Compnrrrr,I .SwwrwerVol. 26. No. S. pp 767-719. 1987 004s7949/87 13.00 + 0.

00
Printedin Gaul Erium. Pcrgamon JoumaL Ltd.

NUMERICAL TECHNIQUES FOR PLASTICITY

COMPUTATIONS IN FINITE ELEMENT ANALYSIS
ROBERTH. DODDS, JR
Department of Civil Engineering, University of Kansas, Lawrence, KS 66045, U.S.A.

(Received 28 March 1986)

Abstract-Common numerical techniques for plasticity computations in finite element analysis are
examined. The plasticity theory considered is the simple rate-independent van Mises criterion for small
strains. Work hardening is represented by a general isotropic model or by a linear, isotropic-kinematic
mixed model. Algorithms to integrate the rate equations, strategies for stress updating over a time (load)
step in implicit codes, and tangent operators consistent with the integration algorithm are discussed. The
elastic predictor-radial return algorithm and a consistent tangent operator satisfy the requirements for
a stable, accurate and efficient numerical procedure. An extension of this model for plane stress with mixed
hardening is described. Two numerical examples are given to demonstrate the accuracy and efficiency for
plane stress analyses.

INTRODUCI’ION dating stresses over a time (load) step in implicit

codes are described. The path-independent strategy
The finite element method is used extensively to eliminates the spurious elastic unloading that occurs
analyze the elastic-plastic response of structures. The under displacement-controlled loading. However, its
accuracy and computational efficiency of an elastic- use requires proper definition of a consistenr tangent
plastic solution depend to a large extent on the operator to maintain convergence of Newton-type
robustness of the numerical techniques to implement solutions. Recent developments by Simo and
the constitutive equations of plasticity. A solid three- Taylor[l] for consistent tangent operators are re-
dimensional model, for example, may have several viewed and a matrix formulation of their tensor result
thousand sampling points at which the plasticity is provided.
computations are performed for every load step (and The paper concludes with the solution of two
corrective iteration). Static and dynamic analyses of example problems that illustrate the accuracy and
such models with implicit codes must use large convergence rates that may be expected for plasticity
time (load) steps to maintain efficiency at the global problems. Both examples are for the more complex
solution level. The large sire steps place severe plane stress condition. The consistent tangent oper-
demands on the plasticity routines to maintain ator is shown to provide much superior convergence
accuracy, numerical stability and efficiency. rates for these plane stress problems to the classical
This paper examines numerical techniques com- rate form of the tangent operator.
monly implemented to perform plasticity compu-
tations for the von Mises yield function with a mixed PLA.STMTY RATE EQUATIONS
isotropic-kinematic hardening rule. This constitutive
model adequately predicts the behavior for many The plasticity theory considered here assumes in-
ductile metals. Attention is limited to the small strain itial isotropy of the material and neglects time-
case. With the definition of suitable measures for dependent and thermal effects. A van Mises yield
strain and stress rates, these same numerical tech- surface and associated flow rule are adopted. For
niques form the basis of mostfinire strain constitutive isotropic hardening, the plastic modulus may vary,
models. The Baushinger effect is incorporated using a mixed
The plasticity rate equations are first summarized. isotropic-kinematic hardening model with constant
Three cases of the generalized trapezoidal rule for plastic modulus[2].
integration of the rate equations are then discussed. The von Mises yield surface is given by
The specific algorithms are: (1) tangent stiffness-
radial corrector; (2) mean normal; and (3) elastic 4 (SE,) = S-S - R2(cJ (1)
predictor-radial return. The elastic predictor-radial
return algorithm is shown to have important advan- where S is the deviatoric part of the shifted stress
tages over the first two algorithms, especially for vector 4, R is the radius of the yield surface in
implicit codes. An extension of the elastic predictor- deviatoric stress space, and et, is the effective plastic
radial return algorithm is presented for plane stress strain. To simplify inner products, strain and stress
with mixed isotropic-kinematic hardening. The path- vectors act in a nine-space and are indicated by a
dependent and path-independent strategies for up- boldface character. Conventional 6 x 1 stress and
768 ROBERT H DODDS, JR

strain vectors are enclosed in { }. R is related to the and using eqns (a), (7) and (1 I), the deviatoric stress
effective tensile stress Y by rate is given by

R =@ Y&J. (2) $ = 2G (e - AS). (13)

The shifted stress { is given by The rate of shifted stress is given by addition of the
volumetric stress rate to eqn (13)
{=a-a (3)

where u is the total stress and a locates the center of ~=S+p[l,l,l,o ,..., O]T (14)
the yield surface, i.e. the back-stress.
where p is the volumetric stress rate, J) = (K/3)
The total strain rate i is decomposed into devi-
(i, + I, + L,), and K is the bulk modulus. Using 6 x 1
atoric and volumetric components. The volumetric
component vectors for the stress and strain rates, eqn
component is treated elastically. The deviatoric strain
(14) is usually written as
rate e is additively decomposed into elastic and
plastic parts

e = e, + er. (4)
where closed forms of [Dr] for three-dimensional and
The elastic part of the deviatoric strain rate is related plane stress conditions were first given by Yamada et
to the deviatoric stress rate by the elastic shear af.[3].
modulus

$ = 2Ge,. (5)
INTEGRATION PROCEDURES
The associated flow rule is used to give
In conventional finite elements analysis, the consti-
e,=In tutive quantities are known at time r,, and the updated
(6)
strains are known at time I,+ 1. For static analyses
where 1 is a proportionality factor and n is a vector time is interpreted as a loading parameter. Over the
normal to the yield surface at the contact point. increment At a load step is applied to the structure.
Components of n are given by The constitutive routine integrates the plasticity rate
equations over the time increment to determine the
updated stresses and plastic state variables.
The generalized trapezoidal rule as described by
Ortiz and Popov[4] encompasses the three most
The plastic hardening rules are popular methods for integration of the plasticity rate
equations. These methods are known as: (1) tangent
p = /IH’($& (8) stiffness-radial corrector; (2) mean normal or secant
stiffness; and (3) elastic predictor-radial return. In
oi = $(l - /?)H’i, (9) this section, each of these methods is discussed from
the viewpoints of numerical stability, accuracy and
where the parameter j? defines the portion of isotropic ease of implementation. Complications that arise for
hardening, and H’(c,) defines the instantaneous plas- the plane stress condition, strain hardening, and
tic modulus. When j? < 1, H’ is assumed constant. combined elastic and elastoplastic response within a
The rate of effective plastic strain consistent with step are also addressed.
the von Mises yield surface and the Prandtl-Reuss Each integration method operates under the re-
relations is given by striction of a constant strain rate vector over the time
step. Higher derivatives of the strain vector are
L, = J(2/3) e;e,. (10) usually not available from the structure level sol-
ution. For finite sixe load steps, this leads to predicted
Using the consistency requirement that 4 = 0 dur- stress and strain paths that are chordal approxi-
ing neutral or plastic loading, the proportionality mations to the generally curved, exact paths. The use
factor is found to be of a linear strain path over each step introduces
discretixation errors. These errors are in addition to
the errors from integration of the plasticity rate
(11) equations over a specified strain increment. To con-
2R2[1 + (H’/3G)l’
trol the load path discretixation errors, Tracey and
Rewriting qn (5) in the form: Freese[S] developed an adaptive scheme that exam-
ines local curvature of the yield surface and direction
$ = 2G [e - e,] (12) of the strain rate vector to select the load step sixes.
 Plasticity computations in finite element analysis 769

Generalized trapezoidal rule

Ortiz and Popov[4] generalized the trapezoidal rule
for integration of plasticity rate equations and in-
vestigated the numerical stability and accuracy. For
the von Mises yield surface, associated flow rule, and
mixed isotropic-kinematic hardening, the generalized
trapexoidal rule has the form

c.+,=L+&-2GAep (16)

4n+I(&+17Y,+,)=O (18)

where A{, is the elastic stress increment correspond-

Fig. 2. Tangent stiffness-radial correction algorithm.
ing to the strain increment At = r, + , - c,,.
The incremental parameter d is determined by
requiring that the updated stresses satisfy the consist- tion and partially along the final plastic flow direc-
tion. Similarly, the updated center of the yield surface
ency condition, dr,+ 1= 0. The algo~thmic parameter
is located by translations along the initial and fhml
y may vary from 0 to 1. The updated state variables
and stress vector are given by plastic flow directions.
The three common procedures used to integrate the
a#+,=a,+f(l -p)H’Ae, plasticity rate equations correspond to the following
0%
cases of the genera&red trapezoidal rule:
Ytl+l = Y, + j?i?(~,)Ac, (20)
y = 0, tangent stiffness-radial corrector;
un+]=tn+I +9+1 (21)
Y= 4, mean normal;
where f?(c,,) denotes an incremental-secant plastic
modulus for a material with variable plastic modulus. y = 1, elastic predictor-radial return.
By combining the first two terms of eqn (16) and
subs~tuting for the plastic strain increment from eqn The value of y determines the stability and accu-
(17), an expression of the form racy of the integration scheme. For y > 0, the method
is implicit and may require iterative solution. For the
von Mises yield surface (and all yield surfaces of
t n+~=t~-~[(1-~)4+~nn+Il cm constant cu~ature), the integration procedure is un-
is obtained, where the quantity CTrepresents a trial conditionally stable for y 2 Ji;for y < 4 the procedure
elastic stress vector. Ortix and Popov noted that this is conditionally stable.
form provides an appealing geometric interpretation
of the integration procedure. As shown in Fig. 1, the
trial stress vector is returned to the updated yield This procedure is used extensively in implicit codes
surface partialiy along the initial plastic flow direc- for static and dynamic analysis where the com-
putation of an elastic-plastic &] is required for
generation of a structure tangent stiffness. The plastic
flow direction is fixed parallel to the yield surface
normal at the contact point 4.. The integration
procedure is usually written in matrix form

Rn+*I = Kl+ ml WI (23)

where R] is given by eqn (15). This corresponds to

a one-step, forward Euler method which is illustrated
in Fig. 2. A significant appeal of this method is the
simplicity of condensing the three-dimensional form
of &] to obtain the corresponding matrix for plane
stress. Computational details for mixed isotropic-
kinematic hardening are given by Hunsaker et aZ.[q.
Nagtegaal and de Jong[7J first proved the con-
Fig. 1. Generalized trapezoidal rule in deviatoric principal ditional stability of this method. The stability limit for
stre4s space. proportional straining of a non-hardening material is
770 ROBERT
H. DODDS. JR

twice the elastic strain vector at yield. The failure of lower bound for the stability limit of twice the elastic
the method for a simple plane stress problem led yield strain imposes severe restrictions on the sub-
Nagtegaal and de Jong to an investigation of the increment size for materials with a large ratio of
instability. The problem that failed is a four-node, Young’s modulus to yield stress. The expense of
isoparametric element oriented at 45” to the X-axis reforming [Dr] and updating the state variables for
and stressed in uniaxial tension. In global coordinates many sub-increments may lead to excessive com-
all three stress components (uX,uY,uXY) are equal. After putation times.
yielding, stresses at the integration points oscillate Two additional complications exist with this
erratically unless very small load steps are used. In method. First, for steps in which both elastic and
contrast the solution is insensitive to load step size for plastic behavior occur, the contact stress must be
an element aligned with the coordinate axes (in which found to permit division of the strain increment
case a, is the only non-zero stress). The appearance vector into purely elastic and elasto-plastic parts.
of such instabilities in complex models has not been Second, the stress state at the end of each sub-
explicitly reported in the literature. Interactions with increment generally lies outside the updated yield
adjacent elements may dampen the oscillations dur- surface. To prevent artificial hardening, the stress
ing equilibrium iterations. More likely, solutions state is scaled radially to the yield surface. Nyssen[ 111,
become slowly convergent with the cause attributed and Mondkar and Powell[l3] note that additional
by the analyst to the frequency of stiffness updates. plastic flow occurs during this correction and that it
Only close study of the stress fluctuations during should be incorporated in the plastic strain vector.
iterations would reveal this type of instability. Moreover, when the plastic modulus varies, an inwn-
The accuracy of this method was studied by Krieg sistency in the updated yield surface radius develops
and Krieg@] for a nonhardening material and by from linearization of the stress-strain curve during
Schreyer, et uf.[9] for isotropic hardening with a each sub-increment. Additional scaling of the
constant plastic modulus. The method is exact if the updated stress vector may be required to prevent
strain increment vector is parallel to the deviatoric accumulation of this error.
component of the contact stress (radial loading). For
other orientations of the strain increment vector, Mean normal method
errors are introduced in both the orientation of the This method was presented first by Rice and
updated stress vector and the radius of the yield Tracey[l4] for the plane strain condition with non-
surface. hardening materials. The method is particularly
To reduce integration errors, the strain increment simple and predicts a stress vector that lies on the
vector is generally partitioned into m sub-increments yield surface for non-hardening materials under
and the elastic-plastic [Dr] matrix reformed at the plane strain, axisymmetric, and threedimensional
beginning of each sub-increment. The updated stress conditions. This is an advantage over the tangent
vector is then given by stiffness method which requires corrective scaling of
the predicted stress vector.
The mean normal method is usually implemented
with sub-incrementation. Equation (23) is again used
but with a rnem [Dr] computed at the contact point
Several criteria have been proposed to select the corresponding to an average vector normal to the
number of sub-increments. Bushnell[lO] selects m by yield surface
limiting the magnitude of an effective uniaxial strain
in each sub-increment. Nyssen[ll] limits the mag- sn + 112 = 0.5(S”+ s-r). (25)
nitude of the deviations from the yield surface in each
sub-increment to a fraction of the deviation for a
single step. Schreyer er a!.[91 select m using the
angular difference between a stress vector normal to
the yield surface at the current contact point and a
normal vector at the contact point computed with a
single step estimate. This angle measures the change
in plastic flow direction due to curvature of the yield
surface. The angle change in each sub-increment is
limited by the value of m. Krieg and Key[lZ] recom-
mend selection of the number of sub-increments
using the direction of the strain increment vector and
the known error magnitudes for the direction taken
from the error studies.
While sub-incrementation clearly improves the
accuracy, an unacceptably large number of sub-
increments may be necessary to ensure stability. The Fig. 3. Mean normal algorithm.
 Plasticity computations in finite element analysis

The trial normal vector, Sr, is the deviatoric com-

ponent of the trial elastic stress vector Cr. Figure 3
graphically illustrates the mean normal method.
The method is unconditionally stable (y = f) and is
exact in the non-hardening case for radial strain
increments. Krieg and Krieg[l] conducted an error
analysis for the procedure. The error in orientation of
the updated stress vector increases monotonically for
increasing angles between the strain increment vector
and the deviatoric component of the contact stress
vector. No error analyses for the hardening case have
been reported.
Sub-incrementation of the strain increment vector
may be employed to reduce the error. Using their
error analyses, Krieg and Krieg were able to develop
a simple error estimator that permits rational selec- Fig. 4. Elastic predictor-radial return algorithm.
tion of the number of sub-increments.
For the hardening case, the updated radius of
previous methods. The orientation error of the
the yield surface may be in error (see Fig. 3).
updated stress vector increases monotonically for
Nagtegaal[ 151reported a significant underestimation
increasingly non-radial strain increments. Without
of the yield surface radius for large, non-radial strain
sub-incrementation, this method exhibits significantly
increments. The error stems from the use of a mean
less error than both the tangent stiffness and mean
normal vector scaled to the yield surface radius R.
normal methods[8]. Sub-incrementation further im-
and to variations in the plastic modulus over the step.
proves the accuracy[9] and is generally required in
Additional schemes (including sub-incrementation)
implicit codes.
are required to control this error.
The method has the same disadvantage for the
The mean normal method has a serious drawback
plane stress case as the mean normal method. The
for the plane stress case. Computation of the trial
strain increment bz is required for computation of
stress vector for use in eqn (25) requires knowledge
the trial elastic stress vector. In the hydrodynamic
of AC,. However, AC,is not known from the plane
codes, AC,is computed by forcing Au, = 0 using the
stress solution. Tracey and Freese[S] described an
corresponding row ot m&
iterative scheme to compute AC,such that Auz= 0 for
contact point on the yield surface. This is an acccpt-
the step. Updated estimates of AC, are found by
able approximation for the small strain increment
forcing Aa, = 0 using the corresponding row of [DT]
vectors that occur in explicit codes (where small time
computed for the current S,, ,12.The improved AC,is
steps are needed at the global level). For implicit
used to compute a new trial elastic stress vector and
codes, which must process much larger strain incre-
a new mean normal vector, etc.
ments, this approximation is not acceptable. Schreyer
Elartic predictor-radial return method et aL[9] presented an iterative scheme for isotropic
hardening. The plane stress conditions are imposed
This method was 6rst described by Wilkins(161and
on the trial stress vector and the plastic strain in-
Mendleson[l7]. It is used extensively in explicit,
crement is iteratively adjusted until the consistency
hydrodynamic codes including HONDO-11[181 and
condition is satisfied by the updated stress vector.
DYNA2D/DYNA3D[l9]. Key and Krieg[ 121 in-
Such an approach for kinematic hardening is far
cluded mixed isotropic-kinematic hardening in a
more complex and has not been reported. The algor-
particularly simple manner.
ithm described in the next section permits rapid
Figure 4 illustrates the method. A trial elfistic stress
solution of the plane stress case with mixed isotropic-
vector & and its deviatoric component Sr are estab-
kinematic hardening.
lished. The plastic strain vector is assumed to lie
parallel to Sr(y = 1). The consistency condition, eqn
ELASTIC PRRDICTOR-RADIAL RJlTlJRN FOR
(18), together with eqns (10) and (16), are solved to PLANE STRESS
determine A,the effective plastic strain increment AcP,
and the incremental plastic strain vector A%. The Of the three integration algorithms, the elastic
updated radius of the yield surface is found using eqn predictor-radial return algorithm appears most suit-
(20). The trial elastic stress vector is returned to the able for application in implicit codes. Conceptually
new yield surface using eqn (16). The updated center simple, it provides the most accuracy for both single
of the yield surface is found using qn (19) and step and sub-increment&on schemes. The method is
the known incremental plastic strain vector. Com- unconditionally stable; it does not require corn-
putational details are given in the next section. putation of the contact stress or ud hoc scaling of the
This procedure is unconditionally stable and is updated stress vector. Mixed isotropic-kinematic
exact for radial strain increments as are the two hardening is easily included.
112 ROBERTH. DODDS, JR

Unfortunately, plane stress conditions significantly the material. If H’ is constant, eqn (33) simplifies to
complicate the procedure. To satisfy the plane stress
requirement, the total stress uz must be zero. Non- Yn+, = Y,+BH’Ac, (34)
zero values of the components {, and a, are generally
necessary for the plane stress solution to match the and eqn (32) becomes
Bauschinger effect predicted by a corresponding Yr - Y”
three-dimensional solution (with plane stress traction AC,=--. (35)
H’+3G
boundary conditions).
In the procedure described below, the strain in- Figure 5 outlines a simple iterative method to com-
crement AC, is iteratively adjusted with a Newton pute AC, if the plastic modulus varies. Equation (35)
algorithm until u, = 0. Details of the algorithm are is also valid for the non-hardening case, H’ = 0.
given following a summary of the computational With known values for I and A.+ the updated
steps for the three-dimensional elastic predictor- back-stress vector and total stress vectors are
radial return method which is used in the iterations.
H’Ac
~.+i=%+(l-8)~sr (36)
T
Three-dimensional algorithm
The yield function is evaluated for the plastic 3GAc
@.+I = a, + tT - 1 (37)
variables at time t. and the trial elastic stress vector YT %
Cr. If 4 (Sr, 5) > 0, then plastic deformation occurs;
Plane stress algorithm
otherwise the updated stress vector is given by
a, + At,. For plastic deformation, y = 1 in eqn (17) The following algorithm provides an updated
with n, + , = &,. The plastic strain increment is then stress vector that satisfies the consistency and plane
given by stress conditions.

AC, = 2%. (26) (a) Compute CTusing the plane stress, elastic consti-
tutive matrix and evaluate d. If 4 > 0, there is
The deviatoric form of eqn (22) becomes simply plastic deformation. Complete the remaining
steps. If 4 S 0, then CT+ a,, is the new Stress,
s “+,=(l -2GL)ST. (27) u,, ,. Update state variables to indicate elastic
unloading from a plastic state if necessary. Exit
Both sides of eqn (27) must satisfy the consistency this process.
condition at I,+, (b) Estimate AC;, where i = 1 for the initial estimate.
(c) Compute CT using the three-dimensional elastic
s II+1‘%,I = (I- 2Gi,)??&& (28) constitutive matrix and perform the elastic
predictor-radial return computations. If
An equivalent form of eqn (28) is ai I 101,then the updated stress vector has been
determined. Otherwise, consider a: to be a re-
Y,,+i=(l -2G1)Y,. (29) sidual stress ai defined by

To eliminate 1, AC, from eqn (26) is substituted into 6: z azucL- ai = - a:. (38)
eqn (10) to give
(d) Compute a first-order change in Afi such that
AC; e ;&L&r (30)
a;+da,=O (39)
or equivalently (using Y: = (&-ST)
wher: da, is defined by
), _3 (31)
-2y,*
da, I

Using this result for 1 in eqn (29) yields

The strain correction for the ith iteration is then
Y“+, = Y, - 3GAc, (32)
AC:= - u;lA (41)
which is solved simultaneously with
and the strain for use in the next iteration is
%
Y($) = ycl+ B H’(c,) dc, (33) AC;+’ = AC; + AC;. (42)
s0

to compute AC,. Y. denotes the tensile yield stress for Return to step (c).
 Plasticity computations in finite element analysis 173

(a) Rewrite equation (32) as

A% = $(Yr- Y,,,)

(b) Estimate AC, for local iteration using

YT- yn
AC; =
H,’ + 3C

(c) Find y’,+t from the supplied curve of stress vs. effective plastic strain
using cj defined by

(d) Compute new estimate of AC, for the next iteration

AC;+’ = & (Y, - r;,,,)

(e) Check for convergence of plastic strain increment

IAcf+’ - Acfl 2 1O-5 A$

if not converged return to (c).

Fig. 5. Local iteration process to compute &,.

The initial estimate for AC,is computed using one cases with zero for the initial estimate converged
of two schemes. If the current state is elastic, then the equally as well.
elastic change in Acz for plane stress conditions is To verify the correctness of this procedure, plane
used. If the current state is plastic, the (IS) row of stress problems for simple extension, shear, and
[DT] is formed with the current stress vector and biaxial loading were analyzed. The strains and
solved (A{, = 0) to yield A+ stresses were compared to the results for the same
The convergence test in step (c) compares a: to the problems analyzed using a threedimensional sol-
maximum of (a,, uv, crJ’_ A tolerance of 10-j has ution with plane stress boundary conditions. The
been used successfully. Alternatively, the magnitude plane stress and three-dimensional solutions were
of Ac: may be compared to AC:. identical in each case.
Simple numerical differentiation yields the required
derivative A in eqn (41). The three-dimensional elas- STRAINPATHS FOR =RRSS UPDATING
tic predictor-radial return computations are repeated
at Ac:(l - rf) and at Ac:(l + q), where 11is a small In implicit codes, the global solution is advanced
constant (n = IO-’ is satisfactory). The correspond- from t, to f,+,using an iterative Newton method. If
ing ur values are used to compute A i denotes the iteration number, then two strain in-
crement vectors for stress updating at material points
may be defined as
A
‘&&+1) - u*I I%!(I - I’)
=
(43)
2?lAc; *
A&=c;+,-c” (44)

Numerical tests have demonstrated that the iter- A&-6:+,-&;‘,. (45)

ative procedure converges very rapidly. After only
3-4 iterations, convergence is obtained for a toler- The strain increment detined by eqn (44) leads to
ance of 10-s. The convergence rate does not appear a path-inakpendent strategy. During each iteration,
to depend strongly on the initial estimate for Acz.Test the plasticity rate equations are integrated over the
774 ROBERT H BDDS, JR

equations are generally integrated over a larger strain

increment. Accuracy and stability of the integration
procedure are essential factors in the solution. In
addition, the tangent operators used for structure
stiffness formation require special consideration as
described in the next section.

CONSISTENTTANGENT OPERATORS

An elasto-plastic tangent, m], is needed to form

the structure stiffness matrix in implicit codes. To
Fig. 6. Strain paths for stress updating. maintain the quadratic rate of asymptotic con-
vergence for Newton’s method, the tangent operator
must be consistent with the numerical method em-
foral strain increment for the step using converged ployed to integrate the plasticity rate equations.
results at I, for the initial conditions. The converged Consistency implies that the stress increment pre-
strain path for the step is a linear segment as shown dicted by the tangent operator acting on the strain
in Fig. 6. increment matches the stress increment predicted by
The strain increment defined by eqn (45) leads the integration procedure to 6rst order.
to a path-dependenr strategy. The plasticity rate The confinuum tangent operator defined by the rate
equations are integrated over the corrective strain form of &] in eqn (15) is used in most implicit codes
increment for iteration i using non-converged results with the tangent stiffness-radial corrector integration
at i - 1 for the initial conditions. The converned algorithm. The rate [DT] is consistent with this inte-
strain path for the step consists of p linear segments, gration algorithm only for the path-dependent up
where p is the number of iterations required for dating strategy. For the path-independent updating
convergence (Fig. 6). strategy, the continuum operator is not the consistent
The path-dependent strategy has two drawbacks. tangent operator. At the beginning of the itb iter-
First, the direction of the strain increments defined in ation, [DT] is formed using the current contact stress
eqn (45) may differ markedly from the actual strain CL;‘, and thus predicts a stress increment to first
path. The corresponding plastic flow direction varies order for the strain increment that occurs during the
sharply from one iteration to the next whereas the itb iteration-not for the total strain increment over
actual plastic flow direction varies smoothly over the the step. Use of the continuum operator with the
step. The integrated effect of variations in plastic flow path-independent strategy severely degrades the
direction over the iterations may not adequately convergence rate unless load steps are very small.
represent an average plastic flow direction for the The continuum operator is also the consistent
step. operator for the elastic predictor-radial return algor-
Second, the sharp variations in the corrective strain ithm when used with the path-dependent strategy.
increments generally cause false indications of elastic However, it is clearly not the consistent tangent
unloading, i.e. the trial elastic stress vector lies inside operator for the path-independent strategy. Use of
the yield surface. Special numerical techniques are the continuum operator for the path-independent
then required to distinguish between real and spurious strategy leads to very poor convergence rates.
elastic unloading. Nyssen[l l] suggests an incremental Nagtegaal[lS] tirst derived a consistent tangent
reversibility concept in which non-linear elastic operator for the elastic predictor-radial return algor-
unloading is allowed until the strain energy density ithm. Simo and Taylor[l] formally derived the consist-
becomes equal to the value at f,. Only then is linear ent operator by expressing the integration algorithm
elastic unloading permitted. Numerical studies[20] in functional form with the strain increments as
indicate that spurious unloading is more pronounced independent variables and the new stresses as the
under displacement controlled loading. Typically, dependent variables. The Jacobian of the function
stresses are overestimated in the first iteration of a represents the consistent tangent operator. Simo and
step which corresponds to the application of the real Taylor expressed the tangent operator in tensor form.
displacement increment. During subsequent iter- The equivalent matrix form, m], is given below for
ations the stresses decrease as the structure stiffness the three-dimensional case. This form is applicable
decreases due to plastic deformation. When spurious for isotropic hardening with a variable plastic modu-
unloading is not controlled, the solution diverges lus and for mixed isotropic-kinematic hardening with
after a few iterations as material points alternately a constant plastic modulus.
yield then unload.
’ The linear approximation for the strain path over Pi1 = PI- 23 7 W4’ W
the step in the path-independent strategy eliminates
saurious unloadinn. However. the nlasticitv rate wherethe non-zero. lower triangles terms of IW are
 Plasticity computations in finite element analysis 775

given by

-=
_r T

I-I-
B

=-
El Y
0
4
0

Pomt L7 r

t-l-
Y,=2625--
ET’260
0 0 0
The 6 x 1 vector {n} contains normalized deviatoric A 262251 0 E = 26.300
B 26251 24 **cl3
components for the contact stress vector at the ith l/c
c
iteration of the step, arranged in the order
{x, Y, 2, V, YZ,xz 1. Thus, Fig. 7. Data for example problem 1.

{“)_{Sb+J
-- (47) Example 1
R:+,
A thin tube is first stressed in tension to the point
where RA,, is the yield surface radius at the ith of yielding and is then twisted under constant axial
iteration. The factors /? and 7 are given by stress. Remote from the ends of the tube, there is no
spatial variation and a single material point is
sufficient to model the problem. Figure 7 provides the
(48) loading path and material properties. For isotropic
hardening, Hill[21] gives the exact solution for the
axial and shear strains as
(49)

For large load steps, the trial elastic stress vector may
lie far outside the yield surface so that fl becomes
significantly < 1. The consistent tangent then differs
markedly from the continuum tangent. Note that the
continuum tangent is recovered for fl = 1, i.e. the
elastic trial stress lies on the current yield surface.
Plane strain and axisymmetric versions are ob- where Y, is the yield stress in tension, E is Young’s
tained by deleting rows 5,6 and columns $6. The modulus, H’ is the plastic modulus, and G is the
plane stress version is obtained by statically condens- elastic shear modulus for v = 0.3. Figure 8 shows the
ing the third (zz) row‘ of the plane strain version. exact strain path which has significant curvature for
The consistent tangent is applied with the path- low values of the shear strain. This problem repre-
independent updating strategy. For iteration one of sents a difficult test case for numerical methods which
each load step, the continuum tangent is used to approximate the strain path with linear segments.
compute the structure stiffness. At the beginning of Numerical results were obtained using a program
each subsequent iteration, the consistent tangent is that ‘exercises’ non-linear material models before
used in the formation of the structure stiffness. they are installed in our production finite element
system (POLO-FINITE[ZZ, 231). The program exe-
cutes interactively and allows users to test material
NUMERICAL EXAMPLES models for arbitrary stress or strain paths with min-

Numerical results for two example problems are

presented in this section. These examples demon-
strate the accuracy of the elastic predictor-radial
return algorithm and the convergence rates for the 6-

continuum and consistent tangent operators. Both

examples are for the more complex plane stress case.
In the first example, a single material point is subject
to a prescribed stress path. Solutions with and with-
out sub-incrementation are computed. The computed
strains are compared to the exact plasticity solution.
The second example considers the analysis of a
fracture mechanics specimen loaded under displace-
ment control. The path-independent updating strat-
Shear Strom, 7 *tOO
egy and the consistent tangent operator are essential
to obtain efficient solutions for this class of problem. Fig. 8. Exact strain path for example 1.
116 ROBERT
H. DODDS, JR

Table 1. Computed strains for example 1 (no sub-incrementation)

*V Exact AT, = 0.50 AT, = 1.0 AT, = 2.0 Ar, = 4.0
(ksi) ~~100 y*lOO c+lOO y*lOO c*lOO y*lOO c*lOO y*lOO c*lOO y*lOO
0 0.093 0.000 0.093 0.000 0.093 0.000 0.093 0.000 0.093 0.000
4 0.405 0.131 0.405 0.140 0.404 0.147 0.401 0.162 0.400 0.177
8 1.234 0.752 1.229 0.780 1.223 0.806 1.212 0.857 1.193 0.936
12 2.352 2.073 2.338 2.123 2.324 2.171 2.300 2.266 2.246 2.420
16 3.568 4.057 3.541 4.126 3.521 4.192 3.459 4.294 3.386 4.541
20 4.773 6.572 4.740 6.653 4.707 6.734 4.631 6.870 4.517 7.161
24 5.917 9.482 5.875 9.574 5.833 9.665 5.744 9.827 5.593 10.150

Table 2. Computed strains for example 1 with sub-incrcmentation (number of sub-increments listed in [ 1)
7XY Exact AT, = 0.50 AT, = 1.0 AT, = 2.0 At, = 4.0
(ksi) ~~100 y*lOO cslO0 ye100 c*lOO y*lOO co100 y*lOO c*lOO y.100
0 0.093 0.000 0.093 0.000[1] 0.093 O.OoO[l] 0.093 0.000 [l] 0.093 0.000 [l]
4 0.405 0.131 0.404 O.MO[l] 0.404 0.142[3] 0.404 0.148 [5] 0.404 0.149m
8 1.234 0.752 1.230 0.773 [3] 1.228 0.782[5] 1.221 0.816[11] 1.204 0.880 [20]
12 2.352 2.073 2.343 2.106[3] 2.337 2.127[7] 2.316 2.197 (141 2.268 2.337 (281
:: 4.773
3.568 4.057
6.572 4.755
3.554 4.100[3] 4.737
3.542 4.133[7]
6.662 4.682
3.504 4.233
6.787 [Is] 4.560
3.419 4.439
1.044 [3 l]
6.622[3] [7] [15] [3 l]
24 5.917 9.482 5.894 9.538[3] 5.871 9.583[6] 5.801 9.726 [14] 5.647 10.021 [29]

imum computational cost. The program has options generally exhibit the same relative error in the shear
to control the frequency of updates for the tangent strain for each load step.
operators and fully simulates the material model These results illustrate the magnitudes of error that
interface of the production finite element system. may be anticipated for realistic load step sixes when
Using the path-independent strategy, com- the strain path is highly nonlinear. Even though
putations were performed with and without sub- equilibrium is satisfied and the computed strains are
incrementation of the strain increment vector. Four converged to four significant figures, it is possible to
analyses were performed for each case. In each have significant errors in numerical values.
analysis the axial stress was increased to 1.OOlc Y, in Solutions computed with the consistent tangent
the first load step. The shear stress was then increased operator converged in many fewer iterations than did
to 24 ksi using a constant size increment in each solutions computed with the continuum tangent
analysis (0.5 ksi, 1.0 ksi, 2.0 ksi and 4.0 ksi). For the operator. Table 3 lists the number of iterations
analyses with sub-incrementation, the number of required for convergence at each load step for
sub-increments was selected to limit the length of the the analysis using 2 ksi shear stress increments with
trial elastic stress vector (deviatoric components) to sub-incrementation.
l/2 the yield surface radius in each sub-increment. Convergence is equally good for the kinematic
The iterative solution continued at each load step hardening model when used with the consistent tan-
until the Euclidean norm of the residual stress vector gent operator. Table 3 also lists the number of
was less than 0.001 ksi. This tolerance also ensured iterations required for convergence at each load step
convergence of the computed strains to four for the kinematic hardening model.
significant figures.
Tables 1 and 2 compare the computed and exact
strains. In both cases (with and without sub- Table 3. Iterations required for convergence in example 1
increments), the errors increase with shear step size. Consistent Continuum Consistent
TXY
Errors in the computed strains are due to the com- Orsi) (isotropic) (isotropic) (kinematic)
bined effects of discretization of the strain path into
2 3 5 3
linear segments and integration of the plasticity rate 4 5 20 5
equations. The largest errors, relative to the mag- 6 5 3
nitude of the shear strain, occur at lower values of 8 4 2 5
shear strain where the curvature of the strain path is 10 4 46 5
12 4 50 5
greatest. At a shear stress of 4 ksi, errors in the shear 4 52 5
strain range from 7 to 35%. At 24 ksi shear stress the :d 5 52 5
errors range from 0.6 to 6%. The comparison in 5 51 5
Table 2 demonstrates the benefits of using a sub- :: 5 50 5
22 5 48 5
increment procedure. The solutions using sub- 5 46 5
24
increments generally have 4&50% less error and
 Plasticitycomputationsin fmitc clement analysis 177

Example 2 edge. Displacement control loading enables con-

A finite element model of a single edge notch tinuation of the analysis beyond limit load. The finite
(SEN) tensile panel is shown in Fig. 9. Plasticity element model for this panel consists of 48 iso-
analyses of such problems are frequently performed parametric elements (each with eight nodes). At the
in fracture mechanics research. Accurate predictions crack tip, four eight-node elements are degenerated
of strains in the panel are essential for determination into triangles by coalescing the nodes common to one
of the fracture parameters. Efficient non-linear sol- edge (see Fig. 9b). This forces the strain field on all
utions that converge rapidly are necessary to mini- rays focusing at the crack tip to exhibit the l/r
mize the costs of parameter studies. singularity characteristic of an elastic-perfectly plas-
For thin panels with crack lengths, a/w > z 0.05, tic material. The collapsed nodes are uncoupled to
a two-dimensional plane stress idealization ade- permit stretching at the crack tip.
quately models the behavior[20]. The panel is loaded In recent fracture mechanics studies[24], plane
by displacement increments imposed across the top stress analyses using the continuum tangent operator
with the path-independent updating strategy ex-
hibited poor convergence. Use of the consistent tan-
gent operator restores excellent convergence rates for
these problems. The path-independent strategy is
dictated by the displacement control loading; the
path-dependent strategy leads immediately to non-
convergent loading-unloading cycles.
To demonstrate the convergence rates, the model
shown in Fig. 9 was analyzed using the continuum
and consistent tangent operators. The material prop-
erties for analysis shown in the figure are representa-
tive of an X70 line-pipe steel. Fifteen, variable size
displacement increments were applied. Iterations at
each step were performed until the convergence test
given by

01 Element Grad IIR II s IIP II* lo-’ (52)

was satisfied, where II 1)denotes the Euclidean norm,

R is the residual load vector, and P is the vector of

Crack

b) Crack TIP Detoll

Lmd St* 4

‘\\ --corltkun
- consiitnt
\\

cl T@nsllr strrw-strain Curvr for x-70

Uwd in Eaamplr ProMom 2

Fig. 9. Data for exampleproblem 2. Fig. 10. Convcrgcnce curves for example 2.
778 ROBERTH. DODDS, JR

Table 4. Iterations required for convergence in example 2 or ad hoc scaling of the updated stress vector. Mixed
A
isotropic-kinematic hardening is easily included.
Load step (in. * 10)) Continuum Consistent For implicit solutions, the path-independent
strategy for stress updating over a time (load) step is
1 1.7 1 1
2 2.5 14 4 essential to prevent spurious numerical unloading.
3 3.3 13 4 Sub-incrementation of the total strain increment vec-
4 4.1 >I5 4 tor for the step considerably improves the accuracy.
5 4.6 14 4 To maintain optimal convergence of the Newton
6 5.1 12 3
I 5.6 13 3 solution, the consistent tangent operator for the
8 6.1 10 3 elastic predictor-radial return algorithm must be
9 6.6 13 3 used to form the structure tangent stiffness.
10 1.6 >15 5
I1 8.6 215 5
12 9.0 >15 3 Acknowledgements-This work was partially sponsored by
13 9.8 >15 3 the National Bureau of Standards, Fracture and Defor-
14 10.6 >15 4 mation Division under contract NB83R4030012. Com-
15 11.4 13 4 putations were performed on the Harris 1000 computer
operated by the School of Engineering at the University
of Kansas. The donation of this computer by the Harris
Corporation is gratefully acknowledged.
reactions at constrained nodes. The tolerance was
selected to insure convergence of the strains to three
REFERENCES
significant figures. Table 4 lists the number of iter-
ations required for convergence at each load step. 1. J. C. Simo and R. L. Taylor, Consistent tangent oper-
The structure stiffness was updated before each iter- ators for rate-independent elastoplasticity. Compuf.
ation. The performance of the consistent tangent Mefh. Appl. Mech. Engng 48, 1101-1118 (1985).
2. R. P. Goel and L. E. Malvem, Biaxial plastic simple
operator is much superior, never requiring more than waves with combined kinematic and isotropic hard-
five iterations. Convergence curves for two load steps ening. Trans. ASME., J. Appl. Mech. 92, 1100-1106
and the corresponding extent of the plastic zones are (1970).
shown in Fig. 10. 3. Y. Yamada, N. Yoshimura and T. Sakurai, Plastic
The load increment size has negligible effect on the stress-strain matrix and its application for the solution
of elastic-plastic problems by the finite element method.
computed strains for this model. At Gauss points in Inl. J. Mech. Sri. 10, 343-354 (1968).
the highly deformed shear band (see Fig. lo), the 4. M. Oritiz and E. P. Popov, Accuracy and stability of
strain paths are nearly linear. The analysis was integration algorithms for elastoplastic constitutive re-
repeated with 60 load steps. The strains changed less lations. Inr. J. Numer. Merh. Engng 21, 1561-1576
(1985).
than 1%. 5. D. M. Tracey and C. Freese, Adaptive load m-
crementation in elasto-plastic finite element analysis,
SUMMARY AND CONCLUSIONS Comput. Struct. 13, 45-53 (1981).
6. B. Hunsaker, W. E. Haisler and J. A. Stricklin, On the
Numerical techniques for plasticity computations use of two hardening rules of plasticity in incremental
and pseudo force analysis. In Constitutive Equations in
in finite element analysis have been examined. The Viscoplasticity: Computational and Engineering Aspects,
widely used von Mises yield function with mixed AMD-20 (Edited by J. A. Stricklin and K. J. Sac&ski),
isotropic-kinematic hardening provided the specific pp. 139-170. ASME, New York (1976).
constitutive relations for study. The discussion 7. J. C. Naatenaal and J. E. de Jona. Some comuutational
aspects if &tic-plastic, large strain analysis. Inr. J.
focused on: (1) integration of the plasticity rate
Numer. Mefh. Engng 17, 15-41 (1981).
equations; (2) strain paths for stress updating in 8. R. D. Krieg and D. B. Krieg, Accuracies of numerical
implicit codes; and (3) consistent tangent operators to solutions for the elastic-perfectly plastic model. Trans.
maintain optimal convergence of Newton-type sol- ASME, J. Press. Vessel Tech. 99, 510-515 (1977).
9. H. L. Schreyer, R. F. Kulak and M. M. Kramer,
utions. An extension of the elastic predictor-radial
Accurate numerical solutions for elastic-plastic models.
return algorithm was presented for solution of plane Trans. ASME, J. Press. Vessel Tech. 101, 226-234
stress models with mixed isotropic-kinematic hard- (1979).
ening. Sufficient details of the algorithms were 10. D. Bushnell, A strategy for the solution of problems
presented to enable rapid implementation into exist- involving large deflections, plasticity, and creep. Int. J.
Numer. Mel. &gng 11, 683-708 (1977).
ing plasticity routines. 11. C. Nyssen, An efficient and accurate iterative method,
The elastic predictor-radial return algorithm ahowing, large incremental steps to solve elasto-plastic
provides a superior basis on which to construct a problems. Cornput. Srrucr. 13, 63-71 (1981).
numerical plasticity model. Conceptually simple, the 12. R. D. Krieg and S. W. Key, Implementation of a time
algorithm provides the most accuracy for both single independent plasticity theory into structural computer
programs. In Constitutive Equations in Vircoplarticity:
step and sub-incrementation schemes. The algorithm Computational and Engineering Aspects, AMD-20 (Bdi-
is unconditionally stable for large strain increments; ted by J. A. Stricklin and K. J. Sacxalski), pp. 125138.
it does not require computation of the contact stress ASME, New York (1976).
 Plasticity computations in finite element analysis 119

13. D. P. Mondkar and G. H. Powell, Evaluation of state 19. G. L. Goudreau and J. 0. Hallquist, Recent develop
determination calculations in nonlinear analysis. Trans ments in large-scale finite element Lagrangian hydro-
4th Int. Con/. Siruct. Mech. React. Tech. (Edited by code technology. Compur. Merh. Appl. Mech. Engng 33,
T. A. Jaeger and B. A. Bolcy), L2/5 (1977). 725-757 (1982).
14. J. R. Rice and D. M. Tracey, Computational fracture 20. S. S. Ibrahim and R. H. Dodds, Finite element study of
mechanics. In Nwnerical h4ithodr G Structural Mech- highly deformed tensile panels with short cracks. Srruc-
anics (Edited bv S. J. Fenves ef al.). DD. 585-623. rural Engineering and Engineering Materiab, SM Rep1
Academic Press,~New York (1973). ” -’ 10, University of Kansas Center for Research, Inc.
15. J. C. Nagtcgaal, On the implementation of inelastic (1984).
constitutive equations with special reference to large 21. R. Hill, The MathematicalTheory o/Piasricity. Oxford
deformation problems. Combur. Meth. Appl. Me&. University Press, London (1950).
EWIP 33.469-484 (19821. 22. R. H. Dodds and L. A. Lopez, A generalized software
16. M.-c Wilkins, Calculation of elastic-plastic flow. In system for nonlinear analysis. Inf. /. Adu. Engng Sofr-
Methods of ComputationalPhysics (Edited by B. Alder ware 2, 161-168 (1980).
er al.), Vol. 3. Academic Press, New York (1964). 23. R. H. Dodds, D. R. Rehak and L. A. Lopez, Devel-
17. A. Mendelson, Plasticity:Theory and Application.Mac- opment methodologies for scientific software. ht. J.
millan, New York (1968). So/rware Pracr. Exp. 12, 1085-l 100 (1982).
18. S. W. Key er al., HONDO II-a finite element computer 24. R. H. Dodds and D. T. Read, Experimental and
program for the large deformation dynamic response analytical estimates of the J-integral for tensile panels
of axisymmetric solids. SAND78-0422, Sandia Labora- containing short center cracks. Inr. J. Fracf. 28, 3954
tories, Albuquerque, NM (1977). (1985).