergence characteristics of the nonlinear so-
lution procedure.

Flow Route Summary


The solution procedure of the modified arc-length method is summarized
in the flow diagram of Fig. 3.
177

J. Struct. Eng. 1992.118:169-185.


Start: Increment no. (IINCSI = 1
Iteration no. (UTERI = 1
Downloaded from ascelibrary.org by QUEEN'S UNIVERSITY on 04/22/13. Copyright ASCE. For personal use only; all rights reserved.

Compute Ao & Al
(Eqs 22 & 24

Compute Ap
(Eg. 231

Store A\ pr = AX
and Ap'"'- Ap

Store AX pr = AX
and Ap pr = Ap

IINCS=IINCStl
IITER = 1

Compute new A!
(Eq. 25)
Optional
Recompute K and 6q
_,_ Reset p 0 & X0
Compute AolEq.111
Compute X (Eq. I/, I
and AX = X - X 0

FIG. 3. Flowchart for Arc-Length Solution Procedure

APPLICATIONS

A finite element program to analyze reinforced concrete plane frame


structures was developed in Fortran 77 on the Cambridge University main-
frame computer IBM3081, and this modified arc-length method was incor-
porated. In the program, two-noded beam elements are used, with three
degrees of freedom at each node, representing axial and transverse dis-
placements and rotation. The axial displacement is assumed to vary linearly
along the element while a cubic relationship is specified for the transverse
component (Lam 1987). Deformations due to shear forces are assumed to
be negligible, member deflections are small in comparison with member
sizes, and no out-of-plane load or displacement can occur.
In a parallel study (Lam 1987) on the modeling of tension stiffening in
178

J. Struct. Eng. 1992.118:169-185.


reinforced concrete beams under service loads, concrete in tension is as-
sumed to carry no stress, but the steel reinforcement in tension is assumed
to carry additional stress, to represent the contribution of the surrounding
concrete. The enhanced stress-strain curve for the tension steel is derived
by comparison with experiments, and is accompanied by a curve giving the
position of the resultant tensile force. So the entire tension zone can be
Downloaded from ascelibrary.org by QUEEN'S UNIVERSITY on 04/22/13. Copyright ASCE. For personal use only; all rights reserved.

represented by one layer in the computer. Similarly, the compressive stress-


strain properties of concrete are preintegrated to produce curves permitting
the resultant compressive force and its position in the beam to be determined
from the neutral axis position and the extreme fiber strain, without the need
for repeated integration across the beam depth during iterations.
To assess the effectiveness of the proposed arc-length algorithms, two
reinforced concrete bending problems are described, based for simplicity
on small deflections and only a few nodes. The material stress-strain relations
are assumed to be reversible [other examples with different loading and
unloading curves have been solved by Lam (1987)]. The method has been
successful in large-deflection slab analysis with snap-through (Morley and
Lam 1986), but further study of its effectiveness in cases with many degrees
of freedom is desirable.

Beam with Fixed and Sliding Supports


The first example (Fig. 4) involves a doubly reinforced concrete beam 2-
m long, with a uniform rectangular section of dimensions 63 mm x 133

10 kN

jm
>//}/
1m 1m

(a) General arrangement

63-0

A s = 56-5mm2 o o
66'5mm
A1-5
reference
axis
41-5
A s = 56-5 mm2 O 0 66-5

(b) Beam.cross section

Gauss points _ 1 2 3 4 5 6 t
Nodes 1 2 3 4
I 350 I 350 I 300 I

(c) Finite element representation

FIG. 4. Analysis of Beam with Fixed and Sliding Supports

179

J. Struct. Eng. 1992.118:169-185.


mm. The beam is fully fixed at one end but horizontal movement (only) is
permitted at the other. A vertical point load is applied at the midspan of
the beam. By symmetry, only half of the beam need be analyzed, repre-
sented by three finite elements; the integrating Gauss points (GP) for the
elements are numbered as in Fig. 4(c).
The concrete strength is 32.2 N/mm2 and the stress-strain relation in
Downloaded from ascelibrary.org by QUEEN'S UNIVERSITY on 04/22/13. Copyright ASCE. For personal use only; all rights reserved.

compression follows the experimental results of Kupfer et al. (1969). High-


yield torbars form the reinforcement, with an initial Young's modulus of
208.6kN/mm2 and a 0.2% proof stress of 479.5 N/mm2, the tensile stress-
strain relation being as measured in uniaxial tension tests. Steel behavior
in compression is assumed to follow the same curve. For steel in tension
within concrete, the chosen enhanced stress-strain (uas es) relation is
shown in Fig. 5(a), in which the sharp drop represents concrete cracking.
The corresponding level arm ratio-strain (zjh, - e,.) relation is given in Fig.
5(b), where z, is the distance from the neutral axis to the resultant tensile
force, and h, is the distance from the neutral axis to the tensile steel layer.
For strains exceeding 0.008, the curve for bare steel bars is followed.
The specified convergence criterion e for the arc-length procedure [(20)]
was 1/1,000.

0-000 0-002 0004 0-006 0008


Steel strain e s
(a) Apparent steel stress-strain relation

1-2

r
I 0-8

| 06

jo,
10-2

0-0 _L 1_
0000 0-002 0-004 0-006 0-008
Steel strain e s
(b) Apparent steel lever arm ratio-strain relation

FIG. 5. Assumed Constitutive Relations for Tension Stiffening

180

J. Struct. Eng. 1992.118:169-185.


The computed applied load is plotted against the central beam deflection
in Fig. 6. Initially, the response is linearly elastic from the origin. Nonlin-
earity sets in at a load of just under 2kN when cracking begins at GPs 1
and 6. There is a load peak at A when cracking occurs at GPs 1, 2, 5, and
6; the structure then restiffens and the load increases with further deflection.
At B, the slight difficulty in convergence, reflected in the automatically
Downloaded from ascelibrary.org by QUEEN'S UNIVERSITY on 04/22/13. Copyright ASCE. For personal use only; all rights reserved.

decreased arc-length increments, is caused by cracking at GP 3. When region


C is reached, the compressive extreme fiber strain at GP 6, followed by
that at GP 1, exceeds the ultimate strain of the concrete stress-strain curve;
the stress at the extreme fiber was then taken to drop instantaneously to
zero, hence the slight load drop following C. At D, localization of strain
takes place at GPs 1 and 6 due to the concrete falling branch in compression,
which then leads to eventual failure with concrete crushing at these Gauss
points.
For comparison, the problem was reanalyzed assuming bare steel behavior
for the tension reinforcement, and the result is also included in Fig. 6. As
expected, the two curves converge as the ultimate load is approached, with
failure occurring in the same manner.

Cantilever with Near-Uniform Bending Moment


Next a cantilever is considered, loaded at the free end by a couple plus
a small transverse force. The magnitude of the transverse force has been
chosen such that the resulting bending moment diagram is fairly uniform,
with the value at the free end 95% of that at the support, thereby creating
a situation in which the concrete cracks successively along the beam as the
load is increased. This gives a similar situation to that considered by Crisfield
(1982b).
The cantilever is 1.2m long, with cross-sectional details and material stress-
strain relations identical to those of the previous example. Three finite
elements were used (Fig. 7) and the convergence tolerance [(20)] was set
at 0.06.
The applied couple is plotted against the rotation at the free end in Fig.

10,-

a solution wilh tension stiffening


o solution assuming bore steel behaviour,
no tension stiffening

10 15 20
Central deflection {mmf

FIG. 6. Load-Deflection Response of Beam

181

J. Struct. Eng. 1992.118:169-185.


100 N

2-28
kNm
1200
Downloaded from ascelibrary.org by QUEEN'S UNIVERSITY on 04/22/13. Copyright ASCE. For personal use only; all rights reserved.

(a) General arrangement

Gauss points 1 2 3 4 5
x * x -)( 0 X-
Nodes 1 2 3
L 400 __L 40^_ . I . 400

(b) Finite element representation

FIG. 7. Analysis of Cantilever

J 1-51

0010
End rotation (radians)

FIG. 8. Cantilever: Load Drops due to Successive Cracking

8. Initially, the behavior is approximately linear until a load peak occurs at


A due to concrete cracking at GPs 1 and 2, then a renewed rise followed
by a second load peak at B with cracking at GPs 3 and 4, a further renewed
rise is then followed by a near-horizontal jump at C, which presumably
bridges the third load drop with cracking at GPs 5 and 6.
When the Gauss points are all cracked, the cantilever restiffens and the
load picks up again. The full results are shown in Fig. 9. A snap-back
response is seen at D when the compressive extreme fiber strain at GPs 1
and 2 exceeds the ultimate strain of the concrete stress-strain relation. Strain
localization and concrete crushing then set in at GPs 1 and 2characterized
by strain increases at these two Gauss points and strain reversals at the
others.
The snap-back response at D is a numerical phenomenon, obtained as a
result of the rather short and steep falling branch of the specified concrete
stress-strain relation in compression and the assumed instantaneous drop to
182

J. Struct. Eng. 1992.118:169-185.


Downloaded from ascelibrary.org by QUEEN'S UNIVERSITY on 04/22/13. Copyright ASCE. For personal use only; all rights reserved.

solution with tension stiffening


solution assuming bare steel behaviour,
no tension stiffening

0-10 0-15
End rotation (radians)

FIG. 9. Moment-Rotation Response of Cantilever

zero stress at the ultimate strain. This snap-back is not predicted if a more
prolonged falling branch is specified for the concrete in compression.
The result of using bare steel behavior for tension steel in the analysis is
also shown in Fig. 9. As expected, the two curves converge at high loads.

Remarks on Examples
Nonlinear solutions can be affected by the assumptions about shape func-
tions for the finite elements. In the second example, Gauss points in each
element were found to crack simultaneously instead of successively. This is
because a linear distribution of axial displacement, i.e., uniform axial strain,
was assumed for each element. With a fairly uniform bending moment along
the cantilever, cracking at one Gauss point often implies cracking at the
other, within the same solution increment. This can be rectified by using a
shorter arc length, or by introducing a midlength third node with extra
freedom in the axial direction, thereby allowing different axial strains at the
two Gauss points.
The examples clearly demonstrate the occurrence of strain localization as
a result of using strain-softening constitutive relations. In Figs. 6 and 9, the
localizations were caused by the assumed short and steep falling branch for
concrete in compression, particularly the sudden drop to zero of the concrete
stress at the specified ultimate compressive strain. In Fig. 8, strain locali-
zation was also obtained on cracking, in this case related to the initial dip
on the apparent steel stress-strain curve [Fig. 5(a)] used to represent the
properties of the tension zone.
These rather simple examples were chosen to show the method working
in cases where other methods would faile.g., load control would probably
not cope with the peaks in Fig. 8, and deflection control would miss a snap-
back such as that in Fig. 9. It is quite possible that unmodified arc-length
methods would cope in these simple casesbut our experience with
more complicated structures (Lam 1987) is that complex roots to the con-
straint equation do regularly occur, so that a method of avoiding them is
advantageous.
183

J. Struct. Eng. 1992.118:169-185.


CONCLUSIONS

This modified constant arc-length method can be implemented fairly sim-


ply and applied to a wide range of structural problems, forming a useful
tool for the analyst in nonlinear situations. The pseudoline-search procedure
further improves the effectiveness of the arc-length method, by overcoming
Downloaded from ascelibrary.org by QUEEN'S UNIVERSITY on 04/22/13. Copyright ASCE. For personal use only; all rights reserved.

the complex root problem, and can be formulated easily without involving
the concept of potential energy.
The examples demonstrate the feasibility and versatility of the arc-length
solution procedure in finite element analysis of reinforced concrete struc-
tures, especially its ability to proceed automatically past load peaks of the
snap-through or snap-back variety.
Although the two examples presented are on reinforced concrete struc-
tures, the proposed modified arc-length method is fairly general, and could
be applied in nonlinear structural calculations for other materials.

APPENDIX I. REFERENCES

Bathe, K. J., and Dvorkin, E. N. (1983). "On the automatic solution of non-linear
finite element equations." Comput. Struct., 17(5-6), 871-879.
Batoz, J. L., and Dhatt, G. (1979). Incremental displacement algorithms for non-
linear problems. Int. J. Numer. Meth. Engrg., 14(8), 1262-1267.
Bergan, P. G. (1980). "Automated incremental-iterative solution schemes." Nu-
merical methods for non-linear problems, C. Taylor, E. Hinton, and D. R. J.
Owen, (eds.), Pineridge Press, Swansea, U.K., 291-305.
Bergan, P. G., and Holand, I. (1979). "Nonlinear finite element analysis of concrete
structures." Comput. Meth. in Appl. Mech. and Engrg., 17/18, 443-467.
Bergan, P. G., Horrigmoe, G., Krakeland, B., and Soreide, T. H. (1978). "Solution
techniques for non-linear finite element problems." Int. J. Numer. Meth. Engrg.,
12(11), 1677-1696.
Cope, R. J., and Rao, P. V. (1981). "Non-linear finite element strategies for bridge
slabs." Advanced mechanics of reinforced concrete, IABSE, Delft, 275-290.
Crisfield, M. A. (1981). "A fast incremental/iterative solution procedure that handles
'snap-through'," Comput. Struct., 13(1-3), 55-62.
Crisfield, M. A. (1982a). "Accelerated solution techniques and concrete cracking."
Comput. Meth. Appl. Mech. and Engrg., 33, 585-607.
Crisfield, M. A. (19826). "Local instabilities in the non-linear analysis of reinforced
concrete beams and slabs. Proc. Inst. Civ. Engrs., 73(2), 135-145.
Crisfield, M. A. (1983). "An arc-length method including line searches and accel-
erations." Int. I. Numer. Meth. Engrg., 19(9), 1269-1289.
Crisfield, M. A. (1984). "Overcoming limit points with material softening and strain
localisation." Numerical methods for non-linear problems, C. Taylor, E. Hinton,
and D. R. J. Owen, eds., Pineridge Press, Swansea, U.K., 244-277.
Kupfer, H., Hilsdorf, H. K., and Rusch, H. (1969). "Behaviour of concrete under
biaxial stresses." /. Am. Concr. Inst., 66(8), 656-666.
Lam, W. F. (1987). "Constitutive relations for finite element analysis of tension
stiffening in reinforced concrete," thesis presented to the Univ. of Cambridge at
Cambridge, England, in partial fulfillment of the requirements for the degree of
Doctor of Philosophy.
Morley, C. T., and Lam, W. F. (1986). "Passing load peaks in finite element cal-
culations for reinforced concrete structures." Proc. Int. Symp. on Fundamental
Theory of Reinforced and Prestressed Concrete (cosponsored by ASCE), Nanjing,
3, 1120-1126.
Phillips, D. V., and Zienkiewicz, O. C. (1976). "Finite element non-linear analysis
of concrete structures." Proc. Inst. Civ. Engrs., 61(2), 59-88.
Ramm, E. (1981). "Strategies for tracing the nonlinear response near limit points."
Non-linear finite element analysis in structural mechanics, W. Wunderlich, E. Stein,
and K. J. Bathe, eds., Springer-Verlag, New York, N.Y. 63-89.

184

J. Struct. Eng. 1992.118:169-185.


Riks, E. (1972). "The application of Newton's method to the problem of elastic
stability." /. Appl. Mech., 39(4), 1060-1066.
Riks, E . (1979). " A n incremental approach to the solution of snapping and buckling
problems." Int. J. Solids Struct, 15(7), 529-551.
Wempner, G. A . (1971). "Discrete approximations related to nonlinear theories of
solids." Int. I. Solids Struct. 7(11), 1581-1599.
Wright, E. W., and Gaylord, E . H . (1968). "Analysis of unbraced multistory steel
Downloaded from ascelibrary.org by QUEEN'S UNIVERSITY on 04/22/13. Copyright ASCE. For personal use only; all rights reserved.

rigid frames." /. Struct. Div., ASCE, 94(5), 1143-1163.

APPENDIX II. NOTATION

The following symbols are used in this paper:

A0 = scalar parameter that governs the relative contribution of displace-


ment and load vector.
As = tension steel area;
e, = out-of-balance loads vector;
E, = vector of external applied loads;
F, = vector of nodal loads in equilibrium with internal forces;
gi = load factor for component of out-of-balance loads parallel to q;
h, = component of out-of-balance loads orthogonal to q;
h, = distance from neutral axis to steel layer;
K = overall stiffness matrix;
A/ = generalized arc length;
P, = force on tension zone;
p = vector of nodal displacements;
p0 = vector of nodal displacements at last converged solution;
Ap, = vector of incremental nodal displacements;
ApPr = previous displacement changes between last two converged solu-
tions;
q = fixed external nodal loads vector;
x, = scalar factor governed by the arc-length constraint equation;
z, = lever arm of P, from neutral axis;
AX,- = incremental load factor;
AXpr = previous load factor changes between the last two converged so-
lutions;
8e, = vector of nodal displacements resulting from eliminating e,-;
8W = vector of nodal displacements resulting from eliminating h,
(= K-%,.);
8, = estimated vector of nodal displacement corrections;
8g = vector of parallel displacements ( = K _ 1 q ) ;
8X,- = estimated load factor correction;
8;, = steel strain;
T|;- = line search factor;
X = load factor;
X0 = load factor at last converged solution; and
cyas = apparent steel stress.

185

J. Struct. Eng. 1992.118:169-185.

S-ar putea să vă placă și